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_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->write_started)
  88                 prt_str(out, "write_started ");
  89         if (buf->write_allocated)
  90                 prt_str(out, "write allocated ");
  91         if (buf->write_done)
  92                 prt_str(out, "write done");
  93         prt_newline(out);
  94
  95         printbuf_indent_sub(out, 2);
  96 }
  97
  98 static void bch2_journal_bufs_to_text(struct printbuf *out, struct journal *j)
  99 {
 100         if (!out->nr_tabstops)
 101                 printbuf_tabstop_push(out, 24);
 102
 103         for (u64 seq = journal_last_unwritten_seq(j);
 104              seq <= journal_cur_seq(j);
 105              seq++)
 106                 bch2_journal_buf_to_text(out, j, seq);
 107         prt_printf(out, "last buf %s\n", journal_entry_is_open(j) ? "open" : "closed");
 108 }
 109
 110 static inline struct journal_buf *
 111 journal_seq_to_buf(struct journal *j, u64 seq)
 112 {
 113         struct journal_buf *buf = NULL;
 114
 115         EBUG_ON(seq > journal_cur_seq(j));
 116
 117         if (journal_seq_unwritten(j, seq)) {
 118                 buf = j->buf + (seq & JOURNAL_BUF_MASK);
 119                 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
 120         }
 121         return buf;
 122 }
 123
 124 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
 125 {
 126         unsigned i;
 127
 128         for (i = 0; i < ARRAY_SIZE(p->list); i++)
 129                 INIT_LIST_HEAD(&p->list[i]);
 130         INIT_LIST_HEAD(&p->flushed);
 131         atomic_set(&p->count, count);
 132         p->devs.nr = 0;
 133 }
 134
 135 /*
 136  * Detect stuck journal conditions and trigger shutdown. Technically the journal
 137  * can end up stuck for a variety of reasons, such as a blocked I/O, journal
 138  * reservation lockup, etc. Since this is a fatal error with potentially
 139  * unpredictable characteristics, we want to be fairly conservative before we
 140  * decide to shut things down.
 141  *
 142  * Consider the journal stuck when it appears full with no ability to commit
 143  * btree transactions, to discard journal buckets, nor acquire priority
 144  * (reserved watermark) reservation.
 145  */
 146 static inline bool
 147 journal_error_check_stuck(struct journal *j, int error, unsigned flags)
 148 {
 149         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 150         bool stuck = false;
 151         struct printbuf buf = PRINTBUF;
 152
 153         if (!(error == JOURNAL_ERR_journal_full ||
 154               error == JOURNAL_ERR_journal_pin_full) ||
 155             nr_unwritten_journal_entries(j) ||
 156             (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim)
 157                 return stuck;
 158
 159         spin_lock(&j->lock);
 160
 161         if (j->can_discard) {
 162                 spin_unlock(&j->lock);
 163                 return stuck;
 164         }
 165
 166         stuck = true;
 167
 168         /*
 169          * The journal shutdown path will set ->err_seq, but do it here first to
 170          * serialize against concurrent failures and avoid duplicate error
 171          * reports.
 172          */
 173         if (j->err_seq) {
 174                 spin_unlock(&j->lock);
 175                 return stuck;
 176         }
 177         j->err_seq = journal_cur_seq(j);
 178         spin_unlock(&j->lock);
 179
 180         bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)",
 181                 bch2_journal_errors[error]);
 182         bch2_journal_debug_to_text(&buf, j);
 183         bch_err(c, "%s", buf.buf);
 184
 185         printbuf_reset(&buf);
 186         bch2_journal_pins_to_text(&buf, j);
 187         bch_err(c, "Journal pins:\n%s", buf.buf);
 188         printbuf_exit(&buf);
 189
 190         bch2_fatal_error(c);
 191         dump_stack();
 192
 193         return stuck;
 194 }
 195
 196 void bch2_journal_do_writes(struct journal *j)
 197 {
 198         for (u64 seq = journal_last_unwritten_seq(j);
 199              seq <= journal_cur_seq(j);
 200              seq++) {
 201                 unsigned idx = seq & JOURNAL_BUF_MASK;
 202                 struct journal_buf *w = j->buf + idx;
 203
 204                 if (w->write_started && !w->write_allocated)
 205                         break;
 206                 if (w->write_started)
 207                         continue;
 208
 209                 if (!journal_state_count(j->reservations, idx)) {
 210                         w->write_started = true;
 211                         closure_call(&w->io, bch2_journal_write, j->wq, NULL);
 212                 }
 213
 214                 break;
 215         }
 216 }
 217
 218 /*
 219  * Final processing when the last reference of a journal buffer has been
 220  * dropped. Drop the pin list reference acquired at journal entry open and write
 221  * the buffer, if requested.
 222  */
 223 void bch2_journal_buf_put_final(struct journal *j, u64 seq)
 224 {
 225         lockdep_assert_held(&j->lock);
 226
 227         if (__bch2_journal_pin_put(j, seq))
 228                 bch2_journal_reclaim_fast(j);
 229         bch2_journal_do_writes(j);
 230 }
 231
 232 /*
 233  * Returns true if journal entry is now closed:
 234  *
 235  * We don't close a journal_buf until the next journal_buf is finished writing,
 236  * and can be opened again - this also initializes the next journal_buf:
 237  */
 238 static void __journal_entry_close(struct journal *j, unsigned closed_val, bool trace)
 239 {
 240         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 241         struct journal_buf *buf = journal_cur_buf(j);
 242         union journal_res_state old, new;
 243         u64 v = atomic64_read(&j->reservations.counter);
 244         unsigned sectors;
 245
 246         BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
 247                closed_val != JOURNAL_ENTRY_ERROR_VAL);
 248
 249         lockdep_assert_held(&j->lock);
 250
 251         do {
 252                 old.v = new.v = v;
 253                 new.cur_entry_offset = closed_val;
 254
 255                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
 256                     old.cur_entry_offset == new.cur_entry_offset)
 257                         return;
 258         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 259                                        old.v, new.v)) != old.v);
 260
 261         if (!__journal_entry_is_open(old))
 262                 return;
 263
 264         /* Close out old buffer: */
 265         buf->data->u64s         = cpu_to_le32(old.cur_entry_offset);
 266
 267         if (trace_journal_entry_close_enabled() && trace) {
 268                 struct printbuf pbuf = PRINTBUF;
 269                 pbuf.atomic++;
 270
 271                 prt_str(&pbuf, "entry size: ");
 272                 prt_human_readable_u64(&pbuf, vstruct_bytes(buf->data));
 273                 prt_newline(&pbuf);
 274                 bch2_prt_task_backtrace(&pbuf, current, 1, GFP_NOWAIT);
 275                 trace_journal_entry_close(c, pbuf.buf);
 276                 printbuf_exit(&pbuf);
 277         }
 278
 279         sectors = vstruct_blocks_plus(buf->data, c->block_bits,
 280                                       buf->u64s_reserved) << c->block_bits;
 281         BUG_ON(sectors > buf->sectors);
 282         buf->sectors = sectors;
 283
 284         /*
 285          * We have to set last_seq here, _before_ opening a new journal entry:
 286          *
 287          * A threads may replace an old pin with a new pin on their current
 288          * journal reservation - the expectation being that the journal will
 289          * contain either what the old pin protected or what the new pin
 290          * protects.
 291          *
 292          * After the old pin is dropped journal_last_seq() won't include the old
 293          * pin, so we can only write the updated last_seq on the entry that
 294          * contains whatever the new pin protects.
 295          *
 296          * Restated, we can _not_ update last_seq for a given entry if there
 297          * could be a newer entry open with reservations/pins that have been
 298          * taken against it.
 299          *
 300          * Hence, we want update/set last_seq on the current journal entry right
 301          * before we open a new one:
 302          */
 303         buf->last_seq           = journal_last_seq(j);
 304         buf->data->last_seq     = cpu_to_le64(buf->last_seq);
 305         BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
 306
 307         cancel_delayed_work(&j->write_work);
 308
 309         bch2_journal_space_available(j);
 310
 311         __bch2_journal_buf_put(j, old.idx, le64_to_cpu(buf->data->seq));
 312 }
 313
 314 void bch2_journal_halt(struct journal *j)
 315 {
 316         spin_lock(&j->lock);
 317         __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL, true);
 318         if (!j->err_seq)
 319                 j->err_seq = journal_cur_seq(j);
 320         journal_wake(j);
 321         spin_unlock(&j->lock);
 322 }
 323
 324 static bool journal_entry_want_write(struct journal *j)
 325 {
 326         bool ret = !journal_entry_is_open(j) ||
 327                 journal_cur_seq(j) == journal_last_unwritten_seq(j);
 328
 329         /* Don't close it yet if we already have a write in flight: */
 330         if (ret)
 331                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 332         else if (nr_unwritten_journal_entries(j)) {
 333                 struct journal_buf *buf = journal_cur_buf(j);
 334
 335                 if (!buf->flush_time) {
 336                         buf->flush_time = local_clock() ?: 1;
 337                         buf->expires = jiffies;
 338                 }
 339         }
 340
 341         return ret;
 342 }
 343
 344 bool bch2_journal_entry_close(struct journal *j)
 345 {
 346         bool ret;
 347
 348         spin_lock(&j->lock);
 349         ret = journal_entry_want_write(j);
 350         spin_unlock(&j->lock);
 351
 352         return ret;
 353 }
 354
 355 /*
 356  * should _only_ called from journal_res_get() - when we actually want a
 357  * journal reservation - journal entry is open means journal is dirty:
 358  */
 359 static int journal_entry_open(struct journal *j)
 360 {
 361         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 362         struct journal_buf *buf = j->buf +
 363                 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
 364         union journal_res_state old, new;
 365         int u64s;
 366         u64 v;
 367
 368         lockdep_assert_held(&j->lock);
 369         BUG_ON(journal_entry_is_open(j));
 370         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
 371
 372         if (j->blocked)
 373                 return JOURNAL_ERR_blocked;
 374
 375         if (j->cur_entry_error)
 376                 return j->cur_entry_error;
 377
 378         if (bch2_journal_error(j))
 379                 return JOURNAL_ERR_insufficient_devices; /* -EROFS */
 380
 381         if (!fifo_free(&j->pin))
 382                 return JOURNAL_ERR_journal_pin_full;
 383
 384         if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
 385                 return JOURNAL_ERR_max_in_flight;
 386
 387         BUG_ON(!j->cur_entry_sectors);
 388
 389         buf->expires            =
 390                 (journal_cur_seq(j) == j->flushed_seq_ondisk
 391                  ? jiffies
 392                  : j->last_flush_write) +
 393                 msecs_to_jiffies(c->opts.journal_flush_delay);
 394
 395         buf->u64s_reserved      = j->entry_u64s_reserved;
 396         buf->disk_sectors       = j->cur_entry_sectors;
 397         buf->sectors            = min(buf->disk_sectors, buf->buf_size >> 9);
 398
 399         u64s = (int) (buf->sectors << 9) / sizeof(u64) -
 400                 journal_entry_overhead(j);
 401         u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
 402
 403         if (u64s <= (ssize_t) j->early_journal_entries.nr)
 404                 return JOURNAL_ERR_journal_full;
 405
 406         if (fifo_empty(&j->pin) && j->reclaim_thread)
 407                 wake_up_process(j->reclaim_thread);
 408
 409         /*
 410          * The fifo_push() needs to happen at the same time as j->seq is
 411          * incremented for journal_last_seq() to be calculated correctly
 412          */
 413         atomic64_inc(&j->seq);
 414         journal_pin_list_init(fifo_push_ref(&j->pin), 1);
 415
 416         BUG_ON(j->pin.back - 1 != atomic64_read(&j->seq));
 417
 418         BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
 419
 420         bkey_extent_init(&buf->key);
 421         buf->noflush            = false;
 422         buf->must_flush         = false;
 423         buf->separate_flush     = false;
 424         buf->flush_time         = 0;
 425         buf->need_flush_to_write_buffer = true;
 426         buf->write_started      = false;
 427         buf->write_allocated    = false;
 428         buf->write_done         = false;
 429
 430         memset(buf->data, 0, sizeof(*buf->data));
 431         buf->data->seq  = cpu_to_le64(journal_cur_seq(j));
 432         buf->data->u64s = 0;
 433
 434         if (j->early_journal_entries.nr) {
 435                 memcpy(buf->data->_data, j->early_journal_entries.data,
 436                        j->early_journal_entries.nr * sizeof(u64));
 437                 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
 438         }
 439
 440         /*
 441          * Must be set before marking the journal entry as open:
 442          */
 443         j->cur_entry_u64s = u64s;
 444
 445         v = atomic64_read(&j->reservations.counter);
 446         do {
 447                 old.v = new.v = v;
 448
 449                 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
 450
 451                 new.idx++;
 452                 BUG_ON(journal_state_count(new, new.idx));
 453                 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
 454
 455                 journal_state_inc(&new);
 456
 457                 /* Handle any already added entries */
 458                 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
 459         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 460                                        old.v, new.v)) != old.v);
 461
 462         if (nr_unwritten_journal_entries(j) == 1)
 463                 mod_delayed_work(j->wq,
 464                                  &j->write_work,
 465                                  msecs_to_jiffies(c->opts.journal_flush_delay));
 466         journal_wake(j);
 467
 468         if (j->early_journal_entries.nr)
 469                 darray_exit(&j->early_journal_entries);
 470         return 0;
 471 }
 472
 473 static bool journal_quiesced(struct journal *j)
 474 {
 475         bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
 476
 477         if (!ret)
 478                 bch2_journal_entry_close(j);
 479         return ret;
 480 }
 481
 482 static void journal_quiesce(struct journal *j)
 483 {
 484         wait_event(j->wait, journal_quiesced(j));
 485 }
 486
 487 static void journal_write_work(struct work_struct *work)
 488 {
 489         struct journal *j = container_of(work, struct journal, write_work.work);
 490
 491         spin_lock(&j->lock);
 492         if (__journal_entry_is_open(j->reservations)) {
 493                 long delta = journal_cur_buf(j)->expires - jiffies;
 494
 495                 if (delta > 0)
 496                         mod_delayed_work(j->wq, &j->write_work, delta);
 497                 else
 498                         __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 499         }
 500         spin_unlock(&j->lock);
 501 }
 502
 503 static int __journal_res_get(struct journal *j, struct journal_res *res,
 504                              unsigned flags)
 505 {
 506         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 507         struct journal_buf *buf;
 508         bool can_discard;
 509         int ret;
 510 retry:
 511         if (journal_res_get_fast(j, res, flags))
 512                 return 0;
 513
 514         if (bch2_journal_error(j))
 515                 return -BCH_ERR_erofs_journal_err;
 516
 517         if (j->blocked)
 518                 return -BCH_ERR_journal_res_get_blocked;
 519
 520         if ((flags & BCH_WATERMARK_MASK) < j->watermark) {
 521                 ret = JOURNAL_ERR_journal_full;
 522                 can_discard = j->can_discard;
 523                 goto out;
 524         }
 525
 526         if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf) && !journal_entry_is_open(j)) {
 527                 ret = JOURNAL_ERR_max_in_flight;
 528                 goto out;
 529         }
 530
 531         spin_lock(&j->lock);
 532
 533         /*
 534          * Recheck after taking the lock, so we don't race with another thread
 535          * that just did journal_entry_open() and call bch2_journal_entry_close()
 536          * unnecessarily
 537          */
 538         if (journal_res_get_fast(j, res, flags)) {
 539                 ret = 0;
 540                 goto unlock;
 541         }
 542
 543         /*
 544          * If we couldn't get a reservation because the current buf filled up,
 545          * and we had room for a bigger entry on disk, signal that we want to
 546          * realloc the journal bufs:
 547          */
 548         buf = journal_cur_buf(j);
 549         if (journal_entry_is_open(j) &&
 550             buf->buf_size >> 9 < buf->disk_sectors &&
 551             buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
 552                 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
 553
 554         __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, false);
 555         ret = journal_entry_open(j) ?: JOURNAL_ERR_retry;
 556 unlock:
 557         can_discard = j->can_discard;
 558         spin_unlock(&j->lock);
 559 out:
 560         if (ret == JOURNAL_ERR_retry)
 561                 goto retry;
 562         if (!ret)
 563                 return 0;
 564
 565         if (journal_error_check_stuck(j, ret, flags))
 566                 ret = -BCH_ERR_journal_res_get_blocked;
 567
 568         if (ret == JOURNAL_ERR_max_in_flight &&
 569             track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight], true)) {
 570
 571                 struct printbuf buf = PRINTBUF;
 572                 prt_printf(&buf, "seq %llu\n", journal_cur_seq(j));
 573                 bch2_journal_bufs_to_text(&buf, j);
 574                 trace_journal_entry_full(c, buf.buf);
 575                 printbuf_exit(&buf);
 576                 count_event(c, journal_entry_full);
 577         }
 578
 579         /*
 580          * Journal is full - can't rely on reclaim from work item due to
 581          * freezing:
 582          */
 583         if ((ret == JOURNAL_ERR_journal_full ||
 584              ret == JOURNAL_ERR_journal_pin_full) &&
 585             !(flags & JOURNAL_RES_GET_NONBLOCK)) {
 586                 if (can_discard) {
 587                         bch2_journal_do_discards(j);
 588                         goto retry;
 589                 }
 590
 591                 if (mutex_trylock(&j->reclaim_lock)) {
 592                         bch2_journal_reclaim(j);
 593                         mutex_unlock(&j->reclaim_lock);
 594                 }
 595         }
 596
 597         return ret == JOURNAL_ERR_insufficient_devices
 598                 ? -BCH_ERR_erofs_journal_err
 599                 : -BCH_ERR_journal_res_get_blocked;
 600 }
 601
 602 /*
 603  * Essentially the entry function to the journaling code. When bcachefs is doing
 604  * a btree insert, it calls this function to get the current journal write.
 605  * Journal write is the structure used set up journal writes. The calling
 606  * function will then add its keys to the structure, queuing them for the next
 607  * write.
 608  *
 609  * To ensure forward progress, the current task must not be holding any
 610  * btree node write locks.
 611  */
 612 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
 613                                   unsigned flags)
 614 {
 615         int ret;
 616
 617         closure_wait_event(&j->async_wait,
 618                    (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
 619                    (flags & JOURNAL_RES_GET_NONBLOCK));
 620         return ret;
 621 }
 622
 623 /* journal_entry_res: */
 624
 625 void bch2_journal_entry_res_resize(struct journal *j,
 626                                    struct journal_entry_res *res,
 627                                    unsigned new_u64s)
 628 {
 629         union journal_res_state state;
 630         int d = new_u64s - res->u64s;
 631
 632         spin_lock(&j->lock);
 633
 634         j->entry_u64s_reserved += d;
 635         if (d <= 0)
 636                 goto out;
 637
 638         j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
 639         smp_mb();
 640         state = READ_ONCE(j->reservations);
 641
 642         if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
 643             state.cur_entry_offset > j->cur_entry_u64s) {
 644                 j->cur_entry_u64s += d;
 645                 /*
 646                  * Not enough room in current journal entry, have to flush it:
 647                  */
 648                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 649         } else {
 650                 journal_cur_buf(j)->u64s_reserved += d;
 651         }
 652 out:
 653         spin_unlock(&j->lock);
 654         res->u64s += d;
 655 }
 656
 657 /* journal flushing: */
 658
 659 /**
 660  * bch2_journal_flush_seq_async - wait for a journal entry to be written
 661  * @j:          journal object
 662  * @seq:        seq to flush
 663  * @parent:     closure object to wait with
 664  * Returns:     1 if @seq has already been flushed, 0 if @seq is being flushed,
 665  *              -EIO if @seq will never be flushed
 666  *
 667  * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if
 668  * necessary
 669  */
 670 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
 671                                  struct closure *parent)
 672 {
 673         struct journal_buf *buf;
 674         int ret = 0;
 675
 676         if (seq <= j->flushed_seq_ondisk)
 677                 return 1;
 678
 679         spin_lock(&j->lock);
 680
 681         if (WARN_ONCE(seq > journal_cur_seq(j),
 682                       "requested to flush journal seq %llu, but currently at %llu",
 683                       seq, journal_cur_seq(j)))
 684                 goto out;
 685
 686         /* Recheck under lock: */
 687         if (j->err_seq && seq >= j->err_seq) {
 688                 ret = -EIO;
 689                 goto out;
 690         }
 691
 692         if (seq <= j->flushed_seq_ondisk) {
 693                 ret = 1;
 694                 goto out;
 695         }
 696
 697         /* if seq was written, but not flushed - flush a newer one instead */
 698         seq = max(seq, journal_last_unwritten_seq(j));
 699
 700 recheck_need_open:
 701         if (seq > journal_cur_seq(j)) {
 702                 struct journal_res res = { 0 };
 703
 704                 if (journal_entry_is_open(j))
 705                         __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 706
 707                 spin_unlock(&j->lock);
 708
 709                 /*
 710                  * We're called from bch2_journal_flush_seq() -> wait_event();
 711                  * but this might block. We won't usually block, so we won't
 712                  * livelock:
 713                  */
 714                 sched_annotate_sleep();
 715                 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 716                 if (ret)
 717                         return ret;
 718
 719                 seq = res.seq;
 720                 buf = journal_seq_to_buf(j, seq);
 721                 buf->must_flush = true;
 722
 723                 if (!buf->flush_time) {
 724                         buf->flush_time = local_clock() ?: 1;
 725                         buf->expires = jiffies;
 726                 }
 727
 728                 if (parent && !closure_wait(&buf->wait, parent))
 729                         BUG();
 730
 731                 bch2_journal_res_put(j, &res);
 732
 733                 spin_lock(&j->lock);
 734                 goto want_write;
 735         }
 736
 737         /*
 738          * if write was kicked off without a flush, or if we promised it
 739          * wouldn't be a flush, flush the next sequence number instead
 740          */
 741         buf = journal_seq_to_buf(j, seq);
 742         if (buf->noflush) {
 743                 seq++;
 744                 goto recheck_need_open;
 745         }
 746
 747         buf->must_flush = true;
 748
 749         if (parent && !closure_wait(&buf->wait, parent))
 750                 BUG();
 751 want_write:
 752         if (seq == journal_cur_seq(j))
 753                 journal_entry_want_write(j);
 754 out:
 755         spin_unlock(&j->lock);
 756         return ret;
 757 }
 758
 759 int bch2_journal_flush_seq(struct journal *j, u64 seq)
 760 {
 761         u64 start_time = local_clock();
 762         int ret, ret2;
 763
 764         /*
 765          * Don't update time_stats when @seq is already flushed:
 766          */
 767         if (seq <= j->flushed_seq_ondisk)
 768                 return 0;
 769
 770         ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
 771
 772         if (!ret)
 773                 bch2_time_stats_update(j->flush_seq_time, start_time);
 774
 775         return ret ?: ret2 < 0 ? ret2 : 0;
 776 }
 777
 778 /*
 779  * bch2_journal_flush_async - if there is an open journal entry, or a journal
 780  * still being written, write it and wait for the write to complete
 781  */
 782 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
 783 {
 784         bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
 785 }
 786
 787 int bch2_journal_flush(struct journal *j)
 788 {
 789         return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
 790 }
 791
 792 /*
 793  * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
 794  * @seq
 795  */
 796 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
 797 {
 798         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 799         u64 unwritten_seq;
 800         bool ret = false;
 801
 802         if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
 803                 return false;
 804
 805         if (seq <= c->journal.flushed_seq_ondisk)
 806                 return false;
 807
 808         spin_lock(&j->lock);
 809         if (seq <= c->journal.flushed_seq_ondisk)
 810                 goto out;
 811
 812         for (unwritten_seq = journal_last_unwritten_seq(j);
 813              unwritten_seq < seq;
 814              unwritten_seq++) {
 815                 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
 816
 817                 /* journal flush already in flight, or flush requseted */
 818                 if (buf->must_flush)
 819                         goto out;
 820
 821                 buf->noflush = true;
 822         }
 823
 824         ret = true;
 825 out:
 826         spin_unlock(&j->lock);
 827         return ret;
 828 }
 829
 830 int bch2_journal_meta(struct journal *j)
 831 {
 832         struct journal_buf *buf;
 833         struct journal_res res;
 834         int ret;
 835
 836         memset(&res, 0, sizeof(res));
 837
 838         ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 839         if (ret)
 840                 return ret;
 841
 842         buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
 843         buf->must_flush = true;
 844
 845         if (!buf->flush_time) {
 846                 buf->flush_time = local_clock() ?: 1;
 847                 buf->expires = jiffies;
 848         }
 849
 850         bch2_journal_res_put(j, &res);
 851
 852         return bch2_journal_flush_seq(j, res.seq);
 853 }
 854
 855 /* block/unlock the journal: */
 856
 857 void bch2_journal_unblock(struct journal *j)
 858 {
 859         spin_lock(&j->lock);
 860         j->blocked--;
 861         spin_unlock(&j->lock);
 862
 863         journal_wake(j);
 864 }
 865
 866 void bch2_journal_block(struct journal *j)
 867 {
 868         spin_lock(&j->lock);
 869         j->blocked++;
 870         spin_unlock(&j->lock);
 871
 872         journal_quiesce(j);
 873 }
 874
 875 static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
 876 {
 877         struct journal_buf *ret = NULL;
 878
 879         /* We're inside wait_event(), but using mutex_lock(: */
 880         sched_annotate_sleep();
 881         mutex_lock(&j->buf_lock);
 882         spin_lock(&j->lock);
 883         max_seq = min(max_seq, journal_cur_seq(j));
 884
 885         for (u64 seq = journal_last_unwritten_seq(j);
 886              seq <= max_seq;
 887              seq++) {
 888                 unsigned idx = seq & JOURNAL_BUF_MASK;
 889                 struct journal_buf *buf = j->buf + idx;
 890
 891                 if (buf->need_flush_to_write_buffer) {
 892                         if (seq == journal_cur_seq(j))
 893                                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 894
 895                         union journal_res_state s;
 896                         s.v = atomic64_read_acquire(&j->reservations.counter);
 897
 898                         ret = journal_state_count(s, idx)
 899                                 ? ERR_PTR(-EAGAIN)
 900                                 : buf;
 901                         break;
 902                 }
 903         }
 904
 905         spin_unlock(&j->lock);
 906         if (IS_ERR_OR_NULL(ret))
 907                 mutex_unlock(&j->buf_lock);
 908         return ret;
 909 }
 910
 911 struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
 912 {
 913         struct journal_buf *ret;
 914
 915         wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j, max_seq)) != ERR_PTR(-EAGAIN));
 916         return ret;
 917 }
 918
 919 /* allocate journal on a device: */
 920
 921 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 922                                          bool new_fs, struct closure *cl)
 923 {
 924         struct bch_fs *c = ca->fs;
 925         struct journal_device *ja = &ca->journal;
 926         u64 *new_bucket_seq = NULL, *new_buckets = NULL;
 927         struct open_bucket **ob = NULL;
 928         long *bu = NULL;
 929         unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
 930         int ret = 0;
 931
 932         BUG_ON(nr <= ja->nr);
 933
 934         bu              = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
 935         ob              = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
 936         new_buckets     = kcalloc(nr, sizeof(u64), GFP_KERNEL);
 937         new_bucket_seq  = kcalloc(nr, sizeof(u64), GFP_KERNEL);
 938         if (!bu || !ob || !new_buckets || !new_bucket_seq) {
 939                 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
 940                 goto err_free;
 941         }
 942
 943         for (nr_got = 0; nr_got < nr_want; nr_got++) {
 944                 if (new_fs) {
 945                         bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
 946                         if (bu[nr_got] < 0) {
 947                                 ret = -BCH_ERR_ENOSPC_bucket_alloc;
 948                                 break;
 949                         }
 950                 } else {
 951                         ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal, cl);
 952                         ret = PTR_ERR_OR_ZERO(ob[nr_got]);
 953                         if (ret)
 954                                 break;
 955
 956                         ret = bch2_trans_run(c,
 957                                 bch2_trans_mark_metadata_bucket(trans, ca,
 958                                                 ob[nr_got]->bucket, BCH_DATA_journal,
 959                                                 ca->mi.bucket_size));
 960                         if (ret) {
 961                                 bch2_open_bucket_put(c, ob[nr_got]);
 962                                 bch_err_msg(c, ret, "marking new journal buckets");
 963                                 break;
 964                         }
 965
 966                         bu[nr_got] = ob[nr_got]->bucket;
 967                 }
 968         }
 969
 970         if (!nr_got)
 971                 goto err_free;
 972
 973         /* Don't return an error if we successfully allocated some buckets: */
 974         ret = 0;
 975
 976         if (c) {
 977                 bch2_journal_flush_all_pins(&c->journal);
 978                 bch2_journal_block(&c->journal);
 979                 mutex_lock(&c->sb_lock);
 980         }
 981
 982         memcpy(new_buckets,     ja->buckets,    ja->nr * sizeof(u64));
 983         memcpy(new_bucket_seq,  ja->bucket_seq, ja->nr * sizeof(u64));
 984
 985         BUG_ON(ja->discard_idx > ja->nr);
 986
 987         pos = ja->discard_idx ?: ja->nr;
 988
 989         memmove(new_buckets + pos + nr_got,
 990                 new_buckets + pos,
 991                 sizeof(new_buckets[0]) * (ja->nr - pos));
 992         memmove(new_bucket_seq + pos + nr_got,
 993                 new_bucket_seq + pos,
 994                 sizeof(new_bucket_seq[0]) * (ja->nr - pos));
 995
 996         for (i = 0; i < nr_got; i++) {
 997                 new_buckets[pos + i] = bu[i];
 998                 new_bucket_seq[pos + i] = 0;
 999         }
1000
1001         nr = ja->nr + nr_got;
1002
1003         ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
1004         if (ret)
1005                 goto err_unblock;
1006
1007         if (!new_fs)
1008                 bch2_write_super(c);
1009
1010         /* Commit: */
1011         if (c)
1012                 spin_lock(&c->journal.lock);
1013
1014         swap(new_buckets,       ja->buckets);
1015         swap(new_bucket_seq,    ja->bucket_seq);
1016         ja->nr = nr;
1017
1018         if (pos <= ja->discard_idx)
1019                 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
1020         if (pos <= ja->dirty_idx_ondisk)
1021                 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
1022         if (pos <= ja->dirty_idx)
1023                 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
1024         if (pos <= ja->cur_idx)
1025                 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
1026
1027         if (c)
1028                 spin_unlock(&c->journal.lock);
1029 err_unblock:
1030         if (c) {
1031                 bch2_journal_unblock(&c->journal);
1032                 mutex_unlock(&c->sb_lock);
1033         }
1034
1035         if (ret && !new_fs)
1036                 for (i = 0; i < nr_got; i++)
1037                         bch2_trans_run(c,
1038                                 bch2_trans_mark_metadata_bucket(trans, ca,
1039                                                 bu[i], BCH_DATA_free, 0));
1040 err_free:
1041         if (!new_fs)
1042                 for (i = 0; i < nr_got; i++)
1043                         bch2_open_bucket_put(c, ob[i]);
1044
1045         kfree(new_bucket_seq);
1046         kfree(new_buckets);
1047         kfree(ob);
1048         kfree(bu);
1049         return ret;
1050 }
1051
1052 /*
1053  * Allocate more journal space at runtime - not currently making use if it, but
1054  * the code works:
1055  */
1056 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
1057                                 unsigned nr)
1058 {
1059         struct journal_device *ja = &ca->journal;
1060         struct closure cl;
1061         int ret = 0;
1062
1063         closure_init_stack(&cl);
1064
1065         down_write(&c->state_lock);
1066
1067         /* don't handle reducing nr of buckets yet: */
1068         if (nr < ja->nr)
1069                 goto unlock;
1070
1071         while (ja->nr < nr) {
1072                 struct disk_reservation disk_res = { 0, 0, 0 };
1073
1074                 /*
1075                  * note: journal buckets aren't really counted as _sectors_ used yet, so
1076                  * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
1077                  * when space used goes up without a reservation - but we do need the
1078                  * reservation to ensure we'll actually be able to allocate:
1079                  *
1080                  * XXX: that's not right, disk reservations only ensure a
1081                  * filesystem-wide allocation will succeed, this is a device
1082                  * specific allocation - we can hang here:
1083                  */
1084
1085                 ret = bch2_disk_reservation_get(c, &disk_res,
1086                                                 bucket_to_sector(ca, nr - ja->nr), 1, 0);
1087                 if (ret)
1088                         break;
1089
1090                 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
1091
1092                 bch2_disk_reservation_put(c, &disk_res);
1093
1094                 closure_sync(&cl);
1095
1096                 if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
1097                         break;
1098         }
1099
1100         bch_err_fn(c, ret);
1101 unlock:
1102         up_write(&c->state_lock);
1103         return ret;
1104 }
1105
1106 int bch2_dev_journal_alloc(struct bch_dev *ca)
1107 {
1108         unsigned nr;
1109         int ret;
1110
1111         if (dynamic_fault("bcachefs:add:journal_alloc")) {
1112                 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
1113                 goto err;
1114         }
1115
1116         /* 1/128th of the device by default: */
1117         nr = ca->mi.nbuckets >> 7;
1118
1119         /*
1120          * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
1121          * is smaller:
1122          */
1123         nr = clamp_t(unsigned, nr,
1124                      BCH_JOURNAL_BUCKETS_MIN,
1125                      min(1 << 13,
1126                          (1 << 24) / ca->mi.bucket_size));
1127
1128         ret = __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
1129 err:
1130         bch_err_fn(ca, ret);
1131         return ret;
1132 }
1133
1134 int bch2_fs_journal_alloc(struct bch_fs *c)
1135 {
1136         for_each_online_member(c, ca) {
1137                 if (ca->journal.nr)
1138                         continue;
1139
1140                 int ret = bch2_dev_journal_alloc(ca);
1141                 if (ret) {
1142                         percpu_ref_put(&ca->io_ref);
1143                         return ret;
1144                 }
1145         }
1146
1147         return 0;
1148 }
1149
1150 /* startup/shutdown: */
1151
1152 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1153 {
1154         bool ret = false;
1155         u64 seq;
1156
1157         spin_lock(&j->lock);
1158         for (seq = journal_last_unwritten_seq(j);
1159              seq <= journal_cur_seq(j) && !ret;
1160              seq++) {
1161                 struct journal_buf *buf = journal_seq_to_buf(j, seq);
1162
1163                 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1164                         ret = true;
1165         }
1166         spin_unlock(&j->lock);
1167
1168         return ret;
1169 }
1170
1171 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1172 {
1173         wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1174 }
1175
1176 void bch2_fs_journal_stop(struct journal *j)
1177 {
1178         bch2_journal_reclaim_stop(j);
1179         bch2_journal_flush_all_pins(j);
1180
1181         wait_event(j->wait, bch2_journal_entry_close(j));
1182
1183         /*
1184          * Always write a new journal entry, to make sure the clock hands are up
1185          * to date (and match the superblock)
1186          */
1187         bch2_journal_meta(j);
1188
1189         journal_quiesce(j);
1190
1191         BUG_ON(!bch2_journal_error(j) &&
1192                test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
1193                j->last_empty_seq != journal_cur_seq(j));
1194
1195         cancel_delayed_work_sync(&j->write_work);
1196 }
1197
1198 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1199 {
1200         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1201         struct journal_entry_pin_list *p;
1202         struct journal_replay *i, **_i;
1203         struct genradix_iter iter;
1204         bool had_entries = false;
1205         u64 last_seq = cur_seq, nr, seq;
1206
1207         genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1208                 i = *_i;
1209
1210                 if (journal_replay_ignore(i))
1211                         continue;
1212
1213                 last_seq = le64_to_cpu(i->j.last_seq);
1214                 break;
1215         }
1216
1217         nr = cur_seq - last_seq;
1218
1219         if (nr + 1 > j->pin.size) {
1220                 free_fifo(&j->pin);
1221                 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1222                 if (!j->pin.data) {
1223                         bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1224                         return -BCH_ERR_ENOMEM_journal_pin_fifo;
1225                 }
1226         }
1227
1228         j->replay_journal_seq   = last_seq;
1229         j->replay_journal_seq_end = cur_seq;
1230         j->last_seq_ondisk      = last_seq;
1231         j->flushed_seq_ondisk   = cur_seq - 1;
1232         j->seq_ondisk           = cur_seq - 1;
1233         j->pin.front            = last_seq;
1234         j->pin.back             = cur_seq;
1235         atomic64_set(&j->seq, cur_seq - 1);
1236
1237         fifo_for_each_entry_ptr(p, &j->pin, seq)
1238                 journal_pin_list_init(p, 1);
1239
1240         genradix_for_each(&c->journal_entries, iter, _i) {
1241                 i = *_i;
1242
1243                 if (journal_replay_ignore(i))
1244                         continue;
1245
1246                 seq = le64_to_cpu(i->j.seq);
1247                 BUG_ON(seq >= cur_seq);
1248
1249                 if (seq < last_seq)
1250                         continue;
1251
1252                 if (journal_entry_empty(&i->j))
1253                         j->last_empty_seq = le64_to_cpu(i->j.seq);
1254
1255                 p = journal_seq_pin(j, seq);
1256
1257                 p->devs.nr = 0;
1258                 darray_for_each(i->ptrs, ptr)
1259                         bch2_dev_list_add_dev(&p->devs, ptr->dev);
1260
1261                 had_entries = true;
1262         }
1263
1264         if (!had_entries)
1265                 j->last_empty_seq = cur_seq;
1266
1267         spin_lock(&j->lock);
1268
1269         set_bit(JOURNAL_STARTED, &j->flags);
1270         j->last_flush_write = jiffies;
1271
1272         j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1273         j->reservations.unwritten_idx++;
1274
1275         c->last_bucket_seq_cleanup = journal_cur_seq(j);
1276
1277         bch2_journal_space_available(j);
1278         spin_unlock(&j->lock);
1279
1280         return bch2_journal_reclaim_start(j);
1281 }
1282
1283 /* init/exit: */
1284
1285 void bch2_dev_journal_exit(struct bch_dev *ca)
1286 {
1287         struct journal_device *ja = &ca->journal;
1288
1289         for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
1290                 kfree(ja->bio[i]);
1291                 ja->bio[i] = NULL;
1292         }
1293
1294         kfree(ja->buckets);
1295         kfree(ja->bucket_seq);
1296         ja->buckets     = NULL;
1297         ja->bucket_seq  = NULL;
1298 }
1299
1300 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1301 {
1302         struct journal_device *ja = &ca->journal;
1303         struct bch_sb_field_journal *journal_buckets =
1304                 bch2_sb_field_get(sb, journal);
1305         struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1306                 bch2_sb_field_get(sb, journal_v2);
1307
1308         ja->nr = 0;
1309
1310         if (journal_buckets_v2) {
1311                 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1312
1313                 for (unsigned i = 0; i < nr; i++)
1314                         ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1315         } else if (journal_buckets) {
1316                 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1317         }
1318
1319         ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1320         if (!ja->bucket_seq)
1321                 return -BCH_ERR_ENOMEM_dev_journal_init;
1322
1323         unsigned nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1324
1325         for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
1326                 ja->bio[i] = kmalloc(struct_size(ja->bio[i], bio.bi_inline_vecs,
1327                                      nr_bvecs), GFP_KERNEL);
1328                 if (!ja->bio[i])
1329                         return -BCH_ERR_ENOMEM_dev_journal_init;
1330
1331                 ja->bio[i]->ca = ca;
1332                 ja->bio[i]->buf_idx = i;
1333                 bio_init(&ja->bio[i]->bio, NULL, ja->bio[i]->bio.bi_inline_vecs, nr_bvecs, 0);
1334         }
1335
1336         ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1337         if (!ja->buckets)
1338                 return -BCH_ERR_ENOMEM_dev_journal_init;
1339
1340         if (journal_buckets_v2) {
1341                 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1342                 unsigned dst = 0;
1343
1344                 for (unsigned i = 0; i < nr; i++)
1345                         for (unsigned j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1346                                 ja->buckets[dst++] =
1347                                         le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1348         } else if (journal_buckets) {
1349                 for (unsigned i = 0; i < ja->nr; i++)
1350                         ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1351         }
1352
1353         return 0;
1354 }
1355
1356 void bch2_fs_journal_exit(struct journal *j)
1357 {
1358         if (j->wq)
1359                 destroy_workqueue(j->wq);
1360
1361         darray_exit(&j->early_journal_entries);
1362
1363         for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++)
1364                 kvfree(j->buf[i].data);
1365         free_fifo(&j->pin);
1366 }
1367
1368 int bch2_fs_journal_init(struct journal *j)
1369 {
1370         static struct lock_class_key res_key;
1371
1372         mutex_init(&j->buf_lock);
1373         spin_lock_init(&j->lock);
1374         spin_lock_init(&j->err_lock);
1375         init_waitqueue_head(&j->wait);
1376         INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1377         init_waitqueue_head(&j->reclaim_wait);
1378         init_waitqueue_head(&j->pin_flush_wait);
1379         mutex_init(&j->reclaim_lock);
1380         mutex_init(&j->discard_lock);
1381
1382         lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1383
1384         atomic64_set(&j->reservations.counter,
1385                 ((union journal_res_state)
1386                  { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1387
1388         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)))
1389                 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1390
1391         for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++) {
1392                 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1393                 j->buf[i].data = kvmalloc(j->buf[i].buf_size, GFP_KERNEL);
1394                 if (!j->buf[i].data)
1395                         return -BCH_ERR_ENOMEM_journal_buf;
1396                 j->buf[i].idx = i;
1397         }
1398
1399         j->pin.front = j->pin.back = 1;
1400
1401         j->wq = alloc_workqueue("bcachefs_journal",
1402                                 WQ_HIGHPRI|WQ_FREEZABLE|WQ_UNBOUND|WQ_MEM_RECLAIM, 512);
1403         if (!j->wq)
1404                 return -BCH_ERR_ENOMEM_fs_other_alloc;
1405         return 0;
1406 }
1407
1408 /* debug: */
1409
1410 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1411 {
1412         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1413         union journal_res_state s;
1414         unsigned long now = jiffies;
1415         u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes;
1416
1417         if (!out->nr_tabstops)
1418                 printbuf_tabstop_push(out, 24);
1419         out->atomic++;
1420
1421         rcu_read_lock();
1422         s = READ_ONCE(j->reservations);
1423
1424         prt_printf(out, "dirty journal entries:\t%llu/%llu\n",  fifo_used(&j->pin), j->pin.size);
1425         prt_printf(out, "seq:\t\t\t%llu\n",                     journal_cur_seq(j));
1426         prt_printf(out, "seq_ondisk:\t\t%llu\n",                j->seq_ondisk);
1427         prt_printf(out, "last_seq:\t\t%llu\n",                  journal_last_seq(j));
1428         prt_printf(out, "last_seq_ondisk:\t%llu\n",             j->last_seq_ondisk);
1429         prt_printf(out, "flushed_seq_ondisk:\t%llu\n",          j->flushed_seq_ondisk);
1430         prt_printf(out, "watermark:\t\t%s\n",                   bch2_watermarks[j->watermark]);
1431         prt_printf(out, "each entry reserved:\t%u\n",           j->entry_u64s_reserved);
1432         prt_printf(out, "nr flush writes:\t%llu\n",             j->nr_flush_writes);
1433         prt_printf(out, "nr noflush writes:\t%llu\n",           j->nr_noflush_writes);
1434         prt_printf(out, "average write size:\t");
1435         prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0);
1436         prt_newline(out);
1437         prt_printf(out, "nr direct reclaim:\t%llu\n",           j->nr_direct_reclaim);
1438         prt_printf(out, "nr background reclaim:\t%llu\n",       j->nr_background_reclaim);
1439         prt_printf(out, "reclaim kicked:\t\t%u\n",              j->reclaim_kicked);
1440         prt_printf(out, "reclaim runs in:\t%u ms\n",            time_after(j->next_reclaim, now)
1441                ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1442         prt_printf(out, "blocked:\t\t%u\n",                     j->blocked);
1443         prt_printf(out, "current entry sectors:\t%u\n",         j->cur_entry_sectors);
1444         prt_printf(out, "current entry error:\t%s\n",           bch2_journal_errors[j->cur_entry_error]);
1445         prt_printf(out, "current entry:\t\t");
1446
1447         switch (s.cur_entry_offset) {
1448         case JOURNAL_ENTRY_ERROR_VAL:
1449                 prt_printf(out, "error");
1450                 break;
1451         case JOURNAL_ENTRY_CLOSED_VAL:
1452                 prt_printf(out, "closed");
1453                 break;
1454         default:
1455                 prt_printf(out, "%u/%u", s.cur_entry_offset, j->cur_entry_u64s);
1456                 break;
1457         }
1458
1459         prt_newline(out);
1460         prt_printf(out, "unwritten entries:");
1461         prt_newline(out);
1462         bch2_journal_bufs_to_text(out, j);
1463
1464         prt_printf(out,
1465                "replay done:\t\t%i\n",
1466                test_bit(JOURNAL_REPLAY_DONE,    &j->flags));
1467
1468         prt_printf(out, "space:\n");
1469         prt_printf(out, "\tdiscarded\t%u:%u\n",
1470                j->space[journal_space_discarded].next_entry,
1471                j->space[journal_space_discarded].total);
1472         prt_printf(out, "\tclean ondisk\t%u:%u\n",
1473                j->space[journal_space_clean_ondisk].next_entry,
1474                j->space[journal_space_clean_ondisk].total);
1475         prt_printf(out, "\tclean\t\t%u:%u\n",
1476                j->space[journal_space_clean].next_entry,
1477                j->space[journal_space_clean].total);
1478         prt_printf(out, "\ttotal\t\t%u:%u\n",
1479                j->space[journal_space_total].next_entry,
1480                j->space[journal_space_total].total);
1481
1482         for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
1483                 struct journal_device *ja = &ca->journal;
1484
1485                 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1486                         continue;
1487
1488                 if (!ja->nr)
1489                         continue;
1490
1491                 prt_printf(out, "dev %u:\n",            ca->dev_idx);
1492                 prt_printf(out, "\tnr\t\t%u\n",         ja->nr);
1493                 prt_printf(out, "\tbucket size\t%u\n",  ca->mi.bucket_size);
1494                 prt_printf(out, "\tavailable\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1495                 prt_printf(out, "\tdiscard_idx\t%u\n",  ja->discard_idx);
1496                 prt_printf(out, "\tdirty_ondisk\t%u (seq %llu)\n", ja->dirty_idx_ondisk,        ja->bucket_seq[ja->dirty_idx_ondisk]);
1497                 prt_printf(out, "\tdirty_idx\t%u (seq %llu)\n", ja->dirty_idx,          ja->bucket_seq[ja->dirty_idx]);
1498                 prt_printf(out, "\tcur_idx\t\t%u (seq %llu)\n", ja->cur_idx,            ja->bucket_seq[ja->cur_idx]);
1499         }
1500
1501         rcu_read_unlock();
1502
1503         --out->atomic;
1504 }
1505
1506 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1507 {
1508         spin_lock(&j->lock);
1509         __bch2_journal_debug_to_text(out, j);
1510         spin_unlock(&j->lock);
1511 }
1512
1513 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1514 {
1515         struct journal_entry_pin_list *pin_list;
1516         struct journal_entry_pin *pin;
1517
1518         spin_lock(&j->lock);
1519         *seq = max(*seq, j->pin.front);
1520
1521         if (*seq >= j->pin.back) {
1522                 spin_unlock(&j->lock);
1523                 return true;
1524         }
1525
1526         out->atomic++;
1527
1528         pin_list = journal_seq_pin(j, *seq);
1529
1530         prt_printf(out, "%llu: count %u", *seq, atomic_read(&pin_list->count));
1531         prt_newline(out);
1532         printbuf_indent_add(out, 2);
1533
1534         for (unsigned i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1535                 list_for_each_entry(pin, &pin_list->list[i], list) {
1536                         prt_printf(out, "\t%px %ps", pin, pin->flush);
1537                         prt_newline(out);
1538                 }
1539
1540         if (!list_empty(&pin_list->flushed)) {
1541                 prt_printf(out, "flushed:");
1542                 prt_newline(out);
1543         }
1544
1545         list_for_each_entry(pin, &pin_list->flushed, list) {
1546                 prt_printf(out, "\t%px %ps", pin, pin->flush);
1547                 prt_newline(out);
1548         }
1549
1550         printbuf_indent_sub(out, 2);
1551
1552         --out->atomic;
1553         spin_unlock(&j->lock);
1554
1555         return false;
1556 }
1557
1558 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1559 {
1560         u64 seq = 0;
1561
1562         while (!bch2_journal_seq_pins_to_text(out, j, &seq))
1563                 seq++;
1564 }