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 "buckets.h"
  13 #include "journal.h"
  14 #include "journal_io.h"
  15 #include "journal_reclaim.h"
  16 #include "journal_seq_blacklist.h"
  17 #include "super-io.h"
  18
  19 #include <trace/events/bcachefs.h>
  20
  21 static u64 last_unwritten_seq(struct journal *j)
  22 {
  23         union journal_res_state s = READ_ONCE(j->reservations);
  24
  25         lockdep_assert_held(&j->lock);
  26
  27         return journal_cur_seq(j) - s.prev_buf_unwritten;
  28 }
  29
  30 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
  31 {
  32         return seq >= last_unwritten_seq(j);
  33 }
  34
  35 static bool __journal_entry_is_open(union journal_res_state state)
  36 {
  37         return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
  38 }
  39
  40 static bool journal_entry_is_open(struct journal *j)
  41 {
  42         return __journal_entry_is_open(j->reservations);
  43 }
  44
  45 static inline struct journal_buf *
  46 journal_seq_to_buf(struct journal *j, u64 seq)
  47 {
  48         struct journal_buf *buf = NULL;
  49
  50         EBUG_ON(seq > journal_cur_seq(j));
  51         EBUG_ON(seq == journal_cur_seq(j) &&
  52                 j->reservations.cur_entry_offset == JOURNAL_ENTRY_CLOSED_VAL);
  53
  54         if (journal_seq_unwritten(j, seq)) {
  55                 buf = j->buf + (seq & 1);
  56                 EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
  57         }
  58         return buf;
  59 }
  60
  61 static void journal_pin_new_entry(struct journal *j, int count)
  62 {
  63         struct journal_entry_pin_list *p;
  64
  65         /*
  66          * The fifo_push() needs to happen at the same time as j->seq is
  67          * incremented for journal_last_seq() to be calculated correctly
  68          */
  69         atomic64_inc(&j->seq);
  70         p = fifo_push_ref(&j->pin);
  71
  72         INIT_LIST_HEAD(&p->list);
  73         INIT_LIST_HEAD(&p->flushed);
  74         atomic_set(&p->count, count);
  75         p->devs.nr = 0;
  76 }
  77
  78 static void bch2_journal_buf_init(struct journal *j)
  79 {
  80         struct journal_buf *buf = journal_cur_buf(j);
  81
  82         bkey_extent_init(&buf->key);
  83
  84         memset(buf->has_inode, 0, sizeof(buf->has_inode));
  85
  86         memset(buf->data, 0, sizeof(*buf->data));
  87         buf->data->seq  = cpu_to_le64(journal_cur_seq(j));
  88         buf->data->u64s = 0;
  89 }
  90
  91 void bch2_journal_halt(struct journal *j)
  92 {
  93         union journal_res_state old, new;
  94         u64 v = atomic64_read(&j->reservations.counter);
  95
  96         do {
  97                 old.v = new.v = v;
  98                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL)
  99                         return;
 100
 101                 new.cur_entry_offset = JOURNAL_ENTRY_ERROR_VAL;
 102         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 103                                        old.v, new.v)) != old.v);
 104
 105         j->err_seq = journal_cur_seq(j);
 106         journal_wake(j);
 107         closure_wake_up(&journal_cur_buf(j)->wait);
 108 }
 109
 110 /* journal entry close/open: */
 111
 112 void __bch2_journal_buf_put(struct journal *j, bool need_write_just_set)
 113 {
 114         if (!need_write_just_set &&
 115             test_bit(JOURNAL_NEED_WRITE, &j->flags))
 116                 bch2_time_stats_update(j->delay_time,
 117                                        j->need_write_time);
 118
 119         clear_bit(JOURNAL_NEED_WRITE, &j->flags);
 120
 121         closure_call(&j->io, bch2_journal_write, system_highpri_wq, NULL);
 122 }
 123
 124 /*
 125  * Returns true if journal entry is now closed:
 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         bool set_need_write = false;
 134         unsigned sectors;
 135
 136         lockdep_assert_held(&j->lock);
 137
 138         do {
 139                 old.v = new.v = v;
 140                 if (old.cur_entry_offset == JOURNAL_ENTRY_CLOSED_VAL)
 141                         return true;
 142
 143                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL) {
 144                         /* this entry will never be written: */
 145                         closure_wake_up(&buf->wait);
 146                         return true;
 147                 }
 148
 149                 if (!test_bit(JOURNAL_NEED_WRITE, &j->flags)) {
 150                         set_bit(JOURNAL_NEED_WRITE, &j->flags);
 151                         j->need_write_time = local_clock();
 152                         set_need_write = true;
 153                 }
 154
 155                 if (new.prev_buf_unwritten)
 156                         return false;
 157
 158                 new.cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL;
 159                 new.idx++;
 160                 new.prev_buf_unwritten = 1;
 161
 162                 BUG_ON(journal_state_count(new, new.idx));
 163         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 164                                        old.v, new.v)) != old.v);
 165
 166         buf->data->u64s         = cpu_to_le32(old.cur_entry_offset);
 167
 168         sectors = vstruct_blocks_plus(buf->data, c->block_bits,
 169                                       buf->u64s_reserved) << c->block_bits;
 170         BUG_ON(sectors > buf->sectors);
 171         buf->sectors = sectors;
 172
 173         /*
 174          * We have to set last_seq here, _before_ opening a new journal entry:
 175          *
 176          * A threads may replace an old pin with a new pin on their current
 177          * journal reservation - the expectation being that the journal will
 178          * contain either what the old pin protected or what the new pin
 179          * protects.
 180          *
 181          * After the old pin is dropped journal_last_seq() won't include the old
 182          * pin, so we can only write the updated last_seq on the entry that
 183          * contains whatever the new pin protects.
 184          *
 185          * Restated, we can _not_ update last_seq for a given entry if there
 186          * could be a newer entry open with reservations/pins that have been
 187          * taken against it.
 188          *
 189          * Hence, we want update/set last_seq on the current journal entry right
 190          * before we open a new one:
 191          */
 192         buf->data->last_seq     = cpu_to_le64(journal_last_seq(j));
 193
 194         journal_pin_new_entry(j, 1);
 195
 196         bch2_journal_buf_init(j);
 197
 198         cancel_delayed_work(&j->write_work);
 199
 200         bch2_journal_space_available(j);
 201
 202         bch2_journal_buf_put(j, old.idx, set_need_write);
 203         return true;
 204 }
 205
 206 static bool journal_entry_close(struct journal *j)
 207 {
 208         bool ret;
 209
 210         spin_lock(&j->lock);
 211         ret = __journal_entry_close(j);
 212         spin_unlock(&j->lock);
 213
 214         return ret;
 215 }
 216
 217 /*
 218  * should _only_ called from journal_res_get() - when we actually want a
 219  * journal reservation - journal entry is open means journal is dirty:
 220  *
 221  * returns:
 222  * 0:           success
 223  * -ENOSPC:     journal currently full, must invoke reclaim
 224  * -EAGAIN:     journal blocked, must wait
 225  * -EROFS:      insufficient rw devices or journal error
 226  */
 227 static int journal_entry_open(struct journal *j)
 228 {
 229         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 230         struct journal_buf *buf = journal_cur_buf(j);
 231         union journal_res_state old, new;
 232         int u64s;
 233         u64 v;
 234
 235         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
 236
 237         lockdep_assert_held(&j->lock);
 238         BUG_ON(journal_entry_is_open(j));
 239
 240         if (j->blocked)
 241                 return cur_entry_blocked;
 242
 243         if (j->cur_entry_error)
 244                 return j->cur_entry_error;
 245
 246         BUG_ON(!j->cur_entry_sectors);
 247
 248         buf->u64s_reserved      = j->entry_u64s_reserved;
 249         buf->disk_sectors       = j->cur_entry_sectors;
 250         buf->sectors            = min(buf->disk_sectors, buf->buf_size >> 9);
 251
 252         u64s = (int) (buf->sectors << 9) / sizeof(u64) -
 253                 journal_entry_overhead(j);
 254         u64s  = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
 255
 256         if (u64s <= le32_to_cpu(buf->data->u64s))
 257                 return cur_entry_journal_full;
 258
 259         /*
 260          * Must be set before marking the journal entry as open:
 261          */
 262         j->cur_entry_u64s = u64s;
 263
 264         v = atomic64_read(&j->reservations.counter);
 265         do {
 266                 old.v = new.v = v;
 267
 268                 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL)
 269                         return cur_entry_insufficient_devices;
 270
 271                 /* Handle any already added entries */
 272                 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
 273
 274                 EBUG_ON(journal_state_count(new, new.idx));
 275                 journal_state_inc(&new);
 276         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
 277                                        old.v, new.v)) != old.v);
 278
 279         if (j->res_get_blocked_start)
 280                 bch2_time_stats_update(j->blocked_time,
 281                                        j->res_get_blocked_start);
 282         j->res_get_blocked_start = 0;
 283
 284         mod_delayed_work(system_freezable_wq,
 285                          &j->write_work,
 286                          msecs_to_jiffies(j->write_delay_ms));
 287         journal_wake(j);
 288         return 0;
 289 }
 290
 291 static bool journal_quiesced(struct journal *j)
 292 {
 293         union journal_res_state state = READ_ONCE(j->reservations);
 294         bool ret = !state.prev_buf_unwritten && !__journal_entry_is_open(state);
 295
 296         if (!ret)
 297                 journal_entry_close(j);
 298         return ret;
 299 }
 300
 301 static void journal_quiesce(struct journal *j)
 302 {
 303         wait_event(j->wait, journal_quiesced(j));
 304 }
 305
 306 static void journal_write_work(struct work_struct *work)
 307 {
 308         struct journal *j = container_of(work, struct journal, write_work.work);
 309
 310         journal_entry_close(j);
 311 }
 312
 313 /*
 314  * Given an inode number, if that inode number has data in the journal that
 315  * hasn't yet been flushed, return the journal sequence number that needs to be
 316  * flushed:
 317  */
 318 u64 bch2_inode_journal_seq(struct journal *j, u64 inode)
 319 {
 320         size_t h = hash_64(inode, ilog2(sizeof(j->buf[0].has_inode) * 8));
 321         u64 seq = 0;
 322
 323         if (!test_bit(h, j->buf[0].has_inode) &&
 324             !test_bit(h, j->buf[1].has_inode))
 325                 return 0;
 326
 327         spin_lock(&j->lock);
 328         if (test_bit(h, journal_cur_buf(j)->has_inode))
 329                 seq = journal_cur_seq(j);
 330         else if (test_bit(h, journal_prev_buf(j)->has_inode))
 331                 seq = journal_cur_seq(j) - 1;
 332         spin_unlock(&j->lock);
 333
 334         return seq;
 335 }
 336
 337 void bch2_journal_set_has_inum(struct journal *j, u64 inode, u64 seq)
 338 {
 339         size_t h = hash_64(inode, ilog2(sizeof(j->buf[0].has_inode) * 8));
 340         struct journal_buf *buf;
 341
 342         spin_lock(&j->lock);
 343
 344         if ((buf = journal_seq_to_buf(j, seq)))
 345                 set_bit(h, buf->has_inode);
 346
 347         spin_unlock(&j->lock);
 348 }
 349
 350 static int __journal_res_get(struct journal *j, struct journal_res *res,
 351                              unsigned flags)
 352 {
 353         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 354         struct journal_buf *buf;
 355         bool can_discard;
 356         int ret;
 357 retry:
 358         if (journal_res_get_fast(j, res, flags))
 359                 return 0;
 360
 361         if (bch2_journal_error(j))
 362                 return -EROFS;
 363
 364         spin_lock(&j->lock);
 365
 366         /*
 367          * Recheck after taking the lock, so we don't race with another thread
 368          * that just did journal_entry_open() and call journal_entry_close()
 369          * unnecessarily
 370          */
 371         if (journal_res_get_fast(j, res, flags)) {
 372                 spin_unlock(&j->lock);
 373                 return 0;
 374         }
 375
 376         if (!(flags & JOURNAL_RES_GET_RESERVED) &&
 377             !test_bit(JOURNAL_MAY_GET_UNRESERVED, &j->flags)) {
 378                 /*
 379                  * Don't want to close current journal entry, just need to
 380                  * invoke reclaim:
 381                  */
 382                 ret = cur_entry_journal_full;
 383                 goto unlock;
 384         }
 385
 386         /*
 387          * If we couldn't get a reservation because the current buf filled up,
 388          * and we had room for a bigger entry on disk, signal that we want to
 389          * realloc the journal bufs:
 390          */
 391         buf = journal_cur_buf(j);
 392         if (journal_entry_is_open(j) &&
 393             buf->buf_size >> 9 < buf->disk_sectors &&
 394             buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
 395                 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
 396
 397         if (journal_entry_is_open(j) &&
 398             !__journal_entry_close(j)) {
 399                 /*
 400                  * We failed to get a reservation on the current open journal
 401                  * entry because it's full, and we can't close it because
 402                  * there's still a previous one in flight:
 403                  */
 404                 trace_journal_entry_full(c);
 405                 ret = cur_entry_blocked;
 406         } else {
 407                 ret = journal_entry_open(j);
 408         }
 409 unlock:
 410         if ((ret && ret != cur_entry_insufficient_devices) &&
 411             !j->res_get_blocked_start) {
 412                 j->res_get_blocked_start = local_clock() ?: 1;
 413                 trace_journal_full(c);
 414         }
 415
 416         can_discard = j->can_discard;
 417         spin_unlock(&j->lock);
 418
 419         if (!ret)
 420                 goto retry;
 421
 422         if (WARN_ONCE(ret == cur_entry_journal_full &&
 423                       !can_discard &&
 424                       (flags & JOURNAL_RES_GET_RESERVED),
 425                       "JOURNAL_RES_GET_RESERVED set but journal full")) {
 426                 char *buf;
 427
 428                 buf = kmalloc(4096, GFP_NOFS);
 429                 if (buf) {
 430                         bch2_journal_debug_to_text(&_PBUF(buf, 4096), j);
 431                         pr_err("\n%s", buf);
 432                         kfree(buf);
 433                 }
 434         }
 435
 436         /*
 437          * Journal is full - can't rely on reclaim from work item due to
 438          * freezing:
 439          */
 440         if ((ret == cur_entry_journal_full ||
 441              ret == cur_entry_journal_pin_full) &&
 442             !(flags & JOURNAL_RES_GET_NONBLOCK)) {
 443                 if (can_discard) {
 444                         bch2_journal_do_discards(j);
 445                         goto retry;
 446                 }
 447
 448                 if (mutex_trylock(&j->reclaim_lock)) {
 449                         bch2_journal_reclaim(j);
 450                         mutex_unlock(&j->reclaim_lock);
 451                 }
 452         }
 453
 454         return ret == cur_entry_insufficient_devices ? -EROFS : -EAGAIN;
 455 }
 456
 457 /*
 458  * Essentially the entry function to the journaling code. When bcachefs is doing
 459  * a btree insert, it calls this function to get the current journal write.
 460  * Journal write is the structure used set up journal writes. The calling
 461  * function will then add its keys to the structure, queuing them for the next
 462  * write.
 463  *
 464  * To ensure forward progress, the current task must not be holding any
 465  * btree node write locks.
 466  */
 467 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
 468                                   unsigned flags)
 469 {
 470         int ret;
 471
 472         closure_wait_event(&j->async_wait,
 473                    (ret = __journal_res_get(j, res, flags)) != -EAGAIN ||
 474                    (flags & JOURNAL_RES_GET_NONBLOCK));
 475         return ret;
 476 }
 477
 478 /* journal_preres: */
 479
 480 static bool journal_preres_available(struct journal *j,
 481                                      struct journal_preres *res,
 482                                      unsigned new_u64s,
 483                                      unsigned flags)
 484 {
 485         bool ret = bch2_journal_preres_get_fast(j, res, new_u64s, flags);
 486
 487         if (!ret && mutex_trylock(&j->reclaim_lock)) {
 488                 bch2_journal_reclaim(j);
 489                 mutex_unlock(&j->reclaim_lock);
 490         }
 491
 492         return ret;
 493 }
 494
 495 int __bch2_journal_preres_get(struct journal *j,
 496                               struct journal_preres *res,
 497                               unsigned new_u64s,
 498                               unsigned flags)
 499 {
 500         int ret;
 501
 502         closure_wait_event(&j->preres_wait,
 503                    (ret = bch2_journal_error(j)) ||
 504                    journal_preres_available(j, res, new_u64s, flags));
 505         return ret;
 506 }
 507
 508 /* journal_entry_res: */
 509
 510 void bch2_journal_entry_res_resize(struct journal *j,
 511                                    struct journal_entry_res *res,
 512                                    unsigned new_u64s)
 513 {
 514         union journal_res_state state;
 515         int d = new_u64s - res->u64s;
 516
 517         spin_lock(&j->lock);
 518
 519         j->entry_u64s_reserved += d;
 520         if (d <= 0)
 521                 goto out;
 522
 523         j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
 524         smp_mb();
 525         state = READ_ONCE(j->reservations);
 526
 527         if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
 528             state.cur_entry_offset > j->cur_entry_u64s) {
 529                 j->cur_entry_u64s += d;
 530                 /*
 531                  * Not enough room in current journal entry, have to flush it:
 532                  */
 533                 __journal_entry_close(j);
 534         } else {
 535                 journal_cur_buf(j)->u64s_reserved += d;
 536         }
 537 out:
 538         spin_unlock(&j->lock);
 539         res->u64s += d;
 540 }
 541
 542 /* journal flushing: */
 543
 544 /**
 545  * bch2_journal_flush_seq_async - wait for a journal entry to be written
 546  *
 547  * like bch2_journal_wait_on_seq, except that it triggers a write immediately if
 548  * necessary
 549  */
 550 int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
 551                                  struct closure *parent)
 552 {
 553         struct journal_buf *buf;
 554         int ret = 0;
 555
 556         if (seq <= j->seq_ondisk)
 557                 return 1;
 558
 559         spin_lock(&j->lock);
 560
 561         /* Recheck under lock: */
 562         if (j->err_seq && seq >= j->err_seq) {
 563                 ret = -EIO;
 564                 goto out;
 565         }
 566
 567         if (seq <= j->seq_ondisk) {
 568                 ret = 1;
 569                 goto out;
 570         }
 571
 572         if (parent &&
 573             (buf = journal_seq_to_buf(j, seq)))
 574                 if (!closure_wait(&buf->wait, parent))
 575                         BUG();
 576
 577         if (seq == journal_cur_seq(j))
 578                 __journal_entry_close(j);
 579 out:
 580         spin_unlock(&j->lock);
 581         return ret;
 582 }
 583
 584 int bch2_journal_flush_seq(struct journal *j, u64 seq)
 585 {
 586         u64 start_time = local_clock();
 587         int ret, ret2;
 588
 589         ret = wait_event_killable(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
 590
 591         bch2_time_stats_update(j->flush_seq_time, start_time);
 592
 593         return ret ?: ret2 < 0 ? ret2 : 0;
 594 }
 595
 596 int bch2_journal_meta(struct journal *j)
 597 {
 598         struct journal_res res;
 599         int ret;
 600
 601         memset(&res, 0, sizeof(res));
 602
 603         ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 604         if (ret)
 605                 return ret;
 606
 607         bch2_journal_res_put(j, &res);
 608
 609         return bch2_journal_flush_seq(j, res.seq);
 610 }
 611
 612 /*
 613  * bch2_journal_flush_async - if there is an open journal entry, or a journal
 614  * still being written, write it and wait for the write to complete
 615  */
 616 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
 617 {
 618         u64 seq, journal_seq;
 619
 620         spin_lock(&j->lock);
 621         journal_seq = journal_cur_seq(j);
 622
 623         if (journal_entry_is_open(j)) {
 624                 seq = journal_seq;
 625         } else if (journal_seq) {
 626                 seq = journal_seq - 1;
 627         } else {
 628                 spin_unlock(&j->lock);
 629                 return;
 630         }
 631         spin_unlock(&j->lock);
 632
 633         bch2_journal_flush_seq_async(j, seq, parent);
 634 }
 635
 636 int bch2_journal_flush(struct journal *j)
 637 {
 638         u64 seq, journal_seq;
 639
 640         spin_lock(&j->lock);
 641         journal_seq = journal_cur_seq(j);
 642
 643         if (journal_entry_is_open(j)) {
 644                 seq = journal_seq;
 645         } else if (journal_seq) {
 646                 seq = journal_seq - 1;
 647         } else {
 648                 spin_unlock(&j->lock);
 649                 return 0;
 650         }
 651         spin_unlock(&j->lock);
 652
 653         return bch2_journal_flush_seq(j, seq);
 654 }
 655
 656 /* block/unlock the journal: */
 657
 658 void bch2_journal_unblock(struct journal *j)
 659 {
 660         spin_lock(&j->lock);
 661         j->blocked--;
 662         spin_unlock(&j->lock);
 663
 664         journal_wake(j);
 665 }
 666
 667 void bch2_journal_block(struct journal *j)
 668 {
 669         spin_lock(&j->lock);
 670         j->blocked++;
 671         spin_unlock(&j->lock);
 672
 673         journal_quiesce(j);
 674 }
 675
 676 /* allocate journal on a device: */
 677
 678 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 679                                          bool new_fs, struct closure *cl)
 680 {
 681         struct bch_fs *c = ca->fs;
 682         struct journal_device *ja = &ca->journal;
 683         struct bch_sb_field_journal *journal_buckets;
 684         u64 *new_bucket_seq = NULL, *new_buckets = NULL;
 685         int ret = 0;
 686
 687         /* don't handle reducing nr of buckets yet: */
 688         if (nr <= ja->nr)
 689                 return 0;
 690
 691         new_buckets     = kzalloc(nr * sizeof(u64), GFP_KERNEL);
 692         new_bucket_seq  = kzalloc(nr * sizeof(u64), GFP_KERNEL);
 693         if (!new_buckets || !new_bucket_seq) {
 694                 ret = -ENOMEM;
 695                 goto err;
 696         }
 697
 698         journal_buckets = bch2_sb_resize_journal(&ca->disk_sb,
 699                                         nr + sizeof(*journal_buckets) / sizeof(u64));
 700         if (!journal_buckets) {
 701                 ret = -ENOSPC;
 702                 goto err;
 703         }
 704
 705         /*
 706          * We may be called from the device add path, before the new device has
 707          * actually been added to the running filesystem:
 708          */
 709         if (c)
 710                 spin_lock(&c->journal.lock);
 711
 712         memcpy(new_buckets,     ja->buckets,    ja->nr * sizeof(u64));
 713         memcpy(new_bucket_seq,  ja->bucket_seq, ja->nr * sizeof(u64));
 714         swap(new_buckets,       ja->buckets);
 715         swap(new_bucket_seq,    ja->bucket_seq);
 716
 717         if (c)
 718                 spin_unlock(&c->journal.lock);
 719
 720         while (ja->nr < nr) {
 721                 struct open_bucket *ob = NULL;
 722                 unsigned pos;
 723                 long bucket;
 724
 725                 if (new_fs) {
 726                         bucket = bch2_bucket_alloc_new_fs(ca);
 727                         if (bucket < 0) {
 728                                 ret = -ENOSPC;
 729                                 goto err;
 730                         }
 731                 } else {
 732                         rcu_read_lock();
 733                         ob = bch2_bucket_alloc(c, ca, RESERVE_ALLOC,
 734                                                false, cl);
 735                         rcu_read_unlock();
 736                         if (IS_ERR(ob)) {
 737                                 ret = cl ? -EAGAIN : -ENOSPC;
 738                                 goto err;
 739                         }
 740
 741                         bucket = sector_to_bucket(ca, ob->ptr.offset);
 742                 }
 743
 744                 if (c) {
 745                         percpu_down_read(&c->mark_lock);
 746                         spin_lock(&c->journal.lock);
 747                 }
 748
 749                 /*
 750                  * XXX
 751                  * For resize at runtime, we should be writing the new
 752                  * superblock before inserting into the journal array
 753                  */
 754
 755                 pos = ja->nr ? (ja->cur_idx + 1) % ja->nr : 0;
 756                 __array_insert_item(ja->buckets,                ja->nr, pos);
 757                 __array_insert_item(ja->bucket_seq,             ja->nr, pos);
 758                 __array_insert_item(journal_buckets->buckets,   ja->nr, pos);
 759                 ja->nr++;
 760
 761                 ja->buckets[pos] = bucket;
 762                 ja->bucket_seq[pos] = 0;
 763                 journal_buckets->buckets[pos] = cpu_to_le64(bucket);
 764
 765                 if (pos <= ja->discard_idx)
 766                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 767                 if (pos <= ja->dirty_idx_ondisk)
 768                         ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
 769                 if (pos <= ja->dirty_idx)
 770                         ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
 771                 if (pos <= ja->cur_idx)
 772                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
 773
 774                 bch2_mark_metadata_bucket(c, ca, bucket, BCH_DATA_journal,
 775                                           ca->mi.bucket_size,
 776                                           gc_phase(GC_PHASE_SB),
 777                                           0);
 778
 779                 if (c) {
 780                         spin_unlock(&c->journal.lock);
 781                         percpu_up_read(&c->mark_lock);
 782                 }
 783
 784                 if (!new_fs)
 785                         bch2_open_bucket_put(c, ob);
 786         }
 787 err:
 788         bch2_sb_resize_journal(&ca->disk_sb,
 789                 ja->nr + sizeof(*journal_buckets) / sizeof(u64));
 790         kfree(new_bucket_seq);
 791         kfree(new_buckets);
 792
 793         return ret;
 794 }
 795
 796 /*
 797  * Allocate more journal space at runtime - not currently making use if it, but
 798  * the code works:
 799  */
 800 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
 801                                 unsigned nr)
 802 {
 803         struct journal_device *ja = &ca->journal;
 804         struct closure cl;
 805         unsigned current_nr;
 806         int ret;
 807
 808         closure_init_stack(&cl);
 809
 810         do {
 811                 struct disk_reservation disk_res = { 0, 0 };
 812
 813                 closure_sync(&cl);
 814
 815                 mutex_lock(&c->sb_lock);
 816                 current_nr = ja->nr;
 817
 818                 /*
 819                  * note: journal buckets aren't really counted as _sectors_ used yet, so
 820                  * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
 821                  * when space used goes up without a reservation - but we do need the
 822                  * reservation to ensure we'll actually be able to allocate:
 823                  */
 824
 825                 if (bch2_disk_reservation_get(c, &disk_res,
 826                                               bucket_to_sector(ca, nr - ja->nr), 1, 0)) {
 827                         mutex_unlock(&c->sb_lock);
 828                         return -ENOSPC;
 829                 }
 830
 831                 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
 832
 833                 bch2_disk_reservation_put(c, &disk_res);
 834
 835                 if (ja->nr != current_nr)
 836                         bch2_write_super(c);
 837                 mutex_unlock(&c->sb_lock);
 838         } while (ret == -EAGAIN);
 839
 840         return ret;
 841 }
 842
 843 int bch2_dev_journal_alloc(struct bch_dev *ca)
 844 {
 845         unsigned nr;
 846
 847         if (dynamic_fault("bcachefs:add:journal_alloc"))
 848                 return -ENOMEM;
 849
 850         /*
 851          * clamp journal size to 1024 buckets or 512MB (in sectors), whichever
 852          * is smaller:
 853          */
 854         nr = clamp_t(unsigned, ca->mi.nbuckets >> 8,
 855                      BCH_JOURNAL_BUCKETS_MIN,
 856                      min(1 << 10,
 857                          (1 << 20) / ca->mi.bucket_size));
 858
 859         return __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
 860 }
 861
 862 /* startup/shutdown: */
 863
 864 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
 865 {
 866         union journal_res_state state;
 867         struct journal_buf *w;
 868         bool ret;
 869
 870         spin_lock(&j->lock);
 871         state = READ_ONCE(j->reservations);
 872         w = j->buf + !state.idx;
 873
 874         ret = state.prev_buf_unwritten &&
 875                 bch2_bkey_has_device(bkey_i_to_s_c(&w->key), dev_idx);
 876         spin_unlock(&j->lock);
 877
 878         return ret;
 879 }
 880
 881 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
 882 {
 883         wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
 884 }
 885
 886 void bch2_fs_journal_stop(struct journal *j)
 887 {
 888         bch2_journal_flush_all_pins(j);
 889
 890         wait_event(j->wait, journal_entry_close(j));
 891
 892         /*
 893          * Always write a new journal entry, to make sure the clock hands are up
 894          * to date (and match the superblock)
 895          */
 896         bch2_journal_meta(j);
 897
 898         journal_quiesce(j);
 899
 900         BUG_ON(!bch2_journal_error(j) &&
 901                (journal_entry_is_open(j) ||
 902                 j->last_empty_seq + 1 != journal_cur_seq(j)));
 903
 904         cancel_delayed_work_sync(&j->write_work);
 905         bch2_journal_reclaim_stop(j);
 906 }
 907
 908 int bch2_fs_journal_start(struct journal *j, u64 cur_seq,
 909                           struct list_head *journal_entries)
 910 {
 911         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 912         struct journal_entry_pin_list *p;
 913         struct journal_replay *i;
 914         u64 last_seq = cur_seq, nr, seq;
 915
 916         if (!list_empty(journal_entries))
 917                 last_seq = le64_to_cpu(list_last_entry(journal_entries,
 918                                 struct journal_replay, list)->j.last_seq);
 919
 920         nr = cur_seq - last_seq;
 921
 922         if (nr + 1 > j->pin.size) {
 923                 free_fifo(&j->pin);
 924                 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
 925                 if (!j->pin.data) {
 926                         bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
 927                         return -ENOMEM;
 928                 }
 929         }
 930
 931         j->replay_journal_seq   = last_seq;
 932         j->replay_journal_seq_end = cur_seq;
 933         j->last_seq_ondisk      = last_seq;
 934         j->pin.front            = last_seq;
 935         j->pin.back             = cur_seq;
 936         atomic64_set(&j->seq, cur_seq - 1);
 937
 938         fifo_for_each_entry_ptr(p, &j->pin, seq) {
 939                 INIT_LIST_HEAD(&p->list);
 940                 INIT_LIST_HEAD(&p->flushed);
 941                 atomic_set(&p->count, 1);
 942                 p->devs.nr = 0;
 943         }
 944
 945         list_for_each_entry(i, journal_entries, list) {
 946                 seq = le64_to_cpu(i->j.seq);
 947                 BUG_ON(seq >= cur_seq);
 948
 949                 if (seq < last_seq)
 950                         continue;
 951
 952                 journal_seq_pin(j, seq)->devs = i->devs;
 953         }
 954
 955         spin_lock(&j->lock);
 956
 957         set_bit(JOURNAL_STARTED, &j->flags);
 958
 959         journal_pin_new_entry(j, 1);
 960
 961         j->reservations.idx = journal_cur_seq(j);
 962
 963         bch2_journal_buf_init(j);
 964
 965         c->last_bucket_seq_cleanup = journal_cur_seq(j);
 966
 967         bch2_journal_space_available(j);
 968         spin_unlock(&j->lock);
 969
 970         return 0;
 971 }
 972
 973 /* init/exit: */
 974
 975 void bch2_dev_journal_exit(struct bch_dev *ca)
 976 {
 977         kfree(ca->journal.bio);
 978         kfree(ca->journal.buckets);
 979         kfree(ca->journal.bucket_seq);
 980
 981         ca->journal.bio         = NULL;
 982         ca->journal.buckets     = NULL;
 983         ca->journal.bucket_seq  = NULL;
 984 }
 985
 986 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
 987 {
 988         struct journal_device *ja = &ca->journal;
 989         struct bch_sb_field_journal *journal_buckets =
 990                 bch2_sb_get_journal(sb);
 991         unsigned i;
 992
 993         ja->nr = bch2_nr_journal_buckets(journal_buckets);
 994
 995         ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
 996         if (!ja->bucket_seq)
 997                 return -ENOMEM;
 998
 999         ca->journal.bio = bio_kmalloc(GFP_KERNEL,
1000                         DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE));
1001         if (!ca->journal.bio)
1002                 return -ENOMEM;
1003
1004         ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1005         if (!ja->buckets)
1006                 return -ENOMEM;
1007
1008         for (i = 0; i < ja->nr; i++)
1009                 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1010
1011         return 0;
1012 }
1013
1014 void bch2_fs_journal_exit(struct journal *j)
1015 {
1016         kvpfree(j->buf[1].data, j->buf[1].buf_size);
1017         kvpfree(j->buf[0].data, j->buf[0].buf_size);
1018         free_fifo(&j->pin);
1019 }
1020
1021 int bch2_fs_journal_init(struct journal *j)
1022 {
1023         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1024         static struct lock_class_key res_key;
1025         int ret = 0;
1026
1027         pr_verbose_init(c->opts, "");
1028
1029         spin_lock_init(&j->lock);
1030         spin_lock_init(&j->err_lock);
1031         init_waitqueue_head(&j->wait);
1032         INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1033         init_waitqueue_head(&j->pin_flush_wait);
1034         mutex_init(&j->reclaim_lock);
1035         mutex_init(&j->discard_lock);
1036
1037         lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1038
1039         j->buf[0].buf_size      = JOURNAL_ENTRY_SIZE_MIN;
1040         j->buf[1].buf_size      = JOURNAL_ENTRY_SIZE_MIN;
1041         j->write_delay_ms       = 1000;
1042         j->reclaim_delay_ms     = 100;
1043
1044         /* Btree roots: */
1045         j->entry_u64s_reserved +=
1046                 BTREE_ID_NR * (JSET_KEYS_U64s + BKEY_EXTENT_U64s_MAX);
1047
1048         atomic64_set(&j->reservations.counter,
1049                 ((union journal_res_state)
1050                  { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1051
1052         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
1053             !(j->buf[0].data = kvpmalloc(j->buf[0].buf_size, GFP_KERNEL)) ||
1054             !(j->buf[1].data = kvpmalloc(j->buf[1].buf_size, GFP_KERNEL))) {
1055                 ret = -ENOMEM;
1056                 goto out;
1057         }
1058
1059         j->pin.front = j->pin.back = 1;
1060 out:
1061         pr_verbose_init(c->opts, "ret %i", ret);
1062         return ret;
1063 }
1064
1065 /* debug: */
1066
1067 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
1068 {
1069         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1070         union journal_res_state s;
1071         struct bch_dev *ca;
1072         unsigned iter;
1073
1074         rcu_read_lock();
1075         spin_lock(&j->lock);
1076         s = READ_ONCE(j->reservations);
1077
1078         pr_buf(out,
1079                "active journal entries:\t%llu\n"
1080                "seq:\t\t\t%llu\n"
1081                "last_seq:\t\t%llu\n"
1082                "last_seq_ondisk:\t%llu\n"
1083                "prereserved:\t\t%u/%u\n"
1084                "nr direct reclaim:\t%llu\n"
1085                "nr background reclaim:\t%llu\n"
1086                "current entry sectors:\t%u\n"
1087                "current entry error:\t%u\n"
1088                "current entry:\t\t",
1089                fifo_used(&j->pin),
1090                journal_cur_seq(j),
1091                journal_last_seq(j),
1092                j->last_seq_ondisk,
1093                j->prereserved.reserved,
1094                j->prereserved.remaining,
1095                j->nr_direct_reclaim,
1096                j->nr_background_reclaim,
1097                j->cur_entry_sectors,
1098                j->cur_entry_error);
1099
1100         switch (s.cur_entry_offset) {
1101         case JOURNAL_ENTRY_ERROR_VAL:
1102                 pr_buf(out, "error\n");
1103                 break;
1104         case JOURNAL_ENTRY_CLOSED_VAL:
1105                 pr_buf(out, "closed\n");
1106                 break;
1107         default:
1108                 pr_buf(out, "%u/%u\n",
1109                        s.cur_entry_offset,
1110                        j->cur_entry_u64s);
1111                 break;
1112         }
1113
1114         pr_buf(out,
1115                "current entry refs:\t%u\n"
1116                "prev entry unwritten:\t",
1117                journal_state_count(s, s.idx));
1118
1119         if (s.prev_buf_unwritten)
1120                 pr_buf(out, "yes, ref %u sectors %u\n",
1121                        journal_state_count(s, !s.idx),
1122                        journal_prev_buf(j)->sectors);
1123         else
1124                 pr_buf(out, "no\n");
1125
1126         pr_buf(out,
1127                "need write:\t\t%i\n"
1128                "replay done:\t\t%i\n",
1129                test_bit(JOURNAL_NEED_WRITE,     &j->flags),
1130                test_bit(JOURNAL_REPLAY_DONE,    &j->flags));
1131
1132         for_each_member_device_rcu(ca, c, iter,
1133                                    &c->rw_devs[BCH_DATA_journal]) {
1134                 struct journal_device *ja = &ca->journal;
1135
1136                 if (!ja->nr)
1137                         continue;
1138
1139                 pr_buf(out,
1140                        "dev %u:\n"
1141                        "\tnr\t\t%u\n"
1142                        "\tavailable\t%u:%u\n"
1143                        "\tdiscard_idx\t\t%u\n"
1144                        "\tdirty_idx_ondisk\t%u (seq %llu)\n"
1145                        "\tdirty_idx\t\t%u (seq %llu)\n"
1146                        "\tcur_idx\t\t%u (seq %llu)\n",
1147                        iter, ja->nr,
1148                        bch2_journal_dev_buckets_available(j, ja, journal_space_discarded),
1149                        ja->sectors_free,
1150                        ja->discard_idx,
1151                        ja->dirty_idx_ondisk,    ja->bucket_seq[ja->dirty_idx_ondisk],
1152                        ja->dirty_idx,           ja->bucket_seq[ja->dirty_idx],
1153                        ja->cur_idx,             ja->bucket_seq[ja->cur_idx]);
1154         }
1155
1156         spin_unlock(&j->lock);
1157         rcu_read_unlock();
1158 }
1159
1160 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
1161 {
1162         struct journal_entry_pin_list *pin_list;
1163         struct journal_entry_pin *pin;
1164         u64 i;
1165
1166         spin_lock(&j->lock);
1167         fifo_for_each_entry_ptr(pin_list, &j->pin, i) {
1168                 pr_buf(out, "%llu: count %u\n",
1169                        i, atomic_read(&pin_list->count));
1170
1171                 list_for_each_entry(pin, &pin_list->list, list)
1172                         pr_buf(out, "\t%px %ps\n",
1173                                pin, pin->flush);
1174
1175                 if (!list_empty(&pin_list->flushed))
1176                         pr_buf(out, "flushed:\n");
1177
1178                 list_for_each_entry(pin, &pin_list->flushed, list)
1179                         pr_buf(out, "\t%px %ps\n",
1180                                pin, pin->flush);
1181         }
1182         spin_unlock(&j->lock);
1183 }