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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * bcachefs journalling code, for btree insertions
   4  *
   5  * Copyright 2012 Google, Inc.
   6  */
   7
   8 #include "bcachefs.h"
   9 #include "alloc_foreground.h"
  10 #include "bkey_methods.h"
  11 #include "btree_gc.h"
  12 #include "btree_update.h"
  13 #include "buckets.h"
  14 #include "error.h"
  15 #include "journal.h"
  16 #include "journal_io.h"
  17 #include "journal_reclaim.h"
  18 #include "journal_seq_blacklist.h"
  19 #include "super-io.h"
  20
  21 #include <trace/events/bcachefs.h>
  22
  23 static u64 last_unwritten_seq(struct journal *j)
  24 {
  25         union journal_res_state s = READ_ONCE(j->reservations);
  26
  27         lockdep_assert_held(&j->lock);
  28
  29         return journal_cur_seq(j) - ((s.idx - s.unwritten_idx) & JOURNAL_BUF_MASK);
  30 }
  31
  32 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
  33 {
  34         return seq >= last_unwritten_seq(j);
  35 }
  36
  37 static bool __journal_entry_is_open(union journal_res_state state)
  38 {
  39         return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
  40 }
  41
  42 static bool journal_entry_is_open(struct journal *j)
  43 {
  44         return __journal_entry_is_open(j->reservations);
  45 }
  46
  47 static inline struct journal_buf *
  48 journal_seq_to_buf(struct journal *j, u64 seq)
  49 {
  50         struct journal_buf *buf = NULL;
  51
  52         EBUG_ON(seq > journal_cur_seq(j));
  53         EBUG_ON(seq == journal_cur_seq(j) &&
  54                 j->reservations.cur_entry_offset == JOURNAL_ENTRY_CLOSED_VAL);
  55
  56         if (journal_seq_unwritten(j, seq)) {
  57                 buf = j->buf + (seq & JOURNAL_BUF_MASK);
  58                 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
  59         }
  60         return buf;
  61 }
  62
  63 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
  64 {
  65         INIT_LIST_HEAD(&p->list);
  66         INIT_LIST_HEAD(&p->key_cache_list);
  67         INIT_LIST_HEAD(&p->flushed);
  68         atomic_set(&p->count, count);
  69         p->devs.nr = 0;
  70 }
  71
  72 static void journal_pin_new_entry(struct journal *j)
  73 {
  74         /*
  75          * The fifo_push() needs to happen at the same time as j->seq is
  76          * incremented for journal_last_seq() to be calculated correctly
  77          */
  78         atomic64_inc(&j->seq);
  79         journal_pin_list_init(fifo_push_ref(&j->pin), 1);
  80 }
  81
  82 static void bch2_journal_buf_init(struct journal *j)
  83 {
  84         struct journal_buf *buf = journal_cur_buf(j);
  85
  86         bkey_extent_init(&buf->key);
  87         buf->noflush    = false;
  88         buf->must_flush = false;
  89         buf->separate_flush = false;
  90
  91         memset(buf->data, 0, sizeof(*buf->data));
  92         buf->data->seq  = cpu_to_le64(journal_cur_seq(j));
  93         buf->data->u64s = 0;
  94 }
  95
  96 void bch2_journal_halt(struct journal *j)
  97 {
  98         union journal_res_state old, new;
  99         u64 v = atomic64_read(&j->reservations.counter);
 100
 101         do {
 102                 old.v = new.v = v;
 103                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL)
 104                         return;
 105
 106                 new.cur_entry_offset = JOURNAL_ENTRY_ERROR_VAL;
 107         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 108                                        old.v, new.v)) != old.v);
 109
 110         /*
 111          * XXX: we're not using j->lock here because this can be called from
 112          * interrupt context, this can race with journal_write_done()
 113          */
 114         if (!j->err_seq)
 115                 j->err_seq = journal_cur_seq(j);
 116         journal_wake(j);
 117         closure_wake_up(&journal_cur_buf(j)->wait);
 118 }
 119
 120 /* journal entry close/open: */
 121
 122 void __bch2_journal_buf_put(struct journal *j)
 123 {
 124         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 125
 126         closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
 127 }
 128
 129 /*
 130  * Returns true if journal entry is now closed:
 131  *
 132  * We don't close a journal_buf until the next journal_buf is finished writing,
 133  * and can be opened again - this also initializes the next journal_buf:
 134  */
 135 static bool __journal_entry_close(struct journal *j)
 136 {
 137         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 138         struct journal_buf *buf = journal_cur_buf(j);
 139         union journal_res_state old, new;
 140         u64 v = atomic64_read(&j->reservations.counter);
 141         unsigned sectors;
 142
 143         lockdep_assert_held(&j->lock);
 144
 145         do {
 146                 old.v = new.v = v;
 147                 if (old.cur_entry_offset == JOURNAL_ENTRY_CLOSED_VAL)
 148                         return true;
 149
 150                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL) {
 151                         /* this entry will never be written: */
 152                         closure_wake_up(&buf->wait);
 153                         return true;
 154                 }
 155
 156                 if (!test_bit(JOURNAL_NEED_WRITE, &j->flags)) {
 157                         set_bit(JOURNAL_NEED_WRITE, &j->flags);
 158                         j->need_write_time = local_clock();
 159                 }
 160
 161                 new.cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL;
 162                 new.idx++;
 163
 164                 if (new.idx == new.unwritten_idx)
 165                         return false;
 166
 167                 BUG_ON(journal_state_count(new, new.idx));
 168         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 169                                        old.v, new.v)) != old.v);
 170
 171         /* Close out old buffer: */
 172         buf->data->u64s         = cpu_to_le32(old.cur_entry_offset);
 173
 174         sectors = vstruct_blocks_plus(buf->data, c->block_bits,
 175                                       buf->u64s_reserved) << c->block_bits;
 176         BUG_ON(sectors > buf->sectors);
 177         buf->sectors = sectors;
 178
 179         /*
 180          * We have to set last_seq here, _before_ opening a new journal entry:
 181          *
 182          * A threads may replace an old pin with a new pin on their current
 183          * journal reservation - the expectation being that the journal will
 184          * contain either what the old pin protected or what the new pin
 185          * protects.
 186          *
 187          * After the old pin is dropped journal_last_seq() won't include the old
 188          * pin, so we can only write the updated last_seq on the entry that
 189          * contains whatever the new pin protects.
 190          *
 191          * Restated, we can _not_ update last_seq for a given entry if there
 192          * could be a newer entry open with reservations/pins that have been
 193          * taken against it.
 194          *
 195          * Hence, we want update/set last_seq on the current journal entry right
 196          * before we open a new one:
 197          */
 198         buf->last_seq           = journal_last_seq(j);
 199         buf->data->last_seq     = cpu_to_le64(buf->last_seq);
 200
 201         __bch2_journal_pin_put(j, le64_to_cpu(buf->data->seq));
 202
 203         /* Initialize new buffer: */
 204         journal_pin_new_entry(j);
 205
 206         bch2_journal_buf_init(j);
 207
 208         cancel_delayed_work(&j->write_work);
 209         clear_bit(JOURNAL_NEED_WRITE, &j->flags);
 210
 211         bch2_journal_space_available(j);
 212
 213         bch2_journal_buf_put(j, old.idx);
 214         return true;
 215 }
 216
 217 static bool journal_entry_want_write(struct journal *j)
 218 {
 219         union journal_res_state s = READ_ONCE(j->reservations);
 220         bool ret = false;
 221
 222         /*
 223          * Don't close it yet if we already have a write in flight, but do set
 224          * NEED_WRITE:
 225          */
 226         if (s.idx != s.unwritten_idx)
 227                 set_bit(JOURNAL_NEED_WRITE, &j->flags);
 228         else
 229                 ret = __journal_entry_close(j);
 230
 231         return ret;
 232 }
 233
 234 static bool journal_entry_close(struct journal *j)
 235 {
 236         bool ret;
 237
 238         spin_lock(&j->lock);
 239         ret = journal_entry_want_write(j);
 240         spin_unlock(&j->lock);
 241
 242         return ret;
 243 }
 244
 245 /*
 246  * should _only_ called from journal_res_get() - when we actually want a
 247  * journal reservation - journal entry is open means journal is dirty:
 248  *
 249  * returns:
 250  * 0:           success
 251  * -ENOSPC:     journal currently full, must invoke reclaim
 252  * -EAGAIN:     journal blocked, must wait
 253  * -EROFS:      insufficient rw devices or journal error
 254  */
 255 static int journal_entry_open(struct journal *j)
 256 {
 257         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 258         struct journal_buf *buf = journal_cur_buf(j);
 259         union journal_res_state old, new;
 260         int u64s;
 261         u64 v;
 262
 263         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
 264
 265         lockdep_assert_held(&j->lock);
 266         BUG_ON(journal_entry_is_open(j));
 267
 268         if (j->blocked)
 269                 return cur_entry_blocked;
 270
 271         if (j->cur_entry_error)
 272                 return j->cur_entry_error;
 273
 274         BUG_ON(!j->cur_entry_sectors);
 275
 276         buf->u64s_reserved      = j->entry_u64s_reserved;
 277         buf->disk_sectors       = j->cur_entry_sectors;
 278         buf->sectors            = min(buf->disk_sectors, buf->buf_size >> 9);
 279
 280         u64s = (int) (buf->sectors << 9) / sizeof(u64) -
 281                 journal_entry_overhead(j);
 282         u64s  = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
 283
 284         if (u64s <= le32_to_cpu(buf->data->u64s))
 285                 return cur_entry_journal_full;
 286
 287         /*
 288          * Must be set before marking the journal entry as open:
 289          */
 290         j->cur_entry_u64s = u64s;
 291
 292         v = atomic64_read(&j->reservations.counter);
 293         do {
 294                 old.v = new.v = v;
 295
 296                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL)
 297                         return cur_entry_insufficient_devices;
 298
 299                 /* Handle any already added entries */
 300                 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
 301
 302                 EBUG_ON(journal_state_count(new, new.idx));
 303                 journal_state_inc(&new);
 304         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 305                                        old.v, new.v)) != old.v);
 306
 307         if (j->res_get_blocked_start)
 308                 bch2_time_stats_update(j->blocked_time,
 309                                        j->res_get_blocked_start);
 310         j->res_get_blocked_start = 0;
 311
 312         mod_delayed_work(c->io_complete_wq,
 313                          &j->write_work,
 314                          msecs_to_jiffies(c->opts.journal_flush_delay));
 315         journal_wake(j);
 316         return 0;
 317 }
 318
 319 static bool journal_quiesced(struct journal *j)
 320 {
 321         union journal_res_state s = READ_ONCE(j->reservations);
 322         bool ret = s.idx == s.unwritten_idx && !__journal_entry_is_open(s);
 323
 324         if (!ret)
 325                 journal_entry_close(j);
 326         return ret;
 327 }
 328
 329 static void journal_quiesce(struct journal *j)
 330 {
 331         wait_event(j->wait, journal_quiesced(j));
 332 }
 333
 334 static void journal_write_work(struct work_struct *work)
 335 {
 336         struct journal *j = container_of(work, struct journal, write_work.work);
 337
 338         journal_entry_close(j);
 339 }
 340
 341 static int __journal_res_get(struct journal *j, struct journal_res *res,
 342                              unsigned flags)
 343 {
 344         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 345         struct journal_buf *buf;
 346         bool can_discard;
 347         int ret;
 348 retry:
 349         if (journal_res_get_fast(j, res, flags))
 350                 return 0;
 351
 352         if (bch2_journal_error(j))
 353                 return -EROFS;
 354
 355         spin_lock(&j->lock);
 356
 357         /*
 358          * Recheck after taking the lock, so we don't race with another thread
 359          * that just did journal_entry_open() and call journal_entry_close()
 360          * unnecessarily
 361          */
 362         if (journal_res_get_fast(j, res, flags)) {
 363                 spin_unlock(&j->lock);
 364                 return 0;
 365         }
 366
 367         if (!(flags & JOURNAL_RES_GET_RESERVED) &&
 368             !test_bit(JOURNAL_MAY_GET_UNRESERVED, &j->flags)) {
 369                 /*
 370                  * Don't want to close current journal entry, just need to
 371                  * invoke reclaim:
 372                  */
 373                 ret = cur_entry_journal_full;
 374                 goto unlock;
 375         }
 376
 377         /*
 378          * If we couldn't get a reservation because the current buf filled up,
 379          * and we had room for a bigger entry on disk, signal that we want to
 380          * realloc the journal bufs:
 381          */
 382         buf = journal_cur_buf(j);
 383         if (journal_entry_is_open(j) &&
 384             buf->buf_size >> 9 < buf->disk_sectors &&
 385             buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
 386                 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
 387
 388         if (journal_entry_is_open(j) &&
 389             !__journal_entry_close(j)) {
 390                 /*
 391                  * We failed to get a reservation on the current open journal
 392                  * entry because it's full, and we can't close it because
 393                  * there's still a previous one in flight:
 394                  */
 395                 trace_journal_entry_full(c);
 396                 ret = cur_entry_blocked;
 397         } else {
 398                 ret = journal_entry_open(j);
 399         }
 400 unlock:
 401         if ((ret && ret != cur_entry_insufficient_devices) &&
 402             !j->res_get_blocked_start) {
 403                 j->res_get_blocked_start = local_clock() ?: 1;
 404                 trace_journal_full(c);
 405         }
 406
 407         can_discard = j->can_discard;
 408         spin_unlock(&j->lock);
 409
 410         if (!ret)
 411                 goto retry;
 412
 413         if ((ret == cur_entry_journal_full ||
 414              ret == cur_entry_journal_pin_full) &&
 415             !can_discard &&
 416             j->reservations.idx == j->reservations.unwritten_idx &&
 417             (flags & JOURNAL_RES_GET_RESERVED)) {
 418                 char *journal_debug_buf = kmalloc(4096, GFP_ATOMIC);
 419
 420                 bch_err(c, "Journal stuck!");
 421                 if (journal_debug_buf) {
 422                         bch2_journal_debug_to_text(&_PBUF(journal_debug_buf, 4096), j);
 423                         bch_err(c, "%s", journal_debug_buf);
 424
 425                         bch2_journal_pins_to_text(&_PBUF(journal_debug_buf, 4096), j);
 426                         bch_err(c, "Journal pins:\n%s", journal_debug_buf);
 427                         kfree(journal_debug_buf);
 428                 }
 429
 430                 bch2_fatal_error(c);
 431                 dump_stack();
 432         }
 433
 434         /*
 435          * Journal is full - can't rely on reclaim from work item due to
 436          * freezing:
 437          */
 438         if ((ret == cur_entry_journal_full ||
 439              ret == cur_entry_journal_pin_full) &&
 440             !(flags & JOURNAL_RES_GET_NONBLOCK)) {
 441                 if (can_discard) {
 442                         bch2_journal_do_discards(j);
 443                         goto retry;
 444                 }
 445
 446                 if (mutex_trylock(&j->reclaim_lock)) {
 447                         bch2_journal_reclaim(j);
 448                         mutex_unlock(&j->reclaim_lock);
 449                 }
 450         }
 451
 452         return ret == cur_entry_insufficient_devices ? -EROFS : -EAGAIN;
 453 }
 454
 455 /*
 456  * Essentially the entry function to the journaling code. When bcachefs is doing
 457  * a btree insert, it calls this function to get the current journal write.
 458  * Journal write is the structure used set up journal writes. The calling
 459  * function will then add its keys to the structure, queuing them for the next
 460  * write.
 461  *
 462  * To ensure forward progress, the current task must not be holding any
 463  * btree node write locks.
 464  */
 465 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
 466                                   unsigned flags)
 467 {
 468         int ret;
 469
 470         closure_wait_event(&j->async_wait,
 471                    (ret = __journal_res_get(j, res, flags)) != -EAGAIN ||
 472                    (flags & JOURNAL_RES_GET_NONBLOCK));
 473         return ret;
 474 }
 475
 476 /* journal_preres: */
 477
 478 static bool journal_preres_available(struct journal *j,
 479                                      struct journal_preres *res,
 480                                      unsigned new_u64s,
 481                                      unsigned flags)
 482 {
 483         bool ret = bch2_journal_preres_get_fast(j, res, new_u64s, flags, true);
 484
 485         if (!ret && mutex_trylock(&j->reclaim_lock)) {
 486                 bch2_journal_reclaim(j);
 487                 mutex_unlock(&j->reclaim_lock);
 488         }
 489
 490         return ret;
 491 }
 492
 493 int __bch2_journal_preres_get(struct journal *j,
 494                               struct journal_preres *res,
 495                               unsigned new_u64s,
 496                               unsigned flags)
 497 {
 498         int ret;
 499
 500         closure_wait_event(&j->preres_wait,
 501                    (ret = bch2_journal_error(j)) ||
 502                    journal_preres_available(j, res, new_u64s, flags));
 503         return ret;
 504 }
 505
 506 /* journal_entry_res: */
 507
 508 void bch2_journal_entry_res_resize(struct journal *j,
 509                                    struct journal_entry_res *res,
 510                                    unsigned new_u64s)
 511 {
 512         union journal_res_state state;
 513         int d = new_u64s - res->u64s;
 514
 515         spin_lock(&j->lock);
 516
 517         j->entry_u64s_reserved += d;
 518         if (d <= 0)
 519                 goto out;
 520
 521         j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
 522         smp_mb();
 523         state = READ_ONCE(j->reservations);
 524
 525         if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
 526             state.cur_entry_offset > j->cur_entry_u64s) {
 527                 j->cur_entry_u64s += d;
 528                 /*
 529                  * Not enough room in current journal entry, have to flush it:
 530                  */
 531                 __journal_entry_close(j);
 532         } else {
 533                 journal_cur_buf(j)->u64s_reserved += d;
 534         }
 535 out:
 536         spin_unlock(&j->lock);
 537         res->u64s += d;
 538 }
 539
 540 /* journal flushing: */
 541
 542 /**
 543  * bch2_journal_flush_seq_async - wait for a journal entry to be written
 544  *
 545  * like bch2_journal_wait_on_seq, except that it triggers a write immediately if
 546  * necessary
 547  */
 548 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
 549                                  struct closure *parent)
 550 {
 551         struct journal_buf *buf;
 552         int ret = 0;
 553
 554         if (seq <= j->flushed_seq_ondisk)
 555                 return 1;
 556
 557         spin_lock(&j->lock);
 558
 559         if (WARN_ONCE(seq > journal_cur_seq(j),
 560                       "requested to flush journal seq %llu, but currently at %llu",
 561                       seq, journal_cur_seq(j)))
 562                 goto out;
 563
 564         /* Recheck under lock: */
 565         if (j->err_seq && seq >= j->err_seq) {
 566                 ret = -EIO;
 567                 goto out;
 568         }
 569
 570         if (seq <= j->flushed_seq_ondisk) {
 571                 ret = 1;
 572                 goto out;
 573         }
 574
 575         /* if seq was written, but not flushed - flush a newer one instead */
 576         seq = max(seq, last_unwritten_seq(j));
 577
 578 recheck_need_open:
 579         if (seq == journal_cur_seq(j) && !journal_entry_is_open(j)) {
 580                 struct journal_res res = { 0 };
 581
 582                 spin_unlock(&j->lock);
 583
 584                 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 585                 if (ret)
 586                         return ret;
 587
 588                 seq = res.seq;
 589                 buf = j->buf + (seq & JOURNAL_BUF_MASK);
 590                 buf->must_flush = true;
 591                 set_bit(JOURNAL_NEED_WRITE, &j->flags);
 592
 593                 if (parent && !closure_wait(&buf->wait, parent))
 594                         BUG();
 595
 596                 bch2_journal_res_put(j, &res);
 597
 598                 spin_lock(&j->lock);
 599                 goto want_write;
 600         }
 601
 602         /*
 603          * if write was kicked off without a flush, flush the next sequence
 604          * number instead
 605          */
 606         buf = journal_seq_to_buf(j, seq);
 607         if (buf->noflush) {
 608                 seq++;
 609                 goto recheck_need_open;
 610         }
 611
 612         buf->must_flush = true;
 613
 614         if (parent && !closure_wait(&buf->wait, parent))
 615                 BUG();
 616 want_write:
 617         if (seq == journal_cur_seq(j))
 618                 journal_entry_want_write(j);
 619 out:
 620         spin_unlock(&j->lock);
 621         return ret;
 622 }
 623
 624 int bch2_journal_flush_seq(struct journal *j, u64 seq)
 625 {
 626         u64 start_time = local_clock();
 627         int ret, ret2;
 628
 629         /*
 630          * Don't update time_stats when @seq is already flushed:
 631          */
 632         if (seq <= j->flushed_seq_ondisk)
 633                 return 0;
 634
 635         ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
 636
 637         if (!ret)
 638                 bch2_time_stats_update(j->flush_seq_time, start_time);
 639
 640         return ret ?: ret2 < 0 ? ret2 : 0;
 641 }
 642
 643 int bch2_journal_meta(struct journal *j)
 644 {
 645         struct journal_res res;
 646         int ret;
 647
 648         memset(&res, 0, sizeof(res));
 649
 650         ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 651         if (ret)
 652                 return ret;
 653
 654         bch2_journal_res_put(j, &res);
 655
 656         return bch2_journal_flush_seq(j, res.seq);
 657 }
 658
 659 /*
 660  * bch2_journal_flush_async - if there is an open journal entry, or a journal
 661  * still being written, write it and wait for the write to complete
 662  */
 663 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
 664 {
 665         u64 seq, journal_seq;
 666
 667         spin_lock(&j->lock);
 668         journal_seq = journal_cur_seq(j);
 669
 670         if (journal_entry_is_open(j)) {
 671                 seq = journal_seq;
 672         } else if (journal_seq) {
 673                 seq = journal_seq - 1;
 674         } else {
 675                 spin_unlock(&j->lock);
 676                 return;
 677         }
 678         spin_unlock(&j->lock);
 679
 680         bch2_journal_flush_seq_async(j, seq, parent);
 681 }
 682
 683 int bch2_journal_flush(struct journal *j)
 684 {
 685         u64 seq, journal_seq;
 686
 687         spin_lock(&j->lock);
 688         journal_seq = journal_cur_seq(j);
 689
 690         if (journal_entry_is_open(j)) {
 691                 seq = journal_seq;
 692         } else if (journal_seq) {
 693                 seq = journal_seq - 1;
 694         } else {
 695                 spin_unlock(&j->lock);
 696                 return 0;
 697         }
 698         spin_unlock(&j->lock);
 699
 700         return bch2_journal_flush_seq(j, seq);
 701 }
 702
 703 /* block/unlock the journal: */
 704
 705 void bch2_journal_unblock(struct journal *j)
 706 {
 707         spin_lock(&j->lock);
 708         j->blocked--;
 709         spin_unlock(&j->lock);
 710
 711         journal_wake(j);
 712 }
 713
 714 void bch2_journal_block(struct journal *j)
 715 {
 716         spin_lock(&j->lock);
 717         j->blocked++;
 718         spin_unlock(&j->lock);
 719
 720         journal_quiesce(j);
 721 }
 722
 723 /* allocate journal on a device: */
 724
 725 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 726                                          bool new_fs, struct closure *cl)
 727 {
 728         struct bch_fs *c = ca->fs;
 729         struct journal_device *ja = &ca->journal;
 730         struct bch_sb_field_journal *journal_buckets;
 731         u64 *new_bucket_seq = NULL, *new_buckets = NULL;
 732         int ret = 0;
 733
 734         /* don't handle reducing nr of buckets yet: */
 735         if (nr <= ja->nr)
 736                 return 0;
 737
 738         new_buckets     = kzalloc(nr * sizeof(u64), GFP_KERNEL);
 739         new_bucket_seq  = kzalloc(nr * sizeof(u64), GFP_KERNEL);
 740         if (!new_buckets || !new_bucket_seq) {
 741                 ret = -ENOMEM;
 742                 goto err;
 743         }
 744
 745         journal_buckets = bch2_sb_resize_journal(&ca->disk_sb,
 746                                         nr + sizeof(*journal_buckets) / sizeof(u64));
 747         if (!journal_buckets) {
 748                 ret = -ENOSPC;
 749                 goto err;
 750         }
 751
 752         /*
 753          * We may be called from the device add path, before the new device has
 754          * actually been added to the running filesystem:
 755          */
 756         if (!new_fs)
 757                 spin_lock(&c->journal.lock);
 758
 759         memcpy(new_buckets,     ja->buckets,    ja->nr * sizeof(u64));
 760         memcpy(new_bucket_seq,  ja->bucket_seq, ja->nr * sizeof(u64));
 761         swap(new_buckets,       ja->buckets);
 762         swap(new_bucket_seq,    ja->bucket_seq);
 763
 764         if (!new_fs)
 765                 spin_unlock(&c->journal.lock);
 766
 767         while (ja->nr < nr) {
 768                 struct open_bucket *ob = NULL;
 769                 unsigned pos;
 770                 long b;
 771
 772                 if (new_fs) {
 773                         if (c)
 774                                 percpu_down_read(&c->mark_lock);
 775                         b = bch2_bucket_alloc_new_fs(ca);
 776                         if (b < 0) {
 777                                 percpu_up_read(&c->mark_lock);
 778                                 ret = -ENOSPC;
 779                                 goto err;
 780                         }
 781                 } else {
 782                         rcu_read_lock();
 783                         ob = bch2_bucket_alloc(c, ca, RESERVE_NONE,
 784                                                false, cl);
 785                         rcu_read_unlock();
 786                         if (IS_ERR(ob)) {
 787                                 ret = cl ? -EAGAIN : -ENOSPC;
 788                                 goto err;
 789                         }
 790
 791                         b = sector_to_bucket(ca, ob->ptr.offset);
 792                 }
 793
 794                 if (c)
 795                         spin_lock(&c->journal.lock);
 796
 797                 /*
 798                  * XXX
 799                  * For resize at runtime, we should be writing the new
 800                  * superblock before inserting into the journal array
 801                  */
 802
 803                 pos = ja->nr ? (ja->cur_idx + 1) % ja->nr : 0;
 804                 __array_insert_item(ja->buckets,                ja->nr, pos);
 805                 __array_insert_item(ja->bucket_seq,             ja->nr, pos);
 806                 __array_insert_item(journal_buckets->buckets,   ja->nr, pos);
 807                 ja->nr++;
 808
 809                 ja->buckets[pos] = b;
 810                 ja->bucket_seq[pos] = 0;
 811                 journal_buckets->buckets[pos] = cpu_to_le64(b);
 812
 813                 if (pos <= ja->discard_idx)
 814                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 815                 if (pos <= ja->dirty_idx_ondisk)
 816                         ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
 817                 if (pos <= ja->dirty_idx)
 818                         ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
 819                 if (pos <= ja->cur_idx)
 820                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
 821
 822                 if (c)
 823                         spin_unlock(&c->journal.lock);
 824
 825                 if (new_fs) {
 826                         bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
 827                                                   ca->mi.bucket_size,
 828                                                   gc_phase(GC_PHASE_SB),
 829                                                   0);
 830                         if (c)
 831                                 percpu_up_read(&c->mark_lock);
 832                 } else {
 833                         ret = bch2_trans_do(c, NULL, NULL, BTREE_INSERT_NOFAIL,
 834                                 bch2_trans_mark_metadata_bucket(&trans, ca,
 835                                                 b, BCH_DATA_journal,
 836                                                 ca->mi.bucket_size));
 837
 838                         bch2_open_bucket_put(c, ob);
 839
 840                         if (ret)
 841                                 goto err;
 842                 }
 843         }
 844 err:
 845         bch2_sb_resize_journal(&ca->disk_sb,
 846                 ja->nr + sizeof(*journal_buckets) / sizeof(u64));
 847         kfree(new_bucket_seq);
 848         kfree(new_buckets);
 849
 850         return ret;
 851 }
 852
 853 /*
 854  * Allocate more journal space at runtime - not currently making use if it, but
 855  * the code works:
 856  */
 857 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
 858                                 unsigned nr)
 859 {
 860         struct journal_device *ja = &ca->journal;
 861         struct closure cl;
 862         unsigned current_nr;
 863         int ret;
 864
 865         closure_init_stack(&cl);
 866
 867         do {
 868                 struct disk_reservation disk_res = { 0, 0 };
 869
 870                 closure_sync(&cl);
 871
 872                 mutex_lock(&c->sb_lock);
 873                 current_nr = ja->nr;
 874
 875                 /*
 876                  * note: journal buckets aren't really counted as _sectors_ used yet, so
 877                  * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
 878                  * when space used goes up without a reservation - but we do need the
 879                  * reservation to ensure we'll actually be able to allocate:
 880                  */
 881
 882                 if (bch2_disk_reservation_get(c, &disk_res,
 883                                               bucket_to_sector(ca, nr - ja->nr), 1, 0)) {
 884                         mutex_unlock(&c->sb_lock);
 885                         return -ENOSPC;
 886                 }
 887
 888                 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
 889
 890                 bch2_disk_reservation_put(c, &disk_res);
 891
 892                 if (ja->nr != current_nr)
 893                         bch2_write_super(c);
 894                 mutex_unlock(&c->sb_lock);
 895         } while (ret == -EAGAIN);
 896
 897         return ret;
 898 }
 899
 900 int bch2_dev_journal_alloc(struct bch_dev *ca)
 901 {
 902         unsigned nr;
 903
 904         if (dynamic_fault("bcachefs:add:journal_alloc"))
 905                 return -ENOMEM;
 906
 907         /* 1/128th of the device by default: */
 908         nr = ca->mi.nbuckets >> 7;
 909
 910         /*
 911          * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
 912          * is smaller:
 913          */
 914         nr = clamp_t(unsigned, nr,
 915                      BCH_JOURNAL_BUCKETS_MIN,
 916                      min(1 << 13,
 917                          (1 << 24) / ca->mi.bucket_size));
 918
 919         return __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
 920 }
 921
 922 /* startup/shutdown: */
 923
 924 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
 925 {
 926         union journal_res_state state;
 927         bool ret = false;
 928         unsigned i;
 929
 930         spin_lock(&j->lock);
 931         state = READ_ONCE(j->reservations);
 932         i = state.idx;
 933
 934         while (i != state.unwritten_idx) {
 935                 i = (i - 1) & JOURNAL_BUF_MASK;
 936                 if (bch2_bkey_has_device(bkey_i_to_s_c(&j->buf[i].key), dev_idx))
 937                         ret = true;
 938         }
 939         spin_unlock(&j->lock);
 940
 941         return ret;
 942 }
 943
 944 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
 945 {
 946         wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
 947 }
 948
 949 void bch2_fs_journal_stop(struct journal *j)
 950 {
 951         bch2_journal_flush_all_pins(j);
 952
 953         wait_event(j->wait, journal_entry_close(j));
 954
 955         /*
 956          * Always write a new journal entry, to make sure the clock hands are up
 957          * to date (and match the superblock)
 958          */
 959         bch2_journal_meta(j);
 960
 961         journal_quiesce(j);
 962
 963         BUG_ON(!bch2_journal_error(j) &&
 964                test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
 965                (journal_entry_is_open(j) ||
 966                 j->last_empty_seq + 1 != journal_cur_seq(j)));
 967
 968         cancel_delayed_work_sync(&j->write_work);
 969         bch2_journal_reclaim_stop(j);
 970 }
 971
 972 int bch2_fs_journal_start(struct journal *j, u64 cur_seq,
 973                           struct list_head *journal_entries)
 974 {
 975         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 976         struct journal_entry_pin_list *p;
 977         struct journal_replay *i;
 978         u64 last_seq = cur_seq, nr, seq;
 979
 980         if (!list_empty(journal_entries))
 981                 last_seq = le64_to_cpu(list_last_entry(journal_entries,
 982                                 struct journal_replay, list)->j.last_seq);
 983
 984         nr = cur_seq - last_seq;
 985
 986         if (nr + 1 > j->pin.size) {
 987                 free_fifo(&j->pin);
 988                 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
 989                 if (!j->pin.data) {
 990                         bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
 991                         return -ENOMEM;
 992                 }
 993         }
 994
 995         j->replay_journal_seq   = last_seq;
 996         j->replay_journal_seq_end = cur_seq;
 997         j->last_seq_ondisk      = last_seq;
 998         j->pin.front            = last_seq;
 999         j->pin.back             = cur_seq;
1000         atomic64_set(&j->seq, cur_seq - 1);
1001
1002         fifo_for_each_entry_ptr(p, &j->pin, seq)
1003                 journal_pin_list_init(p, 1);
1004
1005         list_for_each_entry(i, journal_entries, list) {
1006                 unsigned ptr;
1007
1008                 seq = le64_to_cpu(i->j.seq);
1009                 BUG_ON(seq >= cur_seq);
1010
1011                 if (seq < last_seq)
1012                         continue;
1013
1014                 p = journal_seq_pin(j, seq);
1015
1016                 p->devs.nr = 0;
1017                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1018                         bch2_dev_list_add_dev(&p->devs, i->ptrs[ptr].dev);
1019         }
1020
1021         spin_lock(&j->lock);
1022
1023         set_bit(JOURNAL_STARTED, &j->flags);
1024         j->last_flush_write = jiffies;
1025
1026         journal_pin_new_entry(j);
1027
1028         j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1029
1030         bch2_journal_buf_init(j);
1031
1032         c->last_bucket_seq_cleanup = journal_cur_seq(j);
1033
1034         bch2_journal_space_available(j);
1035         spin_unlock(&j->lock);
1036
1037         return bch2_journal_reclaim_start(j);
1038 }
1039
1040 /* init/exit: */
1041
1042 void bch2_dev_journal_exit(struct bch_dev *ca)
1043 {
1044         kfree(ca->journal.bio);
1045         kfree(ca->journal.buckets);
1046         kfree(ca->journal.bucket_seq);
1047
1048         ca->journal.bio         = NULL;
1049         ca->journal.buckets     = NULL;
1050         ca->journal.bucket_seq  = NULL;
1051 }
1052
1053 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1054 {
1055         struct journal_device *ja = &ca->journal;
1056         struct bch_sb_field_journal *journal_buckets =
1057                 bch2_sb_get_journal(sb);
1058         unsigned i;
1059
1060         ja->nr = bch2_nr_journal_buckets(journal_buckets);
1061
1062         ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1063         if (!ja->bucket_seq)
1064                 return -ENOMEM;
1065
1066         ca->journal.bio = bio_kmalloc(GFP_KERNEL,
1067                         DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE));
1068         if (!ca->journal.bio)
1069                 return -ENOMEM;
1070
1071         ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1072         if (!ja->buckets)
1073                 return -ENOMEM;
1074
1075         for (i = 0; i < ja->nr; i++)
1076                 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1077
1078         return 0;
1079 }
1080
1081 void bch2_fs_journal_exit(struct journal *j)
1082 {
1083         unsigned i;
1084
1085         for (i = 0; i < ARRAY_SIZE(j->buf); i++)
1086                 kvpfree(j->buf[i].data, j->buf[i].buf_size);
1087         free_fifo(&j->pin);
1088 }
1089
1090 int bch2_fs_journal_init(struct journal *j)
1091 {
1092         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1093         static struct lock_class_key res_key;
1094         unsigned i;
1095         int ret = 0;
1096
1097         pr_verbose_init(c->opts, "");
1098
1099         spin_lock_init(&j->lock);
1100         spin_lock_init(&j->err_lock);
1101         init_waitqueue_head(&j->wait);
1102         INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1103         init_waitqueue_head(&j->reclaim_wait);
1104         init_waitqueue_head(&j->pin_flush_wait);
1105         mutex_init(&j->reclaim_lock);
1106         mutex_init(&j->discard_lock);
1107
1108         lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1109
1110         atomic64_set(&j->reservations.counter,
1111                 ((union journal_res_state)
1112                  { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1113
1114         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL))) {
1115                 ret = -ENOMEM;
1116                 goto out;
1117         }
1118
1119         for (i = 0; i < ARRAY_SIZE(j->buf); i++) {
1120                 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1121                 j->buf[i].data = kvpmalloc(j->buf[i].buf_size, GFP_KERNEL);
1122                 if (!j->buf[i].data) {
1123                         ret = -ENOMEM;
1124                         goto out;
1125                 }
1126         }
1127
1128         j->pin.front = j->pin.back = 1;
1129 out:
1130         pr_verbose_init(c->opts, "ret %i", ret);
1131         return ret;
1132 }
1133
1134 /* debug: */
1135
1136 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1137 {
1138         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1139         union journal_res_state s;
1140         struct bch_dev *ca;
1141         unsigned long now = jiffies;
1142         unsigned i;
1143
1144         rcu_read_lock();
1145         s = READ_ONCE(j->reservations);
1146
1147         pr_buf(out, "active journal entries:\t%llu\n",  fifo_used(&j->pin));
1148         pr_buf(out, "seq:\t\t\t%llu\n",                 journal_cur_seq(j));
1149         pr_buf(out, "last_seq:\t\t%llu\n",              journal_last_seq(j));
1150         pr_buf(out, "last_seq_ondisk:\t%llu\n",         j->last_seq_ondisk);
1151         pr_buf(out, "flushed_seq_ondisk:\t%llu\n",      j->flushed_seq_ondisk);
1152         pr_buf(out, "prereserved:\t\t%u/%u\n",          j->prereserved.reserved, j->prereserved.remaining);
1153         pr_buf(out, "each entry reserved:\t%u\n",       j->entry_u64s_reserved);
1154         pr_buf(out, "nr flush writes:\t%llu\n",         j->nr_flush_writes);
1155         pr_buf(out, "nr noflush writes:\t%llu\n",       j->nr_noflush_writes);
1156         pr_buf(out, "nr direct reclaim:\t%llu\n",       j->nr_direct_reclaim);
1157         pr_buf(out, "nr background reclaim:\t%llu\n",   j->nr_background_reclaim);
1158         pr_buf(out, "reclaim kicked:\t\t%u\n",          j->reclaim_kicked);
1159         pr_buf(out, "reclaim runs in:\t%u ms\n",        time_after(j->next_reclaim, now)
1160                ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1161         pr_buf(out, "current entry sectors:\t%u\n",     j->cur_entry_sectors);
1162         pr_buf(out, "current entry error:\t%u\n",       j->cur_entry_error);
1163         pr_buf(out, "current entry:\t\t");
1164
1165         switch (s.cur_entry_offset) {
1166         case JOURNAL_ENTRY_ERROR_VAL:
1167                 pr_buf(out, "error\n");
1168                 break;
1169         case JOURNAL_ENTRY_CLOSED_VAL:
1170                 pr_buf(out, "closed\n");
1171                 break;
1172         default:
1173                 pr_buf(out, "%u/%u\n", s.cur_entry_offset, j->cur_entry_u64s);
1174                 break;
1175         }
1176
1177         pr_buf(out, "current entry:\t\tidx %u refcount %u\n", s.idx, journal_state_count(s, s.idx));
1178
1179         i = s.idx;
1180         while (i != s.unwritten_idx) {
1181                 i = (i - 1) & JOURNAL_BUF_MASK;
1182
1183                 pr_buf(out, "unwritten entry:\tidx %u refcount %u sectors %u\n",
1184                        i, journal_state_count(s, i), j->buf[i].sectors);
1185         }
1186
1187         pr_buf(out,
1188                "need write:\t\t%i\n"
1189                "replay done:\t\t%i\n",
1190                test_bit(JOURNAL_NEED_WRITE,     &j->flags),
1191                test_bit(JOURNAL_REPLAY_DONE,    &j->flags));
1192
1193         pr_buf(out, "space:\n");
1194         pr_buf(out, "\tdiscarded\t%u:%u\n",
1195                j->space[journal_space_discarded].next_entry,
1196                j->space[journal_space_discarded].total);
1197         pr_buf(out, "\tclean ondisk\t%u:%u\n",
1198                j->space[journal_space_clean_ondisk].next_entry,
1199                j->space[journal_space_clean_ondisk].total);
1200         pr_buf(out, "\tclean\t\t%u:%u\n",
1201                j->space[journal_space_clean].next_entry,
1202                j->space[journal_space_clean].total);
1203         pr_buf(out, "\ttotal\t\t%u:%u\n",
1204                j->space[journal_space_total].next_entry,
1205                j->space[journal_space_total].total);
1206
1207         for_each_member_device_rcu(ca, c, i,
1208                                    &c->rw_devs[BCH_DATA_journal]) {
1209                 struct journal_device *ja = &ca->journal;
1210
1211                 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1212                         continue;
1213
1214                 if (!ja->nr)
1215                         continue;
1216
1217                 pr_buf(out, "dev %u:\n",                i);
1218                 pr_buf(out, "\tnr\t\t%u\n",             ja->nr);
1219                 pr_buf(out, "\tbucket size\t%u\n",      ca->mi.bucket_size);
1220                 pr_buf(out, "\tavailable\t%u:%u\n",     bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1221                 pr_buf(out, "\tdiscard_idx\t%u\n",      ja->discard_idx);
1222                 pr_buf(out, "\tdirty_ondisk\t%u (seq %llu)\n", ja->dirty_idx_ondisk,    ja->bucket_seq[ja->dirty_idx_ondisk]);
1223                 pr_buf(out, "\tdirty_idx\t%u (seq %llu)\n", ja->dirty_idx,              ja->bucket_seq[ja->dirty_idx]);
1224                 pr_buf(out, "\tcur_idx\t\t%u (seq %llu)\n", ja->cur_idx,                ja->bucket_seq[ja->cur_idx]);
1225         }
1226
1227         rcu_read_unlock();
1228 }
1229
1230 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1231 {
1232         spin_lock(&j->lock);
1233         __bch2_journal_debug_to_text(out, j);
1234         spin_unlock(&j->lock);
1235 }
1236
1237 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1238 {
1239         struct journal_entry_pin_list *pin_list;
1240         struct journal_entry_pin *pin;
1241         u64 i;
1242
1243         spin_lock(&j->lock);
1244         fifo_for_each_entry_ptr(pin_list, &j->pin, i) {
1245                 pr_buf(out, "%llu: count %u\n",
1246                        i, atomic_read(&pin_list->count));
1247
1248                 list_for_each_entry(pin, &pin_list->list, list)
1249                         pr_buf(out, "\t%px %ps\n",
1250                                pin, pin->flush);
1251
1252                 if (!list_empty(&pin_list->flushed))
1253                         pr_buf(out, "flushed:\n");
1254
1255                 list_for_each_entry(pin, &pin_list->flushed, list)
1256                         pr_buf(out, "\t%px %ps\n",
1257                                pin, pin->flush);
1258         }
1259         spin_unlock(&j->lock);
1260 }