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 * The io_token is required for anything which might race bioops and bio_done 49 * callbacks, with one exception: A successful hammer_try_interlock_norefs(). 50 * the fs_token will be held in all other cases. 51 */ 52 53 #include "hammer.h" 54 #include <sys/fcntl.h> 55 #include <sys/nlookup.h> 56 #include <sys/buf.h> 57 #include <sys/buf2.h> 58 59 static void hammer_io_modify(hammer_io_t io, int count); 60 static void hammer_io_deallocate(struct buf *bp); 61 #if 0 62 static void hammer_io_direct_read_complete(struct bio *nbio); 63 #endif 64 static void hammer_io_direct_write_complete(struct bio *nbio); 65 static int hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data); 66 static void hammer_io_set_modlist(struct hammer_io *io); 67 static void hammer_io_flush_mark(hammer_volume_t volume); 68 69 /* 70 * Initialize a new, already-zero'd hammer_io structure, or reinitialize 71 * an existing hammer_io structure which may have switched to another type. 72 */ 73 void 74 hammer_io_init(hammer_io_t io, hammer_volume_t volume, enum hammer_io_type type) 75 { 76 io->volume = volume; 77 io->hmp = volume->io.hmp; 78 io->type = type; 79 } 80 81 /* 82 * Helper routine to disassociate a buffer cache buffer from an I/O 83 * structure. The io must be interlocked and marked appropriately for 84 * reclamation. 85 * 86 * The io must be in a released state with the io->bp owned and 87 * locked by the caller of this function. When not called from an 88 * io_deallocate() this cannot race an io_deallocate() since the 89 * kernel would be unable to get the buffer lock in that case. 90 * (The released state in this case means we own the bp, not the 91 * hammer_io structure). 92 * 93 * The io may have 0 or 1 references depending on who called us. The 94 * caller is responsible for dealing with the refs. 95 * 96 * This call can only be made when no action is required on the buffer. 97 * 98 * This function is guaranteed not to race against anything because we 99 * own both the io lock and the bp lock and are interlocked with no 100 * references. 101 */ 102 static void 103 hammer_io_disassociate(hammer_io_structure_t iou) 104 { 105 struct buf *bp = iou->io.bp; 106 107 KKASSERT(iou->io.released); 108 KKASSERT(iou->io.modified == 0); 109 KKASSERT(LIST_FIRST(&bp->b_dep) == (void *)iou); 110 buf_dep_init(bp); 111 iou->io.bp = NULL; 112 113 /* 114 * If the buffer was locked someone wanted to get rid of it. 115 */ 116 if (bp->b_flags & B_LOCKED) { 117 atomic_add_int(&hammer_count_io_locked, -1); 118 bp->b_flags &= ~B_LOCKED; 119 } 120 if (iou->io.reclaim) { 121 bp->b_flags |= B_NOCACHE|B_RELBUF; 122 iou->io.reclaim = 0; 123 } 124 125 switch(iou->io.type) { 126 case HAMMER_STRUCTURE_VOLUME: 127 iou->volume.ondisk = NULL; 128 break; 129 case HAMMER_STRUCTURE_DATA_BUFFER: 130 case HAMMER_STRUCTURE_META_BUFFER: 131 case HAMMER_STRUCTURE_UNDO_BUFFER: 132 iou->buffer.ondisk = NULL; 133 break; 134 case HAMMER_STRUCTURE_DUMMY: 135 panic("hammer_io_disassociate: bad io type"); 136 break; 137 } 138 } 139 140 /* 141 * Wait for any physical IO to complete 142 * 143 * XXX we aren't interlocked against a spinlock or anything so there 144 * is a small window in the interlock / io->running == 0 test. 145 */ 146 void 147 hammer_io_wait(hammer_io_t io) 148 { 149 if (io->running) { 150 hammer_mount_t hmp = io->hmp; 151 152 lwkt_gettoken(&hmp->io_token); 153 while (io->running) { 154 io->waiting = 1; 155 tsleep_interlock(io, 0); 156 if (io->running) 157 tsleep(io, PINTERLOCKED, "hmrflw", hz); 158 } 159 lwkt_reltoken(&hmp->io_token); 160 } 161 } 162 163 /* 164 * Wait for all currently queued HAMMER-initiated I/Os to complete. 165 * 166 * This is not supposed to count direct I/O's but some can leak 167 * through (for non-full-sized direct I/Os). 168 */ 169 void 170 hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush) 171 { 172 struct hammer_io iodummy; 173 hammer_io_t io; 174 175 /* 176 * Degenerate case, no I/O is running 177 */ 178 lwkt_gettoken(&hmp->io_token); 179 if (TAILQ_EMPTY(&hmp->iorun_list)) { 180 lwkt_reltoken(&hmp->io_token); 181 if (doflush) 182 hammer_io_flush_sync(hmp); 183 return; 184 } 185 bzero(&iodummy, sizeof(iodummy)); 186 iodummy.type = HAMMER_STRUCTURE_DUMMY; 187 188 /* 189 * Add placemarker and then wait until it becomes the head of 190 * the list. 191 */ 192 TAILQ_INSERT_TAIL(&hmp->iorun_list, &iodummy, iorun_entry); 193 while (TAILQ_FIRST(&hmp->iorun_list) != &iodummy) { 194 tsleep(&iodummy, 0, ident, 0); 195 } 196 197 /* 198 * Chain in case several placemarkers are present. 199 */ 200 TAILQ_REMOVE(&hmp->iorun_list, &iodummy, iorun_entry); 201 io = TAILQ_FIRST(&hmp->iorun_list); 202 if (io && io->type == HAMMER_STRUCTURE_DUMMY) 203 wakeup(io); 204 lwkt_reltoken(&hmp->io_token); 205 206 if (doflush) 207 hammer_io_flush_sync(hmp); 208 } 209 210 /* 211 * Clear a flagged error condition on a I/O buffer. The caller must hold 212 * its own ref on the buffer. 213 */ 214 void 215 hammer_io_clear_error(struct hammer_io *io) 216 { 217 hammer_mount_t hmp = io->hmp; 218 219 lwkt_gettoken(&hmp->io_token); 220 if (io->ioerror) { 221 io->ioerror = 0; 222 hammer_rel(&io->lock); 223 KKASSERT(hammer_isactive(&io->lock)); 224 } 225 lwkt_reltoken(&hmp->io_token); 226 } 227 228 void 229 hammer_io_clear_error_noassert(struct hammer_io *io) 230 { 231 hammer_mount_t hmp = io->hmp; 232 233 lwkt_gettoken(&hmp->io_token); 234 if (io->ioerror) { 235 io->ioerror = 0; 236 hammer_rel(&io->lock); 237 } 238 lwkt_reltoken(&hmp->io_token); 239 } 240 241 /* 242 * This is an advisory function only which tells the buffer cache 243 * the bp is not a meta-data buffer, even though it is backed by 244 * a block device. 245 * 246 * This is used by HAMMER's reblocking code to avoid trying to 247 * swapcache the filesystem's data when it is read or written 248 * by the reblocking code. 249 * 250 * The caller has a ref on the buffer preventing the bp from 251 * being disassociated from it. 252 */ 253 void 254 hammer_io_notmeta(hammer_buffer_t buffer) 255 { 256 if ((buffer->io.bp->b_flags & B_NOTMETA) == 0) { 257 hammer_mount_t hmp = buffer->io.hmp; 258 259 lwkt_gettoken(&hmp->io_token); 260 buffer->io.bp->b_flags |= B_NOTMETA; 261 lwkt_reltoken(&hmp->io_token); 262 } 263 } 264 265 /* 266 * Load bp for a HAMMER structure. The io must be exclusively locked by 267 * the caller. 268 * 269 * This routine is mostly used on meta-data and small-data blocks. Generally 270 * speaking HAMMER assumes some locality of reference and will cluster. 271 * 272 * Note that the caller (hammer_ondisk.c) may place further restrictions 273 * on clusterability via the limit (in bytes). Typically large-data 274 * zones cannot be clustered due to their mixed buffer sizes. This is 275 * not an issue since such clustering occurs in hammer_vnops at the 276 * regular file layer, whereas this is the buffered block device layer. 277 * 278 * No I/O callbacks can occur while we hold the buffer locked. 279 */ 280 int 281 hammer_io_read(struct vnode *devvp, struct hammer_io *io, int limit) 282 { 283 struct buf *bp; 284 int error; 285 286 if ((bp = io->bp) == NULL) { 287 atomic_add_int(&hammer_count_io_running_read, io->bytes); 288 if (hammer_cluster_enable && limit > io->bytes) { 289 error = cluster_read(devvp, io->offset + limit, 290 io->offset, io->bytes, 291 HAMMER_CLUSTER_SIZE, 292 HAMMER_CLUSTER_SIZE, 293 &io->bp); 294 } else { 295 error = bread(devvp, io->offset, io->bytes, &io->bp); 296 } 297 hammer_stats_disk_read += io->bytes; 298 atomic_add_int(&hammer_count_io_running_read, -io->bytes); 299 300 /* 301 * The code generally assumes b_ops/b_dep has been set-up, 302 * even if we error out here. 303 */ 304 bp = io->bp; 305 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) { 306 const char *metatype; 307 308 switch(io->type) { 309 case HAMMER_STRUCTURE_VOLUME: 310 metatype = "volume"; 311 break; 312 case HAMMER_STRUCTURE_META_BUFFER: 313 switch(((struct hammer_buffer *)io)-> 314 zoneX_offset & HAMMER_OFF_ZONE_MASK) { 315 case HAMMER_ZONE_BTREE: 316 metatype = "btree"; 317 break; 318 case HAMMER_ZONE_META: 319 metatype = "meta"; 320 break; 321 case HAMMER_ZONE_FREEMAP: 322 metatype = "freemap"; 323 break; 324 default: 325 metatype = "meta?"; 326 break; 327 } 328 break; 329 case HAMMER_STRUCTURE_DATA_BUFFER: 330 metatype = "data"; 331 break; 332 case HAMMER_STRUCTURE_UNDO_BUFFER: 333 metatype = "undo"; 334 break; 335 default: 336 metatype = "unknown"; 337 break; 338 } 339 kprintf("doff %016jx %s\n", 340 (intmax_t)bp->b_bio2.bio_offset, 341 metatype); 342 } 343 bp->b_flags &= ~B_IODEBUG; 344 bp->b_ops = &hammer_bioops; 345 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 346 347 /* io->worklist is locked by the io lock */ 348 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node); 349 BUF_KERNPROC(bp); 350 KKASSERT(io->modified == 0); 351 KKASSERT(io->running == 0); 352 KKASSERT(io->waiting == 0); 353 io->released = 0; /* we hold an active lock on bp */ 354 } else { 355 error = 0; 356 } 357 return(error); 358 } 359 360 /* 361 * Similar to hammer_io_read() but returns a zero'd out buffer instead. 362 * Must be called with the IO exclusively locked. 363 * 364 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background 365 * I/O by forcing the buffer to not be in a released state before calling 366 * it. 367 * 368 * This function will also mark the IO as modified but it will not 369 * increment the modify_refs count. 370 * 371 * No I/O callbacks can occur while we hold the buffer locked. 372 */ 373 int 374 hammer_io_new(struct vnode *devvp, struct hammer_io *io) 375 { 376 struct buf *bp; 377 378 if ((bp = io->bp) == NULL) { 379 io->bp = getblk(devvp, io->offset, io->bytes, 0, 0); 380 bp = io->bp; 381 bp->b_ops = &hammer_bioops; 382 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 383 384 /* io->worklist is locked by the io lock */ 385 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node); 386 io->released = 0; 387 KKASSERT(io->running == 0); 388 io->waiting = 0; 389 BUF_KERNPROC(bp); 390 } else { 391 if (io->released) { 392 regetblk(bp); 393 BUF_KERNPROC(bp); 394 io->released = 0; 395 } 396 } 397 hammer_io_modify(io, 0); 398 vfs_bio_clrbuf(bp); 399 return(0); 400 } 401 402 /* 403 * Advance the activity count on the underlying buffer because 404 * HAMMER does not getblk/brelse on every access. 405 * 406 * The io->bp cannot go away while the buffer is referenced. 407 */ 408 void 409 hammer_io_advance(struct hammer_io *io) 410 { 411 if (io->bp) 412 buf_act_advance(io->bp); 413 } 414 415 /* 416 * Remove potential device level aliases against buffers managed by high level 417 * vnodes. Aliases can also be created due to mixed buffer sizes or via 418 * direct access to the backing store device. 419 * 420 * This is nasty because the buffers are also VMIO-backed. Even if a buffer 421 * does not exist its backing VM pages might, and we have to invalidate 422 * those as well or a getblk() will reinstate them. 423 * 424 * Buffer cache buffers associated with hammer_buffers cannot be 425 * invalidated. 426 */ 427 int 428 hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset) 429 { 430 hammer_io_structure_t iou; 431 hammer_mount_t hmp; 432 hammer_off_t phys_offset; 433 struct buf *bp; 434 int error; 435 436 hmp = volume->io.hmp; 437 lwkt_gettoken(&hmp->io_token); 438 439 /* 440 * If a device buffer already exists for the specified physical 441 * offset use that, otherwise instantiate a buffer to cover any 442 * related VM pages, set BNOCACHE, and brelse(). 443 */ 444 phys_offset = volume->ondisk->vol_buf_beg + 445 (zone2_offset & HAMMER_OFF_SHORT_MASK); 446 if ((bp = findblk(volume->devvp, phys_offset, 0)) != NULL) 447 bremfree(bp); 448 else 449 bp = getblk(volume->devvp, phys_offset, HAMMER_BUFSIZE, 0, 0); 450 451 if ((iou = (void *)LIST_FIRST(&bp->b_dep)) != NULL) { 452 #if 0 453 hammer_ref(&iou->io.lock); 454 hammer_io_clear_modify(&iou->io, 1); 455 bundirty(bp); 456 iou->io.released = 0; 457 BUF_KERNPROC(bp); 458 iou->io.reclaim = 1; 459 iou->io.waitdep = 1; /* XXX this is a fs_token field */ 460 KKASSERT(hammer_isactive(&iou->io.lock) == 1); 461 hammer_rel_buffer(&iou->buffer, 0); 462 /*hammer_io_deallocate(bp);*/ 463 #endif 464 bqrelse(bp); 465 error = EAGAIN; 466 } else { 467 KKASSERT((bp->b_flags & B_LOCKED) == 0); 468 bundirty(bp); 469 bp->b_flags |= B_NOCACHE|B_RELBUF; 470 brelse(bp); 471 error = 0; 472 } 473 lwkt_reltoken(&hmp->io_token); 474 return(error); 475 } 476 477 /* 478 * This routine is called on the last reference to a hammer structure. 479 * The io must be interlocked with a refcount of zero. The hammer structure 480 * will remain interlocked on return. 481 * 482 * This routine may return a non-NULL bp to the caller for dispoal. 483 * The caller typically brelse()'s the bp. 484 * 485 * The bp may or may not still be passively associated with the IO. It 486 * will remain passively associated if it is unreleasable (e.g. a modified 487 * meta-data buffer). 488 * 489 * The only requirement here is that modified meta-data and volume-header 490 * buffer may NOT be disassociated from the IO structure, and consequently 491 * we also leave such buffers actively associated with the IO if they already 492 * are (since the kernel can't do anything with them anyway). Only the 493 * flusher is allowed to write such buffers out. Modified pure-data and 494 * undo buffers are returned to the kernel but left passively associated 495 * so we can track when the kernel writes the bp out. 496 */ 497 struct buf * 498 hammer_io_release(struct hammer_io *io, int flush) 499 { 500 union hammer_io_structure *iou = (void *)io; 501 struct buf *bp; 502 503 if ((bp = io->bp) == NULL) 504 return(NULL); 505 506 /* 507 * Try to flush a dirty IO to disk if asked to by the 508 * caller or if the kernel tried to flush the buffer in the past. 509 * 510 * Kernel-initiated flushes are only allowed for pure-data buffers. 511 * meta-data and volume buffers can only be flushed explicitly 512 * by HAMMER. 513 */ 514 if (io->modified) { 515 if (flush) { 516 hammer_io_flush(io, 0); 517 } else if (bp->b_flags & B_LOCKED) { 518 switch(io->type) { 519 case HAMMER_STRUCTURE_DATA_BUFFER: 520 hammer_io_flush(io, 0); 521 break; 522 case HAMMER_STRUCTURE_UNDO_BUFFER: 523 hammer_io_flush(io, hammer_undo_reclaim(io)); 524 break; 525 default: 526 break; 527 } 528 } /* else no explicit request to flush the buffer */ 529 } 530 531 /* 532 * Wait for the IO to complete if asked to. This occurs when 533 * the buffer must be disposed of definitively during an umount 534 * or buffer invalidation. 535 */ 536 if (io->waitdep && io->running) { 537 hammer_io_wait(io); 538 } 539 540 /* 541 * Return control of the buffer to the kernel (with the provisio 542 * that our bioops can override kernel decisions with regards to 543 * the buffer). 544 */ 545 if ((flush || io->reclaim) && io->modified == 0 && io->running == 0) { 546 /* 547 * Always disassociate the bp if an explicit flush 548 * was requested and the IO completed with no error 549 * (so unmount can really clean up the structure). 550 */ 551 if (io->released) { 552 regetblk(bp); 553 BUF_KERNPROC(bp); 554 } else { 555 io->released = 1; 556 } 557 hammer_io_disassociate((hammer_io_structure_t)io); 558 /* return the bp */ 559 } else if (io->modified) { 560 /* 561 * Only certain IO types can be released to the kernel if 562 * the buffer has been modified. 563 * 564 * volume and meta-data IO types may only be explicitly 565 * flushed by HAMMER. 566 */ 567 switch(io->type) { 568 case HAMMER_STRUCTURE_DATA_BUFFER: 569 case HAMMER_STRUCTURE_UNDO_BUFFER: 570 if (io->released == 0) { 571 io->released = 1; 572 bdwrite(bp); 573 } 574 break; 575 default: 576 break; 577 } 578 bp = NULL; /* bp left associated */ 579 } else if (io->released == 0) { 580 /* 581 * Clean buffers can be generally released to the kernel. 582 * We leave the bp passively associated with the HAMMER 583 * structure and use bioops to disconnect it later on 584 * if the kernel wants to discard the buffer. 585 * 586 * We can steal the structure's ownership of the bp. 587 */ 588 io->released = 1; 589 if (bp->b_flags & B_LOCKED) { 590 hammer_io_disassociate(iou); 591 /* return the bp */ 592 } else { 593 if (io->reclaim) { 594 hammer_io_disassociate(iou); 595 /* return the bp */ 596 } else { 597 /* return the bp (bp passively associated) */ 598 } 599 } 600 } else { 601 /* 602 * A released buffer is passively associate with our 603 * hammer_io structure. The kernel cannot destroy it 604 * without making a bioops call. If the kernel (B_LOCKED) 605 * or we (reclaim) requested that the buffer be destroyed 606 * we destroy it, otherwise we do a quick get/release to 607 * reset its position in the kernel's LRU list. 608 * 609 * Leaving the buffer passively associated allows us to 610 * use the kernel's LRU buffer flushing mechanisms rather 611 * then rolling our own. 612 * 613 * XXX there are two ways of doing this. We can re-acquire 614 * and passively release to reset the LRU, or not. 615 */ 616 if (io->running == 0) { 617 regetblk(bp); 618 if ((bp->b_flags & B_LOCKED) || io->reclaim) { 619 hammer_io_disassociate(iou); 620 /* return the bp */ 621 } else { 622 /* return the bp (bp passively associated) */ 623 } 624 } else { 625 /* 626 * bp is left passively associated but we do not 627 * try to reacquire it. Interactions with the io 628 * structure will occur on completion of the bp's 629 * I/O. 630 */ 631 bp = NULL; 632 } 633 } 634 return(bp); 635 } 636 637 /* 638 * This routine is called with a locked IO when a flush is desired and 639 * no other references to the structure exists other then ours. This 640 * routine is ONLY called when HAMMER believes it is safe to flush a 641 * potentially modified buffer out. 642 * 643 * The locked io or io reference prevents a flush from being initiated 644 * by the kernel. 645 */ 646 void 647 hammer_io_flush(struct hammer_io *io, int reclaim) 648 { 649 struct buf *bp; 650 hammer_mount_t hmp; 651 652 /* 653 * Degenerate case - nothing to flush if nothing is dirty. 654 */ 655 if (io->modified == 0) 656 return; 657 658 KKASSERT(io->bp); 659 KKASSERT(io->modify_refs <= 0); 660 661 /* 662 * Acquire ownership of the bp, particularly before we clear our 663 * modified flag. 664 * 665 * We are going to bawrite() this bp. Don't leave a window where 666 * io->released is set, we actually own the bp rather then our 667 * buffer. 668 * 669 * The io_token should not be required here as only 670 */ 671 hmp = io->hmp; 672 bp = io->bp; 673 if (io->released) { 674 regetblk(bp); 675 /* BUF_KERNPROC(io->bp); */ 676 /* io->released = 0; */ 677 KKASSERT(io->released); 678 KKASSERT(io->bp == bp); 679 } else { 680 io->released = 1; 681 } 682 683 if (reclaim) { 684 io->reclaim = 1; 685 if ((bp->b_flags & B_LOCKED) == 0) { 686 bp->b_flags |= B_LOCKED; 687 atomic_add_int(&hammer_count_io_locked, 1); 688 } 689 } 690 691 /* 692 * Acquire exclusive access to the bp and then clear the modified 693 * state of the buffer prior to issuing I/O to interlock any 694 * modifications made while the I/O is in progress. This shouldn't 695 * happen anyway but losing data would be worse. The modified bit 696 * will be rechecked after the IO completes. 697 * 698 * NOTE: This call also finalizes the buffer's content (inval == 0). 699 * 700 * This is only legal when lock.refs == 1 (otherwise we might clear 701 * the modified bit while there are still users of the cluster 702 * modifying the data). 703 * 704 * Do this before potentially blocking so any attempt to modify the 705 * ondisk while we are blocked blocks waiting for us. 706 */ 707 hammer_ref(&io->lock); 708 hammer_io_clear_modify(io, 0); 709 hammer_rel(&io->lock); 710 711 if (hammer_debug_io & 0x0002) 712 kprintf("hammer io_write %016jx\n", bp->b_bio1.bio_offset); 713 714 /* 715 * Transfer ownership to the kernel and initiate I/O. 716 * 717 * NOTE: We do not hold io_token so an atomic op is required to 718 * update io_running_space. 719 */ 720 io->running = 1; 721 atomic_add_int(&hmp->io_running_space, io->bytes); 722 atomic_add_int(&hammer_count_io_running_write, io->bytes); 723 lwkt_gettoken(&hmp->io_token); 724 TAILQ_INSERT_TAIL(&hmp->iorun_list, io, iorun_entry); 725 lwkt_reltoken(&hmp->io_token); 726 bawrite(bp); 727 hammer_io_flush_mark(io->volume); 728 } 729 730 /************************************************************************ 731 * BUFFER DIRTYING * 732 ************************************************************************ 733 * 734 * These routines deal with dependancies created when IO buffers get 735 * modified. The caller must call hammer_modify_*() on a referenced 736 * HAMMER structure prior to modifying its on-disk data. 737 * 738 * Any intent to modify an IO buffer acquires the related bp and imposes 739 * various write ordering dependancies. 740 */ 741 742 /* 743 * Mark a HAMMER structure as undergoing modification. Meta-data buffers 744 * are locked until the flusher can deal with them, pure data buffers 745 * can be written out. 746 * 747 * The referenced io prevents races. 748 */ 749 static 750 void 751 hammer_io_modify(hammer_io_t io, int count) 752 { 753 /* 754 * io->modify_refs must be >= 0 755 */ 756 while (io->modify_refs < 0) { 757 io->waitmod = 1; 758 tsleep(io, 0, "hmrmod", 0); 759 } 760 761 /* 762 * Shortcut if nothing to do. 763 */ 764 KKASSERT(hammer_isactive(&io->lock) && io->bp != NULL); 765 io->modify_refs += count; 766 if (io->modified && io->released == 0) 767 return; 768 769 /* 770 * NOTE: It is important not to set the modified bit 771 * until after we have acquired the bp or we risk 772 * racing against checkwrite. 773 */ 774 hammer_lock_ex(&io->lock); 775 if (io->released) { 776 regetblk(io->bp); 777 BUF_KERNPROC(io->bp); 778 io->released = 0; 779 } 780 if (io->modified == 0) { 781 hammer_io_set_modlist(io); 782 io->modified = 1; 783 } 784 hammer_unlock(&io->lock); 785 } 786 787 static __inline 788 void 789 hammer_io_modify_done(hammer_io_t io) 790 { 791 KKASSERT(io->modify_refs > 0); 792 --io->modify_refs; 793 if (io->modify_refs == 0 && io->waitmod) { 794 io->waitmod = 0; 795 wakeup(io); 796 } 797 } 798 799 /* 800 * The write interlock blocks other threads trying to modify a buffer 801 * (they block in hammer_io_modify()) after us, or blocks us while other 802 * threads are in the middle of modifying a buffer. 803 * 804 * The caller also has a ref on the io, however if we are not careful 805 * we will race bioops callbacks (checkwrite). To deal with this 806 * we must at least acquire and release the io_token, and it is probably 807 * better to hold it through the setting of modify_refs. 808 */ 809 void 810 hammer_io_write_interlock(hammer_io_t io) 811 { 812 hammer_mount_t hmp = io->hmp; 813 814 lwkt_gettoken(&hmp->io_token); 815 while (io->modify_refs != 0) { 816 io->waitmod = 1; 817 tsleep(io, 0, "hmrmod", 0); 818 } 819 io->modify_refs = -1; 820 lwkt_reltoken(&hmp->io_token); 821 } 822 823 void 824 hammer_io_done_interlock(hammer_io_t io) 825 { 826 KKASSERT(io->modify_refs == -1); 827 io->modify_refs = 0; 828 if (io->waitmod) { 829 io->waitmod = 0; 830 wakeup(io); 831 } 832 } 833 834 /* 835 * Caller intends to modify a volume's ondisk structure. 836 * 837 * This is only allowed if we are the flusher or we have a ref on the 838 * sync_lock. 839 */ 840 void 841 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume, 842 void *base, int len) 843 { 844 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 845 846 hammer_io_modify(&volume->io, 1); 847 if (len) { 848 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk; 849 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 850 hammer_generate_undo(trans, 851 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset), 852 base, len); 853 } 854 } 855 856 /* 857 * Caller intends to modify a buffer's ondisk structure. 858 * 859 * This is only allowed if we are the flusher or we have a ref on the 860 * sync_lock. 861 */ 862 void 863 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer, 864 void *base, int len) 865 { 866 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 867 868 hammer_io_modify(&buffer->io, 1); 869 if (len) { 870 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk; 871 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 872 hammer_generate_undo(trans, 873 buffer->zone2_offset + rel_offset, 874 base, len); 875 } 876 } 877 878 void 879 hammer_modify_volume_done(hammer_volume_t volume) 880 { 881 hammer_io_modify_done(&volume->io); 882 } 883 884 void 885 hammer_modify_buffer_done(hammer_buffer_t buffer) 886 { 887 hammer_io_modify_done(&buffer->io); 888 } 889 890 /* 891 * Mark an entity as not being dirty any more and finalize any 892 * delayed adjustments to the buffer. 893 * 894 * Delayed adjustments are an important performance enhancement, allowing 895 * us to avoid recalculating B-Tree node CRCs over and over again when 896 * making bulk-modifications to the B-Tree. 897 * 898 * If inval is non-zero delayed adjustments are ignored. 899 * 900 * This routine may dereference related btree nodes and cause the 901 * buffer to be dereferenced. The caller must own a reference on io. 902 */ 903 void 904 hammer_io_clear_modify(struct hammer_io *io, int inval) 905 { 906 hammer_mount_t hmp; 907 908 /* 909 * io_token is needed to avoid races on mod_list 910 */ 911 if (io->modified == 0) 912 return; 913 hmp = io->hmp; 914 lwkt_gettoken(&hmp->io_token); 915 if (io->modified == 0) { 916 lwkt_reltoken(&hmp->io_token); 917 return; 918 } 919 920 /* 921 * Take us off the mod-list and clear the modified bit. 922 */ 923 KKASSERT(io->mod_list != NULL); 924 if (io->mod_list == &io->hmp->volu_list || 925 io->mod_list == &io->hmp->meta_list) { 926 io->hmp->locked_dirty_space -= io->bytes; 927 atomic_add_int(&hammer_count_dirtybufspace, -io->bytes); 928 } 929 TAILQ_REMOVE(io->mod_list, io, mod_entry); 930 io->mod_list = NULL; 931 io->modified = 0; 932 933 lwkt_reltoken(&hmp->io_token); 934 935 /* 936 * If this bit is not set there are no delayed adjustments. 937 */ 938 if (io->gencrc == 0) 939 return; 940 io->gencrc = 0; 941 942 /* 943 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference 944 * on the node (& underlying buffer). Release the node after clearing 945 * the flag. 946 */ 947 if (io->type == HAMMER_STRUCTURE_META_BUFFER) { 948 hammer_buffer_t buffer = (void *)io; 949 hammer_node_t node; 950 951 restart: 952 TAILQ_FOREACH(node, &buffer->clist, entry) { 953 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) 954 continue; 955 node->flags &= ~HAMMER_NODE_NEEDSCRC; 956 KKASSERT(node->ondisk); 957 if (inval == 0) 958 node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE); 959 hammer_rel_node(node); 960 goto restart; 961 } 962 } 963 /* caller must still have ref on io */ 964 KKASSERT(hammer_isactive(&io->lock)); 965 } 966 967 /* 968 * Clear the IO's modify list. Even though the IO is no longer modified 969 * it may still be on the lose_list. This routine is called just before 970 * the governing hammer_buffer is destroyed. 971 * 972 * mod_list requires io_token protection. 973 */ 974 void 975 hammer_io_clear_modlist(struct hammer_io *io) 976 { 977 hammer_mount_t hmp = io->hmp; 978 979 KKASSERT(io->modified == 0); 980 if (io->mod_list) { 981 lwkt_gettoken(&hmp->io_token); 982 if (io->mod_list) { 983 KKASSERT(io->mod_list == &io->hmp->lose_list); 984 TAILQ_REMOVE(io->mod_list, io, mod_entry); 985 io->mod_list = NULL; 986 } 987 lwkt_reltoken(&hmp->io_token); 988 } 989 } 990 991 static void 992 hammer_io_set_modlist(struct hammer_io *io) 993 { 994 struct hammer_mount *hmp = io->hmp; 995 996 lwkt_gettoken(&hmp->io_token); 997 KKASSERT(io->mod_list == NULL); 998 999 switch(io->type) { 1000 case HAMMER_STRUCTURE_VOLUME: 1001 io->mod_list = &hmp->volu_list; 1002 hmp->locked_dirty_space += io->bytes; 1003 atomic_add_int(&hammer_count_dirtybufspace, io->bytes); 1004 break; 1005 case HAMMER_STRUCTURE_META_BUFFER: 1006 io->mod_list = &hmp->meta_list; 1007 hmp->locked_dirty_space += io->bytes; 1008 atomic_add_int(&hammer_count_dirtybufspace, io->bytes); 1009 break; 1010 case HAMMER_STRUCTURE_UNDO_BUFFER: 1011 io->mod_list = &hmp->undo_list; 1012 break; 1013 case HAMMER_STRUCTURE_DATA_BUFFER: 1014 io->mod_list = &hmp->data_list; 1015 break; 1016 case HAMMER_STRUCTURE_DUMMY: 1017 panic("hammer_io_disassociate: bad io type"); 1018 break; 1019 } 1020 TAILQ_INSERT_TAIL(io->mod_list, io, mod_entry); 1021 lwkt_reltoken(&hmp->io_token); 1022 } 1023 1024 /************************************************************************ 1025 * HAMMER_BIOOPS * 1026 ************************************************************************ 1027 * 1028 */ 1029 1030 /* 1031 * Pre-IO initiation kernel callback - cluster build only 1032 * 1033 * bioops callback - hold io_token 1034 */ 1035 static void 1036 hammer_io_start(struct buf *bp) 1037 { 1038 /* nothing to do, so io_token not needed */ 1039 } 1040 1041 /* 1042 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT! 1043 * 1044 * NOTE: HAMMER may modify a data buffer after we have initiated write 1045 * I/O. 1046 * 1047 * NOTE: MPSAFE callback 1048 * 1049 * bioops callback - hold io_token 1050 */ 1051 static void 1052 hammer_io_complete(struct buf *bp) 1053 { 1054 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep); 1055 struct hammer_mount *hmp = iou->io.hmp; 1056 struct hammer_io *ionext; 1057 1058 lwkt_gettoken(&hmp->io_token); 1059 1060 KKASSERT(iou->io.released == 1); 1061 1062 /* 1063 * Deal with people waiting for I/O to drain 1064 */ 1065 if (iou->io.running) { 1066 /* 1067 * Deal with critical write errors. Once a critical error 1068 * has been flagged in hmp the UNDO FIFO will not be updated. 1069 * That way crash recover will give us a consistent 1070 * filesystem. 1071 * 1072 * Because of this we can throw away failed UNDO buffers. If 1073 * we throw away META or DATA buffers we risk corrupting 1074 * the now read-only version of the filesystem visible to 1075 * the user. Clear B_ERROR so the buffer is not re-dirtied 1076 * by the kernel and ref the io so it doesn't get thrown 1077 * away. 1078 */ 1079 if (bp->b_flags & B_ERROR) { 1080 lwkt_gettoken(&hmp->fs_token); 1081 hammer_critical_error(hmp, NULL, bp->b_error, 1082 "while flushing meta-data"); 1083 lwkt_reltoken(&hmp->fs_token); 1084 1085 switch(iou->io.type) { 1086 case HAMMER_STRUCTURE_UNDO_BUFFER: 1087 break; 1088 default: 1089 if (iou->io.ioerror == 0) { 1090 iou->io.ioerror = 1; 1091 hammer_ref(&iou->io.lock); 1092 } 1093 break; 1094 } 1095 bp->b_flags &= ~B_ERROR; 1096 bundirty(bp); 1097 #if 0 1098 hammer_io_set_modlist(&iou->io); 1099 iou->io.modified = 1; 1100 #endif 1101 } 1102 hammer_stats_disk_write += iou->io.bytes; 1103 atomic_add_int(&hammer_count_io_running_write, -iou->io.bytes); 1104 atomic_add_int(&hmp->io_running_space, -iou->io.bytes); 1105 if (hmp->io_running_wakeup && 1106 hmp->io_running_space < hammer_limit_running_io / 2) { 1107 hmp->io_running_wakeup = 0; 1108 wakeup(&hmp->io_running_wakeup); 1109 } 1110 KKASSERT(hmp->io_running_space >= 0); 1111 iou->io.running = 0; 1112 1113 /* 1114 * Remove from iorun list and wakeup any multi-io waiter(s). 1115 */ 1116 if (TAILQ_FIRST(&hmp->iorun_list) == &iou->io) { 1117 ionext = TAILQ_NEXT(&iou->io, iorun_entry); 1118 if (ionext && ionext->type == HAMMER_STRUCTURE_DUMMY) 1119 wakeup(ionext); 1120 } 1121 TAILQ_REMOVE(&hmp->iorun_list, &iou->io, iorun_entry); 1122 } else { 1123 hammer_stats_disk_read += iou->io.bytes; 1124 } 1125 1126 if (iou->io.waiting) { 1127 iou->io.waiting = 0; 1128 wakeup(iou); 1129 } 1130 1131 /* 1132 * If B_LOCKED is set someone wanted to deallocate the bp at some 1133 * point, try to do it now. The operation will fail if there are 1134 * refs or if hammer_io_deallocate() is unable to gain the 1135 * interlock. 1136 */ 1137 if (bp->b_flags & B_LOCKED) { 1138 atomic_add_int(&hammer_count_io_locked, -1); 1139 bp->b_flags &= ~B_LOCKED; 1140 hammer_io_deallocate(bp); 1141 /* structure may be dead now */ 1142 } 1143 lwkt_reltoken(&hmp->io_token); 1144 } 1145 1146 /* 1147 * Callback from kernel when it wishes to deallocate a passively 1148 * associated structure. This mostly occurs with clean buffers 1149 * but it may be possible for a holding structure to be marked dirty 1150 * while its buffer is passively associated. The caller owns the bp. 1151 * 1152 * If we cannot disassociate we set B_LOCKED to prevent the buffer 1153 * from getting reused. 1154 * 1155 * WARNING: Because this can be called directly by getnewbuf we cannot 1156 * recurse into the tree. If a bp cannot be immediately disassociated 1157 * our only recourse is to set B_LOCKED. 1158 * 1159 * WARNING: This may be called from an interrupt via hammer_io_complete() 1160 * 1161 * bioops callback - hold io_token 1162 */ 1163 static void 1164 hammer_io_deallocate(struct buf *bp) 1165 { 1166 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep); 1167 hammer_mount_t hmp; 1168 1169 hmp = iou->io.hmp; 1170 1171 lwkt_gettoken(&hmp->io_token); 1172 1173 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0); 1174 if (hammer_try_interlock_norefs(&iou->io.lock) == 0) { 1175 /* 1176 * We cannot safely disassociate a bp from a referenced 1177 * or interlocked HAMMER structure. 1178 */ 1179 bp->b_flags |= B_LOCKED; 1180 atomic_add_int(&hammer_count_io_locked, 1); 1181 } else if (iou->io.modified) { 1182 /* 1183 * It is not legal to disassociate a modified buffer. This 1184 * case really shouldn't ever occur. 1185 */ 1186 bp->b_flags |= B_LOCKED; 1187 atomic_add_int(&hammer_count_io_locked, 1); 1188 hammer_put_interlock(&iou->io.lock, 0); 1189 } else { 1190 /* 1191 * Disassociate the BP. If the io has no refs left we 1192 * have to add it to the loose list. The kernel has 1193 * locked the buffer and therefore our io must be 1194 * in a released state. 1195 */ 1196 hammer_io_disassociate(iou); 1197 if (iou->io.type != HAMMER_STRUCTURE_VOLUME) { 1198 KKASSERT(iou->io.bp == NULL); 1199 KKASSERT(iou->io.mod_list == NULL); 1200 iou->io.mod_list = &hmp->lose_list; 1201 TAILQ_INSERT_TAIL(iou->io.mod_list, &iou->io, mod_entry); 1202 } 1203 hammer_put_interlock(&iou->io.lock, 1); 1204 } 1205 lwkt_reltoken(&hmp->io_token); 1206 } 1207 1208 /* 1209 * bioops callback - hold io_token 1210 */ 1211 static int 1212 hammer_io_fsync(struct vnode *vp) 1213 { 1214 /* nothing to do, so io_token not needed */ 1215 return(0); 1216 } 1217 1218 /* 1219 * NOTE: will not be called unless we tell the kernel about the 1220 * bioops. Unused... we use the mount's VFS_SYNC instead. 1221 * 1222 * bioops callback - hold io_token 1223 */ 1224 static int 1225 hammer_io_sync(struct mount *mp) 1226 { 1227 /* nothing to do, so io_token not needed */ 1228 return(0); 1229 } 1230 1231 /* 1232 * bioops callback - hold io_token 1233 */ 1234 static void 1235 hammer_io_movedeps(struct buf *bp1, struct buf *bp2) 1236 { 1237 /* nothing to do, so io_token not needed */ 1238 } 1239 1240 /* 1241 * I/O pre-check for reading and writing. HAMMER only uses this for 1242 * B_CACHE buffers so checkread just shouldn't happen, but if it does 1243 * allow it. 1244 * 1245 * Writing is a different case. We don't want the kernel to try to write 1246 * out a buffer that HAMMER may be modifying passively or which has a 1247 * dependancy. In addition, kernel-demanded writes can only proceed for 1248 * certain types of buffers (i.e. UNDO and DATA types). Other dirty 1249 * buffer types can only be explicitly written by the flusher. 1250 * 1251 * checkwrite will only be called for bdwrite()n buffers. If we return 1252 * success the kernel is guaranteed to initiate the buffer write. 1253 * 1254 * bioops callback - hold io_token 1255 */ 1256 static int 1257 hammer_io_checkread(struct buf *bp) 1258 { 1259 /* nothing to do, so io_token not needed */ 1260 return(0); 1261 } 1262 1263 /* 1264 * The kernel is asking us whether it can write out a dirty buffer or not. 1265 * 1266 * bioops callback - hold io_token 1267 */ 1268 static int 1269 hammer_io_checkwrite(struct buf *bp) 1270 { 1271 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep); 1272 hammer_mount_t hmp = io->hmp; 1273 1274 /* 1275 * This shouldn't happen under normal operation. 1276 */ 1277 lwkt_gettoken(&hmp->io_token); 1278 if (io->type == HAMMER_STRUCTURE_VOLUME || 1279 io->type == HAMMER_STRUCTURE_META_BUFFER) { 1280 if (!panicstr) 1281 panic("hammer_io_checkwrite: illegal buffer"); 1282 if ((bp->b_flags & B_LOCKED) == 0) { 1283 bp->b_flags |= B_LOCKED; 1284 atomic_add_int(&hammer_count_io_locked, 1); 1285 } 1286 lwkt_reltoken(&hmp->io_token); 1287 return(1); 1288 } 1289 1290 /* 1291 * We have to be able to interlock the IO to safely modify any 1292 * of its fields without holding the fs_token. If we can't lock 1293 * it then we are racing someone. 1294 * 1295 * Our ownership of the bp lock prevents the io from being ripped 1296 * out from under us. 1297 */ 1298 if (hammer_try_interlock_norefs(&io->lock) == 0) { 1299 bp->b_flags |= B_LOCKED; 1300 atomic_add_int(&hammer_count_io_locked, 1); 1301 lwkt_reltoken(&hmp->io_token); 1302 return(1); 1303 } 1304 1305 /* 1306 * The modified bit must be cleared prior to the initiation of 1307 * any IO (returning 0 initiates the IO). Because this is a 1308 * normal data buffer hammer_io_clear_modify() runs through a 1309 * simple degenerate case. 1310 * 1311 * Return 0 will cause the kernel to initiate the IO, and we 1312 * must normally clear the modified bit before we begin. If 1313 * the io has modify_refs we do not clear the modified bit, 1314 * otherwise we may miss changes. 1315 * 1316 * Only data and undo buffers can reach here. These buffers do 1317 * not have terminal crc functions but we temporarily reference 1318 * the IO anyway, just in case. 1319 */ 1320 if (io->modify_refs == 0 && io->modified) { 1321 hammer_ref(&io->lock); 1322 hammer_io_clear_modify(io, 0); 1323 hammer_rel(&io->lock); 1324 } else if (io->modified) { 1325 KKASSERT(io->type == HAMMER_STRUCTURE_DATA_BUFFER); 1326 } 1327 1328 /* 1329 * The kernel is going to start the IO, set io->running. 1330 */ 1331 KKASSERT(io->running == 0); 1332 io->running = 1; 1333 atomic_add_int(&io->hmp->io_running_space, io->bytes); 1334 atomic_add_int(&hammer_count_io_running_write, io->bytes); 1335 TAILQ_INSERT_TAIL(&io->hmp->iorun_list, io, iorun_entry); 1336 1337 hammer_put_interlock(&io->lock, 1); 1338 lwkt_reltoken(&hmp->io_token); 1339 1340 return(0); 1341 } 1342 1343 /* 1344 * Return non-zero if we wish to delay the kernel's attempt to flush 1345 * this buffer to disk. 1346 * 1347 * bioops callback - hold io_token 1348 */ 1349 static int 1350 hammer_io_countdeps(struct buf *bp, int n) 1351 { 1352 /* nothing to do, so io_token not needed */ 1353 return(0); 1354 } 1355 1356 struct bio_ops hammer_bioops = { 1357 .io_start = hammer_io_start, 1358 .io_complete = hammer_io_complete, 1359 .io_deallocate = hammer_io_deallocate, 1360 .io_fsync = hammer_io_fsync, 1361 .io_sync = hammer_io_sync, 1362 .io_movedeps = hammer_io_movedeps, 1363 .io_countdeps = hammer_io_countdeps, 1364 .io_checkread = hammer_io_checkread, 1365 .io_checkwrite = hammer_io_checkwrite, 1366 }; 1367 1368 /************************************************************************ 1369 * DIRECT IO OPS * 1370 ************************************************************************ 1371 * 1372 * These functions operate directly on the buffer cache buffer associated 1373 * with a front-end vnode rather then a back-end device vnode. 1374 */ 1375 1376 /* 1377 * Read a buffer associated with a front-end vnode directly from the 1378 * disk media. The bio may be issued asynchronously. If leaf is non-NULL 1379 * we validate the CRC. 1380 * 1381 * We must check for the presence of a HAMMER buffer to handle the case 1382 * where the reblocker has rewritten the data (which it does via the HAMMER 1383 * buffer system, not via the high-level vnode buffer cache), but not yet 1384 * committed the buffer to the media. 1385 */ 1386 int 1387 hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio, 1388 hammer_btree_leaf_elm_t leaf) 1389 { 1390 hammer_off_t buf_offset; 1391 hammer_off_t zone2_offset; 1392 hammer_volume_t volume; 1393 struct buf *bp; 1394 struct bio *nbio; 1395 int vol_no; 1396 int error; 1397 1398 buf_offset = bio->bio_offset; 1399 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) == 1400 HAMMER_ZONE_LARGE_DATA); 1401 1402 /* 1403 * The buffer cache may have an aliased buffer (the reblocker can 1404 * write them). If it does we have to sync any dirty data before 1405 * we can build our direct-read. This is a non-critical code path. 1406 */ 1407 bp = bio->bio_buf; 1408 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize); 1409 1410 /* 1411 * Resolve to a zone-2 offset. The conversion just requires 1412 * munging the top 4 bits but we want to abstract it anyway 1413 * so the blockmap code can verify the zone assignment. 1414 */ 1415 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1416 if (error) 1417 goto done; 1418 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) == 1419 HAMMER_ZONE_RAW_BUFFER); 1420 1421 /* 1422 * Resolve volume and raw-offset for 3rd level bio. The 1423 * offset will be specific to the volume. 1424 */ 1425 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1426 volume = hammer_get_volume(hmp, vol_no, &error); 1427 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1428 error = EIO; 1429 1430 if (error == 0) { 1431 /* 1432 * 3rd level bio 1433 */ 1434 nbio = push_bio(bio); 1435 nbio->bio_offset = volume->ondisk->vol_buf_beg + 1436 (zone2_offset & HAMMER_OFF_SHORT_MASK); 1437 #if 0 1438 /* 1439 * XXX disabled - our CRC check doesn't work if the OS 1440 * does bogus_page replacement on the direct-read. 1441 */ 1442 if (leaf && hammer_verify_data) { 1443 nbio->bio_done = hammer_io_direct_read_complete; 1444 nbio->bio_caller_info1.uvalue32 = leaf->data_crc; 1445 } 1446 #endif 1447 hammer_stats_disk_read += bp->b_bufsize; 1448 vn_strategy(volume->devvp, nbio); 1449 } 1450 hammer_rel_volume(volume, 0); 1451 done: 1452 if (error) { 1453 kprintf("hammer_direct_read: failed @ %016llx\n", 1454 (long long)zone2_offset); 1455 bp->b_error = error; 1456 bp->b_flags |= B_ERROR; 1457 biodone(bio); 1458 } 1459 return(error); 1460 } 1461 1462 #if 0 1463 /* 1464 * On completion of the BIO this callback must check the data CRC 1465 * and chain to the previous bio. 1466 * 1467 * MPSAFE - since we do not modify and hammer_records we do not need 1468 * io_token. 1469 * 1470 * NOTE: MPSAFE callback 1471 */ 1472 static 1473 void 1474 hammer_io_direct_read_complete(struct bio *nbio) 1475 { 1476 struct bio *obio; 1477 struct buf *bp; 1478 u_int32_t rec_crc = nbio->bio_caller_info1.uvalue32; 1479 1480 bp = nbio->bio_buf; 1481 if (crc32(bp->b_data, bp->b_bufsize) != rec_crc) { 1482 kprintf("HAMMER: data_crc error @%016llx/%d\n", 1483 nbio->bio_offset, bp->b_bufsize); 1484 if (hammer_debug_critical) 1485 Debugger("data_crc on read"); 1486 bp->b_flags |= B_ERROR; 1487 bp->b_error = EIO; 1488 } 1489 obio = pop_bio(nbio); 1490 biodone(obio); 1491 } 1492 #endif 1493 1494 /* 1495 * Write a buffer associated with a front-end vnode directly to the 1496 * disk media. The bio may be issued asynchronously. 1497 * 1498 * The BIO is associated with the specified record and RECG_DIRECT_IO 1499 * is set. The recorded is added to its object. 1500 */ 1501 int 1502 hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio, 1503 hammer_record_t record) 1504 { 1505 hammer_btree_leaf_elm_t leaf = &record->leaf; 1506 hammer_off_t buf_offset; 1507 hammer_off_t zone2_offset; 1508 hammer_volume_t volume; 1509 hammer_buffer_t buffer; 1510 struct buf *bp; 1511 struct bio *nbio; 1512 char *ptr; 1513 int vol_no; 1514 int error; 1515 1516 buf_offset = leaf->data_offset; 1517 1518 KKASSERT(buf_offset > HAMMER_ZONE_BTREE); 1519 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE); 1520 1521 /* 1522 * Issue or execute the I/O. The new memory record must replace 1523 * the old one before the I/O completes, otherwise a reaquisition of 1524 * the buffer will load the old media data instead of the new. 1525 */ 1526 if ((buf_offset & HAMMER_BUFMASK) == 0 && 1527 leaf->data_len >= HAMMER_BUFSIZE) { 1528 /* 1529 * We are using the vnode's bio to write directly to the 1530 * media, any hammer_buffer at the same zone-X offset will 1531 * now have stale data. 1532 */ 1533 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1534 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1535 volume = hammer_get_volume(hmp, vol_no, &error); 1536 1537 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1538 error = EIO; 1539 if (error == 0) { 1540 bp = bio->bio_buf; 1541 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0); 1542 /* 1543 hammer_del_buffers(hmp, buf_offset, 1544 zone2_offset, bp->b_bufsize); 1545 */ 1546 1547 /* 1548 * Second level bio - cached zone2 offset. 1549 * 1550 * (We can put our bio_done function in either the 1551 * 2nd or 3rd level). 1552 */ 1553 nbio = push_bio(bio); 1554 nbio->bio_offset = zone2_offset; 1555 nbio->bio_done = hammer_io_direct_write_complete; 1556 nbio->bio_caller_info1.ptr = record; 1557 record->zone2_offset = zone2_offset; 1558 record->gflags |= HAMMER_RECG_DIRECT_IO | 1559 HAMMER_RECG_DIRECT_INVAL; 1560 1561 /* 1562 * Third level bio - raw offset specific to the 1563 * correct volume. 1564 */ 1565 zone2_offset &= HAMMER_OFF_SHORT_MASK; 1566 nbio = push_bio(nbio); 1567 nbio->bio_offset = volume->ondisk->vol_buf_beg + 1568 zone2_offset; 1569 hammer_stats_disk_write += bp->b_bufsize; 1570 hammer_ip_replace_bulk(hmp, record); 1571 vn_strategy(volume->devvp, nbio); 1572 hammer_io_flush_mark(volume); 1573 } 1574 hammer_rel_volume(volume, 0); 1575 } else { 1576 /* 1577 * Must fit in a standard HAMMER buffer. In this case all 1578 * consumers use the HAMMER buffer system and RECG_DIRECT_IO 1579 * does not need to be set-up. 1580 */ 1581 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0); 1582 buffer = NULL; 1583 ptr = hammer_bread(hmp, buf_offset, &error, &buffer); 1584 if (error == 0) { 1585 bp = bio->bio_buf; 1586 bp->b_flags |= B_AGE; 1587 hammer_io_modify(&buffer->io, 1); 1588 bcopy(bp->b_data, ptr, leaf->data_len); 1589 hammer_io_modify_done(&buffer->io); 1590 hammer_rel_buffer(buffer, 0); 1591 bp->b_resid = 0; 1592 hammer_ip_replace_bulk(hmp, record); 1593 biodone(bio); 1594 } 1595 } 1596 if (error) { 1597 /* 1598 * Major suckage occured. Also note: The record was 1599 * never added to the tree so we do not have to worry 1600 * about the backend. 1601 */ 1602 kprintf("hammer_direct_write: failed @ %016llx\n", 1603 (long long)leaf->data_offset); 1604 bp = bio->bio_buf; 1605 bp->b_resid = 0; 1606 bp->b_error = EIO; 1607 bp->b_flags |= B_ERROR; 1608 biodone(bio); 1609 record->flags |= HAMMER_RECF_DELETED_FE; 1610 hammer_rel_mem_record(record); 1611 } 1612 return(error); 1613 } 1614 1615 /* 1616 * On completion of the BIO this callback must disconnect 1617 * it from the hammer_record and chain to the previous bio. 1618 * 1619 * An I/O error forces the mount to read-only. Data buffers 1620 * are not B_LOCKED like meta-data buffers are, so we have to 1621 * throw the buffer away to prevent the kernel from retrying. 1622 * 1623 * NOTE: MPSAFE callback, only modify fields we have explicit 1624 * access to (the bp and the record->gflags). 1625 */ 1626 static 1627 void 1628 hammer_io_direct_write_complete(struct bio *nbio) 1629 { 1630 struct bio *obio; 1631 struct buf *bp; 1632 hammer_record_t record; 1633 hammer_mount_t hmp; 1634 1635 record = nbio->bio_caller_info1.ptr; 1636 KKASSERT(record != NULL); 1637 hmp = record->ip->hmp; 1638 1639 lwkt_gettoken(&hmp->io_token); 1640 1641 bp = nbio->bio_buf; 1642 obio = pop_bio(nbio); 1643 if (bp->b_flags & B_ERROR) { 1644 lwkt_gettoken(&hmp->fs_token); 1645 hammer_critical_error(hmp, record->ip, 1646 bp->b_error, 1647 "while writing bulk data"); 1648 lwkt_reltoken(&hmp->fs_token); 1649 bp->b_flags |= B_INVAL; 1650 } 1651 biodone(obio); 1652 1653 KKASSERT(record->gflags & HAMMER_RECG_DIRECT_IO); 1654 if (record->gflags & HAMMER_RECG_DIRECT_WAIT) { 1655 record->gflags &= ~(HAMMER_RECG_DIRECT_IO | 1656 HAMMER_RECG_DIRECT_WAIT); 1657 /* record can disappear once DIRECT_IO flag is cleared */ 1658 wakeup(&record->flags); 1659 } else { 1660 record->gflags &= ~HAMMER_RECG_DIRECT_IO; 1661 /* record can disappear once DIRECT_IO flag is cleared */ 1662 } 1663 lwkt_reltoken(&hmp->io_token); 1664 } 1665 1666 1667 /* 1668 * This is called before a record is either committed to the B-Tree 1669 * or destroyed, to resolve any associated direct-IO. 1670 * 1671 * (1) We must wait for any direct-IO related to the record to complete. 1672 * 1673 * (2) We must remove any buffer cache aliases for data accessed via 1674 * leaf->data_offset or zone2_offset so non-direct-IO consumers 1675 * (the mirroring and reblocking code) do not see stale data. 1676 */ 1677 void 1678 hammer_io_direct_wait(hammer_record_t record) 1679 { 1680 hammer_mount_t hmp = record->ip->hmp; 1681 1682 /* 1683 * Wait for I/O to complete 1684 */ 1685 if (record->gflags & HAMMER_RECG_DIRECT_IO) { 1686 lwkt_gettoken(&hmp->io_token); 1687 while (record->gflags & HAMMER_RECG_DIRECT_IO) { 1688 record->gflags |= HAMMER_RECG_DIRECT_WAIT; 1689 tsleep(&record->flags, 0, "hmdiow", 0); 1690 } 1691 lwkt_reltoken(&hmp->io_token); 1692 } 1693 1694 /* 1695 * Invalidate any related buffer cache aliases associated with the 1696 * backing device. This is needed because the buffer cache buffer 1697 * for file data is associated with the file vnode, not the backing 1698 * device vnode. 1699 * 1700 * XXX I do not think this case can occur any more now that 1701 * reservations ensure that all such buffers are removed before 1702 * an area can be reused. 1703 */ 1704 if (record->gflags & HAMMER_RECG_DIRECT_INVAL) { 1705 KKASSERT(record->leaf.data_offset); 1706 hammer_del_buffers(hmp, record->leaf.data_offset, 1707 record->zone2_offset, record->leaf.data_len, 1708 1); 1709 record->gflags &= ~HAMMER_RECG_DIRECT_INVAL; 1710 } 1711 } 1712 1713 /* 1714 * This is called to remove the second-level cached zone-2 offset from 1715 * frontend buffer cache buffers, now stale due to a data relocation. 1716 * These offsets are generated by cluster_read() via VOP_BMAP, or directly 1717 * by hammer_vop_strategy_read(). 1718 * 1719 * This is rather nasty because here we have something like the reblocker 1720 * scanning the raw B-Tree with no held references on anything, really, 1721 * other then a shared lock on the B-Tree node, and we have to access the 1722 * frontend's buffer cache to check for and clean out the association. 1723 * Specifically, if the reblocker is moving data on the disk, these cached 1724 * offsets will become invalid. 1725 * 1726 * Only data record types associated with the large-data zone are subject 1727 * to direct-io and need to be checked. 1728 * 1729 */ 1730 void 1731 hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf) 1732 { 1733 struct hammer_inode_info iinfo; 1734 int zone; 1735 1736 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA) 1737 return; 1738 zone = HAMMER_ZONE_DECODE(leaf->data_offset); 1739 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX) 1740 return; 1741 iinfo.obj_id = leaf->base.obj_id; 1742 iinfo.obj_asof = 0; /* unused */ 1743 iinfo.obj_localization = leaf->base.localization & 1744 HAMMER_LOCALIZE_PSEUDOFS_MASK; 1745 iinfo.u.leaf = leaf; 1746 hammer_scan_inode_snapshots(hmp, &iinfo, 1747 hammer_io_direct_uncache_callback, 1748 leaf); 1749 } 1750 1751 static int 1752 hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data) 1753 { 1754 hammer_inode_info_t iinfo = data; 1755 hammer_off_t data_offset; 1756 hammer_off_t file_offset; 1757 struct vnode *vp; 1758 struct buf *bp; 1759 int blksize; 1760 1761 if (ip->vp == NULL) 1762 return(0); 1763 data_offset = iinfo->u.leaf->data_offset; 1764 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len; 1765 blksize = iinfo->u.leaf->data_len; 1766 KKASSERT((blksize & HAMMER_BUFMASK) == 0); 1767 1768 /* 1769 * Warning: FINDBLK_TEST return stable storage but not stable 1770 * contents. It happens to be ok in this case. 1771 */ 1772 hammer_ref(&ip->lock); 1773 if (hammer_get_vnode(ip, &vp) == 0) { 1774 if ((bp = findblk(ip->vp, file_offset, FINDBLK_TEST)) != NULL && 1775 bp->b_bio2.bio_offset != NOOFFSET) { 1776 bp = getblk(ip->vp, file_offset, blksize, 0, 0); 1777 bp->b_bio2.bio_offset = NOOFFSET; 1778 brelse(bp); 1779 } 1780 vput(vp); 1781 } 1782 hammer_rel_inode(ip, 0); 1783 return(0); 1784 } 1785 1786 1787 /* 1788 * This function is called when writes may have occured on the volume, 1789 * indicating that the device may be holding cached writes. 1790 */ 1791 static void 1792 hammer_io_flush_mark(hammer_volume_t volume) 1793 { 1794 atomic_set_int(&volume->vol_flags, HAMMER_VOLF_NEEDFLUSH); 1795 } 1796 1797 /* 1798 * This function ensures that the device has flushed any cached writes out. 1799 */ 1800 void 1801 hammer_io_flush_sync(hammer_mount_t hmp) 1802 { 1803 hammer_volume_t volume; 1804 struct buf *bp_base = NULL; 1805 struct buf *bp; 1806 1807 RB_FOREACH(volume, hammer_vol_rb_tree, &hmp->rb_vols_root) { 1808 if (volume->vol_flags & HAMMER_VOLF_NEEDFLUSH) { 1809 atomic_clear_int(&volume->vol_flags, 1810 HAMMER_VOLF_NEEDFLUSH); 1811 bp = getpbuf(NULL); 1812 bp->b_bio1.bio_offset = 0; 1813 bp->b_bufsize = 0; 1814 bp->b_bcount = 0; 1815 bp->b_cmd = BUF_CMD_FLUSH; 1816 bp->b_bio1.bio_caller_info1.cluster_head = bp_base; 1817 bp->b_bio1.bio_done = biodone_sync; 1818 bp->b_bio1.bio_flags |= BIO_SYNC; 1819 bp_base = bp; 1820 vn_strategy(volume->devvp, &bp->b_bio1); 1821 } 1822 } 1823 while ((bp = bp_base) != NULL) { 1824 bp_base = bp->b_bio1.bio_caller_info1.cluster_head; 1825 biowait(&bp->b_bio1, "hmrFLS"); 1826 relpbuf(bp, NULL); 1827 } 1828 } 1829 1830 /* 1831 * Limit the amount of backlog which we allow to build up 1832 */ 1833 void 1834 hammer_io_limit_backlog(hammer_mount_t hmp) 1835 { 1836 while (hmp->io_running_space > hammer_limit_running_io) { 1837 hmp->io_running_wakeup = 1; 1838 tsleep(&hmp->io_running_wakeup, 0, "hmiolm", hz / 10); 1839 } 1840 } 1841