1 /* 2 * Copyright (c) 2008 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * $DragonFly: src/sys/vfs/hammer/hammer_flusher.c,v 1.45 2008/07/31 04:42:04 dillon Exp $ 35 */ 36 /* 37 * HAMMER dependancy flusher thread 38 * 39 * Meta data updates create buffer dependancies which are arranged as a 40 * hierarchy of lists. 41 */ 42 43 #include "hammer.h" 44 45 static void hammer_flusher_master_thread(void *arg); 46 static void hammer_flusher_slave_thread(void *arg); 47 static void hammer_flusher_flush(hammer_mount_t hmp); 48 static void hammer_flusher_flush_inode(hammer_inode_t ip, 49 hammer_transaction_t trans); 50 51 RB_GENERATE(hammer_fls_rb_tree, hammer_inode, rb_flsnode, 52 hammer_ino_rb_compare); 53 54 /* 55 * Inodes are sorted and assigned to slave threads in groups of 128. 56 * We want a flush group size large enough such that the slave threads 57 * are not likely to interfere with each other when accessing the B-Tree, 58 * but not so large that we lose concurrency. 59 */ 60 #define HAMMER_FLUSH_GROUP_SIZE 128 61 62 /* 63 * Support structures for the flusher threads. 64 */ 65 struct hammer_flusher_info { 66 TAILQ_ENTRY(hammer_flusher_info) entry; 67 struct hammer_mount *hmp; 68 thread_t td; 69 int runstate; 70 int count; 71 hammer_flush_group_t flg; 72 hammer_inode_t work_array[HAMMER_FLUSH_GROUP_SIZE]; 73 }; 74 75 typedef struct hammer_flusher_info *hammer_flusher_info_t; 76 77 /* 78 * Sync all inodes pending on the flusher. 79 * 80 * All flush groups will be flushed. This does not queue dirty inodes 81 * to the flush groups, it just flushes out what has already been queued! 82 */ 83 void 84 hammer_flusher_sync(hammer_mount_t hmp) 85 { 86 int seq; 87 88 seq = hammer_flusher_async(hmp, NULL); 89 hammer_flusher_wait(hmp, seq); 90 } 91 92 /* 93 * Sync all flush groups through to close_flg - return immediately. 94 * If close_flg is NULL all flush groups are synced. 95 * 96 * Returns the sequence number of the last closed flush group, 97 * which may be close_flg. When syncing to the end if there 98 * are no flush groups pending we still cycle the flusher, so 99 * we return the next seq number not yet allocated. 100 */ 101 int 102 hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t close_flg) 103 { 104 hammer_flush_group_t flg; 105 int seq; 106 107 /* 108 * Already closed 109 */ 110 if (close_flg && close_flg->closed) 111 return(close_flg->seq); 112 113 /* 114 * Close flush groups until we hit the end of the list 115 * or close_flg. 116 */ 117 while ((flg = hmp->next_flush_group) != NULL) { 118 KKASSERT(flg->closed == 0 && flg->running == 0); 119 flg->closed = 1; 120 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry); 121 if (flg == close_flg) 122 break; 123 } 124 125 if (hmp->flusher.td) { 126 if (hmp->flusher.signal++ == 0) 127 wakeup(&hmp->flusher.signal); 128 seq = flg ? flg->seq : hmp->flusher.next; 129 } else { 130 seq = hmp->flusher.done; 131 } 132 return(seq); 133 } 134 135 /* 136 * Flush the current/next flushable flg. This function is typically called 137 * in a loop along with hammer_flusher_wait(hmp, returned_seq) to iterate 138 * flush groups until specific conditions are met. 139 * 140 * If a flush is currently in progress its seq is returned. 141 * 142 * If no flush is currently in progress the next available flush group 143 * will be flushed and its seq returned. 144 * 145 * If no flush groups are present a dummy seq will be allocated and 146 * returned and the flusher will be activated (e.g. to flush the 147 * undo/redo and the volume header). 148 */ 149 int 150 hammer_flusher_async_one(hammer_mount_t hmp) 151 { 152 hammer_flush_group_t flg; 153 int seq; 154 155 if (hmp->flusher.td) { 156 flg = TAILQ_FIRST(&hmp->flush_group_list); 157 seq = hammer_flusher_async(hmp, flg); 158 } else { 159 seq = hmp->flusher.done; 160 } 161 return(seq); 162 } 163 164 /* 165 * Wait for the flusher to get to the specified sequence number. 166 * Signal the flusher as often as necessary to keep it going. 167 */ 168 void 169 hammer_flusher_wait(hammer_mount_t hmp, int seq) 170 { 171 while ((int)(seq - hmp->flusher.done) > 0) { 172 if ((int)(seq - hmp->flusher.act) > 0) { 173 if (hmp->flusher.signal++ == 0) 174 wakeup(&hmp->flusher.signal); 175 } 176 tsleep(&hmp->flusher.done, 0, "hmrfls", 0); 177 } 178 } 179 180 void 181 hammer_flusher_wait_next(hammer_mount_t hmp) 182 { 183 int seq; 184 185 seq = hammer_flusher_async_one(hmp); 186 hammer_flusher_wait(hmp, seq); 187 } 188 189 void 190 hammer_flusher_create(hammer_mount_t hmp) 191 { 192 hammer_flusher_info_t info; 193 int i; 194 195 hmp->flusher.signal = 0; 196 hmp->flusher.act = 0; 197 hmp->flusher.done = 0; 198 hmp->flusher.next = 1; 199 hammer_ref(&hmp->flusher.finalize_lock); 200 TAILQ_INIT(&hmp->flusher.run_list); 201 TAILQ_INIT(&hmp->flusher.ready_list); 202 203 lwkt_create(hammer_flusher_master_thread, hmp, 204 &hmp->flusher.td, NULL, 0, -1, "hammer-M"); 205 for (i = 0; i < HAMMER_MAX_FLUSHERS; ++i) { 206 info = kmalloc(sizeof(*info), hmp->m_misc, M_WAITOK|M_ZERO); 207 info->hmp = hmp; 208 TAILQ_INSERT_TAIL(&hmp->flusher.ready_list, info, entry); 209 lwkt_create(hammer_flusher_slave_thread, info, 210 &info->td, NULL, 0, -1, "hammer-S%d", i); 211 } 212 } 213 214 void 215 hammer_flusher_destroy(hammer_mount_t hmp) 216 { 217 hammer_flusher_info_t info; 218 219 /* 220 * Kill the master 221 */ 222 hmp->flusher.exiting = 1; 223 while (hmp->flusher.td) { 224 ++hmp->flusher.signal; 225 wakeup(&hmp->flusher.signal); 226 tsleep(&hmp->flusher.exiting, 0, "hmrwex", hz); 227 } 228 229 /* 230 * Kill the slaves 231 */ 232 while ((info = TAILQ_FIRST(&hmp->flusher.ready_list)) != NULL) { 233 KKASSERT(info->runstate == 0); 234 TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry); 235 info->runstate = -1; 236 wakeup(&info->runstate); 237 while (info->td) 238 tsleep(&info->td, 0, "hmrwwc", 0); 239 kfree(info, hmp->m_misc); 240 } 241 } 242 243 /* 244 * The master flusher thread manages the flusher sequence id and 245 * synchronization with the slave work threads. 246 */ 247 static void 248 hammer_flusher_master_thread(void *arg) 249 { 250 hammer_flush_group_t flg; 251 hammer_mount_t hmp; 252 253 hmp = arg; 254 255 lwkt_gettoken(&hmp->fs_token); 256 257 for (;;) { 258 /* 259 * Flush all closed flgs. If no flg's are closed we still 260 * do at least one flush cycle as we may have to update 261 * the UNDO FIFO even if no inodes are queued. 262 */ 263 for (;;) { 264 while (hmp->flusher.group_lock) 265 tsleep(&hmp->flusher.group_lock, 0, "hmrhld", 0); 266 hammer_flusher_clean_loose_ios(hmp); 267 hammer_flusher_flush(hmp); 268 hmp->flusher.done = hmp->flusher.act; 269 wakeup(&hmp->flusher.done); 270 flg = TAILQ_FIRST(&hmp->flush_group_list); 271 if (flg == NULL || flg->closed == 0) 272 break; 273 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 274 break; 275 } 276 277 /* 278 * Wait for activity. 279 */ 280 if (hmp->flusher.exiting && TAILQ_EMPTY(&hmp->flush_group_list)) 281 break; 282 while (hmp->flusher.signal == 0) 283 tsleep(&hmp->flusher.signal, 0, "hmrwwa", 0); 284 285 /* 286 * Flush for each count on signal but only allow one extra 287 * flush request to build up. 288 */ 289 if (--hmp->flusher.signal != 0) 290 hmp->flusher.signal = 1; 291 } 292 293 /* 294 * And we are done. 295 */ 296 hmp->flusher.td = NULL; 297 wakeup(&hmp->flusher.exiting); 298 lwkt_reltoken(&hmp->fs_token); 299 lwkt_exit(); 300 } 301 302 /* 303 * Flush all inodes in the current flush group. 304 */ 305 static void 306 hammer_flusher_flush(hammer_mount_t hmp) 307 { 308 hammer_flusher_info_t info; 309 hammer_flush_group_t flg; 310 hammer_reserve_t resv; 311 hammer_inode_t ip; 312 hammer_inode_t next_ip; 313 int slave_index; 314 int count; 315 316 /* 317 * Just in-case there's a flush race on mount 318 */ 319 if (TAILQ_FIRST(&hmp->flusher.ready_list) == NULL) { 320 return; 321 } 322 323 /* 324 * Set the actively flushing sequence number. If no flushable 325 * groups are present allocate a dummy sequence number for the 326 * operation. 327 */ 328 flg = TAILQ_FIRST(&hmp->flush_group_list); 329 if (flg == NULL) { 330 hmp->flusher.act = hmp->flusher.next; 331 ++hmp->flusher.next; 332 } else if (flg->closed) { 333 KKASSERT(flg->running == 0); 334 flg->running = 1; 335 hmp->flusher.act = flg->seq; 336 if (hmp->fill_flush_group == flg) 337 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry); 338 } 339 340 /* 341 * We only do one flg but we may have to loop/retry. 342 * 343 * Due to various races it is possible to come across a flush 344 * group which as not yet been closed. 345 */ 346 count = 0; 347 while (flg && flg->running) { 348 ++count; 349 if (hammer_debug_general & 0x0001) { 350 kprintf("hammer_flush %d ttl=%d recs=%d\n", 351 hmp->flusher.act, 352 flg->total_count, 353 flg->refs); 354 } 355 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 356 break; 357 hammer_start_transaction_fls(&hmp->flusher.trans, hmp); 358 359 /* 360 * If the previous flush cycle just about exhausted our 361 * UNDO space we may have to do a dummy cycle to move the 362 * first_offset up before actually digging into a new cycle, 363 * or the new cycle will not have sufficient undo space. 364 */ 365 if (hammer_flusher_undo_exhausted(&hmp->flusher.trans, 3)) 366 hammer_flusher_finalize(&hmp->flusher.trans, 0); 367 368 KKASSERT(hmp->next_flush_group != flg); 369 370 /* 371 * Iterate the inodes in the flg's flush_tree and assign 372 * them to slaves. 373 */ 374 slave_index = 0; 375 info = TAILQ_FIRST(&hmp->flusher.ready_list); 376 next_ip = RB_FIRST(hammer_fls_rb_tree, &flg->flush_tree); 377 378 while ((ip = next_ip) != NULL) { 379 next_ip = RB_NEXT(hammer_fls_rb_tree, 380 &flg->flush_tree, ip); 381 382 if (++hmp->check_yield > hammer_yield_check) { 383 hmp->check_yield = 0; 384 lwkt_yield(); 385 } 386 387 /* 388 * Add ip to the slave's work array. The slave is 389 * not currently running. 390 */ 391 info->work_array[info->count++] = ip; 392 if (info->count != HAMMER_FLUSH_GROUP_SIZE) 393 continue; 394 395 /* 396 * Get the slave running 397 */ 398 TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry); 399 TAILQ_INSERT_TAIL(&hmp->flusher.run_list, info, entry); 400 info->flg = flg; 401 info->runstate = 1; 402 wakeup(&info->runstate); 403 404 /* 405 * Get a new slave. We may have to wait for one to 406 * finish running. 407 */ 408 while ((info = TAILQ_FIRST(&hmp->flusher.ready_list)) == NULL) { 409 tsleep(&hmp->flusher.ready_list, 0, "hmrfcc", 0); 410 } 411 } 412 413 /* 414 * Run the current slave if necessary 415 */ 416 if (info->count) { 417 TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry); 418 TAILQ_INSERT_TAIL(&hmp->flusher.run_list, info, entry); 419 info->flg = flg; 420 info->runstate = 1; 421 wakeup(&info->runstate); 422 } 423 424 /* 425 * Wait for all slaves to finish running 426 */ 427 while (TAILQ_FIRST(&hmp->flusher.run_list) != NULL) 428 tsleep(&hmp->flusher.ready_list, 0, "hmrfcc", 0); 429 430 /* 431 * Do the final finalization, clean up 432 */ 433 hammer_flusher_finalize(&hmp->flusher.trans, 1); 434 hmp->flusher.tid = hmp->flusher.trans.tid; 435 436 hammer_done_transaction(&hmp->flusher.trans); 437 438 /* 439 * Loop up on the same flg. If the flg is done clean it up 440 * and break out. We only flush one flg. 441 */ 442 if (RB_EMPTY(&flg->flush_tree)) { 443 KKASSERT(flg->refs == 0); 444 TAILQ_REMOVE(&hmp->flush_group_list, flg, flush_entry); 445 kfree(flg, hmp->m_misc); 446 break; 447 } 448 KKASSERT(TAILQ_FIRST(&hmp->flush_group_list) == flg); 449 } 450 451 /* 452 * We may have pure meta-data to flush, or we may have to finish 453 * cycling the UNDO FIFO, even if there were no flush groups. 454 */ 455 if (count == 0 && hammer_flusher_haswork(hmp)) { 456 hammer_start_transaction_fls(&hmp->flusher.trans, hmp); 457 hammer_flusher_finalize(&hmp->flusher.trans, 1); 458 hammer_done_transaction(&hmp->flusher.trans); 459 } 460 461 /* 462 * Clean up any freed big-blocks (typically zone-2). 463 * resv->flush_group is typically set several flush groups ahead 464 * of the free to ensure that the freed block is not reused until 465 * it can no longer be reused. 466 */ 467 while ((resv = TAILQ_FIRST(&hmp->delay_list)) != NULL) { 468 if (resv->flush_group != hmp->flusher.act) 469 break; 470 hammer_reserve_clrdelay(hmp, resv); 471 } 472 } 473 474 475 /* 476 * The slave flusher thread pulls work off the master flush list until no 477 * work is left. 478 */ 479 static void 480 hammer_flusher_slave_thread(void *arg) 481 { 482 hammer_flush_group_t flg; 483 hammer_flusher_info_t info; 484 hammer_mount_t hmp; 485 hammer_inode_t ip; 486 int i; 487 488 info = arg; 489 hmp = info->hmp; 490 lwkt_gettoken(&hmp->fs_token); 491 492 for (;;) { 493 while (info->runstate == 0) 494 tsleep(&info->runstate, 0, "hmrssw", 0); 495 if (info->runstate < 0) 496 break; 497 flg = info->flg; 498 499 for (i = 0; i < info->count; ++i) { 500 ip = info->work_array[i]; 501 hammer_flusher_flush_inode(ip, &hmp->flusher.trans); 502 ++hammer_stats_inode_flushes; 503 } 504 info->count = 0; 505 info->runstate = 0; 506 TAILQ_REMOVE(&hmp->flusher.run_list, info, entry); 507 TAILQ_INSERT_TAIL(&hmp->flusher.ready_list, info, entry); 508 wakeup(&hmp->flusher.ready_list); 509 } 510 info->td = NULL; 511 wakeup(&info->td); 512 lwkt_reltoken(&hmp->fs_token); 513 lwkt_exit(); 514 } 515 516 void 517 hammer_flusher_clean_loose_ios(hammer_mount_t hmp) 518 { 519 hammer_buffer_t buffer; 520 hammer_io_t io; 521 522 /* 523 * loose ends - buffers without bp's aren't tracked by the kernel 524 * and can build up, so clean them out. This can occur when an 525 * IO completes on a buffer with no references left. 526 * 527 * The io_token is needed to protect the list. 528 */ 529 if ((io = RB_ROOT(&hmp->lose_root)) != NULL) { 530 lwkt_gettoken(&hmp->io_token); 531 while ((io = RB_ROOT(&hmp->lose_root)) != NULL) { 532 KKASSERT(io->mod_root == &hmp->lose_root); 533 RB_REMOVE(hammer_mod_rb_tree, io->mod_root, io); 534 io->mod_root = NULL; 535 hammer_ref(&io->lock); 536 buffer = (void *)io; 537 hammer_rel_buffer(buffer, 0); 538 } 539 lwkt_reltoken(&hmp->io_token); 540 } 541 } 542 543 /* 544 * Flush a single inode that is part of a flush group. 545 * 546 * Flusher errors are extremely serious, even ENOSPC shouldn't occur because 547 * the front-end should have reserved sufficient space on the media. Any 548 * error other then EWOULDBLOCK will force the mount to be read-only. 549 */ 550 static 551 void 552 hammer_flusher_flush_inode(hammer_inode_t ip, hammer_transaction_t trans) 553 { 554 hammer_mount_t hmp = ip->hmp; 555 int error; 556 557 hammer_flusher_clean_loose_ios(hmp); 558 error = hammer_sync_inode(trans, ip); 559 560 /* 561 * EWOULDBLOCK can happen under normal operation, all other errors 562 * are considered extremely serious. We must set WOULDBLOCK 563 * mechanics to deal with the mess left over from the abort of the 564 * previous flush. 565 */ 566 if (error) { 567 ip->flags |= HAMMER_INODE_WOULDBLOCK; 568 if (error == EWOULDBLOCK) 569 error = 0; 570 } 571 hammer_flush_inode_done(ip, error); 572 while (hmp->flusher.finalize_want) 573 tsleep(&hmp->flusher.finalize_want, 0, "hmrsxx", 0); 574 if (hammer_flusher_undo_exhausted(trans, 1)) { 575 kprintf("HAMMER: Warning: UNDO area too small!\n"); 576 hammer_flusher_finalize(trans, 1); 577 } else if (hammer_flusher_meta_limit(trans->hmp)) { 578 hammer_flusher_finalize(trans, 0); 579 } 580 } 581 582 /* 583 * Return non-zero if the UNDO area has less then (QUARTER / 4) of its 584 * space left. 585 * 586 * 1/4 - Emergency free undo space level. Below this point the flusher 587 * will finalize even if directory dependancies have not been resolved. 588 * 589 * 2/4 - Used by the pruning and reblocking code. These functions may be 590 * running in parallel with a flush and cannot be allowed to drop 591 * available undo space to emergency levels. 592 * 593 * 3/4 - Used at the beginning of a flush to force-sync the volume header 594 * to give the flush plenty of runway to work in. 595 */ 596 int 597 hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter) 598 { 599 if (hammer_undo_space(trans) < 600 hammer_undo_max(trans->hmp) * quarter / 4) { 601 return(1); 602 } else { 603 return(0); 604 } 605 } 606 607 /* 608 * Flush all pending UNDOs, wait for write completion, update the volume 609 * header with the new UNDO end position, and flush it. Then 610 * asynchronously flush the meta-data. 611 * 612 * If this is the last finalization in a flush group we also synchronize 613 * our cached blockmap and set hmp->flusher_undo_start and our cached undo 614 * fifo first_offset so the next flush resets the FIFO pointers. 615 * 616 * If this is not final it is being called because too many dirty meta-data 617 * buffers have built up and must be flushed with UNDO synchronization to 618 * avoid a buffer cache deadlock. 619 */ 620 void 621 hammer_flusher_finalize(hammer_transaction_t trans, int final) 622 { 623 hammer_volume_t root_volume; 624 hammer_blockmap_t cundomap, dundomap; 625 hammer_mount_t hmp; 626 hammer_io_t io; 627 hammer_off_t save_undo_next_offset; 628 int count; 629 int i; 630 631 hmp = trans->hmp; 632 root_volume = trans->rootvol; 633 634 /* 635 * Exclusively lock the flusher. This guarantees that all dirty 636 * buffers will be idled (have a mod-count of 0). 637 */ 638 ++hmp->flusher.finalize_want; 639 hammer_lock_ex(&hmp->flusher.finalize_lock); 640 641 /* 642 * If this isn't the final sync several threads may have hit the 643 * meta-limit at the same time and raced. Only sync if we really 644 * have to, after acquiring the lock. 645 */ 646 if (final == 0 && !hammer_flusher_meta_limit(hmp)) 647 goto done; 648 649 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 650 goto done; 651 652 /* 653 * Flush data buffers. This can occur asynchronously and at any 654 * time. We must interlock against the frontend direct-data write 655 * but do not have to acquire the sync-lock yet. 656 * 657 * These data buffers have already been collected prior to the 658 * related inode(s) getting queued to the flush group. 659 */ 660 count = 0; 661 while ((io = RB_FIRST(hammer_mod_rb_tree, &hmp->data_root)) != NULL) { 662 if (io->ioerror) 663 break; 664 hammer_ref(&io->lock); 665 hammer_io_write_interlock(io); 666 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME); 667 hammer_io_flush(io, 0); 668 hammer_io_done_interlock(io); 669 hammer_rel_buffer((hammer_buffer_t)io, 0); 670 hammer_io_limit_backlog(hmp); 671 ++count; 672 } 673 674 /* 675 * The sync-lock is required for the remaining sequence. This lock 676 * prevents meta-data from being modified. 677 */ 678 hammer_sync_lock_ex(trans); 679 680 /* 681 * If we have been asked to finalize the volume header sync the 682 * cached blockmap to the on-disk blockmap. Generate an UNDO 683 * record for the update. 684 */ 685 if (final) { 686 cundomap = &hmp->blockmap[0]; 687 dundomap = &root_volume->ondisk->vol0_blockmap[0]; 688 if (root_volume->io.modified) { 689 hammer_modify_volume(trans, root_volume, 690 dundomap, sizeof(hmp->blockmap)); 691 for (i = 0; i < HAMMER_MAX_ZONES; ++i) 692 hammer_crc_set_blockmap(&cundomap[i]); 693 bcopy(cundomap, dundomap, sizeof(hmp->blockmap)); 694 hammer_modify_volume_done(root_volume); 695 } 696 } 697 698 /* 699 * Flush UNDOs. This can occur concurrently with the data flush 700 * because data writes never overwrite. 701 * 702 * This also waits for I/Os to complete and flushes the cache on 703 * the target disk. 704 * 705 * Record the UNDO append point as this can continue to change 706 * after we have flushed the UNDOs. 707 */ 708 cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX]; 709 hammer_lock_ex(&hmp->undo_lock); 710 save_undo_next_offset = cundomap->next_offset; 711 hammer_unlock(&hmp->undo_lock); 712 hammer_flusher_flush_undos(hmp, HAMMER_FLUSH_UNDOS_FORCED); 713 714 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 715 goto failed; 716 717 /* 718 * HAMMER VERSION < 4: 719 * Update the on-disk volume header with new UNDO FIFO end 720 * position (do not generate new UNDO records for this change). 721 * We have to do this for the UNDO FIFO whether (final) is 722 * set or not in order for the UNDOs to be recognized on 723 * recovery. 724 * 725 * HAMMER VERSION >= 4: 726 * The UNDO FIFO data written above will be recognized on 727 * recovery without us having to sync the volume header. 728 * 729 * Also update the on-disk next_tid field. This does not require 730 * an UNDO. However, because our TID is generated before we get 731 * the sync lock another sync may have beat us to the punch. 732 * 733 * This also has the side effect of updating first_offset based on 734 * a prior finalization when the first finalization of the next flush 735 * cycle occurs, removing any undo info from the prior finalization 736 * from consideration. 737 * 738 * The volume header will be flushed out synchronously. 739 */ 740 dundomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX]; 741 cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX]; 742 743 if (dundomap->first_offset != cundomap->first_offset || 744 dundomap->next_offset != save_undo_next_offset) { 745 hammer_modify_volume(NULL, root_volume, NULL, 0); 746 dundomap->first_offset = cundomap->first_offset; 747 dundomap->next_offset = save_undo_next_offset; 748 hammer_crc_set_blockmap(dundomap); 749 hammer_modify_volume_done(root_volume); 750 } 751 752 /* 753 * vol0_next_tid is used for TID selection and is updated without 754 * an UNDO so we do not reuse a TID that may have been rolled-back. 755 * 756 * vol0_last_tid is the highest fully-synchronized TID. It is 757 * set-up when the UNDO fifo is fully synced, later on (not here). 758 * 759 * The root volume can be open for modification by other threads 760 * generating UNDO or REDO records. For example, reblocking, 761 * pruning, REDO mode fast-fsyncs, so the write interlock is 762 * mandatory. 763 */ 764 if (root_volume->io.modified) { 765 hammer_modify_volume(NULL, root_volume, NULL, 0); 766 if (root_volume->ondisk->vol0_next_tid < trans->tid) 767 root_volume->ondisk->vol0_next_tid = trans->tid; 768 hammer_crc_set_volume(root_volume->ondisk); 769 hammer_modify_volume_done(root_volume); 770 hammer_io_write_interlock(&root_volume->io); 771 hammer_io_flush(&root_volume->io, 0); 772 hammer_io_done_interlock(&root_volume->io); 773 } 774 775 /* 776 * Wait for I/Os to complete. 777 * 778 * For HAMMER VERSION 4+ filesystems we do not have to wait for 779 * the I/O to complete as the new UNDO FIFO entries are recognized 780 * even without the volume header update. This allows the volume 781 * header to flushed along with meta-data, significantly reducing 782 * flush overheads. 783 */ 784 hammer_flusher_clean_loose_ios(hmp); 785 if (hmp->version < HAMMER_VOL_VERSION_FOUR) 786 hammer_io_wait_all(hmp, "hmrfl3", 1); 787 788 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 789 goto failed; 790 791 /* 792 * Flush meta-data. The meta-data will be undone if we crash 793 * so we can safely flush it asynchronously. There is no need 794 * to wait for I/O to complete (or issue a synchronous disk flush). 795 * 796 * In fact, even if we did wait the meta-data will still be undone 797 * by a crash up until the next flush cycle due to the first_offset 798 * in the volume header for the UNDO FIFO not being adjusted until 799 * the following flush cycle. 800 * 801 * No io interlock is needed, bioops callbacks will not mess with 802 * meta data buffers. 803 */ 804 count = 0; 805 while ((io = RB_FIRST(hammer_mod_rb_tree, &hmp->meta_root)) != NULL) { 806 if (io->ioerror) 807 break; 808 KKASSERT(io->modify_refs == 0); 809 hammer_ref(&io->lock); 810 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME); 811 hammer_io_flush(io, 0); 812 hammer_rel_buffer((hammer_buffer_t)io, 0); 813 hammer_io_limit_backlog(hmp); 814 ++count; 815 } 816 817 /* 818 * If this is the final finalization for the flush group set 819 * up for the next sequence by setting a new first_offset in 820 * our cached blockmap and clearing the undo history. 821 * 822 * Even though we have updated our cached first_offset, the on-disk 823 * first_offset still governs available-undo-space calculations. 824 * 825 * We synchronize to save_undo_next_offset rather than 826 * cundomap->next_offset because that is what we flushed out 827 * above. 828 * 829 * NOTE! UNDOs can only be added with the sync_lock held 830 * so we can clear the undo history without racing. 831 * REDOs can be added at any time which is why we 832 * have to be careful and use save_undo_next_offset 833 * when setting the new first_offset. 834 */ 835 if (final) { 836 cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX]; 837 if (cundomap->first_offset != save_undo_next_offset) { 838 cundomap->first_offset = save_undo_next_offset; 839 hmp->hflags |= HMNT_UNDO_DIRTY; 840 } else if (cundomap->first_offset != cundomap->next_offset) { 841 hmp->hflags |= HMNT_UNDO_DIRTY; 842 } else { 843 hmp->hflags &= ~HMNT_UNDO_DIRTY; 844 } 845 hammer_clear_undo_history(hmp); 846 847 /* 848 * Flush tid sequencing. flush_tid1 is fully synchronized, 849 * meaning a crash will not roll it back. flush_tid2 has 850 * been written out asynchronously and a crash will roll 851 * it back. flush_tid1 is used for all mirroring masters. 852 */ 853 if (hmp->flush_tid1 != hmp->flush_tid2) { 854 hmp->flush_tid1 = hmp->flush_tid2; 855 wakeup(&hmp->flush_tid1); 856 } 857 hmp->flush_tid2 = trans->tid; 858 859 /* 860 * Clear the REDO SYNC flag. This flag is used to ensure 861 * that the recovery span in the UNDO/REDO FIFO contains 862 * at least one REDO SYNC record. 863 */ 864 hmp->flags &= ~HAMMER_MOUNT_REDO_SYNC; 865 } 866 867 /* 868 * Cleanup. Report any critical errors. 869 */ 870 failed: 871 hammer_sync_unlock(trans); 872 873 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) { 874 kprintf("HAMMER(%s): Critical write error during flush, " 875 "refusing to sync UNDO FIFO\n", 876 root_volume->ondisk->vol_name); 877 } 878 879 done: 880 hammer_unlock(&hmp->flusher.finalize_lock); 881 882 if (--hmp->flusher.finalize_want == 0) 883 wakeup(&hmp->flusher.finalize_want); 884 hammer_stats_commits += final; 885 } 886 887 /* 888 * Flush UNDOs. 889 */ 890 void 891 hammer_flusher_flush_undos(hammer_mount_t hmp, int mode) 892 { 893 hammer_io_t io; 894 int count; 895 896 count = 0; 897 while ((io = RB_FIRST(hammer_mod_rb_tree, &hmp->undo_root)) != NULL) { 898 if (io->ioerror) 899 break; 900 hammer_ref(&io->lock); 901 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME); 902 hammer_io_write_interlock(io); 903 hammer_io_flush(io, hammer_undo_reclaim(io)); 904 hammer_io_done_interlock(io); 905 hammer_rel_buffer((hammer_buffer_t)io, 0); 906 hammer_io_limit_backlog(hmp); 907 ++count; 908 } 909 hammer_flusher_clean_loose_ios(hmp); 910 if (mode == HAMMER_FLUSH_UNDOS_FORCED || 911 (mode == HAMMER_FLUSH_UNDOS_AUTO && count)) { 912 hammer_io_wait_all(hmp, "hmrfl1", 1); 913 } else { 914 hammer_io_wait_all(hmp, "hmrfl2", 0); 915 } 916 } 917 918 /* 919 * Return non-zero if too many dirty meta-data buffers have built up. 920 * 921 * Since we cannot allow such buffers to flush until we have dealt with 922 * the UNDOs, we risk deadlocking the kernel's buffer cache. 923 */ 924 int 925 hammer_flusher_meta_limit(hammer_mount_t hmp) 926 { 927 if (hmp->locked_dirty_space + hmp->io_running_space > 928 hammer_limit_dirtybufspace) { 929 return(1); 930 } 931 return(0); 932 } 933 934 /* 935 * Return non-zero if too many dirty meta-data buffers have built up. 936 * 937 * This version is used by background operations (mirror, prune, reblock) 938 * to leave room for foreground operations. 939 */ 940 int 941 hammer_flusher_meta_halflimit(hammer_mount_t hmp) 942 { 943 if (hmp->locked_dirty_space + hmp->io_running_space > 944 hammer_limit_dirtybufspace / 2) { 945 return(1); 946 } 947 return(0); 948 } 949 950 /* 951 * Return non-zero if the flusher still has something to flush. 952 */ 953 int 954 hammer_flusher_haswork(hammer_mount_t hmp) 955 { 956 if (hmp->ronly) 957 return(0); 958 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 959 return(0); 960 if (TAILQ_FIRST(&hmp->flush_group_list) || /* dirty inodes */ 961 RB_ROOT(&hmp->volu_root) || /* dirty buffers */ 962 RB_ROOT(&hmp->undo_root) || 963 RB_ROOT(&hmp->data_root) || 964 RB_ROOT(&hmp->meta_root) || 965 (hmp->hflags & HMNT_UNDO_DIRTY) /* UNDO FIFO sync */ 966 ) { 967 return(1); 968 } 969 return(0); 970 } 971 972