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 #include <sys/buf2.h> 56 57 static void hammer_io_modify(hammer_io_t io, int count); 58 static void hammer_io_deallocate(struct buf *bp); 59 #if 0 60 static void hammer_io_direct_read_complete(struct bio *nbio); 61 #endif 62 static void hammer_io_direct_write_complete(struct bio *nbio); 63 static int hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data); 64 static void hammer_io_set_modlist(struct hammer_io *io); 65 static void hammer_io_flush_mark(hammer_volume_t volume); 66 67 static int 68 hammer_mod_rb_compare(hammer_io_t io1, hammer_io_t io2) 69 { 70 hammer_off_t io1_offset; 71 hammer_off_t io2_offset; 72 73 io1_offset = ((io1->offset & HAMMER_OFF_SHORT_MASK) << 8) | 74 HAMMER_VOL_DECODE(io1->offset); 75 io2_offset = ((io2->offset & HAMMER_OFF_SHORT_MASK) << 8) | 76 HAMMER_VOL_DECODE(io2->offset); 77 78 if (io1_offset < io2_offset) 79 return(-1); 80 if (io1_offset > io2_offset) 81 return(1); 82 return(0); 83 } 84 85 RB_GENERATE(hammer_mod_rb_tree, hammer_io, rb_node, hammer_mod_rb_compare); 86 87 /* 88 * Initialize a new, already-zero'd hammer_io structure, or reinitialize 89 * an existing hammer_io structure which may have switched to another type. 90 */ 91 void 92 hammer_io_init(hammer_io_t io, hammer_volume_t volume, enum hammer_io_type type) 93 { 94 io->volume = volume; 95 io->hmp = volume->io.hmp; 96 io->type = type; 97 } 98 99 /* 100 * Helper routine to disassociate a buffer cache buffer from an I/O 101 * structure. The io must be interlocked and marked appropriately for 102 * reclamation. 103 * 104 * The io must be in a released state with the io->bp owned and 105 * locked by the caller of this function. When not called from an 106 * io_deallocate() this cannot race an io_deallocate() since the 107 * kernel would be unable to get the buffer lock in that case. 108 * (The released state in this case means we own the bp, not the 109 * hammer_io structure). 110 * 111 * The io may have 0 or 1 references depending on who called us. The 112 * caller is responsible for dealing with the refs. 113 * 114 * This call can only be made when no action is required on the buffer. 115 * 116 * This function is guaranteed not to race against anything because we 117 * own both the io lock and the bp lock and are interlocked with no 118 * references. 119 */ 120 static void 121 hammer_io_disassociate(hammer_io_structure_t iou) 122 { 123 struct buf *bp = iou->io.bp; 124 125 KKASSERT(iou->io.released); 126 KKASSERT(iou->io.modified == 0); 127 KKASSERT(LIST_FIRST(&bp->b_dep) == (void *)iou); 128 buf_dep_init(bp); 129 iou->io.bp = NULL; 130 131 /* 132 * If the buffer was locked someone wanted to get rid of it. 133 */ 134 if (bp->b_flags & B_LOCKED) { 135 atomic_add_int(&hammer_count_io_locked, -1); 136 bp->b_flags &= ~B_LOCKED; 137 } 138 if (iou->io.reclaim) { 139 bp->b_flags |= B_NOCACHE|B_RELBUF; 140 iou->io.reclaim = 0; 141 } 142 143 switch(iou->io.type) { 144 case HAMMER_STRUCTURE_VOLUME: 145 iou->volume.ondisk = NULL; 146 break; 147 case HAMMER_STRUCTURE_DATA_BUFFER: 148 case HAMMER_STRUCTURE_META_BUFFER: 149 case HAMMER_STRUCTURE_UNDO_BUFFER: 150 iou->buffer.ondisk = NULL; 151 break; 152 case HAMMER_STRUCTURE_DUMMY: 153 panic("hammer_io_disassociate: bad io type"); 154 break; 155 } 156 } 157 158 /* 159 * Wait for any physical IO to complete 160 * 161 * XXX we aren't interlocked against a spinlock or anything so there 162 * is a small window in the interlock / io->running == 0 test. 163 */ 164 void 165 hammer_io_wait(hammer_io_t io) 166 { 167 if (io->running) { 168 hammer_mount_t hmp = io->hmp; 169 170 lwkt_gettoken(&hmp->io_token); 171 while (io->running) { 172 io->waiting = 1; 173 tsleep_interlock(io, 0); 174 if (io->running) 175 tsleep(io, PINTERLOCKED, "hmrflw", hz); 176 } 177 lwkt_reltoken(&hmp->io_token); 178 } 179 } 180 181 /* 182 * Wait for all currently queued HAMMER-initiated I/Os to complete. 183 * 184 * This is not supposed to count direct I/O's but some can leak 185 * through (for non-full-sized direct I/Os). 186 */ 187 void 188 hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush) 189 { 190 struct hammer_io iodummy; 191 hammer_io_t io; 192 193 /* 194 * Degenerate case, no I/O is running 195 */ 196 lwkt_gettoken(&hmp->io_token); 197 if (TAILQ_EMPTY(&hmp->iorun_list)) { 198 lwkt_reltoken(&hmp->io_token); 199 if (doflush) 200 hammer_io_flush_sync(hmp); 201 return; 202 } 203 bzero(&iodummy, sizeof(iodummy)); 204 iodummy.type = HAMMER_STRUCTURE_DUMMY; 205 206 /* 207 * Add placemarker and then wait until it becomes the head of 208 * the list. 209 */ 210 TAILQ_INSERT_TAIL(&hmp->iorun_list, &iodummy, iorun_entry); 211 while (TAILQ_FIRST(&hmp->iorun_list) != &iodummy) { 212 tsleep(&iodummy, 0, ident, 0); 213 } 214 215 /* 216 * Chain in case several placemarkers are present. 217 */ 218 TAILQ_REMOVE(&hmp->iorun_list, &iodummy, iorun_entry); 219 io = TAILQ_FIRST(&hmp->iorun_list); 220 if (io && io->type == HAMMER_STRUCTURE_DUMMY) 221 wakeup(io); 222 lwkt_reltoken(&hmp->io_token); 223 224 if (doflush) 225 hammer_io_flush_sync(hmp); 226 } 227 228 /* 229 * Clear a flagged error condition on a I/O buffer. The caller must hold 230 * its own ref on the buffer. 231 */ 232 void 233 hammer_io_clear_error(struct hammer_io *io) 234 { 235 hammer_mount_t hmp = io->hmp; 236 237 lwkt_gettoken(&hmp->io_token); 238 if (io->ioerror) { 239 io->ioerror = 0; 240 hammer_rel(&io->lock); 241 KKASSERT(hammer_isactive(&io->lock)); 242 } 243 lwkt_reltoken(&hmp->io_token); 244 } 245 246 void 247 hammer_io_clear_error_noassert(struct hammer_io *io) 248 { 249 hammer_mount_t hmp = io->hmp; 250 251 lwkt_gettoken(&hmp->io_token); 252 if (io->ioerror) { 253 io->ioerror = 0; 254 hammer_rel(&io->lock); 255 } 256 lwkt_reltoken(&hmp->io_token); 257 } 258 259 /* 260 * This is an advisory function only which tells the buffer cache 261 * the bp is not a meta-data buffer, even though it is backed by 262 * a block device. 263 * 264 * This is used by HAMMER's reblocking code to avoid trying to 265 * swapcache the filesystem's data when it is read or written 266 * by the reblocking code. 267 * 268 * The caller has a ref on the buffer preventing the bp from 269 * being disassociated from it. 270 */ 271 void 272 hammer_io_notmeta(hammer_buffer_t buffer) 273 { 274 if ((buffer->io.bp->b_flags & B_NOTMETA) == 0) { 275 hammer_mount_t hmp = buffer->io.hmp; 276 277 lwkt_gettoken(&hmp->io_token); 278 buffer->io.bp->b_flags |= B_NOTMETA; 279 lwkt_reltoken(&hmp->io_token); 280 } 281 } 282 283 /* 284 * Load bp for a HAMMER structure. The io must be exclusively locked by 285 * the caller. 286 * 287 * This routine is mostly used on meta-data and small-data blocks. Generally 288 * speaking HAMMER assumes some locality of reference and will cluster. 289 * 290 * Note that the caller (hammer_ondisk.c) may place further restrictions 291 * on clusterability via the limit (in bytes). Typically large-data 292 * zones cannot be clustered due to their mixed buffer sizes. This is 293 * not an issue since such clustering occurs in hammer_vnops at the 294 * regular file layer, whereas this is the buffered block device layer. 295 * 296 * No I/O callbacks can occur while we hold the buffer locked. 297 */ 298 int 299 hammer_io_read(struct vnode *devvp, struct hammer_io *io, int limit) 300 { 301 struct buf *bp; 302 int error; 303 304 if ((bp = io->bp) == NULL) { 305 atomic_add_int(&hammer_count_io_running_read, io->bytes); 306 if (hammer_cluster_enable && limit > io->bytes) { 307 error = cluster_read(devvp, io->offset + limit, 308 io->offset, io->bytes, 309 HAMMER_CLUSTER_SIZE, 310 HAMMER_CLUSTER_SIZE, 311 &io->bp); 312 } else { 313 error = bread(devvp, io->offset, io->bytes, &io->bp); 314 } 315 hammer_stats_disk_read += io->bytes; 316 atomic_add_int(&hammer_count_io_running_read, -io->bytes); 317 318 /* 319 * The code generally assumes b_ops/b_dep has been set-up, 320 * even if we error out here. 321 */ 322 bp = io->bp; 323 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) { 324 const char *metatype; 325 326 switch(io->type) { 327 case HAMMER_STRUCTURE_VOLUME: 328 metatype = "volume"; 329 break; 330 case HAMMER_STRUCTURE_META_BUFFER: 331 switch(((struct hammer_buffer *)io)-> 332 zoneX_offset & HAMMER_OFF_ZONE_MASK) { 333 case HAMMER_ZONE_BTREE: 334 metatype = "btree"; 335 break; 336 case HAMMER_ZONE_META: 337 metatype = "meta"; 338 break; 339 case HAMMER_ZONE_FREEMAP: 340 metatype = "freemap"; 341 break; 342 default: 343 metatype = "meta?"; 344 break; 345 } 346 break; 347 case HAMMER_STRUCTURE_DATA_BUFFER: 348 metatype = "data"; 349 break; 350 case HAMMER_STRUCTURE_UNDO_BUFFER: 351 metatype = "undo"; 352 break; 353 default: 354 metatype = "unknown"; 355 break; 356 } 357 kprintf("doff %016jx %s\n", 358 (intmax_t)bp->b_bio2.bio_offset, 359 metatype); 360 } 361 bp->b_flags &= ~B_IODEBUG; 362 bp->b_ops = &hammer_bioops; 363 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 364 365 /* io->worklist is locked by the io lock */ 366 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node); 367 BUF_KERNPROC(bp); 368 KKASSERT(io->modified == 0); 369 KKASSERT(io->running == 0); 370 KKASSERT(io->waiting == 0); 371 io->released = 0; /* we hold an active lock on bp */ 372 } else { 373 error = 0; 374 } 375 return(error); 376 } 377 378 /* 379 * Similar to hammer_io_read() but returns a zero'd out buffer instead. 380 * Must be called with the IO exclusively locked. 381 * 382 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background 383 * I/O by forcing the buffer to not be in a released state before calling 384 * it. 385 * 386 * This function will also mark the IO as modified but it will not 387 * increment the modify_refs count. 388 * 389 * No I/O callbacks can occur while we hold the buffer locked. 390 */ 391 int 392 hammer_io_new(struct vnode *devvp, struct hammer_io *io) 393 { 394 struct buf *bp; 395 396 if ((bp = io->bp) == NULL) { 397 io->bp = getblk(devvp, io->offset, io->bytes, 0, 0); 398 bp = io->bp; 399 bp->b_ops = &hammer_bioops; 400 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL); 401 402 /* io->worklist is locked by the io lock */ 403 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node); 404 io->released = 0; 405 KKASSERT(io->running == 0); 406 io->waiting = 0; 407 BUF_KERNPROC(bp); 408 } else { 409 if (io->released) { 410 regetblk(bp); 411 BUF_KERNPROC(bp); 412 io->released = 0; 413 } 414 } 415 hammer_io_modify(io, 0); 416 vfs_bio_clrbuf(bp); 417 return(0); 418 } 419 420 /* 421 * Advance the activity count on the underlying buffer because 422 * HAMMER does not getblk/brelse on every access. 423 * 424 * The io->bp cannot go away while the buffer is referenced. 425 */ 426 void 427 hammer_io_advance(struct hammer_io *io) 428 { 429 if (io->bp) 430 buf_act_advance(io->bp); 431 } 432 433 /* 434 * Remove potential device level aliases against buffers managed by high level 435 * vnodes. Aliases can also be created due to mixed buffer sizes or via 436 * direct access to the backing store device. 437 * 438 * This is nasty because the buffers are also VMIO-backed. Even if a buffer 439 * does not exist its backing VM pages might, and we have to invalidate 440 * those as well or a getblk() will reinstate them. 441 * 442 * Buffer cache buffers associated with hammer_buffers cannot be 443 * invalidated. 444 */ 445 int 446 hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset) 447 { 448 hammer_io_structure_t iou; 449 hammer_mount_t hmp; 450 hammer_off_t phys_offset; 451 struct buf *bp; 452 int error; 453 454 hmp = volume->io.hmp; 455 lwkt_gettoken(&hmp->io_token); 456 457 /* 458 * If a device buffer already exists for the specified physical 459 * offset use that, otherwise instantiate a buffer to cover any 460 * related VM pages, set BNOCACHE, and brelse(). 461 */ 462 phys_offset = volume->ondisk->vol_buf_beg + 463 (zone2_offset & HAMMER_OFF_SHORT_MASK); 464 if ((bp = findblk(volume->devvp, phys_offset, 0)) != NULL) 465 bremfree(bp); 466 else 467 bp = getblk(volume->devvp, phys_offset, HAMMER_BUFSIZE, 0, 0); 468 469 if ((iou = (void *)LIST_FIRST(&bp->b_dep)) != NULL) { 470 #if 0 471 hammer_ref(&iou->io.lock); 472 hammer_io_clear_modify(&iou->io, 1); 473 bundirty(bp); 474 iou->io.released = 0; 475 BUF_KERNPROC(bp); 476 iou->io.reclaim = 1; 477 iou->io.waitdep = 1; /* XXX this is a fs_token field */ 478 KKASSERT(hammer_isactive(&iou->io.lock) == 1); 479 hammer_rel_buffer(&iou->buffer, 0); 480 /*hammer_io_deallocate(bp);*/ 481 #endif 482 bqrelse(bp); 483 error = EAGAIN; 484 } else { 485 KKASSERT((bp->b_flags & B_LOCKED) == 0); 486 bundirty(bp); 487 bp->b_flags |= B_NOCACHE|B_RELBUF; 488 brelse(bp); 489 error = 0; 490 } 491 lwkt_reltoken(&hmp->io_token); 492 return(error); 493 } 494 495 /* 496 * This routine is called on the last reference to a hammer structure. 497 * The io must be interlocked with a refcount of zero. The hammer structure 498 * will remain interlocked on return. 499 * 500 * This routine may return a non-NULL bp to the caller for dispoal. 501 * The caller typically brelse()'s the bp. 502 * 503 * The bp may or may not still be passively associated with the IO. It 504 * will remain passively associated if it is unreleasable (e.g. a modified 505 * meta-data buffer). 506 * 507 * The only requirement here is that modified meta-data and volume-header 508 * buffer may NOT be disassociated from the IO structure, and consequently 509 * we also leave such buffers actively associated with the IO if they already 510 * are (since the kernel can't do anything with them anyway). Only the 511 * flusher is allowed to write such buffers out. Modified pure-data and 512 * undo buffers are returned to the kernel but left passively associated 513 * so we can track when the kernel writes the bp out. 514 */ 515 struct buf * 516 hammer_io_release(struct hammer_io *io, int flush) 517 { 518 union hammer_io_structure *iou = (void *)io; 519 struct buf *bp; 520 521 if ((bp = io->bp) == NULL) 522 return(NULL); 523 524 /* 525 * Try to flush a dirty IO to disk if asked to by the 526 * caller or if the kernel tried to flush the buffer in the past. 527 * 528 * Kernel-initiated flushes are only allowed for pure-data buffers. 529 * meta-data and volume buffers can only be flushed explicitly 530 * by HAMMER. 531 */ 532 if (io->modified) { 533 if (flush) { 534 hammer_io_flush(io, 0); 535 } else if (bp->b_flags & B_LOCKED) { 536 switch(io->type) { 537 case HAMMER_STRUCTURE_DATA_BUFFER: 538 hammer_io_flush(io, 0); 539 break; 540 case HAMMER_STRUCTURE_UNDO_BUFFER: 541 hammer_io_flush(io, hammer_undo_reclaim(io)); 542 break; 543 default: 544 break; 545 } 546 } /* else no explicit request to flush the buffer */ 547 } 548 549 /* 550 * Wait for the IO to complete if asked to. This occurs when 551 * the buffer must be disposed of definitively during an umount 552 * or buffer invalidation. 553 */ 554 if (io->waitdep && io->running) { 555 hammer_io_wait(io); 556 } 557 558 /* 559 * Return control of the buffer to the kernel (with the provisio 560 * that our bioops can override kernel decisions with regards to 561 * the buffer). 562 */ 563 if ((flush || io->reclaim) && io->modified == 0 && io->running == 0) { 564 /* 565 * Always disassociate the bp if an explicit flush 566 * was requested and the IO completed with no error 567 * (so unmount can really clean up the structure). 568 */ 569 if (io->released) { 570 regetblk(bp); 571 BUF_KERNPROC(bp); 572 } else { 573 io->released = 1; 574 } 575 hammer_io_disassociate((hammer_io_structure_t)io); 576 /* return the bp */ 577 } else if (io->modified) { 578 /* 579 * Only certain IO types can be released to the kernel if 580 * the buffer has been modified. 581 * 582 * volume and meta-data IO types may only be explicitly 583 * flushed by HAMMER. 584 */ 585 switch(io->type) { 586 case HAMMER_STRUCTURE_DATA_BUFFER: 587 case HAMMER_STRUCTURE_UNDO_BUFFER: 588 if (io->released == 0) { 589 io->released = 1; 590 bdwrite(bp); 591 } 592 break; 593 default: 594 break; 595 } 596 bp = NULL; /* bp left associated */ 597 } else if (io->released == 0) { 598 /* 599 * Clean buffers can be generally released to the kernel. 600 * We leave the bp passively associated with the HAMMER 601 * structure and use bioops to disconnect it later on 602 * if the kernel wants to discard the buffer. 603 * 604 * We can steal the structure's ownership of the bp. 605 */ 606 io->released = 1; 607 if (bp->b_flags & B_LOCKED) { 608 hammer_io_disassociate(iou); 609 /* return the bp */ 610 } else { 611 if (io->reclaim) { 612 hammer_io_disassociate(iou); 613 /* return the bp */ 614 } else { 615 /* return the bp (bp passively associated) */ 616 } 617 } 618 } else { 619 /* 620 * A released buffer is passively associate with our 621 * hammer_io structure. The kernel cannot destroy it 622 * without making a bioops call. If the kernel (B_LOCKED) 623 * or we (reclaim) requested that the buffer be destroyed 624 * we destroy it, otherwise we do a quick get/release to 625 * reset its position in the kernel's LRU list. 626 * 627 * Leaving the buffer passively associated allows us to 628 * use the kernel's LRU buffer flushing mechanisms rather 629 * then rolling our own. 630 * 631 * XXX there are two ways of doing this. We can re-acquire 632 * and passively release to reset the LRU, or not. 633 */ 634 if (io->running == 0) { 635 regetblk(bp); 636 if ((bp->b_flags & B_LOCKED) || io->reclaim) { 637 hammer_io_disassociate(iou); 638 /* return the bp */ 639 } else { 640 /* return the bp (bp passively associated) */ 641 } 642 } else { 643 /* 644 * bp is left passively associated but we do not 645 * try to reacquire it. Interactions with the io 646 * structure will occur on completion of the bp's 647 * I/O. 648 */ 649 bp = NULL; 650 } 651 } 652 return(bp); 653 } 654 655 /* 656 * This routine is called with a locked IO when a flush is desired and 657 * no other references to the structure exists other then ours. This 658 * routine is ONLY called when HAMMER believes it is safe to flush a 659 * potentially modified buffer out. 660 * 661 * The locked io or io reference prevents a flush from being initiated 662 * by the kernel. 663 */ 664 void 665 hammer_io_flush(struct hammer_io *io, int reclaim) 666 { 667 struct buf *bp; 668 hammer_mount_t hmp; 669 670 /* 671 * Degenerate case - nothing to flush if nothing is dirty. 672 */ 673 if (io->modified == 0) 674 return; 675 676 KKASSERT(io->bp); 677 KKASSERT(io->modify_refs <= 0); 678 679 /* 680 * Acquire ownership of the bp, particularly before we clear our 681 * modified flag. 682 * 683 * We are going to bawrite() this bp. Don't leave a window where 684 * io->released is set, we actually own the bp rather then our 685 * buffer. 686 * 687 * The io_token should not be required here as only 688 */ 689 hmp = io->hmp; 690 bp = io->bp; 691 if (io->released) { 692 regetblk(bp); 693 /* BUF_KERNPROC(io->bp); */ 694 /* io->released = 0; */ 695 KKASSERT(io->released); 696 KKASSERT(io->bp == bp); 697 } else { 698 io->released = 1; 699 } 700 701 if (reclaim) { 702 io->reclaim = 1; 703 if ((bp->b_flags & B_LOCKED) == 0) { 704 bp->b_flags |= B_LOCKED; 705 atomic_add_int(&hammer_count_io_locked, 1); 706 } 707 } 708 709 /* 710 * Acquire exclusive access to the bp and then clear the modified 711 * state of the buffer prior to issuing I/O to interlock any 712 * modifications made while the I/O is in progress. This shouldn't 713 * happen anyway but losing data would be worse. The modified bit 714 * will be rechecked after the IO completes. 715 * 716 * NOTE: This call also finalizes the buffer's content (inval == 0). 717 * 718 * This is only legal when lock.refs == 1 (otherwise we might clear 719 * the modified bit while there are still users of the cluster 720 * modifying the data). 721 * 722 * Do this before potentially blocking so any attempt to modify the 723 * ondisk while we are blocked blocks waiting for us. 724 */ 725 hammer_ref(&io->lock); 726 hammer_io_clear_modify(io, 0); 727 hammer_rel(&io->lock); 728 729 if (hammer_debug_io & 0x0002) 730 kprintf("hammer io_write %016jx\n", bp->b_bio1.bio_offset); 731 732 /* 733 * Transfer ownership to the kernel and initiate I/O. 734 * 735 * NOTE: We do not hold io_token so an atomic op is required to 736 * update io_running_space. 737 */ 738 io->running = 1; 739 atomic_add_int(&hmp->io_running_space, io->bytes); 740 atomic_add_int(&hammer_count_io_running_write, io->bytes); 741 lwkt_gettoken(&hmp->io_token); 742 TAILQ_INSERT_TAIL(&hmp->iorun_list, io, iorun_entry); 743 lwkt_reltoken(&hmp->io_token); 744 bawrite(bp); 745 hammer_io_flush_mark(io->volume); 746 } 747 748 /************************************************************************ 749 * BUFFER DIRTYING * 750 ************************************************************************ 751 * 752 * These routines deal with dependancies created when IO buffers get 753 * modified. The caller must call hammer_modify_*() on a referenced 754 * HAMMER structure prior to modifying its on-disk data. 755 * 756 * Any intent to modify an IO buffer acquires the related bp and imposes 757 * various write ordering dependancies. 758 */ 759 760 /* 761 * Mark a HAMMER structure as undergoing modification. Meta-data buffers 762 * are locked until the flusher can deal with them, pure data buffers 763 * can be written out. 764 * 765 * The referenced io prevents races. 766 */ 767 static 768 void 769 hammer_io_modify(hammer_io_t io, int count) 770 { 771 /* 772 * io->modify_refs must be >= 0 773 */ 774 while (io->modify_refs < 0) { 775 io->waitmod = 1; 776 tsleep(io, 0, "hmrmod", 0); 777 } 778 779 /* 780 * Shortcut if nothing to do. 781 */ 782 KKASSERT(hammer_isactive(&io->lock) && io->bp != NULL); 783 io->modify_refs += count; 784 if (io->modified && io->released == 0) 785 return; 786 787 /* 788 * NOTE: It is important not to set the modified bit 789 * until after we have acquired the bp or we risk 790 * racing against checkwrite. 791 */ 792 hammer_lock_ex(&io->lock); 793 if (io->released) { 794 regetblk(io->bp); 795 BUF_KERNPROC(io->bp); 796 io->released = 0; 797 } 798 if (io->modified == 0) { 799 hammer_io_set_modlist(io); 800 io->modified = 1; 801 } 802 hammer_unlock(&io->lock); 803 } 804 805 static __inline 806 void 807 hammer_io_modify_done(hammer_io_t io) 808 { 809 KKASSERT(io->modify_refs > 0); 810 --io->modify_refs; 811 if (io->modify_refs == 0 && io->waitmod) { 812 io->waitmod = 0; 813 wakeup(io); 814 } 815 } 816 817 /* 818 * The write interlock blocks other threads trying to modify a buffer 819 * (they block in hammer_io_modify()) after us, or blocks us while other 820 * threads are in the middle of modifying a buffer. 821 * 822 * The caller also has a ref on the io, however if we are not careful 823 * we will race bioops callbacks (checkwrite). To deal with this 824 * we must at least acquire and release the io_token, and it is probably 825 * better to hold it through the setting of modify_refs. 826 */ 827 void 828 hammer_io_write_interlock(hammer_io_t io) 829 { 830 hammer_mount_t hmp = io->hmp; 831 832 lwkt_gettoken(&hmp->io_token); 833 while (io->modify_refs != 0) { 834 io->waitmod = 1; 835 tsleep(io, 0, "hmrmod", 0); 836 } 837 io->modify_refs = -1; 838 lwkt_reltoken(&hmp->io_token); 839 } 840 841 void 842 hammer_io_done_interlock(hammer_io_t io) 843 { 844 KKASSERT(io->modify_refs == -1); 845 io->modify_refs = 0; 846 if (io->waitmod) { 847 io->waitmod = 0; 848 wakeup(io); 849 } 850 } 851 852 /* 853 * Caller intends to modify a volume's ondisk structure. 854 * 855 * This is only allowed if we are the flusher or we have a ref on the 856 * sync_lock. 857 */ 858 void 859 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume, 860 void *base, int len) 861 { 862 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 863 864 hammer_io_modify(&volume->io, 1); 865 if (len) { 866 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk; 867 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 868 hammer_generate_undo(trans, 869 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset), 870 base, len); 871 } 872 } 873 874 /* 875 * Caller intends to modify a buffer's ondisk structure. 876 * 877 * This is only allowed if we are the flusher or we have a ref on the 878 * sync_lock. 879 */ 880 void 881 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer, 882 void *base, int len) 883 { 884 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 885 886 hammer_io_modify(&buffer->io, 1); 887 if (len) { 888 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk; 889 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 890 hammer_generate_undo(trans, 891 buffer->zone2_offset + rel_offset, 892 base, len); 893 } 894 } 895 896 void 897 hammer_modify_volume_done(hammer_volume_t volume) 898 { 899 hammer_io_modify_done(&volume->io); 900 } 901 902 void 903 hammer_modify_buffer_done(hammer_buffer_t buffer) 904 { 905 hammer_io_modify_done(&buffer->io); 906 } 907 908 /* 909 * Mark an entity as not being dirty any more and finalize any 910 * delayed adjustments to the buffer. 911 * 912 * Delayed adjustments are an important performance enhancement, allowing 913 * us to avoid recalculating B-Tree node CRCs over and over again when 914 * making bulk-modifications to the B-Tree. 915 * 916 * If inval is non-zero delayed adjustments are ignored. 917 * 918 * This routine may dereference related btree nodes and cause the 919 * buffer to be dereferenced. The caller must own a reference on io. 920 */ 921 void 922 hammer_io_clear_modify(struct hammer_io *io, int inval) 923 { 924 hammer_mount_t hmp; 925 926 /* 927 * io_token is needed to avoid races on mod_root 928 */ 929 if (io->modified == 0) 930 return; 931 hmp = io->hmp; 932 lwkt_gettoken(&hmp->io_token); 933 if (io->modified == 0) { 934 lwkt_reltoken(&hmp->io_token); 935 return; 936 } 937 938 /* 939 * Take us off the mod-list and clear the modified bit. 940 */ 941 KKASSERT(io->mod_root != NULL); 942 if (io->mod_root == &io->hmp->volu_root || 943 io->mod_root == &io->hmp->meta_root) { 944 io->hmp->locked_dirty_space -= io->bytes; 945 atomic_add_int(&hammer_count_dirtybufspace, -io->bytes); 946 } 947 RB_REMOVE(hammer_mod_rb_tree, io->mod_root, io); 948 io->mod_root = NULL; 949 io->modified = 0; 950 951 lwkt_reltoken(&hmp->io_token); 952 953 /* 954 * If this bit is not set there are no delayed adjustments. 955 */ 956 if (io->gencrc == 0) 957 return; 958 io->gencrc = 0; 959 960 /* 961 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference 962 * on the node (& underlying buffer). Release the node after clearing 963 * the flag. 964 */ 965 if (io->type == HAMMER_STRUCTURE_META_BUFFER) { 966 hammer_buffer_t buffer = (void *)io; 967 hammer_node_t node; 968 969 restart: 970 TAILQ_FOREACH(node, &buffer->clist, entry) { 971 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) 972 continue; 973 node->flags &= ~HAMMER_NODE_NEEDSCRC; 974 KKASSERT(node->ondisk); 975 if (inval == 0) 976 node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE); 977 hammer_rel_node(node); 978 goto restart; 979 } 980 } 981 /* caller must still have ref on io */ 982 KKASSERT(hammer_isactive(&io->lock)); 983 } 984 985 /* 986 * Clear the IO's modify list. Even though the IO is no longer modified 987 * it may still be on the lose_root. This routine is called just before 988 * the governing hammer_buffer is destroyed. 989 * 990 * mod_root requires io_token protection. 991 */ 992 void 993 hammer_io_clear_modlist(struct hammer_io *io) 994 { 995 hammer_mount_t hmp = io->hmp; 996 997 KKASSERT(io->modified == 0); 998 if (io->mod_root) { 999 lwkt_gettoken(&hmp->io_token); 1000 if (io->mod_root) { 1001 KKASSERT(io->mod_root == &io->hmp->lose_root); 1002 RB_REMOVE(hammer_mod_rb_tree, io->mod_root, io); 1003 io->mod_root = NULL; 1004 } 1005 lwkt_reltoken(&hmp->io_token); 1006 } 1007 } 1008 1009 static void 1010 hammer_io_set_modlist(struct hammer_io *io) 1011 { 1012 struct hammer_mount *hmp = io->hmp; 1013 1014 lwkt_gettoken(&hmp->io_token); 1015 KKASSERT(io->mod_root == NULL); 1016 1017 switch(io->type) { 1018 case HAMMER_STRUCTURE_VOLUME: 1019 io->mod_root = &hmp->volu_root; 1020 hmp->locked_dirty_space += io->bytes; 1021 atomic_add_int(&hammer_count_dirtybufspace, io->bytes); 1022 break; 1023 case HAMMER_STRUCTURE_META_BUFFER: 1024 io->mod_root = &hmp->meta_root; 1025 hmp->locked_dirty_space += io->bytes; 1026 atomic_add_int(&hammer_count_dirtybufspace, io->bytes); 1027 break; 1028 case HAMMER_STRUCTURE_UNDO_BUFFER: 1029 io->mod_root = &hmp->undo_root; 1030 break; 1031 case HAMMER_STRUCTURE_DATA_BUFFER: 1032 io->mod_root = &hmp->data_root; 1033 break; 1034 case HAMMER_STRUCTURE_DUMMY: 1035 panic("hammer_io_set_modlist: bad io type"); 1036 break; /* NOT REACHED */ 1037 } 1038 if (RB_INSERT(hammer_mod_rb_tree, io->mod_root, io)) { 1039 panic("hammer_io_set_modlist: duplicate entry"); 1040 /* NOT REACHED */ 1041 } 1042 lwkt_reltoken(&hmp->io_token); 1043 } 1044 1045 /************************************************************************ 1046 * HAMMER_BIOOPS * 1047 ************************************************************************ 1048 * 1049 */ 1050 1051 /* 1052 * Pre-IO initiation kernel callback - cluster build only 1053 * 1054 * bioops callback - hold io_token 1055 */ 1056 static void 1057 hammer_io_start(struct buf *bp) 1058 { 1059 /* nothing to do, so io_token not needed */ 1060 } 1061 1062 /* 1063 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT! 1064 * 1065 * NOTE: HAMMER may modify a data buffer after we have initiated write 1066 * I/O. 1067 * 1068 * NOTE: MPSAFE callback 1069 * 1070 * bioops callback - hold io_token 1071 */ 1072 static void 1073 hammer_io_complete(struct buf *bp) 1074 { 1075 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep); 1076 struct hammer_mount *hmp = iou->io.hmp; 1077 struct hammer_io *ionext; 1078 1079 lwkt_gettoken(&hmp->io_token); 1080 1081 KKASSERT(iou->io.released == 1); 1082 1083 /* 1084 * Deal with people waiting for I/O to drain 1085 */ 1086 if (iou->io.running) { 1087 /* 1088 * Deal with critical write errors. Once a critical error 1089 * has been flagged in hmp the UNDO FIFO will not be updated. 1090 * That way crash recover will give us a consistent 1091 * filesystem. 1092 * 1093 * Because of this we can throw away failed UNDO buffers. If 1094 * we throw away META or DATA buffers we risk corrupting 1095 * the now read-only version of the filesystem visible to 1096 * the user. Clear B_ERROR so the buffer is not re-dirtied 1097 * by the kernel and ref the io so it doesn't get thrown 1098 * away. 1099 */ 1100 if (bp->b_flags & B_ERROR) { 1101 lwkt_gettoken(&hmp->fs_token); 1102 hammer_critical_error(hmp, NULL, bp->b_error, 1103 "while flushing meta-data"); 1104 lwkt_reltoken(&hmp->fs_token); 1105 1106 switch(iou->io.type) { 1107 case HAMMER_STRUCTURE_UNDO_BUFFER: 1108 break; 1109 default: 1110 if (iou->io.ioerror == 0) { 1111 iou->io.ioerror = 1; 1112 hammer_ref(&iou->io.lock); 1113 } 1114 break; 1115 } 1116 bp->b_flags &= ~B_ERROR; 1117 bundirty(bp); 1118 #if 0 1119 hammer_io_set_modlist(&iou->io); 1120 iou->io.modified = 1; 1121 #endif 1122 } 1123 hammer_stats_disk_write += iou->io.bytes; 1124 atomic_add_int(&hammer_count_io_running_write, -iou->io.bytes); 1125 atomic_add_int(&hmp->io_running_space, -iou->io.bytes); 1126 if (hmp->io_running_wakeup && 1127 hmp->io_running_space < hammer_limit_running_io / 2) { 1128 hmp->io_running_wakeup = 0; 1129 wakeup(&hmp->io_running_wakeup); 1130 } 1131 KKASSERT(hmp->io_running_space >= 0); 1132 iou->io.running = 0; 1133 1134 /* 1135 * Remove from iorun list and wakeup any multi-io waiter(s). 1136 */ 1137 if (TAILQ_FIRST(&hmp->iorun_list) == &iou->io) { 1138 ionext = TAILQ_NEXT(&iou->io, iorun_entry); 1139 if (ionext && ionext->type == HAMMER_STRUCTURE_DUMMY) 1140 wakeup(ionext); 1141 } 1142 TAILQ_REMOVE(&hmp->iorun_list, &iou->io, iorun_entry); 1143 } else { 1144 hammer_stats_disk_read += iou->io.bytes; 1145 } 1146 1147 if (iou->io.waiting) { 1148 iou->io.waiting = 0; 1149 wakeup(iou); 1150 } 1151 1152 /* 1153 * If B_LOCKED is set someone wanted to deallocate the bp at some 1154 * point, try to do it now. The operation will fail if there are 1155 * refs or if hammer_io_deallocate() is unable to gain the 1156 * interlock. 1157 */ 1158 if (bp->b_flags & B_LOCKED) { 1159 atomic_add_int(&hammer_count_io_locked, -1); 1160 bp->b_flags &= ~B_LOCKED; 1161 hammer_io_deallocate(bp); 1162 /* structure may be dead now */ 1163 } 1164 lwkt_reltoken(&hmp->io_token); 1165 } 1166 1167 /* 1168 * Callback from kernel when it wishes to deallocate a passively 1169 * associated structure. This mostly occurs with clean buffers 1170 * but it may be possible for a holding structure to be marked dirty 1171 * while its buffer is passively associated. The caller owns the bp. 1172 * 1173 * If we cannot disassociate we set B_LOCKED to prevent the buffer 1174 * from getting reused. 1175 * 1176 * WARNING: Because this can be called directly by getnewbuf we cannot 1177 * recurse into the tree. If a bp cannot be immediately disassociated 1178 * our only recourse is to set B_LOCKED. 1179 * 1180 * WARNING: This may be called from an interrupt via hammer_io_complete() 1181 * 1182 * bioops callback - hold io_token 1183 */ 1184 static void 1185 hammer_io_deallocate(struct buf *bp) 1186 { 1187 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep); 1188 hammer_mount_t hmp; 1189 1190 hmp = iou->io.hmp; 1191 1192 lwkt_gettoken(&hmp->io_token); 1193 1194 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0); 1195 if (hammer_try_interlock_norefs(&iou->io.lock) == 0) { 1196 /* 1197 * We cannot safely disassociate a bp from a referenced 1198 * or interlocked HAMMER structure. 1199 */ 1200 bp->b_flags |= B_LOCKED; 1201 atomic_add_int(&hammer_count_io_locked, 1); 1202 } else if (iou->io.modified) { 1203 /* 1204 * It is not legal to disassociate a modified buffer. This 1205 * case really shouldn't ever occur. 1206 */ 1207 bp->b_flags |= B_LOCKED; 1208 atomic_add_int(&hammer_count_io_locked, 1); 1209 hammer_put_interlock(&iou->io.lock, 0); 1210 } else { 1211 /* 1212 * Disassociate the BP. If the io has no refs left we 1213 * have to add it to the loose list. The kernel has 1214 * locked the buffer and therefore our io must be 1215 * in a released state. 1216 */ 1217 hammer_io_disassociate(iou); 1218 if (iou->io.type != HAMMER_STRUCTURE_VOLUME) { 1219 KKASSERT(iou->io.bp == NULL); 1220 KKASSERT(iou->io.mod_root == NULL); 1221 iou->io.mod_root = &hmp->lose_root; 1222 if (RB_INSERT(hammer_mod_rb_tree, iou->io.mod_root, 1223 &iou->io)) { 1224 panic("hammer_io_deallocate: duplicate entry"); 1225 } 1226 } 1227 hammer_put_interlock(&iou->io.lock, 1); 1228 } 1229 lwkt_reltoken(&hmp->io_token); 1230 } 1231 1232 /* 1233 * bioops callback - hold io_token 1234 */ 1235 static int 1236 hammer_io_fsync(struct vnode *vp) 1237 { 1238 /* nothing to do, so io_token not needed */ 1239 return(0); 1240 } 1241 1242 /* 1243 * NOTE: will not be called unless we tell the kernel about the 1244 * bioops. Unused... we use the mount's VFS_SYNC instead. 1245 * 1246 * bioops callback - hold io_token 1247 */ 1248 static int 1249 hammer_io_sync(struct mount *mp) 1250 { 1251 /* nothing to do, so io_token not needed */ 1252 return(0); 1253 } 1254 1255 /* 1256 * bioops callback - hold io_token 1257 */ 1258 static void 1259 hammer_io_movedeps(struct buf *bp1, struct buf *bp2) 1260 { 1261 /* nothing to do, so io_token not needed */ 1262 } 1263 1264 /* 1265 * I/O pre-check for reading and writing. HAMMER only uses this for 1266 * B_CACHE buffers so checkread just shouldn't happen, but if it does 1267 * allow it. 1268 * 1269 * Writing is a different case. We don't want the kernel to try to write 1270 * out a buffer that HAMMER may be modifying passively or which has a 1271 * dependancy. In addition, kernel-demanded writes can only proceed for 1272 * certain types of buffers (i.e. UNDO and DATA types). Other dirty 1273 * buffer types can only be explicitly written by the flusher. 1274 * 1275 * checkwrite will only be called for bdwrite()n buffers. If we return 1276 * success the kernel is guaranteed to initiate the buffer write. 1277 * 1278 * bioops callback - hold io_token 1279 */ 1280 static int 1281 hammer_io_checkread(struct buf *bp) 1282 { 1283 /* nothing to do, so io_token not needed */ 1284 return(0); 1285 } 1286 1287 /* 1288 * The kernel is asking us whether it can write out a dirty buffer or not. 1289 * 1290 * bioops callback - hold io_token 1291 */ 1292 static int 1293 hammer_io_checkwrite(struct buf *bp) 1294 { 1295 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep); 1296 hammer_mount_t hmp = io->hmp; 1297 1298 /* 1299 * This shouldn't happen under normal operation. 1300 */ 1301 lwkt_gettoken(&hmp->io_token); 1302 if (io->type == HAMMER_STRUCTURE_VOLUME || 1303 io->type == HAMMER_STRUCTURE_META_BUFFER) { 1304 if (!panicstr) 1305 panic("hammer_io_checkwrite: illegal buffer"); 1306 if ((bp->b_flags & B_LOCKED) == 0) { 1307 bp->b_flags |= B_LOCKED; 1308 atomic_add_int(&hammer_count_io_locked, 1); 1309 } 1310 lwkt_reltoken(&hmp->io_token); 1311 return(1); 1312 } 1313 1314 /* 1315 * We have to be able to interlock the IO to safely modify any 1316 * of its fields without holding the fs_token. If we can't lock 1317 * it then we are racing someone. 1318 * 1319 * Our ownership of the bp lock prevents the io from being ripped 1320 * out from under us. 1321 */ 1322 if (hammer_try_interlock_norefs(&io->lock) == 0) { 1323 bp->b_flags |= B_LOCKED; 1324 atomic_add_int(&hammer_count_io_locked, 1); 1325 lwkt_reltoken(&hmp->io_token); 1326 return(1); 1327 } 1328 1329 /* 1330 * The modified bit must be cleared prior to the initiation of 1331 * any IO (returning 0 initiates the IO). Because this is a 1332 * normal data buffer hammer_io_clear_modify() runs through a 1333 * simple degenerate case. 1334 * 1335 * Return 0 will cause the kernel to initiate the IO, and we 1336 * must normally clear the modified bit before we begin. If 1337 * the io has modify_refs we do not clear the modified bit, 1338 * otherwise we may miss changes. 1339 * 1340 * Only data and undo buffers can reach here. These buffers do 1341 * not have terminal crc functions but we temporarily reference 1342 * the IO anyway, just in case. 1343 */ 1344 if (io->modify_refs == 0 && io->modified) { 1345 hammer_ref(&io->lock); 1346 hammer_io_clear_modify(io, 0); 1347 hammer_rel(&io->lock); 1348 } else if (io->modified) { 1349 KKASSERT(io->type == HAMMER_STRUCTURE_DATA_BUFFER); 1350 } 1351 1352 /* 1353 * The kernel is going to start the IO, set io->running. 1354 */ 1355 KKASSERT(io->running == 0); 1356 io->running = 1; 1357 atomic_add_int(&io->hmp->io_running_space, io->bytes); 1358 atomic_add_int(&hammer_count_io_running_write, io->bytes); 1359 TAILQ_INSERT_TAIL(&io->hmp->iorun_list, io, iorun_entry); 1360 1361 hammer_put_interlock(&io->lock, 1); 1362 lwkt_reltoken(&hmp->io_token); 1363 1364 return(0); 1365 } 1366 1367 /* 1368 * Return non-zero if we wish to delay the kernel's attempt to flush 1369 * this buffer to disk. 1370 * 1371 * bioops callback - hold io_token 1372 */ 1373 static int 1374 hammer_io_countdeps(struct buf *bp, int n) 1375 { 1376 /* nothing to do, so io_token not needed */ 1377 return(0); 1378 } 1379 1380 struct bio_ops hammer_bioops = { 1381 .io_start = hammer_io_start, 1382 .io_complete = hammer_io_complete, 1383 .io_deallocate = hammer_io_deallocate, 1384 .io_fsync = hammer_io_fsync, 1385 .io_sync = hammer_io_sync, 1386 .io_movedeps = hammer_io_movedeps, 1387 .io_countdeps = hammer_io_countdeps, 1388 .io_checkread = hammer_io_checkread, 1389 .io_checkwrite = hammer_io_checkwrite, 1390 }; 1391 1392 /************************************************************************ 1393 * DIRECT IO OPS * 1394 ************************************************************************ 1395 * 1396 * These functions operate directly on the buffer cache buffer associated 1397 * with a front-end vnode rather then a back-end device vnode. 1398 */ 1399 1400 /* 1401 * Read a buffer associated with a front-end vnode directly from the 1402 * disk media. The bio may be issued asynchronously. If leaf is non-NULL 1403 * we validate the CRC. 1404 * 1405 * We must check for the presence of a HAMMER buffer to handle the case 1406 * where the reblocker has rewritten the data (which it does via the HAMMER 1407 * buffer system, not via the high-level vnode buffer cache), but not yet 1408 * committed the buffer to the media. 1409 */ 1410 int 1411 hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio, 1412 hammer_btree_leaf_elm_t leaf) 1413 { 1414 hammer_off_t buf_offset; 1415 hammer_off_t zone2_offset; 1416 hammer_volume_t volume; 1417 struct buf *bp; 1418 struct bio *nbio; 1419 int vol_no; 1420 int error; 1421 1422 buf_offset = bio->bio_offset; 1423 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) == 1424 HAMMER_ZONE_LARGE_DATA); 1425 1426 /* 1427 * The buffer cache may have an aliased buffer (the reblocker can 1428 * write them). If it does we have to sync any dirty data before 1429 * we can build our direct-read. This is a non-critical code path. 1430 */ 1431 bp = bio->bio_buf; 1432 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize); 1433 1434 /* 1435 * Resolve to a zone-2 offset. The conversion just requires 1436 * munging the top 4 bits but we want to abstract it anyway 1437 * so the blockmap code can verify the zone assignment. 1438 */ 1439 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1440 if (error) 1441 goto done; 1442 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) == 1443 HAMMER_ZONE_RAW_BUFFER); 1444 1445 /* 1446 * Resolve volume and raw-offset for 3rd level bio. The 1447 * offset will be specific to the volume. 1448 */ 1449 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1450 volume = hammer_get_volume(hmp, vol_no, &error); 1451 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1452 error = EIO; 1453 1454 if (error == 0) { 1455 /* 1456 * 3rd level bio 1457 */ 1458 nbio = push_bio(bio); 1459 nbio->bio_offset = volume->ondisk->vol_buf_beg + 1460 (zone2_offset & HAMMER_OFF_SHORT_MASK); 1461 #if 0 1462 /* 1463 * XXX disabled - our CRC check doesn't work if the OS 1464 * does bogus_page replacement on the direct-read. 1465 */ 1466 if (leaf && hammer_verify_data) { 1467 nbio->bio_done = hammer_io_direct_read_complete; 1468 nbio->bio_caller_info1.uvalue32 = leaf->data_crc; 1469 } 1470 #endif 1471 hammer_stats_disk_read += bp->b_bufsize; 1472 vn_strategy(volume->devvp, nbio); 1473 } 1474 hammer_rel_volume(volume, 0); 1475 done: 1476 if (error) { 1477 kprintf("hammer_direct_read: failed @ %016llx\n", 1478 (long long)zone2_offset); 1479 bp->b_error = error; 1480 bp->b_flags |= B_ERROR; 1481 biodone(bio); 1482 } 1483 return(error); 1484 } 1485 1486 #if 0 1487 /* 1488 * On completion of the BIO this callback must check the data CRC 1489 * and chain to the previous bio. 1490 * 1491 * MPSAFE - since we do not modify and hammer_records we do not need 1492 * io_token. 1493 * 1494 * NOTE: MPSAFE callback 1495 */ 1496 static 1497 void 1498 hammer_io_direct_read_complete(struct bio *nbio) 1499 { 1500 struct bio *obio; 1501 struct buf *bp; 1502 u_int32_t rec_crc = nbio->bio_caller_info1.uvalue32; 1503 1504 bp = nbio->bio_buf; 1505 if (crc32(bp->b_data, bp->b_bufsize) != rec_crc) { 1506 kprintf("HAMMER: data_crc error @%016llx/%d\n", 1507 nbio->bio_offset, bp->b_bufsize); 1508 if (hammer_debug_critical) 1509 Debugger("data_crc on read"); 1510 bp->b_flags |= B_ERROR; 1511 bp->b_error = EIO; 1512 } 1513 obio = pop_bio(nbio); 1514 biodone(obio); 1515 } 1516 #endif 1517 1518 /* 1519 * Write a buffer associated with a front-end vnode directly to the 1520 * disk media. The bio may be issued asynchronously. 1521 * 1522 * The BIO is associated with the specified record and RECG_DIRECT_IO 1523 * is set. The recorded is added to its object. 1524 */ 1525 int 1526 hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio, 1527 hammer_record_t record) 1528 { 1529 hammer_btree_leaf_elm_t leaf = &record->leaf; 1530 hammer_off_t buf_offset; 1531 hammer_off_t zone2_offset; 1532 hammer_volume_t volume; 1533 hammer_buffer_t buffer; 1534 struct buf *bp; 1535 struct bio *nbio; 1536 char *ptr; 1537 int vol_no; 1538 int error; 1539 1540 buf_offset = leaf->data_offset; 1541 1542 KKASSERT(buf_offset > HAMMER_ZONE_BTREE); 1543 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE); 1544 1545 /* 1546 * Issue or execute the I/O. The new memory record must replace 1547 * the old one before the I/O completes, otherwise a reaquisition of 1548 * the buffer will load the old media data instead of the new. 1549 */ 1550 if ((buf_offset & HAMMER_BUFMASK) == 0 && 1551 leaf->data_len >= HAMMER_BUFSIZE) { 1552 /* 1553 * We are using the vnode's bio to write directly to the 1554 * media, any hammer_buffer at the same zone-X offset will 1555 * now have stale data. 1556 */ 1557 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1558 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1559 volume = hammer_get_volume(hmp, vol_no, &error); 1560 1561 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1562 error = EIO; 1563 if (error == 0) { 1564 bp = bio->bio_buf; 1565 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0); 1566 /* 1567 hammer_del_buffers(hmp, buf_offset, 1568 zone2_offset, bp->b_bufsize); 1569 */ 1570 1571 /* 1572 * Second level bio - cached zone2 offset. 1573 * 1574 * (We can put our bio_done function in either the 1575 * 2nd or 3rd level). 1576 */ 1577 nbio = push_bio(bio); 1578 nbio->bio_offset = zone2_offset; 1579 nbio->bio_done = hammer_io_direct_write_complete; 1580 nbio->bio_caller_info1.ptr = record; 1581 record->zone2_offset = zone2_offset; 1582 record->gflags |= HAMMER_RECG_DIRECT_IO | 1583 HAMMER_RECG_DIRECT_INVAL; 1584 1585 /* 1586 * Third level bio - raw offset specific to the 1587 * correct volume. 1588 */ 1589 zone2_offset &= HAMMER_OFF_SHORT_MASK; 1590 nbio = push_bio(nbio); 1591 nbio->bio_offset = volume->ondisk->vol_buf_beg + 1592 zone2_offset; 1593 hammer_stats_disk_write += bp->b_bufsize; 1594 hammer_ip_replace_bulk(hmp, record); 1595 vn_strategy(volume->devvp, nbio); 1596 hammer_io_flush_mark(volume); 1597 } 1598 hammer_rel_volume(volume, 0); 1599 } else { 1600 /* 1601 * Must fit in a standard HAMMER buffer. In this case all 1602 * consumers use the HAMMER buffer system and RECG_DIRECT_IO 1603 * does not need to be set-up. 1604 */ 1605 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0); 1606 buffer = NULL; 1607 ptr = hammer_bread(hmp, buf_offset, &error, &buffer); 1608 if (error == 0) { 1609 bp = bio->bio_buf; 1610 bp->b_flags |= B_AGE; 1611 hammer_io_modify(&buffer->io, 1); 1612 bcopy(bp->b_data, ptr, leaf->data_len); 1613 hammer_io_modify_done(&buffer->io); 1614 hammer_rel_buffer(buffer, 0); 1615 bp->b_resid = 0; 1616 hammer_ip_replace_bulk(hmp, record); 1617 biodone(bio); 1618 } 1619 } 1620 if (error) { 1621 /* 1622 * Major suckage occured. Also note: The record was 1623 * never added to the tree so we do not have to worry 1624 * about the backend. 1625 */ 1626 kprintf("hammer_direct_write: failed @ %016llx\n", 1627 (long long)leaf->data_offset); 1628 bp = bio->bio_buf; 1629 bp->b_resid = 0; 1630 bp->b_error = EIO; 1631 bp->b_flags |= B_ERROR; 1632 biodone(bio); 1633 record->flags |= HAMMER_RECF_DELETED_FE; 1634 hammer_rel_mem_record(record); 1635 } 1636 return(error); 1637 } 1638 1639 /* 1640 * On completion of the BIO this callback must disconnect 1641 * it from the hammer_record and chain to the previous bio. 1642 * 1643 * An I/O error forces the mount to read-only. Data buffers 1644 * are not B_LOCKED like meta-data buffers are, so we have to 1645 * throw the buffer away to prevent the kernel from retrying. 1646 * 1647 * NOTE: MPSAFE callback, only modify fields we have explicit 1648 * access to (the bp and the record->gflags). 1649 */ 1650 static 1651 void 1652 hammer_io_direct_write_complete(struct bio *nbio) 1653 { 1654 struct bio *obio; 1655 struct buf *bp; 1656 hammer_record_t record; 1657 hammer_mount_t hmp; 1658 1659 record = nbio->bio_caller_info1.ptr; 1660 KKASSERT(record != NULL); 1661 hmp = record->ip->hmp; 1662 1663 lwkt_gettoken(&hmp->io_token); 1664 1665 bp = nbio->bio_buf; 1666 obio = pop_bio(nbio); 1667 if (bp->b_flags & B_ERROR) { 1668 lwkt_gettoken(&hmp->fs_token); 1669 hammer_critical_error(hmp, record->ip, 1670 bp->b_error, 1671 "while writing bulk data"); 1672 lwkt_reltoken(&hmp->fs_token); 1673 bp->b_flags |= B_INVAL; 1674 } 1675 biodone(obio); 1676 1677 KKASSERT(record->gflags & HAMMER_RECG_DIRECT_IO); 1678 if (record->gflags & HAMMER_RECG_DIRECT_WAIT) { 1679 record->gflags &= ~(HAMMER_RECG_DIRECT_IO | 1680 HAMMER_RECG_DIRECT_WAIT); 1681 /* record can disappear once DIRECT_IO flag is cleared */ 1682 wakeup(&record->flags); 1683 } else { 1684 record->gflags &= ~HAMMER_RECG_DIRECT_IO; 1685 /* record can disappear once DIRECT_IO flag is cleared */ 1686 } 1687 lwkt_reltoken(&hmp->io_token); 1688 } 1689 1690 1691 /* 1692 * This is called before a record is either committed to the B-Tree 1693 * or destroyed, to resolve any associated direct-IO. 1694 * 1695 * (1) We must wait for any direct-IO related to the record to complete. 1696 * 1697 * (2) We must remove any buffer cache aliases for data accessed via 1698 * leaf->data_offset or zone2_offset so non-direct-IO consumers 1699 * (the mirroring and reblocking code) do not see stale data. 1700 */ 1701 void 1702 hammer_io_direct_wait(hammer_record_t record) 1703 { 1704 hammer_mount_t hmp = record->ip->hmp; 1705 1706 /* 1707 * Wait for I/O to complete 1708 */ 1709 if (record->gflags & HAMMER_RECG_DIRECT_IO) { 1710 lwkt_gettoken(&hmp->io_token); 1711 while (record->gflags & HAMMER_RECG_DIRECT_IO) { 1712 record->gflags |= HAMMER_RECG_DIRECT_WAIT; 1713 tsleep(&record->flags, 0, "hmdiow", 0); 1714 } 1715 lwkt_reltoken(&hmp->io_token); 1716 } 1717 1718 /* 1719 * Invalidate any related buffer cache aliases associated with the 1720 * backing device. This is needed because the buffer cache buffer 1721 * for file data is associated with the file vnode, not the backing 1722 * device vnode. 1723 * 1724 * XXX I do not think this case can occur any more now that 1725 * reservations ensure that all such buffers are removed before 1726 * an area can be reused. 1727 */ 1728 if (record->gflags & HAMMER_RECG_DIRECT_INVAL) { 1729 KKASSERT(record->leaf.data_offset); 1730 hammer_del_buffers(hmp, record->leaf.data_offset, 1731 record->zone2_offset, record->leaf.data_len, 1732 1); 1733 record->gflags &= ~HAMMER_RECG_DIRECT_INVAL; 1734 } 1735 } 1736 1737 /* 1738 * This is called to remove the second-level cached zone-2 offset from 1739 * frontend buffer cache buffers, now stale due to a data relocation. 1740 * These offsets are generated by cluster_read() via VOP_BMAP, or directly 1741 * by hammer_vop_strategy_read(). 1742 * 1743 * This is rather nasty because here we have something like the reblocker 1744 * scanning the raw B-Tree with no held references on anything, really, 1745 * other then a shared lock on the B-Tree node, and we have to access the 1746 * frontend's buffer cache to check for and clean out the association. 1747 * Specifically, if the reblocker is moving data on the disk, these cached 1748 * offsets will become invalid. 1749 * 1750 * Only data record types associated with the large-data zone are subject 1751 * to direct-io and need to be checked. 1752 * 1753 */ 1754 void 1755 hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf) 1756 { 1757 struct hammer_inode_info iinfo; 1758 int zone; 1759 1760 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA) 1761 return; 1762 zone = HAMMER_ZONE_DECODE(leaf->data_offset); 1763 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX) 1764 return; 1765 iinfo.obj_id = leaf->base.obj_id; 1766 iinfo.obj_asof = 0; /* unused */ 1767 iinfo.obj_localization = leaf->base.localization & 1768 HAMMER_LOCALIZE_PSEUDOFS_MASK; 1769 iinfo.u.leaf = leaf; 1770 hammer_scan_inode_snapshots(hmp, &iinfo, 1771 hammer_io_direct_uncache_callback, 1772 leaf); 1773 } 1774 1775 static int 1776 hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data) 1777 { 1778 hammer_inode_info_t iinfo = data; 1779 hammer_off_t data_offset; 1780 hammer_off_t file_offset; 1781 struct vnode *vp; 1782 struct buf *bp; 1783 int blksize; 1784 1785 if (ip->vp == NULL) 1786 return(0); 1787 data_offset = iinfo->u.leaf->data_offset; 1788 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len; 1789 blksize = iinfo->u.leaf->data_len; 1790 KKASSERT((blksize & HAMMER_BUFMASK) == 0); 1791 1792 /* 1793 * Warning: FINDBLK_TEST return stable storage but not stable 1794 * contents. It happens to be ok in this case. 1795 */ 1796 hammer_ref(&ip->lock); 1797 if (hammer_get_vnode(ip, &vp) == 0) { 1798 if ((bp = findblk(ip->vp, file_offset, FINDBLK_TEST)) != NULL && 1799 bp->b_bio2.bio_offset != NOOFFSET) { 1800 bp = getblk(ip->vp, file_offset, blksize, 0, 0); 1801 bp->b_bio2.bio_offset = NOOFFSET; 1802 brelse(bp); 1803 } 1804 vput(vp); 1805 } 1806 hammer_rel_inode(ip, 0); 1807 return(0); 1808 } 1809 1810 1811 /* 1812 * This function is called when writes may have occured on the volume, 1813 * indicating that the device may be holding cached writes. 1814 */ 1815 static void 1816 hammer_io_flush_mark(hammer_volume_t volume) 1817 { 1818 atomic_set_int(&volume->vol_flags, HAMMER_VOLF_NEEDFLUSH); 1819 } 1820 1821 /* 1822 * This function ensures that the device has flushed any cached writes out. 1823 */ 1824 void 1825 hammer_io_flush_sync(hammer_mount_t hmp) 1826 { 1827 hammer_volume_t volume; 1828 struct buf *bp_base = NULL; 1829 struct buf *bp; 1830 1831 RB_FOREACH(volume, hammer_vol_rb_tree, &hmp->rb_vols_root) { 1832 if (volume->vol_flags & HAMMER_VOLF_NEEDFLUSH) { 1833 atomic_clear_int(&volume->vol_flags, 1834 HAMMER_VOLF_NEEDFLUSH); 1835 bp = getpbuf(NULL); 1836 bp->b_bio1.bio_offset = 0; 1837 bp->b_bufsize = 0; 1838 bp->b_bcount = 0; 1839 bp->b_cmd = BUF_CMD_FLUSH; 1840 bp->b_bio1.bio_caller_info1.cluster_head = bp_base; 1841 bp->b_bio1.bio_done = biodone_sync; 1842 bp->b_bio1.bio_flags |= BIO_SYNC; 1843 bp_base = bp; 1844 vn_strategy(volume->devvp, &bp->b_bio1); 1845 } 1846 } 1847 while ((bp = bp_base) != NULL) { 1848 bp_base = bp->b_bio1.bio_caller_info1.cluster_head; 1849 biowait(&bp->b_bio1, "hmrFLS"); 1850 relpbuf(bp, NULL); 1851 } 1852 } 1853 1854 /* 1855 * Limit the amount of backlog which we allow to build up 1856 */ 1857 void 1858 hammer_io_limit_backlog(hammer_mount_t hmp) 1859 { 1860 while (hmp->io_running_space > hammer_limit_running_io) { 1861 hmp->io_running_wakeup = 1; 1862 tsleep(&hmp->io_running_wakeup, 0, "hmiolm", hz / 10); 1863 } 1864 } 1865