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 "super.h"
  31 #include "super-io.h"
  32
  33 #include <linux/blkdev.h>
  34 #include <linux/random.h>
  35 #include <linux/sched/mm.h>
  36
  37 #include <trace/events/bcachefs.h>
  38
  39 const char *bch2_blk_status_to_str(blk_status_t status)
  40 {
  41         if (status == BLK_STS_REMOVED)
  42                 return "device removed";
  43         return blk_status_to_str(status);
  44 }
  45
  46 static bool bch2_target_congested(struct bch_fs *c, u16 target)
  47 {
  48         const struct bch_devs_mask *devs;
  49         unsigned d, nr = 0, total = 0;
  50         u64 now = local_clock(), last;
  51         s64 congested;
  52         struct bch_dev *ca;
  53
  54         if (!target)
  55                 return false;
  56
  57         rcu_read_lock();
  58         devs = bch2_target_to_mask(c, target) ?:
  59                 &c->rw_devs[BCH_DATA_user];
  60
  61         for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
  62                 ca = rcu_dereference(c->devs[d]);
  63                 if (!ca)
  64                         continue;
  65
  66                 congested = atomic_read(&ca->congested);
  67                 last = READ_ONCE(ca->congested_last);
  68                 if (time_after64(now, last))
  69                         congested -= (now - last) >> 12;
  70
  71                 total += max(congested, 0LL);
  72                 nr++;
  73         }
  74         rcu_read_unlock();
  75
  76         return bch2_rand_range(nr * CONGESTED_MAX) < total;
  77 }
  78
  79 static inline void bch2_congested_acct(struct bch_dev *ca, u64 io_latency,
  80                                        u64 now, int rw)
  81 {
  82         u64 latency_capable =
  83                 ca->io_latency[rw].quantiles.entries[QUANTILE_IDX(1)].m;
  84         /* ideally we'd be taking into account the device's variance here: */
  85         u64 latency_threshold = latency_capable << (rw == READ ? 2 : 3);
  86         s64 latency_over = io_latency - latency_threshold;
  87
  88         if (latency_threshold && latency_over > 0) {
  89                 /*
  90                  * bump up congested by approximately latency_over * 4 /
  91                  * latency_threshold - we don't need much accuracy here so don't
  92                  * bother with the divide:
  93                  */
  94                 if (atomic_read(&ca->congested) < CONGESTED_MAX)
  95                         atomic_add(latency_over >>
  96                                    max_t(int, ilog2(latency_threshold) - 2, 0),
  97                                    &ca->congested);
  98
  99                 ca->congested_last = now;
 100         } else if (atomic_read(&ca->congested) > 0) {
 101                 atomic_dec(&ca->congested);
 102         }
 103 }
 104
 105 void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
 106 {
 107         atomic64_t *latency = &ca->cur_latency[rw];
 108         u64 now = local_clock();
 109         u64 io_latency = time_after64(now, submit_time)
 110                 ? now - submit_time
 111                 : 0;
 112         u64 old, new, v = atomic64_read(latency);
 113
 114         do {
 115                 old = v;
 116
 117                 /*
 118                  * If the io latency was reasonably close to the current
 119                  * latency, skip doing the update and atomic operation - most of
 120                  * the time:
 121                  */
 122                 if (abs((int) (old - io_latency)) < (old >> 1) &&
 123                     now & ~(~0U << 5))
 124                         break;
 125
 126                 new = ewma_add(old, io_latency, 5);
 127         } while ((v = atomic64_cmpxchg(latency, old, new)) != old);
 128
 129         bch2_congested_acct(ca, io_latency, now, rw);
 130
 131         __bch2_time_stats_update(&ca->io_latency[rw], submit_time, now);
 132 }
 133
 134 /* Allocate, free from mempool: */
 135
 136 void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
 137 {
 138         struct bvec_iter_all iter;
 139         struct bio_vec *bv;
 140
 141         bio_for_each_segment_all(bv, bio, iter)
 142                 if (bv->bv_page != ZERO_PAGE(0))
 143                         mempool_free(bv->bv_page, &c->bio_bounce_pages);
 144         bio->bi_vcnt = 0;
 145 }
 146
 147 static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool)
 148 {
 149         struct page *page;
 150
 151         if (likely(!*using_mempool)) {
 152                 page = alloc_page(GFP_NOIO);
 153                 if (unlikely(!page)) {
 154                         mutex_lock(&c->bio_bounce_pages_lock);
 155                         *using_mempool = true;
 156                         goto pool_alloc;
 157
 158                 }
 159         } else {
 160 pool_alloc:
 161                 page = mempool_alloc(&c->bio_bounce_pages, GFP_NOIO);
 162         }
 163
 164         return page;
 165 }
 166
 167 void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
 168                                size_t size)
 169 {
 170         bool using_mempool = false;
 171
 172         while (size) {
 173                 struct page *page = __bio_alloc_page_pool(c, &using_mempool);
 174                 unsigned len = min_t(size_t, PAGE_SIZE, size);
 175
 176                 BUG_ON(!bio_add_page(bio, page, len, 0));
 177                 size -= len;
 178         }
 179
 180         if (using_mempool)
 181                 mutex_unlock(&c->bio_bounce_pages_lock);
 182 }
 183
 184 /* Extent update path: */
 185
 186 int bch2_sum_sector_overwrites(struct btree_trans *trans,
 187                                struct btree_iter *extent_iter,
 188                                struct bkey_i *new,
 189                                bool *maybe_extending,
 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         *maybe_extending        = true;
 202         *usage_increasing       = false;
 203         *i_sectors_delta        = 0;
 204         *disk_sectors_delta     = 0;
 205
 206         iter = bch2_trans_copy_iter(trans, extent_iter);
 207
 208         for_each_btree_key_continue(iter, BTREE_ITER_SLOTS, old, ret) {
 209                 s64 sectors = min(new->k.p.offset, old.k->p.offset) -
 210                         max(bkey_start_offset(&new->k),
 211                             bkey_start_offset(old.k));
 212
 213                 *i_sectors_delta += sectors *
 214                         (bkey_extent_is_allocation(&new->k) -
 215                          bkey_extent_is_allocation(old.k));
 216
 217                 *disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new));
 218                 *disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot
 219                         ? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
 220                         : 0;
 221
 222                 if (!*usage_increasing &&
 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                         /*
 229                          * Check if there's already data above where we're
 230                          * going to be writing to - this means we're definitely
 231                          * not extending the file:
 232                          *
 233                          * Note that it's not sufficient to check if there's
 234                          * data up to the sector offset we're going to be
 235                          * writing to, because i_size could be up to one block
 236                          * less:
 237                          */
 238                         if (!bkey_cmp(old.k->p, new->k.p))
 239                                 old = bch2_btree_iter_next(iter);
 240
 241                         if (old.k && !bkey_err(old) &&
 242                             old.k->p.inode == extent_iter->pos.inode &&
 243                             bkey_extent_is_data(old.k))
 244                                 *maybe_extending = false;
 245
 246                         break;
 247                 }
 248         }
 249
 250         bch2_trans_iter_put(trans, iter);
 251         return ret;
 252 }
 253
 254 int bch2_extent_update(struct btree_trans *trans,
 255                        struct btree_iter *iter,
 256                        struct bkey_i *k,
 257                        struct disk_reservation *disk_res,
 258                        u64 *journal_seq,
 259                        u64 new_i_size,
 260                        s64 *i_sectors_delta_total,
 261                        bool check_enospc)
 262 {
 263         /* this must live until after bch2_trans_commit(): */
 264         struct bkey_inode_buf inode_p;
 265         bool extending = false, usage_increasing;
 266         s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 267         int ret;
 268
 269         ret = bch2_extent_trim_atomic(k, iter);
 270         if (ret)
 271                 return ret;
 272
 273         ret = bch2_sum_sector_overwrites(trans, iter, k,
 274                         &extending,
 275                         &usage_increasing,
 276                         &i_sectors_delta,
 277                         &disk_sectors_delta);
 278         if (ret)
 279                 return ret;
 280
 281         if (!usage_increasing)
 282                 check_enospc = false;
 283
 284         if (disk_res &&
 285             disk_sectors_delta > (s64) disk_res->sectors) {
 286                 ret = bch2_disk_reservation_add(trans->c, disk_res,
 287                                         disk_sectors_delta - disk_res->sectors,
 288                                         !check_enospc
 289                                         ? BCH_DISK_RESERVATION_NOFAIL : 0);
 290                 if (ret)
 291                         return ret;
 292         }
 293
 294         new_i_size = extending
 295                 ? min(k->k.p.offset << 9, new_i_size)
 296                 : 0;
 297
 298         if (i_sectors_delta || new_i_size) {
 299                 struct btree_iter *inode_iter;
 300                 struct bch_inode_unpacked inode_u;
 301
 302                 inode_iter = bch2_inode_peek(trans, &inode_u,
 303                                 k->k.p.inode, BTREE_ITER_INTENT);
 304                 ret = PTR_ERR_OR_ZERO(inode_iter);
 305                 if (ret)
 306                         return ret;
 307
 308                 /*
 309                  * XXX:
 310                  * writeback can race a bit with truncate, because truncate
 311                  * first updates the inode then truncates the pagecache. This is
 312                  * ugly, but lets us preserve the invariant that the in memory
 313                  * i_size is always >= the on disk i_size.
 314                  *
 315                 BUG_ON(new_i_size > inode_u.bi_size &&
 316                        (inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY));
 317                  */
 318                 BUG_ON(new_i_size > inode_u.bi_size && !extending);
 319
 320                 if (!(inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY) &&
 321                     new_i_size > inode_u.bi_size)
 322                         inode_u.bi_size = new_i_size;
 323                 else
 324                         new_i_size = 0;
 325
 326                 inode_u.bi_sectors += i_sectors_delta;
 327
 328                 if (i_sectors_delta || new_i_size) {
 329                         bch2_inode_pack(trans->c, &inode_p, &inode_u);
 330
 331                         inode_p.inode.k.p.snapshot = iter->snapshot;
 332
 333                         ret = bch2_trans_update(trans, inode_iter,
 334                                           &inode_p.inode.k_i, 0);
 335                 }
 336
 337                 bch2_trans_iter_put(trans, inode_iter);
 338
 339                 if (ret)
 340                         return ret;
 341         }
 342
 343         ret =   bch2_trans_update(trans, iter, k, 0) ?:
 344                 bch2_trans_commit(trans, disk_res, journal_seq,
 345                                 BTREE_INSERT_NOCHECK_RW|
 346                                 BTREE_INSERT_NOFAIL);
 347         BUG_ON(ret == -ENOSPC);
 348         if (ret)
 349                 return ret;
 350
 351         if (i_sectors_delta_total)
 352                 *i_sectors_delta_total += i_sectors_delta;
 353         return 0;
 354 }
 355
 356 int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
 357                    struct bpos end, u64 *journal_seq,
 358                    s64 *i_sectors_delta)
 359 {
 360         struct bch_fs *c        = trans->c;
 361         unsigned max_sectors    = KEY_SIZE_MAX & (~0 << c->block_bits);
 362         struct bkey_s_c k;
 363         int ret = 0, ret2 = 0;
 364
 365         while ((k = bch2_btree_iter_peek(iter)).k &&
 366                bkey_cmp(iter->pos, end) < 0) {
 367                 struct disk_reservation disk_res =
 368                         bch2_disk_reservation_init(c, 0);
 369                 struct bkey_i delete;
 370
 371                 bch2_trans_begin(trans);
 372
 373                 ret = bkey_err(k);
 374                 if (ret)
 375                         goto btree_err;
 376
 377                 bkey_init(&delete.k);
 378                 delete.k.p = iter->pos;
 379
 380                 /* create the biggest key we can */
 381                 bch2_key_resize(&delete.k, max_sectors);
 382                 bch2_cut_back(end, &delete);
 383
 384                 ret = bch2_extent_update(trans, iter, &delete,
 385                                 &disk_res, journal_seq,
 386                                 0, i_sectors_delta, false);
 387                 bch2_disk_reservation_put(c, &disk_res);
 388 btree_err:
 389                 if (ret == -EINTR) {
 390                         ret2 = ret;
 391                         ret = 0;
 392                 }
 393                 if (ret)
 394                         break;
 395         }
 396
 397         if (bkey_cmp(iter->pos, end) > 0) {
 398                 bch2_btree_iter_set_pos(iter, end);
 399                 ret = bch2_btree_iter_traverse(iter);
 400         }
 401
 402         return ret ?: ret2;
 403 }
 404
 405 int bch2_fpunch(struct bch_fs *c, u64 inum, u64 start, u64 end,
 406                 u64 *journal_seq, s64 *i_sectors_delta)
 407 {
 408         struct btree_trans trans;
 409         struct btree_iter *iter;
 410         int ret = 0;
 411
 412         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 413         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents,
 414                                    POS(inum, start),
 415                                    BTREE_ITER_INTENT);
 416
 417         ret = bch2_fpunch_at(&trans, iter, POS(inum, end),
 418                              journal_seq, i_sectors_delta);
 419
 420         bch2_trans_iter_put(&trans, iter);
 421         bch2_trans_exit(&trans);
 422
 423         if (ret == -EINTR)
 424                 ret = 0;
 425
 426         return ret;
 427 }
 428
 429 int bch2_write_index_default(struct bch_write_op *op)
 430 {
 431         struct bch_fs *c = op->c;
 432         struct bkey_buf sk;
 433         struct keylist *keys = &op->insert_keys;
 434         struct bkey_i *k = bch2_keylist_front(keys);
 435         struct btree_trans trans;
 436         struct btree_iter *iter;
 437         int ret;
 438
 439         bch2_bkey_buf_init(&sk);
 440         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 441
 442         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents,
 443                                    bkey_start_pos(&k->k),
 444                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 445
 446         do {
 447                 bch2_trans_begin(&trans);
 448
 449                 k = bch2_keylist_front(keys);
 450
 451                 k->k.p.snapshot = iter->snapshot;
 452
 453                 bch2_bkey_buf_realloc(&sk, c, k->k.u64s);
 454                 bkey_copy(sk.k, k);
 455                 bch2_cut_front(iter->pos, sk.k);
 456
 457                 ret = bch2_extent_update(&trans, iter, sk.k,
 458                                          &op->res, op_journal_seq(op),
 459                                          op->new_i_size, &op->i_sectors_delta,
 460                                          op->flags & BCH_WRITE_CHECK_ENOSPC);
 461                 if (ret == -EINTR)
 462                         continue;
 463                 if (ret)
 464                         break;
 465
 466                 if (bkey_cmp(iter->pos, k->k.p) >= 0)
 467                         bch2_keylist_pop_front(keys);
 468         } while (!bch2_keylist_empty(keys));
 469
 470         bch2_trans_iter_put(&trans, iter);
 471         bch2_trans_exit(&trans);
 472         bch2_bkey_buf_exit(&sk, c);
 473
 474         return ret;
 475 }
 476
 477 /* Writes */
 478
 479 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
 480                                enum bch_data_type type,
 481                                const struct bkey_i *k)
 482 {
 483         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
 484         const struct bch_extent_ptr *ptr;
 485         struct bch_write_bio *n;
 486         struct bch_dev *ca;
 487
 488         BUG_ON(c->opts.nochanges);
 489
 490         bkey_for_each_ptr(ptrs, ptr) {
 491                 BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX ||
 492                        !c->devs[ptr->dev]);
 493
 494                 ca = bch_dev_bkey_exists(c, ptr->dev);
 495
 496                 if (to_entry(ptr + 1) < ptrs.end) {
 497                         n = to_wbio(bio_clone_fast(&wbio->bio, GFP_NOIO,
 498                                                    &ca->replica_set));
 499
 500                         n->bio.bi_end_io        = wbio->bio.bi_end_io;
 501                         n->bio.bi_private       = wbio->bio.bi_private;
 502                         n->parent               = wbio;
 503                         n->split                = true;
 504                         n->bounce               = false;
 505                         n->put_bio              = true;
 506                         n->bio.bi_opf           = wbio->bio.bi_opf;
 507                         bio_inc_remaining(&wbio->bio);
 508                 } else {
 509                         n = wbio;
 510                         n->split                = false;
 511                 }
 512
 513                 n->c                    = c;
 514                 n->dev                  = ptr->dev;
 515                 n->have_ioref           = bch2_dev_get_ioref(ca,
 516                                         type == BCH_DATA_btree ? READ : WRITE);
 517                 n->submit_time          = local_clock();
 518                 n->bio.bi_iter.bi_sector = ptr->offset;
 519
 520                 if (likely(n->have_ioref)) {
 521                         this_cpu_add(ca->io_done->sectors[WRITE][type],
 522                                      bio_sectors(&n->bio));
 523
 524                         bio_set_dev(&n->bio, ca->disk_sb.bdev);
 525                         submit_bio(&n->bio);
 526                 } else {
 527                         n->bio.bi_status        = BLK_STS_REMOVED;
 528                         bio_endio(&n->bio);
 529                 }
 530         }
 531 }
 532
 533 static void __bch2_write(struct closure *);
 534
 535 static void bch2_write_done(struct closure *cl)
 536 {
 537         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
 538         struct bch_fs *c = op->c;
 539
 540         if (!op->error && (op->flags & BCH_WRITE_FLUSH))
 541                 op->error = bch2_journal_error(&c->journal);
 542
 543         bch2_disk_reservation_put(c, &op->res);
 544         percpu_ref_put(&c->writes);
 545         bch2_keylist_free(&op->insert_keys, op->inline_keys);
 546
 547         bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
 548
 549         if (op->end_io) {
 550                 EBUG_ON(cl->parent);
 551                 closure_debug_destroy(cl);
 552                 op->end_io(op);
 553         } else {
 554                 closure_return(cl);
 555         }
 556 }
 557
 558 /**
 559  * bch_write_index - after a write, update index to point to new data
 560  */
 561 static void __bch2_write_index(struct bch_write_op *op)
 562 {
 563         struct bch_fs *c = op->c;
 564         struct keylist *keys = &op->insert_keys;
 565         struct bch_extent_ptr *ptr;
 566         struct bkey_i *src, *dst = keys->keys, *n, *k;
 567         unsigned dev;
 568         int ret;
 569
 570         for (src = keys->keys; src != keys->top; src = n) {
 571                 n = bkey_next(src);
 572
 573                 if (bkey_extent_is_direct_data(&src->k)) {
 574                         bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
 575                                             test_bit(ptr->dev, op->failed.d));
 576
 577                         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src))) {
 578                                 ret = -EIO;
 579                                 goto err;
 580                         }
 581                 }
 582
 583                 if (dst != src)
 584                         memmove_u64s_down(dst, src, src->u64s);
 585                 dst = bkey_next(dst);
 586         }
 587
 588         keys->top = dst;
 589
 590         /*
 591          * probably not the ideal place to hook this in, but I don't
 592          * particularly want to plumb io_opts all the way through the btree
 593          * update stack right now
 594          */
 595         for_each_keylist_key(keys, k) {
 596                 bch2_rebalance_add_key(c, bkey_i_to_s_c(k), &op->opts);
 597
 598                 if (bch2_bkey_is_incompressible(bkey_i_to_s_c(k)))
 599                         bch2_check_set_feature(op->c, BCH_FEATURE_incompressible);
 600
 601         }
 602
 603         if (!bch2_keylist_empty(keys)) {
 604                 u64 sectors_start = keylist_sectors(keys);
 605                 int ret = op->index_update_fn(op);
 606
 607                 BUG_ON(ret == -EINTR);
 608                 BUG_ON(keylist_sectors(keys) && !ret);
 609
 610                 op->written += sectors_start - keylist_sectors(keys);
 611
 612                 if (ret) {
 613                         bch_err_inum_ratelimited(c, op->pos.inode,
 614                                 "write error %i from btree update", ret);
 615                         op->error = ret;
 616                 }
 617         }
 618 out:
 619         /* If some a bucket wasn't written, we can't erasure code it: */
 620         for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
 621                 bch2_open_bucket_write_error(c, &op->open_buckets, dev);
 622
 623         bch2_open_buckets_put(c, &op->open_buckets);
 624         return;
 625 err:
 626         keys->top = keys->keys;
 627         op->error = ret;
 628         goto out;
 629 }
 630
 631 static void bch2_write_index(struct closure *cl)
 632 {
 633         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
 634         struct bch_fs *c = op->c;
 635
 636         __bch2_write_index(op);
 637
 638         if (!(op->flags & BCH_WRITE_DONE)) {
 639                 continue_at(cl, __bch2_write, index_update_wq(op));
 640         } else if (!op->error && (op->flags & BCH_WRITE_FLUSH)) {
 641                 bch2_journal_flush_seq_async(&c->journal,
 642                                              *op_journal_seq(op),
 643                                              cl);
 644                 continue_at(cl, bch2_write_done, index_update_wq(op));
 645         } else {
 646                 continue_at_nobarrier(cl, bch2_write_done, NULL);
 647         }
 648 }
 649
 650 static void bch2_write_endio(struct bio *bio)
 651 {
 652         struct closure *cl              = bio->bi_private;
 653         struct bch_write_op *op         = container_of(cl, struct bch_write_op, cl);
 654         struct bch_write_bio *wbio      = to_wbio(bio);
 655         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
 656         struct bch_fs *c                = wbio->c;
 657         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
 658
 659         if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
 660                                     op->pos.inode,
 661                                     op->pos.offset - bio_sectors(bio), /* XXX definitely wrong */
 662                                     "data write error: %s",
 663                                bch2_blk_status_to_str(bio->bi_status)))
 664                 set_bit(wbio->dev, op->failed.d);
 665
 666         if (wbio->have_ioref) {
 667                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
 668                 percpu_ref_put(&ca->io_ref);
 669         }
 670
 671         if (wbio->bounce)
 672                 bch2_bio_free_pages_pool(c, bio);
 673
 674         if (wbio->put_bio)
 675                 bio_put(bio);
 676
 677         if (parent)
 678                 bio_endio(&parent->bio);
 679         else if (!(op->flags & BCH_WRITE_SKIP_CLOSURE_PUT))
 680                 closure_put(cl);
 681         else
 682                 continue_at_nobarrier(cl, bch2_write_index, index_update_wq(op));
 683 }
 684
 685 static void init_append_extent(struct bch_write_op *op,
 686                                struct write_point *wp,
 687                                struct bversion version,
 688                                struct bch_extent_crc_unpacked crc)
 689 {
 690         struct bch_fs *c = op->c;
 691         struct bkey_i_extent *e;
 692         struct open_bucket *ob;
 693         unsigned i;
 694
 695         BUG_ON(crc.compressed_size > wp->sectors_free);
 696         wp->sectors_free -= crc.compressed_size;
 697         op->pos.offset += crc.uncompressed_size;
 698
 699         e = bkey_extent_init(op->insert_keys.top);
 700         e->k.p          = op->pos;
 701         e->k.size       = crc.uncompressed_size;
 702         e->k.version    = version;
 703
 704         if (crc.csum_type ||
 705             crc.compression_type ||
 706             crc.nonce)
 707                 bch2_extent_crc_append(&e->k_i, crc);
 708
 709         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 710                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 711                 union bch_extent_entry *end =
 712                         bkey_val_end(bkey_i_to_s(&e->k_i));
 713
 714                 end->ptr = ob->ptr;
 715                 end->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
 716                 end->ptr.cached = !ca->mi.durability ||
 717                         (op->flags & BCH_WRITE_CACHED) != 0;
 718                 end->ptr.offset += ca->mi.bucket_size - ob->sectors_free;
 719
 720                 e->k.u64s++;
 721
 722                 BUG_ON(crc.compressed_size > ob->sectors_free);
 723                 ob->sectors_free -= crc.compressed_size;
 724         }
 725
 726         bch2_keylist_push(&op->insert_keys);
 727 }
 728
 729 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
 730                                         struct write_point *wp,
 731                                         struct bio *src,
 732                                         bool *page_alloc_failed,
 733                                         void *buf)
 734 {
 735         struct bch_write_bio *wbio;
 736         struct bio *bio;
 737         unsigned output_available =
 738                 min(wp->sectors_free << 9, src->bi_iter.bi_size);
 739         unsigned pages = DIV_ROUND_UP(output_available +
 740                                       (buf
 741                                        ? ((unsigned long) buf & (PAGE_SIZE - 1))
 742                                        : 0), PAGE_SIZE);
 743
 744         bio = bio_alloc_bioset(GFP_NOIO, pages, &c->bio_write);
 745         wbio                    = wbio_init(bio);
 746         wbio->put_bio           = true;
 747         /* copy WRITE_SYNC flag */
 748         wbio->bio.bi_opf        = src->bi_opf;
 749
 750         if (buf) {
 751                 bch2_bio_map(bio, buf, output_available);
 752                 return bio;
 753         }
 754
 755         wbio->bounce            = true;
 756
 757         /*
 758          * We can't use mempool for more than c->sb.encoded_extent_max
 759          * worth of pages, but we'd like to allocate more if we can:
 760          */
 761         bch2_bio_alloc_pages_pool(c, bio,
 762                                   min_t(unsigned, output_available,
 763                                         c->sb.encoded_extent_max << 9));
 764
 765         if (bio->bi_iter.bi_size < output_available)
 766                 *page_alloc_failed =
 767                         bch2_bio_alloc_pages(bio,
 768                                              output_available -
 769                                              bio->bi_iter.bi_size,
 770                                              GFP_NOFS) != 0;
 771
 772         return bio;
 773 }
 774
 775 static int bch2_write_rechecksum(struct bch_fs *c,
 776                                  struct bch_write_op *op,
 777                                  unsigned new_csum_type)
 778 {
 779         struct bio *bio = &op->wbio.bio;
 780         struct bch_extent_crc_unpacked new_crc;
 781         int ret;
 782
 783         /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
 784
 785         if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
 786             bch2_csum_type_is_encryption(new_csum_type))
 787                 new_csum_type = op->crc.csum_type;
 788
 789         ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
 790                                   NULL, &new_crc,
 791                                   op->crc.offset, op->crc.live_size,
 792                                   new_csum_type);
 793         if (ret)
 794                 return ret;
 795
 796         bio_advance(bio, op->crc.offset << 9);
 797         bio->bi_iter.bi_size = op->crc.live_size << 9;
 798         op->crc = new_crc;
 799         return 0;
 800 }
 801
 802 static int bch2_write_decrypt(struct bch_write_op *op)
 803 {
 804         struct bch_fs *c = op->c;
 805         struct nonce nonce = extent_nonce(op->version, op->crc);
 806         struct bch_csum csum;
 807
 808         if (!bch2_csum_type_is_encryption(op->crc.csum_type))
 809                 return 0;
 810
 811         /*
 812          * If we need to decrypt data in the write path, we'll no longer be able
 813          * to verify the existing checksum (poly1305 mac, in this case) after
 814          * it's decrypted - this is the last point we'll be able to reverify the
 815          * checksum:
 816          */
 817         csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 818         if (bch2_crc_cmp(op->crc.csum, csum))
 819                 return -EIO;
 820
 821         bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 822         op->crc.csum_type = 0;
 823         op->crc.csum = (struct bch_csum) { 0, 0 };
 824         return 0;
 825 }
 826
 827 static enum prep_encoded_ret {
 828         PREP_ENCODED_OK,
 829         PREP_ENCODED_ERR,
 830         PREP_ENCODED_CHECKSUM_ERR,
 831         PREP_ENCODED_DO_WRITE,
 832 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
 833 {
 834         struct bch_fs *c = op->c;
 835         struct bio *bio = &op->wbio.bio;
 836
 837         if (!(op->flags & BCH_WRITE_DATA_ENCODED))
 838                 return PREP_ENCODED_OK;
 839
 840         BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
 841
 842         /* Can we just write the entire extent as is? */
 843         if (op->crc.uncompressed_size == op->crc.live_size &&
 844             op->crc.compressed_size <= wp->sectors_free &&
 845             (op->crc.compression_type == op->compression_type ||
 846              op->incompressible)) {
 847                 if (!crc_is_compressed(op->crc) &&
 848                     op->csum_type != op->crc.csum_type &&
 849                     bch2_write_rechecksum(c, op, op->csum_type))
 850                         return PREP_ENCODED_CHECKSUM_ERR;
 851
 852                 return PREP_ENCODED_DO_WRITE;
 853         }
 854
 855         /*
 856          * If the data is compressed and we couldn't write the entire extent as
 857          * is, we have to decompress it:
 858          */
 859         if (crc_is_compressed(op->crc)) {
 860                 struct bch_csum csum;
 861
 862                 if (bch2_write_decrypt(op))
 863                         return PREP_ENCODED_CHECKSUM_ERR;
 864
 865                 /* Last point we can still verify checksum: */
 866                 csum = bch2_checksum_bio(c, op->crc.csum_type,
 867                                          extent_nonce(op->version, op->crc),
 868                                          bio);
 869                 if (bch2_crc_cmp(op->crc.csum, csum))
 870                         return PREP_ENCODED_CHECKSUM_ERR;
 871
 872                 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
 873                         return PREP_ENCODED_ERR;
 874         }
 875
 876         /*
 877          * No longer have compressed data after this point - data might be
 878          * encrypted:
 879          */
 880
 881         /*
 882          * If the data is checksummed and we're only writing a subset,
 883          * rechecksum and adjust bio to point to currently live data:
 884          */
 885         if ((op->crc.live_size != op->crc.uncompressed_size ||
 886              op->crc.csum_type != op->csum_type) &&
 887             bch2_write_rechecksum(c, op, op->csum_type))
 888                 return PREP_ENCODED_CHECKSUM_ERR;
 889
 890         /*
 891          * If we want to compress the data, it has to be decrypted:
 892          */
 893         if ((op->compression_type ||
 894              bch2_csum_type_is_encryption(op->crc.csum_type) !=
 895              bch2_csum_type_is_encryption(op->csum_type)) &&
 896             bch2_write_decrypt(op))
 897                 return PREP_ENCODED_CHECKSUM_ERR;
 898
 899         return PREP_ENCODED_OK;
 900 }
 901
 902 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 903                              struct bio **_dst)
 904 {
 905         struct bch_fs *c = op->c;
 906         struct bio *src = &op->wbio.bio, *dst = src;
 907         struct bvec_iter saved_iter;
 908         void *ec_buf;
 909         struct bpos ec_pos = op->pos;
 910         unsigned total_output = 0, total_input = 0;
 911         bool bounce = false;
 912         bool page_alloc_failed = false;
 913         int ret, more = 0;
 914
 915         BUG_ON(!bio_sectors(src));
 916
 917         ec_buf = bch2_writepoint_ec_buf(c, wp);
 918
 919         switch (bch2_write_prep_encoded_data(op, wp)) {
 920         case PREP_ENCODED_OK:
 921                 break;
 922         case PREP_ENCODED_ERR:
 923                 ret = -EIO;
 924                 goto err;
 925         case PREP_ENCODED_CHECKSUM_ERR:
 926                 BUG();
 927                 goto csum_err;
 928         case PREP_ENCODED_DO_WRITE:
 929                 /* XXX look for bug here */
 930                 if (ec_buf) {
 931                         dst = bch2_write_bio_alloc(c, wp, src,
 932                                                    &page_alloc_failed,
 933                                                    ec_buf);
 934                         bio_copy_data(dst, src);
 935                         bounce = true;
 936                 }
 937                 init_append_extent(op, wp, op->version, op->crc);
 938                 goto do_write;
 939         }
 940
 941         if (ec_buf ||
 942             op->compression_type ||
 943             (op->csum_type &&
 944              !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
 945             (bch2_csum_type_is_encryption(op->csum_type) &&
 946              !(op->flags & BCH_WRITE_PAGES_OWNED))) {
 947                 dst = bch2_write_bio_alloc(c, wp, src,
 948                                            &page_alloc_failed,
 949                                            ec_buf);
 950                 bounce = true;
 951         }
 952
 953         saved_iter = dst->bi_iter;
 954
 955         do {
 956                 struct bch_extent_crc_unpacked crc =
 957                         (struct bch_extent_crc_unpacked) { 0 };
 958                 struct bversion version = op->version;
 959                 size_t dst_len, src_len;
 960
 961                 if (page_alloc_failed &&
 962                     bio_sectors(dst) < wp->sectors_free &&
 963                     bio_sectors(dst) < c->sb.encoded_extent_max)
 964                         break;
 965
 966                 BUG_ON(op->compression_type &&
 967                        (op->flags & BCH_WRITE_DATA_ENCODED) &&
 968                        bch2_csum_type_is_encryption(op->crc.csum_type));
 969                 BUG_ON(op->compression_type && !bounce);
 970
 971                 crc.compression_type = op->incompressible
 972                         ? BCH_COMPRESSION_TYPE_incompressible
 973                         : op->compression_type
 974                         ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
 975                                             op->compression_type)
 976                         : 0;
 977                 if (!crc_is_compressed(crc)) {
 978                         dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
 979                         dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
 980
 981                         if (op->csum_type)
 982                                 dst_len = min_t(unsigned, dst_len,
 983                                                 c->sb.encoded_extent_max << 9);
 984
 985                         if (bounce) {
 986                                 swap(dst->bi_iter.bi_size, dst_len);
 987                                 bio_copy_data(dst, src);
 988                                 swap(dst->bi_iter.bi_size, dst_len);
 989                         }
 990
 991                         src_len = dst_len;
 992                 }
 993
 994                 BUG_ON(!src_len || !dst_len);
 995
 996                 if (bch2_csum_type_is_encryption(op->csum_type)) {
 997                         if (bversion_zero(version)) {
 998                                 version.lo = atomic64_inc_return(&c->key_version);
 999                         } else {
1000                                 crc.nonce = op->nonce;
1001                                 op->nonce += src_len >> 9;
1002                         }
1003                 }
1004
1005                 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1006                     !crc_is_compressed(crc) &&
1007                     bch2_csum_type_is_encryption(op->crc.csum_type) ==
1008                     bch2_csum_type_is_encryption(op->csum_type)) {
1009                         /*
1010                          * Note: when we're using rechecksum(), we need to be
1011                          * checksumming @src because it has all the data our
1012                          * existing checksum covers - if we bounced (because we
1013                          * were trying to compress), @dst will only have the
1014                          * part of the data the new checksum will cover.
1015                          *
1016                          * But normally we want to be checksumming post bounce,
1017                          * because part of the reason for bouncing is so the
1018                          * data can't be modified (by userspace) while it's in
1019                          * flight.
1020                          */
1021                         if (bch2_rechecksum_bio(c, src, version, op->crc,
1022                                         &crc, &op->crc,
1023                                         src_len >> 9,
1024                                         bio_sectors(src) - (src_len >> 9),
1025                                         op->csum_type))
1026                                 goto csum_err;
1027                 } else {
1028                         if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1029                             bch2_rechecksum_bio(c, src, version, op->crc,
1030                                         NULL, &op->crc,
1031                                         src_len >> 9,
1032                                         bio_sectors(src) - (src_len >> 9),
1033                                         op->crc.csum_type))
1034                                 goto csum_err;
1035
1036                         crc.compressed_size     = dst_len >> 9;
1037                         crc.uncompressed_size   = src_len >> 9;
1038                         crc.live_size           = src_len >> 9;
1039
1040                         swap(dst->bi_iter.bi_size, dst_len);
1041                         bch2_encrypt_bio(c, op->csum_type,
1042                                          extent_nonce(version, crc), dst);
1043                         crc.csum = bch2_checksum_bio(c, op->csum_type,
1044                                          extent_nonce(version, crc), dst);
1045                         crc.csum_type = op->csum_type;
1046                         swap(dst->bi_iter.bi_size, dst_len);
1047                 }
1048
1049                 init_append_extent(op, wp, version, crc);
1050
1051                 if (dst != src)
1052                         bio_advance(dst, dst_len);
1053                 bio_advance(src, src_len);
1054                 total_output    += dst_len;
1055                 total_input     += src_len;
1056         } while (dst->bi_iter.bi_size &&
1057                  src->bi_iter.bi_size &&
1058                  wp->sectors_free &&
1059                  !bch2_keylist_realloc(&op->insert_keys,
1060                                       op->inline_keys,
1061                                       ARRAY_SIZE(op->inline_keys),
1062                                       BKEY_EXTENT_U64s_MAX));
1063
1064         more = src->bi_iter.bi_size != 0;
1065
1066         dst->bi_iter = saved_iter;
1067
1068         if (dst == src && more) {
1069                 BUG_ON(total_output != total_input);
1070
1071                 dst = bio_split(src, total_input >> 9,
1072                                 GFP_NOIO, &c->bio_write);
1073                 wbio_init(dst)->put_bio = true;
1074                 /* copy WRITE_SYNC flag */
1075                 dst->bi_opf             = src->bi_opf;
1076         }
1077
1078         dst->bi_iter.bi_size = total_output;
1079 do_write:
1080         /* might have done a realloc... */
1081         bch2_ec_add_backpointer(c, wp, ec_pos, total_input >> 9);
1082
1083         *_dst = dst;
1084         return more;
1085 csum_err:
1086         bch_err(c, "error verifying existing checksum while "
1087                 "rewriting existing data (memory corruption?)");
1088         ret = -EIO;
1089 err:
1090         if (to_wbio(dst)->bounce)
1091                 bch2_bio_free_pages_pool(c, dst);
1092         if (to_wbio(dst)->put_bio)
1093                 bio_put(dst);
1094
1095         return ret;
1096 }
1097
1098 static void __bch2_write(struct closure *cl)
1099 {
1100         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1101         struct bch_fs *c = op->c;
1102         struct write_point *wp;
1103         struct bio *bio;
1104         bool skip_put = true;
1105         unsigned nofs_flags;
1106         int ret;
1107
1108         nofs_flags = memalloc_nofs_save();
1109 again:
1110         memset(&op->failed, 0, sizeof(op->failed));
1111
1112         do {
1113                 struct bkey_i *key_to_write;
1114                 unsigned key_to_write_offset = op->insert_keys.top_p -
1115                         op->insert_keys.keys_p;
1116
1117                 /* +1 for possible cache device: */
1118                 if (op->open_buckets.nr + op->nr_replicas + 1 >
1119                     ARRAY_SIZE(op->open_buckets.v))
1120                         goto flush_io;
1121
1122                 if (bch2_keylist_realloc(&op->insert_keys,
1123                                         op->inline_keys,
1124                                         ARRAY_SIZE(op->inline_keys),
1125                                         BKEY_EXTENT_U64s_MAX))
1126                         goto flush_io;
1127
1128                 if ((op->flags & BCH_WRITE_FROM_INTERNAL) &&
1129                     percpu_ref_is_dying(&c->writes)) {
1130                         ret = -EROFS;
1131                         goto err;
1132                 }
1133
1134                 /*
1135                  * The copygc thread is now global, which means it's no longer
1136                  * freeing up space on specific disks, which means that
1137                  * allocations for specific disks may hang arbitrarily long:
1138                  */
1139                 wp = bch2_alloc_sectors_start(c,
1140                         op->target,
1141                         op->opts.erasure_code,
1142                         op->write_point,
1143                         &op->devs_have,
1144                         op->nr_replicas,
1145                         op->nr_replicas_required,
1146                         op->alloc_reserve,
1147                         op->flags,
1148                         (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1149                                       BCH_WRITE_ONLY_SPECIFIED_DEVS)) ? NULL : cl);
1150                 EBUG_ON(!wp);
1151
1152                 if (unlikely(IS_ERR(wp))) {
1153                         if (unlikely(PTR_ERR(wp) != -EAGAIN)) {
1154                                 ret = PTR_ERR(wp);
1155                                 goto err;
1156                         }
1157
1158                         goto flush_io;
1159                 }
1160
1161                 /*
1162                  * It's possible for the allocator to fail, put us on the
1163                  * freelist waitlist, and then succeed in one of various retry
1164                  * paths: if that happens, we need to disable the skip_put
1165                  * optimization because otherwise there won't necessarily be a
1166                  * barrier before we free the bch_write_op:
1167                  */
1168                 if (atomic_read(&cl->remaining) & CLOSURE_WAITING)
1169                         skip_put = false;
1170
1171                 bch2_open_bucket_get(c, wp, &op->open_buckets);
1172                 ret = bch2_write_extent(op, wp, &bio);
1173                 bch2_alloc_sectors_done(c, wp);
1174
1175                 if (ret < 0)
1176                         goto err;
1177
1178                 if (ret) {
1179                         skip_put = false;
1180                 } else {
1181                         /*
1182                          * for the skip_put optimization this has to be set
1183                          * before we submit the bio:
1184                          */
1185                         op->flags |= BCH_WRITE_DONE;
1186                 }
1187
1188                 bio->bi_end_io  = bch2_write_endio;
1189                 bio->bi_private = &op->cl;
1190                 bio->bi_opf |= REQ_OP_WRITE;
1191
1192                 if (!skip_put)
1193                         closure_get(bio->bi_private);
1194                 else
1195                         op->flags |= BCH_WRITE_SKIP_CLOSURE_PUT;
1196
1197                 key_to_write = (void *) (op->insert_keys.keys_p +
1198                                          key_to_write_offset);
1199
1200                 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1201                                           key_to_write);
1202         } while (ret);
1203
1204         if (!skip_put)
1205                 continue_at(cl, bch2_write_index, index_update_wq(op));
1206 out:
1207         memalloc_nofs_restore(nofs_flags);
1208         return;
1209 err:
1210         op->error = ret;
1211         op->flags |= BCH_WRITE_DONE;
1212
1213         continue_at(cl, bch2_write_index, index_update_wq(op));
1214         goto out;
1215 flush_io:
1216         /*
1217          * If the write can't all be submitted at once, we generally want to
1218          * block synchronously as that signals backpressure to the caller.
1219          *
1220          * However, if we're running out of a workqueue, we can't block here
1221          * because we'll be blocking other work items from completing:
1222          */
1223         if (current->flags & PF_WQ_WORKER) {
1224                 continue_at(cl, bch2_write_index, index_update_wq(op));
1225                 goto out;
1226         }
1227
1228         closure_sync(cl);
1229
1230         if (!bch2_keylist_empty(&op->insert_keys)) {
1231                 __bch2_write_index(op);
1232
1233                 if (op->error) {
1234                         op->flags |= BCH_WRITE_DONE;
1235                         continue_at_nobarrier(cl, bch2_write_done, NULL);
1236                         goto out;
1237                 }
1238         }
1239
1240         goto again;
1241 }
1242
1243 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1244 {
1245         struct closure *cl = &op->cl;
1246         struct bio *bio = &op->wbio.bio;
1247         struct bvec_iter iter;
1248         struct bkey_i_inline_data *id;
1249         unsigned sectors;
1250         int ret;
1251
1252         bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1253
1254         ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1255                                    ARRAY_SIZE(op->inline_keys),
1256                                    BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1257         if (ret) {
1258                 op->error = ret;
1259                 goto err;
1260         }
1261
1262         sectors = bio_sectors(bio);
1263         op->pos.offset += sectors;
1264
1265         id = bkey_inline_data_init(op->insert_keys.top);
1266         id->k.p         = op->pos;
1267         id->k.version   = op->version;
1268         id->k.size      = sectors;
1269
1270         iter = bio->bi_iter;
1271         iter.bi_size = data_len;
1272         memcpy_from_bio(id->v.data, bio, iter);
1273
1274         while (data_len & 7)
1275                 id->v.data[data_len++] = '\0';
1276         set_bkey_val_bytes(&id->k, data_len);
1277         bch2_keylist_push(&op->insert_keys);
1278
1279         op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1280         op->flags |= BCH_WRITE_DONE;
1281
1282         continue_at_nobarrier(cl, bch2_write_index, NULL);
1283         return;
1284 err:
1285         bch2_write_done(&op->cl);
1286 }
1287
1288 /**
1289  * bch_write - handle a write to a cache device or flash only volume
1290  *
1291  * This is the starting point for any data to end up in a cache device; it could
1292  * be from a normal write, or a writeback write, or a write to a flash only
1293  * volume - it's also used by the moving garbage collector to compact data in
1294  * mostly empty buckets.
1295  *
1296  * It first writes the data to the cache, creating a list of keys to be inserted
1297  * (if the data won't fit in a single open bucket, there will be multiple keys);
1298  * after the data is written it calls bch_journal, and after the keys have been
1299  * added to the next journal write they're inserted into the btree.
1300  *
1301  * If op->discard is true, instead of inserting the data it invalidates the
1302  * region of the cache represented by op->bio and op->inode.
1303  */
1304 void bch2_write(struct closure *cl)
1305 {
1306         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1307         struct bio *bio = &op->wbio.bio;
1308         struct bch_fs *c = op->c;
1309         unsigned data_len;
1310
1311         BUG_ON(!op->nr_replicas);
1312         BUG_ON(!op->write_point.v);
1313         BUG_ON(!bkey_cmp(op->pos, POS_MAX));
1314
1315         op->start_time = local_clock();
1316         bch2_keylist_init(&op->insert_keys, op->inline_keys);
1317         wbio_init(bio)->put_bio = false;
1318
1319         if (bio_sectors(bio) & (c->opts.block_size - 1)) {
1320                 bch_err_inum_ratelimited(c, op->pos.inode,
1321                                          "misaligned write");
1322                 op->error = -EIO;
1323                 goto err;
1324         }
1325
1326         if (c->opts.nochanges ||
1327             !percpu_ref_tryget(&c->writes)) {
1328                 op->error = -EROFS;
1329                 goto err;
1330         }
1331
1332         bch2_increment_clock(c, bio_sectors(bio), WRITE);
1333
1334         data_len = min_t(u64, bio->bi_iter.bi_size,
1335                          op->new_i_size - (op->pos.offset << 9));
1336
1337         if (c->opts.inline_data &&
1338             data_len <= min(block_bytes(c) / 2, 1024U)) {
1339                 bch2_write_data_inline(op, data_len);
1340                 return;
1341         }
1342
1343         continue_at_nobarrier(cl, __bch2_write, NULL);
1344         return;
1345 err:
1346         bch2_disk_reservation_put(c, &op->res);
1347
1348         if (op->end_io) {
1349                 EBUG_ON(cl->parent);
1350                 closure_debug_destroy(cl);
1351                 op->end_io(op);
1352         } else {
1353                 closure_return(cl);
1354         }
1355 }
1356
1357 /* Cache promotion on read */
1358
1359 struct promote_op {
1360         struct closure          cl;
1361         struct rcu_head         rcu;
1362         u64                     start_time;
1363
1364         struct rhash_head       hash;
1365         struct bpos             pos;
1366
1367         struct migrate_write    write;
1368         struct bio_vec          bi_inline_vecs[0]; /* must be last */
1369 };
1370
1371 static const struct rhashtable_params bch_promote_params = {
1372         .head_offset    = offsetof(struct promote_op, hash),
1373         .key_offset     = offsetof(struct promote_op, pos),
1374         .key_len        = sizeof(struct bpos),
1375 };
1376
1377 static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
1378                                   struct bpos pos,
1379                                   struct bch_io_opts opts,
1380                                   unsigned flags)
1381 {
1382         if (!(flags & BCH_READ_MAY_PROMOTE))
1383                 return false;
1384
1385         if (!opts.promote_target)
1386                 return false;
1387
1388         if (bch2_bkey_has_target(c, k, opts.promote_target))
1389                 return false;
1390
1391         if (bch2_target_congested(c, opts.promote_target)) {
1392                 /* XXX trace this */
1393                 return false;
1394         }
1395
1396         if (rhashtable_lookup_fast(&c->promote_table, &pos,
1397                                    bch_promote_params))
1398                 return false;
1399
1400         return true;
1401 }
1402
1403 static void promote_free(struct bch_fs *c, struct promote_op *op)
1404 {
1405         int ret;
1406
1407         ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
1408                                      bch_promote_params);
1409         BUG_ON(ret);
1410         percpu_ref_put(&c->writes);
1411         kfree_rcu(op, rcu);
1412 }
1413
1414 static void promote_done(struct closure *cl)
1415 {
1416         struct promote_op *op =
1417                 container_of(cl, struct promote_op, cl);
1418         struct bch_fs *c = op->write.op.c;
1419
1420         bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
1421                                op->start_time);
1422
1423         bch2_bio_free_pages_pool(c, &op->write.op.wbio.bio);
1424         promote_free(c, op);
1425 }
1426
1427 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
1428 {
1429         struct bch_fs *c = rbio->c;
1430         struct closure *cl = &op->cl;
1431         struct bio *bio = &op->write.op.wbio.bio;
1432
1433         trace_promote(&rbio->bio);
1434
1435         /* we now own pages: */
1436         BUG_ON(!rbio->bounce);
1437         BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
1438
1439         memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
1440                sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
1441         swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
1442
1443         bch2_migrate_read_done(&op->write, rbio);
1444
1445         closure_init(cl, NULL);
1446         closure_call(&op->write.op.cl, bch2_write, c->btree_update_wq, cl);
1447         closure_return_with_destructor(cl, promote_done);
1448 }
1449
1450 static struct promote_op *__promote_alloc(struct bch_fs *c,
1451                                           enum btree_id btree_id,
1452                                           struct bkey_s_c k,
1453                                           struct bpos pos,
1454                                           struct extent_ptr_decoded *pick,
1455                                           struct bch_io_opts opts,
1456                                           unsigned sectors,
1457                                           struct bch_read_bio **rbio)
1458 {
1459         struct promote_op *op = NULL;
1460         struct bio *bio;
1461         unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
1462         int ret;
1463
1464         if (!percpu_ref_tryget(&c->writes))
1465                 return NULL;
1466
1467         op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOIO);
1468         if (!op)
1469                 goto err;
1470
1471         op->start_time = local_clock();
1472         op->pos = pos;
1473
1474         /*
1475          * We don't use the mempool here because extents that aren't
1476          * checksummed or compressed can be too big for the mempool:
1477          */
1478         *rbio = kzalloc(sizeof(struct bch_read_bio) +
1479                         sizeof(struct bio_vec) * pages,
1480                         GFP_NOIO);
1481         if (!*rbio)
1482                 goto err;
1483
1484         rbio_init(&(*rbio)->bio, opts);
1485         bio_init(&(*rbio)->bio, (*rbio)->bio.bi_inline_vecs, pages);
1486
1487         if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
1488                                  GFP_NOIO))
1489                 goto err;
1490
1491         (*rbio)->bounce         = true;
1492         (*rbio)->split          = true;
1493         (*rbio)->kmalloc        = true;
1494
1495         if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
1496                                           bch_promote_params))
1497                 goto err;
1498
1499         bio = &op->write.op.wbio.bio;
1500         bio_init(bio, bio->bi_inline_vecs, pages);
1501
1502         ret = bch2_migrate_write_init(c, &op->write,
1503                         writepoint_hashed((unsigned long) current),
1504                         opts,
1505                         DATA_PROMOTE,
1506                         (struct data_opts) {
1507                                 .target         = opts.promote_target,
1508                                 .nr_replicas    = 1,
1509                         },
1510                         btree_id, k);
1511         BUG_ON(ret);
1512
1513         return op;
1514 err:
1515         if (*rbio)
1516                 bio_free_pages(&(*rbio)->bio);
1517         kfree(*rbio);
1518         *rbio = NULL;
1519         kfree(op);
1520         percpu_ref_put(&c->writes);
1521         return NULL;
1522 }
1523
1524 noinline
1525 static struct promote_op *promote_alloc(struct bch_fs *c,
1526                                                struct bvec_iter iter,
1527                                                struct bkey_s_c k,
1528                                                struct extent_ptr_decoded *pick,
1529                                                struct bch_io_opts opts,
1530                                                unsigned flags,
1531                                                struct bch_read_bio **rbio,
1532                                                bool *bounce,
1533                                                bool *read_full)
1534 {
1535         bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
1536         /* data might have to be decompressed in the write path: */
1537         unsigned sectors = promote_full
1538                 ? max(pick->crc.compressed_size, pick->crc.live_size)
1539                 : bvec_iter_sectors(iter);
1540         struct bpos pos = promote_full
1541                 ? bkey_start_pos(k.k)
1542                 : POS(k.k->p.inode, iter.bi_sector);
1543         struct promote_op *promote;
1544
1545         if (!should_promote(c, k, pos, opts, flags))
1546                 return NULL;
1547
1548         promote = __promote_alloc(c,
1549                                   k.k->type == KEY_TYPE_reflink_v
1550                                   ? BTREE_ID_reflink
1551                                   : BTREE_ID_extents,
1552                                   k, pos, pick, opts, sectors, rbio);
1553         if (!promote)
1554                 return NULL;
1555
1556         *bounce         = true;
1557         *read_full      = promote_full;
1558         return promote;
1559 }
1560
1561 /* Read */
1562
1563 #define READ_RETRY_AVOID        1
1564 #define READ_RETRY              2
1565 #define READ_ERR                3
1566
1567 enum rbio_context {
1568         RBIO_CONTEXT_NULL,
1569         RBIO_CONTEXT_HIGHPRI,
1570         RBIO_CONTEXT_UNBOUND,
1571 };
1572
1573 static inline struct bch_read_bio *
1574 bch2_rbio_parent(struct bch_read_bio *rbio)
1575 {
1576         return rbio->split ? rbio->parent : rbio;
1577 }
1578
1579 __always_inline
1580 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
1581                            enum rbio_context context,
1582                            struct workqueue_struct *wq)
1583 {
1584         if (context <= rbio->context) {
1585                 fn(&rbio->work);
1586         } else {
1587                 rbio->work.func         = fn;
1588                 rbio->context           = context;
1589                 queue_work(wq, &rbio->work);
1590         }
1591 }
1592
1593 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
1594 {
1595         BUG_ON(rbio->bounce && !rbio->split);
1596
1597         if (rbio->promote)
1598                 promote_free(rbio->c, rbio->promote);
1599         rbio->promote = NULL;
1600
1601         if (rbio->bounce)
1602                 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
1603
1604         if (rbio->split) {
1605                 struct bch_read_bio *parent = rbio->parent;
1606
1607                 if (rbio->kmalloc)
1608                         kfree(rbio);
1609                 else
1610                         bio_put(&rbio->bio);
1611
1612                 rbio = parent;
1613         }
1614
1615         return rbio;
1616 }
1617
1618 /*
1619  * Only called on a top level bch_read_bio to complete an entire read request,
1620  * not a split:
1621  */
1622 static void bch2_rbio_done(struct bch_read_bio *rbio)
1623 {
1624         if (rbio->start_time)
1625                 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
1626                                        rbio->start_time);
1627         bio_endio(&rbio->bio);
1628 }
1629
1630 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
1631                                      struct bvec_iter bvec_iter, u64 inode,
1632                                      struct bch_io_failures *failed,
1633                                      unsigned flags)
1634 {
1635         struct btree_trans trans;
1636         struct btree_iter *iter;
1637         struct bkey_buf sk;
1638         struct bkey_s_c k;
1639         int ret;
1640
1641         flags &= ~BCH_READ_LAST_FRAGMENT;
1642         flags |= BCH_READ_MUST_CLONE;
1643
1644         bch2_bkey_buf_init(&sk);
1645         bch2_trans_init(&trans, c, 0, 0);
1646
1647         iter = bch2_trans_get_iter(&trans, rbio->data_btree,
1648                                    rbio->read_pos, BTREE_ITER_SLOTS);
1649 retry:
1650         rbio->bio.bi_status = 0;
1651
1652         k = bch2_btree_iter_peek_slot(iter);
1653         if (bkey_err(k))
1654                 goto err;
1655
1656         bch2_bkey_buf_reassemble(&sk, c, k);
1657         k = bkey_i_to_s_c(sk.k);
1658         bch2_trans_unlock(&trans);
1659
1660         if (!bch2_bkey_matches_ptr(c, k,
1661                                    rbio->pick.ptr,
1662                                    rbio->data_pos.offset -
1663                                    rbio->pick.crc.offset)) {
1664                 /* extent we wanted to read no longer exists: */
1665                 rbio->hole = true;
1666                 goto out;
1667         }
1668
1669         ret = __bch2_read_extent(&trans, rbio, bvec_iter,
1670                                  rbio->read_pos,
1671                                  rbio->data_btree,
1672                                  k, 0, failed, flags);
1673         if (ret == READ_RETRY)
1674                 goto retry;
1675         if (ret)
1676                 goto err;
1677 out:
1678         bch2_rbio_done(rbio);
1679         bch2_trans_iter_put(&trans, iter);
1680         bch2_trans_exit(&trans);
1681         bch2_bkey_buf_exit(&sk, c);
1682         return;
1683 err:
1684         rbio->bio.bi_status = BLK_STS_IOERR;
1685         goto out;
1686 }
1687
1688 static void bch2_rbio_retry(struct work_struct *work)
1689 {
1690         struct bch_read_bio *rbio =
1691                 container_of(work, struct bch_read_bio, work);
1692         struct bch_fs *c        = rbio->c;
1693         struct bvec_iter iter   = rbio->bvec_iter;
1694         unsigned flags          = rbio->flags;
1695         u64 inode               = rbio->read_pos.inode;
1696         struct bch_io_failures failed = { .nr = 0 };
1697
1698         trace_read_retry(&rbio->bio);
1699
1700         if (rbio->retry == READ_RETRY_AVOID)
1701                 bch2_mark_io_failure(&failed, &rbio->pick);
1702
1703         rbio->bio.bi_status = 0;
1704
1705         rbio = bch2_rbio_free(rbio);
1706
1707         flags |= BCH_READ_IN_RETRY;
1708         flags &= ~BCH_READ_MAY_PROMOTE;
1709
1710         if (flags & BCH_READ_NODECODE) {
1711                 bch2_read_retry_nodecode(c, rbio, iter, inode, &failed, flags);
1712         } else {
1713                 flags &= ~BCH_READ_LAST_FRAGMENT;
1714                 flags |= BCH_READ_MUST_CLONE;
1715
1716                 __bch2_read(c, rbio, iter, inode, &failed, flags);
1717         }
1718 }
1719
1720 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
1721                             blk_status_t error)
1722 {
1723         rbio->retry = retry;
1724
1725         if (rbio->flags & BCH_READ_IN_RETRY)
1726                 return;
1727
1728         if (retry == READ_ERR) {
1729                 rbio = bch2_rbio_free(rbio);
1730
1731                 rbio->bio.bi_status = error;
1732                 bch2_rbio_done(rbio);
1733         } else {
1734                 bch2_rbio_punt(rbio, bch2_rbio_retry,
1735                                RBIO_CONTEXT_UNBOUND, system_unbound_wq);
1736         }
1737 }
1738
1739 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
1740                                    struct bch_read_bio *rbio)
1741 {
1742         struct bch_fs *c = rbio->c;
1743         u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
1744         struct bch_extent_crc_unpacked new_crc;
1745         struct btree_iter *iter = NULL;
1746         struct bkey_i *new;
1747         struct bkey_s_c k;
1748         int ret = 0;
1749
1750         if (crc_is_compressed(rbio->pick.crc))
1751                 return 0;
1752
1753         iter = bch2_trans_get_iter(trans, rbio->data_btree, rbio->data_pos,
1754                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1755         k = bch2_btree_iter_peek_slot(iter);
1756         if ((ret = bkey_err(k)))
1757                 goto out;
1758
1759         if (bversion_cmp(k.k->version, rbio->version) ||
1760             !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
1761                 goto out;
1762
1763         /* Extent was merged? */
1764         if (bkey_start_offset(k.k) < data_offset ||
1765             k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
1766                 goto out;
1767
1768         if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
1769                         rbio->pick.crc, NULL, &new_crc,
1770                         bkey_start_offset(k.k) - data_offset, k.k->size,
1771                         rbio->pick.crc.csum_type)) {
1772                 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
1773                 ret = 0;
1774                 goto out;
1775         }
1776
1777         /*
1778          * going to be temporarily appending another checksum entry:
1779          */
1780         new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
1781                                  sizeof(struct bch_extent_crc128));
1782         if ((ret = PTR_ERR_OR_ZERO(new)))
1783                 goto out;
1784
1785         bkey_reassemble(new, k);
1786
1787         if (!bch2_bkey_narrow_crcs(new, new_crc))
1788                 goto out;
1789
1790         ret = bch2_trans_update(trans, iter, new, 0);
1791 out:
1792         bch2_trans_iter_put(trans, iter);
1793         return ret;
1794 }
1795
1796 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
1797 {
1798         bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
1799                       __bch2_rbio_narrow_crcs(&trans, rbio));
1800 }
1801
1802 /* Inner part that may run in process context */
1803 static void __bch2_read_endio(struct work_struct *work)
1804 {
1805         struct bch_read_bio *rbio =
1806                 container_of(work, struct bch_read_bio, work);
1807         struct bch_fs *c        = rbio->c;
1808         struct bch_dev *ca      = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1809         struct bio *src         = &rbio->bio;
1810         struct bio *dst         = &bch2_rbio_parent(rbio)->bio;
1811         struct bvec_iter dst_iter = rbio->bvec_iter;
1812         struct bch_extent_crc_unpacked crc = rbio->pick.crc;
1813         struct nonce nonce = extent_nonce(rbio->version, crc);
1814         unsigned nofs_flags;
1815         struct bch_csum csum;
1816
1817         nofs_flags = memalloc_nofs_save();
1818
1819         /* Reset iterator for checksumming and copying bounced data: */
1820         if (rbio->bounce) {
1821                 src->bi_iter.bi_size            = crc.compressed_size << 9;
1822                 src->bi_iter.bi_idx             = 0;
1823                 src->bi_iter.bi_bvec_done       = 0;
1824         } else {
1825                 src->bi_iter                    = rbio->bvec_iter;
1826         }
1827
1828         csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
1829         if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
1830                 goto csum_err;
1831
1832         /*
1833          * XXX
1834          * We need to rework the narrow_crcs path to deliver the read completion
1835          * first, and then punt to a different workqueue, otherwise we're
1836          * holding up reads while doing btree updates which is bad for memory
1837          * reclaim.
1838          */
1839         if (unlikely(rbio->narrow_crcs))
1840                 bch2_rbio_narrow_crcs(rbio);
1841
1842         if (rbio->flags & BCH_READ_NODECODE)
1843                 goto nodecode;
1844
1845         /* Adjust crc to point to subset of data we want: */
1846         crc.offset     += rbio->offset_into_extent;
1847         crc.live_size   = bvec_iter_sectors(rbio->bvec_iter);
1848
1849         if (crc_is_compressed(crc)) {
1850                 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1851                 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
1852                         goto decompression_err;
1853         } else {
1854                 /* don't need to decrypt the entire bio: */
1855                 nonce = nonce_add(nonce, crc.offset << 9);
1856                 bio_advance(src, crc.offset << 9);
1857
1858                 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
1859                 src->bi_iter.bi_size = dst_iter.bi_size;
1860
1861                 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1862
1863                 if (rbio->bounce) {
1864                         struct bvec_iter src_iter = src->bi_iter;
1865                         bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
1866                 }
1867         }
1868
1869         if (rbio->promote) {
1870                 /*
1871                  * Re encrypt data we decrypted, so it's consistent with
1872                  * rbio->crc:
1873                  */
1874                 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1875                 promote_start(rbio->promote, rbio);
1876                 rbio->promote = NULL;
1877         }
1878 nodecode:
1879         if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
1880                 rbio = bch2_rbio_free(rbio);
1881                 bch2_rbio_done(rbio);
1882         }
1883 out:
1884         memalloc_nofs_restore(nofs_flags);
1885         return;
1886 csum_err:
1887         /*
1888          * Checksum error: if the bio wasn't bounced, we may have been
1889          * reading into buffers owned by userspace (that userspace can
1890          * scribble over) - retry the read, bouncing it this time:
1891          */
1892         if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
1893                 rbio->flags |= BCH_READ_MUST_BOUNCE;
1894                 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
1895                 goto out;
1896         }
1897
1898         bch2_dev_inum_io_error(ca, rbio->read_pos.inode, (u64) rbio->bvec_iter.bi_sector,
1899                 "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %u)",
1900                 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
1901                 csum.hi, csum.lo, crc.csum_type);
1902         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1903         goto out;
1904 decompression_err:
1905         bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1906                                  "decompression error");
1907         bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1908         goto out;
1909 }
1910
1911 static void bch2_read_endio(struct bio *bio)
1912 {
1913         struct bch_read_bio *rbio =
1914                 container_of(bio, struct bch_read_bio, bio);
1915         struct bch_fs *c        = rbio->c;
1916         struct bch_dev *ca      = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1917         struct workqueue_struct *wq = NULL;
1918         enum rbio_context context = RBIO_CONTEXT_NULL;
1919
1920         if (rbio->have_ioref) {
1921                 bch2_latency_acct(ca, rbio->submit_time, READ);
1922                 percpu_ref_put(&ca->io_ref);
1923         }
1924
1925         if (!rbio->split)
1926                 rbio->bio.bi_end_io = rbio->end_io;
1927
1928         if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
1929                                     rbio->read_pos.inode,
1930                                     rbio->read_pos.offset,
1931                                     "data read error: %s",
1932                                bch2_blk_status_to_str(bio->bi_status))) {
1933                 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
1934                 return;
1935         }
1936
1937         if (rbio->pick.ptr.cached &&
1938             (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
1939              ptr_stale(ca, &rbio->pick.ptr))) {
1940                 atomic_long_inc(&c->read_realloc_races);
1941
1942                 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
1943                         bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
1944                 else
1945                         bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
1946                 return;
1947         }
1948
1949         if (rbio->narrow_crcs ||
1950             crc_is_compressed(rbio->pick.crc) ||
1951             bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
1952                 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
1953         else if (rbio->pick.crc.csum_type)
1954                 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
1955
1956         bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
1957 }
1958
1959 int __bch2_read_indirect_extent(struct btree_trans *trans,
1960                                 unsigned *offset_into_extent,
1961                                 struct bkey_buf *orig_k)
1962 {
1963         struct btree_iter *iter;
1964         struct bkey_s_c k;
1965         u64 reflink_offset;
1966         int ret;
1967
1968         reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
1969                 *offset_into_extent;
1970
1971         iter = bch2_trans_get_iter(trans, BTREE_ID_reflink,
1972                                    POS(0, reflink_offset),
1973                                    BTREE_ITER_SLOTS);
1974         k = bch2_btree_iter_peek_slot(iter);
1975         ret = bkey_err(k);
1976         if (ret)
1977                 goto err;
1978
1979         if (k.k->type != KEY_TYPE_reflink_v &&
1980             k.k->type != KEY_TYPE_indirect_inline_data) {
1981                 bch_err_inum_ratelimited(trans->c, orig_k->k->k.p.inode,
1982                         "%llu len %u points to nonexistent indirect extent %llu",
1983                         orig_k->k->k.p.offset,
1984                         orig_k->k->k.size,
1985                         reflink_offset);
1986                 bch2_inconsistent_error(trans->c);
1987                 ret = -EIO;
1988                 goto err;
1989         }
1990
1991         *offset_into_extent = iter->pos.offset - bkey_start_offset(k.k);
1992         bch2_bkey_buf_reassemble(orig_k, trans->c, k);
1993 err:
1994         bch2_trans_iter_put(trans, iter);
1995         return ret;
1996 }
1997
1998 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
1999                        struct bvec_iter iter, struct bpos read_pos,
2000                        enum btree_id data_btree, struct bkey_s_c k,
2001                        unsigned offset_into_extent,
2002                        struct bch_io_failures *failed, unsigned flags)
2003 {
2004         struct bch_fs *c = trans->c;
2005         struct extent_ptr_decoded pick;
2006         struct bch_read_bio *rbio = NULL;
2007         struct bch_dev *ca;
2008         struct promote_op *promote = NULL;
2009         bool bounce = false, read_full = false, narrow_crcs = false;
2010         struct bpos data_pos = bkey_start_pos(k.k);
2011         int pick_ret;
2012
2013         if (bkey_extent_is_inline_data(k.k)) {
2014                 unsigned bytes = min_t(unsigned, iter.bi_size,
2015                                        bkey_inline_data_bytes(k.k));
2016
2017                 swap(iter.bi_size, bytes);
2018                 memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
2019                 swap(iter.bi_size, bytes);
2020                 bio_advance_iter(&orig->bio, &iter, bytes);
2021                 zero_fill_bio_iter(&orig->bio, iter);
2022                 goto out_read_done;
2023         }
2024
2025         pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
2026
2027         /* hole or reservation - just zero fill: */
2028         if (!pick_ret)
2029                 goto hole;
2030
2031         if (pick_ret < 0) {
2032                 bch_err_inum_ratelimited(c, k.k->p.inode,
2033                                          "no device to read from");
2034                 goto err;
2035         }
2036
2037         if (pick_ret > 0)
2038                 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
2039
2040         if (flags & BCH_READ_NODECODE) {
2041                 /*
2042                  * can happen if we retry, and the extent we were going to read
2043                  * has been merged in the meantime:
2044                  */
2045                 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
2046                         goto hole;
2047
2048                 iter.bi_size    = pick.crc.compressed_size << 9;
2049                 goto get_bio;
2050         }
2051
2052         if (!(flags & BCH_READ_LAST_FRAGMENT) ||
2053             bio_flagged(&orig->bio, BIO_CHAIN))
2054                 flags |= BCH_READ_MUST_CLONE;
2055
2056         narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
2057                 bch2_can_narrow_extent_crcs(k, pick.crc);
2058
2059         if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
2060                 flags |= BCH_READ_MUST_BOUNCE;
2061
2062         EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
2063
2064         if (crc_is_compressed(pick.crc) ||
2065             (pick.crc.csum_type != BCH_CSUM_NONE &&
2066              (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2067               (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
2068                (flags & BCH_READ_USER_MAPPED)) ||
2069               (flags & BCH_READ_MUST_BOUNCE)))) {
2070                 read_full = true;
2071                 bounce = true;
2072         }
2073
2074         if (orig->opts.promote_target)
2075                 promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
2076                                         &rbio, &bounce, &read_full);
2077
2078         if (!read_full) {
2079                 EBUG_ON(crc_is_compressed(pick.crc));
2080                 EBUG_ON(pick.crc.csum_type &&
2081                         (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2082                          bvec_iter_sectors(iter) != pick.crc.live_size ||
2083                          pick.crc.offset ||
2084                          offset_into_extent));
2085
2086                 data_pos.offset += offset_into_extent;
2087                 pick.ptr.offset += pick.crc.offset +
2088                         offset_into_extent;
2089                 offset_into_extent              = 0;
2090                 pick.crc.compressed_size        = bvec_iter_sectors(iter);
2091                 pick.crc.uncompressed_size      = bvec_iter_sectors(iter);
2092                 pick.crc.offset                 = 0;
2093                 pick.crc.live_size              = bvec_iter_sectors(iter);
2094                 offset_into_extent              = 0;
2095         }
2096 get_bio:
2097         if (rbio) {
2098                 /*
2099                  * promote already allocated bounce rbio:
2100                  * promote needs to allocate a bio big enough for uncompressing
2101                  * data in the write path, but we're not going to use it all
2102                  * here:
2103                  */
2104                 EBUG_ON(rbio->bio.bi_iter.bi_size <
2105                        pick.crc.compressed_size << 9);
2106                 rbio->bio.bi_iter.bi_size =
2107                         pick.crc.compressed_size << 9;
2108         } else if (bounce) {
2109                 unsigned sectors = pick.crc.compressed_size;
2110
2111                 rbio = rbio_init(bio_alloc_bioset(GFP_NOIO,
2112                                                   DIV_ROUND_UP(sectors, PAGE_SECTORS),
2113                                                   &c->bio_read_split),
2114                                  orig->opts);
2115
2116                 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
2117                 rbio->bounce    = true;
2118                 rbio->split     = true;
2119         } else if (flags & BCH_READ_MUST_CLONE) {
2120                 /*
2121                  * Have to clone if there were any splits, due to error
2122                  * reporting issues (if a split errored, and retrying didn't
2123                  * work, when it reports the error to its parent (us) we don't
2124                  * know if the error was from our bio, and we should retry, or
2125                  * from the whole bio, in which case we don't want to retry and
2126                  * lose the error)
2127                  */
2128                 rbio = rbio_init(bio_clone_fast(&orig->bio, GFP_NOIO,
2129                                                 &c->bio_read_split),
2130                                  orig->opts);
2131                 rbio->bio.bi_iter = iter;
2132                 rbio->split     = true;
2133         } else {
2134                 rbio = orig;
2135                 rbio->bio.bi_iter = iter;
2136                 EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
2137         }
2138
2139         EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
2140
2141         rbio->c                 = c;
2142         rbio->submit_time       = local_clock();
2143         if (rbio->split)
2144                 rbio->parent    = orig;
2145         else
2146                 rbio->end_io    = orig->bio.bi_end_io;
2147         rbio->bvec_iter         = iter;
2148         rbio->offset_into_extent= offset_into_extent;
2149         rbio->flags             = flags;
2150         rbio->have_ioref        = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
2151         rbio->narrow_crcs       = narrow_crcs;
2152         rbio->hole              = 0;
2153         rbio->retry             = 0;
2154         rbio->context           = 0;
2155         /* XXX: only initialize this if needed */
2156         rbio->devs_have         = bch2_bkey_devs(k);
2157         rbio->pick              = pick;
2158         rbio->read_pos          = read_pos;
2159         rbio->data_btree        = data_btree;
2160         rbio->data_pos          = data_pos;
2161         rbio->version           = k.k->version;
2162         rbio->promote           = promote;
2163         INIT_WORK(&rbio->work, NULL);
2164
2165         rbio->bio.bi_opf        = orig->bio.bi_opf;
2166         rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
2167         rbio->bio.bi_end_io     = bch2_read_endio;
2168
2169         if (rbio->bounce)
2170                 trace_read_bounce(&rbio->bio);
2171
2172         bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
2173
2174         /*
2175          * If it's being moved internally, we don't want to flag it as a cache
2176          * hit:
2177          */
2178         if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
2179                 bch2_bucket_io_time_reset(trans, pick.ptr.dev,
2180                         PTR_BUCKET_NR(ca, &pick.ptr), READ);
2181
2182         if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
2183                 bio_inc_remaining(&orig->bio);
2184                 trace_read_split(&orig->bio);
2185         }
2186
2187         if (!rbio->pick.idx) {
2188                 if (!rbio->have_ioref) {
2189                         bch_err_inum_ratelimited(c, k.k->p.inode,
2190                                                  "no device to read from");
2191                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2192                         goto out;
2193                 }
2194
2195                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
2196                              bio_sectors(&rbio->bio));
2197                 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
2198
2199                 if (likely(!(flags & BCH_READ_IN_RETRY)))
2200                         submit_bio(&rbio->bio);
2201                 else
2202                         submit_bio_wait(&rbio->bio);
2203         } else {
2204                 /* Attempting reconstruct read: */
2205                 if (bch2_ec_read_extent(c, rbio)) {
2206                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2207                         goto out;
2208                 }
2209
2210                 if (likely(!(flags & BCH_READ_IN_RETRY)))
2211                         bio_endio(&rbio->bio);
2212         }
2213 out:
2214         if (likely(!(flags & BCH_READ_IN_RETRY))) {
2215                 return 0;
2216         } else {
2217                 int ret;
2218
2219                 rbio->context = RBIO_CONTEXT_UNBOUND;
2220                 bch2_read_endio(&rbio->bio);
2221
2222                 ret = rbio->retry;
2223                 rbio = bch2_rbio_free(rbio);
2224
2225                 if (ret == READ_RETRY_AVOID) {
2226                         bch2_mark_io_failure(failed, &pick);
2227                         ret = READ_RETRY;
2228                 }
2229
2230                 if (!ret)
2231                         goto out_read_done;
2232
2233                 return ret;
2234         }
2235
2236 err:
2237         if (flags & BCH_READ_IN_RETRY)
2238                 return READ_ERR;
2239
2240         orig->bio.bi_status = BLK_STS_IOERR;
2241         goto out_read_done;
2242
2243 hole:
2244         /*
2245          * won't normally happen in the BCH_READ_NODECODE
2246          * (bch2_move_extent()) path, but if we retry and the extent we wanted
2247          * to read no longer exists we have to signal that:
2248          */
2249         if (flags & BCH_READ_NODECODE)
2250                 orig->hole = true;
2251
2252         zero_fill_bio_iter(&orig->bio, iter);
2253 out_read_done:
2254         if (flags & BCH_READ_LAST_FRAGMENT)
2255                 bch2_rbio_done(orig);
2256         return 0;
2257 }
2258
2259 void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
2260                  struct bvec_iter bvec_iter, u64 inode,
2261                  struct bch_io_failures *failed, unsigned flags)
2262 {
2263         struct btree_trans trans;
2264         struct btree_iter *iter;
2265         struct bkey_buf sk;
2266         struct bkey_s_c k;
2267         int ret;
2268
2269         BUG_ON(flags & BCH_READ_NODECODE);
2270
2271         bch2_bkey_buf_init(&sk);
2272         bch2_trans_init(&trans, c, 0, 0);
2273 retry:
2274         bch2_trans_begin(&trans);
2275
2276         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents,
2277                                    POS(inode, bvec_iter.bi_sector),
2278                                    BTREE_ITER_SLOTS);
2279         while (1) {
2280                 unsigned bytes, sectors, offset_into_extent;
2281                 enum btree_id data_btree = BTREE_ID_extents;
2282
2283                 bch2_btree_iter_set_pos(iter,
2284                                 POS(inode, bvec_iter.bi_sector));
2285
2286                 k = bch2_btree_iter_peek_slot(iter);
2287                 ret = bkey_err(k);
2288                 if (ret)
2289                         break;
2290
2291                 offset_into_extent = iter->pos.offset -
2292                         bkey_start_offset(k.k);
2293                 sectors = k.k->size - offset_into_extent;
2294
2295                 bch2_bkey_buf_reassemble(&sk, c, k);
2296
2297                 ret = bch2_read_indirect_extent(&trans, &data_btree,
2298                                         &offset_into_extent, &sk);
2299                 if (ret)
2300                         break;
2301
2302                 k = bkey_i_to_s_c(sk.k);
2303
2304                 /*
2305                  * With indirect extents, the amount of data to read is the min
2306                  * of the original extent and the indirect extent:
2307                  */
2308                 sectors = min(sectors, k.k->size - offset_into_extent);
2309
2310                 /*
2311                  * Unlock the iterator while the btree node's lock is still in
2312                  * cache, before doing the IO:
2313                  */
2314                 bch2_trans_unlock(&trans);
2315
2316                 bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
2317                 swap(bvec_iter.bi_size, bytes);
2318
2319                 if (bvec_iter.bi_size == bytes)
2320                         flags |= BCH_READ_LAST_FRAGMENT;
2321
2322                 ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter->pos,
2323                                          data_btree, k,
2324                                          offset_into_extent, failed, flags);
2325                 if (ret)
2326                         break;
2327
2328                 if (flags & BCH_READ_LAST_FRAGMENT)
2329                         break;
2330
2331                 swap(bvec_iter.bi_size, bytes);
2332                 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
2333         }
2334         bch2_trans_iter_put(&trans, iter);
2335
2336         if (ret == -EINTR || ret == READ_RETRY || ret == READ_RETRY_AVOID)
2337                 goto retry;
2338
2339         if (ret) {
2340                 bch_err_inum_ratelimited(c, inode,
2341                                          "read error %i from btree lookup", ret);
2342                 rbio->bio.bi_status = BLK_STS_IOERR;
2343                 bch2_rbio_done(rbio);
2344         }
2345         bch2_trans_exit(&trans);
2346         bch2_bkey_buf_exit(&sk, c);
2347 }
2348
2349 void bch2_fs_io_exit(struct bch_fs *c)
2350 {
2351         if (c->promote_table.tbl)
2352                 rhashtable_destroy(&c->promote_table);
2353         mempool_exit(&c->bio_bounce_pages);
2354         bioset_exit(&c->bio_write);
2355         bioset_exit(&c->bio_read_split);
2356         bioset_exit(&c->bio_read);
2357 }
2358
2359 int bch2_fs_io_init(struct bch_fs *c)
2360 {
2361         if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
2362                         BIOSET_NEED_BVECS) ||
2363             bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
2364                         BIOSET_NEED_BVECS) ||
2365             bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
2366                         BIOSET_NEED_BVECS) ||
2367             mempool_init_page_pool(&c->bio_bounce_pages,
2368                                    max_t(unsigned,
2369                                          c->opts.btree_node_size,
2370                                          c->sb.encoded_extent_max) /
2371                                    PAGE_SECTORS, 0) ||
2372             rhashtable_init(&c->promote_table, &bch_promote_params))
2373                 return -ENOMEM;
2374
2375         return 0;
2376 }