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 "buckets.h"
  14 #include "journal.h"
  15 #include "journal_io.h"
  16 #include "journal_reclaim.h"
  17 #include "journal_seq_blacklist.h"
  18 #include "super-io.h"
  19
  20 #include <trace/events/bcachefs.h>
  21
  22 static u64 last_unwritten_seq(struct journal *j)
  23 {
  24         union journal_res_state s = READ_ONCE(j->reservations);
  25
  26         lockdep_assert_held(&j->lock);
  27
  28         return journal_cur_seq(j) - ((s.idx - s.unwritten_idx) & JOURNAL_BUF_MASK);
  29 }
  30
  31 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
  32 {
  33         return seq >= last_unwritten_seq(j);
  34 }
  35
  36 static bool __journal_entry_is_open(union journal_res_state state)
  37 {
  38         return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
  39 }
  40
  41 static bool journal_entry_is_open(struct journal *j)
  42 {
  43         return __journal_entry_is_open(j->reservations);
  44 }
  45
  46 static inline struct journal_buf *
  47 journal_seq_to_buf(struct journal *j, u64 seq)
  48 {
  49         struct journal_buf *buf = NULL;
  50
  51         EBUG_ON(seq > journal_cur_seq(j));
  52         EBUG_ON(seq == journal_cur_seq(j) &&
  53                 j->reservations.cur_entry_offset == JOURNAL_ENTRY_CLOSED_VAL);
  54
  55         if (journal_seq_unwritten(j, seq)) {
  56                 buf = j->buf + (seq & JOURNAL_BUF_MASK);
  57                 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
  58         }
  59         return buf;
  60 }
  61
  62 static void journal_pin_new_entry(struct journal *j, int count)
  63 {
  64         struct journal_entry_pin_list *p;
  65
  66         /*
  67          * The fifo_push() needs to happen at the same time as j->seq is
  68          * incremented for journal_last_seq() to be calculated correctly
  69          */
  70         atomic64_inc(&j->seq);
  71         p = fifo_push_ref(&j->pin);
  72
  73         INIT_LIST_HEAD(&p->list);
  74         INIT_LIST_HEAD(&p->flushed);
  75         atomic_set(&p->count, count);
  76         p->devs.nr = 0;
  77 }
  78
  79 static void bch2_journal_buf_init(struct journal *j)
  80 {
  81         struct journal_buf *buf = journal_cur_buf(j);
  82
  83         bkey_extent_init(&buf->key);
  84         buf->noflush    = false;
  85         buf->must_flush = false;
  86         buf->separate_flush = false;
  87
  88         memset(buf->has_inode, 0, sizeof(buf->has_inode));
  89
  90         memset(buf->data, 0, sizeof(*buf->data));
  91         buf->data->seq  = cpu_to_le64(journal_cur_seq(j));
  92         buf->data->u64s = 0;
  93 }
  94
  95 void bch2_journal_halt(struct journal *j)
  96 {
  97         union journal_res_state old, new;
  98         u64 v = atomic64_read(&j->reservations.counter);
  99
 100         do {
 101                 old.v = new.v = v;
 102                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL)
 103                         return;
 104
 105                 new.cur_entry_offset = JOURNAL_ENTRY_ERROR_VAL;
 106         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 107                                        old.v, new.v)) != old.v);
 108
 109         j->err_seq = journal_cur_seq(j);
 110         journal_wake(j);
 111         closure_wake_up(&journal_cur_buf(j)->wait);
 112 }
 113
 114 /* journal entry close/open: */
 115
 116 void __bch2_journal_buf_put(struct journal *j)
 117 {
 118         closure_call(&j->io, bch2_journal_write, system_highpri_wq, NULL);
 119 }
 120
 121 /*
 122  * Returns true if journal entry is now closed:
 123  *
 124  * We don't close a journal_buf until the next journal_buf is finished writing,
 125  * and can be opened again - this also initializes the next journal_buf:
 126  */
 127 static bool __journal_entry_close(struct journal *j)
 128 {
 129         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 130         struct journal_buf *buf = journal_cur_buf(j);
 131         union journal_res_state old, new;
 132         u64 v = atomic64_read(&j->reservations.counter);
 133         unsigned sectors;
 134
 135         lockdep_assert_held(&j->lock);
 136
 137         do {
 138                 old.v = new.v = v;
 139                 if (old.cur_entry_offset == JOURNAL_ENTRY_CLOSED_VAL)
 140                         return true;
 141
 142                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL) {
 143                         /* this entry will never be written: */
 144                         closure_wake_up(&buf->wait);
 145                         return true;
 146                 }
 147
 148                 if (!test_bit(JOURNAL_NEED_WRITE, &j->flags)) {
 149                         set_bit(JOURNAL_NEED_WRITE, &j->flags);
 150                         j->need_write_time = local_clock();
 151                 }
 152
 153                 new.cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL;
 154                 new.idx++;
 155
 156                 if (new.idx == new.unwritten_idx)
 157                         return false;
 158
 159                 BUG_ON(journal_state_count(new, new.idx));
 160         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 161                                        old.v, new.v)) != old.v);
 162
 163         /* Close out old buffer: */
 164         buf->data->u64s         = cpu_to_le32(old.cur_entry_offset);
 165
 166         sectors = vstruct_blocks_plus(buf->data, c->block_bits,
 167                                       buf->u64s_reserved) << c->block_bits;
 168         BUG_ON(sectors > buf->sectors);
 169         buf->sectors = sectors;
 170
 171         /*
 172          * We have to set last_seq here, _before_ opening a new journal entry:
 173          *
 174          * A threads may replace an old pin with a new pin on their current
 175          * journal reservation - the expectation being that the journal will
 176          * contain either what the old pin protected or what the new pin
 177          * protects.
 178          *
 179          * After the old pin is dropped journal_last_seq() won't include the old
 180          * pin, so we can only write the updated last_seq on the entry that
 181          * contains whatever the new pin protects.
 182          *
 183          * Restated, we can _not_ update last_seq for a given entry if there
 184          * could be a newer entry open with reservations/pins that have been
 185          * taken against it.
 186          *
 187          * Hence, we want update/set last_seq on the current journal entry right
 188          * before we open a new one:
 189          */
 190         buf->data->last_seq     = cpu_to_le64(journal_last_seq(j));
 191
 192         __bch2_journal_pin_put(j, le64_to_cpu(buf->data->seq));
 193
 194         /* Initialize new buffer: */
 195         journal_pin_new_entry(j, 1);
 196
 197         bch2_journal_buf_init(j);
 198
 199         cancel_delayed_work(&j->write_work);
 200         clear_bit(JOURNAL_NEED_WRITE, &j->flags);
 201
 202         bch2_journal_space_available(j);
 203
 204         bch2_journal_buf_put(j, old.idx);
 205         return true;
 206 }
 207
 208 static bool journal_entry_want_write(struct journal *j)
 209 {
 210         union journal_res_state s = READ_ONCE(j->reservations);
 211         bool ret = false;
 212
 213         /*
 214          * Don't close it yet if we already have a write in flight, but do set
 215          * NEED_WRITE:
 216          */
 217         if (s.idx != s.unwritten_idx)
 218                 set_bit(JOURNAL_NEED_WRITE, &j->flags);
 219         else
 220                 ret = __journal_entry_close(j);
 221
 222         return ret;
 223 }
 224
 225 static bool journal_entry_close(struct journal *j)
 226 {
 227         bool ret;
 228
 229         spin_lock(&j->lock);
 230         ret = journal_entry_want_write(j);
 231         spin_unlock(&j->lock);
 232
 233         return ret;
 234 }
 235
 236 /*
 237  * should _only_ called from journal_res_get() - when we actually want a
 238  * journal reservation - journal entry is open means journal is dirty:
 239  *
 240  * returns:
 241  * 0:           success
 242  * -ENOSPC:     journal currently full, must invoke reclaim
 243  * -EAGAIN:     journal blocked, must wait
 244  * -EROFS:      insufficient rw devices or journal error
 245  */
 246 static int journal_entry_open(struct journal *j)
 247 {
 248         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 249         struct journal_buf *buf = journal_cur_buf(j);
 250         union journal_res_state old, new;
 251         int u64s;
 252         u64 v;
 253
 254         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
 255
 256         lockdep_assert_held(&j->lock);
 257         BUG_ON(journal_entry_is_open(j));
 258
 259         if (j->blocked)
 260                 return cur_entry_blocked;
 261
 262         if (j->cur_entry_error)
 263                 return j->cur_entry_error;
 264
 265         BUG_ON(!j->cur_entry_sectors);
 266
 267         buf->u64s_reserved      = j->entry_u64s_reserved;
 268         buf->disk_sectors       = j->cur_entry_sectors;
 269         buf->sectors            = min(buf->disk_sectors, buf->buf_size >> 9);
 270
 271         u64s = (int) (buf->sectors << 9) / sizeof(u64) -
 272                 journal_entry_overhead(j);
 273         u64s  = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
 274
 275         if (u64s <= le32_to_cpu(buf->data->u64s))
 276                 return cur_entry_journal_full;
 277
 278         /*
 279          * Must be set before marking the journal entry as open:
 280          */
 281         j->cur_entry_u64s = u64s;
 282
 283         v = atomic64_read(&j->reservations.counter);
 284         do {
 285                 old.v = new.v = v;
 286
 287                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL)
 288                         return cur_entry_insufficient_devices;
 289
 290                 /* Handle any already added entries */
 291                 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
 292
 293                 EBUG_ON(journal_state_count(new, new.idx));
 294                 journal_state_inc(&new);
 295         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 296                                        old.v, new.v)) != old.v);
 297
 298         if (j->res_get_blocked_start)
 299                 bch2_time_stats_update(j->blocked_time,
 300                                        j->res_get_blocked_start);
 301         j->res_get_blocked_start = 0;
 302
 303         mod_delayed_work(system_freezable_wq,
 304                          &j->write_work,
 305                          msecs_to_jiffies(j->write_delay_ms));
 306         journal_wake(j);
 307         return 0;
 308 }
 309
 310 static bool journal_quiesced(struct journal *j)
 311 {
 312         union journal_res_state s = READ_ONCE(j->reservations);
 313         bool ret = s.idx == s.unwritten_idx && !__journal_entry_is_open(s);
 314
 315         if (!ret)
 316                 journal_entry_close(j);
 317         return ret;
 318 }
 319
 320 static void journal_quiesce(struct journal *j)
 321 {
 322         wait_event(j->wait, journal_quiesced(j));
 323 }
 324
 325 static void journal_write_work(struct work_struct *work)
 326 {
 327         struct journal *j = container_of(work, struct journal, write_work.work);
 328
 329         journal_entry_close(j);
 330 }
 331
 332 /*
 333  * Given an inode number, if that inode number has data in the journal that
 334  * hasn't yet been flushed, return the journal sequence number that needs to be
 335  * flushed:
 336  */
 337 u64 bch2_inode_journal_seq(struct journal *j, u64 inode)
 338 {
 339         size_t h = hash_64(inode, ilog2(sizeof(j->buf[0].has_inode) * 8));
 340         union journal_res_state s;
 341         unsigned i;
 342         u64 seq;
 343
 344
 345         spin_lock(&j->lock);
 346         seq = journal_cur_seq(j);
 347         s = READ_ONCE(j->reservations);
 348         i = s.idx;
 349
 350         while (1) {
 351                 if (test_bit(h, j->buf[i].has_inode))
 352                         goto out;
 353
 354                 if (i == s.unwritten_idx)
 355                         break;
 356
 357                 i = (i - 1) & JOURNAL_BUF_MASK;
 358                 seq--;
 359         }
 360
 361         seq = 0;
 362 out:
 363         spin_unlock(&j->lock);
 364
 365         return seq;
 366 }
 367
 368 void bch2_journal_set_has_inum(struct journal *j, u64 inode, u64 seq)
 369 {
 370         size_t h = hash_64(inode, ilog2(sizeof(j->buf[0].has_inode) * 8));
 371         struct journal_buf *buf;
 372
 373         spin_lock(&j->lock);
 374
 375         if ((buf = journal_seq_to_buf(j, seq)))
 376                 set_bit(h, buf->has_inode);
 377
 378         spin_unlock(&j->lock);
 379 }
 380
 381 static int __journal_res_get(struct journal *j, struct journal_res *res,
 382                              unsigned flags)
 383 {
 384         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 385         struct journal_buf *buf;
 386         bool can_discard;
 387         int ret;
 388 retry:
 389         if (journal_res_get_fast(j, res, flags))
 390                 return 0;
 391
 392         if (bch2_journal_error(j))
 393                 return -EROFS;
 394
 395         spin_lock(&j->lock);
 396
 397         /*
 398          * Recheck after taking the lock, so we don't race with another thread
 399          * that just did journal_entry_open() and call journal_entry_close()
 400          * unnecessarily
 401          */
 402         if (journal_res_get_fast(j, res, flags)) {
 403                 spin_unlock(&j->lock);
 404                 return 0;
 405         }
 406
 407         if (!(flags & JOURNAL_RES_GET_RESERVED) &&
 408             !test_bit(JOURNAL_MAY_GET_UNRESERVED, &j->flags)) {
 409                 /*
 410                  * Don't want to close current journal entry, just need to
 411                  * invoke reclaim:
 412                  */
 413                 ret = cur_entry_journal_full;
 414                 goto unlock;
 415         }
 416
 417         /*
 418          * If we couldn't get a reservation because the current buf filled up,
 419          * and we had room for a bigger entry on disk, signal that we want to
 420          * realloc the journal bufs:
 421          */
 422         buf = journal_cur_buf(j);
 423         if (journal_entry_is_open(j) &&
 424             buf->buf_size >> 9 < buf->disk_sectors &&
 425             buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
 426                 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
 427
 428         if (journal_entry_is_open(j) &&
 429             !__journal_entry_close(j)) {
 430                 /*
 431                  * We failed to get a reservation on the current open journal
 432                  * entry because it's full, and we can't close it because
 433                  * there's still a previous one in flight:
 434                  */
 435                 trace_journal_entry_full(c);
 436                 ret = cur_entry_blocked;
 437         } else {
 438                 ret = journal_entry_open(j);
 439         }
 440 unlock:
 441         if ((ret && ret != cur_entry_insufficient_devices) &&
 442             !j->res_get_blocked_start) {
 443                 j->res_get_blocked_start = local_clock() ?: 1;
 444                 trace_journal_full(c);
 445         }
 446
 447         can_discard = j->can_discard;
 448         spin_unlock(&j->lock);
 449
 450         if (!ret)
 451                 goto retry;
 452
 453         /*
 454          * Journal is full - can't rely on reclaim from work item due to
 455          * freezing:
 456          */
 457         if ((ret == cur_entry_journal_full ||
 458              ret == cur_entry_journal_pin_full) &&
 459             !(flags & JOURNAL_RES_GET_NONBLOCK)) {
 460                 if (can_discard) {
 461                         bch2_journal_do_discards(j);
 462                         goto retry;
 463                 }
 464
 465                 if (mutex_trylock(&j->reclaim_lock)) {
 466                         bch2_journal_reclaim(j);
 467                         mutex_unlock(&j->reclaim_lock);
 468                 }
 469         }
 470
 471         return ret == cur_entry_insufficient_devices ? -EROFS : -EAGAIN;
 472 }
 473
 474 /*
 475  * Essentially the entry function to the journaling code. When bcachefs is doing
 476  * a btree insert, it calls this function to get the current journal write.
 477  * Journal write is the structure used set up journal writes. The calling
 478  * function will then add its keys to the structure, queuing them for the next
 479  * write.
 480  *
 481  * To ensure forward progress, the current task must not be holding any
 482  * btree node write locks.
 483  */
 484 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
 485                                   unsigned flags)
 486 {
 487         int ret;
 488
 489         closure_wait_event(&j->async_wait,
 490                    (ret = __journal_res_get(j, res, flags)) != -EAGAIN ||
 491                    (flags & JOURNAL_RES_GET_NONBLOCK));
 492         return ret;
 493 }
 494
 495 /* journal_preres: */
 496
 497 static bool journal_preres_available(struct journal *j,
 498                                      struct journal_preres *res,
 499                                      unsigned new_u64s,
 500                                      unsigned flags)
 501 {
 502         bool ret = bch2_journal_preres_get_fast(j, res, new_u64s, flags);
 503
 504         if (!ret && mutex_trylock(&j->reclaim_lock)) {
 505                 bch2_journal_reclaim(j);
 506                 mutex_unlock(&j->reclaim_lock);
 507         }
 508
 509         return ret;
 510 }
 511
 512 int __bch2_journal_preres_get(struct journal *j,
 513                               struct journal_preres *res,
 514                               unsigned new_u64s,
 515                               unsigned flags)
 516 {
 517         int ret;
 518
 519         closure_wait_event(&j->preres_wait,
 520                    (ret = bch2_journal_error(j)) ||
 521                    journal_preres_available(j, res, new_u64s, flags));
 522         return ret;
 523 }
 524
 525 /* journal_entry_res: */
 526
 527 void bch2_journal_entry_res_resize(struct journal *j,
 528                                    struct journal_entry_res *res,
 529                                    unsigned new_u64s)
 530 {
 531         union journal_res_state state;
 532         int d = new_u64s - res->u64s;
 533
 534         spin_lock(&j->lock);
 535
 536         j->entry_u64s_reserved += d;
 537         if (d <= 0)
 538                 goto out;
 539
 540         j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
 541         smp_mb();
 542         state = READ_ONCE(j->reservations);
 543
 544         if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
 545             state.cur_entry_offset > j->cur_entry_u64s) {
 546                 j->cur_entry_u64s += d;
 547                 /*
 548                  * Not enough room in current journal entry, have to flush it:
 549                  */
 550                 __journal_entry_close(j);
 551         } else {
 552                 journal_cur_buf(j)->u64s_reserved += d;
 553         }
 554 out:
 555         spin_unlock(&j->lock);
 556         res->u64s += d;
 557 }
 558
 559 /* journal flushing: */
 560
 561 /**
 562  * bch2_journal_flush_seq_async - wait for a journal entry to be written
 563  *
 564  * like bch2_journal_wait_on_seq, except that it triggers a write immediately if
 565  * necessary
 566  */
 567 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
 568                                  struct closure *parent)
 569 {
 570         struct journal_buf *buf;
 571         int ret = 0;
 572
 573         if (seq <= j->flushed_seq_ondisk)
 574                 return 1;
 575
 576         spin_lock(&j->lock);
 577
 578         BUG_ON(seq > journal_cur_seq(j));
 579
 580         /* Recheck under lock: */
 581         if (j->err_seq && seq >= j->err_seq) {
 582                 ret = -EIO;
 583                 goto out;
 584         }
 585
 586         if (seq <= j->flushed_seq_ondisk) {
 587                 ret = 1;
 588                 goto out;
 589         }
 590
 591         /* if seq was written, but not flushed - flush a newer one instead */
 592         seq = max(seq, last_unwritten_seq(j));
 593
 594 recheck_need_open:
 595         if (seq == journal_cur_seq(j) && !journal_entry_is_open(j)) {
 596                 struct journal_res res = { 0 };
 597
 598                 spin_unlock(&j->lock);
 599
 600                 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 601                 if (ret)
 602                         return ret;
 603
 604                 seq = res.seq;
 605                 buf = j->buf + (seq & JOURNAL_BUF_MASK);
 606                 buf->must_flush = true;
 607                 set_bit(JOURNAL_NEED_WRITE, &j->flags);
 608
 609                 if (parent && !closure_wait(&buf->wait, parent))
 610                         BUG();
 611
 612                 bch2_journal_res_put(j, &res);
 613
 614                 spin_lock(&j->lock);
 615                 goto want_write;
 616         }
 617
 618         /*
 619          * if write was kicked off without a flush, flush the next sequence
 620          * number instead
 621          */
 622         buf = journal_seq_to_buf(j, seq);
 623         if (buf->noflush) {
 624                 seq++;
 625                 goto recheck_need_open;
 626         }
 627
 628         buf->must_flush = true;
 629
 630         if (parent && !closure_wait(&buf->wait, parent))
 631                 BUG();
 632 want_write:
 633         if (seq == journal_cur_seq(j))
 634                 journal_entry_want_write(j);
 635 out:
 636         spin_unlock(&j->lock);
 637         return ret;
 638 }
 639
 640 int bch2_journal_flush_seq(struct journal *j, u64 seq)
 641 {
 642         u64 start_time = local_clock();
 643         int ret, ret2;
 644
 645         ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
 646
 647         if (!ret)
 648                 bch2_time_stats_update(j->flush_seq_time, start_time);
 649
 650         return ret ?: ret2 < 0 ? ret2 : 0;
 651 }
 652
 653 int bch2_journal_meta(struct journal *j)
 654 {
 655         struct journal_res res;
 656         int ret;
 657
 658         memset(&res, 0, sizeof(res));
 659
 660         ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 661         if (ret)
 662                 return ret;
 663
 664         bch2_journal_res_put(j, &res);
 665
 666         return bch2_journal_flush_seq(j, res.seq);
 667 }
 668
 669 /*
 670  * bch2_journal_flush_async - if there is an open journal entry, or a journal
 671  * still being written, write it and wait for the write to complete
 672  */
 673 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
 674 {
 675         u64 seq, journal_seq;
 676
 677         spin_lock(&j->lock);
 678         journal_seq = journal_cur_seq(j);
 679
 680         if (journal_entry_is_open(j)) {
 681                 seq = journal_seq;
 682         } else if (journal_seq) {
 683                 seq = journal_seq - 1;
 684         } else {
 685                 spin_unlock(&j->lock);
 686                 return;
 687         }
 688         spin_unlock(&j->lock);
 689
 690         bch2_journal_flush_seq_async(j, seq, parent);
 691 }
 692
 693 int bch2_journal_flush(struct journal *j)
 694 {
 695         u64 seq, journal_seq;
 696
 697         spin_lock(&j->lock);
 698         journal_seq = journal_cur_seq(j);
 699
 700         if (journal_entry_is_open(j)) {
 701                 seq = journal_seq;
 702         } else if (journal_seq) {
 703                 seq = journal_seq - 1;
 704         } else {
 705                 spin_unlock(&j->lock);
 706                 return 0;
 707         }
 708         spin_unlock(&j->lock);
 709
 710         return bch2_journal_flush_seq(j, seq);
 711 }
 712
 713 /* block/unlock the journal: */
 714
 715 void bch2_journal_unblock(struct journal *j)
 716 {
 717         spin_lock(&j->lock);
 718         j->blocked--;
 719         spin_unlock(&j->lock);
 720
 721         journal_wake(j);
 722 }
 723
 724 void bch2_journal_block(struct journal *j)
 725 {
 726         spin_lock(&j->lock);
 727         j->blocked++;
 728         spin_unlock(&j->lock);
 729
 730         journal_quiesce(j);
 731 }
 732
 733 /* allocate journal on a device: */
 734
 735 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 736                                          bool new_fs, struct closure *cl)
 737 {
 738         struct bch_fs *c = ca->fs;
 739         struct journal_device *ja = &ca->journal;
 740         struct bch_sb_field_journal *journal_buckets;
 741         u64 *new_bucket_seq = NULL, *new_buckets = NULL;
 742         int ret = 0;
 743
 744         /* don't handle reducing nr of buckets yet: */
 745         if (nr <= ja->nr)
 746                 return 0;
 747
 748         new_buckets     = kzalloc(nr * sizeof(u64), GFP_KERNEL);
 749         new_bucket_seq  = kzalloc(nr * sizeof(u64), GFP_KERNEL);
 750         if (!new_buckets || !new_bucket_seq) {
 751                 ret = -ENOMEM;
 752                 goto err;
 753         }
 754
 755         journal_buckets = bch2_sb_resize_journal(&ca->disk_sb,
 756                                         nr + sizeof(*journal_buckets) / sizeof(u64));
 757         if (!journal_buckets) {
 758                 ret = -ENOSPC;
 759                 goto err;
 760         }
 761
 762         /*
 763          * We may be called from the device add path, before the new device has
 764          * actually been added to the running filesystem:
 765          */
 766         if (c)
 767                 spin_lock(&c->journal.lock);
 768
 769         memcpy(new_buckets,     ja->buckets,    ja->nr * sizeof(u64));
 770         memcpy(new_bucket_seq,  ja->bucket_seq, ja->nr * sizeof(u64));
 771         swap(new_buckets,       ja->buckets);
 772         swap(new_bucket_seq,    ja->bucket_seq);
 773
 774         if (c)
 775                 spin_unlock(&c->journal.lock);
 776
 777         while (ja->nr < nr) {
 778                 struct open_bucket *ob = NULL;
 779                 unsigned pos;
 780                 long bucket;
 781
 782                 if (new_fs) {
 783                         bucket = bch2_bucket_alloc_new_fs(ca);
 784                         if (bucket < 0) {
 785                                 ret = -ENOSPC;
 786                                 goto err;
 787                         }
 788                 } else {
 789                         rcu_read_lock();
 790                         ob = bch2_bucket_alloc(c, ca, RESERVE_NONE,
 791                                                false, cl);
 792                         rcu_read_unlock();
 793                         if (IS_ERR(ob)) {
 794                                 ret = cl ? -EAGAIN : -ENOSPC;
 795                                 goto err;
 796                         }
 797
 798                         bucket = sector_to_bucket(ca, ob->ptr.offset);
 799                 }
 800
 801                 if (c) {
 802                         percpu_down_read(&c->mark_lock);
 803                         spin_lock(&c->journal.lock);
 804                 }
 805
 806                 /*
 807                  * XXX
 808                  * For resize at runtime, we should be writing the new
 809                  * superblock before inserting into the journal array
 810                  */
 811
 812                 pos = ja->nr ? (ja->cur_idx + 1) % ja->nr : 0;
 813                 __array_insert_item(ja->buckets,                ja->nr, pos);
 814                 __array_insert_item(ja->bucket_seq,             ja->nr, pos);
 815                 __array_insert_item(journal_buckets->buckets,   ja->nr, pos);
 816                 ja->nr++;
 817
 818                 ja->buckets[pos] = bucket;
 819                 ja->bucket_seq[pos] = 0;
 820                 journal_buckets->buckets[pos] = cpu_to_le64(bucket);
 821
 822                 if (pos <= ja->discard_idx)
 823                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 824                 if (pos <= ja->dirty_idx_ondisk)
 825                         ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
 826                 if (pos <= ja->dirty_idx)
 827                         ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
 828                 if (pos <= ja->cur_idx)
 829                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
 830
 831                 if (!c || new_fs)
 832                         bch2_mark_metadata_bucket(c, ca, bucket, BCH_DATA_journal,
 833                                                   ca->mi.bucket_size,
 834                                                   gc_phase(GC_PHASE_SB),
 835                                                   0);
 836
 837                 if (c) {
 838                         spin_unlock(&c->journal.lock);
 839                         percpu_up_read(&c->mark_lock);
 840                 }
 841
 842                 if (c && !new_fs)
 843                         ret = bch2_trans_do(c, NULL, NULL, BTREE_INSERT_NOFAIL,
 844                                 bch2_trans_mark_metadata_bucket(&trans, NULL, ca,
 845                                                 bucket, BCH_DATA_journal,
 846                                                 ca->mi.bucket_size));
 847
 848                 if (!new_fs)
 849                         bch2_open_bucket_put(c, ob);
 850
 851                 if (ret)
 852                         goto err;
 853         }
 854 err:
 855         bch2_sb_resize_journal(&ca->disk_sb,
 856                 ja->nr + sizeof(*journal_buckets) / sizeof(u64));
 857         kfree(new_bucket_seq);
 858         kfree(new_buckets);
 859
 860         return ret;
 861 }
 862
 863 /*
 864  * Allocate more journal space at runtime - not currently making use if it, but
 865  * the code works:
 866  */
 867 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
 868                                 unsigned nr)
 869 {
 870         struct journal_device *ja = &ca->journal;
 871         struct closure cl;
 872         unsigned current_nr;
 873         int ret;
 874
 875         closure_init_stack(&cl);
 876
 877         do {
 878                 struct disk_reservation disk_res = { 0, 0 };
 879
 880                 closure_sync(&cl);
 881
 882                 mutex_lock(&c->sb_lock);
 883                 current_nr = ja->nr;
 884
 885                 /*
 886                  * note: journal buckets aren't really counted as _sectors_ used yet, so
 887                  * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
 888                  * when space used goes up without a reservation - but we do need the
 889                  * reservation to ensure we'll actually be able to allocate:
 890                  */
 891
 892                 if (bch2_disk_reservation_get(c, &disk_res,
 893                                               bucket_to_sector(ca, nr - ja->nr), 1, 0)) {
 894                         mutex_unlock(&c->sb_lock);
 895                         return -ENOSPC;
 896                 }
 897
 898                 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
 899
 900                 bch2_disk_reservation_put(c, &disk_res);
 901
 902                 if (ja->nr != current_nr)
 903                         bch2_write_super(c);
 904                 mutex_unlock(&c->sb_lock);
 905         } while (ret == -EAGAIN);
 906
 907         return ret;
 908 }
 909
 910 int bch2_dev_journal_alloc(struct bch_dev *ca)
 911 {
 912         unsigned nr;
 913
 914         if (dynamic_fault("bcachefs:add:journal_alloc"))
 915                 return -ENOMEM;
 916
 917         /*
 918          * clamp journal size to 1024 buckets or 512MB (in sectors), whichever
 919          * is smaller:
 920          */
 921         nr = clamp_t(unsigned, ca->mi.nbuckets >> 8,
 922                      BCH_JOURNAL_BUCKETS_MIN,
 923                      min(1 << 10,
 924                          (1 << 20) / ca->mi.bucket_size));
 925
 926         return __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
 927 }
 928
 929 /* startup/shutdown: */
 930
 931 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
 932 {
 933         union journal_res_state state;
 934         bool ret = false;
 935         unsigned i;
 936
 937         spin_lock(&j->lock);
 938         state = READ_ONCE(j->reservations);
 939         i = state.idx;
 940
 941         while (i != state.unwritten_idx) {
 942                 i = (i - 1) & JOURNAL_BUF_MASK;
 943                 if (bch2_bkey_has_device(bkey_i_to_s_c(&j->buf[i].key), dev_idx))
 944                         ret = true;
 945         }
 946         spin_unlock(&j->lock);
 947
 948         return ret;
 949 }
 950
 951 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
 952 {
 953         wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
 954 }
 955
 956 void bch2_fs_journal_stop(struct journal *j)
 957 {
 958         bch2_journal_flush_all_pins(j);
 959
 960         wait_event(j->wait, journal_entry_close(j));
 961
 962         /*
 963          * Always write a new journal entry, to make sure the clock hands are up
 964          * to date (and match the superblock)
 965          */
 966         bch2_journal_meta(j);
 967
 968         journal_quiesce(j);
 969
 970         BUG_ON(!bch2_journal_error(j) &&
 971                test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
 972                (journal_entry_is_open(j) ||
 973                 j->last_empty_seq + 1 != journal_cur_seq(j)));
 974
 975         cancel_delayed_work_sync(&j->write_work);
 976         bch2_journal_reclaim_stop(j);
 977 }
 978
 979 int bch2_fs_journal_start(struct journal *j, u64 cur_seq,
 980                           struct list_head *journal_entries)
 981 {
 982         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 983         struct journal_entry_pin_list *p;
 984         struct journal_replay *i;
 985         u64 last_seq = cur_seq, nr, seq;
 986
 987         if (!list_empty(journal_entries))
 988                 last_seq = le64_to_cpu(list_last_entry(journal_entries,
 989                                 struct journal_replay, list)->j.last_seq);
 990
 991         nr = cur_seq - last_seq;
 992
 993         if (nr + 1 > j->pin.size) {
 994                 free_fifo(&j->pin);
 995                 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
 996                 if (!j->pin.data) {
 997                         bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
 998                         return -ENOMEM;
 999                 }
1000         }
1001
1002         j->replay_journal_seq   = last_seq;
1003         j->replay_journal_seq_end = cur_seq;
1004         j->last_seq_ondisk      = last_seq;
1005         j->pin.front            = last_seq;
1006         j->pin.back             = cur_seq;
1007         atomic64_set(&j->seq, cur_seq - 1);
1008
1009         fifo_for_each_entry_ptr(p, &j->pin, seq) {
1010                 INIT_LIST_HEAD(&p->list);
1011                 INIT_LIST_HEAD(&p->flushed);
1012                 atomic_set(&p->count, 1);
1013                 p->devs.nr = 0;
1014         }
1015
1016         list_for_each_entry(i, journal_entries, list) {
1017                 unsigned ptr;
1018
1019                 seq = le64_to_cpu(i->j.seq);
1020                 BUG_ON(seq >= cur_seq);
1021
1022                 if (seq < last_seq)
1023                         continue;
1024
1025                 p = journal_seq_pin(j, seq);
1026
1027                 p->devs.nr = 0;
1028                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1029                         bch2_dev_list_add_dev(&p->devs, i->ptrs[ptr].dev);
1030         }
1031
1032         spin_lock(&j->lock);
1033
1034         set_bit(JOURNAL_STARTED, &j->flags);
1035         j->last_flush_write = jiffies;
1036
1037         journal_pin_new_entry(j, 1);
1038
1039         j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
1040
1041         bch2_journal_buf_init(j);
1042
1043         c->last_bucket_seq_cleanup = journal_cur_seq(j);
1044
1045         bch2_journal_space_available(j);
1046         spin_unlock(&j->lock);
1047
1048         return 0;
1049 }
1050
1051 /* init/exit: */
1052
1053 void bch2_dev_journal_exit(struct bch_dev *ca)
1054 {
1055         kfree(ca->journal.bio);
1056         kfree(ca->journal.buckets);
1057         kfree(ca->journal.bucket_seq);
1058
1059         ca->journal.bio         = NULL;
1060         ca->journal.buckets     = NULL;
1061         ca->journal.bucket_seq  = NULL;
1062 }
1063
1064 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1065 {
1066         struct journal_device *ja = &ca->journal;
1067         struct bch_sb_field_journal *journal_buckets =
1068                 bch2_sb_get_journal(sb);
1069         unsigned i;
1070
1071         ja->nr = bch2_nr_journal_buckets(journal_buckets);
1072
1073         ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1074         if (!ja->bucket_seq)
1075                 return -ENOMEM;
1076
1077         ca->journal.bio = bio_kmalloc(GFP_KERNEL,
1078                         DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE));
1079         if (!ca->journal.bio)
1080                 return -ENOMEM;
1081
1082         ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1083         if (!ja->buckets)
1084                 return -ENOMEM;
1085
1086         for (i = 0; i < ja->nr; i++)
1087                 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1088
1089         return 0;
1090 }
1091
1092 void bch2_fs_journal_exit(struct journal *j)
1093 {
1094         unsigned i;
1095
1096         for (i = 0; i < ARRAY_SIZE(j->buf); i++)
1097                 kvpfree(j->buf[i].data, j->buf[i].buf_size);
1098         free_fifo(&j->pin);
1099 }
1100
1101 int bch2_fs_journal_init(struct journal *j)
1102 {
1103         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1104         static struct lock_class_key res_key;
1105         unsigned i;
1106         int ret = 0;
1107
1108         pr_verbose_init(c->opts, "");
1109
1110         spin_lock_init(&j->lock);
1111         spin_lock_init(&j->err_lock);
1112         init_waitqueue_head(&j->wait);
1113         INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1114         init_waitqueue_head(&j->pin_flush_wait);
1115         mutex_init(&j->reclaim_lock);
1116         mutex_init(&j->discard_lock);
1117
1118         lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1119
1120         j->write_delay_ms       = 1000;
1121         j->reclaim_delay_ms     = 100;
1122
1123         atomic64_set(&j->reservations.counter,
1124                 ((union journal_res_state)
1125                  { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1126
1127         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL))) {
1128                 ret = -ENOMEM;
1129                 goto out;
1130         }
1131
1132         for (i = 0; i < ARRAY_SIZE(j->buf); i++) {
1133                 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1134                 j->buf[i].data = kvpmalloc(j->buf[i].buf_size, GFP_KERNEL);
1135                 if (!j->buf[i].data) {
1136                         ret = -ENOMEM;
1137                         goto out;
1138                 }
1139         }
1140
1141         j->pin.front = j->pin.back = 1;
1142 out:
1143         pr_verbose_init(c->opts, "ret %i", ret);
1144         return ret;
1145 }
1146
1147 /* debug: */
1148
1149 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1150 {
1151         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1152         union journal_res_state s;
1153         struct bch_dev *ca;
1154         unsigned i;
1155
1156         rcu_read_lock();
1157         s = READ_ONCE(j->reservations);
1158
1159         pr_buf(out,
1160                "active journal entries:\t%llu\n"
1161                "seq:\t\t\t%llu\n"
1162                "last_seq:\t\t%llu\n"
1163                "last_seq_ondisk:\t%llu\n"
1164                "flushed_seq_ondisk:\t%llu\n"
1165                "prereserved:\t\t%u/%u\n"
1166                "nr flush writes:\t%llu\n"
1167                "nr noflush writes:\t%llu\n"
1168                "nr direct reclaim:\t%llu\n"
1169                "nr background reclaim:\t%llu\n"
1170                "current entry sectors:\t%u\n"
1171                "current entry error:\t%u\n"
1172                "current entry:\t\t",
1173                fifo_used(&j->pin),
1174                journal_cur_seq(j),
1175                journal_last_seq(j),
1176                j->last_seq_ondisk,
1177                j->flushed_seq_ondisk,
1178                j->prereserved.reserved,
1179                j->prereserved.remaining,
1180                j->nr_flush_writes,
1181                j->nr_noflush_writes,
1182                j->nr_direct_reclaim,
1183                j->nr_background_reclaim,
1184                j->cur_entry_sectors,
1185                j->cur_entry_error);
1186
1187         switch (s.cur_entry_offset) {
1188         case JOURNAL_ENTRY_ERROR_VAL:
1189                 pr_buf(out, "error\n");
1190                 break;
1191         case JOURNAL_ENTRY_CLOSED_VAL:
1192                 pr_buf(out, "closed\n");
1193                 break;
1194         default:
1195                 pr_buf(out, "%u/%u\n",
1196                        s.cur_entry_offset,
1197                        j->cur_entry_u64s);
1198                 break;
1199         }
1200
1201         pr_buf(out,
1202                "current entry:\t\tidx %u refcount %u\n",
1203                s.idx, journal_state_count(s, s.idx));
1204
1205         i = s.idx;
1206         while (i != s.unwritten_idx) {
1207                 i = (i - 1) & JOURNAL_BUF_MASK;
1208
1209                 pr_buf(out, "unwritten entry:\tidx %u refcount %u sectors %u\n",
1210                        i, journal_state_count(s, i), j->buf[i].sectors);
1211         }
1212
1213         pr_buf(out,
1214                "need write:\t\t%i\n"
1215                "replay done:\t\t%i\n",
1216                test_bit(JOURNAL_NEED_WRITE,     &j->flags),
1217                test_bit(JOURNAL_REPLAY_DONE,    &j->flags));
1218
1219         pr_buf(out, "space:\n");
1220         pr_buf(out, "\tdiscarded\t%u:%u\n",
1221                j->space[journal_space_discarded].next_entry,
1222                j->space[journal_space_discarded].total);
1223         pr_buf(out, "\tclean ondisk\t%u:%u\n",
1224                j->space[journal_space_clean_ondisk].next_entry,
1225                j->space[journal_space_clean_ondisk].total);
1226         pr_buf(out, "\tclean\t\t%u:%u\n",
1227                j->space[journal_space_clean].next_entry,
1228                j->space[journal_space_clean].total);
1229         pr_buf(out, "\ttotal\t\t%u:%u\n",
1230                j->space[journal_space_total].next_entry,
1231                j->space[journal_space_total].total);
1232
1233         for_each_member_device_rcu(ca, c, i,
1234                                    &c->rw_devs[BCH_DATA_journal]) {
1235                 struct journal_device *ja = &ca->journal;
1236
1237                 if (!ja->nr)
1238                         continue;
1239
1240                 pr_buf(out,
1241                        "dev %u:\n"
1242                        "\tnr\t\t%u\n"
1243                        "\tbucket size\t%u\n"
1244                        "\tavailable\t%u:%u\n"
1245                        "\tdiscard_idx\t%u\n"
1246                        "\tdirty_ondisk\t%u (seq %llu)\n"
1247                        "\tdirty_idx\t%u (seq %llu)\n"
1248                        "\tcur_idx\t\t%u (seq %llu)\n",
1249                        i, ja->nr, ca->mi.bucket_size,
1250                        bch2_journal_dev_buckets_available(j, ja, journal_space_discarded),
1251                        ja->sectors_free,
1252                        ja->discard_idx,
1253                        ja->dirty_idx_ondisk,    ja->bucket_seq[ja->dirty_idx_ondisk],
1254                        ja->dirty_idx,           ja->bucket_seq[ja->dirty_idx],
1255                        ja->cur_idx,             ja->bucket_seq[ja->cur_idx]);
1256         }
1257
1258         rcu_read_unlock();
1259 }
1260
1261 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1262 {
1263         spin_lock(&j->lock);
1264         __bch2_journal_debug_to_text(out, j);
1265         spin_unlock(&j->lock);
1266 }
1267
1268 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1269 {
1270         struct journal_entry_pin_list *pin_list;
1271         struct journal_entry_pin *pin;
1272         u64 i;
1273
1274         spin_lock(&j->lock);
1275         fifo_for_each_entry_ptr(pin_list, &j->pin, i) {
1276                 pr_buf(out, "%llu: count %u\n",
1277                        i, atomic_read(&pin_list->count));
1278
1279                 list_for_each_entry(pin, &pin_list->list, list)
1280                         pr_buf(out, "\t%px %ps\n",
1281                                pin, pin->flush);
1282
1283                 if (!list_empty(&pin_list->flushed))
1284                         pr_buf(out, "flushed:\n");
1285
1286                 list_for_each_entry(pin, &pin_list->flushed, list)
1287                         pr_buf(out, "\t%px %ps\n",
1288                                pin, pin->flush);
1289         }
1290         spin_unlock(&j->lock);
1291 }