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 "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_str(out, "seq:");
  57         prt_tab(out);
  58         prt_printf(out, "%llu", seq);
  59         prt_newline(out);
  60         printbuf_indent_add(out, 2);
  61
  62         prt_str(out, "refcount:");
  63         prt_tab(out);
  64         prt_printf(out, "%u", journal_state_count(s, i));
  65         prt_newline(out);
  66
  67         prt_str(out, "size:");
  68         prt_tab(out);
  69         prt_human_readable_u64(out, vstruct_bytes(buf->data));
  70         prt_newline(out);
  71
  72         prt_str(out, "expires:");
  73         prt_tab(out);
  74         prt_printf(out, "%li jiffies", buf->expires - jiffies);
  75         prt_newline(out);
  76
  77         prt_str(out, "flags:");
  78         prt_tab(out);
  79         if (buf->noflush)
  80                 prt_str(out, "noflush ");
  81         if (buf->must_flush)
  82                 prt_str(out, "must_flush ");
  83         if (buf->separate_flush)
  84                 prt_str(out, "separate_flush ");
  85         if (buf->need_flush_to_write_buffer)
  86                 prt_str(out, "need_flush_to_write_buffer ");
  87         if (buf->need_flush_to_write_buffer)
  88                 prt_str(out, "need_flush_to_write_buffer ");
  89         if (buf->write_done)
  90                 prt_str(out, "write done ");
  91         if (buf->write_started)
  92                 prt_str(out, "write started ");
  93         if (buf->write_allocated)
  94                 prt_str(out, "write allocated ");
  95         if (buf->write_done)
  96                 prt_str(out, "write done");
  97         prt_newline(out);
  98
  99         printbuf_indent_sub(out, 2);
 100 }
 101
 102 static void bch2_journal_bufs_to_text(struct printbuf *out, struct journal *j)
 103 {
 104         if (!out->nr_tabstops)
 105                 printbuf_tabstop_push(out, 24);
 106
 107         for (u64 seq = journal_last_unwritten_seq(j);
 108              seq <= journal_cur_seq(j);
 109              seq++)
 110                 bch2_journal_buf_to_text(out, j, seq);
 111         prt_printf(out, "last buf %s\n", journal_entry_is_open(j) ? "open" : "closed");
 112 }
 113
 114 static inline struct journal_buf *
 115 journal_seq_to_buf(struct journal *j, u64 seq)
 116 {
 117         struct journal_buf *buf = NULL;
 118
 119         EBUG_ON(seq > journal_cur_seq(j));
 120
 121         if (journal_seq_unwritten(j, seq)) {
 122                 buf = j->buf + (seq & JOURNAL_BUF_MASK);
 123                 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
 124         }
 125         return buf;
 126 }
 127
 128 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
 129 {
 130         unsigned i;
 131
 132         for (i = 0; i < ARRAY_SIZE(p->list); i++)
 133                 INIT_LIST_HEAD(&p->list[i]);
 134         INIT_LIST_HEAD(&p->flushed);
 135         atomic_set(&p->count, count);
 136         p->devs.nr = 0;
 137 }
 138
 139 /*
 140  * Detect stuck journal conditions and trigger shutdown. Technically the journal
 141  * can end up stuck for a variety of reasons, such as a blocked I/O, journal
 142  * reservation lockup, etc. Since this is a fatal error with potentially
 143  * unpredictable characteristics, we want to be fairly conservative before we
 144  * decide to shut things down.
 145  *
 146  * Consider the journal stuck when it appears full with no ability to commit
 147  * btree transactions, to discard journal buckets, nor acquire priority
 148  * (reserved watermark) reservation.
 149  */
 150 static inline bool
 151 journal_error_check_stuck(struct journal *j, int error, unsigned flags)
 152 {
 153         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 154         bool stuck = false;
 155         struct printbuf buf = PRINTBUF;
 156
 157         if (!(error == JOURNAL_ERR_journal_full ||
 158               error == JOURNAL_ERR_journal_pin_full) ||
 159             nr_unwritten_journal_entries(j) ||
 160             (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim)
 161                 return stuck;
 162
 163         spin_lock(&j->lock);
 164
 165         if (j->can_discard) {
 166                 spin_unlock(&j->lock);
 167                 return stuck;
 168         }
 169
 170         stuck = true;
 171
 172         /*
 173          * The journal shutdown path will set ->err_seq, but do it here first to
 174          * serialize against concurrent failures and avoid duplicate error
 175          * reports.
 176          */
 177         if (j->err_seq) {
 178                 spin_unlock(&j->lock);
 179                 return stuck;
 180         }
 181         j->err_seq = journal_cur_seq(j);
 182         spin_unlock(&j->lock);
 183
 184         bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)",
 185                 bch2_journal_errors[error]);
 186         bch2_journal_debug_to_text(&buf, j);
 187         bch_err(c, "%s", buf.buf);
 188
 189         printbuf_reset(&buf);
 190         bch2_journal_pins_to_text(&buf, j);
 191         bch_err(c, "Journal pins:\n%s", buf.buf);
 192         printbuf_exit(&buf);
 193
 194         bch2_fatal_error(c);
 195         dump_stack();
 196
 197         return stuck;
 198 }
 199
 200 void bch2_journal_do_writes(struct journal *j)
 201 {
 202         for (u64 seq = journal_last_unwritten_seq(j);
 203              seq <= journal_cur_seq(j);
 204              seq++) {
 205                 unsigned idx = seq & JOURNAL_BUF_MASK;
 206                 struct journal_buf *w = j->buf + idx;
 207
 208                 if (w->write_started && !w->write_allocated)
 209                         break;
 210                 if (w->write_started)
 211                         continue;
 212
 213                 if (!journal_state_count(j->reservations, idx)) {
 214                         w->write_started = true;
 215                         closure_call(&w->io, bch2_journal_write, j->wq, NULL);
 216                 }
 217
 218                 break;
 219         }
 220 }
 221
 222 /*
 223  * Final processing when the last reference of a journal buffer has been
 224  * dropped. Drop the pin list reference acquired at journal entry open and write
 225  * the buffer, if requested.
 226  */
 227 void bch2_journal_buf_put_final(struct journal *j, u64 seq)
 228 {
 229         lockdep_assert_held(&j->lock);
 230
 231         if (__bch2_journal_pin_put(j, seq))
 232                 bch2_journal_reclaim_fast(j);
 233         bch2_journal_do_writes(j);
 234 }
 235
 236 /*
 237  * Returns true if journal entry is now closed:
 238  *
 239  * We don't close a journal_buf until the next journal_buf is finished writing,
 240  * and can be opened again - this also initializes the next journal_buf:
 241  */
 242 static void __journal_entry_close(struct journal *j, unsigned closed_val, bool trace)
 243 {
 244         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 245         struct journal_buf *buf = journal_cur_buf(j);
 246         union journal_res_state old, new;
 247         u64 v = atomic64_read(&j->reservations.counter);
 248         unsigned sectors;
 249
 250         BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
 251                closed_val != JOURNAL_ENTRY_ERROR_VAL);
 252
 253         lockdep_assert_held(&j->lock);
 254
 255         do {
 256                 old.v = new.v = v;
 257                 new.cur_entry_offset = closed_val;
 258
 259                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
 260                     old.cur_entry_offset == new.cur_entry_offset)
 261                         return;
 262         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 263                                        old.v, new.v)) != old.v);
 264
 265         if (!__journal_entry_is_open(old))
 266                 return;
 267
 268         /* Close out old buffer: */
 269         buf->data->u64s         = cpu_to_le32(old.cur_entry_offset);
 270
 271         if (trace_journal_entry_close_enabled() && trace) {
 272                 struct printbuf pbuf = PRINTBUF;
 273                 pbuf.atomic++;
 274
 275                 prt_str(&pbuf, "entry size: ");
 276                 prt_human_readable_u64(&pbuf, vstruct_bytes(buf->data));
 277                 prt_newline(&pbuf);
 278                 bch2_prt_task_backtrace(&pbuf, current, 1, GFP_NOWAIT);
 279                 trace_journal_entry_close(c, pbuf.buf);
 280                 printbuf_exit(&pbuf);
 281         }
 282
 283         sectors = vstruct_blocks_plus(buf->data, c->block_bits,
 284                                       buf->u64s_reserved) << c->block_bits;
 285         BUG_ON(sectors > buf->sectors);
 286         buf->sectors = sectors;
 287
 288         /*
 289          * We have to set last_seq here, _before_ opening a new journal entry:
 290          *
 291          * A threads may replace an old pin with a new pin on their current
 292          * journal reservation - the expectation being that the journal will
 293          * contain either what the old pin protected or what the new pin
 294          * protects.
 295          *
 296          * After the old pin is dropped journal_last_seq() won't include the old
 297          * pin, so we can only write the updated last_seq on the entry that
 298          * contains whatever the new pin protects.
 299          *
 300          * Restated, we can _not_ update last_seq for a given entry if there
 301          * could be a newer entry open with reservations/pins that have been
 302          * taken against it.
 303          *
 304          * Hence, we want update/set last_seq on the current journal entry right
 305          * before we open a new one:
 306          */
 307         buf->last_seq           = journal_last_seq(j);
 308         buf->data->last_seq     = cpu_to_le64(buf->last_seq);
 309         BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
 310
 311         cancel_delayed_work(&j->write_work);
 312
 313         bch2_journal_space_available(j);
 314
 315         __bch2_journal_buf_put(j, old.idx, le64_to_cpu(buf->data->seq));
 316 }
 317
 318 void bch2_journal_halt(struct journal *j)
 319 {
 320         spin_lock(&j->lock);
 321         __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL, true);
 322         if (!j->err_seq)
 323                 j->err_seq = journal_cur_seq(j);
 324         journal_wake(j);
 325         spin_unlock(&j->lock);
 326 }
 327
 328 static bool journal_entry_want_write(struct journal *j)
 329 {
 330         bool ret = !journal_entry_is_open(j) ||
 331                 journal_cur_seq(j) == journal_last_unwritten_seq(j);
 332
 333         /* Don't close it yet if we already have a write in flight: */
 334         if (ret)
 335                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 336         else if (nr_unwritten_journal_entries(j)) {
 337                 struct journal_buf *buf = journal_cur_buf(j);
 338
 339                 if (!buf->flush_time) {
 340                         buf->flush_time = local_clock() ?: 1;
 341                         buf->expires = jiffies;
 342                 }
 343         }
 344
 345         return ret;
 346 }
 347
 348 bool bch2_journal_entry_close(struct journal *j)
 349 {
 350         bool ret;
 351
 352         spin_lock(&j->lock);
 353         ret = journal_entry_want_write(j);
 354         spin_unlock(&j->lock);
 355
 356         return ret;
 357 }
 358
 359 /*
 360  * should _only_ called from journal_res_get() - when we actually want a
 361  * journal reservation - journal entry is open means journal is dirty:
 362  */
 363 static int journal_entry_open(struct journal *j)
 364 {
 365         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 366         struct journal_buf *buf = j->buf +
 367                 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
 368         union journal_res_state old, new;
 369         int u64s;
 370         u64 v;
 371
 372         lockdep_assert_held(&j->lock);
 373         BUG_ON(journal_entry_is_open(j));
 374         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
 375
 376         if (j->blocked)
 377                 return JOURNAL_ERR_blocked;
 378
 379         if (j->cur_entry_error)
 380                 return j->cur_entry_error;
 381
 382         if (bch2_journal_error(j))
 383                 return JOURNAL_ERR_insufficient_devices; /* -EROFS */
 384
 385         if (!fifo_free(&j->pin))
 386                 return JOURNAL_ERR_journal_pin_full;
 387
 388         if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
 389                 return JOURNAL_ERR_max_in_flight;
 390
 391         BUG_ON(!j->cur_entry_sectors);
 392
 393         buf->expires            =
 394                 (journal_cur_seq(j) == j->flushed_seq_ondisk
 395                  ? jiffies
 396                  : j->last_flush_write) +
 397                 msecs_to_jiffies(c->opts.journal_flush_delay);
 398
 399         buf->u64s_reserved      = j->entry_u64s_reserved;
 400         buf->disk_sectors       = j->cur_entry_sectors;
 401         buf->sectors            = min(buf->disk_sectors, buf->buf_size >> 9);
 402
 403         u64s = (int) (buf->sectors << 9) / sizeof(u64) -
 404                 journal_entry_overhead(j);
 405         u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
 406
 407         if (u64s <= (ssize_t) j->early_journal_entries.nr)
 408                 return JOURNAL_ERR_journal_full;
 409
 410         if (fifo_empty(&j->pin) && j->reclaim_thread)
 411                 wake_up_process(j->reclaim_thread);
 412
 413         /*
 414          * The fifo_push() needs to happen at the same time as j->seq is
 415          * incremented for journal_last_seq() to be calculated correctly
 416          */
 417         atomic64_inc(&j->seq);
 418         journal_pin_list_init(fifo_push_ref(&j->pin), 1);
 419
 420         BUG_ON(j->pin.back - 1 != atomic64_read(&j->seq));
 421
 422         BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
 423
 424         bkey_extent_init(&buf->key);
 425         buf->noflush            = false;
 426         buf->must_flush         = false;
 427         buf->separate_flush     = false;
 428         buf->flush_time         = 0;
 429         buf->need_flush_to_write_buffer = true;
 430         buf->write_started      = false;
 431         buf->write_allocated    = false;
 432         buf->write_done         = false;
 433
 434         memset(buf->data, 0, sizeof(*buf->data));
 435         buf->data->seq  = cpu_to_le64(journal_cur_seq(j));
 436         buf->data->u64s = 0;
 437
 438         if (j->early_journal_entries.nr) {
 439                 memcpy(buf->data->_data, j->early_journal_entries.data,
 440                        j->early_journal_entries.nr * sizeof(u64));
 441                 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
 442         }
 443
 444         /*
 445          * Must be set before marking the journal entry as open:
 446          */
 447         j->cur_entry_u64s = u64s;
 448
 449         v = atomic64_read(&j->reservations.counter);
 450         do {
 451                 old.v = new.v = v;
 452
 453                 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
 454
 455                 new.idx++;
 456                 BUG_ON(journal_state_count(new, new.idx));
 457                 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
 458
 459                 journal_state_inc(&new);
 460
 461                 /* Handle any already added entries */
 462                 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
 463         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 464                                        old.v, new.v)) != old.v);
 465
 466         if (nr_unwritten_journal_entries(j) == 1)
 467                 mod_delayed_work(j->wq,
 468                                  &j->write_work,
 469                                  msecs_to_jiffies(c->opts.journal_flush_delay));
 470         journal_wake(j);
 471
 472         if (j->early_journal_entries.nr)
 473                 darray_exit(&j->early_journal_entries);
 474         return 0;
 475 }
 476
 477 static bool journal_quiesced(struct journal *j)
 478 {
 479         bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
 480
 481         if (!ret)
 482                 bch2_journal_entry_close(j);
 483         return ret;
 484 }
 485
 486 static void journal_quiesce(struct journal *j)
 487 {
 488         wait_event(j->wait, journal_quiesced(j));
 489 }
 490
 491 static void journal_write_work(struct work_struct *work)
 492 {
 493         struct journal *j = container_of(work, struct journal, write_work.work);
 494
 495         spin_lock(&j->lock);
 496         if (__journal_entry_is_open(j->reservations)) {
 497                 long delta = journal_cur_buf(j)->expires - jiffies;
 498
 499                 if (delta > 0)
 500                         mod_delayed_work(j->wq, &j->write_work, delta);
 501                 else
 502                         __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 503         }
 504         spin_unlock(&j->lock);
 505 }
 506
 507 static int __journal_res_get(struct journal *j, struct journal_res *res,
 508                              unsigned flags)
 509 {
 510         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 511         struct journal_buf *buf;
 512         bool can_discard;
 513         int ret;
 514 retry:
 515         if (journal_res_get_fast(j, res, flags))
 516                 return 0;
 517
 518         if ((flags & BCH_WATERMARK_MASK) < j->watermark) {
 519                 ret = JOURNAL_ERR_journal_full;
 520                 can_discard = j->can_discard;
 521                 goto out;
 522         }
 523
 524         if (j->blocked)
 525                 return -BCH_ERR_journal_res_get_blocked;
 526
 527         if (bch2_journal_error(j))
 528                 return -BCH_ERR_erofs_journal_err;
 529
 530         if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf) && !journal_entry_is_open(j)) {
 531                 ret = JOURNAL_ERR_max_in_flight;
 532                 goto out;
 533         }
 534
 535         spin_lock(&j->lock);
 536
 537         /*
 538          * Recheck after taking the lock, so we don't race with another thread
 539          * that just did journal_entry_open() and call bch2_journal_entry_close()
 540          * unnecessarily
 541          */
 542         if (journal_res_get_fast(j, res, flags)) {
 543                 ret = 0;
 544                 goto unlock;
 545         }
 546
 547         /*
 548          * If we couldn't get a reservation because the current buf filled up,
 549          * and we had room for a bigger entry on disk, signal that we want to
 550          * realloc the journal bufs:
 551          */
 552         buf = journal_cur_buf(j);
 553         if (journal_entry_is_open(j) &&
 554             buf->buf_size >> 9 < buf->disk_sectors &&
 555             buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
 556                 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
 557
 558         __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, false);
 559         ret = journal_entry_open(j) ?: JOURNAL_ERR_retry;
 560 unlock:
 561         can_discard = j->can_discard;
 562         spin_unlock(&j->lock);
 563 out:
 564         if (ret == JOURNAL_ERR_retry)
 565                 goto retry;
 566         if (!ret)
 567                 return 0;
 568
 569         if (journal_error_check_stuck(j, ret, flags))
 570                 ret = -BCH_ERR_journal_res_get_blocked;
 571
 572         if (ret == JOURNAL_ERR_max_in_flight &&
 573             track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight], true)) {
 574
 575                 struct printbuf buf = PRINTBUF;
 576                 prt_printf(&buf, "seq %llu\n", journal_cur_seq(j));
 577                 bch2_journal_bufs_to_text(&buf, j);
 578                 trace_journal_entry_full(c, buf.buf);
 579                 printbuf_exit(&buf);
 580                 count_event(c, journal_entry_full);
 581         }
 582
 583         /*
 584          * Journal is full - can't rely on reclaim from work item due to
 585          * freezing:
 586          */
 587         if ((ret == JOURNAL_ERR_journal_full ||
 588              ret == JOURNAL_ERR_journal_pin_full) &&
 589             !(flags & JOURNAL_RES_GET_NONBLOCK)) {
 590                 if (can_discard) {
 591                         bch2_journal_do_discards(j);
 592                         goto retry;
 593                 }
 594
 595                 if (mutex_trylock(&j->reclaim_lock)) {
 596                         bch2_journal_reclaim(j);
 597                         mutex_unlock(&j->reclaim_lock);
 598                 }
 599         }
 600
 601         return ret == JOURNAL_ERR_insufficient_devices
 602                 ? -BCH_ERR_erofs_journal_err
 603                 : -BCH_ERR_journal_res_get_blocked;
 604 }
 605
 606 /*
 607  * Essentially the entry function to the journaling code. When bcachefs is doing
 608  * a btree insert, it calls this function to get the current journal write.
 609  * Journal write is the structure used set up journal writes. The calling
 610  * function will then add its keys to the structure, queuing them for the next
 611  * write.
 612  *
 613  * To ensure forward progress, the current task must not be holding any
 614  * btree node write locks.
 615  */
 616 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
 617                                   unsigned flags)
 618 {
 619         int ret;
 620
 621         closure_wait_event(&j->async_wait,
 622                    (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
 623                    (flags & JOURNAL_RES_GET_NONBLOCK));
 624         return ret;
 625 }
 626
 627 /* journal_entry_res: */
 628
 629 void bch2_journal_entry_res_resize(struct journal *j,
 630                                    struct journal_entry_res *res,
 631                                    unsigned new_u64s)
 632 {
 633         union journal_res_state state;
 634         int d = new_u64s - res->u64s;
 635
 636         spin_lock(&j->lock);
 637
 638         j->entry_u64s_reserved += d;
 639         if (d <= 0)
 640                 goto out;
 641
 642         j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
 643         smp_mb();
 644         state = READ_ONCE(j->reservations);
 645
 646         if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
 647             state.cur_entry_offset > j->cur_entry_u64s) {
 648                 j->cur_entry_u64s += d;
 649                 /*
 650                  * Not enough room in current journal entry, have to flush it:
 651                  */
 652                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 653         } else {
 654                 journal_cur_buf(j)->u64s_reserved += d;
 655         }
 656 out:
 657         spin_unlock(&j->lock);
 658         res->u64s += d;
 659 }
 660
 661 /* journal flushing: */
 662
 663 /**
 664  * bch2_journal_flush_seq_async - wait for a journal entry to be written
 665  * @j:          journal object
 666  * @seq:        seq to flush
 667  * @parent:     closure object to wait with
 668  * Returns:     1 if @seq has already been flushed, 0 if @seq is being flushed,
 669  *              -EIO if @seq will never be flushed
 670  *
 671  * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if
 672  * necessary
 673  */
 674 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
 675                                  struct closure *parent)
 676 {
 677         struct journal_buf *buf;
 678         int ret = 0;
 679
 680         if (seq <= j->flushed_seq_ondisk)
 681                 return 1;
 682
 683         spin_lock(&j->lock);
 684
 685         if (WARN_ONCE(seq > journal_cur_seq(j),
 686                       "requested to flush journal seq %llu, but currently at %llu",
 687                       seq, journal_cur_seq(j)))
 688                 goto out;
 689
 690         /* Recheck under lock: */
 691         if (j->err_seq && seq >= j->err_seq) {
 692                 ret = -EIO;
 693                 goto out;
 694         }
 695
 696         if (seq <= j->flushed_seq_ondisk) {
 697                 ret = 1;
 698                 goto out;
 699         }
 700
 701         /* if seq was written, but not flushed - flush a newer one instead */
 702         seq = max(seq, journal_last_unwritten_seq(j));
 703
 704 recheck_need_open:
 705         if (seq > journal_cur_seq(j)) {
 706                 struct journal_res res = { 0 };
 707
 708                 if (journal_entry_is_open(j))
 709                         __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 710
 711                 spin_unlock(&j->lock);
 712
 713                 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 714                 if (ret)
 715                         return ret;
 716
 717                 seq = res.seq;
 718                 buf = j->buf + (seq & JOURNAL_BUF_MASK);
 719                 buf->must_flush = true;
 720
 721                 if (!buf->flush_time) {
 722                         buf->flush_time = local_clock() ?: 1;
 723                         buf->expires = jiffies;
 724                 }
 725
 726                 if (parent && !closure_wait(&buf->wait, parent))
 727                         BUG();
 728
 729                 bch2_journal_res_put(j, &res);
 730
 731                 spin_lock(&j->lock);
 732                 goto want_write;
 733         }
 734
 735         /*
 736          * if write was kicked off without a flush, flush the next sequence
 737          * number instead
 738          */
 739         buf = journal_seq_to_buf(j, seq);
 740         if (buf->noflush) {
 741                 seq++;
 742                 goto recheck_need_open;
 743         }
 744
 745         buf->must_flush = true;
 746
 747         if (parent && !closure_wait(&buf->wait, parent))
 748                 BUG();
 749 want_write:
 750         if (seq == journal_cur_seq(j))
 751                 journal_entry_want_write(j);
 752 out:
 753         spin_unlock(&j->lock);
 754         return ret;
 755 }
 756
 757 int bch2_journal_flush_seq(struct journal *j, u64 seq)
 758 {
 759         u64 start_time = local_clock();
 760         int ret, ret2;
 761
 762         /*
 763          * Don't update time_stats when @seq is already flushed:
 764          */
 765         if (seq <= j->flushed_seq_ondisk)
 766                 return 0;
 767
 768         ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
 769
 770         if (!ret)
 771                 time_stats_update(j->flush_seq_time, start_time);
 772
 773         return ret ?: ret2 < 0 ? ret2 : 0;
 774 }
 775
 776 /*
 777  * bch2_journal_flush_async - if there is an open journal entry, or a journal
 778  * still being written, write it and wait for the write to complete
 779  */
 780 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
 781 {
 782         bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
 783 }
 784
 785 int bch2_journal_flush(struct journal *j)
 786 {
 787         return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
 788 }
 789
 790 /*
 791  * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
 792  * @seq
 793  */
 794 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
 795 {
 796         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 797         u64 unwritten_seq;
 798         bool ret = false;
 799
 800         if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
 801                 return false;
 802
 803         if (seq <= c->journal.flushed_seq_ondisk)
 804                 return false;
 805
 806         spin_lock(&j->lock);
 807         if (seq <= c->journal.flushed_seq_ondisk)
 808                 goto out;
 809
 810         for (unwritten_seq = journal_last_unwritten_seq(j);
 811              unwritten_seq < seq;
 812              unwritten_seq++) {
 813                 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
 814
 815                 /* journal write is already in flight, and was a flush write: */
 816                 if (unwritten_seq == journal_last_unwritten_seq(j) && !buf->noflush)
 817                         goto out;
 818
 819                 buf->noflush = true;
 820         }
 821
 822         ret = true;
 823 out:
 824         spin_unlock(&j->lock);
 825         return ret;
 826 }
 827
 828 int bch2_journal_meta(struct journal *j)
 829 {
 830         struct journal_buf *buf;
 831         struct journal_res res;
 832         int ret;
 833
 834         memset(&res, 0, sizeof(res));
 835
 836         ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 837         if (ret)
 838                 return ret;
 839
 840         buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
 841         buf->must_flush = true;
 842
 843         if (!buf->flush_time) {
 844                 buf->flush_time = local_clock() ?: 1;
 845                 buf->expires = jiffies;
 846         }
 847
 848         bch2_journal_res_put(j, &res);
 849
 850         return bch2_journal_flush_seq(j, res.seq);
 851 }
 852
 853 /* block/unlock the journal: */
 854
 855 void bch2_journal_unblock(struct journal *j)
 856 {
 857         spin_lock(&j->lock);
 858         j->blocked--;
 859         spin_unlock(&j->lock);
 860
 861         journal_wake(j);
 862 }
 863
 864 void bch2_journal_block(struct journal *j)
 865 {
 866         spin_lock(&j->lock);
 867         j->blocked++;
 868         spin_unlock(&j->lock);
 869
 870         journal_quiesce(j);
 871 }
 872
 873 static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
 874 {
 875         struct journal_buf *ret = NULL;
 876
 877         mutex_lock(&j->buf_lock);
 878         spin_lock(&j->lock);
 879         max_seq = min(max_seq, journal_cur_seq(j));
 880
 881         for (u64 seq = journal_last_unwritten_seq(j);
 882              seq <= max_seq;
 883              seq++) {
 884                 unsigned idx = seq & JOURNAL_BUF_MASK;
 885                 struct journal_buf *buf = j->buf + idx;
 886
 887                 if (buf->need_flush_to_write_buffer) {
 888                         if (seq == journal_cur_seq(j))
 889                                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 890
 891                         union journal_res_state s;
 892                         s.v = atomic64_read_acquire(&j->reservations.counter);
 893
 894                         ret = journal_state_count(s, idx)
 895                                 ? ERR_PTR(-EAGAIN)
 896                                 : buf;
 897                         break;
 898                 }
 899         }
 900
 901         spin_unlock(&j->lock);
 902         if (IS_ERR_OR_NULL(ret))
 903                 mutex_unlock(&j->buf_lock);
 904         return ret;
 905 }
 906
 907 struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
 908 {
 909         struct journal_buf *ret;
 910
 911         wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j, max_seq)) != ERR_PTR(-EAGAIN));
 912         return ret;
 913 }
 914
 915 /* allocate journal on a device: */
 916
 917 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 918                                          bool new_fs, struct closure *cl)
 919 {
 920         struct bch_fs *c = ca->fs;
 921         struct journal_device *ja = &ca->journal;
 922         u64 *new_bucket_seq = NULL, *new_buckets = NULL;
 923         struct open_bucket **ob = NULL;
 924         long *bu = NULL;
 925         unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
 926         int ret = 0;
 927
 928         BUG_ON(nr <= ja->nr);
 929
 930         bu              = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
 931         ob              = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
 932         new_buckets     = kcalloc(nr, sizeof(u64), GFP_KERNEL);
 933         new_bucket_seq  = kcalloc(nr, sizeof(u64), GFP_KERNEL);
 934         if (!bu || !ob || !new_buckets || !new_bucket_seq) {
 935                 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
 936                 goto err_free;
 937         }
 938
 939         for (nr_got = 0; nr_got < nr_want; nr_got++) {
 940                 if (new_fs) {
 941                         bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
 942                         if (bu[nr_got] < 0) {
 943                                 ret = -BCH_ERR_ENOSPC_bucket_alloc;
 944                                 break;
 945                         }
 946                 } else {
 947                         ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal, cl);
 948                         ret = PTR_ERR_OR_ZERO(ob[nr_got]);
 949                         if (ret)
 950                                 break;
 951
 952                         ret = bch2_trans_run(c,
 953                                 bch2_trans_mark_metadata_bucket(trans, ca,
 954                                                 ob[nr_got]->bucket, BCH_DATA_journal,
 955                                                 ca->mi.bucket_size));
 956                         if (ret) {
 957                                 bch2_open_bucket_put(c, ob[nr_got]);
 958                                 bch_err_msg(c, ret, "marking new journal buckets");
 959                                 break;
 960                         }
 961
 962                         bu[nr_got] = ob[nr_got]->bucket;
 963                 }
 964         }
 965
 966         if (!nr_got)
 967                 goto err_free;
 968
 969         /* Don't return an error if we successfully allocated some buckets: */
 970         ret = 0;
 971
 972         if (c) {
 973                 bch2_journal_flush_all_pins(&c->journal);
 974                 bch2_journal_block(&c->journal);
 975                 mutex_lock(&c->sb_lock);
 976         }
 977
 978         memcpy(new_buckets,     ja->buckets,    ja->nr * sizeof(u64));
 979         memcpy(new_bucket_seq,  ja->bucket_seq, ja->nr * sizeof(u64));
 980
 981         BUG_ON(ja->discard_idx > ja->nr);
 982
 983         pos = ja->discard_idx ?: ja->nr;
 984
 985         memmove(new_buckets + pos + nr_got,
 986                 new_buckets + pos,
 987                 sizeof(new_buckets[0]) * (ja->nr - pos));
 988         memmove(new_bucket_seq + pos + nr_got,
 989                 new_bucket_seq + pos,
 990                 sizeof(new_bucket_seq[0]) * (ja->nr - pos));
 991
 992         for (i = 0; i < nr_got; i++) {
 993                 new_buckets[pos + i] = bu[i];
 994                 new_bucket_seq[pos + i] = 0;
 995         }
 996
 997         nr = ja->nr + nr_got;
 998
 999         ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
1000         if (ret)
1001                 goto err_unblock;
1002
1003         if (!new_fs)
1004                 bch2_write_super(c);
1005
1006         /* Commit: */
1007         if (c)
1008                 spin_lock(&c->journal.lock);
1009
1010         swap(new_buckets,       ja->buckets);
1011         swap(new_bucket_seq,    ja->bucket_seq);
1012         ja->nr = nr;
1013
1014         if (pos <= ja->discard_idx)
1015                 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
1016         if (pos <= ja->dirty_idx_ondisk)
1017                 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
1018         if (pos <= ja->dirty_idx)
1019                 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
1020         if (pos <= ja->cur_idx)
1021                 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
1022
1023         if (c)
1024                 spin_unlock(&c->journal.lock);
1025 err_unblock:
1026         if (c) {
1027                 bch2_journal_unblock(&c->journal);
1028                 mutex_unlock(&c->sb_lock);
1029         }
1030
1031         if (ret && !new_fs)
1032                 for (i = 0; i < nr_got; i++)
1033                         bch2_trans_run(c,
1034                                 bch2_trans_mark_metadata_bucket(trans, ca,
1035                                                 bu[i], BCH_DATA_free, 0));
1036 err_free:
1037         if (!new_fs)
1038                 for (i = 0; i < nr_got; i++)
1039                         bch2_open_bucket_put(c, ob[i]);
1040
1041         kfree(new_bucket_seq);
1042         kfree(new_buckets);
1043         kfree(ob);
1044         kfree(bu);
1045         return ret;
1046 }
1047
1048 /*
1049  * Allocate more journal space at runtime - not currently making use if it, but
1050  * the code works:
1051  */
1052 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
1053                                 unsigned nr)
1054 {
1055         struct journal_device *ja = &ca->journal;
1056         struct closure cl;
1057         int ret = 0;
1058
1059         closure_init_stack(&cl);
1060
1061         down_write(&c->state_lock);
1062
1063         /* don't handle reducing nr of buckets yet: */
1064         if (nr < ja->nr)
1065                 goto unlock;
1066
1067         while (ja->nr < nr) {
1068                 struct disk_reservation disk_res = { 0, 0, 0 };
1069
1070                 /*
1071                  * note: journal buckets aren't really counted as _sectors_ used yet, so
1072                  * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
1073                  * when space used goes up without a reservation - but we do need the
1074                  * reservation to ensure we'll actually be able to allocate:
1075                  *
1076                  * XXX: that's not right, disk reservations only ensure a
1077                  * filesystem-wide allocation will succeed, this is a device
1078                  * specific allocation - we can hang here:
1079                  */
1080
1081                 ret = bch2_disk_reservation_get(c, &disk_res,
1082                                                 bucket_to_sector(ca, nr - ja->nr), 1, 0);
1083                 if (ret)
1084                         break;
1085
1086                 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
1087
1088                 bch2_disk_reservation_put(c, &disk_res);
1089
1090                 closure_sync(&cl);
1091
1092                 if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
1093                         break;
1094         }
1095
1096         bch_err_fn(c, ret);
1097 unlock:
1098         up_write(&c->state_lock);
1099         return ret;
1100 }
1101
1102 int bch2_dev_journal_alloc(struct bch_dev *ca)
1103 {
1104         unsigned nr;
1105         int ret;
1106
1107         if (dynamic_fault("bcachefs:add:journal_alloc")) {
1108                 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
1109                 goto err;
1110         }
1111
1112         /* 1/128th of the device by default: */
1113         nr = ca->mi.nbuckets >> 7;
1114
1115         /*
1116          * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
1117          * is smaller:
1118          */
1119         nr = clamp_t(unsigned, nr,
1120                      BCH_JOURNAL_BUCKETS_MIN,
1121                      min(1 << 13,
1122                          (1 << 24) / ca->mi.bucket_size));
1123
1124         ret = __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
1125 err:
1126         bch_err_fn(ca, ret);
1127         return ret;
1128 }
1129
1130 int bch2_fs_journal_alloc(struct bch_fs *c)
1131 {
1132         for_each_online_member(c, ca) {
1133                 if (ca->journal.nr)
1134                         continue;
1135
1136                 int ret = bch2_dev_journal_alloc(ca);
1137                 if (ret) {
1138                         percpu_ref_put(&ca->io_ref);
1139                         return ret;
1140                 }
1141         }
1142
1143         return 0;
1144 }
1145
1146 /* startup/shutdown: */
1147
1148 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1149 {
1150         bool ret = false;
1151         u64 seq;
1152
1153         spin_lock(&j->lock);
1154         for (seq = journal_last_unwritten_seq(j);
1155              seq <= journal_cur_seq(j) && !ret;
1156              seq++) {
1157                 struct journal_buf *buf = journal_seq_to_buf(j, seq);
1158
1159                 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1160                         ret = true;
1161         }
1162         spin_unlock(&j->lock);
1163
1164         return ret;
1165 }
1166
1167 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1168 {
1169         wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1170 }
1171
1172 void bch2_fs_journal_stop(struct journal *j)
1173 {
1174         bch2_journal_reclaim_stop(j);
1175         bch2_journal_flush_all_pins(j);
1176
1177         wait_event(j->wait, bch2_journal_entry_close(j));
1178
1179         /*
1180          * Always write a new journal entry, to make sure the clock hands are up
1181          * to date (and match the superblock)
1182          */
1183         bch2_journal_meta(j);
1184
1185         journal_quiesce(j);
1186
1187         BUG_ON(!bch2_journal_error(j) &&
1188                test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
1189                j->last_empty_seq != journal_cur_seq(j));
1190
1191         cancel_delayed_work_sync(&j->write_work);
1192 }
1193
1194 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1195 {
1196         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1197         struct journal_entry_pin_list *p;
1198         struct journal_replay *i, **_i;
1199         struct genradix_iter iter;
1200         bool had_entries = false;
1201         u64 last_seq = cur_seq, nr, seq;
1202
1203         genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1204                 i = *_i;
1205
1206                 if (!i || i->ignore)
1207                         continue;
1208
1209                 last_seq = le64_to_cpu(i->j.last_seq);
1210                 break;
1211         }
1212
1213         nr = cur_seq - last_seq;
1214
1215         if (nr + 1 > j->pin.size) {
1216                 free_fifo(&j->pin);
1217                 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1218                 if (!j->pin.data) {
1219                         bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1220                         return -BCH_ERR_ENOMEM_journal_pin_fifo;
1221                 }
1222         }
1223
1224         j->replay_journal_seq   = last_seq;
1225         j->replay_journal_seq_end = cur_seq;
1226         j->last_seq_ondisk      = last_seq;
1227         j->flushed_seq_ondisk   = cur_seq - 1;
1228         j->seq_ondisk           = cur_seq - 1;
1229         j->pin.front            = last_seq;
1230         j->pin.back             = cur_seq;
1231         atomic64_set(&j->seq, cur_seq - 1);
1232
1233         fifo_for_each_entry_ptr(p, &j->pin, seq)
1234                 journal_pin_list_init(p, 1);
1235
1236         genradix_for_each(&c->journal_entries, iter, _i) {
1237                 i = *_i;
1238
1239                 if (!i || i->ignore)
1240                         continue;
1241
1242                 seq = le64_to_cpu(i->j.seq);
1243                 BUG_ON(seq >= cur_seq);
1244
1245                 if (seq < last_seq)
1246                         continue;
1247
1248                 if (journal_entry_empty(&i->j))
1249                         j->last_empty_seq = le64_to_cpu(i->j.seq);
1250
1251                 p = journal_seq_pin(j, seq);
1252
1253                 p->devs.nr = 0;
1254                 darray_for_each(i->ptrs, ptr)
1255                         bch2_dev_list_add_dev(&p->devs, ptr->dev);
1256
1257                 had_entries = true;
1258         }
1259
1260         if (!had_entries)
1261                 j->last_empty_seq = cur_seq;
1262
1263         spin_lock(&j->lock);
1264
1265         set_bit(JOURNAL_STARTED, &j->flags);
1266         j->last_flush_write = jiffies;
1267
1268         j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1269         j->reservations.unwritten_idx++;
1270
1271         c->last_bucket_seq_cleanup = journal_cur_seq(j);
1272
1273         bch2_journal_space_available(j);
1274         spin_unlock(&j->lock);
1275
1276         return bch2_journal_reclaim_start(j);
1277 }
1278
1279 /* init/exit: */
1280
1281 void bch2_dev_journal_exit(struct bch_dev *ca)
1282 {
1283         struct journal_device *ja = &ca->journal;
1284
1285         for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
1286                 kfree(ja->bio[i]);
1287                 ja->bio[i] = NULL;
1288         }
1289
1290         kfree(ja->buckets);
1291         kfree(ja->bucket_seq);
1292         ja->buckets     = NULL;
1293         ja->bucket_seq  = NULL;
1294 }
1295
1296 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1297 {
1298         struct journal_device *ja = &ca->journal;
1299         struct bch_sb_field_journal *journal_buckets =
1300                 bch2_sb_field_get(sb, journal);
1301         struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1302                 bch2_sb_field_get(sb, journal_v2);
1303
1304         ja->nr = 0;
1305
1306         if (journal_buckets_v2) {
1307                 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1308
1309                 for (unsigned i = 0; i < nr; i++)
1310                         ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1311         } else if (journal_buckets) {
1312                 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1313         }
1314
1315         ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1316         if (!ja->bucket_seq)
1317                 return -BCH_ERR_ENOMEM_dev_journal_init;
1318
1319         unsigned nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1320
1321         for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
1322                 ja->bio[i] = kmalloc(struct_size(ja->bio[i], bio.bi_inline_vecs,
1323                                      nr_bvecs), GFP_KERNEL);
1324                 if (!ja->bio[i])
1325                         return -BCH_ERR_ENOMEM_dev_journal_init;
1326
1327                 ja->bio[i]->ca = ca;
1328                 ja->bio[i]->buf_idx = i;
1329                 bio_init(&ja->bio[i]->bio, NULL, ja->bio[i]->bio.bi_inline_vecs, nr_bvecs, 0);
1330         }
1331
1332         ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1333         if (!ja->buckets)
1334                 return -BCH_ERR_ENOMEM_dev_journal_init;
1335
1336         if (journal_buckets_v2) {
1337                 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1338                 unsigned dst = 0;
1339
1340                 for (unsigned i = 0; i < nr; i++)
1341                         for (unsigned j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1342                                 ja->buckets[dst++] =
1343                                         le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1344         } else if (journal_buckets) {
1345                 for (unsigned i = 0; i < ja->nr; i++)
1346                         ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1347         }
1348
1349         return 0;
1350 }
1351
1352 void bch2_fs_journal_exit(struct journal *j)
1353 {
1354         if (j->wq)
1355                 destroy_workqueue(j->wq);
1356
1357         darray_exit(&j->early_journal_entries);
1358
1359         for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++)
1360                 kvfree(j->buf[i].data);
1361         free_fifo(&j->pin);
1362 }
1363
1364 int bch2_fs_journal_init(struct journal *j)
1365 {
1366         static struct lock_class_key res_key;
1367
1368         mutex_init(&j->buf_lock);
1369         spin_lock_init(&j->lock);
1370         spin_lock_init(&j->err_lock);
1371         init_waitqueue_head(&j->wait);
1372         INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1373         init_waitqueue_head(&j->reclaim_wait);
1374         init_waitqueue_head(&j->pin_flush_wait);
1375         mutex_init(&j->reclaim_lock);
1376         mutex_init(&j->discard_lock);
1377
1378         lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1379
1380         atomic64_set(&j->reservations.counter,
1381                 ((union journal_res_state)
1382                  { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1383
1384         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)))
1385                 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1386
1387         for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++) {
1388                 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1389                 j->buf[i].data = kvmalloc(j->buf[i].buf_size, GFP_KERNEL);
1390                 if (!j->buf[i].data)
1391                         return -BCH_ERR_ENOMEM_journal_buf;
1392                 j->buf[i].idx = i;
1393         }
1394
1395         j->pin.front = j->pin.back = 1;
1396
1397         j->wq = alloc_workqueue("bcachefs_journal",
1398                                 WQ_HIGHPRI|WQ_FREEZABLE|WQ_UNBOUND|WQ_MEM_RECLAIM, 512);
1399         if (!j->wq)
1400                 return -BCH_ERR_ENOMEM_fs_other_alloc;
1401         return 0;
1402 }
1403
1404 /* debug: */
1405
1406 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1407 {
1408         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1409         union journal_res_state s;
1410         unsigned long now = jiffies;
1411         u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes;
1412
1413         if (!out->nr_tabstops)
1414                 printbuf_tabstop_push(out, 24);
1415         out->atomic++;
1416
1417         rcu_read_lock();
1418         s = READ_ONCE(j->reservations);
1419
1420         prt_printf(out, "dirty journal entries:\t%llu/%llu\n",  fifo_used(&j->pin), j->pin.size);
1421         prt_printf(out, "seq:\t\t\t%llu\n",                     journal_cur_seq(j));
1422         prt_printf(out, "seq_ondisk:\t\t%llu\n",                j->seq_ondisk);
1423         prt_printf(out, "last_seq:\t\t%llu\n",                  journal_last_seq(j));
1424         prt_printf(out, "last_seq_ondisk:\t%llu\n",             j->last_seq_ondisk);
1425         prt_printf(out, "flushed_seq_ondisk:\t%llu\n",          j->flushed_seq_ondisk);
1426         prt_printf(out, "watermark:\t\t%s\n",                   bch2_watermarks[j->watermark]);
1427         prt_printf(out, "each entry reserved:\t%u\n",           j->entry_u64s_reserved);
1428         prt_printf(out, "nr flush writes:\t%llu\n",             j->nr_flush_writes);
1429         prt_printf(out, "nr noflush writes:\t%llu\n",           j->nr_noflush_writes);
1430         prt_printf(out, "average write size:\t");
1431         prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0);
1432         prt_newline(out);
1433         prt_printf(out, "nr direct reclaim:\t%llu\n",           j->nr_direct_reclaim);
1434         prt_printf(out, "nr background reclaim:\t%llu\n",       j->nr_background_reclaim);
1435         prt_printf(out, "reclaim kicked:\t\t%u\n",              j->reclaim_kicked);
1436         prt_printf(out, "reclaim runs in:\t%u ms\n",            time_after(j->next_reclaim, now)
1437                ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1438         prt_printf(out, "blocked:\t\t%u\n",                     j->blocked);
1439         prt_printf(out, "current entry sectors:\t%u\n",         j->cur_entry_sectors);
1440         prt_printf(out, "current entry error:\t%s\n",           bch2_journal_errors[j->cur_entry_error]);
1441         prt_printf(out, "current entry:\t\t");
1442
1443         switch (s.cur_entry_offset) {
1444         case JOURNAL_ENTRY_ERROR_VAL:
1445                 prt_printf(out, "error");
1446                 break;
1447         case JOURNAL_ENTRY_CLOSED_VAL:
1448                 prt_printf(out, "closed");
1449                 break;
1450         default:
1451                 prt_printf(out, "%u/%u", s.cur_entry_offset, j->cur_entry_u64s);
1452                 break;
1453         }
1454
1455         prt_newline(out);
1456         prt_printf(out, "unwritten entries:");
1457         prt_newline(out);
1458         bch2_journal_bufs_to_text(out, j);
1459
1460         prt_printf(out,
1461                "replay done:\t\t%i\n",
1462                test_bit(JOURNAL_REPLAY_DONE,    &j->flags));
1463
1464         prt_printf(out, "space:\n");
1465         prt_printf(out, "\tdiscarded\t%u:%u\n",
1466                j->space[journal_space_discarded].next_entry,
1467                j->space[journal_space_discarded].total);
1468         prt_printf(out, "\tclean ondisk\t%u:%u\n",
1469                j->space[journal_space_clean_ondisk].next_entry,
1470                j->space[journal_space_clean_ondisk].total);
1471         prt_printf(out, "\tclean\t\t%u:%u\n",
1472                j->space[journal_space_clean].next_entry,
1473                j->space[journal_space_clean].total);
1474         prt_printf(out, "\ttotal\t\t%u:%u\n",
1475                j->space[journal_space_total].next_entry,
1476                j->space[journal_space_total].total);
1477
1478         for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
1479                 struct journal_device *ja = &ca->journal;
1480
1481                 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1482                         continue;
1483
1484                 if (!ja->nr)
1485                         continue;
1486
1487                 prt_printf(out, "dev %u:\n",            ca->dev_idx);
1488                 prt_printf(out, "\tnr\t\t%u\n",         ja->nr);
1489                 prt_printf(out, "\tbucket size\t%u\n",  ca->mi.bucket_size);
1490                 prt_printf(out, "\tavailable\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1491                 prt_printf(out, "\tdiscard_idx\t%u\n",  ja->discard_idx);
1492                 prt_printf(out, "\tdirty_ondisk\t%u (seq %llu)\n", ja->dirty_idx_ondisk,        ja->bucket_seq[ja->dirty_idx_ondisk]);
1493                 prt_printf(out, "\tdirty_idx\t%u (seq %llu)\n", ja->dirty_idx,          ja->bucket_seq[ja->dirty_idx]);
1494                 prt_printf(out, "\tcur_idx\t\t%u (seq %llu)\n", ja->cur_idx,            ja->bucket_seq[ja->cur_idx]);
1495         }
1496
1497         rcu_read_unlock();
1498
1499         --out->atomic;
1500 }
1501
1502 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1503 {
1504         spin_lock(&j->lock);
1505         __bch2_journal_debug_to_text(out, j);
1506         spin_unlock(&j->lock);
1507 }
1508
1509 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1510 {
1511         struct journal_entry_pin_list *pin_list;
1512         struct journal_entry_pin *pin;
1513
1514         spin_lock(&j->lock);
1515         *seq = max(*seq, j->pin.front);
1516
1517         if (*seq >= j->pin.back) {
1518                 spin_unlock(&j->lock);
1519                 return true;
1520         }
1521
1522         out->atomic++;
1523
1524         pin_list = journal_seq_pin(j, *seq);
1525
1526         prt_printf(out, "%llu: count %u", *seq, atomic_read(&pin_list->count));
1527         prt_newline(out);
1528         printbuf_indent_add(out, 2);
1529
1530         for (unsigned i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1531                 list_for_each_entry(pin, &pin_list->list[i], list) {
1532                         prt_printf(out, "\t%px %ps", pin, pin->flush);
1533                         prt_newline(out);
1534                 }
1535
1536         if (!list_empty(&pin_list->flushed)) {
1537                 prt_printf(out, "flushed:");
1538                 prt_newline(out);
1539         }
1540
1541         list_for_each_entry(pin, &pin_list->flushed, list) {
1542                 prt_printf(out, "\t%px %ps", pin, pin->flush);
1543                 prt_newline(out);
1544         }
1545
1546         printbuf_indent_sub(out, 2);
1547
1548         --out->atomic;
1549         spin_unlock(&j->lock);
1550
1551         return false;
1552 }
1553
1554 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1555 {
1556         u64 seq = 0;
1557
1558         while (!bch2_journal_seq_pins_to_text(out, j, &seq))
1559                 seq++;
1560 }