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 inline bool journal_seq_unwritten(struct journal *j, u64 seq) 24 { 25 return seq > j->seq_ondisk; 26 } 27 28 static bool __journal_entry_is_open(union journal_res_state state) 29 { 30 return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL; 31 } 32 33 static inline unsigned nr_unwritten_journal_entries(struct journal *j) 34 { 35 return atomic64_read(&j->seq) - j->seq_ondisk; 36 } 37 38 static bool journal_entry_is_open(struct journal *j) 39 { 40 return __journal_entry_is_open(j->reservations); 41 } 42 43 static void bch2_journal_buf_to_text(struct printbuf *out, struct journal *j, u64 seq) 44 { 45 union journal_res_state s = READ_ONCE(j->reservations); 46 unsigned i = seq & JOURNAL_BUF_MASK; 47 struct journal_buf *buf = j->buf + i; 48 49 prt_printf(out, "seq:\t%llu\n", seq); 50 printbuf_indent_add(out, 2); 51 52 if (!buf->write_started) 53 prt_printf(out, "refcount:\t%u\n", journal_state_count(s, i & JOURNAL_STATE_BUF_MASK)); 54 55 struct closure *cl = &buf->io; 56 int r = atomic_read(&cl->remaining); 57 prt_printf(out, "io:\t%pS r %i\n", cl->fn, r & CLOSURE_REMAINING_MASK); 58 59 if (buf->data) { 60 prt_printf(out, "size:\t"); 61 prt_human_readable_u64(out, vstruct_bytes(buf->data)); 62 prt_newline(out); 63 } 64 65 prt_printf(out, "expires:\t%li jiffies\n", buf->expires - jiffies); 66 67 prt_printf(out, "flags:\t"); 68 if (buf->noflush) 69 prt_str(out, "noflush "); 70 if (buf->must_flush) 71 prt_str(out, "must_flush "); 72 if (buf->separate_flush) 73 prt_str(out, "separate_flush "); 74 if (buf->need_flush_to_write_buffer) 75 prt_str(out, "need_flush_to_write_buffer "); 76 if (buf->write_started) 77 prt_str(out, "write_started "); 78 if (buf->write_allocated) 79 prt_str(out, "write_allocated "); 80 if (buf->write_done) 81 prt_str(out, "write_done"); 82 prt_newline(out); 83 84 printbuf_indent_sub(out, 2); 85 } 86 87 static void bch2_journal_bufs_to_text(struct printbuf *out, struct journal *j) 88 { 89 lockdep_assert_held(&j->lock); 90 out->atomic++; 91 92 if (!out->nr_tabstops) 93 printbuf_tabstop_push(out, 24); 94 95 for (u64 seq = journal_last_unwritten_seq(j); 96 seq <= journal_cur_seq(j); 97 seq++) 98 bch2_journal_buf_to_text(out, j, seq); 99 prt_printf(out, "last buf %s\n", journal_entry_is_open(j) ? "open" : "closed"); 100 101 --out->atomic; 102 } 103 104 static inline struct journal_buf * 105 journal_seq_to_buf(struct journal *j, u64 seq) 106 { 107 struct journal_buf *buf = NULL; 108 109 EBUG_ON(seq > journal_cur_seq(j)); 110 111 if (journal_seq_unwritten(j, seq)) 112 buf = j->buf + (seq & JOURNAL_BUF_MASK); 113 return buf; 114 } 115 116 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count) 117 { 118 for (unsigned i = 0; i < ARRAY_SIZE(p->unflushed); i++) 119 INIT_LIST_HEAD(&p->unflushed[i]); 120 for (unsigned i = 0; i < ARRAY_SIZE(p->flushed); i++) 121 INIT_LIST_HEAD(&p->flushed[i]); 122 atomic_set(&p->count, count); 123 p->devs.nr = 0; 124 } 125 126 /* 127 * Detect stuck journal conditions and trigger shutdown. Technically the journal 128 * can end up stuck for a variety of reasons, such as a blocked I/O, journal 129 * reservation lockup, etc. Since this is a fatal error with potentially 130 * unpredictable characteristics, we want to be fairly conservative before we 131 * decide to shut things down. 132 * 133 * Consider the journal stuck when it appears full with no ability to commit 134 * btree transactions, to discard journal buckets, nor acquire priority 135 * (reserved watermark) reservation. 136 */ 137 static inline bool 138 journal_error_check_stuck(struct journal *j, int error, unsigned flags) 139 { 140 struct bch_fs *c = container_of(j, struct bch_fs, journal); 141 bool stuck = false; 142 struct printbuf buf = PRINTBUF; 143 144 buf.atomic++; 145 146 if (!(error == -BCH_ERR_journal_full || 147 error == -BCH_ERR_journal_pin_full) || 148 nr_unwritten_journal_entries(j) || 149 (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) 150 return stuck; 151 152 spin_lock(&j->lock); 153 154 if (j->can_discard) { 155 spin_unlock(&j->lock); 156 return stuck; 157 } 158 159 stuck = true; 160 161 /* 162 * The journal shutdown path will set ->err_seq, but do it here first to 163 * serialize against concurrent failures and avoid duplicate error 164 * reports. 165 */ 166 if (j->err_seq) { 167 spin_unlock(&j->lock); 168 return stuck; 169 } 170 j->err_seq = journal_cur_seq(j); 171 172 __bch2_journal_debug_to_text(&buf, j); 173 spin_unlock(&j->lock); 174 prt_printf(&buf, bch2_fmt(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)"), 175 bch2_err_str(error)); 176 bch2_print_string_as_lines(KERN_ERR, buf.buf); 177 178 printbuf_reset(&buf); 179 bch2_journal_pins_to_text(&buf, j); 180 bch_err(c, "Journal pins:\n%s", buf.buf); 181 printbuf_exit(&buf); 182 183 bch2_fatal_error(c); 184 dump_stack(); 185 186 return stuck; 187 } 188 189 void bch2_journal_do_writes(struct journal *j) 190 { 191 for (u64 seq = journal_last_unwritten_seq(j); 192 seq <= journal_cur_seq(j); 193 seq++) { 194 unsigned idx = seq & JOURNAL_BUF_MASK; 195 struct journal_buf *w = j->buf + idx; 196 197 if (w->write_started && !w->write_allocated) 198 break; 199 if (w->write_started) 200 continue; 201 202 if (!journal_state_seq_count(j, j->reservations, seq)) { 203 j->seq_write_started = seq; 204 w->write_started = true; 205 closure_call(&w->io, bch2_journal_write, j->wq, NULL); 206 } 207 208 break; 209 } 210 } 211 212 /* 213 * Final processing when the last reference of a journal buffer has been 214 * dropped. Drop the pin list reference acquired at journal entry open and write 215 * the buffer, if requested. 216 */ 217 void bch2_journal_buf_put_final(struct journal *j, u64 seq) 218 { 219 lockdep_assert_held(&j->lock); 220 221 if (__bch2_journal_pin_put(j, seq)) 222 bch2_journal_reclaim_fast(j); 223 bch2_journal_do_writes(j); 224 225 /* 226 * for __bch2_next_write_buffer_flush_journal_buf(), when quiescing an 227 * open journal entry 228 */ 229 wake_up(&j->wait); 230 } 231 232 /* 233 * Returns true if journal entry is now closed: 234 * 235 * We don't close a journal_buf until the next journal_buf is finished writing, 236 * and can be opened again - this also initializes the next journal_buf: 237 */ 238 static void __journal_entry_close(struct journal *j, unsigned closed_val, bool trace) 239 { 240 struct bch_fs *c = container_of(j, struct bch_fs, journal); 241 struct journal_buf *buf = journal_cur_buf(j); 242 union journal_res_state old, new; 243 unsigned sectors; 244 245 BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL && 246 closed_val != JOURNAL_ENTRY_ERROR_VAL); 247 248 lockdep_assert_held(&j->lock); 249 250 old.v = atomic64_read(&j->reservations.counter); 251 do { 252 new.v = old.v; 253 new.cur_entry_offset = closed_val; 254 255 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL || 256 old.cur_entry_offset == new.cur_entry_offset) 257 return; 258 } while (!atomic64_try_cmpxchg(&j->reservations.counter, 259 &old.v, new.v)); 260 261 if (!__journal_entry_is_open(old)) 262 return; 263 264 if (old.cur_entry_offset == JOURNAL_ENTRY_BLOCKED_VAL) 265 old.cur_entry_offset = j->cur_entry_offset_if_blocked; 266 267 /* Close out old buffer: */ 268 buf->data->u64s = cpu_to_le32(old.cur_entry_offset); 269 270 if (trace_journal_entry_close_enabled() && trace) { 271 struct printbuf pbuf = PRINTBUF; 272 pbuf.atomic++; 273 274 prt_str(&pbuf, "entry size: "); 275 prt_human_readable_u64(&pbuf, vstruct_bytes(buf->data)); 276 prt_newline(&pbuf); 277 bch2_prt_task_backtrace(&pbuf, current, 1, GFP_NOWAIT); 278 trace_journal_entry_close(c, pbuf.buf); 279 printbuf_exit(&pbuf); 280 } 281 282 sectors = vstruct_blocks_plus(buf->data, c->block_bits, 283 buf->u64s_reserved) << c->block_bits; 284 if (unlikely(sectors > buf->sectors)) { 285 struct printbuf err = PRINTBUF; 286 err.atomic++; 287 288 prt_printf(&err, "journal entry overran reserved space: %u > %u\n", 289 sectors, buf->sectors); 290 prt_printf(&err, "buf u64s %u u64s reserved %u cur_entry_u64s %u block_bits %u\n", 291 le32_to_cpu(buf->data->u64s), buf->u64s_reserved, 292 j->cur_entry_u64s, 293 c->block_bits); 294 prt_printf(&err, "fatal error - emergency read only"); 295 bch2_journal_halt_locked(j); 296 297 bch_err(c, "%s", err.buf); 298 printbuf_exit(&err); 299 return; 300 } 301 302 buf->sectors = sectors; 303 304 /* 305 * We have to set last_seq here, _before_ opening a new journal entry: 306 * 307 * A threads may replace an old pin with a new pin on their current 308 * journal reservation - the expectation being that the journal will 309 * contain either what the old pin protected or what the new pin 310 * protects. 311 * 312 * After the old pin is dropped journal_last_seq() won't include the old 313 * pin, so we can only write the updated last_seq on the entry that 314 * contains whatever the new pin protects. 315 * 316 * Restated, we can _not_ update last_seq for a given entry if there 317 * could be a newer entry open with reservations/pins that have been 318 * taken against it. 319 * 320 * Hence, we want update/set last_seq on the current journal entry right 321 * before we open a new one: 322 */ 323 buf->last_seq = journal_last_seq(j); 324 buf->data->last_seq = cpu_to_le64(buf->last_seq); 325 BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq)); 326 327 cancel_delayed_work(&j->write_work); 328 329 bch2_journal_space_available(j); 330 331 __bch2_journal_buf_put(j, le64_to_cpu(buf->data->seq)); 332 } 333 334 void bch2_journal_halt(struct journal *j) 335 { 336 spin_lock(&j->lock); 337 __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL, true); 338 if (!j->err_seq) 339 j->err_seq = journal_cur_seq(j); 340 journal_wake(j); 341 spin_unlock(&j->lock); 342 } 343 344 void bch2_journal_halt_locked(struct journal *j) 345 { 346 lockdep_assert_held(&j->lock); 347 348 __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL, true); 349 if (!j->err_seq) 350 j->err_seq = journal_cur_seq(j); 351 journal_wake(j); 352 } 353 354 static bool journal_entry_want_write(struct journal *j) 355 { 356 bool ret = !journal_entry_is_open(j) || 357 journal_cur_seq(j) == journal_last_unwritten_seq(j); 358 359 /* Don't close it yet if we already have a write in flight: */ 360 if (ret) 361 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true); 362 else if (nr_unwritten_journal_entries(j)) { 363 struct journal_buf *buf = journal_cur_buf(j); 364 365 if (!buf->flush_time) { 366 buf->flush_time = local_clock() ?: 1; 367 buf->expires = jiffies; 368 } 369 } 370 371 return ret; 372 } 373 374 bool bch2_journal_entry_close(struct journal *j) 375 { 376 bool ret; 377 378 spin_lock(&j->lock); 379 ret = journal_entry_want_write(j); 380 spin_unlock(&j->lock); 381 382 return ret; 383 } 384 385 /* 386 * should _only_ called from journal_res_get() - when we actually want a 387 * journal reservation - journal entry is open means journal is dirty: 388 */ 389 static int journal_entry_open(struct journal *j) 390 { 391 struct bch_fs *c = container_of(j, struct bch_fs, journal); 392 struct journal_buf *buf = j->buf + 393 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK); 394 union journal_res_state old, new; 395 int u64s; 396 397 lockdep_assert_held(&j->lock); 398 BUG_ON(journal_entry_is_open(j)); 399 BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb)); 400 401 if (j->blocked) 402 return -BCH_ERR_journal_blocked; 403 404 if (j->cur_entry_error) 405 return j->cur_entry_error; 406 407 int ret = bch2_journal_error(j); 408 if (unlikely(ret)) 409 return ret; 410 411 if (!fifo_free(&j->pin)) 412 return -BCH_ERR_journal_pin_full; 413 414 if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf)) 415 return -BCH_ERR_journal_max_in_flight; 416 417 if (atomic64_read(&j->seq) - j->seq_write_started == JOURNAL_STATE_BUF_NR) 418 return -BCH_ERR_journal_max_open; 419 420 if (journal_cur_seq(j) >= JOURNAL_SEQ_MAX) { 421 bch_err(c, "cannot start: journal seq overflow"); 422 if (bch2_fs_emergency_read_only_locked(c)) 423 bch_err(c, "fatal error - emergency read only"); 424 return -BCH_ERR_journal_shutdown; 425 } 426 427 if (!j->free_buf && !buf->data) 428 return -BCH_ERR_journal_buf_enomem; /* will retry after write completion frees up a buf */ 429 430 BUG_ON(!j->cur_entry_sectors); 431 432 if (!buf->data) { 433 swap(buf->data, j->free_buf); 434 swap(buf->buf_size, j->free_buf_size); 435 } 436 437 buf->expires = 438 (journal_cur_seq(j) == j->flushed_seq_ondisk 439 ? jiffies 440 : j->last_flush_write) + 441 msecs_to_jiffies(c->opts.journal_flush_delay); 442 443 buf->u64s_reserved = j->entry_u64s_reserved; 444 buf->disk_sectors = j->cur_entry_sectors; 445 buf->sectors = min(buf->disk_sectors, buf->buf_size >> 9); 446 447 u64s = (int) (buf->sectors << 9) / sizeof(u64) - 448 journal_entry_overhead(j); 449 u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1); 450 451 if (u64s <= (ssize_t) j->early_journal_entries.nr) 452 return -BCH_ERR_journal_full; 453 454 if (fifo_empty(&j->pin) && j->reclaim_thread) 455 wake_up_process(j->reclaim_thread); 456 457 /* 458 * The fifo_push() needs to happen at the same time as j->seq is 459 * incremented for journal_last_seq() to be calculated correctly 460 */ 461 atomic64_inc(&j->seq); 462 journal_pin_list_init(fifo_push_ref(&j->pin), 1); 463 464 BUG_ON(j->pin.back - 1 != atomic64_read(&j->seq)); 465 466 BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf); 467 468 bkey_extent_init(&buf->key); 469 buf->noflush = false; 470 buf->must_flush = false; 471 buf->separate_flush = false; 472 buf->flush_time = 0; 473 buf->need_flush_to_write_buffer = true; 474 buf->write_started = false; 475 buf->write_allocated = false; 476 buf->write_done = false; 477 478 memset(buf->data, 0, sizeof(*buf->data)); 479 buf->data->seq = cpu_to_le64(journal_cur_seq(j)); 480 buf->data->u64s = 0; 481 482 if (j->early_journal_entries.nr) { 483 memcpy(buf->data->_data, j->early_journal_entries.data, 484 j->early_journal_entries.nr * sizeof(u64)); 485 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr); 486 } 487 488 /* 489 * Must be set before marking the journal entry as open: 490 */ 491 j->cur_entry_u64s = u64s; 492 493 old.v = atomic64_read(&j->reservations.counter); 494 do { 495 new.v = old.v; 496 497 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL); 498 499 new.idx++; 500 BUG_ON(journal_state_count(new, new.idx)); 501 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_STATE_BUF_MASK)); 502 503 journal_state_inc(&new); 504 505 /* Handle any already added entries */ 506 new.cur_entry_offset = le32_to_cpu(buf->data->u64s); 507 } while (!atomic64_try_cmpxchg(&j->reservations.counter, 508 &old.v, new.v)); 509 510 if (nr_unwritten_journal_entries(j) == 1) 511 mod_delayed_work(j->wq, 512 &j->write_work, 513 msecs_to_jiffies(c->opts.journal_flush_delay)); 514 journal_wake(j); 515 516 if (j->early_journal_entries.nr) 517 darray_exit(&j->early_journal_entries); 518 return 0; 519 } 520 521 static bool journal_quiesced(struct journal *j) 522 { 523 bool ret = atomic64_read(&j->seq) == j->seq_ondisk; 524 525 if (!ret) 526 bch2_journal_entry_close(j); 527 return ret; 528 } 529 530 static void journal_quiesce(struct journal *j) 531 { 532 wait_event(j->wait, journal_quiesced(j)); 533 } 534 535 static void journal_write_work(struct work_struct *work) 536 { 537 struct journal *j = container_of(work, struct journal, write_work.work); 538 539 spin_lock(&j->lock); 540 if (__journal_entry_is_open(j->reservations)) { 541 long delta = journal_cur_buf(j)->expires - jiffies; 542 543 if (delta > 0) 544 mod_delayed_work(j->wq, &j->write_work, delta); 545 else 546 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true); 547 } 548 spin_unlock(&j->lock); 549 } 550 551 static void journal_buf_prealloc(struct journal *j) 552 { 553 if (j->free_buf && 554 j->free_buf_size >= j->buf_size_want) 555 return; 556 557 unsigned buf_size = j->buf_size_want; 558 559 spin_unlock(&j->lock); 560 void *buf = kvmalloc(buf_size, GFP_NOFS); 561 spin_lock(&j->lock); 562 563 if (buf && 564 (!j->free_buf || 565 buf_size > j->free_buf_size)) { 566 swap(buf, j->free_buf); 567 swap(buf_size, j->free_buf_size); 568 } 569 570 if (unlikely(buf)) { 571 spin_unlock(&j->lock); 572 /* kvfree can sleep */ 573 kvfree(buf); 574 spin_lock(&j->lock); 575 } 576 } 577 578 static int __journal_res_get(struct journal *j, struct journal_res *res, 579 unsigned flags) 580 { 581 struct bch_fs *c = container_of(j, struct bch_fs, journal); 582 struct journal_buf *buf; 583 bool can_discard; 584 int ret; 585 retry: 586 if (journal_res_get_fast(j, res, flags)) 587 return 0; 588 589 ret = bch2_journal_error(j); 590 if (unlikely(ret)) 591 return ret; 592 593 if (j->blocked) 594 return -BCH_ERR_journal_blocked; 595 596 if ((flags & BCH_WATERMARK_MASK) < j->watermark) { 597 ret = -BCH_ERR_journal_full; 598 can_discard = j->can_discard; 599 goto out; 600 } 601 602 if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf) && !journal_entry_is_open(j)) { 603 ret = -BCH_ERR_journal_max_in_flight; 604 goto out; 605 } 606 607 spin_lock(&j->lock); 608 609 journal_buf_prealloc(j); 610 611 /* 612 * Recheck after taking the lock, so we don't race with another thread 613 * that just did journal_entry_open() and call bch2_journal_entry_close() 614 * unnecessarily 615 */ 616 if (journal_res_get_fast(j, res, flags)) { 617 ret = 0; 618 goto unlock; 619 } 620 621 /* 622 * If we couldn't get a reservation because the current buf filled up, 623 * and we had room for a bigger entry on disk, signal that we want to 624 * realloc the journal bufs: 625 */ 626 buf = journal_cur_buf(j); 627 if (journal_entry_is_open(j) && 628 buf->buf_size >> 9 < buf->disk_sectors && 629 buf->buf_size < JOURNAL_ENTRY_SIZE_MAX) 630 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1); 631 632 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, false); 633 ret = journal_entry_open(j) ?: -BCH_ERR_journal_retry_open; 634 unlock: 635 can_discard = j->can_discard; 636 spin_unlock(&j->lock); 637 out: 638 if (likely(!ret)) 639 return 0; 640 if (ret == -BCH_ERR_journal_retry_open) 641 goto retry; 642 643 if (journal_error_check_stuck(j, ret, flags)) 644 ret = -BCH_ERR_journal_stuck; 645 646 if (ret == -BCH_ERR_journal_max_in_flight && 647 track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight], true) && 648 trace_journal_entry_full_enabled()) { 649 struct printbuf buf = PRINTBUF; 650 651 bch2_printbuf_make_room(&buf, 4096); 652 653 spin_lock(&j->lock); 654 prt_printf(&buf, "seq %llu\n", journal_cur_seq(j)); 655 bch2_journal_bufs_to_text(&buf, j); 656 spin_unlock(&j->lock); 657 658 trace_journal_entry_full(c, buf.buf); 659 printbuf_exit(&buf); 660 count_event(c, journal_entry_full); 661 } 662 663 if (ret == -BCH_ERR_journal_max_open && 664 track_event_change(&c->times[BCH_TIME_blocked_journal_max_open], true) && 665 trace_journal_entry_full_enabled()) { 666 struct printbuf buf = PRINTBUF; 667 668 bch2_printbuf_make_room(&buf, 4096); 669 670 spin_lock(&j->lock); 671 prt_printf(&buf, "seq %llu\n", journal_cur_seq(j)); 672 bch2_journal_bufs_to_text(&buf, j); 673 spin_unlock(&j->lock); 674 675 trace_journal_entry_full(c, buf.buf); 676 printbuf_exit(&buf); 677 count_event(c, journal_entry_full); 678 } 679 680 /* 681 * Journal is full - can't rely on reclaim from work item due to 682 * freezing: 683 */ 684 if ((ret == -BCH_ERR_journal_full || 685 ret == -BCH_ERR_journal_pin_full) && 686 !(flags & JOURNAL_RES_GET_NONBLOCK)) { 687 if (can_discard) { 688 bch2_journal_do_discards(j); 689 goto retry; 690 } 691 692 if (mutex_trylock(&j->reclaim_lock)) { 693 bch2_journal_reclaim(j); 694 mutex_unlock(&j->reclaim_lock); 695 } 696 } 697 698 return ret; 699 } 700 701 static unsigned max_dev_latency(struct bch_fs *c) 702 { 703 u64 nsecs = 0; 704 705 for_each_rw_member(c, ca) 706 nsecs = max(nsecs, ca->io_latency[WRITE].stats.max_duration); 707 708 return nsecs_to_jiffies(nsecs); 709 } 710 711 /* 712 * Essentially the entry function to the journaling code. When bcachefs is doing 713 * a btree insert, it calls this function to get the current journal write. 714 * Journal write is the structure used set up journal writes. The calling 715 * function will then add its keys to the structure, queuing them for the next 716 * write. 717 * 718 * To ensure forward progress, the current task must not be holding any 719 * btree node write locks. 720 */ 721 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res, 722 unsigned flags, 723 struct btree_trans *trans) 724 { 725 int ret; 726 727 if (closure_wait_event_timeout(&j->async_wait, 728 !bch2_err_matches(ret = __journal_res_get(j, res, flags), BCH_ERR_operation_blocked) || 729 (flags & JOURNAL_RES_GET_NONBLOCK), 730 HZ)) 731 return ret; 732 733 if (trans) 734 bch2_trans_unlock_long(trans); 735 736 struct bch_fs *c = container_of(j, struct bch_fs, journal); 737 int remaining_wait = max(max_dev_latency(c) * 2, HZ * 10); 738 739 remaining_wait = max(0, remaining_wait - HZ); 740 741 if (closure_wait_event_timeout(&j->async_wait, 742 !bch2_err_matches(ret = __journal_res_get(j, res, flags), BCH_ERR_operation_blocked) || 743 (flags & JOURNAL_RES_GET_NONBLOCK), 744 remaining_wait)) 745 return ret; 746 747 struct printbuf buf = PRINTBUF; 748 bch2_journal_debug_to_text(&buf, j); 749 bch2_print_string_as_lines(KERN_ERR, buf.buf); 750 prt_printf(&buf, bch2_fmt(c, "Journal stuck? Waited for 10 seconds, err %s"), bch2_err_str(ret)); 751 printbuf_exit(&buf); 752 753 closure_wait_event(&j->async_wait, 754 !bch2_err_matches(ret = __journal_res_get(j, res, flags), BCH_ERR_operation_blocked) || 755 (flags & JOURNAL_RES_GET_NONBLOCK)); 756 return ret; 757 } 758 759 /* journal_entry_res: */ 760 761 void bch2_journal_entry_res_resize(struct journal *j, 762 struct journal_entry_res *res, 763 unsigned new_u64s) 764 { 765 union journal_res_state state; 766 int d = new_u64s - res->u64s; 767 768 spin_lock(&j->lock); 769 770 j->entry_u64s_reserved += d; 771 if (d <= 0) 772 goto out; 773 774 j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d); 775 state = READ_ONCE(j->reservations); 776 777 if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL && 778 state.cur_entry_offset > j->cur_entry_u64s) { 779 j->cur_entry_u64s += d; 780 /* 781 * Not enough room in current journal entry, have to flush it: 782 */ 783 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true); 784 } else { 785 journal_cur_buf(j)->u64s_reserved += d; 786 } 787 out: 788 spin_unlock(&j->lock); 789 res->u64s += d; 790 } 791 792 /* journal flushing: */ 793 794 /** 795 * bch2_journal_flush_seq_async - wait for a journal entry to be written 796 * @j: journal object 797 * @seq: seq to flush 798 * @parent: closure object to wait with 799 * Returns: 1 if @seq has already been flushed, 0 if @seq is being flushed, 800 * -BCH_ERR_journal_flush_err if @seq will never be flushed 801 * 802 * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if 803 * necessary 804 */ 805 int bch2_journal_flush_seq_async(struct journal *j, u64 seq, 806 struct closure *parent) 807 { 808 struct journal_buf *buf; 809 int ret = 0; 810 811 if (seq <= j->flushed_seq_ondisk) 812 return 1; 813 814 spin_lock(&j->lock); 815 816 if (WARN_ONCE(seq > journal_cur_seq(j), 817 "requested to flush journal seq %llu, but currently at %llu", 818 seq, journal_cur_seq(j))) 819 goto out; 820 821 /* Recheck under lock: */ 822 if (j->err_seq && seq >= j->err_seq) { 823 ret = -BCH_ERR_journal_flush_err; 824 goto out; 825 } 826 827 if (seq <= j->flushed_seq_ondisk) { 828 ret = 1; 829 goto out; 830 } 831 832 /* if seq was written, but not flushed - flush a newer one instead */ 833 seq = max(seq, journal_last_unwritten_seq(j)); 834 835 recheck_need_open: 836 if (seq > journal_cur_seq(j)) { 837 struct journal_res res = { 0 }; 838 839 if (journal_entry_is_open(j)) 840 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true); 841 842 spin_unlock(&j->lock); 843 844 /* 845 * We're called from bch2_journal_flush_seq() -> wait_event(); 846 * but this might block. We won't usually block, so we won't 847 * livelock: 848 */ 849 sched_annotate_sleep(); 850 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0, NULL); 851 if (ret) 852 return ret; 853 854 seq = res.seq; 855 buf = journal_seq_to_buf(j, seq); 856 buf->must_flush = true; 857 858 if (!buf->flush_time) { 859 buf->flush_time = local_clock() ?: 1; 860 buf->expires = jiffies; 861 } 862 863 if (parent && !closure_wait(&buf->wait, parent)) 864 BUG(); 865 866 bch2_journal_res_put(j, &res); 867 868 spin_lock(&j->lock); 869 goto want_write; 870 } 871 872 /* 873 * if write was kicked off without a flush, or if we promised it 874 * wouldn't be a flush, flush the next sequence number instead 875 */ 876 buf = journal_seq_to_buf(j, seq); 877 if (buf->noflush) { 878 seq++; 879 goto recheck_need_open; 880 } 881 882 buf->must_flush = true; 883 j->flushing_seq = max(j->flushing_seq, seq); 884 885 if (parent && !closure_wait(&buf->wait, parent)) 886 BUG(); 887 want_write: 888 if (seq == journal_cur_seq(j)) 889 journal_entry_want_write(j); 890 out: 891 spin_unlock(&j->lock); 892 return ret; 893 } 894 895 int bch2_journal_flush_seq(struct journal *j, u64 seq, unsigned task_state) 896 { 897 u64 start_time = local_clock(); 898 int ret, ret2; 899 900 /* 901 * Don't update time_stats when @seq is already flushed: 902 */ 903 if (seq <= j->flushed_seq_ondisk) 904 return 0; 905 906 ret = wait_event_state(j->wait, 907 (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)), 908 task_state); 909 910 if (!ret) 911 bch2_time_stats_update(j->flush_seq_time, start_time); 912 913 return ret ?: ret2 < 0 ? ret2 : 0; 914 } 915 916 /* 917 * bch2_journal_flush_async - if there is an open journal entry, or a journal 918 * still being written, write it and wait for the write to complete 919 */ 920 void bch2_journal_flush_async(struct journal *j, struct closure *parent) 921 { 922 bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent); 923 } 924 925 int bch2_journal_flush(struct journal *j) 926 { 927 return bch2_journal_flush_seq(j, atomic64_read(&j->seq), TASK_UNINTERRUPTIBLE); 928 } 929 930 /* 931 * bch2_journal_noflush_seq - ask the journal not to issue any flushes in the 932 * range [start, end) 933 * @seq 934 */ 935 bool bch2_journal_noflush_seq(struct journal *j, u64 start, u64 end) 936 { 937 struct bch_fs *c = container_of(j, struct bch_fs, journal); 938 u64 unwritten_seq; 939 bool ret = false; 940 941 if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush))) 942 return false; 943 944 if (c->journal.flushed_seq_ondisk >= start) 945 return false; 946 947 spin_lock(&j->lock); 948 if (c->journal.flushed_seq_ondisk >= start) 949 goto out; 950 951 for (unwritten_seq = journal_last_unwritten_seq(j); 952 unwritten_seq < end; 953 unwritten_seq++) { 954 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq); 955 956 /* journal flush already in flight, or flush requseted */ 957 if (buf->must_flush) 958 goto out; 959 960 buf->noflush = true; 961 } 962 963 ret = true; 964 out: 965 spin_unlock(&j->lock); 966 return ret; 967 } 968 969 static int __bch2_journal_meta(struct journal *j) 970 { 971 struct journal_res res = {}; 972 int ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0, NULL); 973 if (ret) 974 return ret; 975 976 struct journal_buf *buf = j->buf + (res.seq & JOURNAL_BUF_MASK); 977 buf->must_flush = true; 978 979 if (!buf->flush_time) { 980 buf->flush_time = local_clock() ?: 1; 981 buf->expires = jiffies; 982 } 983 984 bch2_journal_res_put(j, &res); 985 986 return bch2_journal_flush_seq(j, res.seq, TASK_UNINTERRUPTIBLE); 987 } 988 989 int bch2_journal_meta(struct journal *j) 990 { 991 struct bch_fs *c = container_of(j, struct bch_fs, journal); 992 993 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_journal)) 994 return -BCH_ERR_erofs_no_writes; 995 996 int ret = __bch2_journal_meta(j); 997 bch2_write_ref_put(c, BCH_WRITE_REF_journal); 998 return ret; 999 } 1000 1001 /* block/unlock the journal: */ 1002 1003 void bch2_journal_unblock(struct journal *j) 1004 { 1005 spin_lock(&j->lock); 1006 if (!--j->blocked && 1007 j->cur_entry_offset_if_blocked < JOURNAL_ENTRY_CLOSED_VAL && 1008 j->reservations.cur_entry_offset == JOURNAL_ENTRY_BLOCKED_VAL) { 1009 union journal_res_state old, new; 1010 1011 old.v = atomic64_read(&j->reservations.counter); 1012 do { 1013 new.v = old.v; 1014 new.cur_entry_offset = j->cur_entry_offset_if_blocked; 1015 } while (!atomic64_try_cmpxchg(&j->reservations.counter, &old.v, new.v)); 1016 } 1017 spin_unlock(&j->lock); 1018 1019 journal_wake(j); 1020 } 1021 1022 static void __bch2_journal_block(struct journal *j) 1023 { 1024 if (!j->blocked++) { 1025 union journal_res_state old, new; 1026 1027 old.v = atomic64_read(&j->reservations.counter); 1028 do { 1029 j->cur_entry_offset_if_blocked = old.cur_entry_offset; 1030 1031 if (j->cur_entry_offset_if_blocked >= JOURNAL_ENTRY_CLOSED_VAL) 1032 break; 1033 1034 new.v = old.v; 1035 new.cur_entry_offset = JOURNAL_ENTRY_BLOCKED_VAL; 1036 } while (!atomic64_try_cmpxchg(&j->reservations.counter, &old.v, new.v)); 1037 1038 if (old.cur_entry_offset < JOURNAL_ENTRY_BLOCKED_VAL) 1039 journal_cur_buf(j)->data->u64s = cpu_to_le32(old.cur_entry_offset); 1040 } 1041 } 1042 1043 void bch2_journal_block(struct journal *j) 1044 { 1045 spin_lock(&j->lock); 1046 __bch2_journal_block(j); 1047 spin_unlock(&j->lock); 1048 1049 journal_quiesce(j); 1050 } 1051 1052 static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j, 1053 u64 max_seq, bool *blocked) 1054 { 1055 struct journal_buf *ret = NULL; 1056 1057 /* We're inside wait_event(), but using mutex_lock(: */ 1058 sched_annotate_sleep(); 1059 mutex_lock(&j->buf_lock); 1060 spin_lock(&j->lock); 1061 max_seq = min(max_seq, journal_cur_seq(j)); 1062 1063 for (u64 seq = journal_last_unwritten_seq(j); 1064 seq <= max_seq; 1065 seq++) { 1066 unsigned idx = seq & JOURNAL_BUF_MASK; 1067 struct journal_buf *buf = j->buf + idx; 1068 1069 if (buf->need_flush_to_write_buffer) { 1070 union journal_res_state s; 1071 s.v = atomic64_read_acquire(&j->reservations.counter); 1072 1073 unsigned open = seq == journal_cur_seq(j) && __journal_entry_is_open(s); 1074 1075 if (open && !*blocked) { 1076 __bch2_journal_block(j); 1077 *blocked = true; 1078 } 1079 1080 ret = journal_state_count(s, idx & JOURNAL_STATE_BUF_MASK) > open 1081 ? ERR_PTR(-EAGAIN) 1082 : buf; 1083 break; 1084 } 1085 } 1086 1087 spin_unlock(&j->lock); 1088 if (IS_ERR_OR_NULL(ret)) 1089 mutex_unlock(&j->buf_lock); 1090 return ret; 1091 } 1092 1093 struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, 1094 u64 max_seq, bool *blocked) 1095 { 1096 struct journal_buf *ret; 1097 *blocked = false; 1098 1099 wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j, 1100 max_seq, blocked)) != ERR_PTR(-EAGAIN)); 1101 if (IS_ERR_OR_NULL(ret) && *blocked) 1102 bch2_journal_unblock(j); 1103 1104 return ret; 1105 } 1106 1107 /* allocate journal on a device: */ 1108 1109 static int bch2_set_nr_journal_buckets_iter(struct bch_dev *ca, unsigned nr, 1110 bool new_fs, struct closure *cl) 1111 { 1112 struct bch_fs *c = ca->fs; 1113 struct journal_device *ja = &ca->journal; 1114 u64 *new_bucket_seq = NULL, *new_buckets = NULL; 1115 struct open_bucket **ob = NULL; 1116 long *bu = NULL; 1117 unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr; 1118 int ret = 0; 1119 1120 BUG_ON(nr <= ja->nr); 1121 1122 bu = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL); 1123 ob = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL); 1124 new_buckets = kcalloc(nr, sizeof(u64), GFP_KERNEL); 1125 new_bucket_seq = kcalloc(nr, sizeof(u64), GFP_KERNEL); 1126 if (!bu || !ob || !new_buckets || !new_bucket_seq) { 1127 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets; 1128 goto err_free; 1129 } 1130 1131 for (nr_got = 0; nr_got < nr_want; nr_got++) { 1132 enum bch_watermark watermark = new_fs 1133 ? BCH_WATERMARK_btree 1134 : BCH_WATERMARK_normal; 1135 1136 ob[nr_got] = bch2_bucket_alloc(c, ca, watermark, 1137 BCH_DATA_journal, cl); 1138 ret = PTR_ERR_OR_ZERO(ob[nr_got]); 1139 if (ret) 1140 break; 1141 1142 if (!new_fs) { 1143 ret = bch2_trans_run(c, 1144 bch2_trans_mark_metadata_bucket(trans, ca, 1145 ob[nr_got]->bucket, BCH_DATA_journal, 1146 ca->mi.bucket_size, BTREE_TRIGGER_transactional)); 1147 if (ret) { 1148 bch2_open_bucket_put(c, ob[nr_got]); 1149 bch_err_msg(c, ret, "marking new journal buckets"); 1150 break; 1151 } 1152 } 1153 1154 bu[nr_got] = ob[nr_got]->bucket; 1155 } 1156 1157 if (!nr_got) 1158 goto err_free; 1159 1160 /* Don't return an error if we successfully allocated some buckets: */ 1161 ret = 0; 1162 1163 if (c) { 1164 bch2_journal_flush_all_pins(&c->journal); 1165 bch2_journal_block(&c->journal); 1166 mutex_lock(&c->sb_lock); 1167 } 1168 1169 memcpy(new_buckets, ja->buckets, ja->nr * sizeof(u64)); 1170 memcpy(new_bucket_seq, ja->bucket_seq, ja->nr * sizeof(u64)); 1171 1172 BUG_ON(ja->discard_idx > ja->nr); 1173 1174 pos = ja->discard_idx ?: ja->nr; 1175 1176 memmove(new_buckets + pos + nr_got, 1177 new_buckets + pos, 1178 sizeof(new_buckets[0]) * (ja->nr - pos)); 1179 memmove(new_bucket_seq + pos + nr_got, 1180 new_bucket_seq + pos, 1181 sizeof(new_bucket_seq[0]) * (ja->nr - pos)); 1182 1183 for (i = 0; i < nr_got; i++) { 1184 new_buckets[pos + i] = bu[i]; 1185 new_bucket_seq[pos + i] = 0; 1186 } 1187 1188 nr = ja->nr + nr_got; 1189 1190 ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr); 1191 if (ret) 1192 goto err_unblock; 1193 1194 bch2_write_super(c); 1195 1196 /* Commit: */ 1197 if (c) 1198 spin_lock(&c->journal.lock); 1199 1200 swap(new_buckets, ja->buckets); 1201 swap(new_bucket_seq, ja->bucket_seq); 1202 ja->nr = nr; 1203 1204 if (pos <= ja->discard_idx) 1205 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr; 1206 if (pos <= ja->dirty_idx_ondisk) 1207 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr; 1208 if (pos <= ja->dirty_idx) 1209 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr; 1210 if (pos <= ja->cur_idx) 1211 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr; 1212 1213 if (c) 1214 spin_unlock(&c->journal.lock); 1215 err_unblock: 1216 if (c) { 1217 bch2_journal_unblock(&c->journal); 1218 mutex_unlock(&c->sb_lock); 1219 } 1220 1221 if (ret && !new_fs) 1222 for (i = 0; i < nr_got; i++) 1223 bch2_trans_run(c, 1224 bch2_trans_mark_metadata_bucket(trans, ca, 1225 bu[i], BCH_DATA_free, 0, 1226 BTREE_TRIGGER_transactional)); 1227 err_free: 1228 for (i = 0; i < nr_got; i++) 1229 bch2_open_bucket_put(c, ob[i]); 1230 1231 kfree(new_bucket_seq); 1232 kfree(new_buckets); 1233 kfree(ob); 1234 kfree(bu); 1235 return ret; 1236 } 1237 1238 static int bch2_set_nr_journal_buckets_loop(struct bch_fs *c, struct bch_dev *ca, 1239 unsigned nr, bool new_fs) 1240 { 1241 struct journal_device *ja = &ca->journal; 1242 int ret = 0; 1243 1244 struct closure cl; 1245 closure_init_stack(&cl); 1246 1247 /* don't handle reducing nr of buckets yet: */ 1248 if (nr < ja->nr) 1249 return 0; 1250 1251 while (!ret && ja->nr < nr) { 1252 struct disk_reservation disk_res = { 0, 0, 0 }; 1253 1254 /* 1255 * note: journal buckets aren't really counted as _sectors_ used yet, so 1256 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c 1257 * when space used goes up without a reservation - but we do need the 1258 * reservation to ensure we'll actually be able to allocate: 1259 * 1260 * XXX: that's not right, disk reservations only ensure a 1261 * filesystem-wide allocation will succeed, this is a device 1262 * specific allocation - we can hang here: 1263 */ 1264 if (!new_fs) { 1265 ret = bch2_disk_reservation_get(c, &disk_res, 1266 bucket_to_sector(ca, nr - ja->nr), 1, 0); 1267 if (ret) 1268 break; 1269 } 1270 1271 ret = bch2_set_nr_journal_buckets_iter(ca, nr, new_fs, &cl); 1272 1273 if (ret == -BCH_ERR_bucket_alloc_blocked || 1274 ret == -BCH_ERR_open_buckets_empty) 1275 ret = 0; /* wait and retry */ 1276 1277 bch2_disk_reservation_put(c, &disk_res); 1278 closure_sync(&cl); 1279 } 1280 1281 return ret; 1282 } 1283 1284 /* 1285 * Allocate more journal space at runtime - not currently making use if it, but 1286 * the code works: 1287 */ 1288 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca, 1289 unsigned nr) 1290 { 1291 down_write(&c->state_lock); 1292 int ret = bch2_set_nr_journal_buckets_loop(c, ca, nr, false); 1293 up_write(&c->state_lock); 1294 1295 bch_err_fn(c, ret); 1296 return ret; 1297 } 1298 1299 int bch2_dev_journal_alloc(struct bch_dev *ca, bool new_fs) 1300 { 1301 unsigned nr; 1302 int ret; 1303 1304 if (dynamic_fault("bcachefs:add:journal_alloc")) { 1305 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets; 1306 goto err; 1307 } 1308 1309 /* 1/128th of the device by default: */ 1310 nr = ca->mi.nbuckets >> 7; 1311 1312 /* 1313 * clamp journal size to 8192 buckets or 8GB (in sectors), whichever 1314 * is smaller: 1315 */ 1316 nr = clamp_t(unsigned, nr, 1317 BCH_JOURNAL_BUCKETS_MIN, 1318 min(1 << 13, 1319 (1 << 24) / ca->mi.bucket_size)); 1320 1321 ret = bch2_set_nr_journal_buckets_loop(ca->fs, ca, nr, new_fs); 1322 err: 1323 bch_err_fn(ca, ret); 1324 return ret; 1325 } 1326 1327 int bch2_fs_journal_alloc(struct bch_fs *c) 1328 { 1329 for_each_online_member(c, ca) { 1330 if (ca->journal.nr) 1331 continue; 1332 1333 int ret = bch2_dev_journal_alloc(ca, true); 1334 if (ret) { 1335 percpu_ref_put(&ca->io_ref[READ]); 1336 return ret; 1337 } 1338 } 1339 1340 return 0; 1341 } 1342 1343 /* startup/shutdown: */ 1344 1345 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx) 1346 { 1347 bool ret = false; 1348 u64 seq; 1349 1350 spin_lock(&j->lock); 1351 for (seq = journal_last_unwritten_seq(j); 1352 seq <= journal_cur_seq(j) && !ret; 1353 seq++) { 1354 struct journal_buf *buf = journal_seq_to_buf(j, seq); 1355 1356 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx)) 1357 ret = true; 1358 } 1359 spin_unlock(&j->lock); 1360 1361 return ret; 1362 } 1363 1364 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca) 1365 { 1366 wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx)); 1367 } 1368 1369 void bch2_fs_journal_stop(struct journal *j) 1370 { 1371 if (!test_bit(JOURNAL_running, &j->flags)) 1372 return; 1373 1374 bch2_journal_reclaim_stop(j); 1375 bch2_journal_flush_all_pins(j); 1376 1377 wait_event(j->wait, bch2_journal_entry_close(j)); 1378 1379 /* 1380 * Always write a new journal entry, to make sure the clock hands are up 1381 * to date (and match the superblock) 1382 */ 1383 __bch2_journal_meta(j); 1384 1385 journal_quiesce(j); 1386 cancel_delayed_work_sync(&j->write_work); 1387 1388 WARN(!bch2_journal_error(j) && 1389 test_bit(JOURNAL_replay_done, &j->flags) && 1390 j->last_empty_seq != journal_cur_seq(j), 1391 "journal shutdown error: cur seq %llu but last empty seq %llu", 1392 journal_cur_seq(j), j->last_empty_seq); 1393 1394 if (!bch2_journal_error(j)) 1395 clear_bit(JOURNAL_running, &j->flags); 1396 } 1397 1398 int bch2_fs_journal_start(struct journal *j, u64 cur_seq) 1399 { 1400 struct bch_fs *c = container_of(j, struct bch_fs, journal); 1401 struct journal_entry_pin_list *p; 1402 struct journal_replay *i, **_i; 1403 struct genradix_iter iter; 1404 bool had_entries = false; 1405 u64 last_seq = cur_seq, nr, seq; 1406 1407 if (cur_seq >= JOURNAL_SEQ_MAX) { 1408 bch_err(c, "cannot start: journal seq overflow"); 1409 return -EINVAL; 1410 } 1411 1412 genradix_for_each_reverse(&c->journal_entries, iter, _i) { 1413 i = *_i; 1414 1415 if (journal_replay_ignore(i)) 1416 continue; 1417 1418 last_seq = le64_to_cpu(i->j.last_seq); 1419 break; 1420 } 1421 1422 nr = cur_seq - last_seq; 1423 1424 /* 1425 * Extra fudge factor, in case we crashed when the journal pin fifo was 1426 * nearly or completely full. We'll need to be able to open additional 1427 * journal entries (at least a few) in order for journal replay to get 1428 * going: 1429 */ 1430 nr += nr / 4; 1431 1432 if (nr + 1 > j->pin.size) { 1433 free_fifo(&j->pin); 1434 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL); 1435 if (!j->pin.data) { 1436 bch_err(c, "error reallocating journal fifo (%llu open entries)", nr); 1437 return -BCH_ERR_ENOMEM_journal_pin_fifo; 1438 } 1439 } 1440 1441 j->replay_journal_seq = last_seq; 1442 j->replay_journal_seq_end = cur_seq; 1443 j->last_seq_ondisk = last_seq; 1444 j->flushed_seq_ondisk = cur_seq - 1; 1445 j->seq_write_started = cur_seq - 1; 1446 j->seq_ondisk = cur_seq - 1; 1447 j->pin.front = last_seq; 1448 j->pin.back = cur_seq; 1449 atomic64_set(&j->seq, cur_seq - 1); 1450 1451 fifo_for_each_entry_ptr(p, &j->pin, seq) 1452 journal_pin_list_init(p, 1); 1453 1454 genradix_for_each(&c->journal_entries, iter, _i) { 1455 i = *_i; 1456 1457 if (journal_replay_ignore(i)) 1458 continue; 1459 1460 seq = le64_to_cpu(i->j.seq); 1461 BUG_ON(seq >= cur_seq); 1462 1463 if (seq < last_seq) 1464 continue; 1465 1466 if (journal_entry_empty(&i->j)) 1467 j->last_empty_seq = le64_to_cpu(i->j.seq); 1468 1469 p = journal_seq_pin(j, seq); 1470 1471 p->devs.nr = 0; 1472 darray_for_each(i->ptrs, ptr) 1473 bch2_dev_list_add_dev(&p->devs, ptr->dev); 1474 1475 had_entries = true; 1476 } 1477 1478 if (!had_entries) 1479 j->last_empty_seq = cur_seq - 1; /* to match j->seq */ 1480 1481 spin_lock(&j->lock); 1482 j->last_flush_write = jiffies; 1483 1484 j->reservations.idx = journal_cur_seq(j); 1485 1486 c->last_bucket_seq_cleanup = journal_cur_seq(j); 1487 spin_unlock(&j->lock); 1488 1489 return 0; 1490 } 1491 1492 void bch2_journal_set_replay_done(struct journal *j) 1493 { 1494 /* 1495 * journal_space_available must happen before setting JOURNAL_running 1496 * JOURNAL_running must happen before JOURNAL_replay_done 1497 */ 1498 spin_lock(&j->lock); 1499 bch2_journal_space_available(j); 1500 1501 set_bit(JOURNAL_need_flush_write, &j->flags); 1502 set_bit(JOURNAL_running, &j->flags); 1503 set_bit(JOURNAL_replay_done, &j->flags); 1504 spin_unlock(&j->lock); 1505 } 1506 1507 /* init/exit: */ 1508 1509 void bch2_dev_journal_exit(struct bch_dev *ca) 1510 { 1511 struct journal_device *ja = &ca->journal; 1512 1513 for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) { 1514 kfree(ja->bio[i]); 1515 ja->bio[i] = NULL; 1516 } 1517 1518 kfree(ja->buckets); 1519 kfree(ja->bucket_seq); 1520 ja->buckets = NULL; 1521 ja->bucket_seq = NULL; 1522 } 1523 1524 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb) 1525 { 1526 struct journal_device *ja = &ca->journal; 1527 struct bch_sb_field_journal *journal_buckets = 1528 bch2_sb_field_get(sb, journal); 1529 struct bch_sb_field_journal_v2 *journal_buckets_v2 = 1530 bch2_sb_field_get(sb, journal_v2); 1531 1532 ja->nr = 0; 1533 1534 if (journal_buckets_v2) { 1535 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2); 1536 1537 for (unsigned i = 0; i < nr; i++) 1538 ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr); 1539 } else if (journal_buckets) { 1540 ja->nr = bch2_nr_journal_buckets(journal_buckets); 1541 } 1542 1543 ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL); 1544 if (!ja->bucket_seq) 1545 return -BCH_ERR_ENOMEM_dev_journal_init; 1546 1547 unsigned nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE); 1548 1549 for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) { 1550 ja->bio[i] = kzalloc(struct_size(ja->bio[i], bio.bi_inline_vecs, 1551 nr_bvecs), GFP_KERNEL); 1552 if (!ja->bio[i]) 1553 return -BCH_ERR_ENOMEM_dev_journal_init; 1554 1555 ja->bio[i]->ca = ca; 1556 ja->bio[i]->buf_idx = i; 1557 bio_init(&ja->bio[i]->bio, NULL, ja->bio[i]->bio.bi_inline_vecs, nr_bvecs, 0); 1558 } 1559 1560 ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL); 1561 if (!ja->buckets) 1562 return -BCH_ERR_ENOMEM_dev_journal_init; 1563 1564 if (journal_buckets_v2) { 1565 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2); 1566 unsigned dst = 0; 1567 1568 for (unsigned i = 0; i < nr; i++) 1569 for (unsigned j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++) 1570 ja->buckets[dst++] = 1571 le64_to_cpu(journal_buckets_v2->d[i].start) + j; 1572 } else if (journal_buckets) { 1573 for (unsigned i = 0; i < ja->nr; i++) 1574 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]); 1575 } 1576 1577 return 0; 1578 } 1579 1580 void bch2_fs_journal_exit(struct journal *j) 1581 { 1582 if (j->wq) 1583 destroy_workqueue(j->wq); 1584 1585 darray_exit(&j->early_journal_entries); 1586 1587 for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++) 1588 kvfree(j->buf[i].data); 1589 kvfree(j->free_buf); 1590 free_fifo(&j->pin); 1591 } 1592 1593 int bch2_fs_journal_init(struct journal *j) 1594 { 1595 static struct lock_class_key res_key; 1596 1597 mutex_init(&j->buf_lock); 1598 spin_lock_init(&j->lock); 1599 spin_lock_init(&j->err_lock); 1600 init_waitqueue_head(&j->wait); 1601 INIT_DELAYED_WORK(&j->write_work, journal_write_work); 1602 init_waitqueue_head(&j->reclaim_wait); 1603 init_waitqueue_head(&j->pin_flush_wait); 1604 mutex_init(&j->reclaim_lock); 1605 mutex_init(&j->discard_lock); 1606 1607 lockdep_init_map(&j->res_map, "journal res", &res_key, 0); 1608 1609 atomic64_set(&j->reservations.counter, 1610 ((union journal_res_state) 1611 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v); 1612 1613 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL))) 1614 return -BCH_ERR_ENOMEM_journal_pin_fifo; 1615 1616 j->free_buf_size = j->buf_size_want = JOURNAL_ENTRY_SIZE_MIN; 1617 j->free_buf = kvmalloc(j->free_buf_size, GFP_KERNEL); 1618 if (!j->free_buf) 1619 return -BCH_ERR_ENOMEM_journal_buf; 1620 1621 for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++) 1622 j->buf[i].idx = i; 1623 1624 j->pin.front = j->pin.back = 1; 1625 1626 j->wq = alloc_workqueue("bcachefs_journal", 1627 WQ_HIGHPRI|WQ_FREEZABLE|WQ_UNBOUND|WQ_MEM_RECLAIM, 512); 1628 if (!j->wq) 1629 return -BCH_ERR_ENOMEM_fs_other_alloc; 1630 return 0; 1631 } 1632 1633 /* debug: */ 1634 1635 static const char * const bch2_journal_flags_strs[] = { 1636 #define x(n) #n, 1637 JOURNAL_FLAGS() 1638 #undef x 1639 NULL 1640 }; 1641 1642 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j) 1643 { 1644 struct bch_fs *c = container_of(j, struct bch_fs, journal); 1645 union journal_res_state s; 1646 unsigned long now = jiffies; 1647 u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes; 1648 1649 printbuf_tabstops_reset(out); 1650 printbuf_tabstop_push(out, 28); 1651 out->atomic++; 1652 1653 rcu_read_lock(); 1654 s = READ_ONCE(j->reservations); 1655 1656 prt_printf(out, "flags:\t"); 1657 prt_bitflags(out, bch2_journal_flags_strs, j->flags); 1658 prt_newline(out); 1659 prt_printf(out, "dirty journal entries:\t%llu/%llu\n", fifo_used(&j->pin), j->pin.size); 1660 prt_printf(out, "seq:\t%llu\n", journal_cur_seq(j)); 1661 prt_printf(out, "seq_ondisk:\t%llu\n", j->seq_ondisk); 1662 prt_printf(out, "last_seq:\t%llu\n", journal_last_seq(j)); 1663 prt_printf(out, "last_seq_ondisk:\t%llu\n", j->last_seq_ondisk); 1664 prt_printf(out, "flushed_seq_ondisk:\t%llu\n", j->flushed_seq_ondisk); 1665 prt_printf(out, "watermark:\t%s\n", bch2_watermarks[j->watermark]); 1666 prt_printf(out, "each entry reserved:\t%u\n", j->entry_u64s_reserved); 1667 prt_printf(out, "nr flush writes:\t%llu\n", j->nr_flush_writes); 1668 prt_printf(out, "nr noflush writes:\t%llu\n", j->nr_noflush_writes); 1669 prt_printf(out, "average write size:\t"); 1670 prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0); 1671 prt_newline(out); 1672 prt_printf(out, "free buf:\t%u\n", j->free_buf ? j->free_buf_size : 0); 1673 prt_printf(out, "nr direct reclaim:\t%llu\n", j->nr_direct_reclaim); 1674 prt_printf(out, "nr background reclaim:\t%llu\n", j->nr_background_reclaim); 1675 prt_printf(out, "reclaim kicked:\t%u\n", j->reclaim_kicked); 1676 prt_printf(out, "reclaim runs in:\t%u ms\n", time_after(j->next_reclaim, now) 1677 ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0); 1678 prt_printf(out, "blocked:\t%u\n", j->blocked); 1679 prt_printf(out, "current entry sectors:\t%u\n", j->cur_entry_sectors); 1680 prt_printf(out, "current entry error:\t%s\n", bch2_err_str(j->cur_entry_error)); 1681 prt_printf(out, "current entry:\t"); 1682 1683 switch (s.cur_entry_offset) { 1684 case JOURNAL_ENTRY_ERROR_VAL: 1685 prt_printf(out, "error\n"); 1686 break; 1687 case JOURNAL_ENTRY_CLOSED_VAL: 1688 prt_printf(out, "closed\n"); 1689 break; 1690 case JOURNAL_ENTRY_BLOCKED_VAL: 1691 prt_printf(out, "blocked\n"); 1692 break; 1693 default: 1694 prt_printf(out, "%u/%u\n", s.cur_entry_offset, j->cur_entry_u64s); 1695 break; 1696 } 1697 1698 prt_printf(out, "unwritten entries:\n"); 1699 bch2_journal_bufs_to_text(out, j); 1700 1701 prt_printf(out, "space:\n"); 1702 printbuf_indent_add(out, 2); 1703 prt_printf(out, "discarded\t%u:%u\n", 1704 j->space[journal_space_discarded].next_entry, 1705 j->space[journal_space_discarded].total); 1706 prt_printf(out, "clean ondisk\t%u:%u\n", 1707 j->space[journal_space_clean_ondisk].next_entry, 1708 j->space[journal_space_clean_ondisk].total); 1709 prt_printf(out, "clean\t%u:%u\n", 1710 j->space[journal_space_clean].next_entry, 1711 j->space[journal_space_clean].total); 1712 prt_printf(out, "total\t%u:%u\n", 1713 j->space[journal_space_total].next_entry, 1714 j->space[journal_space_total].total); 1715 printbuf_indent_sub(out, 2); 1716 1717 for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) { 1718 if (!ca->mi.durability) 1719 continue; 1720 1721 struct journal_device *ja = &ca->journal; 1722 1723 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d)) 1724 continue; 1725 1726 if (!ja->nr) 1727 continue; 1728 1729 prt_printf(out, "dev %u:\n", ca->dev_idx); 1730 prt_printf(out, "durability %u:\n", ca->mi.durability); 1731 printbuf_indent_add(out, 2); 1732 prt_printf(out, "nr\t%u\n", ja->nr); 1733 prt_printf(out, "bucket size\t%u\n", ca->mi.bucket_size); 1734 prt_printf(out, "available\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free); 1735 prt_printf(out, "discard_idx\t%u\n", ja->discard_idx); 1736 prt_printf(out, "dirty_ondisk\t%u (seq %llu)\n",ja->dirty_idx_ondisk, ja->bucket_seq[ja->dirty_idx_ondisk]); 1737 prt_printf(out, "dirty_idx\t%u (seq %llu)\n", ja->dirty_idx, ja->bucket_seq[ja->dirty_idx]); 1738 prt_printf(out, "cur_idx\t%u (seq %llu)\n", ja->cur_idx, ja->bucket_seq[ja->cur_idx]); 1739 printbuf_indent_sub(out, 2); 1740 } 1741 1742 prt_printf(out, "replicas want %u need %u\n", c->opts.metadata_replicas, c->opts.metadata_replicas_required); 1743 1744 rcu_read_unlock(); 1745 1746 --out->atomic; 1747 } 1748 1749 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j) 1750 { 1751 spin_lock(&j->lock); 1752 __bch2_journal_debug_to_text(out, j); 1753 spin_unlock(&j->lock); 1754 } 1755