1 /* 2 * Copyright (c) 2007-2008 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_io.c,v 1.55 2008/09/15 17:02:49 dillon Exp $ 35 */ 36 /* 37 * IO Primitives and buffer cache management 38 * 39 * All major data-tracking structures in HAMMER contain a struct hammer_io 40 * which is used to manage their backing store. We use filesystem buffers 41 * for backing store and we leave them passively associated with their 42 * HAMMER structures. 43 * 44 * If the kernel tries to destroy a passively associated buf which we cannot 45 * yet let go we set B_LOCKED in the buffer and then actively released it 46 * later when we can. 47 */ 48 49 #include "hammer.h" 50 #include <sys/fcntl.h> 51 #include <sys/nlookup.h> 52 #include <sys/buf.h> 53 #include <sys/buf2.h> 54 55 static void hammer_io_modify(hammer_io_t io, int count); 56 static void hammer_io_deallocate(struct buf *bp); 57 #if 0 58 static void hammer_io_direct_read_complete(struct bio *nbio); 59 #endif 60 static void hammer_io_direct_write_complete(struct bio *nbio); 61 static int hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data); 62 static void hammer_io_set_modlist(struct hammer_io *io); 63 static void hammer_io_flush_mark(hammer_volume_t volume); 64 static void hammer_io_flush_sync_done(struct bio *bio); 65 66 67 /* 68 * Initialize a new, already-zero'd hammer_io structure, or reinitialize 69 * an existing hammer_io structure which may have switched to another type. 70 */ 71 void 72 hammer_io_init(hammer_io_t io, hammer_volume_t volume, enum hammer_io_type type) 73 { 74 io->volume = volume; 75 io->hmp = volume->io.hmp; 76 io->type = type; 77 } 78 79 /* 80 * Helper routine to disassociate a buffer cache buffer from an I/O 81 * structure. The buffer is unlocked and marked appropriate for reclamation. 82 * 83 * The io may have 0 or 1 references depending on who called us. The 84 * caller is responsible for dealing with the refs. 85 * 86 * This call can only be made when no action is required on the buffer. 87 * 88 * The caller must own the buffer and the IO must indicate that the 89 * structure no longer owns it (io.released != 0). 90 */ 91 static void 92 hammer_io_disassociate(hammer_io_structure_t iou) 93 { 94 struct buf *bp = iou->io.bp; 95 96 KKASSERT(iou->io.released); 97 KKASSERT(iou->io.modified == 0); 98 KKASSERT(LIST_FIRST(&bp->b_dep) == (void *)iou); 99 buf_dep_init(bp); 100 iou->io.bp = NULL; 101 102 /* 103 * If the buffer was locked someone wanted to get rid of it. 104 */ 105 if (bp->b_flags & B_LOCKED) { 106 --hammer_count_io_locked; 107 bp->b_flags &= ~B_LOCKED; 108 } 109 if (iou->io.reclaim) { 110 bp->b_flags |= B_NOCACHE|B_RELBUF; 111 iou->io.reclaim = 0; 112 } 113 114 switch(iou->io.type) { 115 case HAMMER_STRUCTURE_VOLUME: 116 iou->volume.ondisk = NULL; 117 break; 118 case HAMMER_STRUCTURE_DATA_BUFFER: 119 case HAMMER_STRUCTURE_META_BUFFER: 120 case HAMMER_STRUCTURE_UNDO_BUFFER: 121 iou->buffer.ondisk = NULL; 122 break; 123 } 124 } 125 126 /* 127 * Wait for any physical IO to complete 128 */ 129 void 130 hammer_io_wait(hammer_io_t io) 131 { 132 if (io->running) { 133 crit_enter(); 134 tsleep_interlock(io); 135 io->waiting = 1; 136 for (;;) { 137 tsleep(io, 0, "hmrflw", 0); 138 if (io->running == 0) 139 break; 140 tsleep_interlock(io); 141 io->waiting = 1; 142 if (io->running == 0) 143 break; 144 } 145 crit_exit(); 146 } 147 } 148 149 /* 150 * Wait for all hammer_io-initated write I/O's to complete. This is not 151 * supposed to count direct I/O's but some can leak through (for 152 * non-full-sized direct I/Os). 153 */ 154 void 155 hammer_io_wait_all(hammer_mount_t hmp, const char *ident) 156 { 157 hammer_io_flush_sync(hmp); 158 crit_enter(); 159 while (hmp->io_running_space) 160 tsleep(&hmp->io_running_space, 0, ident, 0); 161 crit_exit(); 162 } 163 164 #define HAMMER_MAXRA 4 165 166 /* 167 * Load bp for a HAMMER structure. The io must be exclusively locked by 168 * the caller. 169 * 170 * This routine is mostly used on meta-data and small-data blocks. Generally 171 * speaking HAMMER assumes some locality of reference and will cluster 172 * a 64K read. 173 * 174 * Note that clustering occurs at the device layer, not the logical layer. 175 * If the buffers do not apply to the current operation they may apply to 176 * some other. 177 */ 178 int 179 hammer_io_read(struct vnode *devvp, struct hammer_io *io, hammer_off_t limit) 180 { 181 struct buf *bp; 182 int error; 183 184 if ((bp = io->bp) == NULL) { 185 hammer_count_io_running_read += io->bytes; 186 if (hammer_cluster_enable) { 187 error = cluster_read(devvp, limit, 188 io->offset, io->bytes, 189 HAMMER_CLUSTER_SIZE, 190 HAMMER_CLUSTER_BUFS, &io->bp); 191 } else { 192 error = bread(devvp, io->offset, io->bytes, &io->bp); 193 } 194 hammer_stats_disk_read += io->bytes; 195 hammer_count_io_running_read -= io->bytes; 196 197 /* 198 * The code generally assumes b_ops/b_dep has been set-up, 199 * even if we error out here. 200 */ 201 bp = io->bp; 202 bp->b_ops = &hammer_bioops; 203 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 204 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node); 205 BUF_KERNPROC(bp); 206 KKASSERT(io->modified == 0); 207 KKASSERT(io->running == 0); 208 KKASSERT(io->waiting == 0); 209 io->released = 0; /* we hold an active lock on bp */ 210 } else { 211 error = 0; 212 } 213 return(error); 214 } 215 216 /* 217 * Similar to hammer_io_read() but returns a zero'd out buffer instead. 218 * Must be called with the IO exclusively locked. 219 * 220 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background 221 * I/O by forcing the buffer to not be in a released state before calling 222 * it. 223 * 224 * This function will also mark the IO as modified but it will not 225 * increment the modify_refs count. 226 */ 227 int 228 hammer_io_new(struct vnode *devvp, struct hammer_io *io) 229 { 230 struct buf *bp; 231 232 if ((bp = io->bp) == NULL) { 233 io->bp = getblk(devvp, io->offset, io->bytes, 0, 0); 234 bp = io->bp; 235 bp->b_ops = &hammer_bioops; 236 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 237 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node); 238 io->released = 0; 239 KKASSERT(io->running == 0); 240 io->waiting = 0; 241 BUF_KERNPROC(bp); 242 } else { 243 if (io->released) { 244 regetblk(bp); 245 BUF_KERNPROC(bp); 246 io->released = 0; 247 } 248 } 249 hammer_io_modify(io, 0); 250 vfs_bio_clrbuf(bp); 251 return(0); 252 } 253 254 /* 255 * Remove potential device level aliases against buffers managed by high level 256 * vnodes. Aliases can also be created due to mixed buffer sizes or via 257 * direct access to the backing store device. 258 * 259 * This is nasty because the buffers are also VMIO-backed. Even if a buffer 260 * does not exist its backing VM pages might, and we have to invalidate 261 * those as well or a getblk() will reinstate them. 262 * 263 * Buffer cache buffers associated with hammer_buffers cannot be 264 * invalidated. 265 */ 266 int 267 hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset) 268 { 269 hammer_io_structure_t iou; 270 hammer_off_t phys_offset; 271 struct buf *bp; 272 int error; 273 274 phys_offset = volume->ondisk->vol_buf_beg + 275 (zone2_offset & HAMMER_OFF_SHORT_MASK); 276 crit_enter(); 277 if ((bp = findblk(volume->devvp, phys_offset)) != NULL) 278 bp = getblk(volume->devvp, phys_offset, bp->b_bufsize, 0, 0); 279 else 280 bp = getblk(volume->devvp, phys_offset, HAMMER_BUFSIZE, 0, 0); 281 if ((iou = (void *)LIST_FIRST(&bp->b_dep)) != NULL) { 282 #if 0 283 hammer_ref(&iou->io.lock); 284 hammer_io_clear_modify(&iou->io, 1); 285 bundirty(bp); 286 iou->io.released = 0; 287 BUF_KERNPROC(bp); 288 iou->io.reclaim = 1; 289 iou->io.waitdep = 1; 290 KKASSERT(iou->io.lock.refs == 1); 291 hammer_rel_buffer(&iou->buffer, 0); 292 /*hammer_io_deallocate(bp);*/ 293 #endif 294 bqrelse(bp); 295 error = EAGAIN; 296 } else { 297 KKASSERT((bp->b_flags & B_LOCKED) == 0); 298 bundirty(bp); 299 bp->b_flags |= B_NOCACHE|B_RELBUF; 300 brelse(bp); 301 error = 0; 302 } 303 crit_exit(); 304 return(error); 305 } 306 307 /* 308 * This routine is called on the last reference to a hammer structure. 309 * The io is usually interlocked with io.loading and io.refs must be 1. 310 * 311 * This routine may return a non-NULL bp to the caller for dispoal. Disposal 312 * simply means the caller finishes decrementing the ref-count on the 313 * IO structure then brelse()'s the bp. The bp may or may not still be 314 * passively associated with the IO. 315 * 316 * The only requirement here is that modified meta-data and volume-header 317 * buffer may NOT be disassociated from the IO structure, and consequently 318 * we also leave such buffers actively associated with the IO if they already 319 * are (since the kernel can't do anything with them anyway). Only the 320 * flusher is allowed to write such buffers out. Modified pure-data and 321 * undo buffers are returned to the kernel but left passively associated 322 * so we can track when the kernel writes the bp out. 323 */ 324 struct buf * 325 hammer_io_release(struct hammer_io *io, int flush) 326 { 327 union hammer_io_structure *iou = (void *)io; 328 struct buf *bp; 329 330 if ((bp = io->bp) == NULL) 331 return(NULL); 332 333 /* 334 * Try to flush a dirty IO to disk if asked to by the 335 * caller or if the kernel tried to flush the buffer in the past. 336 * 337 * Kernel-initiated flushes are only allowed for pure-data buffers. 338 * meta-data and volume buffers can only be flushed explicitly 339 * by HAMMER. 340 */ 341 if (io->modified) { 342 if (flush) { 343 hammer_io_flush(io); 344 } else if (bp->b_flags & B_LOCKED) { 345 switch(io->type) { 346 case HAMMER_STRUCTURE_DATA_BUFFER: 347 case HAMMER_STRUCTURE_UNDO_BUFFER: 348 hammer_io_flush(io); 349 break; 350 default: 351 break; 352 } 353 } /* else no explicit request to flush the buffer */ 354 } 355 356 /* 357 * Wait for the IO to complete if asked to. This occurs when 358 * the buffer must be disposed of definitively during an umount 359 * or buffer invalidation. 360 */ 361 if (io->waitdep && io->running) { 362 hammer_io_wait(io); 363 } 364 365 /* 366 * Return control of the buffer to the kernel (with the provisio 367 * that our bioops can override kernel decisions with regards to 368 * the buffer). 369 */ 370 if ((flush || io->reclaim) && io->modified == 0 && io->running == 0) { 371 /* 372 * Always disassociate the bp if an explicit flush 373 * was requested and the IO completed with no error 374 * (so unmount can really clean up the structure). 375 */ 376 if (io->released) { 377 regetblk(bp); 378 BUF_KERNPROC(bp); 379 } else { 380 io->released = 1; 381 } 382 hammer_io_disassociate((hammer_io_structure_t)io); 383 /* return the bp */ 384 } else if (io->modified) { 385 /* 386 * Only certain IO types can be released to the kernel if 387 * the buffer has been modified. 388 * 389 * volume and meta-data IO types may only be explicitly 390 * flushed by HAMMER. 391 */ 392 switch(io->type) { 393 case HAMMER_STRUCTURE_DATA_BUFFER: 394 case HAMMER_STRUCTURE_UNDO_BUFFER: 395 if (io->released == 0) { 396 io->released = 1; 397 bdwrite(bp); 398 } 399 break; 400 default: 401 break; 402 } 403 bp = NULL; /* bp left associated */ 404 } else if (io->released == 0) { 405 /* 406 * Clean buffers can be generally released to the kernel. 407 * We leave the bp passively associated with the HAMMER 408 * structure and use bioops to disconnect it later on 409 * if the kernel wants to discard the buffer. 410 * 411 * We can steal the structure's ownership of the bp. 412 */ 413 io->released = 1; 414 if (bp->b_flags & B_LOCKED) { 415 hammer_io_disassociate(iou); 416 /* return the bp */ 417 } else { 418 if (io->reclaim) { 419 hammer_io_disassociate(iou); 420 /* return the bp */ 421 } else { 422 /* return the bp (bp passively associated) */ 423 } 424 } 425 } else { 426 /* 427 * A released buffer is passively associate with our 428 * hammer_io structure. The kernel cannot destroy it 429 * without making a bioops call. If the kernel (B_LOCKED) 430 * or we (reclaim) requested that the buffer be destroyed 431 * we destroy it, otherwise we do a quick get/release to 432 * reset its position in the kernel's LRU list. 433 * 434 * Leaving the buffer passively associated allows us to 435 * use the kernel's LRU buffer flushing mechanisms rather 436 * then rolling our own. 437 * 438 * XXX there are two ways of doing this. We can re-acquire 439 * and passively release to reset the LRU, or not. 440 */ 441 if (io->running == 0) { 442 regetblk(bp); 443 if ((bp->b_flags & B_LOCKED) || io->reclaim) { 444 hammer_io_disassociate(iou); 445 /* return the bp */ 446 } else { 447 /* return the bp (bp passively associated) */ 448 } 449 } else { 450 /* 451 * bp is left passively associated but we do not 452 * try to reacquire it. Interactions with the io 453 * structure will occur on completion of the bp's 454 * I/O. 455 */ 456 bp = NULL; 457 } 458 } 459 return(bp); 460 } 461 462 /* 463 * This routine is called with a locked IO when a flush is desired and 464 * no other references to the structure exists other then ours. This 465 * routine is ONLY called when HAMMER believes it is safe to flush a 466 * potentially modified buffer out. 467 */ 468 void 469 hammer_io_flush(struct hammer_io *io) 470 { 471 struct buf *bp; 472 473 /* 474 * Degenerate case - nothing to flush if nothing is dirty. 475 */ 476 if (io->modified == 0) { 477 return; 478 } 479 480 KKASSERT(io->bp); 481 KKASSERT(io->modify_refs <= 0); 482 483 /* 484 * Acquire ownership of the bp, particularly before we clear our 485 * modified flag. 486 * 487 * We are going to bawrite() this bp. Don't leave a window where 488 * io->released is set, we actually own the bp rather then our 489 * buffer. 490 */ 491 bp = io->bp; 492 if (io->released) { 493 regetblk(bp); 494 /* BUF_KERNPROC(io->bp); */ 495 /* io->released = 0; */ 496 KKASSERT(io->released); 497 KKASSERT(io->bp == bp); 498 } 499 io->released = 1; 500 501 /* 502 * Acquire exclusive access to the bp and then clear the modified 503 * state of the buffer prior to issuing I/O to interlock any 504 * modifications made while the I/O is in progress. This shouldn't 505 * happen anyway but losing data would be worse. The modified bit 506 * will be rechecked after the IO completes. 507 * 508 * NOTE: This call also finalizes the buffer's content (inval == 0). 509 * 510 * This is only legal when lock.refs == 1 (otherwise we might clear 511 * the modified bit while there are still users of the cluster 512 * modifying the data). 513 * 514 * Do this before potentially blocking so any attempt to modify the 515 * ondisk while we are blocked blocks waiting for us. 516 */ 517 hammer_ref(&io->lock); 518 hammer_io_clear_modify(io, 0); 519 hammer_unref(&io->lock); 520 521 /* 522 * Transfer ownership to the kernel and initiate I/O. 523 */ 524 io->running = 1; 525 io->hmp->io_running_space += io->bytes; 526 hammer_count_io_running_write += io->bytes; 527 bawrite(bp); 528 hammer_io_flush_mark(io->volume); 529 } 530 531 /************************************************************************ 532 * BUFFER DIRTYING * 533 ************************************************************************ 534 * 535 * These routines deal with dependancies created when IO buffers get 536 * modified. The caller must call hammer_modify_*() on a referenced 537 * HAMMER structure prior to modifying its on-disk data. 538 * 539 * Any intent to modify an IO buffer acquires the related bp and imposes 540 * various write ordering dependancies. 541 */ 542 543 /* 544 * Mark a HAMMER structure as undergoing modification. Meta-data buffers 545 * are locked until the flusher can deal with them, pure data buffers 546 * can be written out. 547 */ 548 static 549 void 550 hammer_io_modify(hammer_io_t io, int count) 551 { 552 /* 553 * io->modify_refs must be >= 0 554 */ 555 while (io->modify_refs < 0) { 556 io->waitmod = 1; 557 tsleep(io, 0, "hmrmod", 0); 558 } 559 560 /* 561 * Shortcut if nothing to do. 562 */ 563 KKASSERT(io->lock.refs != 0 && io->bp != NULL); 564 io->modify_refs += count; 565 if (io->modified && io->released == 0) 566 return; 567 568 hammer_lock_ex(&io->lock); 569 if (io->modified == 0) { 570 hammer_io_set_modlist(io); 571 io->modified = 1; 572 } 573 if (io->released) { 574 regetblk(io->bp); 575 BUF_KERNPROC(io->bp); 576 io->released = 0; 577 KKASSERT(io->modified != 0); 578 } 579 hammer_unlock(&io->lock); 580 } 581 582 static __inline 583 void 584 hammer_io_modify_done(hammer_io_t io) 585 { 586 KKASSERT(io->modify_refs > 0); 587 --io->modify_refs; 588 if (io->modify_refs == 0 && io->waitmod) { 589 io->waitmod = 0; 590 wakeup(io); 591 } 592 } 593 594 void 595 hammer_io_write_interlock(hammer_io_t io) 596 { 597 while (io->modify_refs != 0) { 598 io->waitmod = 1; 599 tsleep(io, 0, "hmrmod", 0); 600 } 601 io->modify_refs = -1; 602 } 603 604 void 605 hammer_io_done_interlock(hammer_io_t io) 606 { 607 KKASSERT(io->modify_refs == -1); 608 io->modify_refs = 0; 609 if (io->waitmod) { 610 io->waitmod = 0; 611 wakeup(io); 612 } 613 } 614 615 /* 616 * Caller intends to modify a volume's ondisk structure. 617 * 618 * This is only allowed if we are the flusher or we have a ref on the 619 * sync_lock. 620 */ 621 void 622 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume, 623 void *base, int len) 624 { 625 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 626 627 hammer_io_modify(&volume->io, 1); 628 if (len) { 629 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk; 630 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 631 hammer_generate_undo(trans, &volume->io, 632 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset), 633 base, len); 634 } 635 } 636 637 /* 638 * Caller intends to modify a buffer's ondisk structure. 639 * 640 * This is only allowed if we are the flusher or we have a ref on the 641 * sync_lock. 642 */ 643 void 644 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer, 645 void *base, int len) 646 { 647 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 648 649 hammer_io_modify(&buffer->io, 1); 650 if (len) { 651 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk; 652 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 653 hammer_generate_undo(trans, &buffer->io, 654 buffer->zone2_offset + rel_offset, 655 base, len); 656 } 657 } 658 659 void 660 hammer_modify_volume_done(hammer_volume_t volume) 661 { 662 hammer_io_modify_done(&volume->io); 663 } 664 665 void 666 hammer_modify_buffer_done(hammer_buffer_t buffer) 667 { 668 hammer_io_modify_done(&buffer->io); 669 } 670 671 /* 672 * Mark an entity as not being dirty any more and finalize any 673 * delayed adjustments to the buffer. 674 * 675 * Delayed adjustments are an important performance enhancement, allowing 676 * us to avoid recalculating B-Tree node CRCs over and over again when 677 * making bulk-modifications to the B-Tree. 678 * 679 * If inval is non-zero delayed adjustments are ignored. 680 * 681 * This routine may dereference related btree nodes and cause the 682 * buffer to be dereferenced. The caller must own a reference on io. 683 */ 684 void 685 hammer_io_clear_modify(struct hammer_io *io, int inval) 686 { 687 if (io->modified == 0) 688 return; 689 690 /* 691 * Take us off the mod-list and clear the modified bit. 692 */ 693 KKASSERT(io->mod_list != NULL); 694 if (io->mod_list == &io->hmp->volu_list || 695 io->mod_list == &io->hmp->meta_list) { 696 io->hmp->locked_dirty_space -= io->bytes; 697 hammer_count_dirtybufspace -= io->bytes; 698 } 699 TAILQ_REMOVE(io->mod_list, io, mod_entry); 700 io->mod_list = NULL; 701 io->modified = 0; 702 703 /* 704 * If this bit is not set there are no delayed adjustments. 705 */ 706 if (io->gencrc == 0) 707 return; 708 io->gencrc = 0; 709 710 /* 711 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference 712 * on the node (& underlying buffer). Release the node after clearing 713 * the flag. 714 */ 715 if (io->type == HAMMER_STRUCTURE_META_BUFFER) { 716 hammer_buffer_t buffer = (void *)io; 717 hammer_node_t node; 718 719 restart: 720 TAILQ_FOREACH(node, &buffer->clist, entry) { 721 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) 722 continue; 723 node->flags &= ~HAMMER_NODE_NEEDSCRC; 724 KKASSERT(node->ondisk); 725 if (inval == 0) 726 node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE); 727 hammer_rel_node(node); 728 goto restart; 729 } 730 } 731 /* caller must still have ref on io */ 732 KKASSERT(io->lock.refs > 0); 733 } 734 735 /* 736 * Clear the IO's modify list. Even though the IO is no longer modified 737 * it may still be on the lose_list. This routine is called just before 738 * the governing hammer_buffer is destroyed. 739 */ 740 void 741 hammer_io_clear_modlist(struct hammer_io *io) 742 { 743 KKASSERT(io->modified == 0); 744 if (io->mod_list) { 745 crit_enter(); /* biodone race against list */ 746 KKASSERT(io->mod_list == &io->hmp->lose_list); 747 TAILQ_REMOVE(io->mod_list, io, mod_entry); 748 io->mod_list = NULL; 749 crit_exit(); 750 } 751 } 752 753 static void 754 hammer_io_set_modlist(struct hammer_io *io) 755 { 756 struct hammer_mount *hmp = io->hmp; 757 758 KKASSERT(io->mod_list == NULL); 759 760 switch(io->type) { 761 case HAMMER_STRUCTURE_VOLUME: 762 io->mod_list = &hmp->volu_list; 763 hmp->locked_dirty_space += io->bytes; 764 hammer_count_dirtybufspace += io->bytes; 765 break; 766 case HAMMER_STRUCTURE_META_BUFFER: 767 io->mod_list = &hmp->meta_list; 768 hmp->locked_dirty_space += io->bytes; 769 hammer_count_dirtybufspace += io->bytes; 770 break; 771 case HAMMER_STRUCTURE_UNDO_BUFFER: 772 io->mod_list = &hmp->undo_list; 773 break; 774 case HAMMER_STRUCTURE_DATA_BUFFER: 775 io->mod_list = &hmp->data_list; 776 break; 777 } 778 TAILQ_INSERT_TAIL(io->mod_list, io, mod_entry); 779 } 780 781 /************************************************************************ 782 * HAMMER_BIOOPS * 783 ************************************************************************ 784 * 785 */ 786 787 /* 788 * Pre-IO initiation kernel callback - cluster build only 789 */ 790 static void 791 hammer_io_start(struct buf *bp) 792 { 793 } 794 795 /* 796 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT! 797 * 798 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit 799 * may also be set if we were marking a cluster header open. Only remove 800 * our dependancy if the modified bit is clear. 801 */ 802 static void 803 hammer_io_complete(struct buf *bp) 804 { 805 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep); 806 807 KKASSERT(iou->io.released == 1); 808 809 /* 810 * Deal with people waiting for I/O to drain 811 */ 812 if (iou->io.running) { 813 /* 814 * Deal with critical write errors. Once a critical error 815 * has been flagged in hmp the UNDO FIFO will not be updated. 816 * That way crash recover will give us a consistent 817 * filesystem. 818 * 819 * Because of this we can throw away failed UNDO buffers. If 820 * we throw away META or DATA buffers we risk corrupting 821 * the now read-only version of the filesystem visible to 822 * the user. Clear B_ERROR so the buffer is not re-dirtied 823 * by the kernel and ref the io so it doesn't get thrown 824 * away. 825 */ 826 if (bp->b_flags & B_ERROR) { 827 hammer_critical_error(iou->io.hmp, NULL, bp->b_error, 828 "while flushing meta-data"); 829 switch(iou->io.type) { 830 case HAMMER_STRUCTURE_UNDO_BUFFER: 831 break; 832 default: 833 if (iou->io.ioerror == 0) { 834 iou->io.ioerror = 1; 835 if (iou->io.lock.refs == 0) 836 ++hammer_count_refedbufs; 837 hammer_ref(&iou->io.lock); 838 } 839 break; 840 } 841 bp->b_flags &= ~B_ERROR; 842 bundirty(bp); 843 #if 0 844 hammer_io_set_modlist(&iou->io); 845 iou->io.modified = 1; 846 #endif 847 } 848 hammer_stats_disk_write += iou->io.bytes; 849 hammer_count_io_running_write -= iou->io.bytes; 850 iou->io.hmp->io_running_space -= iou->io.bytes; 851 if (iou->io.hmp->io_running_space == 0) 852 wakeup(&iou->io.hmp->io_running_space); 853 KKASSERT(iou->io.hmp->io_running_space >= 0); 854 iou->io.running = 0; 855 } else { 856 hammer_stats_disk_read += iou->io.bytes; 857 } 858 859 if (iou->io.waiting) { 860 iou->io.waiting = 0; 861 wakeup(iou); 862 } 863 864 /* 865 * If B_LOCKED is set someone wanted to deallocate the bp at some 866 * point, do it now if refs has become zero. 867 */ 868 if ((bp->b_flags & B_LOCKED) && iou->io.lock.refs == 0) { 869 KKASSERT(iou->io.modified == 0); 870 --hammer_count_io_locked; 871 bp->b_flags &= ~B_LOCKED; 872 hammer_io_deallocate(bp); 873 /* structure may be dead now */ 874 } 875 } 876 877 /* 878 * Callback from kernel when it wishes to deallocate a passively 879 * associated structure. This mostly occurs with clean buffers 880 * but it may be possible for a holding structure to be marked dirty 881 * while its buffer is passively associated. The caller owns the bp. 882 * 883 * If we cannot disassociate we set B_LOCKED to prevent the buffer 884 * from getting reused. 885 * 886 * WARNING: Because this can be called directly by getnewbuf we cannot 887 * recurse into the tree. If a bp cannot be immediately disassociated 888 * our only recourse is to set B_LOCKED. 889 * 890 * WARNING: This may be called from an interrupt via hammer_io_complete() 891 */ 892 static void 893 hammer_io_deallocate(struct buf *bp) 894 { 895 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep); 896 897 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0); 898 if (iou->io.lock.refs > 0 || iou->io.modified) { 899 /* 900 * It is not legal to disassociate a modified buffer. This 901 * case really shouldn't ever occur. 902 */ 903 bp->b_flags |= B_LOCKED; 904 ++hammer_count_io_locked; 905 } else { 906 /* 907 * Disassociate the BP. If the io has no refs left we 908 * have to add it to the loose list. 909 */ 910 hammer_io_disassociate(iou); 911 if (iou->io.type != HAMMER_STRUCTURE_VOLUME) { 912 KKASSERT(iou->io.bp == NULL); 913 KKASSERT(iou->io.mod_list == NULL); 914 crit_enter(); /* biodone race against list */ 915 iou->io.mod_list = &iou->io.hmp->lose_list; 916 TAILQ_INSERT_TAIL(iou->io.mod_list, &iou->io, mod_entry); 917 crit_exit(); 918 } 919 } 920 } 921 922 static int 923 hammer_io_fsync(struct vnode *vp) 924 { 925 return(0); 926 } 927 928 /* 929 * NOTE: will not be called unless we tell the kernel about the 930 * bioops. Unused... we use the mount's VFS_SYNC instead. 931 */ 932 static int 933 hammer_io_sync(struct mount *mp) 934 { 935 return(0); 936 } 937 938 static void 939 hammer_io_movedeps(struct buf *bp1, struct buf *bp2) 940 { 941 } 942 943 /* 944 * I/O pre-check for reading and writing. HAMMER only uses this for 945 * B_CACHE buffers so checkread just shouldn't happen, but if it does 946 * allow it. 947 * 948 * Writing is a different case. We don't want the kernel to try to write 949 * out a buffer that HAMMER may be modifying passively or which has a 950 * dependancy. In addition, kernel-demanded writes can only proceed for 951 * certain types of buffers (i.e. UNDO and DATA types). Other dirty 952 * buffer types can only be explicitly written by the flusher. 953 * 954 * checkwrite will only be called for bdwrite()n buffers. If we return 955 * success the kernel is guaranteed to initiate the buffer write. 956 */ 957 static int 958 hammer_io_checkread(struct buf *bp) 959 { 960 return(0); 961 } 962 963 static int 964 hammer_io_checkwrite(struct buf *bp) 965 { 966 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep); 967 968 /* 969 * This shouldn't happen under normal operation. 970 */ 971 if (io->type == HAMMER_STRUCTURE_VOLUME || 972 io->type == HAMMER_STRUCTURE_META_BUFFER) { 973 if (!panicstr) 974 panic("hammer_io_checkwrite: illegal buffer"); 975 if ((bp->b_flags & B_LOCKED) == 0) { 976 bp->b_flags |= B_LOCKED; 977 ++hammer_count_io_locked; 978 } 979 return(1); 980 } 981 982 /* 983 * We can only clear the modified bit if the IO is not currently 984 * undergoing modification. Otherwise we may miss changes. 985 * 986 * Only data and undo buffers can reach here. These buffers do 987 * not have terminal crc functions but we temporarily reference 988 * the IO anyway, just in case. 989 */ 990 if (io->modify_refs == 0 && io->modified) { 991 hammer_ref(&io->lock); 992 hammer_io_clear_modify(io, 0); 993 hammer_unref(&io->lock); 994 } else if (io->modified) { 995 KKASSERT(io->type == HAMMER_STRUCTURE_DATA_BUFFER); 996 } 997 998 /* 999 * The kernel is going to start the IO, set io->running. 1000 */ 1001 KKASSERT(io->running == 0); 1002 io->running = 1; 1003 io->hmp->io_running_space += io->bytes; 1004 hammer_count_io_running_write += io->bytes; 1005 return(0); 1006 } 1007 1008 /* 1009 * Return non-zero if we wish to delay the kernel's attempt to flush 1010 * this buffer to disk. 1011 */ 1012 static int 1013 hammer_io_countdeps(struct buf *bp, int n) 1014 { 1015 return(0); 1016 } 1017 1018 struct bio_ops hammer_bioops = { 1019 .io_start = hammer_io_start, 1020 .io_complete = hammer_io_complete, 1021 .io_deallocate = hammer_io_deallocate, 1022 .io_fsync = hammer_io_fsync, 1023 .io_sync = hammer_io_sync, 1024 .io_movedeps = hammer_io_movedeps, 1025 .io_countdeps = hammer_io_countdeps, 1026 .io_checkread = hammer_io_checkread, 1027 .io_checkwrite = hammer_io_checkwrite, 1028 }; 1029 1030 /************************************************************************ 1031 * DIRECT IO OPS * 1032 ************************************************************************ 1033 * 1034 * These functions operate directly on the buffer cache buffer associated 1035 * with a front-end vnode rather then a back-end device vnode. 1036 */ 1037 1038 /* 1039 * Read a buffer associated with a front-end vnode directly from the 1040 * disk media. The bio may be issued asynchronously. If leaf is non-NULL 1041 * we validate the CRC. 1042 * 1043 * We must check for the presence of a HAMMER buffer to handle the case 1044 * where the reblocker has rewritten the data (which it does via the HAMMER 1045 * buffer system, not via the high-level vnode buffer cache), but not yet 1046 * committed the buffer to the media. 1047 */ 1048 int 1049 hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio, 1050 hammer_btree_leaf_elm_t leaf) 1051 { 1052 hammer_off_t buf_offset; 1053 hammer_off_t zone2_offset; 1054 hammer_volume_t volume; 1055 struct buf *bp; 1056 struct bio *nbio; 1057 int vol_no; 1058 int error; 1059 1060 buf_offset = bio->bio_offset; 1061 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) == 1062 HAMMER_ZONE_LARGE_DATA); 1063 1064 /* 1065 * The buffer cache may have an aliased buffer (the reblocker can 1066 * write them). If it does we have to sync any dirty data before 1067 * we can build our direct-read. This is a non-critical code path. 1068 */ 1069 bp = bio->bio_buf; 1070 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize); 1071 1072 /* 1073 * Resolve to a zone-2 offset. The conversion just requires 1074 * munging the top 4 bits but we want to abstract it anyway 1075 * so the blockmap code can verify the zone assignment. 1076 */ 1077 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1078 if (error) 1079 goto done; 1080 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) == 1081 HAMMER_ZONE_RAW_BUFFER); 1082 1083 /* 1084 * Resolve volume and raw-offset for 3rd level bio. The 1085 * offset will be specific to the volume. 1086 */ 1087 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1088 volume = hammer_get_volume(hmp, vol_no, &error); 1089 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1090 error = EIO; 1091 1092 if (error == 0) { 1093 /* 1094 * 3rd level bio 1095 */ 1096 nbio = push_bio(bio); 1097 nbio->bio_offset = volume->ondisk->vol_buf_beg + 1098 (zone2_offset & HAMMER_OFF_SHORT_MASK); 1099 #if 0 1100 /* 1101 * XXX disabled - our CRC check doesn't work if the OS 1102 * does bogus_page replacement on the direct-read. 1103 */ 1104 if (leaf && hammer_verify_data) { 1105 nbio->bio_done = hammer_io_direct_read_complete; 1106 nbio->bio_caller_info1.uvalue32 = leaf->data_crc; 1107 } 1108 #endif 1109 hammer_stats_disk_read += bp->b_bufsize; 1110 vn_strategy(volume->devvp, nbio); 1111 } 1112 hammer_rel_volume(volume, 0); 1113 done: 1114 if (error) { 1115 kprintf("hammer_direct_read: failed @ %016llx\n", 1116 zone2_offset); 1117 bp->b_error = error; 1118 bp->b_flags |= B_ERROR; 1119 biodone(bio); 1120 } 1121 return(error); 1122 } 1123 1124 #if 0 1125 /* 1126 * On completion of the BIO this callback must check the data CRC 1127 * and chain to the previous bio. 1128 */ 1129 static 1130 void 1131 hammer_io_direct_read_complete(struct bio *nbio) 1132 { 1133 struct bio *obio; 1134 struct buf *bp; 1135 u_int32_t rec_crc = nbio->bio_caller_info1.uvalue32; 1136 1137 bp = nbio->bio_buf; 1138 if (crc32(bp->b_data, bp->b_bufsize) != rec_crc) { 1139 kprintf("HAMMER: data_crc error @%016llx/%d\n", 1140 nbio->bio_offset, bp->b_bufsize); 1141 if (hammer_debug_debug) 1142 Debugger(""); 1143 bp->b_flags |= B_ERROR; 1144 bp->b_error = EIO; 1145 } 1146 obio = pop_bio(nbio); 1147 biodone(obio); 1148 } 1149 #endif 1150 1151 /* 1152 * Write a buffer associated with a front-end vnode directly to the 1153 * disk media. The bio may be issued asynchronously. 1154 * 1155 * The BIO is associated with the specified record and RECF_DIRECT_IO 1156 * is set. The recorded is added to its object. 1157 */ 1158 int 1159 hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record, 1160 struct bio *bio) 1161 { 1162 hammer_btree_leaf_elm_t leaf = &record->leaf; 1163 hammer_off_t buf_offset; 1164 hammer_off_t zone2_offset; 1165 hammer_volume_t volume; 1166 hammer_buffer_t buffer; 1167 struct buf *bp; 1168 struct bio *nbio; 1169 char *ptr; 1170 int vol_no; 1171 int error; 1172 1173 buf_offset = leaf->data_offset; 1174 1175 KKASSERT(buf_offset > HAMMER_ZONE_BTREE); 1176 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE); 1177 1178 if ((buf_offset & HAMMER_BUFMASK) == 0 && 1179 leaf->data_len >= HAMMER_BUFSIZE) { 1180 /* 1181 * We are using the vnode's bio to write directly to the 1182 * media, any hammer_buffer at the same zone-X offset will 1183 * now have stale data. 1184 */ 1185 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1186 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1187 volume = hammer_get_volume(hmp, vol_no, &error); 1188 1189 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1190 error = EIO; 1191 if (error == 0) { 1192 bp = bio->bio_buf; 1193 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0); 1194 /* 1195 hammer_del_buffers(hmp, buf_offset, 1196 zone2_offset, bp->b_bufsize); 1197 */ 1198 1199 /* 1200 * Second level bio - cached zone2 offset. 1201 * 1202 * (We can put our bio_done function in either the 1203 * 2nd or 3rd level). 1204 */ 1205 nbio = push_bio(bio); 1206 nbio->bio_offset = zone2_offset; 1207 nbio->bio_done = hammer_io_direct_write_complete; 1208 nbio->bio_caller_info1.ptr = record; 1209 record->zone2_offset = zone2_offset; 1210 record->flags |= HAMMER_RECF_DIRECT_IO | 1211 HAMMER_RECF_DIRECT_INVAL; 1212 1213 /* 1214 * Third level bio - raw offset specific to the 1215 * correct volume. 1216 */ 1217 zone2_offset &= HAMMER_OFF_SHORT_MASK; 1218 nbio = push_bio(nbio); 1219 nbio->bio_offset = volume->ondisk->vol_buf_beg + 1220 zone2_offset; 1221 hammer_stats_disk_write += bp->b_bufsize; 1222 vn_strategy(volume->devvp, nbio); 1223 hammer_io_flush_mark(volume); 1224 } 1225 hammer_rel_volume(volume, 0); 1226 } else { 1227 /* 1228 * Must fit in a standard HAMMER buffer. In this case all 1229 * consumers use the HAMMER buffer system and RECF_DIRECT_IO 1230 * does not need to be set-up. 1231 */ 1232 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0); 1233 buffer = NULL; 1234 ptr = hammer_bread(hmp, buf_offset, &error, &buffer); 1235 if (error == 0) { 1236 bp = bio->bio_buf; 1237 bp->b_flags |= B_AGE; 1238 hammer_io_modify(&buffer->io, 1); 1239 bcopy(bp->b_data, ptr, leaf->data_len); 1240 hammer_io_modify_done(&buffer->io); 1241 hammer_rel_buffer(buffer, 0); 1242 bp->b_resid = 0; 1243 biodone(bio); 1244 } 1245 } 1246 if (error == 0) { 1247 /* 1248 * The record is all setup now, add it. Potential conflics 1249 * have already been dealt with. 1250 */ 1251 error = hammer_mem_add(record); 1252 KKASSERT(error == 0); 1253 } else { 1254 /* 1255 * Major suckage occured. 1256 */ 1257 kprintf("hammer_direct_write: failed @ %016llx\n", 1258 leaf->data_offset); 1259 bp = bio->bio_buf; 1260 bp->b_resid = 0; 1261 bp->b_error = EIO; 1262 bp->b_flags |= B_ERROR; 1263 biodone(bio); 1264 record->flags |= HAMMER_RECF_DELETED_FE; 1265 hammer_rel_mem_record(record); 1266 } 1267 return(error); 1268 } 1269 1270 /* 1271 * On completion of the BIO this callback must disconnect 1272 * it from the hammer_record and chain to the previous bio. 1273 * 1274 * An I/O error forces the mount to read-only. Data buffers 1275 * are not B_LOCKED like meta-data buffers are, so we have to 1276 * throw the buffer away to prevent the kernel from retrying. 1277 */ 1278 static 1279 void 1280 hammer_io_direct_write_complete(struct bio *nbio) 1281 { 1282 struct bio *obio; 1283 struct buf *bp; 1284 hammer_record_t record = nbio->bio_caller_info1.ptr; 1285 1286 bp = nbio->bio_buf; 1287 obio = pop_bio(nbio); 1288 if (bp->b_flags & B_ERROR) { 1289 hammer_critical_error(record->ip->hmp, record->ip, 1290 bp->b_error, 1291 "while writing bulk data"); 1292 bp->b_flags |= B_INVAL; 1293 } 1294 biodone(obio); 1295 1296 KKASSERT(record != NULL); 1297 KKASSERT(record->flags & HAMMER_RECF_DIRECT_IO); 1298 record->flags &= ~HAMMER_RECF_DIRECT_IO; 1299 if (record->flags & HAMMER_RECF_DIRECT_WAIT) { 1300 record->flags &= ~HAMMER_RECF_DIRECT_WAIT; 1301 wakeup(&record->flags); 1302 } 1303 } 1304 1305 1306 /* 1307 * This is called before a record is either committed to the B-Tree 1308 * or destroyed, to resolve any associated direct-IO. 1309 * 1310 * (1) We must wait for any direct-IO related to the record to complete. 1311 * 1312 * (2) We must remove any buffer cache aliases for data accessed via 1313 * leaf->data_offset or zone2_offset so non-direct-IO consumers 1314 * (the mirroring and reblocking code) do not see stale data. 1315 */ 1316 void 1317 hammer_io_direct_wait(hammer_record_t record) 1318 { 1319 /* 1320 * Wait for I/O to complete 1321 */ 1322 if (record->flags & HAMMER_RECF_DIRECT_IO) { 1323 crit_enter(); 1324 while (record->flags & HAMMER_RECF_DIRECT_IO) { 1325 record->flags |= HAMMER_RECF_DIRECT_WAIT; 1326 tsleep(&record->flags, 0, "hmdiow", 0); 1327 } 1328 crit_exit(); 1329 } 1330 1331 /* 1332 * Invalidate any related buffer cache aliases associated with the 1333 * backing device. This is needed because the buffer cache buffer 1334 * for file data is associated with the file vnode, not the backing 1335 * device vnode. 1336 * 1337 * XXX I do not think this case can occur any more now that 1338 * reservations ensure that all such buffers are removed before 1339 * an area can be reused. 1340 */ 1341 if (record->flags & HAMMER_RECF_DIRECT_INVAL) { 1342 KKASSERT(record->leaf.data_offset); 1343 hammer_del_buffers(record->ip->hmp, record->leaf.data_offset, 1344 record->zone2_offset, record->leaf.data_len, 1345 1); 1346 record->flags &= ~HAMMER_RECF_DIRECT_INVAL; 1347 } 1348 } 1349 1350 /* 1351 * This is called to remove the second-level cached zone-2 offset from 1352 * frontend buffer cache buffers, now stale due to a data relocation. 1353 * These offsets are generated by cluster_read() via VOP_BMAP, or directly 1354 * by hammer_vop_strategy_read(). 1355 * 1356 * This is rather nasty because here we have something like the reblocker 1357 * scanning the raw B-Tree with no held references on anything, really, 1358 * other then a shared lock on the B-Tree node, and we have to access the 1359 * frontend's buffer cache to check for and clean out the association. 1360 * Specifically, if the reblocker is moving data on the disk, these cached 1361 * offsets will become invalid. 1362 * 1363 * Only data record types associated with the large-data zone are subject 1364 * to direct-io and need to be checked. 1365 * 1366 */ 1367 void 1368 hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf) 1369 { 1370 struct hammer_inode_info iinfo; 1371 int zone; 1372 1373 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA) 1374 return; 1375 zone = HAMMER_ZONE_DECODE(leaf->data_offset); 1376 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX) 1377 return; 1378 iinfo.obj_id = leaf->base.obj_id; 1379 iinfo.obj_asof = 0; /* unused */ 1380 iinfo.obj_localization = leaf->base.localization & 1381 HAMMER_LOCALIZE_PSEUDOFS_MASK; 1382 iinfo.u.leaf = leaf; 1383 hammer_scan_inode_snapshots(hmp, &iinfo, 1384 hammer_io_direct_uncache_callback, 1385 leaf); 1386 } 1387 1388 static int 1389 hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data) 1390 { 1391 hammer_inode_info_t iinfo = data; 1392 hammer_off_t data_offset; 1393 hammer_off_t file_offset; 1394 struct vnode *vp; 1395 struct buf *bp; 1396 int blksize; 1397 1398 if (ip->vp == NULL) 1399 return(0); 1400 data_offset = iinfo->u.leaf->data_offset; 1401 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len; 1402 blksize = iinfo->u.leaf->data_len; 1403 KKASSERT((blksize & HAMMER_BUFMASK) == 0); 1404 1405 hammer_ref(&ip->lock); 1406 if (hammer_get_vnode(ip, &vp) == 0) { 1407 if ((bp = findblk(ip->vp, file_offset)) != NULL && 1408 bp->b_bio2.bio_offset != NOOFFSET) { 1409 bp = getblk(ip->vp, file_offset, blksize, 0, 0); 1410 bp->b_bio2.bio_offset = NOOFFSET; 1411 brelse(bp); 1412 } 1413 vput(vp); 1414 } 1415 hammer_rel_inode(ip, 0); 1416 return(0); 1417 } 1418 1419 1420 /* 1421 * This function is called when writes may have occured on the volume, 1422 * indicating that the device may be holding cached writes. 1423 */ 1424 static void 1425 hammer_io_flush_mark(hammer_volume_t volume) 1426 { 1427 volume->vol_flags |= HAMMER_VOLF_NEEDFLUSH; 1428 } 1429 1430 /* 1431 * This function ensures that the device has flushed any cached writes out. 1432 */ 1433 void 1434 hammer_io_flush_sync(hammer_mount_t hmp) 1435 { 1436 hammer_volume_t volume; 1437 struct buf *bp_base = NULL; 1438 struct buf *bp; 1439 1440 RB_FOREACH(volume, hammer_vol_rb_tree, &hmp->rb_vols_root) { 1441 if (volume->vol_flags & HAMMER_VOLF_NEEDFLUSH) { 1442 volume->vol_flags &= ~HAMMER_VOLF_NEEDFLUSH; 1443 bp = getpbuf(NULL); 1444 bp->b_bio1.bio_offset = 0; 1445 bp->b_bufsize = 0; 1446 bp->b_bcount = 0; 1447 bp->b_cmd = BUF_CMD_FLUSH; 1448 bp->b_bio1.bio_caller_info1.cluster_head = bp_base; 1449 bp->b_bio1.bio_done = hammer_io_flush_sync_done; 1450 bp->b_flags |= B_ASYNC; 1451 bp_base = bp; 1452 vn_strategy(volume->devvp, &bp->b_bio1); 1453 } 1454 } 1455 while ((bp = bp_base) != NULL) { 1456 bp_base = bp->b_bio1.bio_caller_info1.cluster_head; 1457 while (bp->b_cmd != BUF_CMD_DONE) { 1458 crit_enter(); 1459 tsleep_interlock(&bp->b_cmd); 1460 if (bp->b_cmd != BUF_CMD_DONE) 1461 tsleep(&bp->b_cmd, 0, "hmrFLS", 0); 1462 crit_exit(); 1463 } 1464 bp->b_flags &= ~B_ASYNC; 1465 relpbuf(bp, NULL); 1466 } 1467 } 1468 1469 /* 1470 * Callback to deal with completed flush commands to the device. 1471 */ 1472 static void 1473 hammer_io_flush_sync_done(struct bio *bio) 1474 { 1475 struct buf *bp; 1476 1477 bp = bio->bio_buf; 1478 bp->b_cmd = BUF_CMD_DONE; 1479 wakeup(&bp->b_cmd); 1480 } 1481 1482