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 /* 35 * IO Primitives and buffer cache management 36 * 37 * All major data-tracking structures in HAMMER contain a struct hammer_io 38 * which is used to manage their backing store. We use filesystem buffers 39 * for backing store and we leave them passively associated with their 40 * HAMMER structures. 41 * 42 * If the kernel tries to destroy a passively associated buf which we cannot 43 * yet let go we set B_LOCKED in the buffer and then actively released it 44 * later when we can. 45 * 46 * The io_token is required for anything which might race bioops and bio_done 47 * callbacks, with one exception: A successful hammer_try_interlock_norefs(). 48 * the fs_token will be held in all other cases. 49 */ 50 51 #include "hammer.h" 52 #include <sys/fcntl.h> 53 #include <sys/nlookup.h> 54 #include <sys/buf.h> 55 56 #include <sys/buf2.h> 57 58 static void hammer_io_modify(hammer_io_t io, int count); 59 static void hammer_io_deallocate(struct buf *bp); 60 static void hammer_indirect_callback(struct bio *bio); 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 static int 70 hammer_mod_rb_compare(hammer_io_t io1, hammer_io_t io2) 71 { 72 hammer_off_t io1_offset; 73 hammer_off_t io2_offset; 74 75 io1_offset = ((io1->offset & HAMMER_OFF_SHORT_MASK) << 8) | 76 HAMMER_VOL_DECODE(io1->offset); 77 io2_offset = ((io2->offset & HAMMER_OFF_SHORT_MASK) << 8) | 78 HAMMER_VOL_DECODE(io2->offset); 79 80 if (io1_offset < io2_offset) 81 return(-1); 82 if (io1_offset > io2_offset) 83 return(1); 84 return(0); 85 } 86 87 RB_GENERATE(hammer_mod_rb_tree, hammer_io, rb_node, hammer_mod_rb_compare); 88 89 /* 90 * Initialize a new, already-zero'd hammer_io structure, or reinitialize 91 * an existing hammer_io structure which may have switched to another type. 92 */ 93 void 94 hammer_io_init(hammer_io_t io, hammer_volume_t volume, enum hammer_io_type type) 95 { 96 io->volume = volume; 97 io->hmp = volume->io.hmp; 98 io->type = type; 99 } 100 101 /* 102 * Helper routine to disassociate a buffer cache buffer from an I/O 103 * structure. The io must be interlocked and marked appropriately for 104 * reclamation. 105 * 106 * The io must be in a released state with the io->bp owned and 107 * locked by the caller of this function. When not called from an 108 * io_deallocate() this cannot race an io_deallocate() since the 109 * kernel would be unable to get the buffer lock in that case. 110 * (The released state in this case means we own the bp, not the 111 * hammer_io structure). 112 * 113 * The io may have 0 or 1 references depending on who called us. The 114 * caller is responsible for dealing with the refs. 115 * 116 * This call can only be made when no action is required on the buffer. 117 * 118 * This function is guaranteed not to race against anything because we 119 * own both the io lock and the bp lock and are interlocked with no 120 * references. 121 */ 122 static void 123 hammer_io_disassociate(hammer_io_structure_t iou) 124 { 125 struct buf *bp = iou->io.bp; 126 127 KKASSERT(iou->io.released); 128 KKASSERT(iou->io.modified == 0); 129 KKASSERT(LIST_FIRST(&bp->b_dep) == (void *)iou); 130 buf_dep_init(bp); 131 iou->io.bp = NULL; 132 133 /* 134 * If the buffer was locked someone wanted to get rid of it. 135 */ 136 if (bp->b_flags & B_LOCKED) { 137 atomic_add_int(&hammer_count_io_locked, -1); 138 bp->b_flags &= ~B_LOCKED; 139 } 140 if (iou->io.reclaim) { 141 bp->b_flags |= B_NOCACHE|B_RELBUF; 142 iou->io.reclaim = 0; 143 } 144 145 switch(iou->io.type) { 146 case HAMMER_STRUCTURE_VOLUME: 147 iou->volume.ondisk = NULL; 148 break; 149 case HAMMER_STRUCTURE_DATA_BUFFER: 150 case HAMMER_STRUCTURE_META_BUFFER: 151 case HAMMER_STRUCTURE_UNDO_BUFFER: 152 iou->buffer.ondisk = NULL; 153 break; 154 case HAMMER_STRUCTURE_DUMMY: 155 panic("hammer_io_disassociate: bad io type"); 156 break; 157 } 158 } 159 160 /* 161 * Wait for any physical IO to complete 162 * 163 * XXX we aren't interlocked against a spinlock or anything so there 164 * is a small window in the interlock / io->running == 0 test. 165 */ 166 void 167 hammer_io_wait(hammer_io_t io) 168 { 169 if (io->running) { 170 hammer_mount_t hmp = io->hmp; 171 172 lwkt_gettoken(&hmp->io_token); 173 while (io->running) { 174 io->waiting = 1; 175 tsleep_interlock(io, 0); 176 if (io->running) 177 tsleep(io, PINTERLOCKED, "hmrflw", hz); 178 } 179 lwkt_reltoken(&hmp->io_token); 180 } 181 } 182 183 /* 184 * Wait for all currently queued HAMMER-initiated I/Os to complete. 185 * 186 * This is not supposed to count direct I/O's but some can leak 187 * through (for non-full-sized direct I/Os). 188 */ 189 void 190 hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush) 191 { 192 struct hammer_io iodummy; 193 hammer_io_t io; 194 195 /* 196 * Degenerate case, no I/O is running 197 */ 198 lwkt_gettoken(&hmp->io_token); 199 if (TAILQ_EMPTY(&hmp->iorun_list)) { 200 lwkt_reltoken(&hmp->io_token); 201 if (doflush) 202 hammer_io_flush_sync(hmp); 203 return; 204 } 205 bzero(&iodummy, sizeof(iodummy)); 206 iodummy.type = HAMMER_STRUCTURE_DUMMY; 207 208 /* 209 * Add placemarker and then wait until it becomes the head of 210 * the list. 211 */ 212 TAILQ_INSERT_TAIL(&hmp->iorun_list, &iodummy, iorun_entry); 213 while (TAILQ_FIRST(&hmp->iorun_list) != &iodummy) { 214 tsleep(&iodummy, 0, ident, 0); 215 } 216 217 /* 218 * Chain in case several placemarkers are present. 219 */ 220 TAILQ_REMOVE(&hmp->iorun_list, &iodummy, iorun_entry); 221 io = TAILQ_FIRST(&hmp->iorun_list); 222 if (io && io->type == HAMMER_STRUCTURE_DUMMY) 223 wakeup(io); 224 lwkt_reltoken(&hmp->io_token); 225 226 if (doflush) 227 hammer_io_flush_sync(hmp); 228 } 229 230 /* 231 * Clear a flagged error condition on a I/O buffer. The caller must hold 232 * its own ref on the buffer. 233 */ 234 void 235 hammer_io_clear_error(struct hammer_io *io) 236 { 237 hammer_mount_t hmp = io->hmp; 238 239 lwkt_gettoken(&hmp->io_token); 240 if (io->ioerror) { 241 io->ioerror = 0; 242 hammer_rel(&io->lock); 243 KKASSERT(hammer_isactive(&io->lock)); 244 } 245 lwkt_reltoken(&hmp->io_token); 246 } 247 248 void 249 hammer_io_clear_error_noassert(struct hammer_io *io) 250 { 251 hammer_mount_t hmp = io->hmp; 252 253 lwkt_gettoken(&hmp->io_token); 254 if (io->ioerror) { 255 io->ioerror = 0; 256 hammer_rel(&io->lock); 257 } 258 lwkt_reltoken(&hmp->io_token); 259 } 260 261 /* 262 * This is an advisory function only which tells the buffer cache 263 * the bp is not a meta-data buffer, even though it is backed by 264 * a block device. 265 * 266 * This is used by HAMMER's reblocking code to avoid trying to 267 * swapcache the filesystem's data when it is read or written 268 * by the reblocking code. 269 * 270 * The caller has a ref on the buffer preventing the bp from 271 * being disassociated from it. 272 */ 273 void 274 hammer_io_notmeta(hammer_buffer_t buffer) 275 { 276 if ((buffer->io.bp->b_flags & B_NOTMETA) == 0) { 277 hammer_mount_t hmp = buffer->io.hmp; 278 279 lwkt_gettoken(&hmp->io_token); 280 buffer->io.bp->b_flags |= B_NOTMETA; 281 lwkt_reltoken(&hmp->io_token); 282 } 283 } 284 285 /* 286 * Load bp for a HAMMER structure. The io must be exclusively locked by 287 * the caller. 288 * 289 * This routine is mostly used on meta-data and small-data blocks. Generally 290 * speaking HAMMER assumes some locality of reference and will cluster. 291 * 292 * Note that the caller (hammer_ondisk.c) may place further restrictions 293 * on clusterability via the limit (in bytes). Typically large-data 294 * zones cannot be clustered due to their mixed buffer sizes. This is 295 * not an issue since such clustering occurs in hammer_vnops at the 296 * regular file layer, whereas this is the buffered block device layer. 297 * 298 * No I/O callbacks can occur while we hold the buffer locked. 299 */ 300 int 301 hammer_io_read(struct vnode *devvp, struct hammer_io *io, int limit) 302 { 303 struct buf *bp; 304 int error; 305 306 if ((bp = io->bp) == NULL) { 307 atomic_add_long(&hammer_count_io_running_read, io->bytes); 308 if (hammer_cluster_enable && limit > io->bytes) { 309 error = cluster_read(devvp, io->offset + limit, 310 io->offset, io->bytes, 311 HAMMER_CLUSTER_SIZE, 312 HAMMER_CLUSTER_SIZE, 313 &io->bp); 314 } else { 315 error = bread(devvp, io->offset, io->bytes, &io->bp); 316 } 317 hammer_stats_disk_read += io->bytes; 318 atomic_add_long(&hammer_count_io_running_read, -io->bytes); 319 320 /* 321 * The code generally assumes b_ops/b_dep has been set-up, 322 * even if we error out here. 323 */ 324 bp = io->bp; 325 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) { 326 const char *metatype; 327 328 switch(io->type) { 329 case HAMMER_STRUCTURE_VOLUME: 330 metatype = "volume"; 331 break; 332 case HAMMER_STRUCTURE_META_BUFFER: 333 switch(((struct hammer_buffer *)io)-> 334 zoneX_offset & HAMMER_OFF_ZONE_MASK) { 335 case HAMMER_ZONE_BTREE: 336 metatype = "btree"; 337 break; 338 case HAMMER_ZONE_META: 339 metatype = "meta"; 340 break; 341 case HAMMER_ZONE_FREEMAP: 342 metatype = "freemap"; 343 break; 344 default: 345 metatype = "meta?"; 346 break; 347 } 348 break; 349 case HAMMER_STRUCTURE_DATA_BUFFER: 350 metatype = "data"; 351 break; 352 case HAMMER_STRUCTURE_UNDO_BUFFER: 353 metatype = "undo"; 354 break; 355 default: 356 metatype = "unknown"; 357 break; 358 } 359 kprintf("doff %016jx %s\n", 360 (intmax_t)bp->b_bio2.bio_offset, 361 metatype); 362 } 363 bp->b_flags &= ~B_IODEBUG; 364 bp->b_ops = &hammer_bioops; 365 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 366 367 /* io->worklist is locked by the io lock */ 368 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node); 369 BUF_KERNPROC(bp); 370 KKASSERT(io->modified == 0); 371 KKASSERT(io->running == 0); 372 KKASSERT(io->waiting == 0); 373 io->released = 0; /* we hold an active lock on bp */ 374 } else { 375 error = 0; 376 } 377 return(error); 378 } 379 380 /* 381 * Similar to hammer_io_read() but returns a zero'd out buffer instead. 382 * Must be called with the IO exclusively locked. 383 * 384 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background 385 * I/O by forcing the buffer to not be in a released state before calling 386 * it. 387 * 388 * This function will also mark the IO as modified but it will not 389 * increment the modify_refs count. 390 * 391 * No I/O callbacks can occur while we hold the buffer locked. 392 */ 393 int 394 hammer_io_new(struct vnode *devvp, struct hammer_io *io) 395 { 396 struct buf *bp; 397 398 if ((bp = io->bp) == NULL) { 399 io->bp = getblk(devvp, io->offset, io->bytes, 0, 0); 400 bp = io->bp; 401 bp->b_ops = &hammer_bioops; 402 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 403 404 /* io->worklist is locked by the io lock */ 405 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node); 406 io->released = 0; 407 KKASSERT(io->running == 0); 408 io->waiting = 0; 409 BUF_KERNPROC(bp); 410 } else { 411 if (io->released) { 412 regetblk(bp); 413 BUF_KERNPROC(bp); 414 io->released = 0; 415 } 416 } 417 hammer_io_modify(io, 0); 418 vfs_bio_clrbuf(bp); 419 return(0); 420 } 421 422 /* 423 * Advance the activity count on the underlying buffer because 424 * HAMMER does not getblk/brelse on every access. 425 * 426 * The io->bp cannot go away while the buffer is referenced. 427 */ 428 void 429 hammer_io_advance(struct hammer_io *io) 430 { 431 if (io->bp) 432 buf_act_advance(io->bp); 433 } 434 435 /* 436 * Remove potential device level aliases against buffers managed by high level 437 * vnodes. Aliases can also be created due to mixed buffer sizes or via 438 * direct access to the backing store device. 439 * 440 * This is nasty because the buffers are also VMIO-backed. Even if a buffer 441 * does not exist its backing VM pages might, and we have to invalidate 442 * those as well or a getblk() will reinstate them. 443 * 444 * Buffer cache buffers associated with hammer_buffers cannot be 445 * invalidated. 446 */ 447 int 448 hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset) 449 { 450 hammer_io_structure_t iou; 451 hammer_mount_t hmp; 452 hammer_off_t phys_offset; 453 struct buf *bp; 454 int error; 455 456 hmp = volume->io.hmp; 457 lwkt_gettoken(&hmp->io_token); 458 459 /* 460 * If a device buffer already exists for the specified physical 461 * offset use that, otherwise instantiate a buffer to cover any 462 * related VM pages, set BNOCACHE, and brelse(). 463 */ 464 phys_offset = volume->ondisk->vol_buf_beg + 465 (zone2_offset & HAMMER_OFF_SHORT_MASK); 466 if ((bp = findblk(volume->devvp, phys_offset, 0)) != NULL) 467 bremfree(bp); 468 else 469 bp = getblk(volume->devvp, phys_offset, HAMMER_BUFSIZE, 0, 0); 470 471 if ((iou = (void *)LIST_FIRST(&bp->b_dep)) != NULL) { 472 #if 0 473 hammer_ref(&iou->io.lock); 474 hammer_io_clear_modify(&iou->io, 1); 475 bundirty(bp); 476 iou->io.released = 0; 477 BUF_KERNPROC(bp); 478 iou->io.reclaim = 1; 479 iou->io.waitdep = 1; /* XXX this is a fs_token field */ 480 KKASSERT(hammer_isactive(&iou->io.lock) == 1); 481 hammer_rel_buffer(&iou->buffer, 0); 482 /*hammer_io_deallocate(bp);*/ 483 #endif 484 bqrelse(bp); 485 error = EAGAIN; 486 } else { 487 KKASSERT((bp->b_flags & B_LOCKED) == 0); 488 bundirty(bp); 489 bp->b_flags |= B_NOCACHE|B_RELBUF; 490 brelse(bp); 491 error = 0; 492 } 493 lwkt_reltoken(&hmp->io_token); 494 return(error); 495 } 496 497 /* 498 * This routine is called on the last reference to a hammer structure. 499 * The io must be interlocked with a refcount of zero. The hammer structure 500 * will remain interlocked on return. 501 * 502 * This routine may return a non-NULL bp to the caller for dispoal. 503 * The caller typically brelse()'s the bp. 504 * 505 * The bp may or may not still be passively associated with the IO. It 506 * will remain passively associated if it is unreleasable (e.g. a modified 507 * meta-data buffer). 508 * 509 * The only requirement here is that modified meta-data and volume-header 510 * buffer may NOT be disassociated from the IO structure, and consequently 511 * we also leave such buffers actively associated with the IO if they already 512 * are (since the kernel can't do anything with them anyway). Only the 513 * flusher is allowed to write such buffers out. Modified pure-data and 514 * undo buffers are returned to the kernel but left passively associated 515 * so we can track when the kernel writes the bp out. 516 */ 517 struct buf * 518 hammer_io_release(struct hammer_io *io, int flush) 519 { 520 union hammer_io_structure *iou = (void *)io; 521 struct buf *bp; 522 523 if ((bp = io->bp) == NULL) 524 return(NULL); 525 526 /* 527 * Try to flush a dirty IO to disk if asked to by the 528 * caller or if the kernel tried to flush the buffer in the past. 529 * 530 * Kernel-initiated flushes are only allowed for pure-data buffers. 531 * meta-data and volume buffers can only be flushed explicitly 532 * by HAMMER. 533 */ 534 if (io->modified) { 535 if (flush) { 536 hammer_io_flush(io, 0); 537 } else if (bp->b_flags & B_LOCKED) { 538 switch(io->type) { 539 case HAMMER_STRUCTURE_DATA_BUFFER: 540 hammer_io_flush(io, 0); 541 break; 542 case HAMMER_STRUCTURE_UNDO_BUFFER: 543 hammer_io_flush(io, hammer_undo_reclaim(io)); 544 break; 545 default: 546 break; 547 } 548 } /* else no explicit request to flush the buffer */ 549 } 550 551 /* 552 * Wait for the IO to complete if asked to. This occurs when 553 * the buffer must be disposed of definitively during an umount 554 * or buffer invalidation. 555 */ 556 if (io->waitdep && io->running) { 557 hammer_io_wait(io); 558 } 559 560 /* 561 * Return control of the buffer to the kernel (with the provisio 562 * that our bioops can override kernel decisions with regards to 563 * the buffer). 564 */ 565 if ((flush || io->reclaim) && io->modified == 0 && io->running == 0) { 566 /* 567 * Always disassociate the bp if an explicit flush 568 * was requested and the IO completed with no error 569 * (so unmount can really clean up the structure). 570 */ 571 if (io->released) { 572 regetblk(bp); 573 BUF_KERNPROC(bp); 574 } else { 575 io->released = 1; 576 } 577 hammer_io_disassociate((hammer_io_structure_t)io); 578 /* return the bp */ 579 } else if (io->modified) { 580 /* 581 * Only certain IO types can be released to the kernel if 582 * the buffer has been modified. 583 * 584 * volume and meta-data IO types may only be explicitly 585 * flushed by HAMMER. 586 */ 587 switch(io->type) { 588 case HAMMER_STRUCTURE_DATA_BUFFER: 589 case HAMMER_STRUCTURE_UNDO_BUFFER: 590 if (io->released == 0) { 591 io->released = 1; 592 bdwrite(bp); 593 } 594 break; 595 default: 596 break; 597 } 598 bp = NULL; /* bp left associated */ 599 } else if (io->released == 0) { 600 /* 601 * Clean buffers can be generally released to the kernel. 602 * We leave the bp passively associated with the HAMMER 603 * structure and use bioops to disconnect it later on 604 * if the kernel wants to discard the buffer. 605 * 606 * We can steal the structure's ownership of the bp. 607 */ 608 io->released = 1; 609 if (bp->b_flags & B_LOCKED) { 610 hammer_io_disassociate(iou); 611 /* return the bp */ 612 } else { 613 if (io->reclaim) { 614 hammer_io_disassociate(iou); 615 /* return the bp */ 616 } else { 617 /* return the bp (bp passively associated) */ 618 } 619 } 620 } else { 621 /* 622 * A released buffer is passively associate with our 623 * hammer_io structure. The kernel cannot destroy it 624 * without making a bioops call. If the kernel (B_LOCKED) 625 * or we (reclaim) requested that the buffer be destroyed 626 * we destroy it, otherwise we do a quick get/release to 627 * reset its position in the kernel's LRU list. 628 * 629 * Leaving the buffer passively associated allows us to 630 * use the kernel's LRU buffer flushing mechanisms rather 631 * then rolling our own. 632 * 633 * XXX there are two ways of doing this. We can re-acquire 634 * and passively release to reset the LRU, or not. 635 */ 636 if (io->running == 0) { 637 regetblk(bp); 638 if ((bp->b_flags & B_LOCKED) || io->reclaim) { 639 hammer_io_disassociate(iou); 640 /* return the bp */ 641 } else { 642 /* return the bp (bp passively associated) */ 643 } 644 } else { 645 /* 646 * bp is left passively associated but we do not 647 * try to reacquire it. Interactions with the io 648 * structure will occur on completion of the bp's 649 * I/O. 650 */ 651 bp = NULL; 652 } 653 } 654 return(bp); 655 } 656 657 /* 658 * This routine is called with a locked IO when a flush is desired and 659 * no other references to the structure exists other then ours. This 660 * routine is ONLY called when HAMMER believes it is safe to flush a 661 * potentially modified buffer out. 662 * 663 * The locked io or io reference prevents a flush from being initiated 664 * by the kernel. 665 */ 666 void 667 hammer_io_flush(struct hammer_io *io, int reclaim) 668 { 669 struct buf *bp; 670 hammer_mount_t hmp; 671 672 /* 673 * Degenerate case - nothing to flush if nothing is dirty. 674 */ 675 if (io->modified == 0) 676 return; 677 678 KKASSERT(io->bp); 679 KKASSERT(io->modify_refs <= 0); 680 681 /* 682 * Acquire ownership of the bp, particularly before we clear our 683 * modified flag. 684 * 685 * We are going to bawrite() this bp. Don't leave a window where 686 * io->released is set, we actually own the bp rather then our 687 * buffer. 688 * 689 * The io_token should not be required here as only 690 */ 691 hmp = io->hmp; 692 bp = io->bp; 693 if (io->released) { 694 regetblk(bp); 695 /* BUF_KERNPROC(io->bp); */ 696 /* io->released = 0; */ 697 KKASSERT(io->released); 698 KKASSERT(io->bp == bp); 699 } else { 700 io->released = 1; 701 } 702 703 if (reclaim) { 704 io->reclaim = 1; 705 if ((bp->b_flags & B_LOCKED) == 0) { 706 bp->b_flags |= B_LOCKED; 707 atomic_add_int(&hammer_count_io_locked, 1); 708 } 709 } 710 711 /* 712 * Acquire exclusive access to the bp and then clear the modified 713 * state of the buffer prior to issuing I/O to interlock any 714 * modifications made while the I/O is in progress. This shouldn't 715 * happen anyway but losing data would be worse. The modified bit 716 * will be rechecked after the IO completes. 717 * 718 * NOTE: This call also finalizes the buffer's content (inval == 0). 719 * 720 * This is only legal when lock.refs == 1 (otherwise we might clear 721 * the modified bit while there are still users of the cluster 722 * modifying the data). 723 * 724 * Do this before potentially blocking so any attempt to modify the 725 * ondisk while we are blocked blocks waiting for us. 726 */ 727 hammer_ref(&io->lock); 728 hammer_io_clear_modify(io, 0); 729 hammer_rel(&io->lock); 730 731 if (hammer_debug_io & 0x0002) 732 kprintf("hammer io_write %016jx\n", bp->b_bio1.bio_offset); 733 734 /* 735 * Transfer ownership to the kernel and initiate I/O. 736 * 737 * NOTE: We do not hold io_token so an atomic op is required to 738 * update io_running_space. 739 */ 740 io->running = 1; 741 atomic_add_long(&hmp->io_running_space, io->bytes); 742 atomic_add_long(&hammer_count_io_running_write, io->bytes); 743 lwkt_gettoken(&hmp->io_token); 744 TAILQ_INSERT_TAIL(&hmp->iorun_list, io, iorun_entry); 745 lwkt_reltoken(&hmp->io_token); 746 bawrite(bp); 747 hammer_io_flush_mark(io->volume); 748 } 749 750 /************************************************************************ 751 * BUFFER DIRTYING * 752 ************************************************************************ 753 * 754 * These routines deal with dependancies created when IO buffers get 755 * modified. The caller must call hammer_modify_*() on a referenced 756 * HAMMER structure prior to modifying its on-disk data. 757 * 758 * Any intent to modify an IO buffer acquires the related bp and imposes 759 * various write ordering dependancies. 760 */ 761 762 /* 763 * Mark a HAMMER structure as undergoing modification. Meta-data buffers 764 * are locked until the flusher can deal with them, pure data buffers 765 * can be written out. 766 * 767 * The referenced io prevents races. 768 */ 769 static 770 void 771 hammer_io_modify(hammer_io_t io, int count) 772 { 773 /* 774 * io->modify_refs must be >= 0 775 */ 776 while (io->modify_refs < 0) { 777 io->waitmod = 1; 778 tsleep(io, 0, "hmrmod", 0); 779 } 780 781 /* 782 * Shortcut if nothing to do. 783 */ 784 KKASSERT(hammer_isactive(&io->lock) && io->bp != NULL); 785 io->modify_refs += count; 786 if (io->modified && io->released == 0) 787 return; 788 789 /* 790 * NOTE: It is important not to set the modified bit 791 * until after we have acquired the bp or we risk 792 * racing against checkwrite. 793 */ 794 hammer_lock_ex(&io->lock); 795 if (io->released) { 796 regetblk(io->bp); 797 BUF_KERNPROC(io->bp); 798 io->released = 0; 799 } 800 if (io->modified == 0) { 801 hammer_io_set_modlist(io); 802 io->modified = 1; 803 } 804 hammer_unlock(&io->lock); 805 } 806 807 static __inline 808 void 809 hammer_io_modify_done(hammer_io_t io) 810 { 811 KKASSERT(io->modify_refs > 0); 812 --io->modify_refs; 813 if (io->modify_refs == 0 && io->waitmod) { 814 io->waitmod = 0; 815 wakeup(io); 816 } 817 } 818 819 /* 820 * The write interlock blocks other threads trying to modify a buffer 821 * (they block in hammer_io_modify()) after us, or blocks us while other 822 * threads are in the middle of modifying a buffer. 823 * 824 * The caller also has a ref on the io, however if we are not careful 825 * we will race bioops callbacks (checkwrite). To deal with this 826 * we must at least acquire and release the io_token, and it is probably 827 * better to hold it through the setting of modify_refs. 828 */ 829 void 830 hammer_io_write_interlock(hammer_io_t io) 831 { 832 hammer_mount_t hmp = io->hmp; 833 834 lwkt_gettoken(&hmp->io_token); 835 while (io->modify_refs != 0) { 836 io->waitmod = 1; 837 tsleep(io, 0, "hmrmod", 0); 838 } 839 io->modify_refs = -1; 840 lwkt_reltoken(&hmp->io_token); 841 } 842 843 void 844 hammer_io_done_interlock(hammer_io_t io) 845 { 846 KKASSERT(io->modify_refs == -1); 847 io->modify_refs = 0; 848 if (io->waitmod) { 849 io->waitmod = 0; 850 wakeup(io); 851 } 852 } 853 854 /* 855 * Caller intends to modify a volume's ondisk structure. 856 * 857 * This is only allowed if we are the flusher or we have a ref on the 858 * sync_lock. 859 */ 860 void 861 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume, 862 void *base, int len) 863 { 864 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 865 866 hammer_io_modify(&volume->io, 1); 867 if (len) { 868 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk; 869 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 870 hammer_generate_undo(trans, 871 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset), 872 base, len); 873 } 874 } 875 876 /* 877 * Caller intends to modify a buffer's ondisk structure. 878 * 879 * This is only allowed if we are the flusher or we have a ref on the 880 * sync_lock. 881 */ 882 void 883 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer, 884 void *base, int len) 885 { 886 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 887 888 hammer_io_modify(&buffer->io, 1); 889 if (len) { 890 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk; 891 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 892 hammer_generate_undo(trans, 893 buffer->zone2_offset + rel_offset, 894 base, len); 895 } 896 } 897 898 void 899 hammer_modify_volume_done(hammer_volume_t volume) 900 { 901 hammer_io_modify_done(&volume->io); 902 } 903 904 void 905 hammer_modify_buffer_done(hammer_buffer_t buffer) 906 { 907 hammer_io_modify_done(&buffer->io); 908 } 909 910 /* 911 * Mark an entity as not being dirty any more and finalize any 912 * delayed adjustments to the buffer. 913 * 914 * Delayed adjustments are an important performance enhancement, allowing 915 * us to avoid recalculating B-Tree node CRCs over and over again when 916 * making bulk-modifications to the B-Tree. 917 * 918 * If inval is non-zero delayed adjustments are ignored. 919 * 920 * This routine may dereference related btree nodes and cause the 921 * buffer to be dereferenced. The caller must own a reference on io. 922 */ 923 void 924 hammer_io_clear_modify(struct hammer_io *io, int inval) 925 { 926 hammer_mount_t hmp; 927 928 /* 929 * io_token is needed to avoid races on mod_root 930 */ 931 if (io->modified == 0) 932 return; 933 hmp = io->hmp; 934 lwkt_gettoken(&hmp->io_token); 935 if (io->modified == 0) { 936 lwkt_reltoken(&hmp->io_token); 937 return; 938 } 939 940 /* 941 * Take us off the mod-list and clear the modified bit. 942 */ 943 KKASSERT(io->mod_root != NULL); 944 if (io->mod_root == &io->hmp->volu_root || 945 io->mod_root == &io->hmp->meta_root) { 946 io->hmp->locked_dirty_space -= io->bytes; 947 atomic_add_long(&hammer_count_dirtybufspace, -io->bytes); 948 } 949 RB_REMOVE(hammer_mod_rb_tree, io->mod_root, io); 950 io->mod_root = NULL; 951 io->modified = 0; 952 953 lwkt_reltoken(&hmp->io_token); 954 955 /* 956 * If this bit is not set there are no delayed adjustments. 957 */ 958 if (io->gencrc == 0) 959 return; 960 io->gencrc = 0; 961 962 /* 963 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference 964 * on the node (& underlying buffer). Release the node after clearing 965 * the flag. 966 */ 967 if (io->type == HAMMER_STRUCTURE_META_BUFFER) { 968 hammer_buffer_t buffer = (void *)io; 969 hammer_node_t node; 970 971 restart: 972 TAILQ_FOREACH(node, &buffer->clist, entry) { 973 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) 974 continue; 975 node->flags &= ~HAMMER_NODE_NEEDSCRC; 976 KKASSERT(node->ondisk); 977 if (inval == 0) 978 node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE); 979 hammer_rel_node(node); 980 goto restart; 981 } 982 } 983 /* caller must still have ref on io */ 984 KKASSERT(hammer_isactive(&io->lock)); 985 } 986 987 /* 988 * Clear the IO's modify list. Even though the IO is no longer modified 989 * it may still be on the lose_root. This routine is called just before 990 * the governing hammer_buffer is destroyed. 991 * 992 * mod_root requires io_token protection. 993 */ 994 void 995 hammer_io_clear_modlist(struct hammer_io *io) 996 { 997 hammer_mount_t hmp = io->hmp; 998 999 KKASSERT(io->modified == 0); 1000 if (io->mod_root) { 1001 lwkt_gettoken(&hmp->io_token); 1002 if (io->mod_root) { 1003 KKASSERT(io->mod_root == &io->hmp->lose_root); 1004 RB_REMOVE(hammer_mod_rb_tree, io->mod_root, io); 1005 io->mod_root = NULL; 1006 } 1007 lwkt_reltoken(&hmp->io_token); 1008 } 1009 } 1010 1011 static void 1012 hammer_io_set_modlist(struct hammer_io *io) 1013 { 1014 struct hammer_mount *hmp = io->hmp; 1015 1016 lwkt_gettoken(&hmp->io_token); 1017 KKASSERT(io->mod_root == NULL); 1018 1019 switch(io->type) { 1020 case HAMMER_STRUCTURE_VOLUME: 1021 io->mod_root = &hmp->volu_root; 1022 hmp->locked_dirty_space += io->bytes; 1023 atomic_add_long(&hammer_count_dirtybufspace, io->bytes); 1024 break; 1025 case HAMMER_STRUCTURE_META_BUFFER: 1026 io->mod_root = &hmp->meta_root; 1027 hmp->locked_dirty_space += io->bytes; 1028 atomic_add_long(&hammer_count_dirtybufspace, io->bytes); 1029 break; 1030 case HAMMER_STRUCTURE_UNDO_BUFFER: 1031 io->mod_root = &hmp->undo_root; 1032 break; 1033 case HAMMER_STRUCTURE_DATA_BUFFER: 1034 io->mod_root = &hmp->data_root; 1035 break; 1036 case HAMMER_STRUCTURE_DUMMY: 1037 panic("hammer_io_set_modlist: bad io type"); 1038 break; /* NOT REACHED */ 1039 } 1040 if (RB_INSERT(hammer_mod_rb_tree, io->mod_root, io)) { 1041 panic("hammer_io_set_modlist: duplicate entry"); 1042 /* NOT REACHED */ 1043 } 1044 lwkt_reltoken(&hmp->io_token); 1045 } 1046 1047 /************************************************************************ 1048 * HAMMER_BIOOPS * 1049 ************************************************************************ 1050 * 1051 */ 1052 1053 /* 1054 * Pre-IO initiation kernel callback - cluster build only 1055 * 1056 * bioops callback - hold io_token 1057 */ 1058 static void 1059 hammer_io_start(struct buf *bp) 1060 { 1061 /* nothing to do, so io_token not needed */ 1062 } 1063 1064 /* 1065 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT! 1066 * 1067 * NOTE: HAMMER may modify a data buffer after we have initiated write 1068 * I/O. 1069 * 1070 * NOTE: MPSAFE callback 1071 * 1072 * bioops callback - hold io_token 1073 */ 1074 static void 1075 hammer_io_complete(struct buf *bp) 1076 { 1077 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep); 1078 struct hammer_mount *hmp = iou->io.hmp; 1079 struct hammer_io *ionext; 1080 1081 lwkt_gettoken(&hmp->io_token); 1082 1083 KKASSERT(iou->io.released == 1); 1084 1085 /* 1086 * Deal with people waiting for I/O to drain 1087 */ 1088 if (iou->io.running) { 1089 /* 1090 * Deal with critical write errors. Once a critical error 1091 * has been flagged in hmp the UNDO FIFO will not be updated. 1092 * That way crash recover will give us a consistent 1093 * filesystem. 1094 * 1095 * Because of this we can throw away failed UNDO buffers. If 1096 * we throw away META or DATA buffers we risk corrupting 1097 * the now read-only version of the filesystem visible to 1098 * the user. Clear B_ERROR so the buffer is not re-dirtied 1099 * by the kernel and ref the io so it doesn't get thrown 1100 * away. 1101 */ 1102 if (bp->b_flags & B_ERROR) { 1103 lwkt_gettoken(&hmp->fs_token); 1104 hammer_critical_error(hmp, NULL, bp->b_error, 1105 "while flushing meta-data"); 1106 lwkt_reltoken(&hmp->fs_token); 1107 1108 switch(iou->io.type) { 1109 case HAMMER_STRUCTURE_UNDO_BUFFER: 1110 break; 1111 default: 1112 if (iou->io.ioerror == 0) { 1113 iou->io.ioerror = 1; 1114 hammer_ref(&iou->io.lock); 1115 } 1116 break; 1117 } 1118 bp->b_flags &= ~B_ERROR; 1119 bundirty(bp); 1120 #if 0 1121 hammer_io_set_modlist(&iou->io); 1122 iou->io.modified = 1; 1123 #endif 1124 } 1125 hammer_stats_disk_write += iou->io.bytes; 1126 atomic_add_long(&hammer_count_io_running_write, -iou->io.bytes); 1127 atomic_add_long(&hmp->io_running_space, -iou->io.bytes); 1128 KKASSERT(hmp->io_running_space >= 0); 1129 iou->io.running = 0; 1130 1131 /* 1132 * Remove from iorun list and wakeup any multi-io waiter(s). 1133 */ 1134 if (TAILQ_FIRST(&hmp->iorun_list) == &iou->io) { 1135 ionext = TAILQ_NEXT(&iou->io, iorun_entry); 1136 if (ionext && ionext->type == HAMMER_STRUCTURE_DUMMY) 1137 wakeup(ionext); 1138 } 1139 TAILQ_REMOVE(&hmp->iorun_list, &iou->io, iorun_entry); 1140 } else { 1141 hammer_stats_disk_read += iou->io.bytes; 1142 } 1143 1144 if (iou->io.waiting) { 1145 iou->io.waiting = 0; 1146 wakeup(iou); 1147 } 1148 1149 /* 1150 * If B_LOCKED is set someone wanted to deallocate the bp at some 1151 * point, try to do it now. The operation will fail if there are 1152 * refs or if hammer_io_deallocate() is unable to gain the 1153 * interlock. 1154 */ 1155 if (bp->b_flags & B_LOCKED) { 1156 atomic_add_int(&hammer_count_io_locked, -1); 1157 bp->b_flags &= ~B_LOCKED; 1158 hammer_io_deallocate(bp); 1159 /* structure may be dead now */ 1160 } 1161 lwkt_reltoken(&hmp->io_token); 1162 } 1163 1164 /* 1165 * Callback from kernel when it wishes to deallocate a passively 1166 * associated structure. This mostly occurs with clean buffers 1167 * but it may be possible for a holding structure to be marked dirty 1168 * while its buffer is passively associated. The caller owns the bp. 1169 * 1170 * If we cannot disassociate we set B_LOCKED to prevent the buffer 1171 * from getting reused. 1172 * 1173 * WARNING: Because this can be called directly by getnewbuf we cannot 1174 * recurse into the tree. If a bp cannot be immediately disassociated 1175 * our only recourse is to set B_LOCKED. 1176 * 1177 * WARNING: This may be called from an interrupt via hammer_io_complete() 1178 * 1179 * bioops callback - hold io_token 1180 */ 1181 static void 1182 hammer_io_deallocate(struct buf *bp) 1183 { 1184 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep); 1185 hammer_mount_t hmp; 1186 1187 hmp = iou->io.hmp; 1188 1189 lwkt_gettoken(&hmp->io_token); 1190 1191 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0); 1192 if (hammer_try_interlock_norefs(&iou->io.lock) == 0) { 1193 /* 1194 * We cannot safely disassociate a bp from a referenced 1195 * or interlocked HAMMER structure. 1196 */ 1197 bp->b_flags |= B_LOCKED; 1198 atomic_add_int(&hammer_count_io_locked, 1); 1199 } else if (iou->io.modified) { 1200 /* 1201 * It is not legal to disassociate a modified buffer. This 1202 * case really shouldn't ever occur. 1203 */ 1204 bp->b_flags |= B_LOCKED; 1205 atomic_add_int(&hammer_count_io_locked, 1); 1206 hammer_put_interlock(&iou->io.lock, 0); 1207 } else { 1208 /* 1209 * Disassociate the BP. If the io has no refs left we 1210 * have to add it to the loose list. The kernel has 1211 * locked the buffer and therefore our io must be 1212 * in a released state. 1213 */ 1214 hammer_io_disassociate(iou); 1215 if (iou->io.type != HAMMER_STRUCTURE_VOLUME) { 1216 KKASSERT(iou->io.bp == NULL); 1217 KKASSERT(iou->io.mod_root == NULL); 1218 iou->io.mod_root = &hmp->lose_root; 1219 if (RB_INSERT(hammer_mod_rb_tree, iou->io.mod_root, 1220 &iou->io)) { 1221 panic("hammer_io_deallocate: duplicate entry"); 1222 } 1223 } 1224 hammer_put_interlock(&iou->io.lock, 1); 1225 } 1226 lwkt_reltoken(&hmp->io_token); 1227 } 1228 1229 /* 1230 * bioops callback - hold io_token 1231 */ 1232 static int 1233 hammer_io_fsync(struct vnode *vp) 1234 { 1235 /* nothing to do, so io_token not needed */ 1236 return(0); 1237 } 1238 1239 /* 1240 * NOTE: will not be called unless we tell the kernel about the 1241 * bioops. Unused... we use the mount's VFS_SYNC instead. 1242 * 1243 * bioops callback - hold io_token 1244 */ 1245 static int 1246 hammer_io_sync(struct mount *mp) 1247 { 1248 /* nothing to do, so io_token not needed */ 1249 return(0); 1250 } 1251 1252 /* 1253 * bioops callback - hold io_token 1254 */ 1255 static void 1256 hammer_io_movedeps(struct buf *bp1, struct buf *bp2) 1257 { 1258 /* nothing to do, so io_token not needed */ 1259 } 1260 1261 /* 1262 * I/O pre-check for reading and writing. HAMMER only uses this for 1263 * B_CACHE buffers so checkread just shouldn't happen, but if it does 1264 * allow it. 1265 * 1266 * Writing is a different case. We don't want the kernel to try to write 1267 * out a buffer that HAMMER may be modifying passively or which has a 1268 * dependancy. In addition, kernel-demanded writes can only proceed for 1269 * certain types of buffers (i.e. UNDO and DATA types). Other dirty 1270 * buffer types can only be explicitly written by the flusher. 1271 * 1272 * checkwrite will only be called for bdwrite()n buffers. If we return 1273 * success the kernel is guaranteed to initiate the buffer write. 1274 * 1275 * bioops callback - hold io_token 1276 */ 1277 static int 1278 hammer_io_checkread(struct buf *bp) 1279 { 1280 /* nothing to do, so io_token not needed */ 1281 return(0); 1282 } 1283 1284 /* 1285 * The kernel is asking us whether it can write out a dirty buffer or not. 1286 * 1287 * bioops callback - hold io_token 1288 */ 1289 static int 1290 hammer_io_checkwrite(struct buf *bp) 1291 { 1292 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep); 1293 hammer_mount_t hmp = io->hmp; 1294 1295 /* 1296 * This shouldn't happen under normal operation. 1297 */ 1298 lwkt_gettoken(&hmp->io_token); 1299 if (io->type == HAMMER_STRUCTURE_VOLUME || 1300 io->type == HAMMER_STRUCTURE_META_BUFFER) { 1301 if (!panicstr) 1302 panic("hammer_io_checkwrite: illegal buffer"); 1303 if ((bp->b_flags & B_LOCKED) == 0) { 1304 bp->b_flags |= B_LOCKED; 1305 atomic_add_int(&hammer_count_io_locked, 1); 1306 } 1307 lwkt_reltoken(&hmp->io_token); 1308 return(1); 1309 } 1310 1311 /* 1312 * We have to be able to interlock the IO to safely modify any 1313 * of its fields without holding the fs_token. If we can't lock 1314 * it then we are racing someone. 1315 * 1316 * Our ownership of the bp lock prevents the io from being ripped 1317 * out from under us. 1318 */ 1319 if (hammer_try_interlock_norefs(&io->lock) == 0) { 1320 bp->b_flags |= B_LOCKED; 1321 atomic_add_int(&hammer_count_io_locked, 1); 1322 lwkt_reltoken(&hmp->io_token); 1323 return(1); 1324 } 1325 1326 /* 1327 * The modified bit must be cleared prior to the initiation of 1328 * any IO (returning 0 initiates the IO). Because this is a 1329 * normal data buffer hammer_io_clear_modify() runs through a 1330 * simple degenerate case. 1331 * 1332 * Return 0 will cause the kernel to initiate the IO, and we 1333 * must normally clear the modified bit before we begin. If 1334 * the io has modify_refs we do not clear the modified bit, 1335 * otherwise we may miss changes. 1336 * 1337 * Only data and undo buffers can reach here. These buffers do 1338 * not have terminal crc functions but we temporarily reference 1339 * the IO anyway, just in case. 1340 */ 1341 if (io->modify_refs == 0 && io->modified) { 1342 hammer_ref(&io->lock); 1343 hammer_io_clear_modify(io, 0); 1344 hammer_rel(&io->lock); 1345 } else if (io->modified) { 1346 KKASSERT(io->type == HAMMER_STRUCTURE_DATA_BUFFER); 1347 } 1348 1349 /* 1350 * The kernel is going to start the IO, set io->running. 1351 */ 1352 KKASSERT(io->running == 0); 1353 io->running = 1; 1354 atomic_add_long(&io->hmp->io_running_space, io->bytes); 1355 atomic_add_long(&hammer_count_io_running_write, io->bytes); 1356 TAILQ_INSERT_TAIL(&io->hmp->iorun_list, io, iorun_entry); 1357 1358 hammer_put_interlock(&io->lock, 1); 1359 lwkt_reltoken(&hmp->io_token); 1360 1361 return(0); 1362 } 1363 1364 /* 1365 * Return non-zero if we wish to delay the kernel's attempt to flush 1366 * this buffer to disk. 1367 * 1368 * bioops callback - hold io_token 1369 */ 1370 static int 1371 hammer_io_countdeps(struct buf *bp, int n) 1372 { 1373 /* nothing to do, so io_token not needed */ 1374 return(0); 1375 } 1376 1377 struct bio_ops hammer_bioops = { 1378 .io_start = hammer_io_start, 1379 .io_complete = hammer_io_complete, 1380 .io_deallocate = hammer_io_deallocate, 1381 .io_fsync = hammer_io_fsync, 1382 .io_sync = hammer_io_sync, 1383 .io_movedeps = hammer_io_movedeps, 1384 .io_countdeps = hammer_io_countdeps, 1385 .io_checkread = hammer_io_checkread, 1386 .io_checkwrite = hammer_io_checkwrite, 1387 }; 1388 1389 /************************************************************************ 1390 * DIRECT IO OPS * 1391 ************************************************************************ 1392 * 1393 * These functions operate directly on the buffer cache buffer associated 1394 * with a front-end vnode rather then a back-end device vnode. 1395 */ 1396 1397 /* 1398 * Read a buffer associated with a front-end vnode directly from the 1399 * disk media. The bio may be issued asynchronously. If leaf is non-NULL 1400 * we validate the CRC. 1401 * 1402 * We must check for the presence of a HAMMER buffer to handle the case 1403 * where the reblocker has rewritten the data (which it does via the HAMMER 1404 * buffer system, not via the high-level vnode buffer cache), but not yet 1405 * committed the buffer to the media. 1406 */ 1407 int 1408 hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio, 1409 hammer_btree_leaf_elm_t leaf) 1410 { 1411 hammer_off_t buf_offset; 1412 hammer_off_t zone2_offset; 1413 hammer_volume_t volume; 1414 struct buf *bp; 1415 struct bio *nbio; 1416 int vol_no; 1417 int error; 1418 1419 buf_offset = bio->bio_offset; 1420 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) == 1421 HAMMER_ZONE_LARGE_DATA); 1422 1423 /* 1424 * The buffer cache may have an aliased buffer (the reblocker can 1425 * write them). If it does we have to sync any dirty data before 1426 * we can build our direct-read. This is a non-critical code path. 1427 */ 1428 bp = bio->bio_buf; 1429 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize); 1430 1431 /* 1432 * Resolve to a zone-2 offset. The conversion just requires 1433 * munging the top 4 bits but we want to abstract it anyway 1434 * so the blockmap code can verify the zone assignment. 1435 */ 1436 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1437 if (error) 1438 goto done; 1439 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) == 1440 HAMMER_ZONE_RAW_BUFFER); 1441 1442 /* 1443 * Resolve volume and raw-offset for 3rd level bio. The 1444 * offset will be specific to the volume. 1445 */ 1446 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1447 volume = hammer_get_volume(hmp, vol_no, &error); 1448 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1449 error = EIO; 1450 1451 if (error == 0) { 1452 /* 1453 * 3rd level bio 1454 */ 1455 nbio = push_bio(bio); 1456 nbio->bio_offset = volume->ondisk->vol_buf_beg + 1457 (zone2_offset & HAMMER_OFF_SHORT_MASK); 1458 #if 0 1459 /* 1460 * XXX disabled - our CRC check doesn't work if the OS 1461 * does bogus_page replacement on the direct-read. 1462 */ 1463 if (leaf && hammer_verify_data) { 1464 nbio->bio_done = hammer_io_direct_read_complete; 1465 nbio->bio_caller_info1.uvalue32 = leaf->data_crc; 1466 } 1467 #endif 1468 hammer_stats_disk_read += bp->b_bufsize; 1469 vn_strategy(volume->devvp, nbio); 1470 } 1471 hammer_rel_volume(volume, 0); 1472 done: 1473 if (error) { 1474 kprintf("hammer_direct_read: failed @ %016llx\n", 1475 (long long)zone2_offset); 1476 bp->b_error = error; 1477 bp->b_flags |= B_ERROR; 1478 biodone(bio); 1479 } 1480 return(error); 1481 } 1482 1483 /* 1484 * This works similarly to hammer_io_direct_read() except instead of 1485 * directly reading from the device into the bio we instead indirectly 1486 * read through the device's buffer cache and then copy the data into 1487 * the bio. 1488 * 1489 * If leaf is non-NULL and validation is enabled, the CRC will be checked. 1490 * 1491 * This routine also executes asynchronously. It allows hammer strategy 1492 * calls to operate asynchronously when in double_buffer mode (in addition 1493 * to operating asynchronously when in normal mode). 1494 */ 1495 int 1496 hammer_io_indirect_read(hammer_mount_t hmp, struct bio *bio, 1497 hammer_btree_leaf_elm_t leaf) 1498 { 1499 hammer_off_t buf_offset; 1500 hammer_off_t zone2_offset; 1501 hammer_volume_t volume; 1502 struct buf *bp; 1503 int vol_no; 1504 int error; 1505 1506 buf_offset = bio->bio_offset; 1507 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) == 1508 HAMMER_ZONE_LARGE_DATA); 1509 1510 /* 1511 * The buffer cache may have an aliased buffer (the reblocker can 1512 * write them). If it does we have to sync any dirty data before 1513 * we can build our direct-read. This is a non-critical code path. 1514 */ 1515 bp = bio->bio_buf; 1516 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize); 1517 1518 /* 1519 * Resolve to a zone-2 offset. The conversion just requires 1520 * munging the top 4 bits but we want to abstract it anyway 1521 * so the blockmap code can verify the zone assignment. 1522 */ 1523 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1524 if (error) 1525 goto done; 1526 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) == 1527 HAMMER_ZONE_RAW_BUFFER); 1528 1529 /* 1530 * Resolve volume and raw-offset for 3rd level bio. The 1531 * offset will be specific to the volume. 1532 */ 1533 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1534 volume = hammer_get_volume(hmp, vol_no, &error); 1535 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1536 error = EIO; 1537 1538 if (error == 0) { 1539 /* 1540 * Convert to the raw volume->devvp offset and acquire 1541 * the buf, issuing async I/O if necessary. 1542 */ 1543 buf_offset = volume->ondisk->vol_buf_beg + 1544 (zone2_offset & HAMMER_OFF_SHORT_MASK); 1545 1546 if (leaf && hammer_verify_data) { 1547 bio->bio_caller_info1.uvalue32 = leaf->data_crc; 1548 bio->bio_caller_info2.index = 1; 1549 } else { 1550 bio->bio_caller_info2.index = 0; 1551 } 1552 breadcb(volume->devvp, buf_offset, bp->b_bufsize, 1553 hammer_indirect_callback, bio); 1554 } 1555 hammer_rel_volume(volume, 0); 1556 done: 1557 if (error) { 1558 kprintf("hammer_direct_read: failed @ %016llx\n", 1559 (long long)zone2_offset); 1560 bp->b_error = error; 1561 bp->b_flags |= B_ERROR; 1562 biodone(bio); 1563 } 1564 return(error); 1565 } 1566 1567 /* 1568 * Indirect callback on completion. bio/bp specify the device-backed 1569 * buffer. bio->bio_caller_info1.ptr holds obio. 1570 * 1571 * obio/obp is the original regular file buffer. obio->bio_caller_info* 1572 * contains the crc specification. 1573 * 1574 * We are responsible for calling bpdone() and bqrelse() on bio/bp, and 1575 * for calling biodone() on obio. 1576 */ 1577 static void 1578 hammer_indirect_callback(struct bio *bio) 1579 { 1580 struct buf *bp = bio->bio_buf; 1581 struct buf *obp; 1582 struct bio *obio; 1583 1584 /* 1585 * If BIO_DONE is already set the device buffer was already 1586 * fully valid (B_CACHE). If it is not set then I/O was issued 1587 * and we have to run I/O completion as the last bio. 1588 * 1589 * Nobody is waiting for our device I/O to complete, we are 1590 * responsible for bqrelse()ing it which means we also have to do 1591 * the equivalent of biowait() and clear BIO_DONE (which breadcb() 1592 * may have set). 1593 * 1594 * Any preexisting device buffer should match the requested size, 1595 * but due to bigblock recycling and other factors there is some 1596 * fragility there, so we assert that the device buffer covers 1597 * the request. 1598 */ 1599 if ((bio->bio_flags & BIO_DONE) == 0) 1600 bpdone(bp, 0); 1601 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC); 1602 1603 obio = bio->bio_caller_info1.ptr; 1604 obp = obio->bio_buf; 1605 1606 if (bp->b_flags & B_ERROR) { 1607 obp->b_flags |= B_ERROR; 1608 obp->b_error = bp->b_error; 1609 } else if (obio->bio_caller_info2.index && 1610 obio->bio_caller_info1.uvalue32 != 1611 crc32(bp->b_data, bp->b_bufsize)) { 1612 obp->b_flags |= B_ERROR; 1613 obp->b_error = EIO; 1614 } else { 1615 KKASSERT(bp->b_bufsize >= obp->b_bufsize); 1616 bcopy(bp->b_data, obp->b_data, obp->b_bufsize); 1617 obp->b_resid = 0; 1618 obp->b_flags |= B_AGE; 1619 } 1620 biodone(obio); 1621 bqrelse(bp); 1622 } 1623 1624 #if 0 1625 /* 1626 * On completion of the BIO this callback must check the data CRC 1627 * and chain to the previous bio. 1628 * 1629 * MPSAFE - since we do not modify and hammer_records we do not need 1630 * io_token. 1631 * 1632 * NOTE: MPSAFE callback 1633 */ 1634 static 1635 void 1636 hammer_io_direct_read_complete(struct bio *nbio) 1637 { 1638 struct bio *obio; 1639 struct buf *bp; 1640 u_int32_t rec_crc = nbio->bio_caller_info1.uvalue32; 1641 1642 bp = nbio->bio_buf; 1643 if (crc32(bp->b_data, bp->b_bufsize) != rec_crc) { 1644 kprintf("HAMMER: data_crc error @%016llx/%d\n", 1645 nbio->bio_offset, bp->b_bufsize); 1646 if (hammer_debug_critical) 1647 Debugger("data_crc on read"); 1648 bp->b_flags |= B_ERROR; 1649 bp->b_error = EIO; 1650 } 1651 obio = pop_bio(nbio); 1652 biodone(obio); 1653 } 1654 #endif 1655 1656 /* 1657 * Write a buffer associated with a front-end vnode directly to the 1658 * disk media. The bio may be issued asynchronously. 1659 * 1660 * The BIO is associated with the specified record and RECG_DIRECT_IO 1661 * is set. The recorded is added to its object. 1662 */ 1663 int 1664 hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio, 1665 hammer_record_t record) 1666 { 1667 hammer_btree_leaf_elm_t leaf = &record->leaf; 1668 hammer_off_t buf_offset; 1669 hammer_off_t zone2_offset; 1670 hammer_volume_t volume; 1671 hammer_buffer_t buffer; 1672 struct buf *bp; 1673 struct bio *nbio; 1674 char *ptr; 1675 int vol_no; 1676 int error; 1677 1678 buf_offset = leaf->data_offset; 1679 1680 KKASSERT(buf_offset > HAMMER_ZONE_BTREE); 1681 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE); 1682 1683 /* 1684 * Issue or execute the I/O. The new memory record must replace 1685 * the old one before the I/O completes, otherwise a reaquisition of 1686 * the buffer will load the old media data instead of the new. 1687 */ 1688 if ((buf_offset & HAMMER_BUFMASK) == 0 && 1689 leaf->data_len >= HAMMER_BUFSIZE) { 1690 /* 1691 * We are using the vnode's bio to write directly to the 1692 * media, any hammer_buffer at the same zone-X offset will 1693 * now have stale data. 1694 */ 1695 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1696 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1697 volume = hammer_get_volume(hmp, vol_no, &error); 1698 1699 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1700 error = EIO; 1701 if (error == 0) { 1702 bp = bio->bio_buf; 1703 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0); 1704 /* 1705 hammer_del_buffers(hmp, buf_offset, 1706 zone2_offset, bp->b_bufsize); 1707 */ 1708 1709 /* 1710 * Second level bio - cached zone2 offset. 1711 * 1712 * (We can put our bio_done function in either the 1713 * 2nd or 3rd level). 1714 */ 1715 nbio = push_bio(bio); 1716 nbio->bio_offset = zone2_offset; 1717 nbio->bio_done = hammer_io_direct_write_complete; 1718 nbio->bio_caller_info1.ptr = record; 1719 record->zone2_offset = zone2_offset; 1720 record->gflags |= HAMMER_RECG_DIRECT_IO | 1721 HAMMER_RECG_DIRECT_INVAL; 1722 1723 /* 1724 * Third level bio - raw offset specific to the 1725 * correct volume. 1726 */ 1727 zone2_offset &= HAMMER_OFF_SHORT_MASK; 1728 nbio = push_bio(nbio); 1729 nbio->bio_offset = volume->ondisk->vol_buf_beg + 1730 zone2_offset; 1731 hammer_stats_disk_write += bp->b_bufsize; 1732 hammer_ip_replace_bulk(hmp, record); 1733 vn_strategy(volume->devvp, nbio); 1734 hammer_io_flush_mark(volume); 1735 } 1736 hammer_rel_volume(volume, 0); 1737 } else { 1738 /* 1739 * Must fit in a standard HAMMER buffer. In this case all 1740 * consumers use the HAMMER buffer system and RECG_DIRECT_IO 1741 * does not need to be set-up. 1742 */ 1743 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0); 1744 buffer = NULL; 1745 ptr = hammer_bread(hmp, buf_offset, &error, &buffer); 1746 if (error == 0) { 1747 bp = bio->bio_buf; 1748 bp->b_flags |= B_AGE; 1749 hammer_io_modify(&buffer->io, 1); 1750 bcopy(bp->b_data, ptr, leaf->data_len); 1751 hammer_io_modify_done(&buffer->io); 1752 hammer_rel_buffer(buffer, 0); 1753 bp->b_resid = 0; 1754 hammer_ip_replace_bulk(hmp, record); 1755 biodone(bio); 1756 } 1757 } 1758 if (error) { 1759 /* 1760 * Major suckage occured. Also note: The record was 1761 * never added to the tree so we do not have to worry 1762 * about the backend. 1763 */ 1764 kprintf("hammer_direct_write: failed @ %016llx\n", 1765 (long long)leaf->data_offset); 1766 bp = bio->bio_buf; 1767 bp->b_resid = 0; 1768 bp->b_error = EIO; 1769 bp->b_flags |= B_ERROR; 1770 biodone(bio); 1771 record->flags |= HAMMER_RECF_DELETED_FE; 1772 hammer_rel_mem_record(record); 1773 } 1774 return(error); 1775 } 1776 1777 /* 1778 * On completion of the BIO this callback must disconnect 1779 * it from the hammer_record and chain to the previous bio. 1780 * 1781 * An I/O error forces the mount to read-only. Data buffers 1782 * are not B_LOCKED like meta-data buffers are, so we have to 1783 * throw the buffer away to prevent the kernel from retrying. 1784 * 1785 * NOTE: MPSAFE callback, only modify fields we have explicit 1786 * access to (the bp and the record->gflags). 1787 */ 1788 static 1789 void 1790 hammer_io_direct_write_complete(struct bio *nbio) 1791 { 1792 struct bio *obio; 1793 struct buf *bp; 1794 hammer_record_t record; 1795 hammer_mount_t hmp; 1796 1797 record = nbio->bio_caller_info1.ptr; 1798 KKASSERT(record != NULL); 1799 hmp = record->ip->hmp; 1800 1801 lwkt_gettoken(&hmp->io_token); 1802 1803 bp = nbio->bio_buf; 1804 obio = pop_bio(nbio); 1805 if (bp->b_flags & B_ERROR) { 1806 lwkt_gettoken(&hmp->fs_token); 1807 hammer_critical_error(hmp, record->ip, 1808 bp->b_error, 1809 "while writing bulk data"); 1810 lwkt_reltoken(&hmp->fs_token); 1811 bp->b_flags |= B_INVAL; 1812 } 1813 biodone(obio); 1814 1815 KKASSERT(record->gflags & HAMMER_RECG_DIRECT_IO); 1816 if (record->gflags & HAMMER_RECG_DIRECT_WAIT) { 1817 record->gflags &= ~(HAMMER_RECG_DIRECT_IO | 1818 HAMMER_RECG_DIRECT_WAIT); 1819 /* record can disappear once DIRECT_IO flag is cleared */ 1820 wakeup(&record->flags); 1821 } else { 1822 record->gflags &= ~HAMMER_RECG_DIRECT_IO; 1823 /* record can disappear once DIRECT_IO flag is cleared */ 1824 } 1825 lwkt_reltoken(&hmp->io_token); 1826 } 1827 1828 1829 /* 1830 * This is called before a record is either committed to the B-Tree 1831 * or destroyed, to resolve any associated direct-IO. 1832 * 1833 * (1) We must wait for any direct-IO related to the record to complete. 1834 * 1835 * (2) We must remove any buffer cache aliases for data accessed via 1836 * leaf->data_offset or zone2_offset so non-direct-IO consumers 1837 * (the mirroring and reblocking code) do not see stale data. 1838 */ 1839 void 1840 hammer_io_direct_wait(hammer_record_t record) 1841 { 1842 hammer_mount_t hmp = record->ip->hmp; 1843 1844 /* 1845 * Wait for I/O to complete 1846 */ 1847 if (record->gflags & HAMMER_RECG_DIRECT_IO) { 1848 lwkt_gettoken(&hmp->io_token); 1849 while (record->gflags & HAMMER_RECG_DIRECT_IO) { 1850 record->gflags |= HAMMER_RECG_DIRECT_WAIT; 1851 tsleep(&record->flags, 0, "hmdiow", 0); 1852 } 1853 lwkt_reltoken(&hmp->io_token); 1854 } 1855 1856 /* 1857 * Invalidate any related buffer cache aliases associated with the 1858 * backing device. This is needed because the buffer cache buffer 1859 * for file data is associated with the file vnode, not the backing 1860 * device vnode. 1861 * 1862 * XXX I do not think this case can occur any more now that 1863 * reservations ensure that all such buffers are removed before 1864 * an area can be reused. 1865 */ 1866 if (record->gflags & HAMMER_RECG_DIRECT_INVAL) { 1867 KKASSERT(record->leaf.data_offset); 1868 hammer_del_buffers(hmp, record->leaf.data_offset, 1869 record->zone2_offset, record->leaf.data_len, 1870 1); 1871 record->gflags &= ~HAMMER_RECG_DIRECT_INVAL; 1872 } 1873 } 1874 1875 /* 1876 * This is called to remove the second-level cached zone-2 offset from 1877 * frontend buffer cache buffers, now stale due to a data relocation. 1878 * These offsets are generated by cluster_read() via VOP_BMAP, or directly 1879 * by hammer_vop_strategy_read(). 1880 * 1881 * This is rather nasty because here we have something like the reblocker 1882 * scanning the raw B-Tree with no held references on anything, really, 1883 * other then a shared lock on the B-Tree node, and we have to access the 1884 * frontend's buffer cache to check for and clean out the association. 1885 * Specifically, if the reblocker is moving data on the disk, these cached 1886 * offsets will become invalid. 1887 * 1888 * Only data record types associated with the large-data zone are subject 1889 * to direct-io and need to be checked. 1890 * 1891 */ 1892 void 1893 hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf) 1894 { 1895 struct hammer_inode_info iinfo; 1896 int zone; 1897 1898 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA) 1899 return; 1900 zone = HAMMER_ZONE_DECODE(leaf->data_offset); 1901 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX) 1902 return; 1903 iinfo.obj_id = leaf->base.obj_id; 1904 iinfo.obj_asof = 0; /* unused */ 1905 iinfo.obj_localization = leaf->base.localization & 1906 HAMMER_LOCALIZE_PSEUDOFS_MASK; 1907 iinfo.u.leaf = leaf; 1908 hammer_scan_inode_snapshots(hmp, &iinfo, 1909 hammer_io_direct_uncache_callback, 1910 leaf); 1911 } 1912 1913 static int 1914 hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data) 1915 { 1916 hammer_inode_info_t iinfo = data; 1917 hammer_off_t data_offset; 1918 hammer_off_t file_offset; 1919 struct vnode *vp; 1920 struct buf *bp; 1921 int blksize; 1922 1923 if (ip->vp == NULL) 1924 return(0); 1925 data_offset = iinfo->u.leaf->data_offset; 1926 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len; 1927 blksize = iinfo->u.leaf->data_len; 1928 KKASSERT((blksize & HAMMER_BUFMASK) == 0); 1929 1930 /* 1931 * Warning: FINDBLK_TEST return stable storage but not stable 1932 * contents. It happens to be ok in this case. 1933 */ 1934 hammer_ref(&ip->lock); 1935 if (hammer_get_vnode(ip, &vp) == 0) { 1936 if ((bp = findblk(ip->vp, file_offset, FINDBLK_TEST)) != NULL && 1937 bp->b_bio2.bio_offset != NOOFFSET) { 1938 bp = getblk(ip->vp, file_offset, blksize, 0, 0); 1939 bp->b_bio2.bio_offset = NOOFFSET; 1940 brelse(bp); 1941 } 1942 vput(vp); 1943 } 1944 hammer_rel_inode(ip, 0); 1945 return(0); 1946 } 1947 1948 1949 /* 1950 * This function is called when writes may have occured on the volume, 1951 * indicating that the device may be holding cached writes. 1952 */ 1953 static void 1954 hammer_io_flush_mark(hammer_volume_t volume) 1955 { 1956 atomic_set_int(&volume->vol_flags, HAMMER_VOLF_NEEDFLUSH); 1957 } 1958 1959 /* 1960 * This function ensures that the device has flushed any cached writes out. 1961 */ 1962 void 1963 hammer_io_flush_sync(hammer_mount_t hmp) 1964 { 1965 hammer_volume_t volume; 1966 struct buf *bp_base = NULL; 1967 struct buf *bp; 1968 1969 RB_FOREACH(volume, hammer_vol_rb_tree, &hmp->rb_vols_root) { 1970 if (volume->vol_flags & HAMMER_VOLF_NEEDFLUSH) { 1971 atomic_clear_int(&volume->vol_flags, 1972 HAMMER_VOLF_NEEDFLUSH); 1973 bp = getpbuf(NULL); 1974 bp->b_bio1.bio_offset = 0; 1975 bp->b_bufsize = 0; 1976 bp->b_bcount = 0; 1977 bp->b_cmd = BUF_CMD_FLUSH; 1978 bp->b_bio1.bio_caller_info1.cluster_head = bp_base; 1979 bp->b_bio1.bio_done = biodone_sync; 1980 bp->b_bio1.bio_flags |= BIO_SYNC; 1981 bp_base = bp; 1982 vn_strategy(volume->devvp, &bp->b_bio1); 1983 } 1984 } 1985 while ((bp = bp_base) != NULL) { 1986 bp_base = bp->b_bio1.bio_caller_info1.cluster_head; 1987 biowait(&bp->b_bio1, "hmrFLS"); 1988 relpbuf(bp, NULL); 1989 } 1990 } 1991 1992 /* 1993 * Limit the amount of backlog which we allow to build up 1994 */ 1995 void 1996 hammer_io_limit_backlog(hammer_mount_t hmp) 1997 { 1998 waitrunningbufspace(); 1999 } 2000