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 65 66 /* 67 * Initialize a new, already-zero'd hammer_io structure, or reinitialize 68 * an existing hammer_io structure which may have switched to another type. 69 */ 70 void 71 hammer_io_init(hammer_io_t io, hammer_volume_t volume, enum hammer_io_type type) 72 { 73 io->volume = volume; 74 io->hmp = volume->io.hmp; 75 io->type = type; 76 } 77 78 /* 79 * Helper routine to disassociate a buffer cache buffer from an I/O 80 * structure. The buffer is unlocked and marked appropriate for reclamation. 81 * 82 * The io may have 0 or 1 references depending on who called us. The 83 * caller is responsible for dealing with the refs. 84 * 85 * This call can only be made when no action is required on the buffer. 86 * 87 * The caller must own the buffer and the IO must indicate that the 88 * structure no longer owns it (io.released != 0). 89 */ 90 static void 91 hammer_io_disassociate(hammer_io_structure_t iou) 92 { 93 struct buf *bp = iou->io.bp; 94 95 KKASSERT(iou->io.released); 96 KKASSERT(iou->io.modified == 0); 97 KKASSERT(LIST_FIRST(&bp->b_dep) == (void *)iou); 98 buf_dep_init(bp); 99 iou->io.bp = NULL; 100 101 /* 102 * If the buffer was locked someone wanted to get rid of it. 103 */ 104 if (bp->b_flags & B_LOCKED) { 105 --hammer_count_io_locked; 106 bp->b_flags &= ~B_LOCKED; 107 } 108 if (iou->io.reclaim) { 109 bp->b_flags |= B_NOCACHE|B_RELBUF; 110 iou->io.reclaim = 0; 111 } 112 113 switch(iou->io.type) { 114 case HAMMER_STRUCTURE_VOLUME: 115 iou->volume.ondisk = NULL; 116 break; 117 case HAMMER_STRUCTURE_DATA_BUFFER: 118 case HAMMER_STRUCTURE_META_BUFFER: 119 case HAMMER_STRUCTURE_UNDO_BUFFER: 120 iou->buffer.ondisk = NULL; 121 break; 122 } 123 } 124 125 /* 126 * Wait for any physical IO to complete 127 * 128 * XXX we aren't interlocked against a spinlock or anything so there 129 * is a small window in the interlock / io->running == 0 test. 130 */ 131 void 132 hammer_io_wait(hammer_io_t io) 133 { 134 if (io->running) { 135 for (;;) { 136 io->waiting = 1; 137 tsleep_interlock(io, 0); 138 if (io->running == 0) 139 break; 140 tsleep(io, PINTERLOCKED, "hmrflw", hz); 141 if (io->running == 0) 142 break; 143 } 144 } 145 } 146 147 /* 148 * Wait for all hammer_io-initated write I/O's to complete. This is not 149 * supposed to count direct I/O's but some can leak through (for 150 * non-full-sized direct I/Os). 151 */ 152 void 153 hammer_io_wait_all(hammer_mount_t hmp, const char *ident) 154 { 155 hammer_io_flush_sync(hmp); 156 crit_enter(); 157 while (hmp->io_running_space) 158 tsleep(&hmp->io_running_space, 0, ident, 0); 159 crit_exit(); 160 } 161 162 /* 163 * Clear a flagged error condition on a I/O buffer. The caller must hold 164 * its own ref on the buffer. 165 */ 166 void 167 hammer_io_clear_error(struct hammer_io *io) 168 { 169 if (io->ioerror) { 170 io->ioerror = 0; 171 hammer_unref(&io->lock); 172 KKASSERT(io->lock.refs > 0); 173 } 174 } 175 176 177 #define HAMMER_MAXRA 4 178 179 /* 180 * Load bp for a HAMMER structure. The io must be exclusively locked by 181 * the caller. 182 * 183 * This routine is mostly used on meta-data and small-data blocks. Generally 184 * speaking HAMMER assumes some locality of reference and will cluster 185 * a 64K read. 186 * 187 * Note that clustering occurs at the device layer, not the logical layer. 188 * If the buffers do not apply to the current operation they may apply to 189 * some other. 190 */ 191 int 192 hammer_io_read(struct vnode *devvp, struct hammer_io *io, hammer_off_t limit) 193 { 194 struct buf *bp; 195 int error; 196 197 if ((bp = io->bp) == NULL) { 198 hammer_count_io_running_read += io->bytes; 199 if (hammer_cluster_enable) { 200 error = cluster_read(devvp, limit, 201 io->offset, io->bytes, 202 HAMMER_CLUSTER_SIZE, 203 HAMMER_CLUSTER_BUFS, &io->bp); 204 } else { 205 error = bread(devvp, io->offset, io->bytes, &io->bp); 206 } 207 hammer_stats_disk_read += io->bytes; 208 hammer_count_io_running_read -= io->bytes; 209 210 /* 211 * The code generally assumes b_ops/b_dep has been set-up, 212 * even if we error out here. 213 */ 214 bp = io->bp; 215 bp->b_ops = &hammer_bioops; 216 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 217 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node); 218 BUF_KERNPROC(bp); 219 KKASSERT(io->modified == 0); 220 KKASSERT(io->running == 0); 221 KKASSERT(io->waiting == 0); 222 io->released = 0; /* we hold an active lock on bp */ 223 } else { 224 error = 0; 225 } 226 return(error); 227 } 228 229 /* 230 * Similar to hammer_io_read() but returns a zero'd out buffer instead. 231 * Must be called with the IO exclusively locked. 232 * 233 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background 234 * I/O by forcing the buffer to not be in a released state before calling 235 * it. 236 * 237 * This function will also mark the IO as modified but it will not 238 * increment the modify_refs count. 239 */ 240 int 241 hammer_io_new(struct vnode *devvp, struct hammer_io *io) 242 { 243 struct buf *bp; 244 245 if ((bp = io->bp) == NULL) { 246 io->bp = getblk(devvp, io->offset, io->bytes, 0, 0); 247 bp = io->bp; 248 bp->b_ops = &hammer_bioops; 249 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 250 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node); 251 io->released = 0; 252 KKASSERT(io->running == 0); 253 io->waiting = 0; 254 BUF_KERNPROC(bp); 255 } else { 256 if (io->released) { 257 regetblk(bp); 258 BUF_KERNPROC(bp); 259 io->released = 0; 260 } 261 } 262 hammer_io_modify(io, 0); 263 vfs_bio_clrbuf(bp); 264 return(0); 265 } 266 267 /* 268 * Advance the activity count on the underlying buffer because 269 * HAMMER does not getblk/brelse on every access. 270 */ 271 void 272 hammer_io_advance(struct hammer_io *io) 273 { 274 if (io->bp) 275 buf_act_advance(io->bp); 276 } 277 278 /* 279 * Remove potential device level aliases against buffers managed by high level 280 * vnodes. Aliases can also be created due to mixed buffer sizes or via 281 * direct access to the backing store device. 282 * 283 * This is nasty because the buffers are also VMIO-backed. Even if a buffer 284 * does not exist its backing VM pages might, and we have to invalidate 285 * those as well or a getblk() will reinstate them. 286 * 287 * Buffer cache buffers associated with hammer_buffers cannot be 288 * invalidated. 289 */ 290 int 291 hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset) 292 { 293 hammer_io_structure_t iou; 294 hammer_off_t phys_offset; 295 struct buf *bp; 296 int error; 297 298 phys_offset = volume->ondisk->vol_buf_beg + 299 (zone2_offset & HAMMER_OFF_SHORT_MASK); 300 crit_enter(); 301 if ((bp = findblk(volume->devvp, phys_offset, FINDBLK_TEST)) != NULL) 302 bp = getblk(volume->devvp, phys_offset, bp->b_bufsize, 0, 0); 303 else 304 bp = getblk(volume->devvp, phys_offset, HAMMER_BUFSIZE, 0, 0); 305 if ((iou = (void *)LIST_FIRST(&bp->b_dep)) != NULL) { 306 #if 0 307 hammer_ref(&iou->io.lock); 308 hammer_io_clear_modify(&iou->io, 1); 309 bundirty(bp); 310 iou->io.released = 0; 311 BUF_KERNPROC(bp); 312 iou->io.reclaim = 1; 313 iou->io.waitdep = 1; 314 KKASSERT(iou->io.lock.refs == 1); 315 hammer_rel_buffer(&iou->buffer, 0); 316 /*hammer_io_deallocate(bp);*/ 317 #endif 318 bqrelse(bp); 319 error = EAGAIN; 320 } else { 321 KKASSERT((bp->b_flags & B_LOCKED) == 0); 322 bundirty(bp); 323 bp->b_flags |= B_NOCACHE|B_RELBUF; 324 brelse(bp); 325 error = 0; 326 } 327 crit_exit(); 328 return(error); 329 } 330 331 /* 332 * This routine is called on the last reference to a hammer structure. 333 * The io is usually interlocked with io.loading and io.refs must be 1. 334 * 335 * This routine may return a non-NULL bp to the caller for dispoal. Disposal 336 * simply means the caller finishes decrementing the ref-count on the 337 * IO structure then brelse()'s the bp. The bp may or may not still be 338 * passively associated with the IO. 339 * 340 * The only requirement here is that modified meta-data and volume-header 341 * buffer may NOT be disassociated from the IO structure, and consequently 342 * we also leave such buffers actively associated with the IO if they already 343 * are (since the kernel can't do anything with them anyway). Only the 344 * flusher is allowed to write such buffers out. Modified pure-data and 345 * undo buffers are returned to the kernel but left passively associated 346 * so we can track when the kernel writes the bp out. 347 */ 348 struct buf * 349 hammer_io_release(struct hammer_io *io, int flush) 350 { 351 union hammer_io_structure *iou = (void *)io; 352 struct buf *bp; 353 354 if ((bp = io->bp) == NULL) 355 return(NULL); 356 357 /* 358 * Try to flush a dirty IO to disk if asked to by the 359 * caller or if the kernel tried to flush the buffer in the past. 360 * 361 * Kernel-initiated flushes are only allowed for pure-data buffers. 362 * meta-data and volume buffers can only be flushed explicitly 363 * by HAMMER. 364 */ 365 if (io->modified) { 366 if (flush) { 367 hammer_io_flush(io, 0); 368 } else if (bp->b_flags & B_LOCKED) { 369 switch(io->type) { 370 case HAMMER_STRUCTURE_DATA_BUFFER: 371 hammer_io_flush(io, 0); 372 break; 373 case HAMMER_STRUCTURE_UNDO_BUFFER: 374 hammer_io_flush(io, hammer_undo_reclaim(io)); 375 break; 376 default: 377 break; 378 } 379 } /* else no explicit request to flush the buffer */ 380 } 381 382 /* 383 * Wait for the IO to complete if asked to. This occurs when 384 * the buffer must be disposed of definitively during an umount 385 * or buffer invalidation. 386 */ 387 if (io->waitdep && io->running) { 388 hammer_io_wait(io); 389 } 390 391 /* 392 * Return control of the buffer to the kernel (with the provisio 393 * that our bioops can override kernel decisions with regards to 394 * the buffer). 395 */ 396 if ((flush || io->reclaim) && io->modified == 0 && io->running == 0) { 397 /* 398 * Always disassociate the bp if an explicit flush 399 * was requested and the IO completed with no error 400 * (so unmount can really clean up the structure). 401 */ 402 if (io->released) { 403 regetblk(bp); 404 BUF_KERNPROC(bp); 405 } else { 406 io->released = 1; 407 } 408 hammer_io_disassociate((hammer_io_structure_t)io); 409 /* return the bp */ 410 } else if (io->modified) { 411 /* 412 * Only certain IO types can be released to the kernel if 413 * the buffer has been modified. 414 * 415 * volume and meta-data IO types may only be explicitly 416 * flushed by HAMMER. 417 */ 418 switch(io->type) { 419 case HAMMER_STRUCTURE_DATA_BUFFER: 420 case HAMMER_STRUCTURE_UNDO_BUFFER: 421 if (io->released == 0) { 422 io->released = 1; 423 bdwrite(bp); 424 } 425 break; 426 default: 427 break; 428 } 429 bp = NULL; /* bp left associated */ 430 } else if (io->released == 0) { 431 /* 432 * Clean buffers can be generally released to the kernel. 433 * We leave the bp passively associated with the HAMMER 434 * structure and use bioops to disconnect it later on 435 * if the kernel wants to discard the buffer. 436 * 437 * We can steal the structure's ownership of the bp. 438 */ 439 io->released = 1; 440 if (bp->b_flags & B_LOCKED) { 441 hammer_io_disassociate(iou); 442 /* return the bp */ 443 } else { 444 if (io->reclaim) { 445 hammer_io_disassociate(iou); 446 /* return the bp */ 447 } else { 448 /* return the bp (bp passively associated) */ 449 } 450 } 451 } else { 452 /* 453 * A released buffer is passively associate with our 454 * hammer_io structure. The kernel cannot destroy it 455 * without making a bioops call. If the kernel (B_LOCKED) 456 * or we (reclaim) requested that the buffer be destroyed 457 * we destroy it, otherwise we do a quick get/release to 458 * reset its position in the kernel's LRU list. 459 * 460 * Leaving the buffer passively associated allows us to 461 * use the kernel's LRU buffer flushing mechanisms rather 462 * then rolling our own. 463 * 464 * XXX there are two ways of doing this. We can re-acquire 465 * and passively release to reset the LRU, or not. 466 */ 467 if (io->running == 0) { 468 regetblk(bp); 469 if ((bp->b_flags & B_LOCKED) || io->reclaim) { 470 hammer_io_disassociate(iou); 471 /* return the bp */ 472 } else { 473 /* return the bp (bp passively associated) */ 474 } 475 } else { 476 /* 477 * bp is left passively associated but we do not 478 * try to reacquire it. Interactions with the io 479 * structure will occur on completion of the bp's 480 * I/O. 481 */ 482 bp = NULL; 483 } 484 } 485 return(bp); 486 } 487 488 /* 489 * This routine is called with a locked IO when a flush is desired and 490 * no other references to the structure exists other then ours. This 491 * routine is ONLY called when HAMMER believes it is safe to flush a 492 * potentially modified buffer out. 493 */ 494 void 495 hammer_io_flush(struct hammer_io *io, int reclaim) 496 { 497 struct buf *bp; 498 499 /* 500 * Degenerate case - nothing to flush if nothing is dirty. 501 */ 502 if (io->modified == 0) { 503 return; 504 } 505 506 KKASSERT(io->bp); 507 KKASSERT(io->modify_refs <= 0); 508 509 /* 510 * Acquire ownership of the bp, particularly before we clear our 511 * modified flag. 512 * 513 * We are going to bawrite() this bp. Don't leave a window where 514 * io->released is set, we actually own the bp rather then our 515 * buffer. 516 */ 517 bp = io->bp; 518 if (io->released) { 519 regetblk(bp); 520 /* BUF_KERNPROC(io->bp); */ 521 /* io->released = 0; */ 522 KKASSERT(io->released); 523 KKASSERT(io->bp == bp); 524 } 525 io->released = 1; 526 527 if (reclaim) { 528 io->reclaim = 1; 529 if ((bp->b_flags & B_LOCKED) == 0) { 530 bp->b_flags |= B_LOCKED; 531 ++hammer_count_io_locked; 532 } 533 } 534 535 /* 536 * Acquire exclusive access to the bp and then clear the modified 537 * state of the buffer prior to issuing I/O to interlock any 538 * modifications made while the I/O is in progress. This shouldn't 539 * happen anyway but losing data would be worse. The modified bit 540 * will be rechecked after the IO completes. 541 * 542 * NOTE: This call also finalizes the buffer's content (inval == 0). 543 * 544 * This is only legal when lock.refs == 1 (otherwise we might clear 545 * the modified bit while there are still users of the cluster 546 * modifying the data). 547 * 548 * Do this before potentially blocking so any attempt to modify the 549 * ondisk while we are blocked blocks waiting for us. 550 */ 551 hammer_ref(&io->lock); 552 hammer_io_clear_modify(io, 0); 553 hammer_unref(&io->lock); 554 555 /* 556 * Transfer ownership to the kernel and initiate I/O. 557 */ 558 io->running = 1; 559 io->hmp->io_running_space += io->bytes; 560 hammer_count_io_running_write += io->bytes; 561 bawrite(bp); 562 hammer_io_flush_mark(io->volume); 563 } 564 565 /************************************************************************ 566 * BUFFER DIRTYING * 567 ************************************************************************ 568 * 569 * These routines deal with dependancies created when IO buffers get 570 * modified. The caller must call hammer_modify_*() on a referenced 571 * HAMMER structure prior to modifying its on-disk data. 572 * 573 * Any intent to modify an IO buffer acquires the related bp and imposes 574 * various write ordering dependancies. 575 */ 576 577 /* 578 * Mark a HAMMER structure as undergoing modification. Meta-data buffers 579 * are locked until the flusher can deal with them, pure data buffers 580 * can be written out. 581 */ 582 static 583 void 584 hammer_io_modify(hammer_io_t io, int count) 585 { 586 /* 587 * io->modify_refs must be >= 0 588 */ 589 while (io->modify_refs < 0) { 590 io->waitmod = 1; 591 tsleep(io, 0, "hmrmod", 0); 592 } 593 594 /* 595 * Shortcut if nothing to do. 596 */ 597 KKASSERT(io->lock.refs != 0 && io->bp != NULL); 598 io->modify_refs += count; 599 if (io->modified && io->released == 0) 600 return; 601 602 hammer_lock_ex(&io->lock); 603 if (io->modified == 0) { 604 hammer_io_set_modlist(io); 605 io->modified = 1; 606 } 607 if (io->released) { 608 regetblk(io->bp); 609 BUF_KERNPROC(io->bp); 610 io->released = 0; 611 KKASSERT(io->modified != 0); 612 } 613 hammer_unlock(&io->lock); 614 } 615 616 static __inline 617 void 618 hammer_io_modify_done(hammer_io_t io) 619 { 620 KKASSERT(io->modify_refs > 0); 621 --io->modify_refs; 622 if (io->modify_refs == 0 && io->waitmod) { 623 io->waitmod = 0; 624 wakeup(io); 625 } 626 } 627 628 void 629 hammer_io_write_interlock(hammer_io_t io) 630 { 631 while (io->modify_refs != 0) { 632 io->waitmod = 1; 633 tsleep(io, 0, "hmrmod", 0); 634 } 635 io->modify_refs = -1; 636 } 637 638 void 639 hammer_io_done_interlock(hammer_io_t io) 640 { 641 KKASSERT(io->modify_refs == -1); 642 io->modify_refs = 0; 643 if (io->waitmod) { 644 io->waitmod = 0; 645 wakeup(io); 646 } 647 } 648 649 /* 650 * Caller intends to modify a volume's ondisk structure. 651 * 652 * This is only allowed if we are the flusher or we have a ref on the 653 * sync_lock. 654 */ 655 void 656 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume, 657 void *base, int len) 658 { 659 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 660 661 hammer_io_modify(&volume->io, 1); 662 if (len) { 663 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk; 664 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 665 hammer_generate_undo(trans, 666 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset), 667 base, len); 668 } 669 } 670 671 /* 672 * Caller intends to modify a buffer's ondisk structure. 673 * 674 * This is only allowed if we are the flusher or we have a ref on the 675 * sync_lock. 676 */ 677 void 678 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer, 679 void *base, int len) 680 { 681 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 682 683 hammer_io_modify(&buffer->io, 1); 684 if (len) { 685 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk; 686 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 687 hammer_generate_undo(trans, 688 buffer->zone2_offset + rel_offset, 689 base, len); 690 } 691 } 692 693 void 694 hammer_modify_volume_done(hammer_volume_t volume) 695 { 696 hammer_io_modify_done(&volume->io); 697 } 698 699 void 700 hammer_modify_buffer_done(hammer_buffer_t buffer) 701 { 702 hammer_io_modify_done(&buffer->io); 703 } 704 705 /* 706 * Mark an entity as not being dirty any more and finalize any 707 * delayed adjustments to the buffer. 708 * 709 * Delayed adjustments are an important performance enhancement, allowing 710 * us to avoid recalculating B-Tree node CRCs over and over again when 711 * making bulk-modifications to the B-Tree. 712 * 713 * If inval is non-zero delayed adjustments are ignored. 714 * 715 * This routine may dereference related btree nodes and cause the 716 * buffer to be dereferenced. The caller must own a reference on io. 717 */ 718 void 719 hammer_io_clear_modify(struct hammer_io *io, int inval) 720 { 721 if (io->modified == 0) 722 return; 723 724 /* 725 * Take us off the mod-list and clear the modified bit. 726 */ 727 KKASSERT(io->mod_list != NULL); 728 if (io->mod_list == &io->hmp->volu_list || 729 io->mod_list == &io->hmp->meta_list) { 730 io->hmp->locked_dirty_space -= io->bytes; 731 hammer_count_dirtybufspace -= io->bytes; 732 } 733 TAILQ_REMOVE(io->mod_list, io, mod_entry); 734 io->mod_list = NULL; 735 io->modified = 0; 736 737 /* 738 * If this bit is not set there are no delayed adjustments. 739 */ 740 if (io->gencrc == 0) 741 return; 742 io->gencrc = 0; 743 744 /* 745 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference 746 * on the node (& underlying buffer). Release the node after clearing 747 * the flag. 748 */ 749 if (io->type == HAMMER_STRUCTURE_META_BUFFER) { 750 hammer_buffer_t buffer = (void *)io; 751 hammer_node_t node; 752 753 restart: 754 TAILQ_FOREACH(node, &buffer->clist, entry) { 755 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) 756 continue; 757 node->flags &= ~HAMMER_NODE_NEEDSCRC; 758 KKASSERT(node->ondisk); 759 if (inval == 0) 760 node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE); 761 hammer_rel_node(node); 762 goto restart; 763 } 764 } 765 /* caller must still have ref on io */ 766 KKASSERT(io->lock.refs > 0); 767 } 768 769 /* 770 * Clear the IO's modify list. Even though the IO is no longer modified 771 * it may still be on the lose_list. This routine is called just before 772 * the governing hammer_buffer is destroyed. 773 */ 774 void 775 hammer_io_clear_modlist(struct hammer_io *io) 776 { 777 KKASSERT(io->modified == 0); 778 if (io->mod_list) { 779 crit_enter(); /* biodone race against list */ 780 KKASSERT(io->mod_list == &io->hmp->lose_list); 781 TAILQ_REMOVE(io->mod_list, io, mod_entry); 782 io->mod_list = NULL; 783 crit_exit(); 784 } 785 } 786 787 static void 788 hammer_io_set_modlist(struct hammer_io *io) 789 { 790 struct hammer_mount *hmp = io->hmp; 791 792 KKASSERT(io->mod_list == NULL); 793 794 switch(io->type) { 795 case HAMMER_STRUCTURE_VOLUME: 796 io->mod_list = &hmp->volu_list; 797 hmp->locked_dirty_space += io->bytes; 798 hammer_count_dirtybufspace += io->bytes; 799 break; 800 case HAMMER_STRUCTURE_META_BUFFER: 801 io->mod_list = &hmp->meta_list; 802 hmp->locked_dirty_space += io->bytes; 803 hammer_count_dirtybufspace += io->bytes; 804 break; 805 case HAMMER_STRUCTURE_UNDO_BUFFER: 806 io->mod_list = &hmp->undo_list; 807 break; 808 case HAMMER_STRUCTURE_DATA_BUFFER: 809 io->mod_list = &hmp->data_list; 810 break; 811 } 812 TAILQ_INSERT_TAIL(io->mod_list, io, mod_entry); 813 } 814 815 /************************************************************************ 816 * HAMMER_BIOOPS * 817 ************************************************************************ 818 * 819 */ 820 821 /* 822 * Pre-IO initiation kernel callback - cluster build only 823 */ 824 static void 825 hammer_io_start(struct buf *bp) 826 { 827 } 828 829 /* 830 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT! 831 * 832 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit 833 * may also be set if we were marking a cluster header open. Only remove 834 * our dependancy if the modified bit is clear. 835 */ 836 static void 837 hammer_io_complete(struct buf *bp) 838 { 839 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep); 840 841 KKASSERT(iou->io.released == 1); 842 843 /* 844 * Deal with people waiting for I/O to drain 845 */ 846 if (iou->io.running) { 847 /* 848 * Deal with critical write errors. Once a critical error 849 * has been flagged in hmp the UNDO FIFO will not be updated. 850 * That way crash recover will give us a consistent 851 * filesystem. 852 * 853 * Because of this we can throw away failed UNDO buffers. If 854 * we throw away META or DATA buffers we risk corrupting 855 * the now read-only version of the filesystem visible to 856 * the user. Clear B_ERROR so the buffer is not re-dirtied 857 * by the kernel and ref the io so it doesn't get thrown 858 * away. 859 */ 860 if (bp->b_flags & B_ERROR) { 861 hammer_critical_error(iou->io.hmp, NULL, bp->b_error, 862 "while flushing meta-data"); 863 switch(iou->io.type) { 864 case HAMMER_STRUCTURE_UNDO_BUFFER: 865 break; 866 default: 867 if (iou->io.ioerror == 0) { 868 iou->io.ioerror = 1; 869 if (iou->io.lock.refs == 0) 870 ++hammer_count_refedbufs; 871 hammer_ref(&iou->io.lock); 872 } 873 break; 874 } 875 bp->b_flags &= ~B_ERROR; 876 bundirty(bp); 877 #if 0 878 hammer_io_set_modlist(&iou->io); 879 iou->io.modified = 1; 880 #endif 881 } 882 hammer_stats_disk_write += iou->io.bytes; 883 hammer_count_io_running_write -= iou->io.bytes; 884 iou->io.hmp->io_running_space -= iou->io.bytes; 885 if (iou->io.hmp->io_running_space == 0) 886 wakeup(&iou->io.hmp->io_running_space); 887 KKASSERT(iou->io.hmp->io_running_space >= 0); 888 iou->io.running = 0; 889 } else { 890 hammer_stats_disk_read += iou->io.bytes; 891 } 892 893 if (iou->io.waiting) { 894 iou->io.waiting = 0; 895 wakeup(iou); 896 } 897 898 /* 899 * If B_LOCKED is set someone wanted to deallocate the bp at some 900 * point, do it now if refs has become zero. 901 */ 902 if ((bp->b_flags & B_LOCKED) && iou->io.lock.refs == 0) { 903 KKASSERT(iou->io.modified == 0); 904 --hammer_count_io_locked; 905 bp->b_flags &= ~B_LOCKED; 906 hammer_io_deallocate(bp); 907 /* structure may be dead now */ 908 } 909 } 910 911 /* 912 * Callback from kernel when it wishes to deallocate a passively 913 * associated structure. This mostly occurs with clean buffers 914 * but it may be possible for a holding structure to be marked dirty 915 * while its buffer is passively associated. The caller owns the bp. 916 * 917 * If we cannot disassociate we set B_LOCKED to prevent the buffer 918 * from getting reused. 919 * 920 * WARNING: Because this can be called directly by getnewbuf we cannot 921 * recurse into the tree. If a bp cannot be immediately disassociated 922 * our only recourse is to set B_LOCKED. 923 * 924 * WARNING: This may be called from an interrupt via hammer_io_complete() 925 */ 926 static void 927 hammer_io_deallocate(struct buf *bp) 928 { 929 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep); 930 931 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0); 932 if (iou->io.lock.refs > 0 || iou->io.modified) { 933 /* 934 * It is not legal to disassociate a modified buffer. This 935 * case really shouldn't ever occur. 936 */ 937 bp->b_flags |= B_LOCKED; 938 ++hammer_count_io_locked; 939 } else { 940 /* 941 * Disassociate the BP. If the io has no refs left we 942 * have to add it to the loose list. 943 */ 944 hammer_io_disassociate(iou); 945 if (iou->io.type != HAMMER_STRUCTURE_VOLUME) { 946 KKASSERT(iou->io.bp == NULL); 947 KKASSERT(iou->io.mod_list == NULL); 948 crit_enter(); /* biodone race against list */ 949 iou->io.mod_list = &iou->io.hmp->lose_list; 950 TAILQ_INSERT_TAIL(iou->io.mod_list, &iou->io, mod_entry); 951 crit_exit(); 952 } 953 } 954 } 955 956 static int 957 hammer_io_fsync(struct vnode *vp) 958 { 959 return(0); 960 } 961 962 /* 963 * NOTE: will not be called unless we tell the kernel about the 964 * bioops. Unused... we use the mount's VFS_SYNC instead. 965 */ 966 static int 967 hammer_io_sync(struct mount *mp) 968 { 969 return(0); 970 } 971 972 static void 973 hammer_io_movedeps(struct buf *bp1, struct buf *bp2) 974 { 975 } 976 977 /* 978 * I/O pre-check for reading and writing. HAMMER only uses this for 979 * B_CACHE buffers so checkread just shouldn't happen, but if it does 980 * allow it. 981 * 982 * Writing is a different case. We don't want the kernel to try to write 983 * out a buffer that HAMMER may be modifying passively or which has a 984 * dependancy. In addition, kernel-demanded writes can only proceed for 985 * certain types of buffers (i.e. UNDO and DATA types). Other dirty 986 * buffer types can only be explicitly written by the flusher. 987 * 988 * checkwrite will only be called for bdwrite()n buffers. If we return 989 * success the kernel is guaranteed to initiate the buffer write. 990 */ 991 static int 992 hammer_io_checkread(struct buf *bp) 993 { 994 return(0); 995 } 996 997 static int 998 hammer_io_checkwrite(struct buf *bp) 999 { 1000 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep); 1001 1002 /* 1003 * This shouldn't happen under normal operation. 1004 */ 1005 if (io->type == HAMMER_STRUCTURE_VOLUME || 1006 io->type == HAMMER_STRUCTURE_META_BUFFER) { 1007 if (!panicstr) 1008 panic("hammer_io_checkwrite: illegal buffer"); 1009 if ((bp->b_flags & B_LOCKED) == 0) { 1010 bp->b_flags |= B_LOCKED; 1011 ++hammer_count_io_locked; 1012 } 1013 return(1); 1014 } 1015 1016 /* 1017 * We can only clear the modified bit if the IO is not currently 1018 * undergoing modification. Otherwise we may miss changes. 1019 * 1020 * Only data and undo buffers can reach here. These buffers do 1021 * not have terminal crc functions but we temporarily reference 1022 * the IO anyway, just in case. 1023 */ 1024 if (io->modify_refs == 0 && io->modified) { 1025 hammer_ref(&io->lock); 1026 hammer_io_clear_modify(io, 0); 1027 hammer_unref(&io->lock); 1028 } else if (io->modified) { 1029 KKASSERT(io->type == HAMMER_STRUCTURE_DATA_BUFFER); 1030 } 1031 1032 /* 1033 * The kernel is going to start the IO, set io->running. 1034 */ 1035 KKASSERT(io->running == 0); 1036 io->running = 1; 1037 io->hmp->io_running_space += io->bytes; 1038 hammer_count_io_running_write += io->bytes; 1039 return(0); 1040 } 1041 1042 /* 1043 * Return non-zero if we wish to delay the kernel's attempt to flush 1044 * this buffer to disk. 1045 */ 1046 static int 1047 hammer_io_countdeps(struct buf *bp, int n) 1048 { 1049 return(0); 1050 } 1051 1052 struct bio_ops hammer_bioops = { 1053 .io_start = hammer_io_start, 1054 .io_complete = hammer_io_complete, 1055 .io_deallocate = hammer_io_deallocate, 1056 .io_fsync = hammer_io_fsync, 1057 .io_sync = hammer_io_sync, 1058 .io_movedeps = hammer_io_movedeps, 1059 .io_countdeps = hammer_io_countdeps, 1060 .io_checkread = hammer_io_checkread, 1061 .io_checkwrite = hammer_io_checkwrite, 1062 }; 1063 1064 /************************************************************************ 1065 * DIRECT IO OPS * 1066 ************************************************************************ 1067 * 1068 * These functions operate directly on the buffer cache buffer associated 1069 * with a front-end vnode rather then a back-end device vnode. 1070 */ 1071 1072 /* 1073 * Read a buffer associated with a front-end vnode directly from the 1074 * disk media. The bio may be issued asynchronously. If leaf is non-NULL 1075 * we validate the CRC. 1076 * 1077 * We must check for the presence of a HAMMER buffer to handle the case 1078 * where the reblocker has rewritten the data (which it does via the HAMMER 1079 * buffer system, not via the high-level vnode buffer cache), but not yet 1080 * committed the buffer to the media. 1081 */ 1082 int 1083 hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio, 1084 hammer_btree_leaf_elm_t leaf) 1085 { 1086 hammer_off_t buf_offset; 1087 hammer_off_t zone2_offset; 1088 hammer_volume_t volume; 1089 struct buf *bp; 1090 struct bio *nbio; 1091 int vol_no; 1092 int error; 1093 1094 buf_offset = bio->bio_offset; 1095 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) == 1096 HAMMER_ZONE_LARGE_DATA); 1097 1098 /* 1099 * The buffer cache may have an aliased buffer (the reblocker can 1100 * write them). If it does we have to sync any dirty data before 1101 * we can build our direct-read. This is a non-critical code path. 1102 */ 1103 bp = bio->bio_buf; 1104 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize); 1105 1106 /* 1107 * Resolve to a zone-2 offset. The conversion just requires 1108 * munging the top 4 bits but we want to abstract it anyway 1109 * so the blockmap code can verify the zone assignment. 1110 */ 1111 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1112 if (error) 1113 goto done; 1114 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) == 1115 HAMMER_ZONE_RAW_BUFFER); 1116 1117 /* 1118 * Resolve volume and raw-offset for 3rd level bio. The 1119 * offset will be specific to the volume. 1120 */ 1121 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1122 volume = hammer_get_volume(hmp, vol_no, &error); 1123 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1124 error = EIO; 1125 1126 if (error == 0) { 1127 /* 1128 * 3rd level bio 1129 */ 1130 nbio = push_bio(bio); 1131 nbio->bio_offset = volume->ondisk->vol_buf_beg + 1132 (zone2_offset & HAMMER_OFF_SHORT_MASK); 1133 #if 0 1134 /* 1135 * XXX disabled - our CRC check doesn't work if the OS 1136 * does bogus_page replacement on the direct-read. 1137 */ 1138 if (leaf && hammer_verify_data) { 1139 nbio->bio_done = hammer_io_direct_read_complete; 1140 nbio->bio_caller_info1.uvalue32 = leaf->data_crc; 1141 } 1142 #endif 1143 hammer_stats_disk_read += bp->b_bufsize; 1144 vn_strategy(volume->devvp, nbio); 1145 } 1146 hammer_rel_volume(volume, 0); 1147 done: 1148 if (error) { 1149 kprintf("hammer_direct_read: failed @ %016llx\n", 1150 (long long)zone2_offset); 1151 bp->b_error = error; 1152 bp->b_flags |= B_ERROR; 1153 biodone(bio); 1154 } 1155 return(error); 1156 } 1157 1158 #if 0 1159 /* 1160 * On completion of the BIO this callback must check the data CRC 1161 * and chain to the previous bio. 1162 */ 1163 static 1164 void 1165 hammer_io_direct_read_complete(struct bio *nbio) 1166 { 1167 struct bio *obio; 1168 struct buf *bp; 1169 u_int32_t rec_crc = nbio->bio_caller_info1.uvalue32; 1170 1171 bp = nbio->bio_buf; 1172 if (crc32(bp->b_data, bp->b_bufsize) != rec_crc) { 1173 kprintf("HAMMER: data_crc error @%016llx/%d\n", 1174 nbio->bio_offset, bp->b_bufsize); 1175 if (hammer_debug_critical) 1176 Debugger("data_crc on read"); 1177 bp->b_flags |= B_ERROR; 1178 bp->b_error = EIO; 1179 } 1180 obio = pop_bio(nbio); 1181 biodone(obio); 1182 } 1183 #endif 1184 1185 /* 1186 * Write a buffer associated with a front-end vnode directly to the 1187 * disk media. The bio may be issued asynchronously. 1188 * 1189 * The BIO is associated with the specified record and RECF_DIRECT_IO 1190 * is set. The recorded is added to its object. 1191 */ 1192 int 1193 hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record, 1194 struct bio *bio) 1195 { 1196 hammer_btree_leaf_elm_t leaf = &record->leaf; 1197 hammer_off_t buf_offset; 1198 hammer_off_t zone2_offset; 1199 hammer_volume_t volume; 1200 hammer_buffer_t buffer; 1201 struct buf *bp; 1202 struct bio *nbio; 1203 char *ptr; 1204 int vol_no; 1205 int error; 1206 1207 buf_offset = leaf->data_offset; 1208 1209 KKASSERT(buf_offset > HAMMER_ZONE_BTREE); 1210 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE); 1211 1212 if ((buf_offset & HAMMER_BUFMASK) == 0 && 1213 leaf->data_len >= HAMMER_BUFSIZE) { 1214 /* 1215 * We are using the vnode's bio to write directly to the 1216 * media, any hammer_buffer at the same zone-X offset will 1217 * now have stale data. 1218 */ 1219 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1220 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1221 volume = hammer_get_volume(hmp, vol_no, &error); 1222 1223 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1224 error = EIO; 1225 if (error == 0) { 1226 bp = bio->bio_buf; 1227 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0); 1228 /* 1229 hammer_del_buffers(hmp, buf_offset, 1230 zone2_offset, bp->b_bufsize); 1231 */ 1232 1233 /* 1234 * Second level bio - cached zone2 offset. 1235 * 1236 * (We can put our bio_done function in either the 1237 * 2nd or 3rd level). 1238 */ 1239 nbio = push_bio(bio); 1240 nbio->bio_offset = zone2_offset; 1241 nbio->bio_done = hammer_io_direct_write_complete; 1242 nbio->bio_caller_info1.ptr = record; 1243 record->zone2_offset = zone2_offset; 1244 record->flags |= HAMMER_RECF_DIRECT_IO | 1245 HAMMER_RECF_DIRECT_INVAL; 1246 1247 /* 1248 * Third level bio - raw offset specific to the 1249 * correct volume. 1250 */ 1251 zone2_offset &= HAMMER_OFF_SHORT_MASK; 1252 nbio = push_bio(nbio); 1253 nbio->bio_offset = volume->ondisk->vol_buf_beg + 1254 zone2_offset; 1255 hammer_stats_disk_write += bp->b_bufsize; 1256 vn_strategy(volume->devvp, nbio); 1257 hammer_io_flush_mark(volume); 1258 } 1259 hammer_rel_volume(volume, 0); 1260 } else { 1261 /* 1262 * Must fit in a standard HAMMER buffer. In this case all 1263 * consumers use the HAMMER buffer system and RECF_DIRECT_IO 1264 * does not need to be set-up. 1265 */ 1266 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0); 1267 buffer = NULL; 1268 ptr = hammer_bread(hmp, buf_offset, &error, &buffer); 1269 if (error == 0) { 1270 bp = bio->bio_buf; 1271 bp->b_flags |= B_AGE; 1272 hammer_io_modify(&buffer->io, 1); 1273 bcopy(bp->b_data, ptr, leaf->data_len); 1274 hammer_io_modify_done(&buffer->io); 1275 hammer_rel_buffer(buffer, 0); 1276 bp->b_resid = 0; 1277 biodone(bio); 1278 } 1279 } 1280 if (error == 0) { 1281 /* 1282 * The record is all setup now, add it. Potential conflics 1283 * have already been dealt with. 1284 */ 1285 error = hammer_mem_add(record); 1286 KKASSERT(error == 0); 1287 } else { 1288 /* 1289 * Major suckage occured. Also note: The record was never added 1290 * to the tree so we do not have to worry about the backend. 1291 */ 1292 kprintf("hammer_direct_write: failed @ %016llx\n", 1293 (long long)leaf->data_offset); 1294 bp = bio->bio_buf; 1295 bp->b_resid = 0; 1296 bp->b_error = EIO; 1297 bp->b_flags |= B_ERROR; 1298 biodone(bio); 1299 record->flags |= HAMMER_RECF_DELETED_FE; 1300 hammer_rel_mem_record(record); 1301 } 1302 return(error); 1303 } 1304 1305 /* 1306 * On completion of the BIO this callback must disconnect 1307 * it from the hammer_record and chain to the previous bio. 1308 * 1309 * An I/O error forces the mount to read-only. Data buffers 1310 * are not B_LOCKED like meta-data buffers are, so we have to 1311 * throw the buffer away to prevent the kernel from retrying. 1312 */ 1313 static 1314 void 1315 hammer_io_direct_write_complete(struct bio *nbio) 1316 { 1317 struct bio *obio; 1318 struct buf *bp; 1319 hammer_record_t record = nbio->bio_caller_info1.ptr; 1320 1321 bp = nbio->bio_buf; 1322 obio = pop_bio(nbio); 1323 if (bp->b_flags & B_ERROR) { 1324 hammer_critical_error(record->ip->hmp, record->ip, 1325 bp->b_error, 1326 "while writing bulk data"); 1327 bp->b_flags |= B_INVAL; 1328 } 1329 biodone(obio); 1330 1331 KKASSERT(record != NULL); 1332 KKASSERT(record->flags & HAMMER_RECF_DIRECT_IO); 1333 if (record->flags & HAMMER_RECF_DIRECT_WAIT) { 1334 record->flags &= ~(HAMMER_RECF_DIRECT_IO | 1335 HAMMER_RECF_DIRECT_WAIT); 1336 /* record can disappear once DIRECT_IO flag is cleared */ 1337 wakeup(&record->flags); 1338 } else { 1339 record->flags &= ~HAMMER_RECF_DIRECT_IO; 1340 /* record can disappear once DIRECT_IO flag is cleared */ 1341 } 1342 } 1343 1344 1345 /* 1346 * This is called before a record is either committed to the B-Tree 1347 * or destroyed, to resolve any associated direct-IO. 1348 * 1349 * (1) We must wait for any direct-IO related to the record to complete. 1350 * 1351 * (2) We must remove any buffer cache aliases for data accessed via 1352 * leaf->data_offset or zone2_offset so non-direct-IO consumers 1353 * (the mirroring and reblocking code) do not see stale data. 1354 */ 1355 void 1356 hammer_io_direct_wait(hammer_record_t record) 1357 { 1358 /* 1359 * Wait for I/O to complete 1360 */ 1361 if (record->flags & HAMMER_RECF_DIRECT_IO) { 1362 crit_enter(); 1363 while (record->flags & HAMMER_RECF_DIRECT_IO) { 1364 record->flags |= HAMMER_RECF_DIRECT_WAIT; 1365 tsleep(&record->flags, 0, "hmdiow", 0); 1366 } 1367 crit_exit(); 1368 } 1369 1370 /* 1371 * Invalidate any related buffer cache aliases associated with the 1372 * backing device. This is needed because the buffer cache buffer 1373 * for file data is associated with the file vnode, not the backing 1374 * device vnode. 1375 * 1376 * XXX I do not think this case can occur any more now that 1377 * reservations ensure that all such buffers are removed before 1378 * an area can be reused. 1379 */ 1380 if (record->flags & HAMMER_RECF_DIRECT_INVAL) { 1381 KKASSERT(record->leaf.data_offset); 1382 hammer_del_buffers(record->ip->hmp, record->leaf.data_offset, 1383 record->zone2_offset, record->leaf.data_len, 1384 1); 1385 record->flags &= ~HAMMER_RECF_DIRECT_INVAL; 1386 } 1387 } 1388 1389 /* 1390 * This is called to remove the second-level cached zone-2 offset from 1391 * frontend buffer cache buffers, now stale due to a data relocation. 1392 * These offsets are generated by cluster_read() via VOP_BMAP, or directly 1393 * by hammer_vop_strategy_read(). 1394 * 1395 * This is rather nasty because here we have something like the reblocker 1396 * scanning the raw B-Tree with no held references on anything, really, 1397 * other then a shared lock on the B-Tree node, and we have to access the 1398 * frontend's buffer cache to check for and clean out the association. 1399 * Specifically, if the reblocker is moving data on the disk, these cached 1400 * offsets will become invalid. 1401 * 1402 * Only data record types associated with the large-data zone are subject 1403 * to direct-io and need to be checked. 1404 * 1405 */ 1406 void 1407 hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf) 1408 { 1409 struct hammer_inode_info iinfo; 1410 int zone; 1411 1412 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA) 1413 return; 1414 zone = HAMMER_ZONE_DECODE(leaf->data_offset); 1415 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX) 1416 return; 1417 iinfo.obj_id = leaf->base.obj_id; 1418 iinfo.obj_asof = 0; /* unused */ 1419 iinfo.obj_localization = leaf->base.localization & 1420 HAMMER_LOCALIZE_PSEUDOFS_MASK; 1421 iinfo.u.leaf = leaf; 1422 hammer_scan_inode_snapshots(hmp, &iinfo, 1423 hammer_io_direct_uncache_callback, 1424 leaf); 1425 } 1426 1427 static int 1428 hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data) 1429 { 1430 hammer_inode_info_t iinfo = data; 1431 hammer_off_t data_offset; 1432 hammer_off_t file_offset; 1433 struct vnode *vp; 1434 struct buf *bp; 1435 int blksize; 1436 1437 if (ip->vp == NULL) 1438 return(0); 1439 data_offset = iinfo->u.leaf->data_offset; 1440 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len; 1441 blksize = iinfo->u.leaf->data_len; 1442 KKASSERT((blksize & HAMMER_BUFMASK) == 0); 1443 1444 hammer_ref(&ip->lock); 1445 if (hammer_get_vnode(ip, &vp) == 0) { 1446 if ((bp = findblk(ip->vp, file_offset, FINDBLK_TEST)) != NULL && 1447 bp->b_bio2.bio_offset != NOOFFSET) { 1448 bp = getblk(ip->vp, file_offset, blksize, 0, 0); 1449 bp->b_bio2.bio_offset = NOOFFSET; 1450 brelse(bp); 1451 } 1452 vput(vp); 1453 } 1454 hammer_rel_inode(ip, 0); 1455 return(0); 1456 } 1457 1458 1459 /* 1460 * This function is called when writes may have occured on the volume, 1461 * indicating that the device may be holding cached writes. 1462 */ 1463 static void 1464 hammer_io_flush_mark(hammer_volume_t volume) 1465 { 1466 volume->vol_flags |= HAMMER_VOLF_NEEDFLUSH; 1467 } 1468 1469 /* 1470 * This function ensures that the device has flushed any cached writes out. 1471 */ 1472 void 1473 hammer_io_flush_sync(hammer_mount_t hmp) 1474 { 1475 hammer_volume_t volume; 1476 struct buf *bp_base = NULL; 1477 struct buf *bp; 1478 1479 RB_FOREACH(volume, hammer_vol_rb_tree, &hmp->rb_vols_root) { 1480 if (volume->vol_flags & HAMMER_VOLF_NEEDFLUSH) { 1481 volume->vol_flags &= ~HAMMER_VOLF_NEEDFLUSH; 1482 bp = getpbuf(NULL); 1483 bp->b_bio1.bio_offset = 0; 1484 bp->b_bufsize = 0; 1485 bp->b_bcount = 0; 1486 bp->b_cmd = BUF_CMD_FLUSH; 1487 bp->b_bio1.bio_caller_info1.cluster_head = bp_base; 1488 bp->b_bio1.bio_done = biodone_sync; 1489 bp->b_bio1.bio_flags |= BIO_SYNC; 1490 bp_base = bp; 1491 vn_strategy(volume->devvp, &bp->b_bio1); 1492 } 1493 } 1494 while ((bp = bp_base) != NULL) { 1495 bp_base = bp->b_bio1.bio_caller_info1.cluster_head; 1496 biowait(&bp->b_bio1, "hmrFLS"); 1497 relpbuf(bp, NULL); 1498 } 1499 } 1500