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 void bch2_journal_buf_to_text(struct printbuf *out, struct journal *j, u64 seq)
  31 {
  32         union journal_res_state s = READ_ONCE(j->reservations);
  33         unsigned i = seq & JOURNAL_BUF_MASK;
  34         struct journal_buf *buf = j->buf + i;
  35
  36         prt_printf(out, "seq:");
  37         prt_tab(out);
  38         prt_printf(out, "%llu", seq);
  39         prt_newline(out);
  40         printbuf_indent_add(out, 2);
  41
  42         prt_printf(out, "refcount:");
  43         prt_tab(out);
  44         prt_printf(out, "%u", journal_state_count(s, i));
  45         prt_newline(out);
  46
  47         prt_printf(out, "size:");
  48         prt_tab(out);
  49         prt_human_readable_u64(out, vstruct_bytes(buf->data));
  50         prt_newline(out);
  51
  52         prt_printf(out, "expires");
  53         prt_tab(out);
  54         prt_printf(out, "%li jiffies", buf->expires - jiffies);
  55         prt_newline(out);
  56
  57         printbuf_indent_sub(out, 2);
  58 }
  59
  60 static void bch2_journal_bufs_to_text(struct printbuf *out, struct journal *j)
  61 {
  62         if (!out->nr_tabstops)
  63                 printbuf_tabstop_push(out, 24);
  64
  65         for (u64 seq = journal_last_unwritten_seq(j);
  66              seq <= journal_cur_seq(j);
  67              seq++)
  68                 bch2_journal_buf_to_text(out, j, seq);
  69 }
  70
  71 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
  72 {
  73         return seq > j->seq_ondisk;
  74 }
  75
  76 static bool __journal_entry_is_open(union journal_res_state state)
  77 {
  78         return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
  79 }
  80
  81 static inline unsigned nr_unwritten_journal_entries(struct journal *j)
  82 {
  83         return atomic64_read(&j->seq) - j->seq_ondisk;
  84 }
  85
  86 static bool journal_entry_is_open(struct journal *j)
  87 {
  88         return __journal_entry_is_open(j->reservations);
  89 }
  90
  91 static inline struct journal_buf *
  92 journal_seq_to_buf(struct journal *j, u64 seq)
  93 {
  94         struct journal_buf *buf = NULL;
  95
  96         EBUG_ON(seq > journal_cur_seq(j));
  97
  98         if (journal_seq_unwritten(j, seq)) {
  99                 buf = j->buf + (seq & JOURNAL_BUF_MASK);
 100                 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
 101         }
 102         return buf;
 103 }
 104
 105 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
 106 {
 107         unsigned i;
 108
 109         for (i = 0; i < ARRAY_SIZE(p->list); i++)
 110                 INIT_LIST_HEAD(&p->list[i]);
 111         INIT_LIST_HEAD(&p->flushed);
 112         atomic_set(&p->count, count);
 113         p->devs.nr = 0;
 114 }
 115
 116 /*
 117  * Detect stuck journal conditions and trigger shutdown. Technically the journal
 118  * can end up stuck for a variety of reasons, such as a blocked I/O, journal
 119  * reservation lockup, etc. Since this is a fatal error with potentially
 120  * unpredictable characteristics, we want to be fairly conservative before we
 121  * decide to shut things down.
 122  *
 123  * Consider the journal stuck when it appears full with no ability to commit
 124  * btree transactions, to discard journal buckets, nor acquire priority
 125  * (reserved watermark) reservation.
 126  */
 127 static inline bool
 128 journal_error_check_stuck(struct journal *j, int error, unsigned flags)
 129 {
 130         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 131         bool stuck = false;
 132         struct printbuf buf = PRINTBUF;
 133
 134         if (!(error == JOURNAL_ERR_journal_full ||
 135               error == JOURNAL_ERR_journal_pin_full) ||
 136             nr_unwritten_journal_entries(j) ||
 137             (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim)
 138                 return stuck;
 139
 140         spin_lock(&j->lock);
 141
 142         if (j->can_discard) {
 143                 spin_unlock(&j->lock);
 144                 return stuck;
 145         }
 146
 147         stuck = true;
 148
 149         /*
 150          * The journal shutdown path will set ->err_seq, but do it here first to
 151          * serialize against concurrent failures and avoid duplicate error
 152          * reports.
 153          */
 154         if (j->err_seq) {
 155                 spin_unlock(&j->lock);
 156                 return stuck;
 157         }
 158         j->err_seq = journal_cur_seq(j);
 159         spin_unlock(&j->lock);
 160
 161         bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)",
 162                 bch2_journal_errors[error]);
 163         bch2_journal_debug_to_text(&buf, j);
 164         bch_err(c, "%s", buf.buf);
 165
 166         printbuf_reset(&buf);
 167         bch2_journal_pins_to_text(&buf, j);
 168         bch_err(c, "Journal pins:\n%s", buf.buf);
 169         printbuf_exit(&buf);
 170
 171         bch2_fatal_error(c);
 172         dump_stack();
 173
 174         return stuck;
 175 }
 176
 177 /*
 178  * Final processing when the last reference of a journal buffer has been
 179  * dropped. Drop the pin list reference acquired at journal entry open and write
 180  * the buffer, if requested.
 181  */
 182 void bch2_journal_buf_put_final(struct journal *j, u64 seq, bool write)
 183 {
 184         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 185
 186         lockdep_assert_held(&j->lock);
 187
 188         if (__bch2_journal_pin_put(j, seq))
 189                 bch2_journal_reclaim_fast(j);
 190         if (write)
 191                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
 192 }
 193
 194 /*
 195  * Returns true if journal entry is now closed:
 196  *
 197  * We don't close a journal_buf until the next journal_buf is finished writing,
 198  * and can be opened again - this also initializes the next journal_buf:
 199  */
 200 static void __journal_entry_close(struct journal *j, unsigned closed_val, bool trace)
 201 {
 202         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 203         struct journal_buf *buf = journal_cur_buf(j);
 204         union journal_res_state old, new;
 205         u64 v = atomic64_read(&j->reservations.counter);
 206         unsigned sectors;
 207
 208         BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
 209                closed_val != JOURNAL_ENTRY_ERROR_VAL);
 210
 211         lockdep_assert_held(&j->lock);
 212
 213         do {
 214                 old.v = new.v = v;
 215                 new.cur_entry_offset = closed_val;
 216
 217                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
 218                     old.cur_entry_offset == new.cur_entry_offset)
 219                         return;
 220         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 221                                        old.v, new.v)) != old.v);
 222
 223         if (!__journal_entry_is_open(old))
 224                 return;
 225
 226         /* Close out old buffer: */
 227         buf->data->u64s         = cpu_to_le32(old.cur_entry_offset);
 228
 229         if (trace_journal_entry_close_enabled() && trace) {
 230                 struct printbuf pbuf = PRINTBUF;
 231                 pbuf.atomic++;
 232
 233                 prt_str(&pbuf, "entry size: ");
 234                 prt_human_readable_u64(&pbuf, vstruct_bytes(buf->data));
 235                 prt_newline(&pbuf);
 236                 bch2_prt_task_backtrace(&pbuf, current, 1, GFP_NOWAIT);
 237                 trace_journal_entry_close(c, pbuf.buf);
 238                 printbuf_exit(&pbuf);
 239         }
 240
 241         sectors = vstruct_blocks_plus(buf->data, c->block_bits,
 242                                       buf->u64s_reserved) << c->block_bits;
 243         BUG_ON(sectors > buf->sectors);
 244         buf->sectors = sectors;
 245
 246         /*
 247          * We have to set last_seq here, _before_ opening a new journal entry:
 248          *
 249          * A threads may replace an old pin with a new pin on their current
 250          * journal reservation - the expectation being that the journal will
 251          * contain either what the old pin protected or what the new pin
 252          * protects.
 253          *
 254          * After the old pin is dropped journal_last_seq() won't include the old
 255          * pin, so we can only write the updated last_seq on the entry that
 256          * contains whatever the new pin protects.
 257          *
 258          * Restated, we can _not_ update last_seq for a given entry if there
 259          * could be a newer entry open with reservations/pins that have been
 260          * taken against it.
 261          *
 262          * Hence, we want update/set last_seq on the current journal entry right
 263          * before we open a new one:
 264          */
 265         buf->last_seq           = journal_last_seq(j);
 266         buf->data->last_seq     = cpu_to_le64(buf->last_seq);
 267         BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq));
 268
 269         cancel_delayed_work(&j->write_work);
 270
 271         bch2_journal_space_available(j);
 272
 273         __bch2_journal_buf_put(j, old.idx, le64_to_cpu(buf->data->seq));
 274 }
 275
 276 void bch2_journal_halt(struct journal *j)
 277 {
 278         spin_lock(&j->lock);
 279         __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL, true);
 280         if (!j->err_seq)
 281                 j->err_seq = journal_cur_seq(j);
 282         journal_wake(j);
 283         spin_unlock(&j->lock);
 284 }
 285
 286 static bool journal_entry_want_write(struct journal *j)
 287 {
 288         bool ret = !journal_entry_is_open(j) ||
 289                 journal_cur_seq(j) == journal_last_unwritten_seq(j);
 290
 291         /* Don't close it yet if we already have a write in flight: */
 292         if (ret)
 293                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 294         else if (nr_unwritten_journal_entries(j)) {
 295                 struct journal_buf *buf = journal_cur_buf(j);
 296
 297                 if (!buf->flush_time) {
 298                         buf->flush_time = local_clock() ?: 1;
 299                         buf->expires = jiffies;
 300                 }
 301         }
 302
 303         return ret;
 304 }
 305
 306 bool bch2_journal_entry_close(struct journal *j)
 307 {
 308         bool ret;
 309
 310         spin_lock(&j->lock);
 311         ret = journal_entry_want_write(j);
 312         spin_unlock(&j->lock);
 313
 314         return ret;
 315 }
 316
 317 /*
 318  * should _only_ called from journal_res_get() - when we actually want a
 319  * journal reservation - journal entry is open means journal is dirty:
 320  */
 321 static int journal_entry_open(struct journal *j)
 322 {
 323         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 324         struct journal_buf *buf = j->buf +
 325                 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK);
 326         union journal_res_state old, new;
 327         int u64s;
 328         u64 v;
 329
 330         lockdep_assert_held(&j->lock);
 331         BUG_ON(journal_entry_is_open(j));
 332         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
 333
 334         if (j->blocked)
 335                 return JOURNAL_ERR_blocked;
 336
 337         if (j->cur_entry_error)
 338                 return j->cur_entry_error;
 339
 340         if (bch2_journal_error(j))
 341                 return JOURNAL_ERR_insufficient_devices; /* -EROFS */
 342
 343         if (!fifo_free(&j->pin))
 344                 return JOURNAL_ERR_journal_pin_full;
 345
 346         if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
 347                 return JOURNAL_ERR_max_in_flight;
 348
 349         BUG_ON(!j->cur_entry_sectors);
 350
 351         buf->expires            =
 352                 (journal_cur_seq(j) == j->flushed_seq_ondisk
 353                  ? jiffies
 354                  : j->last_flush_write) +
 355                 msecs_to_jiffies(c->opts.journal_flush_delay);
 356
 357         buf->u64s_reserved      = j->entry_u64s_reserved;
 358         buf->disk_sectors       = j->cur_entry_sectors;
 359         buf->sectors            = min(buf->disk_sectors, buf->buf_size >> 9);
 360
 361         u64s = (int) (buf->sectors << 9) / sizeof(u64) -
 362                 journal_entry_overhead(j);
 363         u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
 364
 365         if (u64s <= (ssize_t) j->early_journal_entries.nr)
 366                 return JOURNAL_ERR_journal_full;
 367
 368         if (fifo_empty(&j->pin) && j->reclaim_thread)
 369                 wake_up_process(j->reclaim_thread);
 370
 371         /*
 372          * The fifo_push() needs to happen at the same time as j->seq is
 373          * incremented for journal_last_seq() to be calculated correctly
 374          */
 375         atomic64_inc(&j->seq);
 376         journal_pin_list_init(fifo_push_ref(&j->pin), 1);
 377
 378         BUG_ON(j->pin.back - 1 != atomic64_read(&j->seq));
 379
 380         BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf);
 381
 382         bkey_extent_init(&buf->key);
 383         buf->noflush    = false;
 384         buf->must_flush = false;
 385         buf->separate_flush = false;
 386         buf->flush_time = 0;
 387         buf->need_flush_to_write_buffer = true;
 388
 389         memset(buf->data, 0, sizeof(*buf->data));
 390         buf->data->seq  = cpu_to_le64(journal_cur_seq(j));
 391         buf->data->u64s = 0;
 392
 393         if (j->early_journal_entries.nr) {
 394                 memcpy(buf->data->_data, j->early_journal_entries.data,
 395                        j->early_journal_entries.nr * sizeof(u64));
 396                 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr);
 397         }
 398
 399         /*
 400          * Must be set before marking the journal entry as open:
 401          */
 402         j->cur_entry_u64s = u64s;
 403
 404         v = atomic64_read(&j->reservations.counter);
 405         do {
 406                 old.v = new.v = v;
 407
 408                 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL);
 409
 410                 new.idx++;
 411                 BUG_ON(journal_state_count(new, new.idx));
 412                 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK));
 413
 414                 journal_state_inc(&new);
 415
 416                 /* Handle any already added entries */
 417                 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
 418         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 419                                        old.v, new.v)) != old.v);
 420
 421         mod_delayed_work(c->io_complete_wq,
 422                          &j->write_work,
 423                          msecs_to_jiffies(c->opts.journal_flush_delay));
 424         journal_wake(j);
 425
 426         if (j->early_journal_entries.nr)
 427                 darray_exit(&j->early_journal_entries);
 428         return 0;
 429 }
 430
 431 static bool journal_quiesced(struct journal *j)
 432 {
 433         bool ret = atomic64_read(&j->seq) == j->seq_ondisk;
 434
 435         if (!ret)
 436                 bch2_journal_entry_close(j);
 437         return ret;
 438 }
 439
 440 static void journal_quiesce(struct journal *j)
 441 {
 442         wait_event(j->wait, journal_quiesced(j));
 443 }
 444
 445 static void journal_write_work(struct work_struct *work)
 446 {
 447         struct journal *j = container_of(work, struct journal, write_work.work);
 448         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 449         long delta;
 450
 451         spin_lock(&j->lock);
 452         if (!__journal_entry_is_open(j->reservations))
 453                 goto unlock;
 454
 455         delta = journal_cur_buf(j)->expires - jiffies;
 456
 457         if (delta > 0)
 458                 mod_delayed_work(c->io_complete_wq, &j->write_work, delta);
 459         else
 460                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 461 unlock:
 462         spin_unlock(&j->lock);
 463 }
 464
 465 static int __journal_res_get(struct journal *j, struct journal_res *res,
 466                              unsigned flags)
 467 {
 468         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 469         struct journal_buf *buf;
 470         bool can_discard;
 471         int ret;
 472 retry:
 473         if (journal_res_get_fast(j, res, flags))
 474                 return 0;
 475
 476         if (bch2_journal_error(j))
 477                 return -BCH_ERR_erofs_journal_err;
 478
 479         spin_lock(&j->lock);
 480
 481         /* check once more in case somebody else shut things down... */
 482         if (bch2_journal_error(j)) {
 483                 spin_unlock(&j->lock);
 484                 return -BCH_ERR_erofs_journal_err;
 485         }
 486
 487         /*
 488          * Recheck after taking the lock, so we don't race with another thread
 489          * that just did journal_entry_open() and call bch2_journal_entry_close()
 490          * unnecessarily
 491          */
 492         if (journal_res_get_fast(j, res, flags)) {
 493                 spin_unlock(&j->lock);
 494                 return 0;
 495         }
 496
 497         if ((flags & BCH_WATERMARK_MASK) < j->watermark) {
 498                 /*
 499                  * Don't want to close current journal entry, just need to
 500                  * invoke reclaim:
 501                  */
 502                 ret = JOURNAL_ERR_journal_full;
 503                 goto unlock;
 504         }
 505
 506         /*
 507          * If we couldn't get a reservation because the current buf filled up,
 508          * and we had room for a bigger entry on disk, signal that we want to
 509          * realloc the journal bufs:
 510          */
 511         buf = journal_cur_buf(j);
 512         if (journal_entry_is_open(j) &&
 513             buf->buf_size >> 9 < buf->disk_sectors &&
 514             buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
 515                 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
 516
 517         __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, false);
 518         ret = journal_entry_open(j);
 519
 520         if (ret == JOURNAL_ERR_max_in_flight) {
 521                 track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight],
 522                                    &j->max_in_flight_start, true);
 523                 if (trace_journal_entry_full_enabled()) {
 524                         struct printbuf buf = PRINTBUF;
 525                         buf.atomic++;
 526
 527                         bch2_journal_bufs_to_text(&buf, j);
 528                         trace_journal_entry_full(c, buf.buf);
 529                         printbuf_exit(&buf);
 530                 }
 531                 count_event(c, journal_entry_full);
 532         }
 533 unlock:
 534         can_discard = j->can_discard;
 535         spin_unlock(&j->lock);
 536
 537         if (!ret)
 538                 goto retry;
 539         if (journal_error_check_stuck(j, ret, flags))
 540                 ret = -BCH_ERR_journal_res_get_blocked;
 541
 542         /*
 543          * Journal is full - can't rely on reclaim from work item due to
 544          * freezing:
 545          */
 546         if ((ret == JOURNAL_ERR_journal_full ||
 547              ret == JOURNAL_ERR_journal_pin_full) &&
 548             !(flags & JOURNAL_RES_GET_NONBLOCK)) {
 549                 if (can_discard) {
 550                         bch2_journal_do_discards(j);
 551                         goto retry;
 552                 }
 553
 554                 if (mutex_trylock(&j->reclaim_lock)) {
 555                         bch2_journal_reclaim(j);
 556                         mutex_unlock(&j->reclaim_lock);
 557                 }
 558         }
 559
 560         return ret == JOURNAL_ERR_insufficient_devices
 561                 ? -BCH_ERR_erofs_journal_err
 562                 : -BCH_ERR_journal_res_get_blocked;
 563 }
 564
 565 /*
 566  * Essentially the entry function to the journaling code. When bcachefs is doing
 567  * a btree insert, it calls this function to get the current journal write.
 568  * Journal write is the structure used set up journal writes. The calling
 569  * function will then add its keys to the structure, queuing them for the next
 570  * write.
 571  *
 572  * To ensure forward progress, the current task must not be holding any
 573  * btree node write locks.
 574  */
 575 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
 576                                   unsigned flags)
 577 {
 578         int ret;
 579
 580         closure_wait_event(&j->async_wait,
 581                    (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked ||
 582                    (flags & JOURNAL_RES_GET_NONBLOCK));
 583         return ret;
 584 }
 585
 586 /* journal_entry_res: */
 587
 588 void bch2_journal_entry_res_resize(struct journal *j,
 589                                    struct journal_entry_res *res,
 590                                    unsigned new_u64s)
 591 {
 592         union journal_res_state state;
 593         int d = new_u64s - res->u64s;
 594
 595         spin_lock(&j->lock);
 596
 597         j->entry_u64s_reserved += d;
 598         if (d <= 0)
 599                 goto out;
 600
 601         j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
 602         smp_mb();
 603         state = READ_ONCE(j->reservations);
 604
 605         if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
 606             state.cur_entry_offset > j->cur_entry_u64s) {
 607                 j->cur_entry_u64s += d;
 608                 /*
 609                  * Not enough room in current journal entry, have to flush it:
 610                  */
 611                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 612         } else {
 613                 journal_cur_buf(j)->u64s_reserved += d;
 614         }
 615 out:
 616         spin_unlock(&j->lock);
 617         res->u64s += d;
 618 }
 619
 620 /* journal flushing: */
 621
 622 /**
 623  * bch2_journal_flush_seq_async - wait for a journal entry to be written
 624  * @j:          journal object
 625  * @seq:        seq to flush
 626  * @parent:     closure object to wait with
 627  * Returns:     1 if @seq has already been flushed, 0 if @seq is being flushed,
 628  *              -EIO if @seq will never be flushed
 629  *
 630  * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if
 631  * necessary
 632  */
 633 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
 634                                  struct closure *parent)
 635 {
 636         struct journal_buf *buf;
 637         int ret = 0;
 638
 639         if (seq <= j->flushed_seq_ondisk)
 640                 return 1;
 641
 642         spin_lock(&j->lock);
 643
 644         if (WARN_ONCE(seq > journal_cur_seq(j),
 645                       "requested to flush journal seq %llu, but currently at %llu",
 646                       seq, journal_cur_seq(j)))
 647                 goto out;
 648
 649         /* Recheck under lock: */
 650         if (j->err_seq && seq >= j->err_seq) {
 651                 ret = -EIO;
 652                 goto out;
 653         }
 654
 655         if (seq <= j->flushed_seq_ondisk) {
 656                 ret = 1;
 657                 goto out;
 658         }
 659
 660         /* if seq was written, but not flushed - flush a newer one instead */
 661         seq = max(seq, journal_last_unwritten_seq(j));
 662
 663 recheck_need_open:
 664         if (seq > journal_cur_seq(j)) {
 665                 struct journal_res res = { 0 };
 666
 667                 if (journal_entry_is_open(j))
 668                         __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 669
 670                 spin_unlock(&j->lock);
 671
 672                 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 673                 if (ret)
 674                         return ret;
 675
 676                 seq = res.seq;
 677                 buf = j->buf + (seq & JOURNAL_BUF_MASK);
 678                 buf->must_flush = true;
 679
 680                 if (!buf->flush_time) {
 681                         buf->flush_time = local_clock() ?: 1;
 682                         buf->expires = jiffies;
 683                 }
 684
 685                 if (parent && !closure_wait(&buf->wait, parent))
 686                         BUG();
 687
 688                 bch2_journal_res_put(j, &res);
 689
 690                 spin_lock(&j->lock);
 691                 goto want_write;
 692         }
 693
 694         /*
 695          * if write was kicked off without a flush, flush the next sequence
 696          * number instead
 697          */
 698         buf = journal_seq_to_buf(j, seq);
 699         if (buf->noflush) {
 700                 seq++;
 701                 goto recheck_need_open;
 702         }
 703
 704         buf->must_flush = true;
 705
 706         if (parent && !closure_wait(&buf->wait, parent))
 707                 BUG();
 708 want_write:
 709         if (seq == journal_cur_seq(j))
 710                 journal_entry_want_write(j);
 711 out:
 712         spin_unlock(&j->lock);
 713         return ret;
 714 }
 715
 716 int bch2_journal_flush_seq(struct journal *j, u64 seq)
 717 {
 718         u64 start_time = local_clock();
 719         int ret, ret2;
 720
 721         /*
 722          * Don't update time_stats when @seq is already flushed:
 723          */
 724         if (seq <= j->flushed_seq_ondisk)
 725                 return 0;
 726
 727         ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
 728
 729         if (!ret)
 730                 bch2_time_stats_update(j->flush_seq_time, start_time);
 731
 732         return ret ?: ret2 < 0 ? ret2 : 0;
 733 }
 734
 735 /*
 736  * bch2_journal_flush_async - if there is an open journal entry, or a journal
 737  * still being written, write it and wait for the write to complete
 738  */
 739 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
 740 {
 741         bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent);
 742 }
 743
 744 int bch2_journal_flush(struct journal *j)
 745 {
 746         return bch2_journal_flush_seq(j, atomic64_read(&j->seq));
 747 }
 748
 749 /*
 750  * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
 751  * @seq
 752  */
 753 bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
 754 {
 755         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 756         u64 unwritten_seq;
 757         bool ret = false;
 758
 759         if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
 760                 return false;
 761
 762         if (seq <= c->journal.flushed_seq_ondisk)
 763                 return false;
 764
 765         spin_lock(&j->lock);
 766         if (seq <= c->journal.flushed_seq_ondisk)
 767                 goto out;
 768
 769         for (unwritten_seq = journal_last_unwritten_seq(j);
 770              unwritten_seq < seq;
 771              unwritten_seq++) {
 772                 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
 773
 774                 /* journal write is already in flight, and was a flush write: */
 775                 if (unwritten_seq == journal_last_unwritten_seq(j) && !buf->noflush)
 776                         goto out;
 777
 778                 buf->noflush = true;
 779         }
 780
 781         ret = true;
 782 out:
 783         spin_unlock(&j->lock);
 784         return ret;
 785 }
 786
 787 int bch2_journal_meta(struct journal *j)
 788 {
 789         struct journal_buf *buf;
 790         struct journal_res res;
 791         int ret;
 792
 793         memset(&res, 0, sizeof(res));
 794
 795         ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 796         if (ret)
 797                 return ret;
 798
 799         buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
 800         buf->must_flush = true;
 801
 802         if (!buf->flush_time) {
 803                 buf->flush_time = local_clock() ?: 1;
 804                 buf->expires = jiffies;
 805         }
 806
 807         bch2_journal_res_put(j, &res);
 808
 809         return bch2_journal_flush_seq(j, res.seq);
 810 }
 811
 812 /* block/unlock the journal: */
 813
 814 void bch2_journal_unblock(struct journal *j)
 815 {
 816         spin_lock(&j->lock);
 817         j->blocked--;
 818         spin_unlock(&j->lock);
 819
 820         journal_wake(j);
 821 }
 822
 823 void bch2_journal_block(struct journal *j)
 824 {
 825         spin_lock(&j->lock);
 826         j->blocked++;
 827         spin_unlock(&j->lock);
 828
 829         journal_quiesce(j);
 830 }
 831
 832 static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
 833 {
 834         struct journal_buf *ret = NULL;
 835
 836         mutex_lock(&j->buf_lock);
 837         spin_lock(&j->lock);
 838         max_seq = min(max_seq, journal_cur_seq(j));
 839
 840         for (u64 seq = journal_last_unwritten_seq(j);
 841              seq <= max_seq;
 842              seq++) {
 843                 unsigned idx = seq & JOURNAL_BUF_MASK;
 844                 struct journal_buf *buf = j->buf + idx;
 845
 846                 if (buf->need_flush_to_write_buffer) {
 847                         if (seq == journal_cur_seq(j))
 848                                 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
 849
 850                         union journal_res_state s;
 851                         s.v = atomic64_read_acquire(&j->reservations.counter);
 852
 853                         ret = journal_state_count(s, idx)
 854                                 ? ERR_PTR(-EAGAIN)
 855                                 : buf;
 856                         break;
 857                 }
 858         }
 859
 860         spin_unlock(&j->lock);
 861         if (IS_ERR_OR_NULL(ret))
 862                 mutex_unlock(&j->buf_lock);
 863         return ret;
 864 }
 865
 866 struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
 867 {
 868         struct journal_buf *ret;
 869
 870         wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j, max_seq)) != ERR_PTR(-EAGAIN));
 871         return ret;
 872 }
 873
 874 /* allocate journal on a device: */
 875
 876 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 877                                          bool new_fs, struct closure *cl)
 878 {
 879         struct bch_fs *c = ca->fs;
 880         struct journal_device *ja = &ca->journal;
 881         u64 *new_bucket_seq = NULL, *new_buckets = NULL;
 882         struct open_bucket **ob = NULL;
 883         long *bu = NULL;
 884         unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr;
 885         int ret = 0;
 886
 887         BUG_ON(nr <= ja->nr);
 888
 889         bu              = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL);
 890         ob              = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL);
 891         new_buckets     = kcalloc(nr, sizeof(u64), GFP_KERNEL);
 892         new_bucket_seq  = kcalloc(nr, sizeof(u64), GFP_KERNEL);
 893         if (!bu || !ob || !new_buckets || !new_bucket_seq) {
 894                 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
 895                 goto err_free;
 896         }
 897
 898         for (nr_got = 0; nr_got < nr_want; nr_got++) {
 899                 if (new_fs) {
 900                         bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
 901                         if (bu[nr_got] < 0) {
 902                                 ret = -BCH_ERR_ENOSPC_bucket_alloc;
 903                                 break;
 904                         }
 905                 } else {
 906                         ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal, cl);
 907                         ret = PTR_ERR_OR_ZERO(ob[nr_got]);
 908                         if (ret)
 909                                 break;
 910
 911                         ret = bch2_trans_run(c,
 912                                 bch2_trans_mark_metadata_bucket(trans, ca,
 913                                                 ob[nr_got]->bucket, BCH_DATA_journal,
 914                                                 ca->mi.bucket_size));
 915                         if (ret) {
 916                                 bch2_open_bucket_put(c, ob[nr_got]);
 917                                 bch_err_msg(c, ret, "marking new journal buckets");
 918                                 break;
 919                         }
 920
 921                         bu[nr_got] = ob[nr_got]->bucket;
 922                 }
 923         }
 924
 925         if (!nr_got)
 926                 goto err_free;
 927
 928         /* Don't return an error if we successfully allocated some buckets: */
 929         ret = 0;
 930
 931         if (c) {
 932                 bch2_journal_flush_all_pins(&c->journal);
 933                 bch2_journal_block(&c->journal);
 934                 mutex_lock(&c->sb_lock);
 935         }
 936
 937         memcpy(new_buckets,     ja->buckets,    ja->nr * sizeof(u64));
 938         memcpy(new_bucket_seq,  ja->bucket_seq, ja->nr * sizeof(u64));
 939
 940         BUG_ON(ja->discard_idx > ja->nr);
 941
 942         pos = ja->discard_idx ?: ja->nr;
 943
 944         memmove(new_buckets + pos + nr_got,
 945                 new_buckets + pos,
 946                 sizeof(new_buckets[0]) * (ja->nr - pos));
 947         memmove(new_bucket_seq + pos + nr_got,
 948                 new_bucket_seq + pos,
 949                 sizeof(new_bucket_seq[0]) * (ja->nr - pos));
 950
 951         for (i = 0; i < nr_got; i++) {
 952                 new_buckets[pos + i] = bu[i];
 953                 new_bucket_seq[pos + i] = 0;
 954         }
 955
 956         nr = ja->nr + nr_got;
 957
 958         ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr);
 959         if (ret)
 960                 goto err_unblock;
 961
 962         if (!new_fs)
 963                 bch2_write_super(c);
 964
 965         /* Commit: */
 966         if (c)
 967                 spin_lock(&c->journal.lock);
 968
 969         swap(new_buckets,       ja->buckets);
 970         swap(new_bucket_seq,    ja->bucket_seq);
 971         ja->nr = nr;
 972
 973         if (pos <= ja->discard_idx)
 974                 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr;
 975         if (pos <= ja->dirty_idx_ondisk)
 976                 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr;
 977         if (pos <= ja->dirty_idx)
 978                 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr;
 979         if (pos <= ja->cur_idx)
 980                 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr;
 981
 982         if (c)
 983                 spin_unlock(&c->journal.lock);
 984 err_unblock:
 985         if (c) {
 986                 bch2_journal_unblock(&c->journal);
 987                 mutex_unlock(&c->sb_lock);
 988         }
 989
 990         if (ret && !new_fs)
 991                 for (i = 0; i < nr_got; i++)
 992                         bch2_trans_run(c,
 993                                 bch2_trans_mark_metadata_bucket(trans, ca,
 994                                                 bu[i], BCH_DATA_free, 0));
 995 err_free:
 996         if (!new_fs)
 997                 for (i = 0; i < nr_got; i++)
 998                         bch2_open_bucket_put(c, ob[i]);
 999
1000         kfree(new_bucket_seq);
1001         kfree(new_buckets);
1002         kfree(ob);
1003         kfree(bu);
1004         return ret;
1005 }
1006
1007 /*
1008  * Allocate more journal space at runtime - not currently making use if it, but
1009  * the code works:
1010  */
1011 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
1012                                 unsigned nr)
1013 {
1014         struct journal_device *ja = &ca->journal;
1015         struct closure cl;
1016         int ret = 0;
1017
1018         closure_init_stack(&cl);
1019
1020         down_write(&c->state_lock);
1021
1022         /* don't handle reducing nr of buckets yet: */
1023         if (nr < ja->nr)
1024                 goto unlock;
1025
1026         while (ja->nr < nr) {
1027                 struct disk_reservation disk_res = { 0, 0, 0 };
1028
1029                 /*
1030                  * note: journal buckets aren't really counted as _sectors_ used yet, so
1031                  * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
1032                  * when space used goes up without a reservation - but we do need the
1033                  * reservation to ensure we'll actually be able to allocate:
1034                  *
1035                  * XXX: that's not right, disk reservations only ensure a
1036                  * filesystem-wide allocation will succeed, this is a device
1037                  * specific allocation - we can hang here:
1038                  */
1039
1040                 ret = bch2_disk_reservation_get(c, &disk_res,
1041                                                 bucket_to_sector(ca, nr - ja->nr), 1, 0);
1042                 if (ret)
1043                         break;
1044
1045                 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
1046
1047                 bch2_disk_reservation_put(c, &disk_res);
1048
1049                 closure_sync(&cl);
1050
1051                 if (ret && ret != -BCH_ERR_bucket_alloc_blocked)
1052                         break;
1053         }
1054
1055         bch_err_fn(c, ret);
1056 unlock:
1057         up_write(&c->state_lock);
1058         return ret;
1059 }
1060
1061 int bch2_dev_journal_alloc(struct bch_dev *ca)
1062 {
1063         unsigned nr;
1064         int ret;
1065
1066         if (dynamic_fault("bcachefs:add:journal_alloc")) {
1067                 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets;
1068                 goto err;
1069         }
1070
1071         /* 1/128th of the device by default: */
1072         nr = ca->mi.nbuckets >> 7;
1073
1074         /*
1075          * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
1076          * is smaller:
1077          */
1078         nr = clamp_t(unsigned, nr,
1079                      BCH_JOURNAL_BUCKETS_MIN,
1080                      min(1 << 13,
1081                          (1 << 24) / ca->mi.bucket_size));
1082
1083         ret = __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
1084 err:
1085         bch_err_fn(ca, ret);
1086         return ret;
1087 }
1088
1089 int bch2_fs_journal_alloc(struct bch_fs *c)
1090 {
1091         for_each_online_member(c, ca) {
1092                 if (ca->journal.nr)
1093                         continue;
1094
1095                 int ret = bch2_dev_journal_alloc(ca);
1096                 if (ret) {
1097                         percpu_ref_put(&ca->io_ref);
1098                         return ret;
1099                 }
1100         }
1101
1102         return 0;
1103 }
1104
1105 /* startup/shutdown: */
1106
1107 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
1108 {
1109         bool ret = false;
1110         u64 seq;
1111
1112         spin_lock(&j->lock);
1113         for (seq = journal_last_unwritten_seq(j);
1114              seq <= journal_cur_seq(j) && !ret;
1115              seq++) {
1116                 struct journal_buf *buf = journal_seq_to_buf(j, seq);
1117
1118                 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx))
1119                         ret = true;
1120         }
1121         spin_unlock(&j->lock);
1122
1123         return ret;
1124 }
1125
1126 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
1127 {
1128         wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
1129 }
1130
1131 void bch2_fs_journal_stop(struct journal *j)
1132 {
1133         bch2_journal_reclaim_stop(j);
1134         bch2_journal_flush_all_pins(j);
1135
1136         wait_event(j->wait, bch2_journal_entry_close(j));
1137
1138         /*
1139          * Always write a new journal entry, to make sure the clock hands are up
1140          * to date (and match the superblock)
1141          */
1142         bch2_journal_meta(j);
1143
1144         journal_quiesce(j);
1145
1146         BUG_ON(!bch2_journal_error(j) &&
1147                test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
1148                j->last_empty_seq != journal_cur_seq(j));
1149
1150         cancel_delayed_work_sync(&j->write_work);
1151 }
1152
1153 int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
1154 {
1155         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1156         struct journal_entry_pin_list *p;
1157         struct journal_replay *i, **_i;
1158         struct genradix_iter iter;
1159         bool had_entries = false;
1160         unsigned ptr;
1161         u64 last_seq = cur_seq, nr, seq;
1162
1163         genradix_for_each_reverse(&c->journal_entries, iter, _i) {
1164                 i = *_i;
1165
1166                 if (!i || i->ignore)
1167                         continue;
1168
1169                 last_seq = le64_to_cpu(i->j.last_seq);
1170                 break;
1171         }
1172
1173         nr = cur_seq - last_seq;
1174
1175         if (nr + 1 > j->pin.size) {
1176                 free_fifo(&j->pin);
1177                 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1178                 if (!j->pin.data) {
1179                         bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1180                         return -BCH_ERR_ENOMEM_journal_pin_fifo;
1181                 }
1182         }
1183
1184         j->replay_journal_seq   = last_seq;
1185         j->replay_journal_seq_end = cur_seq;
1186         j->last_seq_ondisk      = last_seq;
1187         j->flushed_seq_ondisk   = cur_seq - 1;
1188         j->seq_ondisk           = cur_seq - 1;
1189         j->pin.front            = last_seq;
1190         j->pin.back             = cur_seq;
1191         atomic64_set(&j->seq, cur_seq - 1);
1192
1193         fifo_for_each_entry_ptr(p, &j->pin, seq)
1194                 journal_pin_list_init(p, 1);
1195
1196         genradix_for_each(&c->journal_entries, iter, _i) {
1197                 i = *_i;
1198
1199                 if (!i || i->ignore)
1200                         continue;
1201
1202                 seq = le64_to_cpu(i->j.seq);
1203                 BUG_ON(seq >= cur_seq);
1204
1205                 if (seq < last_seq)
1206                         continue;
1207
1208                 if (journal_entry_empty(&i->j))
1209                         j->last_empty_seq = le64_to_cpu(i->j.seq);
1210
1211                 p = journal_seq_pin(j, seq);
1212
1213                 p->devs.nr = 0;
1214                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1215                         bch2_dev_list_add_dev(&p->devs, i->ptrs[ptr].dev);
1216
1217                 had_entries = true;
1218         }
1219
1220         if (!had_entries)
1221                 j->last_empty_seq = cur_seq;
1222
1223         spin_lock(&j->lock);
1224
1225         set_bit(JOURNAL_STARTED, &j->flags);
1226         j->last_flush_write = jiffies;
1227
1228         j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1229         j->reservations.unwritten_idx++;
1230
1231         c->last_bucket_seq_cleanup = journal_cur_seq(j);
1232
1233         bch2_journal_space_available(j);
1234         spin_unlock(&j->lock);
1235
1236         return bch2_journal_reclaim_start(j);
1237 }
1238
1239 /* init/exit: */
1240
1241 void bch2_dev_journal_exit(struct bch_dev *ca)
1242 {
1243         kfree(ca->journal.bio);
1244         kfree(ca->journal.buckets);
1245         kfree(ca->journal.bucket_seq);
1246
1247         ca->journal.bio         = NULL;
1248         ca->journal.buckets     = NULL;
1249         ca->journal.bucket_seq  = NULL;
1250 }
1251
1252 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1253 {
1254         struct journal_device *ja = &ca->journal;
1255         struct bch_sb_field_journal *journal_buckets =
1256                 bch2_sb_field_get(sb, journal);
1257         struct bch_sb_field_journal_v2 *journal_buckets_v2 =
1258                 bch2_sb_field_get(sb, journal_v2);
1259         unsigned i, nr_bvecs;
1260
1261         ja->nr = 0;
1262
1263         if (journal_buckets_v2) {
1264                 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1265
1266                 for (i = 0; i < nr; i++)
1267                         ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr);
1268         } else if (journal_buckets) {
1269                 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1270         }
1271
1272         ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1273         if (!ja->bucket_seq)
1274                 return -BCH_ERR_ENOMEM_dev_journal_init;
1275
1276         nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
1277
1278         ca->journal.bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
1279         if (!ca->journal.bio)
1280                 return -BCH_ERR_ENOMEM_dev_journal_init;
1281
1282         bio_init(ca->journal.bio, NULL, ca->journal.bio->bi_inline_vecs, nr_bvecs, 0);
1283
1284         ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1285         if (!ja->buckets)
1286                 return -BCH_ERR_ENOMEM_dev_journal_init;
1287
1288         if (journal_buckets_v2) {
1289                 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2);
1290                 unsigned j, dst = 0;
1291
1292                 for (i = 0; i < nr; i++)
1293                         for (j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++)
1294                                 ja->buckets[dst++] =
1295                                         le64_to_cpu(journal_buckets_v2->d[i].start) + j;
1296         } else if (journal_buckets) {
1297                 for (i = 0; i < ja->nr; i++)
1298                         ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1299         }
1300
1301         return 0;
1302 }
1303
1304 void bch2_fs_journal_exit(struct journal *j)
1305 {
1306         unsigned i;
1307
1308         darray_exit(&j->early_journal_entries);
1309
1310         for (i = 0; i < ARRAY_SIZE(j->buf); i++)
1311                 kvpfree(j->buf[i].data, j->buf[i].buf_size);
1312         free_fifo(&j->pin);
1313 }
1314
1315 int bch2_fs_journal_init(struct journal *j)
1316 {
1317         static struct lock_class_key res_key;
1318         unsigned i;
1319
1320         mutex_init(&j->buf_lock);
1321         spin_lock_init(&j->lock);
1322         spin_lock_init(&j->err_lock);
1323         init_waitqueue_head(&j->wait);
1324         INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1325         init_waitqueue_head(&j->reclaim_wait);
1326         init_waitqueue_head(&j->pin_flush_wait);
1327         mutex_init(&j->reclaim_lock);
1328         mutex_init(&j->discard_lock);
1329
1330         lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1331
1332         atomic64_set(&j->reservations.counter,
1333                 ((union journal_res_state)
1334                  { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1335
1336         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)))
1337                 return -BCH_ERR_ENOMEM_journal_pin_fifo;
1338
1339         for (i = 0; i < ARRAY_SIZE(j->buf); i++) {
1340                 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1341                 j->buf[i].data = kvpmalloc(j->buf[i].buf_size, GFP_KERNEL);
1342                 if (!j->buf[i].data)
1343                         return -BCH_ERR_ENOMEM_journal_buf;
1344         }
1345
1346         j->pin.front = j->pin.back = 1;
1347         return 0;
1348 }
1349
1350 /* debug: */
1351
1352 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1353 {
1354         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1355         union journal_res_state s;
1356         unsigned long now = jiffies;
1357         u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes;
1358
1359         if (!out->nr_tabstops)
1360                 printbuf_tabstop_push(out, 24);
1361         out->atomic++;
1362
1363         rcu_read_lock();
1364         s = READ_ONCE(j->reservations);
1365
1366         prt_printf(out, "dirty journal entries:\t%llu/%llu\n",  fifo_used(&j->pin), j->pin.size);
1367         prt_printf(out, "seq:\t\t\t%llu\n",                     journal_cur_seq(j));
1368         prt_printf(out, "seq_ondisk:\t\t%llu\n",                j->seq_ondisk);
1369         prt_printf(out, "last_seq:\t\t%llu\n",                  journal_last_seq(j));
1370         prt_printf(out, "last_seq_ondisk:\t%llu\n",             j->last_seq_ondisk);
1371         prt_printf(out, "flushed_seq_ondisk:\t%llu\n",          j->flushed_seq_ondisk);
1372         prt_printf(out, "watermark:\t\t%s\n",                   bch2_watermarks[j->watermark]);
1373         prt_printf(out, "each entry reserved:\t%u\n",           j->entry_u64s_reserved);
1374         prt_printf(out, "nr flush writes:\t%llu\n",             j->nr_flush_writes);
1375         prt_printf(out, "nr noflush writes:\t%llu\n",           j->nr_noflush_writes);
1376         prt_printf(out, "average write size:\t");
1377         prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0);
1378         prt_newline(out);
1379         prt_printf(out, "nr direct reclaim:\t%llu\n",           j->nr_direct_reclaim);
1380         prt_printf(out, "nr background reclaim:\t%llu\n",       j->nr_background_reclaim);
1381         prt_printf(out, "reclaim kicked:\t\t%u\n",              j->reclaim_kicked);
1382         prt_printf(out, "reclaim runs in:\t%u ms\n",            time_after(j->next_reclaim, now)
1383                ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
1384         prt_printf(out, "current entry sectors:\t%u\n",         j->cur_entry_sectors);
1385         prt_printf(out, "current entry error:\t%s\n",           bch2_journal_errors[j->cur_entry_error]);
1386         prt_printf(out, "current entry:\t\t");
1387
1388         switch (s.cur_entry_offset) {
1389         case JOURNAL_ENTRY_ERROR_VAL:
1390                 prt_printf(out, "error");
1391                 break;
1392         case JOURNAL_ENTRY_CLOSED_VAL:
1393                 prt_printf(out, "closed");
1394                 break;
1395         default:
1396                 prt_printf(out, "%u/%u", s.cur_entry_offset, j->cur_entry_u64s);
1397                 break;
1398         }
1399
1400         prt_newline(out);
1401         prt_printf(out, "unwritten entries:");
1402         prt_newline(out);
1403         bch2_journal_bufs_to_text(out, j);
1404
1405         prt_printf(out,
1406                "replay done:\t\t%i\n",
1407                test_bit(JOURNAL_REPLAY_DONE,    &j->flags));
1408
1409         prt_printf(out, "space:\n");
1410         prt_printf(out, "\tdiscarded\t%u:%u\n",
1411                j->space[journal_space_discarded].next_entry,
1412                j->space[journal_space_discarded].total);
1413         prt_printf(out, "\tclean ondisk\t%u:%u\n",
1414                j->space[journal_space_clean_ondisk].next_entry,
1415                j->space[journal_space_clean_ondisk].total);
1416         prt_printf(out, "\tclean\t\t%u:%u\n",
1417                j->space[journal_space_clean].next_entry,
1418                j->space[journal_space_clean].total);
1419         prt_printf(out, "\ttotal\t\t%u:%u\n",
1420                j->space[journal_space_total].next_entry,
1421                j->space[journal_space_total].total);
1422
1423         for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
1424                 struct journal_device *ja = &ca->journal;
1425
1426                 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
1427                         continue;
1428
1429                 if (!ja->nr)
1430                         continue;
1431
1432                 prt_printf(out, "dev %u:\n",            ca->dev_idx);
1433                 prt_printf(out, "\tnr\t\t%u\n",         ja->nr);
1434                 prt_printf(out, "\tbucket size\t%u\n",  ca->mi.bucket_size);
1435                 prt_printf(out, "\tavailable\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free);
1436                 prt_printf(out, "\tdiscard_idx\t%u\n",  ja->discard_idx);
1437                 prt_printf(out, "\tdirty_ondisk\t%u (seq %llu)\n", ja->dirty_idx_ondisk,        ja->bucket_seq[ja->dirty_idx_ondisk]);
1438                 prt_printf(out, "\tdirty_idx\t%u (seq %llu)\n", ja->dirty_idx,          ja->bucket_seq[ja->dirty_idx]);
1439                 prt_printf(out, "\tcur_idx\t\t%u (seq %llu)\n", ja->cur_idx,            ja->bucket_seq[ja->cur_idx]);
1440         }
1441
1442         rcu_read_unlock();
1443
1444         --out->atomic;
1445 }
1446
1447 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1448 {
1449         spin_lock(&j->lock);
1450         __bch2_journal_debug_to_text(out, j);
1451         spin_unlock(&j->lock);
1452 }
1453
1454 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq)
1455 {
1456         struct journal_entry_pin_list *pin_list;
1457         struct journal_entry_pin *pin;
1458         unsigned i;
1459
1460         spin_lock(&j->lock);
1461         *seq = max(*seq, j->pin.front);
1462
1463         if (*seq >= j->pin.back) {
1464                 spin_unlock(&j->lock);
1465                 return true;
1466         }
1467
1468         out->atomic++;
1469
1470         pin_list = journal_seq_pin(j, *seq);
1471
1472         prt_printf(out, "%llu: count %u", *seq, atomic_read(&pin_list->count));
1473         prt_newline(out);
1474         printbuf_indent_add(out, 2);
1475
1476         for (i = 0; i < ARRAY_SIZE(pin_list->list); i++)
1477                 list_for_each_entry(pin, &pin_list->list[i], list) {
1478                         prt_printf(out, "\t%px %ps", pin, pin->flush);
1479                         prt_newline(out);
1480                 }
1481
1482         if (!list_empty(&pin_list->flushed)) {
1483                 prt_printf(out, "flushed:");
1484                 prt_newline(out);
1485         }
1486
1487         list_for_each_entry(pin, &pin_list->flushed, list) {
1488                 prt_printf(out, "\t%px %ps", pin, pin->flush);
1489                 prt_newline(out);
1490         }
1491
1492         printbuf_indent_sub(out, 2);
1493
1494         --out->atomic;
1495         spin_unlock(&j->lock);
1496
1497         return false;
1498 }
1499
1500 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1501 {
1502         u64 seq = 0;
1503
1504         while (!bch2_journal_seq_pins_to_text(out, j, &seq))
1505                 seq++;
1506 }