1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * bcachefs journalling code, for btree insertions 4 * 5 * Copyright 2012 Google, Inc. 6 */ 7 8 #include "bcachefs.h" 9 #include "alloc_foreground.h" 10 #include "bkey_methods.h" 11 #include "btree_gc.h" 12 #include "btree_update.h" 13 #include "btree_write_buffer.h" 14 #include "buckets.h" 15 #include "error.h" 16 #include "journal.h" 17 #include "journal_io.h" 18 #include "journal_reclaim.h" 19 #include "journal_sb.h" 20 #include "journal_seq_blacklist.h" 21 #include "trace.h" 22 23 static const char * const bch2_journal_errors[] = { 24 #define x(n) #n, 25 JOURNAL_ERRORS() 26 #undef x 27 NULL 28 }; 29 30 static inline bool journal_seq_unwritten(struct journal *j, u64 seq) 31 { 32 return seq > j->seq_ondisk; 33 } 34 35 static bool __journal_entry_is_open(union journal_res_state state) 36 { 37 return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL; 38 } 39 40 static inline unsigned nr_unwritten_journal_entries(struct journal *j) 41 { 42 return atomic64_read(&j->seq) - j->seq_ondisk; 43 } 44 45 static bool journal_entry_is_open(struct journal *j) 46 { 47 return __journal_entry_is_open(j->reservations); 48 } 49 50 static void bch2_journal_buf_to_text(struct printbuf *out, struct journal *j, u64 seq) 51 { 52 union journal_res_state s = READ_ONCE(j->reservations); 53 unsigned i = seq & JOURNAL_BUF_MASK; 54 struct journal_buf *buf = j->buf + i; 55 56 prt_printf(out, "seq:\t%llu\n", seq); 57 printbuf_indent_add(out, 2); 58 59 prt_printf(out, "refcount:\t%u\n", journal_state_count(s, i)); 60 61 prt_printf(out, "size:\t"); 62 prt_human_readable_u64(out, vstruct_bytes(buf->data)); 63 prt_newline(out); 64 65 prt_printf(out, "expires:\t"); 66 prt_printf(out, "%li jiffies\n", buf->expires - jiffies); 67 68 prt_printf(out, "flags:\t"); 69 if (buf->noflush) 70 prt_str(out, "noflush "); 71 if (buf->must_flush) 72 prt_str(out, "must_flush "); 73 if (buf->separate_flush) 74 prt_str(out, "separate_flush "); 75 if (buf->need_flush_to_write_buffer) 76 prt_str(out, "need_flush_to_write_buffer "); 77 if (buf->write_started) 78 prt_str(out, "write_started "); 79 if (buf->write_allocated) 80 prt_str(out, "write_allocated "); 81 if (buf->write_done) 82 prt_str(out, "write_done"); 83 prt_newline(out); 84 85 printbuf_indent_sub(out, 2); 86 } 87 88 static void bch2_journal_bufs_to_text(struct printbuf *out, struct journal *j) 89 { 90 if (!out->nr_tabstops) 91 printbuf_tabstop_push(out, 24); 92 93 for (u64 seq = journal_last_unwritten_seq(j); 94 seq <= journal_cur_seq(j); 95 seq++) 96 bch2_journal_buf_to_text(out, j, seq); 97 prt_printf(out, "last buf %s\n", journal_entry_is_open(j) ? "open" : "closed"); 98 } 99 100 static inline struct journal_buf * 101 journal_seq_to_buf(struct journal *j, u64 seq) 102 { 103 struct journal_buf *buf = NULL; 104 105 EBUG_ON(seq > journal_cur_seq(j)); 106 107 if (journal_seq_unwritten(j, seq)) { 108 buf = j->buf + (seq & JOURNAL_BUF_MASK); 109 EBUG_ON(le64_to_cpu(buf->data->seq) != seq); 110 } 111 return buf; 112 } 113 114 static void journal_pin_list_init(struct journal_entry_pin_list *p, int count) 115 { 116 unsigned i; 117 118 for (i = 0; i < ARRAY_SIZE(p->list); i++) 119 INIT_LIST_HEAD(&p->list[i]); 120 INIT_LIST_HEAD(&p->flushed); 121 atomic_set(&p->count, count); 122 p->devs.nr = 0; 123 } 124 125 /* 126 * Detect stuck journal conditions and trigger shutdown. Technically the journal 127 * can end up stuck for a variety of reasons, such as a blocked I/O, journal 128 * reservation lockup, etc. Since this is a fatal error with potentially 129 * unpredictable characteristics, we want to be fairly conservative before we 130 * decide to shut things down. 131 * 132 * Consider the journal stuck when it appears full with no ability to commit 133 * btree transactions, to discard journal buckets, nor acquire priority 134 * (reserved watermark) reservation. 135 */ 136 static inline bool 137 journal_error_check_stuck(struct journal *j, int error, unsigned flags) 138 { 139 struct bch_fs *c = container_of(j, struct bch_fs, journal); 140 bool stuck = false; 141 struct printbuf buf = PRINTBUF; 142 143 if (!(error == JOURNAL_ERR_journal_full || 144 error == JOURNAL_ERR_journal_pin_full) || 145 nr_unwritten_journal_entries(j) || 146 (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) 147 return stuck; 148 149 spin_lock(&j->lock); 150 151 if (j->can_discard) { 152 spin_unlock(&j->lock); 153 return stuck; 154 } 155 156 stuck = true; 157 158 /* 159 * The journal shutdown path will set ->err_seq, but do it here first to 160 * serialize against concurrent failures and avoid duplicate error 161 * reports. 162 */ 163 if (j->err_seq) { 164 spin_unlock(&j->lock); 165 return stuck; 166 } 167 j->err_seq = journal_cur_seq(j); 168 spin_unlock(&j->lock); 169 170 bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (error %s)", 171 bch2_journal_errors[error]); 172 bch2_journal_debug_to_text(&buf, j); 173 bch_err(c, "%s", buf.buf); 174 175 printbuf_reset(&buf); 176 bch2_journal_pins_to_text(&buf, j); 177 bch_err(c, "Journal pins:\n%s", buf.buf); 178 printbuf_exit(&buf); 179 180 bch2_fatal_error(c); 181 dump_stack(); 182 183 return stuck; 184 } 185 186 void bch2_journal_do_writes(struct journal *j) 187 { 188 for (u64 seq = journal_last_unwritten_seq(j); 189 seq <= journal_cur_seq(j); 190 seq++) { 191 unsigned idx = seq & JOURNAL_BUF_MASK; 192 struct journal_buf *w = j->buf + idx; 193 194 if (w->write_started && !w->write_allocated) 195 break; 196 if (w->write_started) 197 continue; 198 199 if (!journal_state_count(j->reservations, idx)) { 200 w->write_started = true; 201 closure_call(&w->io, bch2_journal_write, j->wq, NULL); 202 } 203 204 break; 205 } 206 } 207 208 /* 209 * Final processing when the last reference of a journal buffer has been 210 * dropped. Drop the pin list reference acquired at journal entry open and write 211 * the buffer, if requested. 212 */ 213 void bch2_journal_buf_put_final(struct journal *j, u64 seq) 214 { 215 lockdep_assert_held(&j->lock); 216 217 if (__bch2_journal_pin_put(j, seq)) 218 bch2_journal_reclaim_fast(j); 219 bch2_journal_do_writes(j); 220 } 221 222 /* 223 * Returns true if journal entry is now closed: 224 * 225 * We don't close a journal_buf until the next journal_buf is finished writing, 226 * and can be opened again - this also initializes the next journal_buf: 227 */ 228 static void __journal_entry_close(struct journal *j, unsigned closed_val, bool trace) 229 { 230 struct bch_fs *c = container_of(j, struct bch_fs, journal); 231 struct journal_buf *buf = journal_cur_buf(j); 232 union journal_res_state old, new; 233 unsigned sectors; 234 235 BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL && 236 closed_val != JOURNAL_ENTRY_ERROR_VAL); 237 238 lockdep_assert_held(&j->lock); 239 240 old.v = atomic64_read(&j->reservations.counter); 241 do { 242 new.v = old.v; 243 new.cur_entry_offset = closed_val; 244 245 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL || 246 old.cur_entry_offset == new.cur_entry_offset) 247 return; 248 } while (!atomic64_try_cmpxchg(&j->reservations.counter, 249 &old.v, new.v)); 250 251 if (!__journal_entry_is_open(old)) 252 return; 253 254 /* Close out old buffer: */ 255 buf->data->u64s = cpu_to_le32(old.cur_entry_offset); 256 257 if (trace_journal_entry_close_enabled() && trace) { 258 struct printbuf pbuf = PRINTBUF; 259 pbuf.atomic++; 260 261 prt_str(&pbuf, "entry size: "); 262 prt_human_readable_u64(&pbuf, vstruct_bytes(buf->data)); 263 prt_newline(&pbuf); 264 bch2_prt_task_backtrace(&pbuf, current, 1, GFP_NOWAIT); 265 trace_journal_entry_close(c, pbuf.buf); 266 printbuf_exit(&pbuf); 267 } 268 269 sectors = vstruct_blocks_plus(buf->data, c->block_bits, 270 buf->u64s_reserved) << c->block_bits; 271 BUG_ON(sectors > buf->sectors); 272 buf->sectors = sectors; 273 274 /* 275 * We have to set last_seq here, _before_ opening a new journal entry: 276 * 277 * A threads may replace an old pin with a new pin on their current 278 * journal reservation - the expectation being that the journal will 279 * contain either what the old pin protected or what the new pin 280 * protects. 281 * 282 * After the old pin is dropped journal_last_seq() won't include the old 283 * pin, so we can only write the updated last_seq on the entry that 284 * contains whatever the new pin protects. 285 * 286 * Restated, we can _not_ update last_seq for a given entry if there 287 * could be a newer entry open with reservations/pins that have been 288 * taken against it. 289 * 290 * Hence, we want update/set last_seq on the current journal entry right 291 * before we open a new one: 292 */ 293 buf->last_seq = journal_last_seq(j); 294 buf->data->last_seq = cpu_to_le64(buf->last_seq); 295 BUG_ON(buf->last_seq > le64_to_cpu(buf->data->seq)); 296 297 cancel_delayed_work(&j->write_work); 298 299 bch2_journal_space_available(j); 300 301 __bch2_journal_buf_put(j, old.idx, le64_to_cpu(buf->data->seq)); 302 } 303 304 void bch2_journal_halt(struct journal *j) 305 { 306 spin_lock(&j->lock); 307 __journal_entry_close(j, JOURNAL_ENTRY_ERROR_VAL, true); 308 if (!j->err_seq) 309 j->err_seq = journal_cur_seq(j); 310 journal_wake(j); 311 spin_unlock(&j->lock); 312 } 313 314 static bool journal_entry_want_write(struct journal *j) 315 { 316 bool ret = !journal_entry_is_open(j) || 317 journal_cur_seq(j) == journal_last_unwritten_seq(j); 318 319 /* Don't close it yet if we already have a write in flight: */ 320 if (ret) 321 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true); 322 else if (nr_unwritten_journal_entries(j)) { 323 struct journal_buf *buf = journal_cur_buf(j); 324 325 if (!buf->flush_time) { 326 buf->flush_time = local_clock() ?: 1; 327 buf->expires = jiffies; 328 } 329 } 330 331 return ret; 332 } 333 334 bool bch2_journal_entry_close(struct journal *j) 335 { 336 bool ret; 337 338 spin_lock(&j->lock); 339 ret = journal_entry_want_write(j); 340 spin_unlock(&j->lock); 341 342 return ret; 343 } 344 345 /* 346 * should _only_ called from journal_res_get() - when we actually want a 347 * journal reservation - journal entry is open means journal is dirty: 348 */ 349 static int journal_entry_open(struct journal *j) 350 { 351 struct bch_fs *c = container_of(j, struct bch_fs, journal); 352 struct journal_buf *buf = j->buf + 353 ((journal_cur_seq(j) + 1) & JOURNAL_BUF_MASK); 354 union journal_res_state old, new; 355 int u64s; 356 357 lockdep_assert_held(&j->lock); 358 BUG_ON(journal_entry_is_open(j)); 359 BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb)); 360 361 if (j->blocked) 362 return JOURNAL_ERR_blocked; 363 364 if (j->cur_entry_error) 365 return j->cur_entry_error; 366 367 if (bch2_journal_error(j)) 368 return JOURNAL_ERR_insufficient_devices; /* -EROFS */ 369 370 if (!fifo_free(&j->pin)) 371 return JOURNAL_ERR_journal_pin_full; 372 373 if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf)) 374 return JOURNAL_ERR_max_in_flight; 375 376 BUG_ON(!j->cur_entry_sectors); 377 378 buf->expires = 379 (journal_cur_seq(j) == j->flushed_seq_ondisk 380 ? jiffies 381 : j->last_flush_write) + 382 msecs_to_jiffies(c->opts.journal_flush_delay); 383 384 buf->u64s_reserved = j->entry_u64s_reserved; 385 buf->disk_sectors = j->cur_entry_sectors; 386 buf->sectors = min(buf->disk_sectors, buf->buf_size >> 9); 387 388 u64s = (int) (buf->sectors << 9) / sizeof(u64) - 389 journal_entry_overhead(j); 390 u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1); 391 392 if (u64s <= (ssize_t) j->early_journal_entries.nr) 393 return JOURNAL_ERR_journal_full; 394 395 if (fifo_empty(&j->pin) && j->reclaim_thread) 396 wake_up_process(j->reclaim_thread); 397 398 /* 399 * The fifo_push() needs to happen at the same time as j->seq is 400 * incremented for journal_last_seq() to be calculated correctly 401 */ 402 atomic64_inc(&j->seq); 403 journal_pin_list_init(fifo_push_ref(&j->pin), 1); 404 405 BUG_ON(j->pin.back - 1 != atomic64_read(&j->seq)); 406 407 BUG_ON(j->buf + (journal_cur_seq(j) & JOURNAL_BUF_MASK) != buf); 408 409 bkey_extent_init(&buf->key); 410 buf->noflush = false; 411 buf->must_flush = false; 412 buf->separate_flush = false; 413 buf->flush_time = 0; 414 buf->need_flush_to_write_buffer = true; 415 buf->write_started = false; 416 buf->write_allocated = false; 417 buf->write_done = false; 418 419 memset(buf->data, 0, sizeof(*buf->data)); 420 buf->data->seq = cpu_to_le64(journal_cur_seq(j)); 421 buf->data->u64s = 0; 422 423 if (j->early_journal_entries.nr) { 424 memcpy(buf->data->_data, j->early_journal_entries.data, 425 j->early_journal_entries.nr * sizeof(u64)); 426 le32_add_cpu(&buf->data->u64s, j->early_journal_entries.nr); 427 } 428 429 /* 430 * Must be set before marking the journal entry as open: 431 */ 432 j->cur_entry_u64s = u64s; 433 434 old.v = atomic64_read(&j->reservations.counter); 435 do { 436 new.v = old.v; 437 438 BUG_ON(old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL); 439 440 new.idx++; 441 BUG_ON(journal_state_count(new, new.idx)); 442 BUG_ON(new.idx != (journal_cur_seq(j) & JOURNAL_BUF_MASK)); 443 444 journal_state_inc(&new); 445 446 /* Handle any already added entries */ 447 new.cur_entry_offset = le32_to_cpu(buf->data->u64s); 448 } while (!atomic64_try_cmpxchg(&j->reservations.counter, 449 &old.v, new.v)); 450 451 if (nr_unwritten_journal_entries(j) == 1) 452 mod_delayed_work(j->wq, 453 &j->write_work, 454 msecs_to_jiffies(c->opts.journal_flush_delay)); 455 journal_wake(j); 456 457 if (j->early_journal_entries.nr) 458 darray_exit(&j->early_journal_entries); 459 return 0; 460 } 461 462 static bool journal_quiesced(struct journal *j) 463 { 464 bool ret = atomic64_read(&j->seq) == j->seq_ondisk; 465 466 if (!ret) 467 bch2_journal_entry_close(j); 468 return ret; 469 } 470 471 static void journal_quiesce(struct journal *j) 472 { 473 wait_event(j->wait, journal_quiesced(j)); 474 } 475 476 static void journal_write_work(struct work_struct *work) 477 { 478 struct journal *j = container_of(work, struct journal, write_work.work); 479 480 spin_lock(&j->lock); 481 if (__journal_entry_is_open(j->reservations)) { 482 long delta = journal_cur_buf(j)->expires - jiffies; 483 484 if (delta > 0) 485 mod_delayed_work(j->wq, &j->write_work, delta); 486 else 487 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true); 488 } 489 spin_unlock(&j->lock); 490 } 491 492 static int __journal_res_get(struct journal *j, struct journal_res *res, 493 unsigned flags) 494 { 495 struct bch_fs *c = container_of(j, struct bch_fs, journal); 496 struct journal_buf *buf; 497 bool can_discard; 498 int ret; 499 retry: 500 if (journal_res_get_fast(j, res, flags)) 501 return 0; 502 503 if (bch2_journal_error(j)) 504 return -BCH_ERR_erofs_journal_err; 505 506 if (j->blocked) 507 return -BCH_ERR_journal_res_get_blocked; 508 509 if ((flags & BCH_WATERMARK_MASK) < j->watermark) { 510 ret = JOURNAL_ERR_journal_full; 511 can_discard = j->can_discard; 512 goto out; 513 } 514 515 if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf) && !journal_entry_is_open(j)) { 516 ret = JOURNAL_ERR_max_in_flight; 517 goto out; 518 } 519 520 spin_lock(&j->lock); 521 522 /* 523 * Recheck after taking the lock, so we don't race with another thread 524 * that just did journal_entry_open() and call bch2_journal_entry_close() 525 * unnecessarily 526 */ 527 if (journal_res_get_fast(j, res, flags)) { 528 ret = 0; 529 goto unlock; 530 } 531 532 /* 533 * If we couldn't get a reservation because the current buf filled up, 534 * and we had room for a bigger entry on disk, signal that we want to 535 * realloc the journal bufs: 536 */ 537 buf = journal_cur_buf(j); 538 if (journal_entry_is_open(j) && 539 buf->buf_size >> 9 < buf->disk_sectors && 540 buf->buf_size < JOURNAL_ENTRY_SIZE_MAX) 541 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1); 542 543 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, false); 544 ret = journal_entry_open(j) ?: JOURNAL_ERR_retry; 545 unlock: 546 can_discard = j->can_discard; 547 spin_unlock(&j->lock); 548 out: 549 if (ret == JOURNAL_ERR_retry) 550 goto retry; 551 if (!ret) 552 return 0; 553 554 if (journal_error_check_stuck(j, ret, flags)) 555 ret = -BCH_ERR_journal_res_get_blocked; 556 557 if (ret == JOURNAL_ERR_max_in_flight && 558 track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight], true)) { 559 560 struct printbuf buf = PRINTBUF; 561 prt_printf(&buf, "seq %llu\n", journal_cur_seq(j)); 562 bch2_journal_bufs_to_text(&buf, j); 563 trace_journal_entry_full(c, buf.buf); 564 printbuf_exit(&buf); 565 count_event(c, journal_entry_full); 566 } 567 568 /* 569 * Journal is full - can't rely on reclaim from work item due to 570 * freezing: 571 */ 572 if ((ret == JOURNAL_ERR_journal_full || 573 ret == JOURNAL_ERR_journal_pin_full) && 574 !(flags & JOURNAL_RES_GET_NONBLOCK)) { 575 if (can_discard) { 576 bch2_journal_do_discards(j); 577 goto retry; 578 } 579 580 if (mutex_trylock(&j->reclaim_lock)) { 581 bch2_journal_reclaim(j); 582 mutex_unlock(&j->reclaim_lock); 583 } 584 } 585 586 return ret == JOURNAL_ERR_insufficient_devices 587 ? -BCH_ERR_erofs_journal_err 588 : -BCH_ERR_journal_res_get_blocked; 589 } 590 591 /* 592 * Essentially the entry function to the journaling code. When bcachefs is doing 593 * a btree insert, it calls this function to get the current journal write. 594 * Journal write is the structure used set up journal writes. The calling 595 * function will then add its keys to the structure, queuing them for the next 596 * write. 597 * 598 * To ensure forward progress, the current task must not be holding any 599 * btree node write locks. 600 */ 601 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res, 602 unsigned flags) 603 { 604 int ret; 605 606 closure_wait_event(&j->async_wait, 607 (ret = __journal_res_get(j, res, flags)) != -BCH_ERR_journal_res_get_blocked || 608 (flags & JOURNAL_RES_GET_NONBLOCK)); 609 return ret; 610 } 611 612 /* journal_entry_res: */ 613 614 void bch2_journal_entry_res_resize(struct journal *j, 615 struct journal_entry_res *res, 616 unsigned new_u64s) 617 { 618 union journal_res_state state; 619 int d = new_u64s - res->u64s; 620 621 spin_lock(&j->lock); 622 623 j->entry_u64s_reserved += d; 624 if (d <= 0) 625 goto out; 626 627 j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d); 628 smp_mb(); 629 state = READ_ONCE(j->reservations); 630 631 if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL && 632 state.cur_entry_offset > j->cur_entry_u64s) { 633 j->cur_entry_u64s += d; 634 /* 635 * Not enough room in current journal entry, have to flush it: 636 */ 637 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true); 638 } else { 639 journal_cur_buf(j)->u64s_reserved += d; 640 } 641 out: 642 spin_unlock(&j->lock); 643 res->u64s += d; 644 } 645 646 /* journal flushing: */ 647 648 /** 649 * bch2_journal_flush_seq_async - wait for a journal entry to be written 650 * @j: journal object 651 * @seq: seq to flush 652 * @parent: closure object to wait with 653 * Returns: 1 if @seq has already been flushed, 0 if @seq is being flushed, 654 * -EIO if @seq will never be flushed 655 * 656 * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if 657 * necessary 658 */ 659 int bch2_journal_flush_seq_async(struct journal *j, u64 seq, 660 struct closure *parent) 661 { 662 struct journal_buf *buf; 663 int ret = 0; 664 665 if (seq <= j->flushed_seq_ondisk) 666 return 1; 667 668 spin_lock(&j->lock); 669 670 if (WARN_ONCE(seq > journal_cur_seq(j), 671 "requested to flush journal seq %llu, but currently at %llu", 672 seq, journal_cur_seq(j))) 673 goto out; 674 675 /* Recheck under lock: */ 676 if (j->err_seq && seq >= j->err_seq) { 677 ret = -EIO; 678 goto out; 679 } 680 681 if (seq <= j->flushed_seq_ondisk) { 682 ret = 1; 683 goto out; 684 } 685 686 /* if seq was written, but not flushed - flush a newer one instead */ 687 seq = max(seq, journal_last_unwritten_seq(j)); 688 689 recheck_need_open: 690 if (seq > journal_cur_seq(j)) { 691 struct journal_res res = { 0 }; 692 693 if (journal_entry_is_open(j)) 694 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true); 695 696 spin_unlock(&j->lock); 697 698 /* 699 * We're called from bch2_journal_flush_seq() -> wait_event(); 700 * but this might block. We won't usually block, so we won't 701 * livelock: 702 */ 703 sched_annotate_sleep(); 704 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0); 705 if (ret) 706 return ret; 707 708 seq = res.seq; 709 buf = journal_seq_to_buf(j, seq); 710 buf->must_flush = true; 711 712 if (!buf->flush_time) { 713 buf->flush_time = local_clock() ?: 1; 714 buf->expires = jiffies; 715 } 716 717 if (parent && !closure_wait(&buf->wait, parent)) 718 BUG(); 719 720 bch2_journal_res_put(j, &res); 721 722 spin_lock(&j->lock); 723 goto want_write; 724 } 725 726 /* 727 * if write was kicked off without a flush, or if we promised it 728 * wouldn't be a flush, flush the next sequence number instead 729 */ 730 buf = journal_seq_to_buf(j, seq); 731 if (buf->noflush) { 732 seq++; 733 goto recheck_need_open; 734 } 735 736 buf->must_flush = true; 737 738 if (parent && !closure_wait(&buf->wait, parent)) 739 BUG(); 740 want_write: 741 if (seq == journal_cur_seq(j)) 742 journal_entry_want_write(j); 743 out: 744 spin_unlock(&j->lock); 745 return ret; 746 } 747 748 int bch2_journal_flush_seq(struct journal *j, u64 seq) 749 { 750 u64 start_time = local_clock(); 751 int ret, ret2; 752 753 /* 754 * Don't update time_stats when @seq is already flushed: 755 */ 756 if (seq <= j->flushed_seq_ondisk) 757 return 0; 758 759 ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL))); 760 761 if (!ret) 762 bch2_time_stats_update(j->flush_seq_time, start_time); 763 764 return ret ?: ret2 < 0 ? ret2 : 0; 765 } 766 767 /* 768 * bch2_journal_flush_async - if there is an open journal entry, or a journal 769 * still being written, write it and wait for the write to complete 770 */ 771 void bch2_journal_flush_async(struct journal *j, struct closure *parent) 772 { 773 bch2_journal_flush_seq_async(j, atomic64_read(&j->seq), parent); 774 } 775 776 int bch2_journal_flush(struct journal *j) 777 { 778 return bch2_journal_flush_seq(j, atomic64_read(&j->seq)); 779 } 780 781 /* 782 * bch2_journal_noflush_seq - tell the journal not to issue any flushes before 783 * @seq 784 */ 785 bool bch2_journal_noflush_seq(struct journal *j, u64 seq) 786 { 787 struct bch_fs *c = container_of(j, struct bch_fs, journal); 788 u64 unwritten_seq; 789 bool ret = false; 790 791 if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush))) 792 return false; 793 794 if (seq <= c->journal.flushed_seq_ondisk) 795 return false; 796 797 spin_lock(&j->lock); 798 if (seq <= c->journal.flushed_seq_ondisk) 799 goto out; 800 801 for (unwritten_seq = journal_last_unwritten_seq(j); 802 unwritten_seq < seq; 803 unwritten_seq++) { 804 struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq); 805 806 /* journal flush already in flight, or flush requseted */ 807 if (buf->must_flush) 808 goto out; 809 810 buf->noflush = true; 811 } 812 813 ret = true; 814 out: 815 spin_unlock(&j->lock); 816 return ret; 817 } 818 819 int bch2_journal_meta(struct journal *j) 820 { 821 struct journal_buf *buf; 822 struct journal_res res; 823 int ret; 824 825 memset(&res, 0, sizeof(res)); 826 827 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0); 828 if (ret) 829 return ret; 830 831 buf = j->buf + (res.seq & JOURNAL_BUF_MASK); 832 buf->must_flush = true; 833 834 if (!buf->flush_time) { 835 buf->flush_time = local_clock() ?: 1; 836 buf->expires = jiffies; 837 } 838 839 bch2_journal_res_put(j, &res); 840 841 return bch2_journal_flush_seq(j, res.seq); 842 } 843 844 /* block/unlock the journal: */ 845 846 void bch2_journal_unblock(struct journal *j) 847 { 848 spin_lock(&j->lock); 849 j->blocked--; 850 spin_unlock(&j->lock); 851 852 journal_wake(j); 853 } 854 855 void bch2_journal_block(struct journal *j) 856 { 857 spin_lock(&j->lock); 858 j->blocked++; 859 spin_unlock(&j->lock); 860 861 journal_quiesce(j); 862 } 863 864 static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq) 865 { 866 struct journal_buf *ret = NULL; 867 868 /* We're inside wait_event(), but using mutex_lock(: */ 869 sched_annotate_sleep(); 870 mutex_lock(&j->buf_lock); 871 spin_lock(&j->lock); 872 max_seq = min(max_seq, journal_cur_seq(j)); 873 874 for (u64 seq = journal_last_unwritten_seq(j); 875 seq <= max_seq; 876 seq++) { 877 unsigned idx = seq & JOURNAL_BUF_MASK; 878 struct journal_buf *buf = j->buf + idx; 879 880 if (buf->need_flush_to_write_buffer) { 881 if (seq == journal_cur_seq(j)) 882 __journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true); 883 884 union journal_res_state s; 885 s.v = atomic64_read_acquire(&j->reservations.counter); 886 887 ret = journal_state_count(s, idx) 888 ? ERR_PTR(-EAGAIN) 889 : buf; 890 break; 891 } 892 } 893 894 spin_unlock(&j->lock); 895 if (IS_ERR_OR_NULL(ret)) 896 mutex_unlock(&j->buf_lock); 897 return ret; 898 } 899 900 struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq) 901 { 902 struct journal_buf *ret; 903 904 wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j, max_seq)) != ERR_PTR(-EAGAIN)); 905 return ret; 906 } 907 908 /* allocate journal on a device: */ 909 910 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr, 911 bool new_fs, struct closure *cl) 912 { 913 struct bch_fs *c = ca->fs; 914 struct journal_device *ja = &ca->journal; 915 u64 *new_bucket_seq = NULL, *new_buckets = NULL; 916 struct open_bucket **ob = NULL; 917 long *bu = NULL; 918 unsigned i, pos, nr_got = 0, nr_want = nr - ja->nr; 919 int ret = 0; 920 921 BUG_ON(nr <= ja->nr); 922 923 bu = kcalloc(nr_want, sizeof(*bu), GFP_KERNEL); 924 ob = kcalloc(nr_want, sizeof(*ob), GFP_KERNEL); 925 new_buckets = kcalloc(nr, sizeof(u64), GFP_KERNEL); 926 new_bucket_seq = kcalloc(nr, sizeof(u64), GFP_KERNEL); 927 if (!bu || !ob || !new_buckets || !new_bucket_seq) { 928 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets; 929 goto err_free; 930 } 931 932 for (nr_got = 0; nr_got < nr_want; nr_got++) { 933 if (new_fs) { 934 bu[nr_got] = bch2_bucket_alloc_new_fs(ca); 935 if (bu[nr_got] < 0) { 936 ret = -BCH_ERR_ENOSPC_bucket_alloc; 937 break; 938 } 939 } else { 940 ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal, 941 BCH_DATA_journal, cl); 942 ret = PTR_ERR_OR_ZERO(ob[nr_got]); 943 if (ret) 944 break; 945 946 ret = bch2_trans_run(c, 947 bch2_trans_mark_metadata_bucket(trans, ca, 948 ob[nr_got]->bucket, BCH_DATA_journal, 949 ca->mi.bucket_size, BTREE_TRIGGER_transactional)); 950 if (ret) { 951 bch2_open_bucket_put(c, ob[nr_got]); 952 bch_err_msg(c, ret, "marking new journal buckets"); 953 break; 954 } 955 956 bu[nr_got] = ob[nr_got]->bucket; 957 } 958 } 959 960 if (!nr_got) 961 goto err_free; 962 963 /* Don't return an error if we successfully allocated some buckets: */ 964 ret = 0; 965 966 if (c) { 967 bch2_journal_flush_all_pins(&c->journal); 968 bch2_journal_block(&c->journal); 969 mutex_lock(&c->sb_lock); 970 } 971 972 memcpy(new_buckets, ja->buckets, ja->nr * sizeof(u64)); 973 memcpy(new_bucket_seq, ja->bucket_seq, ja->nr * sizeof(u64)); 974 975 BUG_ON(ja->discard_idx > ja->nr); 976 977 pos = ja->discard_idx ?: ja->nr; 978 979 memmove(new_buckets + pos + nr_got, 980 new_buckets + pos, 981 sizeof(new_buckets[0]) * (ja->nr - pos)); 982 memmove(new_bucket_seq + pos + nr_got, 983 new_bucket_seq + pos, 984 sizeof(new_bucket_seq[0]) * (ja->nr - pos)); 985 986 for (i = 0; i < nr_got; i++) { 987 new_buckets[pos + i] = bu[i]; 988 new_bucket_seq[pos + i] = 0; 989 } 990 991 nr = ja->nr + nr_got; 992 993 ret = bch2_journal_buckets_to_sb(c, ca, new_buckets, nr); 994 if (ret) 995 goto err_unblock; 996 997 if (!new_fs) 998 bch2_write_super(c); 999 1000 /* Commit: */ 1001 if (c) 1002 spin_lock(&c->journal.lock); 1003 1004 swap(new_buckets, ja->buckets); 1005 swap(new_bucket_seq, ja->bucket_seq); 1006 ja->nr = nr; 1007 1008 if (pos <= ja->discard_idx) 1009 ja->discard_idx = (ja->discard_idx + nr_got) % ja->nr; 1010 if (pos <= ja->dirty_idx_ondisk) 1011 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + nr_got) % ja->nr; 1012 if (pos <= ja->dirty_idx) 1013 ja->dirty_idx = (ja->dirty_idx + nr_got) % ja->nr; 1014 if (pos <= ja->cur_idx) 1015 ja->cur_idx = (ja->cur_idx + nr_got) % ja->nr; 1016 1017 if (c) 1018 spin_unlock(&c->journal.lock); 1019 err_unblock: 1020 if (c) { 1021 bch2_journal_unblock(&c->journal); 1022 mutex_unlock(&c->sb_lock); 1023 } 1024 1025 if (ret && !new_fs) 1026 for (i = 0; i < nr_got; i++) 1027 bch2_trans_run(c, 1028 bch2_trans_mark_metadata_bucket(trans, ca, 1029 bu[i], BCH_DATA_free, 0, 1030 BTREE_TRIGGER_transactional)); 1031 err_free: 1032 if (!new_fs) 1033 for (i = 0; i < nr_got; i++) 1034 bch2_open_bucket_put(c, ob[i]); 1035 1036 kfree(new_bucket_seq); 1037 kfree(new_buckets); 1038 kfree(ob); 1039 kfree(bu); 1040 return ret; 1041 } 1042 1043 /* 1044 * Allocate more journal space at runtime - not currently making use if it, but 1045 * the code works: 1046 */ 1047 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca, 1048 unsigned nr) 1049 { 1050 struct journal_device *ja = &ca->journal; 1051 struct closure cl; 1052 int ret = 0; 1053 1054 closure_init_stack(&cl); 1055 1056 down_write(&c->state_lock); 1057 1058 /* don't handle reducing nr of buckets yet: */ 1059 if (nr < ja->nr) 1060 goto unlock; 1061 1062 while (ja->nr < nr) { 1063 struct disk_reservation disk_res = { 0, 0, 0 }; 1064 1065 /* 1066 * note: journal buckets aren't really counted as _sectors_ used yet, so 1067 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c 1068 * when space used goes up without a reservation - but we do need the 1069 * reservation to ensure we'll actually be able to allocate: 1070 * 1071 * XXX: that's not right, disk reservations only ensure a 1072 * filesystem-wide allocation will succeed, this is a device 1073 * specific allocation - we can hang here: 1074 */ 1075 1076 ret = bch2_disk_reservation_get(c, &disk_res, 1077 bucket_to_sector(ca, nr - ja->nr), 1, 0); 1078 if (ret) 1079 break; 1080 1081 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl); 1082 1083 bch2_disk_reservation_put(c, &disk_res); 1084 1085 closure_sync(&cl); 1086 1087 if (ret && ret != -BCH_ERR_bucket_alloc_blocked) 1088 break; 1089 } 1090 1091 bch_err_fn(c, ret); 1092 unlock: 1093 up_write(&c->state_lock); 1094 return ret; 1095 } 1096 1097 int bch2_dev_journal_alloc(struct bch_dev *ca, bool new_fs) 1098 { 1099 unsigned nr; 1100 int ret; 1101 1102 if (dynamic_fault("bcachefs:add:journal_alloc")) { 1103 ret = -BCH_ERR_ENOMEM_set_nr_journal_buckets; 1104 goto err; 1105 } 1106 1107 /* 1/128th of the device by default: */ 1108 nr = ca->mi.nbuckets >> 7; 1109 1110 /* 1111 * clamp journal size to 8192 buckets or 8GB (in sectors), whichever 1112 * is smaller: 1113 */ 1114 nr = clamp_t(unsigned, nr, 1115 BCH_JOURNAL_BUCKETS_MIN, 1116 min(1 << 13, 1117 (1 << 24) / ca->mi.bucket_size)); 1118 1119 ret = __bch2_set_nr_journal_buckets(ca, nr, new_fs, NULL); 1120 err: 1121 bch_err_fn(ca, ret); 1122 return ret; 1123 } 1124 1125 int bch2_fs_journal_alloc(struct bch_fs *c) 1126 { 1127 for_each_online_member(c, ca) { 1128 if (ca->journal.nr) 1129 continue; 1130 1131 int ret = bch2_dev_journal_alloc(ca, true); 1132 if (ret) { 1133 percpu_ref_put(&ca->io_ref); 1134 return ret; 1135 } 1136 } 1137 1138 return 0; 1139 } 1140 1141 /* startup/shutdown: */ 1142 1143 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx) 1144 { 1145 bool ret = false; 1146 u64 seq; 1147 1148 spin_lock(&j->lock); 1149 for (seq = journal_last_unwritten_seq(j); 1150 seq <= journal_cur_seq(j) && !ret; 1151 seq++) { 1152 struct journal_buf *buf = journal_seq_to_buf(j, seq); 1153 1154 if (bch2_bkey_has_device_c(bkey_i_to_s_c(&buf->key), dev_idx)) 1155 ret = true; 1156 } 1157 spin_unlock(&j->lock); 1158 1159 return ret; 1160 } 1161 1162 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca) 1163 { 1164 wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx)); 1165 } 1166 1167 void bch2_fs_journal_stop(struct journal *j) 1168 { 1169 if (!test_bit(JOURNAL_running, &j->flags)) 1170 return; 1171 1172 bch2_journal_reclaim_stop(j); 1173 bch2_journal_flush_all_pins(j); 1174 1175 wait_event(j->wait, bch2_journal_entry_close(j)); 1176 1177 /* 1178 * Always write a new journal entry, to make sure the clock hands are up 1179 * to date (and match the superblock) 1180 */ 1181 bch2_journal_meta(j); 1182 1183 journal_quiesce(j); 1184 cancel_delayed_work_sync(&j->write_work); 1185 1186 WARN(!bch2_journal_error(j) && 1187 test_bit(JOURNAL_replay_done, &j->flags) && 1188 j->last_empty_seq != journal_cur_seq(j), 1189 "journal shutdown error: cur seq %llu but last empty seq %llu", 1190 journal_cur_seq(j), j->last_empty_seq); 1191 1192 if (!bch2_journal_error(j)) 1193 clear_bit(JOURNAL_running, &j->flags); 1194 } 1195 1196 int bch2_fs_journal_start(struct journal *j, u64 cur_seq) 1197 { 1198 struct bch_fs *c = container_of(j, struct bch_fs, journal); 1199 struct journal_entry_pin_list *p; 1200 struct journal_replay *i, **_i; 1201 struct genradix_iter iter; 1202 bool had_entries = false; 1203 u64 last_seq = cur_seq, nr, seq; 1204 1205 genradix_for_each_reverse(&c->journal_entries, iter, _i) { 1206 i = *_i; 1207 1208 if (journal_replay_ignore(i)) 1209 continue; 1210 1211 last_seq = le64_to_cpu(i->j.last_seq); 1212 break; 1213 } 1214 1215 nr = cur_seq - last_seq; 1216 1217 if (nr + 1 > j->pin.size) { 1218 free_fifo(&j->pin); 1219 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL); 1220 if (!j->pin.data) { 1221 bch_err(c, "error reallocating journal fifo (%llu open entries)", nr); 1222 return -BCH_ERR_ENOMEM_journal_pin_fifo; 1223 } 1224 } 1225 1226 j->replay_journal_seq = last_seq; 1227 j->replay_journal_seq_end = cur_seq; 1228 j->last_seq_ondisk = last_seq; 1229 j->flushed_seq_ondisk = cur_seq - 1; 1230 j->seq_ondisk = cur_seq - 1; 1231 j->pin.front = last_seq; 1232 j->pin.back = cur_seq; 1233 atomic64_set(&j->seq, cur_seq - 1); 1234 1235 fifo_for_each_entry_ptr(p, &j->pin, seq) 1236 journal_pin_list_init(p, 1); 1237 1238 genradix_for_each(&c->journal_entries, iter, _i) { 1239 i = *_i; 1240 1241 if (journal_replay_ignore(i)) 1242 continue; 1243 1244 seq = le64_to_cpu(i->j.seq); 1245 BUG_ON(seq >= cur_seq); 1246 1247 if (seq < last_seq) 1248 continue; 1249 1250 if (journal_entry_empty(&i->j)) 1251 j->last_empty_seq = le64_to_cpu(i->j.seq); 1252 1253 p = journal_seq_pin(j, seq); 1254 1255 p->devs.nr = 0; 1256 darray_for_each(i->ptrs, ptr) 1257 bch2_dev_list_add_dev(&p->devs, ptr->dev); 1258 1259 had_entries = true; 1260 } 1261 1262 if (!had_entries) 1263 j->last_empty_seq = cur_seq - 1; /* to match j->seq */ 1264 1265 spin_lock(&j->lock); 1266 1267 set_bit(JOURNAL_running, &j->flags); 1268 j->last_flush_write = jiffies; 1269 1270 j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j); 1271 j->reservations.unwritten_idx++; 1272 1273 c->last_bucket_seq_cleanup = journal_cur_seq(j); 1274 1275 bch2_journal_space_available(j); 1276 spin_unlock(&j->lock); 1277 1278 return bch2_journal_reclaim_start(j); 1279 } 1280 1281 /* init/exit: */ 1282 1283 void bch2_dev_journal_exit(struct bch_dev *ca) 1284 { 1285 struct journal_device *ja = &ca->journal; 1286 1287 for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) { 1288 kfree(ja->bio[i]); 1289 ja->bio[i] = NULL; 1290 } 1291 1292 kfree(ja->buckets); 1293 kfree(ja->bucket_seq); 1294 ja->buckets = NULL; 1295 ja->bucket_seq = NULL; 1296 } 1297 1298 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb) 1299 { 1300 struct journal_device *ja = &ca->journal; 1301 struct bch_sb_field_journal *journal_buckets = 1302 bch2_sb_field_get(sb, journal); 1303 struct bch_sb_field_journal_v2 *journal_buckets_v2 = 1304 bch2_sb_field_get(sb, journal_v2); 1305 1306 ja->nr = 0; 1307 1308 if (journal_buckets_v2) { 1309 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2); 1310 1311 for (unsigned i = 0; i < nr; i++) 1312 ja->nr += le64_to_cpu(journal_buckets_v2->d[i].nr); 1313 } else if (journal_buckets) { 1314 ja->nr = bch2_nr_journal_buckets(journal_buckets); 1315 } 1316 1317 ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL); 1318 if (!ja->bucket_seq) 1319 return -BCH_ERR_ENOMEM_dev_journal_init; 1320 1321 unsigned nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE); 1322 1323 for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) { 1324 ja->bio[i] = kmalloc(struct_size(ja->bio[i], bio.bi_inline_vecs, 1325 nr_bvecs), GFP_KERNEL); 1326 if (!ja->bio[i]) 1327 return -BCH_ERR_ENOMEM_dev_journal_init; 1328 1329 ja->bio[i]->ca = ca; 1330 ja->bio[i]->buf_idx = i; 1331 bio_init(&ja->bio[i]->bio, NULL, ja->bio[i]->bio.bi_inline_vecs, nr_bvecs, 0); 1332 } 1333 1334 ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL); 1335 if (!ja->buckets) 1336 return -BCH_ERR_ENOMEM_dev_journal_init; 1337 1338 if (journal_buckets_v2) { 1339 unsigned nr = bch2_sb_field_journal_v2_nr_entries(journal_buckets_v2); 1340 unsigned dst = 0; 1341 1342 for (unsigned i = 0; i < nr; i++) 1343 for (unsigned j = 0; j < le64_to_cpu(journal_buckets_v2->d[i].nr); j++) 1344 ja->buckets[dst++] = 1345 le64_to_cpu(journal_buckets_v2->d[i].start) + j; 1346 } else if (journal_buckets) { 1347 for (unsigned i = 0; i < ja->nr; i++) 1348 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]); 1349 } 1350 1351 return 0; 1352 } 1353 1354 void bch2_fs_journal_exit(struct journal *j) 1355 { 1356 if (j->wq) 1357 destroy_workqueue(j->wq); 1358 1359 darray_exit(&j->early_journal_entries); 1360 1361 for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++) 1362 kvfree(j->buf[i].data); 1363 free_fifo(&j->pin); 1364 } 1365 1366 int bch2_fs_journal_init(struct journal *j) 1367 { 1368 static struct lock_class_key res_key; 1369 1370 mutex_init(&j->buf_lock); 1371 spin_lock_init(&j->lock); 1372 spin_lock_init(&j->err_lock); 1373 init_waitqueue_head(&j->wait); 1374 INIT_DELAYED_WORK(&j->write_work, journal_write_work); 1375 init_waitqueue_head(&j->reclaim_wait); 1376 init_waitqueue_head(&j->pin_flush_wait); 1377 mutex_init(&j->reclaim_lock); 1378 mutex_init(&j->discard_lock); 1379 1380 lockdep_init_map(&j->res_map, "journal res", &res_key, 0); 1381 1382 atomic64_set(&j->reservations.counter, 1383 ((union journal_res_state) 1384 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v); 1385 1386 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL))) 1387 return -BCH_ERR_ENOMEM_journal_pin_fifo; 1388 1389 for (unsigned i = 0; i < ARRAY_SIZE(j->buf); i++) { 1390 j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN; 1391 j->buf[i].data = kvmalloc(j->buf[i].buf_size, GFP_KERNEL); 1392 if (!j->buf[i].data) 1393 return -BCH_ERR_ENOMEM_journal_buf; 1394 j->buf[i].idx = i; 1395 } 1396 1397 j->pin.front = j->pin.back = 1; 1398 1399 j->wq = alloc_workqueue("bcachefs_journal", 1400 WQ_HIGHPRI|WQ_FREEZABLE|WQ_UNBOUND|WQ_MEM_RECLAIM, 512); 1401 if (!j->wq) 1402 return -BCH_ERR_ENOMEM_fs_other_alloc; 1403 return 0; 1404 } 1405 1406 /* debug: */ 1407 1408 static const char * const bch2_journal_flags_strs[] = { 1409 #define x(n) #n, 1410 JOURNAL_FLAGS() 1411 #undef x 1412 NULL 1413 }; 1414 1415 void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j) 1416 { 1417 struct bch_fs *c = container_of(j, struct bch_fs, journal); 1418 union journal_res_state s; 1419 unsigned long now = jiffies; 1420 u64 nr_writes = j->nr_flush_writes + j->nr_noflush_writes; 1421 1422 printbuf_tabstops_reset(out); 1423 printbuf_tabstop_push(out, 28); 1424 out->atomic++; 1425 1426 rcu_read_lock(); 1427 s = READ_ONCE(j->reservations); 1428 1429 prt_printf(out, "flags:\t"); 1430 prt_bitflags(out, bch2_journal_flags_strs, j->flags); 1431 prt_newline(out); 1432 prt_printf(out, "dirty journal entries:\t%llu/%llu\n", fifo_used(&j->pin), j->pin.size); 1433 prt_printf(out, "seq:\t%llu\n", journal_cur_seq(j)); 1434 prt_printf(out, "seq_ondisk:\t%llu\n", j->seq_ondisk); 1435 prt_printf(out, "last_seq:\t%llu\n", journal_last_seq(j)); 1436 prt_printf(out, "last_seq_ondisk:\t%llu\n", j->last_seq_ondisk); 1437 prt_printf(out, "flushed_seq_ondisk:\t%llu\n", j->flushed_seq_ondisk); 1438 prt_printf(out, "watermark:\t%s\n", bch2_watermarks[j->watermark]); 1439 prt_printf(out, "each entry reserved:\t%u\n", j->entry_u64s_reserved); 1440 prt_printf(out, "nr flush writes:\t%llu\n", j->nr_flush_writes); 1441 prt_printf(out, "nr noflush writes:\t%llu\n", j->nr_noflush_writes); 1442 prt_printf(out, "average write size:\t"); 1443 prt_human_readable_u64(out, nr_writes ? div64_u64(j->entry_bytes_written, nr_writes) : 0); 1444 prt_newline(out); 1445 prt_printf(out, "nr direct reclaim:\t%llu\n", j->nr_direct_reclaim); 1446 prt_printf(out, "nr background reclaim:\t%llu\n", j->nr_background_reclaim); 1447 prt_printf(out, "reclaim kicked:\t%u\n", j->reclaim_kicked); 1448 prt_printf(out, "reclaim runs in:\t%u ms\n", time_after(j->next_reclaim, now) 1449 ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0); 1450 prt_printf(out, "blocked:\t%u\n", j->blocked); 1451 prt_printf(out, "current entry sectors:\t%u\n", j->cur_entry_sectors); 1452 prt_printf(out, "current entry error:\t%s\n", bch2_journal_errors[j->cur_entry_error]); 1453 prt_printf(out, "current entry:\t"); 1454 1455 switch (s.cur_entry_offset) { 1456 case JOURNAL_ENTRY_ERROR_VAL: 1457 prt_printf(out, "error\n"); 1458 break; 1459 case JOURNAL_ENTRY_CLOSED_VAL: 1460 prt_printf(out, "closed\n"); 1461 break; 1462 default: 1463 prt_printf(out, "%u/%u\n", s.cur_entry_offset, j->cur_entry_u64s); 1464 break; 1465 } 1466 1467 prt_printf(out, "unwritten entries:\n"); 1468 bch2_journal_bufs_to_text(out, j); 1469 1470 prt_printf(out, "space:\n"); 1471 printbuf_indent_add(out, 2); 1472 prt_printf(out, "discarded\t%u:%u\n", 1473 j->space[journal_space_discarded].next_entry, 1474 j->space[journal_space_discarded].total); 1475 prt_printf(out, "clean ondisk\t%u:%u\n", 1476 j->space[journal_space_clean_ondisk].next_entry, 1477 j->space[journal_space_clean_ondisk].total); 1478 prt_printf(out, "clean\t%u:%u\n", 1479 j->space[journal_space_clean].next_entry, 1480 j->space[journal_space_clean].total); 1481 prt_printf(out, "total\t%u:%u\n", 1482 j->space[journal_space_total].next_entry, 1483 j->space[journal_space_total].total); 1484 printbuf_indent_sub(out, 2); 1485 1486 for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) { 1487 struct journal_device *ja = &ca->journal; 1488 1489 if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d)) 1490 continue; 1491 1492 if (!ja->nr) 1493 continue; 1494 1495 prt_printf(out, "dev %u:\n", ca->dev_idx); 1496 printbuf_indent_add(out, 2); 1497 prt_printf(out, "nr\t%u\n", ja->nr); 1498 prt_printf(out, "bucket size\t%u\n", ca->mi.bucket_size); 1499 prt_printf(out, "available\t%u:%u\n", bch2_journal_dev_buckets_available(j, ja, journal_space_discarded), ja->sectors_free); 1500 prt_printf(out, "discard_idx\t%u\n", ja->discard_idx); 1501 prt_printf(out, "dirty_ondisk\t%u (seq %llu)\n",ja->dirty_idx_ondisk, ja->bucket_seq[ja->dirty_idx_ondisk]); 1502 prt_printf(out, "dirty_idx\t%u (seq %llu)\n", ja->dirty_idx, ja->bucket_seq[ja->dirty_idx]); 1503 prt_printf(out, "cur_idx\t%u (seq %llu)\n", ja->cur_idx, ja->bucket_seq[ja->cur_idx]); 1504 printbuf_indent_sub(out, 2); 1505 } 1506 1507 rcu_read_unlock(); 1508 1509 --out->atomic; 1510 } 1511 1512 void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j) 1513 { 1514 spin_lock(&j->lock); 1515 __bch2_journal_debug_to_text(out, j); 1516 spin_unlock(&j->lock); 1517 } 1518 1519 bool bch2_journal_seq_pins_to_text(struct printbuf *out, struct journal *j, u64 *seq) 1520 { 1521 struct journal_entry_pin_list *pin_list; 1522 struct journal_entry_pin *pin; 1523 1524 spin_lock(&j->lock); 1525 if (!test_bit(JOURNAL_running, &j->flags)) { 1526 spin_unlock(&j->lock); 1527 return true; 1528 } 1529 1530 *seq = max(*seq, j->pin.front); 1531 1532 if (*seq >= j->pin.back) { 1533 spin_unlock(&j->lock); 1534 return true; 1535 } 1536 1537 out->atomic++; 1538 1539 pin_list = journal_seq_pin(j, *seq); 1540 1541 prt_printf(out, "%llu: count %u\n", *seq, atomic_read(&pin_list->count)); 1542 printbuf_indent_add(out, 2); 1543 1544 for (unsigned i = 0; i < ARRAY_SIZE(pin_list->list); i++) 1545 list_for_each_entry(pin, &pin_list->list[i], list) 1546 prt_printf(out, "\t%px %ps\n", pin, pin->flush); 1547 1548 if (!list_empty(&pin_list->flushed)) 1549 prt_printf(out, "flushed:\n"); 1550 1551 list_for_each_entry(pin, &pin_list->flushed, list) 1552 prt_printf(out, "\t%px %ps\n", pin, pin->flush); 1553 1554 printbuf_indent_sub(out, 2); 1555 1556 --out->atomic; 1557 spin_unlock(&j->lock); 1558 1559 return false; 1560 } 1561 1562 void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j) 1563 { 1564 u64 seq = 0; 1565 1566 while (!bch2_journal_seq_pins_to_text(out, j, &seq)) 1567 seq++; 1568 } 1569