fs/bcachefs/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 "btree_write_buffer.h"
  14 #include "buckets.h"
  15 #include "error.h"
  16 #include "journal.h"
  17 #include "journal_io.h"
  18 #include "journal_reclaim.h"
  19 #include "journal_sb.h"
  20 #include "journal_seq_blacklist.h"
  21 #include "trace.h"
  22
  23 static const char * const bch2_journal_errors[] = {
  24 #define x(n)    #n,
  25         JOURNAL_ERRORS()
  26 #undef x
  27         NULL
  28 };
  29
  30 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
  31 {
  32         return seq > j->seq_ondisk;
  33 }
  34
  35 static bool __journal_entry_is_open(union journal_res_state state)
  36 {
  37         return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
  38 }
  39
  40 static inline unsigned nr_unwritten_journal_entries(struct journal *j)
  41 {
  42         return atomic64_read(&j->seq) - j->seq_ondisk;
  43 }
  44
  45 static bool journal_entry_is_open(struct journal *j)
  46 {
  47         return __journal_entry_is_open(j->reservations);
  48 }
  49
  50 static void bch2_journal_buf_to_text(struct printbuf *out, struct journal *j, u64 seq)
  51 {
  52         union journal_res_state s = READ_ONCE(j->reservations);
  53         unsigned i = seq & JOURNAL_BUF_MASK;
  54         struct journal_buf *buf = j->buf + i;
  55
  56         prt_printf(out, "seq:\t%llu\n", seq);
  57         printbuf_indent_add(out, 2);
  58
  59         prt_printf(out, "refcount:\t%u\n", journal_state_count(s, i));
  60
  61         prt_printf(out, "size:\t");
  62         prt_human_readable_u64(out, vstruct_bytes(buf->data));
  63         prt_newline(out);
  64
  65         prt_printf(out, "expires:\t");
  66         prt_printf(out, "%li jiffies\n", buf->expires - jiffies);
  67
  68         prt_printf(out, "flags:\t");
  69         if (buf->noflush)
  70                 prt_str(out, "noflush ");
  71         if (buf->must_flush)
  72                 prt_str(out, "must_flush ");
  73         if (buf->separate_flush)
  74                 prt_str(out, "separate_flush ");
  75         if (buf->need_flush_to_write_buffer)
  76                 prt_str(out, "need_flush_to_write_buffer ");
  77         if (buf->write_started)
  78                 prt_str(out, "write_started ");
  79         if (buf->write_allocated)
  80                 prt_str(out, "write_allocated ");
  81         if (buf->write_done)
  82                 prt_str(out, "write_done");
  83         prt_newline(out);
  84
  85         printbuf_indent_sub(out, 2);
  86 }
  87
  88 static void bch2_journal_bufs_to_text(struct printbuf *out, struct journal *j)
  89 {
  90         if (!out->nr_tabstops)
  91                 printbuf_tabstop_push(out, 24);
  92
  93         for (u64 seq = journal_last_unwritten_seq(j);
  94              seq <= journal_cur_seq(j);
  95              seq++)
  96                 bch2_journal_buf_to_text(out, j, seq);
  97         prt_printf(out, "last buf %s\n", journal_entry_is_open(j) ? "open" : "closed");
  98 }
  99
 100 static inline struct journal_buf *
 101 journal_seq_to_buf(struct journal *j, u64 seq)
 102 {
 103         struct journal_buf *buf = NULL;
 104
 105         EBUG_ON(seq > journal_cur_seq(j));
 106
 107         if (journal_seq_unwritten(j, seq)) {
 108                 buf = j->buf + (seq & JOURNAL_BUF_MASK);
 109                 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
 110         }
 111         return buf;
 112 }
 113
 114 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
 115 {
 116         unsigned i;
 117
 118         for (i = 0; i < ARRAY_SIZE(p->list); i++)
 119                 INIT_LIST_HEAD(&p->list[i]);
 120         INIT_LIST_HEAD(&p->flushed);
 121         atomic_set(&p->count, count);
 122         p->devs.nr = 0;
 123 }
 124
 125 /*
 126  * Detect stuck journal conditions and trigger shutdown. Technically the journal
 127  * can end up stuck for a variety of reasons, such as a blocked I/O, journal
 128  * reservation lockup, etc. Since this is a fatal error with potentially
 129  * unpredictable characteristics, we want to be fairly conservative before we
 130  * decide to shut things down.
 131  *
 132  * Consider the journal stuck when it appears full with no ability to commit
 133  * btree transactions, to discard journal buckets, nor acquire priority
 134  * (reserved watermark) reservation.
 135  */
 136 static inline bool
 137 journal_error_check_stuck(struct journal *j, int error, unsigned flags)
 138 {
 139         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 140         bool stuck = false;
 141         struct printbuf buf = PRINTBUF;
 142
 143         if (!(error == JOURNAL_ERR_journal_full ||
 144               error == JOURNAL_ERR_journal_pin_full) ||
 145             nr_unwritten_journal_entries(j) ||
 146             (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim)
 147                 return stuck;
 148
 149         spin_lock(&j->lock);
 150
 151         if (j->can_discard) {
 152                 spin_unlock(&j->lock);
 153                 return stuck;
 154         }
 155
 156         stuck = true;
 157
 158         /*
 159          * The journal shutdown path will set ->err_seq, but do it here first to
 160          * serialize against concurrent failures and avoid duplicate error
 161          * reports.
 162          */
 163         if (j->err_seq) {
 164                 spin_unlock(&j->lock);
 165                 return stuck;
 166         }
 167         j->err_seq = journal_cur_seq(j);
 168         spin_unlock(&j->lock);
 169
 170         bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)",
 171                 bch2_journal_errors[error]);
 172         bch2_journal_debug_to_text(&buf, j);
 173         bch_err(c, "%s", buf.buf);
 174
 175         printbuf_reset(&buf);
 176         bch2_journal_pins_to_text(&buf, j);
 177         bch_err(c, "Journal pins:\n%s", buf.buf);
 178         printbuf_exit(&buf);
 179
 180         bch2_fatal_error(c);
 181         dump_stack();
 182
 183         return stuck;
 184 }
 185
 186 void bch2_journal_do_writes(struct journal *j)
 187 {
 188         for (u64 seq = journal_last_unwritten_seq(j);
 189              seq <= journal_cur_seq(j);
 190              seq++) {
 191                 unsigned idx = seq & JOURNAL_BUF_MASK;
 192                 struct journal_buf *w = j->buf + idx;
 193
 194                 if (w->write_started && !w->write_allocated)
 195                         break;
 196                 if (w->write_started)
 197                         continue;
 198
 199                 if (!journal_state_count(j->reservations, idx)) {
 200                         w->write_started = true;
 201                         closure_call(&w->io, bch2_journal_write, j->wq, NULL);
 202                 }
 203
 204                 break;
 205         }
 206 }
 207
 208 /*
 209  * Final processing when the last reference of a journal buffer has been
 210  * dropped. Drop the pin list reference acquired at journal entry open and write
 211  * the buffer, if requested.
 212  */
 213 void bch2_journal_buf_put_final(struct journal *j, u64 seq)
 214 {
 215         lockdep_assert_held(&j->lock);
 216
 217         if (__bch2_journal_pin_put(j, seq))
 218                 bch2_journal_reclaim_fast(j);
 219         bch2_journal_do_writes(j);
 220 }
 221
 222 /*
 223  * Returns true if journal entry is now closed:
 224  *
 225  * We don't close a journal_buf until the next journal_buf is finished writing,
 226  * and can be opened again - this also initializes the next journal_buf:
 227  */
 228 static void __journal_entry_close(struct journal *j, unsigned closed_val, bool trace)
 229 {
 230         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 231         struct journal_buf *buf = journal_cur_buf(j);
 232         union journal_res_state old, new;
 233         unsigned sectors;
 234
 235         BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
 236                closed_val != JOURNAL_ENTRY_ERROR_VAL);
 237
 238         lockdep_assert_held(&j->lock);
 239
 240         old.v = atomic64_read(&j->reservations.counter);
 241         do {
 242                 new.v = old.v;
 243                 new.cur_entry_offset = closed_val;
 244
 245                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
 246                     old.cur_entry_offset == new.cur_entry_offset)
 247                         return;
 248         } while (!atomic64_try_cmpxchg(&j->reservations.counter,
 249                                        &old.v, new.v));
 250
 251         if (!__journal_entry_is_open(old))
 252                 return;
 253
 254         /* Close out old buffer: */
 255         buf->data->u64s         = cpu_to_le32(old.cur_entry_offset);
 256
 257         if (trace_journal_entry_close_enabled() && trace) {
 258                 struct printbuf pbuf = PRINTBUF;
 259                 pbuf.atomic++;
 260
 261                 prt_str(&pbuf, "entry size: ");
 262                 prt_human_readable_u64(&pbuf, vstruct_bytes(buf->data));
 263                 prt_newline(&pbuf);
 264                 bch2_prt_task_backtrace(&pbuf, current, 1, GFP_NOWAIT);
 265                 trace_journal_entry_close(c, pbuf.buf);
 266                 printbuf_exit(&pbuf);
 267         }
 268
 269         sectors = vstruct_blocks_plus(buf->data, c->block_bits,
 270                                       buf->u64s_reserved) << c->block_bits;
 271         BUG_ON(sectors > buf->sectors);
 272         buf->sectors = sectors;
 273
 274         /*
 275          * We have to set last_seq here, _before_ opening a new journal entry:
 276          *
 277          * A threads may replace an old pin with a new pin on their current
 278          * journal reservation - the expectation being that the journal will
 279          * contain either what the old pin protected or what the new pin
 280          * protects.
 281          *
 282          * After the old pin is dropped journal_last_seq() won't include the old
 283          * pin, so we can only write the updated last_seq on the entry that
 284          * contains whatever the new pin protects.
 285          *
 286          * Restated, we can _not_ update last_seq for a given entry if there
 287          * could be a newer entry open with reservations/pins that have been
 288          * taken against it.
 289          *
 290          * Hence, we want update/set last_seq on the current journal entry right
 291          * before we open a new one:
 292          */
 293         buf->last_seq           = journal_last_seq(j);
 294         buf->data->last_seq     = cpu_to_le64(buf->last_seq);
 295         BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
 296
 297         cancel_delayed_work(&j->write_work);
 298
 299         bch2_journal_space_available(j);
 300
 301         __bch2_journal_buf_put(j, old.idx, le64_to_cpu(buf->data->seq));
 302 }
 303
 304 void bch2_journal_halt(struct journal *j)
 305 {
 306         spin_lock(&j->lock);
 307         __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL, true);
 308         if (!j->err_seq)
 309                 j->err_seq = journal_cur_seq(j);
 310         journal_wake(j);
 311         spin_unlock(&j->lock);
 312 }
 313
 314 static bool journal_entry_want_write(struct journal *j)
 315 {
 316         bool ret = !journal_entry_is_open(j) ||
 317                 journal_cur_seq(j) == journal_last_unwritten_seq(j);
 318
 319         /* Don't close it yet if we already have a write in flight: */
 320         if (ret)
 321                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 322         else if (nr_unwritten_journal_entries(j)) {
 323                 struct journal_buf *buf = journal_cur_buf(j);
 324
 325                 if (!buf->flush_time) {
 326                         buf->flush_time = local_clock() ?: 1;
 327                         buf->expires = jiffies;
 328                 }
 329         }
 330
 331         return ret;
 332 }
 333
 334 bool bch2_journal_entry_close(struct journal *j)
 335 {
 336         bool ret;
 337
 338         spin_lock(&j->lock);
 339         ret = journal_entry_want_write(j);
 340         spin_unlock(&j->lock);
 341
 342         return ret;
 343 }
 344
 345 /*
 346  * should _only_ called from journal_res_get() - when we actually want a
 347  * journal reservation - journal entry is open means journal is dirty:
 348  */
 349 static int journal_entry_open(struct journal *j)
 350 {
 351         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 352         struct journal_buf *buf = j->buf +
 353                 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
 354         union journal_res_state old, new;
 355         int u64s;
 356
 357         lockdep_assert_held(&j->lock);
 358         BUG_ON(journal_entry_is_open(j));
 359         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
 360
 361         if (j->blocked)
 362                 return JOURNAL_ERR_blocked;
 363
 364         if (j->cur_entry_error)
 365                 return j->cur_entry_error;
 366
 367         if (bch2_journal_error(j))
 368                 return JOURNAL_ERR_insufficient_devices; /* -EROFS */
 369
 370         if (!fifo_free(&j->pin))
 371                 return JOURNAL_ERR_journal_pin_full;
 372
 373         if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
 374                 return JOURNAL_ERR_max_in_flight;
 375
 376         BUG_ON(!j->cur_entry_sectors);
 377
 378         buf->expires            =
 379                 (journal_cur_seq(j) == j->flushed_seq_ondisk
 380                  ? jiffies
 381                  : j->last_flush_write) +
 382                 msecs_to_jiffies(c->opts.journal_flush_delay);
 383
 384         buf->u64s_reserved      = j->entry_u64s_reserved;
 385         buf->disk_sectors       = j->cur_entry_sectors;
 386         buf->sectors            = min(buf->disk_sectors, buf->buf_size >> 9);
 387
 388         u64s = (int) (buf->sectors << 9) / sizeof(u64) -
 389                 journal_entry_overhead(j);
 390         u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
 391
 392         if (u64s <= (ssize_t) j->early_journal_entries.nr)
 393                 return JOURNAL_ERR_journal_full;
 394
 395         if (fifo_empty(&j->pin) && j->reclaim_thread)
 396                 wake_up_process(j->reclaim_thread);
 397
 398         /*
 399          * The fifo_push() needs to happen at the same time as j->seq is
 400          * incremented for journal_last_seq() to be calculated correctly
 401          */
 402         atomic64_inc(&j->seq);
 403         journal_pin_list_init(fifo_push_ref(&j->pin), 1);
 404
 405         BUG_ON(j->pin.back - 1 != atomic64_read(&j->seq));
 406
 407         BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
 408
 409         bkey_extent_init(&buf->key);
 410         buf->noflush            = false;
 411         buf->must_flush         = false;
 412         buf->separate_flush     = false;
 413         buf->flush_time         = 0;
 414         buf->need_flush_to_write_buffer = true;
 415         buf->write_started      = false;
 416         buf->write_allocated    = false;
 417         buf->write_done         = false;
 418
 419         memset(buf->data, 0, sizeof(*buf->data));
 420         buf->data->seq  = cpu_to_le64(journal_cur_seq(j));
 421         buf->data->u64s = 0;
 422
 423         if (j->early_journal_entries.nr) {
 424                 memcpy(buf->data->_data, j->early_journal_entries.data,
 425                        j->early_journal_entries.nr * sizeof(u64));
 426                 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
 427         }
 428
 429         /*
 430          * Must be set before marking the journal entry as open:
 431          */
 432         j->cur_entry_u64s = u64s;
 433
 434         old.v = atomic64_read(&j->reservations.counter);
 435         do {
 436                 new.v = old.v;
 437
 438                 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
 439
 440                 new.idx++;
 441                 BUG_ON(journal_state_count(new, new.idx));
 442                 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
 443
 444                 journal_state_inc(&new);
 445
 446                 /* Handle any already added entries */
 447                 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
 448         } while (!atomic64_try_cmpxchg(&j->reservations.counter,
 449                                        &old.v, new.v));
 450
 451         if (nr_unwritten_journal_entries(j) == 1)
 452                 mod_delayed_work(j->wq,
 453                                  &j->write_work,
 454                                  msecs_to_jiffies(c->opts.journal_flush_delay));
 455         journal_wake(j);
 456
 457         if (j->early_journal_entries.nr)
 458                 darray_exit(&j->early_journal_entries);
 459         return 0;
 460 }
 461
 462 static bool journal_quiesced(struct journal *j)
 463 {
 464         bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
 465
 466         if (!ret)
 467                 bch2_journal_entry_close(j);
 468         return ret;
 469 }
 470
 471 static void journal_quiesce(struct journal *j)
 472 {
 473         wait_event(j->wait, journal_quiesced(j));
 474 }
 475
 476 static void journal_write_work(struct work_struct *work)
 477 {
 478         struct journal *j = container_of(work, struct journal, write_work.work);
 479
 480         spin_lock(&j->lock);
 481         if (__journal_entry_is_open(j->reservations)) {
 482                 long delta = journal_cur_buf(j)->expires - jiffies;
 483
 484                 if (delta > 0)
 485                         mod_delayed_work(j->wq, &j->write_work, delta);
 486                 else
 487                         __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 488         }
 489         spin_unlock(&j->lock);
 490 }
 491
 492 static int __journal_res_get(struct journal *j, struct journal_res *res,
 493                              unsigned flags)
 494 {
 495         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 496         struct journal_buf *buf;
 497         bool can_discard;
 498         int ret;
 499 retry:
 500         if (journal_res_get_fast(j, res, flags))
 501                 return 0;
 502
 503         if (bch2_journal_error(j))
 504                 return -BCH_ERR_erofs_journal_err;
 505
 506         if (j->blocked)
 507                 return -BCH_ERR_journal_res_get_blocked;
 508
 509         if ((flags & BCH_WATERMARK_MASK) < j->watermark) {
 510                 ret = JOURNAL_ERR_journal_full;
 511                 can_discard = j->can_discard;
 512                 goto out;
 513         }
 514
 515         if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf) && !journal_entry_is_open(j)) {
 516                 ret = JOURNAL_ERR_max_in_flight;
 517                 goto out;
 518         }
 519
 520         spin_lock(&j->lock);
 521
 522         /*
 523          * Recheck after taking the lock, so we don't race with another thread
 524          * that just did journal_entry_open() and call bch2_journal_entry_close()
 525          * unnecessarily
 526          */
 527         if (journal_res_get_fast(j, res, flags)) {
 528                 ret = 0;
 529                 goto unlock;
 530         }
 531
 532         /*
 533          * If we couldn't get a reservation because the current buf filled up,
 534          * and we had room for a bigger entry on disk, signal that we want to
 535          * realloc the journal bufs:
 536          */
 537         buf = journal_cur_buf(j);
 538         if (journal_entry_is_open(j) &&
 539             buf->buf_size >> 9 < buf->disk_sectors &&
 540             buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
 541                 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
 542
 543         __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, false);
 544         ret = journal_entry_open(j) ?: JOURNAL_ERR_retry;
 545 unlock:
 546         can_discard = j->can_discard;
 547         spin_unlock(&j->lock);
 548 out:
 549         if (ret == JOURNAL_ERR_retry)
 550                 goto retry;
 551         if (!ret)
 552                 return 0;
 553
 554         if (journal_error_check_stuck(j, ret, flags))
 555                 ret = -BCH_ERR_journal_res_get_blocked;
 556
 557         if (ret == JOURNAL_ERR_max_in_flight &&
 558             track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight], true)) {
 559
 560                 struct printbuf buf = PRINTBUF;
 561                 prt_printf(&buf, "seq %llu\n", journal_cur_seq(j));
 562                 bch2_journal_bufs_to_text(&buf, j);
 563                 trace_journal_entry_full(c, buf.buf);
 564                 printbuf_exit(&buf);
 565                 count_event(c, journal_entry_full);
 566         }
 567
 568         /*
 569          * Journal is full - can't rely on reclaim from work item due to
 570          * freezing:
 571          */
 572         if ((ret == JOURNAL_ERR_journal_full ||
 573              ret == JOURNAL_ERR_journal_pin_full) &&
 574             !(flags & JOURNAL_RES_GET_NONBLOCK)) {
 575                 if (can_discard) {
 576                         bch2_journal_do_discards(j);
 577                         goto retry;
 578                 }
 579
 580                 if (mutex_trylock(&j->reclaim_lock)) {
 581                         bch2_journal_reclaim(j);
 582                         mutex_unlock(&j->reclaim_lock);
 583                 }
 584         }
 585
 586         return ret == JOURNAL_ERR_insufficient_devices
 587                 ? -BCH_ERR_erofs_journal_err
 588                 : -BCH_ERR_journal_res_get_blocked;
 589 }
 590
 591 /*
 592  * Essentially the entry function to the journaling code. When bcachefs is doing
 593  * a btree insert, it calls this function to get the current journal write.
 594  * Journal write is the structure used set up journal writes. The calling
 595  * function will then add its keys to the structure, queuing them for the next
 596  * write.
 597  *
 598  * To ensure forward progress, the current task must not be holding any
 599  * btree node write locks.
 600  */
 601 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
 602                                   unsigned flags)
 603 {
 604         int ret;
 605
 606         if (closure_wait_event_timeout(&j->async_wait,
 607                    (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
 608                    (flags & JOURNAL_RES_GET_NONBLOCK),
 609                    HZ * 10))
 610                 return ret;
 611
 612         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 613         struct printbuf buf = PRINTBUF;
 614         bch2_journal_debug_to_text(&buf, j);
 615         bch_err(c, "Journal stuck? Waited for 10 seconds...\n%s",
 616                 buf.buf);
 617         printbuf_exit(&buf);
 618
 619         closure_wait_event(&j->async_wait,
 620                    (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
 621                    (flags & JOURNAL_RES_GET_NONBLOCK));
 622         return ret;
 623 }
 624
 625 /* journal_entry_res: */
 626
 627 void bch2_journal_entry_res_resize(struct journal *j,
 628                                    struct journal_entry_res *res,
 629                                    unsigned new_u64s)
 630 {
 631         union journal_res_state state;
 632         int d = new_u64s - res->u64s;
 633
 634         spin_lock(&j->lock);
 635
 636         j->entry_u64s_reserved += d;
 637         if (d <= 0)
 638                 goto out;
 639
 640         j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
 641         smp_mb();
 642         state = READ_ONCE(j->reservations);
 643
 644         if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
 645             state.cur_entry_offset > j->cur_entry_u64s) {
 646                 j->cur_entry_u64s += d;
 647                 /*
 648                  * Not enough room in current journal entry, have to flush it:
 649                  */
 650                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 651         } else {
 652                 journal_cur_buf(j)->u64s_reserved += d;
 653         }
 654 out:
 655         spin_unlock(&j->lock);
 656         res->u64s += d;
 657 }
 658
 659 /* journal flushing: */
 660
 661 /**
 662  * bch2_journal_flush_seq_async - wait for a journal entry to be written
 663  * @j:          journal object
 664  * @seq:        seq to flush
 665  * @parent:     closure object to wait with
 666  * Returns:     1 if @seq has already been flushed, 0 if @seq is being flushed,
 667  *              -EIO if @seq will never be flushed
 668  *
 669  * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if
 670  * necessary
 671  */
 672 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
 673                                  struct closure *parent)
 674 {
 675         struct journal_buf *buf;
 676         int ret = 0;
 677
 678         if (seq <= j->flushed_seq_ondisk)
 679                 return 1;
 680
 681         spin_lock(&j->lock);
 682
 683         if (WARN_ONCE(seq > journal_cur_seq(j),
 684                       "requested to flush journal seq %llu, but currently at %llu",
 685                       seq, journal_cur_seq(j)))
 686                 goto out;
 687
 688         /* Recheck under lock: */
 689         if (j->err_seq && seq >= j->err_seq) {
 690                 ret = -EIO;
 691                 goto out;
 692         }
 693
 694         if (seq <= j->flushed_seq_ondisk) {
 695                 ret = 1;
 696                 goto out;
 697         }
 698
 699         /* if seq was written, but not flushed - flush a newer one instead */
 700         seq = max(seq, journal_last_unwritten_seq(j));
 701
 702 recheck_need_open:
 703         if (seq > journal_cur_seq(j)) {
 704                 struct journal_res res = { 0 };
 705
 706                 if (journal_entry_is_open(j))
 707                         __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 708
 709                 spin_unlock(&j->lock);
 710
 711                 /*
 712                  * We're called from bch2_journal_flush_seq() -> wait_event();
 713                  * but this might block. We won't usually block, so we won't
 714                  * livelock:
 715                  */
 716                 sched_annotate_sleep();
 717                 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 718                 if (ret)
 719                         return ret;
 720
 721                 seq = res.seq;
 722                 buf = journal_seq_to_buf(j, seq);
 723                 buf->must_flush = true;
 724
 725                 if (!buf->flush_time) {
 726                         buf->flush_time = local_clock() ?: 1;
 727                         buf->expires = jiffies;
 728                 }
 729
 730                 if (parent && !closure_wait(&buf->wait, parent))
 731                         BUG();
 732
 733                 bch2_journal_res_put(j, &res);
 734
 735                 spin_lock(&j->lock);
 736                 goto want_write;
 737         }
 738
 739         /*
 740          * if write was kicked off without a flush, or if we promised it
 741          * wouldn't be a flush, flush the next sequence number instead
 742          */
 743         buf = journal_seq_to_buf(j, seq);
 744         if (buf->noflush) {
 745                 seq++;
 746                 goto recheck_need_open;
 747         }
 748
 749         buf->must_flush = true;
 750
 751         if (parent && !closure_wait(&buf->wait, parent))
 752                 BUG();
 753 want_write:
 754         if (seq == journal_cur_seq(j))
 755                 journal_entry_want_write(j);
 756 out:
 757         spin_unlock(&j->lock);
 758         return ret;
 759 }
 760
 761 int bch2_journal_flush_seq(struct journal *j, u64 seq, unsigned task_state)
 762 {
 763         u64 start_time = local_clock();
 764         int ret, ret2;
 765
 766         /*
 767          * Don't update time_stats when @seq is already flushed:
 768          */
 769         if (seq <= j->flushed_seq_ondisk)
 770                 return 0;
 771
 772         ret = wait_event_state(j->wait,
 773                                (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)),
 774                                task_state);
 775
 776         if (!ret)
 777                 bch2_time_stats_update(j->flush_seq_time, start_time);
 778
 779         return ret ?: ret2 < 0 ? ret2 : 0;
 780 }
 781
 782 /*
 783  * bch2_journal_flush_async - if there is an open journal entry, or a journal
 784  * still being written, write it and wait for the write to complete
 785  */
 786 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
 787 {
 788         bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
 789 }
 790
 791 int bch2_journal_flush(struct journal *j)
 792 {
 793         return bch2_journal_flush_seq(j, atomic64_read(&j->seq), TASK_UNINTERRUPTIBLE);
 794 }
 795
 796 /*
 797  * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
 798  * @seq
 799  */
 800 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
 801 {
 802         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 803         u64 unwritten_seq;
 804         bool ret = false;
 805
 806         if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
 807                 return false;
 808
 809         if (seq <= c->journal.flushed_seq_ondisk)
 810                 return false;
 811
 812         spin_lock(&j->lock);
 813         if (seq <= c->journal.flushed_seq_ondisk)
 814                 goto out;
 815
 816         for (unwritten_seq = journal_last_unwritten_seq(j);
 817              unwritten_seq < seq;
 818              unwritten_seq++) {
 819                 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
 820
 821                 /* journal flush already in flight, or flush requseted */
 822                 if (buf->must_flush)
 823                         goto out;
 824
 825                 buf->noflush = true;
 826         }
 827
 828         ret = true;
 829 out:
 830         spin_unlock(&j->lock);
 831         return ret;
 832 }
 833
 834 int bch2_journal_meta(struct journal *j)
 835 {
 836         struct journal_buf *buf;
 837         struct journal_res res;
 838         int ret;
 839
 840         memset(&res, 0, sizeof(res));
 841
 842         ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 843         if (ret)
 844                 return ret;
 845
 846         buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
 847         buf->must_flush = true;
 848
 849         if (!buf->flush_time) {
 850                 buf->flush_time = local_clock() ?: 1;
 851                 buf->expires = jiffies;
 852         }
 853
 854         bch2_journal_res_put(j, &res);
 855
 856         return bch2_journal_flush_seq(j, res.seq, TASK_UNINTERRUPTIBLE);
 857 }
 858
 859 /* block/unlock the journal: */
 860
 861 void bch2_journal_unblock(struct journal *j)
 862 {
 863         spin_lock(&j->lock);
 864         j->blocked--;
 865         spin_unlock(&j->lock);
 866
 867         journal_wake(j);
 868 }
 869
 870 void bch2_journal_block(struct journal *j)
 871 {
 872         spin_lock(&j->lock);
 873         j->blocked++;
 874         spin_unlock(&j->lock);
 875
 876         journal_quiesce(j);
 877 }
 878
 879 static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
 880 {
 881         struct journal_buf *ret = NULL;
 882
 883         /* We're inside wait_event(), but using mutex_lock(: */
 884         sched_annotate_sleep();
 885         mutex_lock(&j->buf_lock);
 886         spin_lock(&j->lock);
 887         max_seq = min(max_seq, journal_cur_seq(j));
 888
 889         for (u64 seq = journal_last_unwritten_seq(j);
 890              seq <= max_seq;
 891              seq++) {
 892                 unsigned idx = seq & JOURNAL_BUF_MASK;
 893                 struct journal_buf *buf = j->buf + idx;
 894
 895                 if (buf->need_flush_to_write_buffer) {
 896                         if (seq == journal_cur_seq(j))
 897                                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 898
 899                         union journal_res_state s;
 900                         s.v = atomic64_read_acquire(&j->reservations.counter);
 901
 902                         ret = journal_state_count(s, idx)
 903                                 ? ERR_PTR(-EAGAIN)
 904                                 : buf;
 905                         break;
 906                 }
 907         }
 908
 909         spin_unlock(&j->lock);
 910         if (IS_ERR_OR_NULL(ret))
 911                 mutex_unlock(&j->buf_lock);
 912         return ret;
 913 }
 914
 915 struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
 916 {
 917         struct journal_buf *ret;
 918
 919         wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j, max_seq)) != ERR_PTR(-EAGAIN));
 920         return ret;
 921 }
 922
 923 /* allocate journal on a device: */
 924
 925 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 926                                          bool new_fs, struct closure *cl)
 927 {
 928         struct bch_fs *c = ca->fs;
 929         struct journal_device *ja = &ca->journal;
 930         u64 *new_bucket_seq = NULL, *new_buckets = NULL;
 931         struct open_bucket **ob = NULL;
 932         long *bu = NULL;
 933         unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
 934         int ret = 0;
 935
 936         BUG_ON(nr <= ja->nr);
 937
 938         bu              = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
 939         ob              = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
 940         new_buckets     = kcalloc(nr, sizeof(u64), GFP_KERNEL);
 941         new_bucket_seq  = kcalloc(nr, sizeof(u64), GFP_KERNEL);
 942         if (!bu || !ob || !new_buckets || !new_bucket_seq) {
 943                 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
 944                 goto err_free;
 945         }
 946
 947         for (nr_got = 0; nr_got < nr_want; nr_got++) {
 948                 if (new_fs) {
 949                         bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
 950                         if (bu[nr_got] < 0) {
 951                                 ret = -BCH_ERR_ENOSPC_bucket_alloc;
 952                                 break;
 953                         }
 954                 } else {
 955                         ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal,
 956                                                        BCH_DATA_journal, cl);
 957                         ret = PTR_ERR_OR_ZERO(ob[nr_got]);
 958                         if (ret)
 959                                 break;
 960
 961                         ret = bch2_trans_run(c,
 962                                 bch2_trans_mark_metadata_bucket(trans, ca,
 963                                                 ob[nr_got]->bucket, BCH_DATA_journal,
 964                                                 ca->mi.bucket_size, BTREE_TRIGGER_transactional));
 965                         if (ret) {
 966                                 bch2_open_bucket_put(c, ob[nr_got]);
 967                                 bch_err_msg(c, ret, "marking new journal buckets");
 968                                 break;
 969                         }
 970
 971                         bu[nr_got] = ob[nr_got]->bucket;
 972                 }
 973         }
 974
 975         if (!nr_got)
 976                 goto err_free;
 977
 978         /* Don't return an error if we successfully allocated some buckets: */
 979         ret = 0;
 980
 981         if (c) {
 982                 bch2_journal_flush_all_pins(&c->journal);
 983                 bch2_journal_block(&c->journal);
 984                 mutex_lock(&c->sb_lock);
 985         }
 986
 987         memcpy(new_buckets,     ja->buckets,    ja->nr * sizeof(u64));
 988         memcpy(new_bucket_seq,  ja->bucket_seq, ja->nr * sizeof(u64));
 989
 990         BUG_ON(ja->discard_idx > ja->nr);
 991
 992         pos = ja->discard_idx ?: ja->nr;
 993
 994         memmove(new_buckets + pos + nr_got,
 995                 new_buckets + pos,
 996                 sizeof(new_buckets[0]) * (ja->nr - pos));
 997         memmove(new_bucket_seq + pos + nr_got,
 998                 new_bucket_seq + pos,
 999                 sizeof(new_bucket_seq[0]) * (ja->nr - pos));
1000
1001         for (i = 0; i < nr_got; i++) {
1002                 new_buckets[pos + i] = bu[i];
1003                 new_bucket_seq[pos + i] = 0;
1004         }
1005
1006         nr = ja->nr + nr_got;
1007
1008         ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
1009         if (ret)
1010                 goto err_unblock;
1011
1012         if (!new_fs)
1013                 bch2_write_super(c);
1014
1015         /* Commit: */
1016         if (c)
1017                 spin_lock(&c->journal.lock);
1018
1019         swap(new_buckets,       ja->buckets);
1020         swap(new_bucket_seq,    ja->bucket_seq);
1021         ja->nr = nr;
1022
1023         if (pos <= ja->discard_idx)
1024                 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
1025         if (pos <= ja->dirty_idx_ondisk)
1026                 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
1027         if (pos <= ja->dirty_idx)
1028                 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
1029         if (pos <= ja->cur_idx)
1030                 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
1031
1032         if (c)
1033                 spin_unlock(&c->journal.lock);
1034 err_unblock:
1035         if (c) {
1036                 bch2_journal_unblock(&c->journal);
1037                 mutex_unlock(&c->sb_lock);
1038         }
1039
1040         if (ret && !new_fs)
1041                 for (i = 0; i < nr_got; i++)
1042                         bch2_trans_run(c,
1043                                 bch2_trans_mark_metadata_bucket(trans, ca,
1044                                                 bu[i], BCH_DATA_free, 0,
1045                                                 BTREE_TRIGGER_transactional));
1046 err_free:
1047         if (!new_fs)
1048                 for (i = 0; i < nr_got; i++)
1049                         bch2_open_bucket_put(c, ob[i]);
1050
1051         kfree(new_bucket_seq);
1052         kfree(new_buckets);
1053         kfree(ob);
1054         kfree(bu);
1055         return ret;
1056 }
1057
1058 /*
1059  * Allocate more journal space at runtime - not currently making use if it, but
1060  * the code works:
1061  */
1062 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
1063                                 unsigned nr)
1064 {
1065         struct journal_device *ja = &ca->journal;
1066         struct closure cl;
1067         int ret = 0;
1068
1069         closure_init_stack(&cl);
1070
1071         down_write(&c->state_lock);
1072
1073         /* don't handle reducing nr of buckets yet: */
1074         if (nr < ja->nr)
1075                 goto unlock;
1076
1077         while (ja->nr < nr) {
1078                 struct disk_reservation disk_res = { 0, 0, 0 };
1079
1080                 /*
1081                  * note: journal buckets aren't really counted as _sectors_ used yet, so
1082                  * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
1083                  * when space used goes up without a reservation - but we do need the
1084                  * reservation to ensure we'll actually be able to allocate:
1085                  *
1086                  * XXX: that's not right, disk reservations only ensure a
1087                  * filesystem-wide allocation will succeed, this is a device
1088                  * specific allocation - we can hang here:
1089                  */
1090
1091                 ret = bch2_disk_reservation_get(c, &disk_res,
1092                                                 bucket_to_sector(ca, nr - ja->nr), 1, 0);
1093                 if (ret)
1094                         break;
1095
1096                 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
1097
1098                 bch2_disk_reservation_put(c, &disk_res);
1099
1100                 closure_sync(&cl);
1101
1102                 if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
1103                         break;
1104         }
1105
1106         bch_err_fn(c, ret);
1107 unlock:
1108         up_write(&c->state_lock);
1109         return ret;
1110 }
1111
1112 int bch2_dev_journal_alloc(struct bch_dev *ca, bool new_fs)
1113 {
1114         unsigned nr;
1115         int ret;
1116
1117         if (dynamic_fault("bcachefs:add:journal_alloc")) {
1118                 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
1119                 goto err;
1120         }
1121
1122         /* 1/128th of the device by default: */
1123         nr = ca->mi.nbuckets >> 7;
1124
1125         /*
1126          * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
1127          * is smaller:
1128          */
1129         nr = clamp_t(unsigned, nr,
1130                      BCH_JOURNAL_BUCKETS_MIN,
1131                      min(1 << 13,
1132                          (1 << 24) / ca->mi.bucket_size));
1133
1134         ret = __bch2_set_nr_journal_buckets(ca, nr, new_fs, NULL);
1135 err:
1136         bch_err_fn(ca, ret);
1137         return ret;
1138 }
1139
1140 int bch2_fs_journal_alloc(struct bch_fs *c)
1141 {
1142         for_each_online_member(c, ca) {
1143                 if (ca->journal.nr)
1144                         continue;
1145
1146                 int ret = bch2_dev_journal_alloc(ca, true);
1147                 if (ret) {
1148                         percpu_ref_put(&ca->io_ref);
1149                         return ret;
1150                 }
1151         }
1152
1153         return 0;
1154 }
1155
1156 /* startup/shutdown: */
1157
1158 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1159 {
1160         bool ret = false;
1161         u64 seq;
1162
1163         spin_lock(&j->lock);
1164         for (seq = journal_last_unwritten_seq(j);
1165              seq <= journal_cur_seq(j) && !ret;
1166              seq++) {
1167                 struct journal_buf *buf = journal_seq_to_buf(j, seq);
1168
1169                 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1170                         ret = true;
1171         }
1172         spin_unlock(&j->lock);
1173
1174         return ret;
1175 }
1176
1177 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1178 {
1179         wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1180 }
1181
1182 void bch2_fs_journal_stop(struct journal *j)
1183 {
1184         if (!test_bit(JOURNAL_running, &j->flags))
1185                 return;
1186
1187         bch2_journal_reclaim_stop(j);
1188         bch2_journal_flush_all_pins(j);
1189
1190         wait_event(j->wait, bch2_journal_entry_close(j));
1191
1192         /*
1193          * Always write a new journal entry, to make sure the clock hands are up
1194          * to date (and match the superblock)
1195          */
1196         bch2_journal_meta(j);
1197
1198         journal_quiesce(j);
1199         cancel_delayed_work_sync(&j->write_work);
1200
1201         WARN(!bch2_journal_error(j) &&
1202              test_bit(JOURNAL_replay_done, &j->flags) &&
1203              j->last_empty_seq != journal_cur_seq(j),
1204              "journal shutdown error: cur seq %llu but last empty seq %llu",
1205              journal_cur_seq(j), j->last_empty_seq);
1206
1207         if (!bch2_journal_error(j))
1208                 clear_bit(JOURNAL_running, &j->flags);
1209 }
1210
1211 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1212 {
1213         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1214         struct journal_entry_pin_list *p;
1215         struct journal_replay *i, **_i;
1216         struct genradix_iter iter;
1217         bool had_entries = false;
1218         u64 last_seq = cur_seq, nr, seq;
1219
1220         genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1221                 i = *_i;
1222
1223                 if (journal_replay_ignore(i))
1224                         continue;
1225
1226                 last_seq = le64_to_cpu(i->j.last_seq);
1227                 break;
1228         }
1229
1230         nr = cur_seq - last_seq;
1231
1232         if (nr + 1 > j->pin.size) {
1233                 free_fifo(&j->pin);
1234                 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1235                 if (!j->pin.data) {
1236                         bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1237                         return -BCH_ERR_ENOMEM_journal_pin_fifo;
1238                 }
1239         }
1240
1241         j->replay_journal_seq   = last_seq;
1242         j->replay_journal_seq_end = cur_seq;
1243         j->last_seq_ondisk      = last_seq;
1244         j->flushed_seq_ondisk   = cur_seq - 1;
1245         j->seq_ondisk           = cur_seq - 1;
1246         j->pin.front            = last_seq;
1247         j->pin.back             = cur_seq;
1248         atomic64_set(&j->seq, cur_seq - 1);
1249
1250         fifo_for_each_entry_ptr(p, &j->pin, seq)
1251                 journal_pin_list_init(p, 1);
1252
1253         genradix_for_each(&c->journal_entries, iter, _i) {
1254                 i = *_i;
1255
1256                 if (journal_replay_ignore(i))
1257                         continue;
1258
1259                 seq = le64_to_cpu(i->j.seq);
1260                 BUG_ON(seq >= cur_seq);
1261
1262                 if (seq < last_seq)
1263                         continue;
1264
1265                 if (journal_entry_empty(&i->j))
1266                         j->last_empty_seq = le64_to_cpu(i->j.seq);
1267
1268                 p = journal_seq_pin(j, seq);
1269
1270                 p->devs.nr = 0;
1271                 darray_for_each(i->ptrs, ptr)
1272                         bch2_dev_list_add_dev(&p->devs, ptr->dev);
1273
1274                 had_entries = true;
1275         }
1276
1277         if (!had_entries)
1278                 j->last_empty_seq = cur_seq - 1; /* to match j->seq */
1279
1280         spin_lock(&j->lock);
1281
1282         set_bit(JOURNAL_running, &j->flags);
1283         j->last_flush_write = jiffies;
1284
1285         j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1286         j->reservations.unwritten_idx++;
1287
1288         c->last_bucket_seq_cleanup = journal_cur_seq(j);
1289
1290         bch2_journal_space_available(j);
1291         spin_unlock(&j->lock);
1292
1293         return bch2_journal_reclaim_start(j);
1294 }
1295
1296 /* init/exit: */
1297
1298 void bch2_dev_journal_exit(struct bch_dev *ca)
1299 {
1300         struct journal_device *ja = &ca->journal;
1301
1302         for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
1303                 kfree(ja->bio[i]);
1304                 ja->bio[i] = NULL;
1305         }
1306
1307         kfree(ja->buckets);
1308         kfree(ja->bucket_seq);
1309         ja->buckets     = NULL;
1310         ja->bucket_seq  = NULL;
1311 }
1312
1313 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1314 {
1315         struct journal_device *ja = &ca->journal;
1316         struct bch_sb_field_journal *journal_buckets =
1317                 bch2_sb_field_get(sb, journal);
1318         struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1319                 bch2_sb_field_get(sb, journal_v2);
1320
1321         ja->nr = 0;
1322
1323         if (journal_buckets_v2) {
1324                 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1325
1326                 for (unsigned i = 0; i < nr; i++)
1327                         ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1328         } else if (journal_buckets) {
1329                 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1330         }
1331
1332         ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1333         if (!ja->bucket_seq)
1334                 return -BCH_ERR_ENOMEM_dev_journal_init;
1335
1336         unsigned nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1337
1338         for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
1339                 ja->bio[i] = kmalloc(struct_size(ja->bio[i], bio.bi_inline_vecs,
1340                                      nr_bvecs), GFP_KERNEL);
1341                 if (!ja->bio[i])
1342                         return -BCH_ERR_ENOMEM_dev_journal_init;
1343
1344                 ja->bio[i]->ca = ca;
1345                 ja->bio[i]->buf_idx = i;
1346                 bio_init(&ja->bio[i]->bio, NULL, ja->bio[i]->bio.bi_inline_vecs, nr_bvecs, 0);
1347         }
1348
1349         ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1350         if (!ja->buckets)
1351                 return -BCH_ERR_ENOMEM_dev_journal_init;
1352
1353         if (journal_buckets_v2) {
1354                 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1355                 unsigned dst = 0;
1356
1357                 for (unsigned i = 0; i < nr; i++)
1358                         for (unsigned j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1359                                 ja->buckets[dst++] =
1360                                         le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1361         } else if (journal_buckets) {
1362                 for (unsigned i = 0; i < ja->nr; i++)
1363                         ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1364         }
1365
1366         return 0;
1367 }
1368
1369 void bch2_fs_journal_exit(struct journal *j)
1370 {
1371         if (j->wq)
1372                 destroy_workqueue(j->wq);
1373
1374         darray_exit(&j->early_journal_entries);
1375
1376         for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++)
1377                 kvfree(j->buf[i].data);
1378         free_fifo(&j->pin);
1379 }
1380
1381 int bch2_fs_journal_init(struct journal *j)
1382 {
1383         static struct lock_class_key res_key;
1384
1385         mutex_init(&j->buf_lock);
1386         spin_lock_init(&j->lock);
1387         spin_lock_init(&j->err_lock);
1388         init_waitqueue_head(&j->wait);
1389         INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1390         init_waitqueue_head(&j->reclaim_wait);
1391         init_waitqueue_head(&j->pin_flush_wait);
1392         mutex_init(&j->reclaim_lock);
1393         mutex_init(&j->discard_lock);
1394
1395         lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1396
1397         atomic64_set(&j->reservations.counter,
1398                 ((union journal_res_state)
1399                  { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1400
1401         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)))
1402                 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1403
1404         for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++) {
1405                 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1406                 j->buf[i].data = kvmalloc(j->buf[i].buf_size, GFP_KERNEL);
1407                 if (!j->buf[i].data)
1408                         return -BCH_ERR_ENOMEM_journal_buf;
1409                 j->buf[i].idx = i;
1410         }
1411
1412         j->pin.front = j->pin.back = 1;
1413
1414         j->wq = alloc_workqueue("bcachefs_journal",
1415                                 WQ_HIGHPRI|WQ_FREEZABLE|WQ_UNBOUND|WQ_MEM_RECLAIM, 512);
1416         if (!j->wq)
1417                 return -BCH_ERR_ENOMEM_fs_other_alloc;
1418         return 0;
1419 }
1420
1421 /* debug: */
1422
1423 static const char * const bch2_journal_flags_strs[] = {
1424 #define x(n)    #n,
1425         JOURNAL_FLAGS()
1426 #undef x
1427         NULL
1428 };
1429
1430 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1431 {
1432         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1433         union journal_res_state s;
1434         unsigned long now = jiffies;
1435         u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes;
1436
1437         printbuf_tabstops_reset(out);
1438         printbuf_tabstop_push(out, 28);
1439         out->atomic++;
1440
1441         rcu_read_lock();
1442         s = READ_ONCE(j->reservations);
1443
1444         prt_printf(out, "flags:\t");
1445         prt_bitflags(out, bch2_journal_flags_strs, j->flags);
1446         prt_newline(out);
1447         prt_printf(out, "dirty journal entries:\t%llu/%llu\n",  fifo_used(&j->pin), j->pin.size);
1448         prt_printf(out, "seq:\t%llu\n",                         journal_cur_seq(j));
1449         prt_printf(out, "seq_ondisk:\t%llu\n",                  j->seq_ondisk);
1450         prt_printf(out, "last_seq:\t%llu\n",                    journal_last_seq(j));
1451         prt_printf(out, "last_seq_ondisk:\t%llu\n",             j->last_seq_ondisk);
1452         prt_printf(out, "flushed_seq_ondisk:\t%llu\n",          j->flushed_seq_ondisk);
1453         prt_printf(out, "watermark:\t%s\n",                     bch2_watermarks[j->watermark]);
1454         prt_printf(out, "each entry reserved:\t%u\n",           j->entry_u64s_reserved);
1455         prt_printf(out, "nr flush writes:\t%llu\n",             j->nr_flush_writes);
1456         prt_printf(out, "nr noflush writes:\t%llu\n",           j->nr_noflush_writes);
1457         prt_printf(out, "average write size:\t");
1458         prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0);
1459         prt_newline(out);
1460         prt_printf(out, "nr direct reclaim:\t%llu\n",           j->nr_direct_reclaim);
1461         prt_printf(out, "nr background reclaim:\t%llu\n",       j->nr_background_reclaim);
1462         prt_printf(out, "reclaim kicked:\t%u\n",                j->reclaim_kicked);
1463         prt_printf(out, "reclaim runs in:\t%u ms\n",            time_after(j->next_reclaim, now)
1464                ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1465         prt_printf(out, "blocked:\t%u\n",                       j->blocked);
1466         prt_printf(out, "current entry sectors:\t%u\n",         j->cur_entry_sectors);
1467         prt_printf(out, "current entry error:\t%s\n",           bch2_journal_errors[j->cur_entry_error]);
1468         prt_printf(out, "current entry:\t");
1469
1470         switch (s.cur_entry_offset) {
1471         case JOURNAL_ENTRY_ERROR_VAL:
1472                 prt_printf(out, "error\n");
1473                 break;
1474         case JOURNAL_ENTRY_CLOSED_VAL:
1475                 prt_printf(out, "closed\n");
1476                 break;
1477         default:
1478                 prt_printf(out, "%u/%u\n", s.cur_entry_offset, j->cur_entry_u64s);
1479                 break;
1480         }
1481
1482         prt_printf(out, "unwritten entries:\n");
1483         bch2_journal_bufs_to_text(out, j);
1484
1485         prt_printf(out, "space:\n");
1486         printbuf_indent_add(out, 2);
1487         prt_printf(out, "discarded\t%u:%u\n",
1488                j->space[journal_space_discarded].next_entry,
1489                j->space[journal_space_discarded].total);
1490         prt_printf(out, "clean ondisk\t%u:%u\n",
1491                j->space[journal_space_clean_ondisk].next_entry,
1492                j->space[journal_space_clean_ondisk].total);
1493         prt_printf(out, "clean\t%u:%u\n",
1494                j->space[journal_space_clean].next_entry,
1495                j->space[journal_space_clean].total);
1496         prt_printf(out, "total\t%u:%u\n",
1497                j->space[journal_space_total].next_entry,
1498                j->space[journal_space_total].total);
1499         printbuf_indent_sub(out, 2);
1500
1501         for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
1502                 struct journal_device *ja = &ca->journal;
1503
1504                 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1505                         continue;
1506
1507                 if (!ja->nr)
1508                         continue;
1509
1510                 prt_printf(out, "dev %u:\n",                    ca->dev_idx);
1511                 printbuf_indent_add(out, 2);
1512                 prt_printf(out, "nr\t%u\n",                     ja->nr);
1513                 prt_printf(out, "bucket size\t%u\n",            ca->mi.bucket_size);
1514                 prt_printf(out, "available\t%u:%u\n",           bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1515                 prt_printf(out, "discard_idx\t%u\n",            ja->discard_idx);
1516                 prt_printf(out, "dirty_ondisk\t%u (seq %llu)\n",ja->dirty_idx_ondisk,   ja->bucket_seq[ja->dirty_idx_ondisk]);
1517                 prt_printf(out, "dirty_idx\t%u (seq %llu)\n",   ja->dirty_idx,          ja->bucket_seq[ja->dirty_idx]);
1518                 prt_printf(out, "cur_idx\t%u (seq %llu)\n",     ja->cur_idx,            ja->bucket_seq[ja->cur_idx]);
1519                 printbuf_indent_sub(out, 2);
1520         }
1521
1522         rcu_read_unlock();
1523
1524         --out->atomic;
1525 }
1526
1527 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1528 {
1529         spin_lock(&j->lock);
1530         __bch2_journal_debug_to_text(out, j);
1531         spin_unlock(&j->lock);
1532 }
1533
1534 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1535 {
1536         struct journal_entry_pin_list *pin_list;
1537         struct journal_entry_pin *pin;
1538
1539         spin_lock(&j->lock);
1540         if (!test_bit(JOURNAL_running, &j->flags)) {
1541                 spin_unlock(&j->lock);
1542                 return true;
1543         }
1544
1545         *seq = max(*seq, j->pin.front);
1546
1547         if (*seq >= j->pin.back) {
1548                 spin_unlock(&j->lock);
1549                 return true;
1550         }
1551
1552         out->atomic++;
1553
1554         pin_list = journal_seq_pin(j, *seq);
1555
1556         prt_printf(out, "%llu: count %u\n", *seq, atomic_read(&pin_list->count));
1557         printbuf_indent_add(out, 2);
1558
1559         for (unsigned i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1560                 list_for_each_entry(pin, &pin_list->list[i], list)
1561                         prt_printf(out, "\t%px %ps\n", pin, pin->flush);
1562
1563         if (!list_empty(&pin_list->flushed))
1564                 prt_printf(out, "flushed:\n");
1565
1566         list_for_each_entry(pin, &pin_list->flushed, list)
1567                 prt_printf(out, "\t%px %ps\n", pin, pin->flush);
1568
1569         printbuf_indent_sub(out, 2);
1570
1571         --out->atomic;
1572         spin_unlock(&j->lock);
1573
1574         return false;
1575 }
1576
1577 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1578 {
1579         u64 seq = 0;
1580
1581         while (!bch2_journal_seq_pins_to_text(out, j, &seq))
1582                 seq++;
1583 }