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