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 /* 52 * Support structures for the flusher threads. 53 */ 54 struct hammer_flusher_info { 55 TAILQ_ENTRY(hammer_flusher_info) entry; 56 struct hammer_mount *hmp; 57 thread_t td; 58 int runstate; 59 int count; 60 hammer_flush_group_t flg; 61 hammer_inode_t work_array[HAMMER_FLUSH_GROUP_SIZE]; 62 }; 63 64 typedef struct hammer_flusher_info *hammer_flusher_info_t; 65 66 /* 67 * Sync all inodes pending on the flusher. 68 * 69 * All flush groups will be flushed. This does not queue dirty inodes 70 * to the flush groups, it just flushes out what has already been queued! 71 */ 72 void 73 hammer_flusher_sync(hammer_mount_t hmp) 74 { 75 int seq; 76 77 seq = hammer_flusher_async(hmp, NULL); 78 hammer_flusher_wait(hmp, seq); 79 } 80 81 /* 82 * Sync all inodes pending on the flusher - return immediately. 83 * 84 * All flush groups will be flushed. 85 */ 86 int 87 hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t close_flg) 88 { 89 hammer_flush_group_t flg; 90 int seq = hmp->flusher.next; 91 92 TAILQ_FOREACH(flg, &hmp->flush_group_list, flush_entry) { 93 if (flg->running == 0) 94 ++seq; 95 flg->closed = 1; 96 if (flg == close_flg) 97 break; 98 } 99 if (hmp->flusher.td) { 100 if (hmp->flusher.signal++ == 0) 101 wakeup(&hmp->flusher.signal); 102 } else { 103 seq = hmp->flusher.done; 104 } 105 return(seq); 106 } 107 108 int 109 hammer_flusher_async_one(hammer_mount_t hmp) 110 { 111 int seq; 112 113 if (hmp->flusher.td) { 114 seq = hmp->flusher.next; 115 if (hmp->flusher.signal++ == 0) 116 wakeup(&hmp->flusher.signal); 117 } else { 118 seq = hmp->flusher.done; 119 } 120 return(seq); 121 } 122 123 /* 124 * Wait for the flusher to get to the specified sequence number. 125 * Signal the flusher as often as necessary to keep it going. 126 */ 127 void 128 hammer_flusher_wait(hammer_mount_t hmp, int seq) 129 { 130 while ((int)(seq - hmp->flusher.done) > 0) { 131 if (hmp->flusher.act != seq) { 132 if (hmp->flusher.signal++ == 0) 133 wakeup(&hmp->flusher.signal); 134 } 135 tsleep(&hmp->flusher.done, 0, "hmrfls", 0); 136 } 137 } 138 139 void 140 hammer_flusher_wait_next(hammer_mount_t hmp) 141 { 142 int seq; 143 144 seq = hammer_flusher_async_one(hmp); 145 hammer_flusher_wait(hmp, seq); 146 } 147 148 void 149 hammer_flusher_create(hammer_mount_t hmp) 150 { 151 hammer_flusher_info_t info; 152 int i; 153 154 hmp->flusher.signal = 0; 155 hmp->flusher.act = 0; 156 hmp->flusher.done = 0; 157 hmp->flusher.next = 1; 158 hammer_ref(&hmp->flusher.finalize_lock); 159 TAILQ_INIT(&hmp->flusher.run_list); 160 TAILQ_INIT(&hmp->flusher.ready_list); 161 162 lwkt_create(hammer_flusher_master_thread, hmp, 163 &hmp->flusher.td, NULL, 0, -1, "hammer-M"); 164 for (i = 0; i < HAMMER_MAX_FLUSHERS; ++i) { 165 info = kmalloc(sizeof(*info), hmp->m_misc, M_WAITOK|M_ZERO); 166 info->hmp = hmp; 167 TAILQ_INSERT_TAIL(&hmp->flusher.ready_list, info, entry); 168 lwkt_create(hammer_flusher_slave_thread, info, 169 &info->td, NULL, 0, -1, "hammer-S%d", i); 170 } 171 } 172 173 void 174 hammer_flusher_destroy(hammer_mount_t hmp) 175 { 176 hammer_flusher_info_t info; 177 178 /* 179 * Kill the master 180 */ 181 hmp->flusher.exiting = 1; 182 while (hmp->flusher.td) { 183 ++hmp->flusher.signal; 184 wakeup(&hmp->flusher.signal); 185 tsleep(&hmp->flusher.exiting, 0, "hmrwex", hz); 186 } 187 188 /* 189 * Kill the slaves 190 */ 191 while ((info = TAILQ_FIRST(&hmp->flusher.ready_list)) != NULL) { 192 KKASSERT(info->runstate == 0); 193 TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry); 194 info->runstate = -1; 195 wakeup(&info->runstate); 196 while (info->td) 197 tsleep(&info->td, 0, "hmrwwc", 0); 198 kfree(info, hmp->m_misc); 199 } 200 } 201 202 /* 203 * The master flusher thread manages the flusher sequence id and 204 * synchronization with the slave work threads. 205 */ 206 static void 207 hammer_flusher_master_thread(void *arg) 208 { 209 hammer_flush_group_t flg; 210 hammer_mount_t hmp; 211 212 hmp = arg; 213 214 for (;;) { 215 /* 216 * Do at least one flush cycle. We may have to update the 217 * UNDO FIFO even if no inodes are queued. 218 */ 219 for (;;) { 220 while (hmp->flusher.group_lock) 221 tsleep(&hmp->flusher.group_lock, 0, "hmrhld", 0); 222 hmp->flusher.act = hmp->flusher.next; 223 ++hmp->flusher.next; 224 hammer_flusher_clean_loose_ios(hmp); 225 hammer_flusher_flush(hmp); 226 hmp->flusher.done = hmp->flusher.act; 227 wakeup(&hmp->flusher.done); 228 flg = TAILQ_FIRST(&hmp->flush_group_list); 229 if (flg == NULL || flg->closed == 0) 230 break; 231 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 232 break; 233 } 234 235 /* 236 * Wait for activity. 237 */ 238 if (hmp->flusher.exiting && TAILQ_EMPTY(&hmp->flush_group_list)) 239 break; 240 while (hmp->flusher.signal == 0) 241 tsleep(&hmp->flusher.signal, 0, "hmrwwa", 0); 242 243 /* 244 * Flush for each count on signal but only allow one extra 245 * flush request to build up. 246 */ 247 if (--hmp->flusher.signal != 0) 248 hmp->flusher.signal = 1; 249 } 250 251 /* 252 * And we are done. 253 */ 254 hmp->flusher.td = NULL; 255 wakeup(&hmp->flusher.exiting); 256 lwkt_exit(); 257 } 258 259 /* 260 * Flush all inodes in the current flush group. 261 */ 262 static void 263 hammer_flusher_flush(hammer_mount_t hmp) 264 { 265 hammer_flusher_info_t info; 266 hammer_flush_group_t flg; 267 hammer_reserve_t resv; 268 hammer_inode_t ip; 269 hammer_inode_t next_ip; 270 int slave_index; 271 int count; 272 273 /* 274 * Just in-case there's a flush race on mount 275 */ 276 if (TAILQ_FIRST(&hmp->flusher.ready_list) == NULL) 277 return; 278 279 /* 280 * We only do one flg but we may have to loop/retry. 281 */ 282 count = 0; 283 while ((flg = TAILQ_FIRST(&hmp->flush_group_list)) != NULL) { 284 ++count; 285 if (hammer_debug_general & 0x0001) { 286 kprintf("hammer_flush %d ttl=%d recs=%d\n", 287 hmp->flusher.act, 288 flg->total_count, flg->refs); 289 } 290 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 291 break; 292 hammer_start_transaction_fls(&hmp->flusher.trans, hmp); 293 294 /* 295 * If the previous flush cycle just about exhausted our 296 * UNDO space we may have to do a dummy cycle to move the 297 * first_offset up before actually digging into a new cycle, 298 * or the new cycle will not have sufficient undo space. 299 */ 300 if (hammer_flusher_undo_exhausted(&hmp->flusher.trans, 3)) 301 hammer_flusher_finalize(&hmp->flusher.trans, 0); 302 303 /* 304 * Ok, we are running this flush group now (this prevents new 305 * additions to it). 306 */ 307 flg->running = 1; 308 if (hmp->next_flush_group == flg) 309 hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry); 310 311 /* 312 * Iterate the inodes in the flg's flush_list and assign 313 * them to slaves. 314 */ 315 slave_index = 0; 316 info = TAILQ_FIRST(&hmp->flusher.ready_list); 317 next_ip = TAILQ_FIRST(&flg->flush_list); 318 319 while ((ip = next_ip) != NULL) { 320 next_ip = TAILQ_NEXT(ip, flush_entry); 321 322 /* 323 * Add ip to the slave's work array. The slave is 324 * not currently running. 325 */ 326 info->work_array[info->count++] = ip; 327 if (info->count != HAMMER_FLUSH_GROUP_SIZE) 328 continue; 329 330 /* 331 * Get the slave running 332 */ 333 TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry); 334 TAILQ_INSERT_TAIL(&hmp->flusher.run_list, info, entry); 335 info->flg = flg; 336 info->runstate = 1; 337 wakeup(&info->runstate); 338 339 /* 340 * Get a new slave. We may have to wait for one to 341 * finish running. 342 */ 343 while ((info = TAILQ_FIRST(&hmp->flusher.ready_list)) == NULL) { 344 tsleep(&hmp->flusher.ready_list, 0, "hmrfcc", 0); 345 } 346 } 347 348 /* 349 * Run the current slave if necessary 350 */ 351 if (info->count) { 352 TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry); 353 TAILQ_INSERT_TAIL(&hmp->flusher.run_list, info, entry); 354 info->flg = flg; 355 info->runstate = 1; 356 wakeup(&info->runstate); 357 } 358 359 /* 360 * Wait for all slaves to finish running 361 */ 362 while (TAILQ_FIRST(&hmp->flusher.run_list) != NULL) 363 tsleep(&hmp->flusher.ready_list, 0, "hmrfcc", 0); 364 365 /* 366 * Do the final finalization, clean up 367 */ 368 hammer_flusher_finalize(&hmp->flusher.trans, 1); 369 hmp->flusher.tid = hmp->flusher.trans.tid; 370 371 hammer_done_transaction(&hmp->flusher.trans); 372 373 /* 374 * Loop up on the same flg. If the flg is done clean it up 375 * and break out. We only flush one flg. 376 */ 377 if (TAILQ_FIRST(&flg->flush_list) == NULL) { 378 KKASSERT(TAILQ_EMPTY(&flg->flush_list)); 379 KKASSERT(flg->refs == 0); 380 TAILQ_REMOVE(&hmp->flush_group_list, flg, flush_entry); 381 kfree(flg, hmp->m_misc); 382 break; 383 } 384 } 385 386 /* 387 * We may have pure meta-data to flush, or we may have to finish 388 * cycling the UNDO FIFO, even if there were no flush groups. 389 */ 390 if (count == 0 && hammer_flusher_haswork(hmp)) { 391 hammer_start_transaction_fls(&hmp->flusher.trans, hmp); 392 hammer_flusher_finalize(&hmp->flusher.trans, 1); 393 hammer_done_transaction(&hmp->flusher.trans); 394 } 395 396 /* 397 * Clean up any freed big-blocks (typically zone-2). 398 * resv->flush_group is typically set several flush groups ahead 399 * of the free to ensure that the freed block is not reused until 400 * it can no longer be reused. 401 */ 402 while ((resv = TAILQ_FIRST(&hmp->delay_list)) != NULL) { 403 if (resv->flush_group != hmp->flusher.act) 404 break; 405 hammer_reserve_clrdelay(hmp, resv); 406 } 407 } 408 409 410 /* 411 * The slave flusher thread pulls work off the master flush_list until no 412 * work is left. 413 */ 414 static void 415 hammer_flusher_slave_thread(void *arg) 416 { 417 hammer_flush_group_t flg; 418 hammer_flusher_info_t info; 419 hammer_mount_t hmp; 420 hammer_inode_t ip; 421 int i; 422 423 info = arg; 424 hmp = info->hmp; 425 426 for (;;) { 427 while (info->runstate == 0) 428 tsleep(&info->runstate, 0, "hmrssw", 0); 429 if (info->runstate < 0) 430 break; 431 flg = info->flg; 432 433 for (i = 0; i < info->count; ++i) { 434 ip = info->work_array[i]; 435 hammer_flusher_flush_inode(ip, &hmp->flusher.trans); 436 ++hammer_stats_inode_flushes; 437 } 438 info->count = 0; 439 info->runstate = 0; 440 TAILQ_REMOVE(&hmp->flusher.run_list, info, entry); 441 TAILQ_INSERT_TAIL(&hmp->flusher.ready_list, info, entry); 442 wakeup(&hmp->flusher.ready_list); 443 } 444 info->td = NULL; 445 wakeup(&info->td); 446 lwkt_exit(); 447 } 448 449 void 450 hammer_flusher_clean_loose_ios(hammer_mount_t hmp) 451 { 452 hammer_buffer_t buffer; 453 hammer_io_t io; 454 455 /* 456 * loose ends - buffers without bp's aren't tracked by the kernel 457 * and can build up, so clean them out. This can occur when an 458 * IO completes on a buffer with no references left. 459 */ 460 if ((io = TAILQ_FIRST(&hmp->lose_list)) != NULL) { 461 crit_enter(); /* biodone() race */ 462 while ((io = TAILQ_FIRST(&hmp->lose_list)) != NULL) { 463 KKASSERT(io->mod_list == &hmp->lose_list); 464 TAILQ_REMOVE(&hmp->lose_list, io, mod_entry); 465 io->mod_list = NULL; 466 if (io->lock.refs == 0) 467 ++hammer_count_refedbufs; 468 hammer_ref(&io->lock); 469 buffer = (void *)io; 470 hammer_rel_buffer(buffer, 0); 471 } 472 crit_exit(); 473 } 474 } 475 476 /* 477 * Flush a single inode that is part of a flush group. 478 * 479 * Flusher errors are extremely serious, even ENOSPC shouldn't occur because 480 * the front-end should have reserved sufficient space on the media. Any 481 * error other then EWOULDBLOCK will force the mount to be read-only. 482 */ 483 static 484 void 485 hammer_flusher_flush_inode(hammer_inode_t ip, hammer_transaction_t trans) 486 { 487 hammer_mount_t hmp = ip->hmp; 488 int error; 489 490 hammer_flusher_clean_loose_ios(hmp); 491 error = hammer_sync_inode(trans, ip); 492 493 /* 494 * EWOULDBLOCK can happen under normal operation, all other errors 495 * are considered extremely serious. We must set WOULDBLOCK 496 * mechanics to deal with the mess left over from the abort of the 497 * previous flush. 498 */ 499 if (error) { 500 ip->flags |= HAMMER_INODE_WOULDBLOCK; 501 if (error == EWOULDBLOCK) 502 error = 0; 503 } 504 hammer_flush_inode_done(ip, error); 505 while (hmp->flusher.finalize_want) 506 tsleep(&hmp->flusher.finalize_want, 0, "hmrsxx", 0); 507 if (hammer_flusher_undo_exhausted(trans, 1)) { 508 kprintf("HAMMER: Warning: UNDO area too small!\n"); 509 hammer_flusher_finalize(trans, 1); 510 } else if (hammer_flusher_meta_limit(trans->hmp)) { 511 hammer_flusher_finalize(trans, 0); 512 } 513 } 514 515 /* 516 * Return non-zero if the UNDO area has less then (QUARTER / 4) of its 517 * space left. 518 * 519 * 1/4 - Emergency free undo space level. Below this point the flusher 520 * will finalize even if directory dependancies have not been resolved. 521 * 522 * 2/4 - Used by the pruning and reblocking code. These functions may be 523 * running in parallel with a flush and cannot be allowed to drop 524 * available undo space to emergency levels. 525 * 526 * 3/4 - Used at the beginning of a flush to force-sync the volume header 527 * to give the flush plenty of runway to work in. 528 */ 529 int 530 hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter) 531 { 532 if (hammer_undo_space(trans) < 533 hammer_undo_max(trans->hmp) * quarter / 4) { 534 return(1); 535 } else { 536 return(0); 537 } 538 } 539 540 /* 541 * Flush all pending UNDOs, wait for write completion, update the volume 542 * header with the new UNDO end position, and flush it. Then 543 * asynchronously flush the meta-data. 544 * 545 * If this is the last finalization in a flush group we also synchronize 546 * our cached blockmap and set hmp->flusher_undo_start and our cached undo 547 * fifo first_offset so the next flush resets the FIFO pointers. 548 * 549 * If this is not final it is being called because too many dirty meta-data 550 * buffers have built up and must be flushed with UNDO synchronization to 551 * avoid a buffer cache deadlock. 552 */ 553 void 554 hammer_flusher_finalize(hammer_transaction_t trans, int final) 555 { 556 hammer_volume_t root_volume; 557 hammer_blockmap_t cundomap, dundomap; 558 hammer_mount_t hmp; 559 hammer_io_t io; 560 int count; 561 int i; 562 563 hmp = trans->hmp; 564 root_volume = trans->rootvol; 565 566 /* 567 * Exclusively lock the flusher. This guarantees that all dirty 568 * buffers will be idled (have a mod-count of 0). 569 */ 570 ++hmp->flusher.finalize_want; 571 hammer_lock_ex(&hmp->flusher.finalize_lock); 572 573 /* 574 * If this isn't the final sync several threads may have hit the 575 * meta-limit at the same time and raced. Only sync if we really 576 * have to, after acquiring the lock. 577 */ 578 if (final == 0 && !hammer_flusher_meta_limit(hmp)) 579 goto done; 580 581 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 582 goto done; 583 584 /* 585 * Flush data buffers. This can occur asynchronously and at any 586 * time. We must interlock against the frontend direct-data write 587 * but do not have to acquire the sync-lock yet. 588 */ 589 count = 0; 590 while ((io = TAILQ_FIRST(&hmp->data_list)) != NULL) { 591 if (io->ioerror) 592 break; 593 if (io->lock.refs == 0) 594 ++hammer_count_refedbufs; 595 hammer_ref(&io->lock); 596 hammer_io_write_interlock(io); 597 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME); 598 hammer_io_flush(io); 599 hammer_io_done_interlock(io); 600 hammer_rel_buffer((hammer_buffer_t)io, 0); 601 ++count; 602 } 603 604 /* 605 * The sync-lock is required for the remaining sequence. This lock 606 * prevents meta-data from being modified. 607 */ 608 hammer_sync_lock_ex(trans); 609 610 /* 611 * If we have been asked to finalize the volume header sync the 612 * cached blockmap to the on-disk blockmap. Generate an UNDO 613 * record for the update. 614 */ 615 if (final) { 616 cundomap = &hmp->blockmap[0]; 617 dundomap = &root_volume->ondisk->vol0_blockmap[0]; 618 if (root_volume->io.modified) { 619 hammer_modify_volume(trans, root_volume, 620 dundomap, sizeof(hmp->blockmap)); 621 for (i = 0; i < HAMMER_MAX_ZONES; ++i) 622 hammer_crc_set_blockmap(&cundomap[i]); 623 bcopy(cundomap, dundomap, sizeof(hmp->blockmap)); 624 hammer_modify_volume_done(root_volume); 625 } 626 } 627 628 /* 629 * Flush UNDOs 630 */ 631 count = 0; 632 while ((io = TAILQ_FIRST(&hmp->undo_list)) != NULL) { 633 if (io->ioerror) 634 break; 635 KKASSERT(io->modify_refs == 0); 636 if (io->lock.refs == 0) 637 ++hammer_count_refedbufs; 638 hammer_ref(&io->lock); 639 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME); 640 hammer_io_flush(io); 641 hammer_rel_buffer((hammer_buffer_t)io, 0); 642 ++count; 643 } 644 645 /* 646 * Wait for I/Os to complete 647 */ 648 hammer_flusher_clean_loose_ios(hmp); 649 hammer_io_wait_all(hmp, "hmrfl1"); 650 651 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 652 goto failed; 653 654 /* 655 * Update the on-disk volume header with new UNDO FIFO end position 656 * (do not generate new UNDO records for this change). We have to 657 * do this for the UNDO FIFO whether (final) is set or not. 658 * 659 * Also update the on-disk next_tid field. This does not require 660 * an UNDO. However, because our TID is generated before we get 661 * the sync lock another sync may have beat us to the punch. 662 * 663 * This also has the side effect of updating first_offset based on 664 * a prior finalization when the first finalization of the next flush 665 * cycle occurs, removing any undo info from the prior finalization 666 * from consideration. 667 * 668 * The volume header will be flushed out synchronously. 669 */ 670 dundomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX]; 671 cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX]; 672 673 if (dundomap->first_offset != cundomap->first_offset || 674 dundomap->next_offset != cundomap->next_offset) { 675 hammer_modify_volume(NULL, root_volume, NULL, 0); 676 dundomap->first_offset = cundomap->first_offset; 677 dundomap->next_offset = cundomap->next_offset; 678 hammer_crc_set_blockmap(dundomap); 679 hammer_modify_volume_done(root_volume); 680 } 681 682 /* 683 * vol0_next_tid is used for TID selection and is updated without 684 * an UNDO so we do not reuse a TID that may have been rolled-back. 685 * 686 * vol0_last_tid is the highest fully-synchronized TID. It is 687 * set-up when the UNDO fifo is fully synced, later on (not here). 688 */ 689 if (root_volume->io.modified) { 690 hammer_modify_volume(NULL, root_volume, NULL, 0); 691 if (root_volume->ondisk->vol0_next_tid < trans->tid) 692 root_volume->ondisk->vol0_next_tid = trans->tid; 693 hammer_crc_set_volume(root_volume->ondisk); 694 hammer_modify_volume_done(root_volume); 695 hammer_io_flush(&root_volume->io); 696 } 697 698 /* 699 * Wait for I/Os to complete 700 */ 701 hammer_flusher_clean_loose_ios(hmp); 702 hammer_io_wait_all(hmp, "hmrfl2"); 703 704 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 705 goto failed; 706 707 /* 708 * Flush meta-data. The meta-data will be undone if we crash 709 * so we can safely flush it asynchronously. 710 * 711 * Repeated catchups will wind up flushing this update's meta-data 712 * and the UNDO buffers for the next update simultaniously. This 713 * is ok. 714 */ 715 count = 0; 716 while ((io = TAILQ_FIRST(&hmp->meta_list)) != NULL) { 717 if (io->ioerror) 718 break; 719 KKASSERT(io->modify_refs == 0); 720 if (io->lock.refs == 0) 721 ++hammer_count_refedbufs; 722 hammer_ref(&io->lock); 723 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME); 724 hammer_io_flush(io); 725 hammer_rel_buffer((hammer_buffer_t)io, 0); 726 ++count; 727 } 728 729 /* 730 * If this is the final finalization for the flush group set 731 * up for the next sequence by setting a new first_offset in 732 * our cached blockmap and clearing the undo history. 733 * 734 * Even though we have updated our cached first_offset, the on-disk 735 * first_offset still governs available-undo-space calculations. 736 */ 737 if (final) { 738 cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX]; 739 if (cundomap->first_offset == cundomap->next_offset) { 740 hmp->hflags &= ~HMNT_UNDO_DIRTY; 741 } else { 742 cundomap->first_offset = cundomap->next_offset; 743 hmp->hflags |= HMNT_UNDO_DIRTY; 744 } 745 hammer_clear_undo_history(hmp); 746 747 /* 748 * Flush tid sequencing. flush_tid1 is fully synchronized, 749 * meaning a crash will not roll it back. flush_tid2 has 750 * been written out asynchronously and a crash will roll 751 * it back. flush_tid1 is used for all mirroring masters. 752 */ 753 if (hmp->flush_tid1 != hmp->flush_tid2) { 754 hmp->flush_tid1 = hmp->flush_tid2; 755 wakeup(&hmp->flush_tid1); 756 } 757 hmp->flush_tid2 = trans->tid; 758 } 759 760 /* 761 * Cleanup. Report any critical errors. 762 */ 763 failed: 764 hammer_sync_unlock(trans); 765 766 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) { 767 kprintf("HAMMER(%s): Critical write error during flush, " 768 "refusing to sync UNDO FIFO\n", 769 root_volume->ondisk->vol_name); 770 } 771 772 done: 773 hammer_unlock(&hmp->flusher.finalize_lock); 774 775 if (--hmp->flusher.finalize_want == 0) 776 wakeup(&hmp->flusher.finalize_want); 777 hammer_stats_commits += final; 778 } 779 780 /* 781 * Return non-zero if too many dirty meta-data buffers have built up. 782 * 783 * Since we cannot allow such buffers to flush until we have dealt with 784 * the UNDOs, we risk deadlocking the kernel's buffer cache. 785 */ 786 int 787 hammer_flusher_meta_limit(hammer_mount_t hmp) 788 { 789 if (hmp->locked_dirty_space + hmp->io_running_space > 790 hammer_limit_dirtybufspace) { 791 return(1); 792 } 793 return(0); 794 } 795 796 /* 797 * Return non-zero if too many dirty meta-data buffers have built up. 798 * 799 * This version is used by background operations (mirror, prune, reblock) 800 * to leave room for foreground operations. 801 */ 802 int 803 hammer_flusher_meta_halflimit(hammer_mount_t hmp) 804 { 805 if (hmp->locked_dirty_space + hmp->io_running_space > 806 hammer_limit_dirtybufspace / 2) { 807 return(1); 808 } 809 return(0); 810 } 811 812 /* 813 * Return non-zero if the flusher still has something to flush. 814 */ 815 int 816 hammer_flusher_haswork(hammer_mount_t hmp) 817 { 818 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 819 return(0); 820 if (TAILQ_FIRST(&hmp->flush_group_list) || /* dirty inodes */ 821 TAILQ_FIRST(&hmp->volu_list) || /* dirty bufffers */ 822 TAILQ_FIRST(&hmp->undo_list) || 823 TAILQ_FIRST(&hmp->data_list) || 824 TAILQ_FIRST(&hmp->meta_list) || 825 (hmp->hflags & HMNT_UNDO_DIRTY) /* UNDO FIFO sync */ 826 ) { 827 return(1); 828 } 829 return(0); 830 } 831 832