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 <sys/buf2.h> 52 53 #include "hammer.h" 54 55 static void hammer_io_modify(hammer_io_t io, int count); 56 static void hammer_io_deallocate(struct buf *bp); 57 static void hammer_indirect_callback(struct bio *bio); 58 static void hammer_io_direct_write_complete(struct bio *nbio); 59 static int hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data); 60 static void hammer_io_set_modlist(hammer_io_t io); 61 static __inline void hammer_io_flush_mark(hammer_volume_t volume); 62 static struct bio_ops hammer_bioops; 63 64 static int 65 hammer_mod_rb_compare(hammer_io_t io1, hammer_io_t io2) 66 { 67 hammer_off_t io1_offset; 68 hammer_off_t io2_offset; 69 70 /* 71 * Encoded offsets are neither valid block device offsets 72 * nor valid zone-X offsets. 73 */ 74 io1_offset = HAMMER_ENCODE(0, io1->volume->vol_no, io1->offset); 75 io2_offset = HAMMER_ENCODE(0, io2->volume->vol_no, io2->offset); 76 77 if (io1_offset < io2_offset) 78 return(-1); 79 if (io1_offset > io2_offset) 80 return(1); 81 return(0); 82 } 83 84 RB_GENERATE(hammer_mod_rb_tree, hammer_io, rb_node, hammer_mod_rb_compare); 85 86 /* 87 * Initialize a new, already-zero'd hammer_io structure, or reinitialize 88 * an existing hammer_io structure which may have switched to another type. 89 */ 90 void 91 hammer_io_init(hammer_io_t io, hammer_volume_t volume, hammer_io_type_t type) 92 { 93 io->volume = volume; 94 io->hmp = volume->io.hmp; 95 io->type = type; 96 } 97 98 hammer_io_type_t 99 hammer_zone_to_iotype(int zone) 100 { 101 hammer_io_type_t iotype; 102 103 switch(zone) { 104 case HAMMER_ZONE_RAW_VOLUME_INDEX: 105 iotype = HAMMER_IOTYPE_VOLUME; 106 break; 107 case HAMMER_ZONE_RAW_BUFFER_INDEX: 108 case HAMMER_ZONE_FREEMAP_INDEX: 109 case HAMMER_ZONE_BTREE_INDEX: 110 case HAMMER_ZONE_META_INDEX: 111 iotype = HAMMER_IOTYPE_META_BUFFER; 112 break; 113 case HAMMER_ZONE_UNDO_INDEX: 114 iotype = HAMMER_IOTYPE_UNDO_BUFFER; 115 break; 116 case HAMMER_ZONE_LARGE_DATA_INDEX: 117 case HAMMER_ZONE_SMALL_DATA_INDEX: 118 iotype = HAMMER_IOTYPE_DATA_BUFFER; 119 break; 120 default: 121 iotype = HAMMER_IOTYPE_DUMMY; 122 break; 123 } 124 125 return(iotype); 126 } 127 128 static const char* 129 hammer_io_to_iostring(hammer_io_t io) 130 { 131 const char *iostr = NULL; 132 133 switch(io->type) { 134 case HAMMER_IOTYPE_VOLUME: 135 iostr = "volume"; 136 break; 137 case HAMMER_IOTYPE_META_BUFFER: 138 switch(HAMMER_ZONE(HAMMER_ITOB(io)->zoneX_offset)) { 139 case HAMMER_ZONE_RAW_BUFFER: 140 iostr = "meta/raw_buffer"; 141 break; 142 case HAMMER_ZONE_FREEMAP: 143 iostr = "meta/freemap"; 144 break; 145 case HAMMER_ZONE_BTREE: 146 iostr = "meta/btree"; 147 break; 148 case HAMMER_ZONE_META: 149 iostr = "meta/meta"; 150 break; 151 } 152 break; 153 case HAMMER_IOTYPE_UNDO_BUFFER: 154 iostr = "undo"; 155 break; 156 case HAMMER_IOTYPE_DATA_BUFFER: 157 switch(HAMMER_ZONE(HAMMER_ITOB(io)->zoneX_offset)) { 158 case HAMMER_ZONE_LARGE_DATA: 159 iostr = "data/large_data"; 160 break; 161 case HAMMER_ZONE_SMALL_DATA: 162 iostr = "data/small_data"; 163 break; 164 } 165 break; 166 case HAMMER_IOTYPE_DUMMY: 167 iostr = "dummy"; 168 break; 169 default: 170 hpanic("bad io type"); 171 break; 172 } 173 174 return(iostr); 175 } 176 177 /* 178 * Helper routine to disassociate a buffer cache buffer from an I/O 179 * structure. The io must be interlocked and marked appropriately for 180 * reclamation. 181 * 182 * The io must be in a released state with the io->bp owned and 183 * locked by the caller of this function. When not called from an 184 * io_deallocate() this cannot race an io_deallocate() since the 185 * kernel would be unable to get the buffer lock in that case. 186 * (The released state in this case means we own the bp, not the 187 * hammer_io structure). 188 * 189 * The io may have 0 or 1 references depending on who called us. The 190 * caller is responsible for dealing with the refs. 191 * 192 * This call can only be made when no action is required on the buffer. 193 * 194 * This function is guaranteed not to race against anything because we 195 * own both the io lock and the bp lock and are interlocked with no 196 * references. 197 */ 198 static void 199 hammer_io_disassociate(hammer_io_t io) 200 { 201 struct buf *bp = io->bp; 202 203 KKASSERT(io->released); 204 KKASSERT(io->modified == 0); 205 KKASSERT(hammer_buf_peek_io(bp) == io); 206 buf_dep_init(bp); 207 io->bp = NULL; 208 209 /* 210 * If the buffer was locked someone wanted to get rid of it. 211 */ 212 if (bp->b_flags & B_LOCKED) { 213 atomic_add_int(&hammer_count_io_locked, -1); 214 bp->b_flags &= ~B_LOCKED; 215 } 216 if (io->reclaim) { 217 bp->b_flags |= B_NOCACHE|B_RELBUF; 218 io->reclaim = 0; 219 } 220 221 switch(io->type) { 222 case HAMMER_IOTYPE_VOLUME: 223 HAMMER_ITOV(io)->ondisk = NULL; 224 break; 225 case HAMMER_IOTYPE_DATA_BUFFER: 226 case HAMMER_IOTYPE_META_BUFFER: 227 case HAMMER_IOTYPE_UNDO_BUFFER: 228 HAMMER_ITOB(io)->ondisk = NULL; 229 break; 230 case HAMMER_IOTYPE_DUMMY: 231 hpanic("bad io type"); 232 break; 233 } 234 } 235 236 /* 237 * Wait for any physical IO to complete 238 * 239 * XXX we aren't interlocked against a spinlock or anything so there 240 * is a small window in the interlock / io->running == 0 test. 241 */ 242 void 243 hammer_io_wait(hammer_io_t io) 244 { 245 if (io->running) { 246 hammer_mount_t hmp = io->hmp; 247 248 lwkt_gettoken(&hmp->io_token); 249 while (io->running) { 250 io->waiting = 1; 251 tsleep_interlock(io, 0); 252 if (io->running) 253 tsleep(io, PINTERLOCKED, "hmrflw", hz); 254 } 255 lwkt_reltoken(&hmp->io_token); 256 } 257 } 258 259 /* 260 * Wait for all currently queued HAMMER-initiated I/Os to complete. 261 * 262 * This is not supposed to count direct I/O's but some can leak 263 * through (for non-full-sized direct I/Os). 264 */ 265 void 266 hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush) 267 { 268 struct hammer_io iodummy; 269 hammer_io_t io; 270 271 /* 272 * Degenerate case, no I/O is running 273 */ 274 lwkt_gettoken(&hmp->io_token); 275 if (TAILQ_EMPTY(&hmp->iorun_list)) { 276 lwkt_reltoken(&hmp->io_token); 277 if (doflush) 278 hammer_io_flush_sync(hmp); 279 return; 280 } 281 bzero(&iodummy, sizeof(iodummy)); 282 iodummy.type = HAMMER_IOTYPE_DUMMY; 283 284 /* 285 * Add placemarker and then wait until it becomes the head of 286 * the list. 287 */ 288 TAILQ_INSERT_TAIL(&hmp->iorun_list, &iodummy, iorun_entry); 289 while (TAILQ_FIRST(&hmp->iorun_list) != &iodummy) { 290 tsleep(&iodummy, 0, ident, 0); 291 } 292 293 /* 294 * Chain in case several placemarkers are present. 295 */ 296 TAILQ_REMOVE(&hmp->iorun_list, &iodummy, iorun_entry); 297 io = TAILQ_FIRST(&hmp->iorun_list); 298 if (io && io->type == HAMMER_IOTYPE_DUMMY) 299 wakeup(io); 300 lwkt_reltoken(&hmp->io_token); 301 302 if (doflush) 303 hammer_io_flush_sync(hmp); 304 } 305 306 /* 307 * Clear a flagged error condition on a I/O buffer. The caller must hold 308 * its own ref on the buffer. 309 */ 310 void 311 hammer_io_clear_error(hammer_io_t io) 312 { 313 hammer_mount_t hmp = io->hmp; 314 315 lwkt_gettoken(&hmp->io_token); 316 if (io->ioerror) { 317 io->ioerror = 0; 318 hammer_rel(&io->lock); 319 KKASSERT(hammer_isactive(&io->lock)); 320 } 321 lwkt_reltoken(&hmp->io_token); 322 } 323 324 void 325 hammer_io_clear_error_noassert(hammer_io_t io) 326 { 327 hammer_mount_t hmp = io->hmp; 328 329 lwkt_gettoken(&hmp->io_token); 330 if (io->ioerror) { 331 io->ioerror = 0; 332 hammer_rel(&io->lock); 333 } 334 lwkt_reltoken(&hmp->io_token); 335 } 336 337 /* 338 * This is an advisory function only which tells the buffer cache 339 * the bp is not a meta-data buffer, even though it is backed by 340 * a block device. 341 * 342 * This is used by HAMMER's reblocking code to avoid trying to 343 * swapcache the filesystem's data when it is read or written 344 * by the reblocking code. 345 * 346 * The caller has a ref on the buffer preventing the bp from 347 * being disassociated from it. 348 */ 349 void 350 hammer_io_notmeta(hammer_buffer_t buffer) 351 { 352 if ((buffer->io.bp->b_flags & B_NOTMETA) == 0) { 353 hammer_mount_t hmp = buffer->io.hmp; 354 355 lwkt_gettoken(&hmp->io_token); 356 buffer->io.bp->b_flags |= B_NOTMETA; 357 lwkt_reltoken(&hmp->io_token); 358 } 359 } 360 361 /* 362 * Load bp for a HAMMER structure. The io must be exclusively locked by 363 * the caller. 364 * 365 * This routine is mostly used on meta-data and small-data blocks. Generally 366 * speaking HAMMER assumes some locality of reference and will cluster. 367 * 368 * Note that the caller (hammer_ondisk.c) may place further restrictions 369 * on clusterability via the limit (in bytes). Typically large-data 370 * zones cannot be clustered due to their mixed buffer sizes. This is 371 * not an issue since such clustering occurs in hammer_vnops at the 372 * regular file layer, whereas this is the buffered block device layer. 373 * 374 * No I/O callbacks can occur while we hold the buffer locked. 375 */ 376 int 377 hammer_io_read(struct vnode *devvp, hammer_io_t io, int limit) 378 { 379 struct buf *bp; 380 int error; 381 382 if ((bp = io->bp) == NULL) { 383 int hce = hammer_cluster_enable; 384 385 atomic_add_long(&hammer_count_io_running_read, io->bytes); 386 if (hce && limit > io->bytes) { 387 error = cluster_read(devvp, io->offset + limit, 388 io->offset, io->bytes, 389 HAMMER_CLUSTER_SIZE, 390 HAMMER_CLUSTER_SIZE * hce, 391 &io->bp); 392 } else { 393 error = bread(devvp, io->offset, io->bytes, &io->bp); 394 } 395 hammer_stats_disk_read += io->bytes; 396 atomic_add_long(&hammer_count_io_running_read, -io->bytes); 397 398 /* 399 * The code generally assumes b_ops/b_dep has been set-up, 400 * even if we error out here. 401 */ 402 bp = io->bp; 403 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IOISSUED)) { 404 hdkprintf("zone2_offset %016jx %s\n", 405 (intmax_t)bp->b_bio2.bio_offset, 406 hammer_io_to_iostring(io)); 407 } 408 bp->b_flags &= ~B_IOISSUED; 409 bp->b_ops = &hammer_bioops; 410 411 hammer_buf_attach_io(bp, io); /* locked by the io lock */ 412 BUF_KERNPROC(bp); 413 KKASSERT(io->modified == 0); 414 KKASSERT(io->running == 0); 415 KKASSERT(io->waiting == 0); 416 io->released = 0; /* we hold an active lock on bp */ 417 } else { 418 error = 0; 419 } 420 return(error); 421 } 422 423 /* 424 * Similar to hammer_io_read() but returns a zero'd out buffer instead. 425 * Must be called with the IO exclusively locked. 426 * 427 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background 428 * I/O by forcing the buffer to not be in a released state before calling 429 * it. 430 * 431 * This function will also mark the IO as modified but it will not 432 * increment the modify_refs count. 433 * 434 * No I/O callbacks can occur while we hold the buffer locked. 435 */ 436 int 437 hammer_io_new(struct vnode *devvp, hammer_io_t io) 438 { 439 struct buf *bp; 440 441 if ((bp = io->bp) == NULL) { 442 io->bp = getblk(devvp, io->offset, io->bytes, 0, 0); 443 bp = io->bp; 444 bp->b_ops = &hammer_bioops; 445 446 hammer_buf_attach_io(bp, io); /* locked by the io lock */ 447 io->released = 0; 448 KKASSERT(io->running == 0); 449 io->waiting = 0; 450 BUF_KERNPROC(bp); 451 } else { 452 if (io->released) { 453 regetblk(bp); 454 BUF_KERNPROC(bp); 455 io->released = 0; 456 } 457 } 458 hammer_io_modify(io, 0); 459 vfs_bio_clrbuf(bp); 460 return(0); 461 } 462 463 /* 464 * Advance the activity count on the underlying buffer because 465 * HAMMER does not getblk/brelse on every access. 466 * 467 * The io->bp cannot go away while the buffer is referenced. 468 */ 469 void 470 hammer_io_advance(hammer_io_t io) 471 { 472 if (io->bp) 473 buf_act_advance(io->bp); 474 } 475 476 /* 477 * Remove potential device level aliases against buffers managed by high level 478 * vnodes. Aliases can also be created due to mixed buffer sizes or via 479 * direct access to the backing store device. 480 * 481 * This is nasty because the buffers are also VMIO-backed. Even if a buffer 482 * does not exist its backing VM pages might, and we have to invalidate 483 * those as well or a getblk() will reinstate them. 484 * 485 * Buffer cache buffers associated with hammer_buffers cannot be 486 * invalidated. 487 */ 488 int 489 hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset) 490 { 491 hammer_io_t io; 492 hammer_mount_t hmp; 493 hammer_off_t phys_offset; 494 struct buf *bp; 495 int error; 496 497 hmp = volume->io.hmp; 498 lwkt_gettoken(&hmp->io_token); 499 500 /* 501 * If a device buffer already exists for the specified physical 502 * offset use that, otherwise instantiate a buffer to cover any 503 * related VM pages, set BNOCACHE, and brelse(). 504 */ 505 phys_offset = hammer_xlate_to_phys(volume->ondisk, zone2_offset); 506 if ((bp = findblk(volume->devvp, phys_offset, 0)) != NULL) 507 bremfree(bp); 508 else 509 bp = getblk(volume->devvp, phys_offset, HAMMER_BUFSIZE, 0, 0); 510 511 if ((io = hammer_buf_peek_io(bp)) != NULL) { 512 #if 0 513 hammer_ref(&io->lock); 514 hammer_io_clear_modify(io, 1); 515 bundirty(bp); 516 io->released = 0; 517 BUF_KERNPROC(bp); 518 io->reclaim = 1; 519 io->waitdep = 1; /* XXX this is a fs_token field */ 520 KKASSERT(hammer_isactive(&io->lock) == 1); 521 hammer_rel_buffer(HAMMER_ITOB(io), 0); 522 /*hammer_io_deallocate(bp);*/ 523 #endif 524 bqrelse(bp); 525 error = EAGAIN; 526 } else { 527 KKASSERT((bp->b_flags & B_LOCKED) == 0); 528 bundirty(bp); 529 bp->b_flags |= B_NOCACHE|B_RELBUF; 530 brelse(bp); 531 error = 0; 532 } 533 lwkt_reltoken(&hmp->io_token); 534 return(error); 535 } 536 537 /* 538 * This routine is called on the last reference to a hammer structure. 539 * The io must be interlocked with a refcount of zero. The hammer structure 540 * will remain interlocked on return. 541 * 542 * This routine may return a non-NULL bp to the caller for dispoal. 543 * The caller typically brelse()'s the bp. 544 * 545 * The bp may or may not still be passively associated with the IO. It 546 * will remain passively associated if it is unreleasable (e.g. a modified 547 * meta-data buffer). 548 * 549 * The only requirement here is that modified meta-data and volume-header 550 * buffer may NOT be disassociated from the IO structure, and consequently 551 * we also leave such buffers actively associated with the IO if they already 552 * are (since the kernel can't do anything with them anyway). Only the 553 * flusher is allowed to write such buffers out. Modified pure-data and 554 * undo buffers are returned to the kernel but left passively associated 555 * so we can track when the kernel writes the bp out. 556 */ 557 struct buf * 558 hammer_io_release(hammer_io_t io, int flush) 559 { 560 struct buf *bp; 561 562 if ((bp = io->bp) == NULL) 563 return(NULL); 564 565 /* 566 * Try to flush a dirty IO to disk if asked to by the 567 * caller or if the kernel tried to flush the buffer in the past. 568 * 569 * Kernel-initiated flushes are only allowed for pure-data buffers. 570 * meta-data and volume buffers can only be flushed explicitly 571 * by HAMMER. 572 */ 573 if (io->modified) { 574 if (flush) { 575 hammer_io_flush(io, 0); 576 } else if (bp->b_flags & B_LOCKED) { 577 switch(io->type) { 578 case HAMMER_IOTYPE_DATA_BUFFER: 579 hammer_io_flush(io, 0); 580 break; 581 case HAMMER_IOTYPE_UNDO_BUFFER: 582 hammer_io_flush(io, hammer_undo_reclaim(io)); 583 break; 584 default: 585 break; 586 } 587 } /* else no explicit request to flush the buffer */ 588 } 589 590 /* 591 * Wait for the IO to complete if asked to. This occurs when 592 * the buffer must be disposed of definitively during an umount 593 * or buffer invalidation. 594 */ 595 if (io->waitdep && io->running) { 596 hammer_io_wait(io); 597 } 598 599 /* 600 * Return control of the buffer to the kernel (with the provisio 601 * that our bioops can override kernel decisions with regards to 602 * the buffer). 603 */ 604 if ((flush || io->reclaim) && io->modified == 0 && io->running == 0) { 605 /* 606 * Always disassociate the bp if an explicit flush 607 * was requested and the IO completed with no error 608 * (so unmount can really clean up the structure). 609 */ 610 if (io->released) { 611 regetblk(bp); 612 BUF_KERNPROC(bp); 613 } else { 614 io->released = 1; 615 } 616 hammer_io_disassociate(io); 617 /* return the bp */ 618 } else if (io->modified) { 619 /* 620 * Only certain IO types can be released to the kernel if 621 * the buffer has been modified. 622 * 623 * volume and meta-data IO types may only be explicitly 624 * flushed by HAMMER. 625 */ 626 switch(io->type) { 627 case HAMMER_IOTYPE_DATA_BUFFER: 628 case HAMMER_IOTYPE_UNDO_BUFFER: 629 if (io->released == 0) { 630 io->released = 1; 631 bp->b_flags |= B_CLUSTEROK; 632 bdwrite(bp); 633 } 634 break; 635 default: 636 break; 637 } 638 bp = NULL; /* bp left associated */ 639 } else if (io->released == 0) { 640 /* 641 * Clean buffers can be generally released to the kernel. 642 * We leave the bp passively associated with the HAMMER 643 * structure and use bioops to disconnect it later on 644 * if the kernel wants to discard the buffer. 645 * 646 * We can steal the structure's ownership of the bp. 647 */ 648 io->released = 1; 649 if (bp->b_flags & B_LOCKED) { 650 hammer_io_disassociate(io); 651 /* return the bp */ 652 } else { 653 if (io->reclaim) { 654 hammer_io_disassociate(io); 655 /* return the bp */ 656 } else { 657 /* return the bp (bp passively associated) */ 658 } 659 } 660 } else { 661 /* 662 * A released buffer is passively associate with our 663 * hammer_io structure. The kernel cannot destroy it 664 * without making a bioops call. If the kernel (B_LOCKED) 665 * or we (reclaim) requested that the buffer be destroyed 666 * we destroy it, otherwise we do a quick get/release to 667 * reset its position in the kernel's LRU list. 668 * 669 * Leaving the buffer passively associated allows us to 670 * use the kernel's LRU buffer flushing mechanisms rather 671 * then rolling our own. 672 * 673 * XXX there are two ways of doing this. We can re-acquire 674 * and passively release to reset the LRU, or not. 675 */ 676 if (io->running == 0) { 677 regetblk(bp); 678 if ((bp->b_flags & B_LOCKED) || io->reclaim) { 679 hammer_io_disassociate(io); 680 /* return the bp */ 681 } else { 682 /* return the bp (bp passively associated) */ 683 } 684 } else { 685 /* 686 * bp is left passively associated but we do not 687 * try to reacquire it. Interactions with the io 688 * structure will occur on completion of the bp's 689 * I/O. 690 */ 691 bp = NULL; 692 } 693 } 694 return(bp); 695 } 696 697 /* 698 * This routine is called with a locked IO when a flush is desired and 699 * no other references to the structure exists other then ours. This 700 * routine is ONLY called when HAMMER believes it is safe to flush a 701 * potentially modified buffer out. 702 * 703 * The locked io or io reference prevents a flush from being initiated 704 * by the kernel. 705 */ 706 void 707 hammer_io_flush(hammer_io_t io, int reclaim) 708 { 709 struct buf *bp; 710 hammer_mount_t hmp; 711 712 /* 713 * Degenerate case - nothing to flush if nothing is dirty. 714 */ 715 if (io->modified == 0) 716 return; 717 718 KKASSERT(io->bp); 719 KKASSERT(io->modify_refs <= 0); 720 721 /* 722 * Acquire ownership of the bp, particularly before we clear our 723 * modified flag. 724 * 725 * We are going to bawrite() this bp. Don't leave a window where 726 * io->released is set, we actually own the bp rather then our 727 * buffer. 728 * 729 * The io_token should not be required here as only 730 */ 731 hmp = io->hmp; 732 bp = io->bp; 733 if (io->released) { 734 regetblk(bp); 735 /* BUF_KERNPROC(io->bp); */ 736 /* io->released = 0; */ 737 KKASSERT(io->released); 738 KKASSERT(io->bp == bp); 739 } else { 740 io->released = 1; 741 } 742 743 if (reclaim) { 744 io->reclaim = 1; 745 if ((bp->b_flags & B_LOCKED) == 0) { 746 bp->b_flags |= B_LOCKED; 747 atomic_add_int(&hammer_count_io_locked, 1); 748 } 749 } 750 751 /* 752 * Acquire exclusive access to the bp and then clear the modified 753 * state of the buffer prior to issuing I/O to interlock any 754 * modifications made while the I/O is in progress. This shouldn't 755 * happen anyway but losing data would be worse. The modified bit 756 * will be rechecked after the IO completes. 757 * 758 * NOTE: This call also finalizes the buffer's content (inval == 0). 759 * 760 * This is only legal when lock.refs == 1 (otherwise we might clear 761 * the modified bit while there are still users of the cluster 762 * modifying the data). 763 * 764 * Do this before potentially blocking so any attempt to modify the 765 * ondisk while we are blocked blocks waiting for us. 766 */ 767 hammer_ref(&io->lock); 768 hammer_io_clear_modify(io, 0); 769 hammer_rel(&io->lock); 770 771 if (hammer_debug_io & 0x0002) 772 hdkprintf("%016jx\n", bp->b_bio1.bio_offset); 773 774 /* 775 * Transfer ownership to the kernel and initiate I/O. 776 * 777 * NOTE: We do not hold io_token so an atomic op is required to 778 * update io_running_space. 779 */ 780 io->running = 1; 781 atomic_add_long(&hmp->io_running_space, io->bytes); 782 atomic_add_long(&hammer_count_io_running_write, io->bytes); 783 lwkt_gettoken(&hmp->io_token); 784 TAILQ_INSERT_TAIL(&hmp->iorun_list, io, iorun_entry); 785 lwkt_reltoken(&hmp->io_token); 786 cluster_awrite(bp); 787 hammer_io_flush_mark(io->volume); 788 } 789 790 /************************************************************************ 791 * BUFFER DIRTYING * 792 ************************************************************************ 793 * 794 * These routines deal with dependancies created when IO buffers get 795 * modified. The caller must call hammer_modify_*() on a referenced 796 * HAMMER structure prior to modifying its on-disk data. 797 * 798 * Any intent to modify an IO buffer acquires the related bp and imposes 799 * various write ordering dependancies. 800 */ 801 802 /* 803 * Mark a HAMMER structure as undergoing modification. Meta-data buffers 804 * are locked until the flusher can deal with them, pure data buffers 805 * can be written out. 806 * 807 * The referenced io prevents races. 808 */ 809 static 810 void 811 hammer_io_modify(hammer_io_t io, int count) 812 { 813 /* 814 * io->modify_refs must be >= 0 815 */ 816 while (io->modify_refs < 0) { 817 io->waitmod = 1; 818 tsleep(io, 0, "hmrmod", 0); 819 } 820 821 /* 822 * Shortcut if nothing to do. 823 */ 824 KKASSERT(hammer_isactive(&io->lock) && io->bp != NULL); 825 io->modify_refs += count; 826 if (io->modified && io->released == 0) 827 return; 828 829 /* 830 * NOTE: It is important not to set the modified bit 831 * until after we have acquired the bp or we risk 832 * racing against checkwrite. 833 */ 834 hammer_lock_ex(&io->lock); 835 if (io->released) { 836 regetblk(io->bp); 837 BUF_KERNPROC(io->bp); 838 io->released = 0; 839 } 840 if (io->modified == 0) { 841 hammer_io_set_modlist(io); 842 io->modified = 1; 843 } 844 hammer_unlock(&io->lock); 845 } 846 847 static __inline 848 void 849 hammer_io_modify_done(hammer_io_t io) 850 { 851 KKASSERT(io->modify_refs > 0); 852 --io->modify_refs; 853 if (io->modify_refs == 0 && io->waitmod) { 854 io->waitmod = 0; 855 wakeup(io); 856 } 857 } 858 859 /* 860 * The write interlock blocks other threads trying to modify a buffer 861 * (they block in hammer_io_modify()) after us, or blocks us while other 862 * threads are in the middle of modifying a buffer. 863 * 864 * The caller also has a ref on the io, however if we are not careful 865 * we will race bioops callbacks (checkwrite). To deal with this 866 * we must at least acquire and release the io_token, and it is probably 867 * better to hold it through the setting of modify_refs. 868 */ 869 void 870 hammer_io_write_interlock(hammer_io_t io) 871 { 872 hammer_mount_t hmp = io->hmp; 873 874 lwkt_gettoken(&hmp->io_token); 875 while (io->modify_refs != 0) { 876 io->waitmod = 1; 877 tsleep(io, 0, "hmrmod", 0); 878 } 879 io->modify_refs = -1; 880 lwkt_reltoken(&hmp->io_token); 881 } 882 883 void 884 hammer_io_done_interlock(hammer_io_t io) 885 { 886 KKASSERT(io->modify_refs == -1); 887 io->modify_refs = 0; 888 if (io->waitmod) { 889 io->waitmod = 0; 890 wakeup(io); 891 } 892 } 893 894 /* 895 * Caller intends to modify a volume's ondisk structure. 896 * 897 * This is only allowed if we are the flusher or we have a ref on the 898 * sync_lock. 899 */ 900 void 901 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume, 902 void *base, int len) 903 { 904 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 905 906 hammer_io_modify(&volume->io, 1); 907 if (len) { 908 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk; 909 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 910 hammer_generate_undo(trans, 911 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset), 912 base, len); 913 } 914 } 915 916 /* 917 * Caller intends to modify a buffer's ondisk structure. 918 * 919 * This is only allowed if we are the flusher or we have a ref on the 920 * sync_lock. 921 */ 922 void 923 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer, 924 void *base, int len) 925 { 926 KKASSERT (trans == NULL || trans->sync_lock_refs > 0); 927 928 hammer_io_modify(&buffer->io, 1); 929 if (len) { 930 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk; 931 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0); 932 hammer_generate_undo(trans, 933 buffer->zone2_offset + rel_offset, 934 base, len); 935 } 936 } 937 938 void 939 hammer_modify_volume_done(hammer_volume_t volume) 940 { 941 hammer_io_modify_done(&volume->io); 942 } 943 944 void 945 hammer_modify_buffer_done(hammer_buffer_t buffer) 946 { 947 hammer_io_modify_done(&buffer->io); 948 } 949 950 /* 951 * Mark an entity as not being dirty any more and finalize any 952 * delayed adjustments to the buffer. 953 * 954 * Delayed adjustments are an important performance enhancement, allowing 955 * us to avoid recalculating B-Tree node CRCs over and over again when 956 * making bulk-modifications to the B-Tree. 957 * 958 * If inval is non-zero delayed adjustments are ignored. 959 * 960 * This routine may dereference related btree nodes and cause the 961 * buffer to be dereferenced. The caller must own a reference on io. 962 */ 963 void 964 hammer_io_clear_modify(hammer_io_t io, int inval) 965 { 966 hammer_mount_t hmp; 967 968 /* 969 * io_token is needed to avoid races on mod_root 970 */ 971 if (io->modified == 0) 972 return; 973 hmp = io->hmp; 974 lwkt_gettoken(&hmp->io_token); 975 if (io->modified == 0) { 976 lwkt_reltoken(&hmp->io_token); 977 return; 978 } 979 980 /* 981 * Take us off the mod-list and clear the modified bit. 982 */ 983 KKASSERT(io->mod_root != NULL); 984 if (io->mod_root == &io->hmp->volu_root || 985 io->mod_root == &io->hmp->meta_root) { 986 io->hmp->locked_dirty_space -= io->bytes; 987 atomic_add_long(&hammer_count_dirtybufspace, -io->bytes); 988 } 989 RB_REMOVE(hammer_mod_rb_tree, io->mod_root, io); 990 io->mod_root = NULL; 991 io->modified = 0; 992 993 lwkt_reltoken(&hmp->io_token); 994 995 /* 996 * If this bit is not set there are no delayed adjustments. 997 */ 998 if (io->gencrc == 0) 999 return; 1000 io->gencrc = 0; 1001 1002 /* 1003 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference 1004 * on the node (& underlying buffer). Release the node after clearing 1005 * the flag. 1006 */ 1007 if (io->type == HAMMER_IOTYPE_META_BUFFER) { 1008 hammer_buffer_t buffer = HAMMER_ITOB(io); 1009 hammer_node_t node; 1010 1011 restart: 1012 TAILQ_FOREACH(node, &buffer->node_list, entry) { 1013 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) 1014 continue; 1015 node->flags &= ~HAMMER_NODE_NEEDSCRC; 1016 KKASSERT(node->ondisk); 1017 if (inval == 0) 1018 hammer_crc_set_btree(hmp->version, node->ondisk); 1019 hammer_rel_node(node); 1020 goto restart; 1021 } 1022 } 1023 /* caller must still have ref on io */ 1024 KKASSERT(hammer_isactive(&io->lock)); 1025 } 1026 1027 /* 1028 * Clear the IO's modify list. Even though the IO is no longer modified 1029 * it may still be on the lose_root. This routine is called just before 1030 * the governing hammer_buffer is destroyed. 1031 * 1032 * mod_root requires io_token protection. 1033 */ 1034 void 1035 hammer_io_clear_modlist(hammer_io_t io) 1036 { 1037 hammer_mount_t hmp = io->hmp; 1038 1039 KKASSERT(io->modified == 0); 1040 if (io->mod_root) { 1041 lwkt_gettoken(&hmp->io_token); 1042 if (io->mod_root) { 1043 KKASSERT(io->mod_root == &io->hmp->lose_root); 1044 RB_REMOVE(hammer_mod_rb_tree, io->mod_root, io); 1045 io->mod_root = NULL; 1046 } 1047 lwkt_reltoken(&hmp->io_token); 1048 } 1049 } 1050 1051 static void 1052 hammer_io_set_modlist(hammer_io_t io) 1053 { 1054 hammer_mount_t hmp = io->hmp; 1055 1056 lwkt_gettoken(&hmp->io_token); 1057 KKASSERT(io->mod_root == NULL); 1058 1059 switch(io->type) { 1060 case HAMMER_IOTYPE_VOLUME: 1061 io->mod_root = &hmp->volu_root; 1062 hmp->locked_dirty_space += io->bytes; 1063 atomic_add_long(&hammer_count_dirtybufspace, io->bytes); 1064 break; 1065 case HAMMER_IOTYPE_META_BUFFER: 1066 io->mod_root = &hmp->meta_root; 1067 hmp->locked_dirty_space += io->bytes; 1068 atomic_add_long(&hammer_count_dirtybufspace, io->bytes); 1069 break; 1070 case HAMMER_IOTYPE_UNDO_BUFFER: 1071 io->mod_root = &hmp->undo_root; 1072 break; 1073 case HAMMER_IOTYPE_DATA_BUFFER: 1074 io->mod_root = &hmp->data_root; 1075 break; 1076 case HAMMER_IOTYPE_DUMMY: 1077 hpanic("bad io type"); 1078 break; /* NOT REACHED */ 1079 } 1080 if (RB_INSERT(hammer_mod_rb_tree, io->mod_root, io)) { 1081 hpanic("duplicate entry @ %d:%015jx", 1082 io->volume->vol_no, io->offset); 1083 /* NOT REACHED */ 1084 } 1085 lwkt_reltoken(&hmp->io_token); 1086 } 1087 1088 /************************************************************************ 1089 * HAMMER_BIOOPS * 1090 ************************************************************************ 1091 * 1092 */ 1093 1094 /* 1095 * Pre-IO initiation kernel callback - cluster build only 1096 * 1097 * bioops callback - hold io_token 1098 */ 1099 static void 1100 hammer_io_start(struct buf *bp) 1101 { 1102 /* nothing to do, so io_token not needed */ 1103 } 1104 1105 /* 1106 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT! 1107 * 1108 * NOTE: HAMMER may modify a data buffer after we have initiated write 1109 * I/O. 1110 * 1111 * NOTE: MPSAFE callback 1112 * 1113 * bioops callback - hold io_token 1114 */ 1115 static void 1116 hammer_io_complete(struct buf *bp) 1117 { 1118 hammer_io_t io = hammer_buf_peek_io(bp); 1119 hammer_mount_t hmp = io->hmp; 1120 hammer_io_t ionext; 1121 1122 lwkt_gettoken(&hmp->io_token); 1123 1124 KKASSERT(io->released == 1); 1125 1126 /* 1127 * Deal with people waiting for I/O to drain 1128 */ 1129 if (io->running) { 1130 /* 1131 * Deal with critical write errors. Once a critical error 1132 * has been flagged in hmp the UNDO FIFO will not be updated. 1133 * That way crash recover will give us a consistent 1134 * filesystem. 1135 * 1136 * Because of this we can throw away failed UNDO buffers. If 1137 * we throw away META or DATA buffers we risk corrupting 1138 * the now read-only version of the filesystem visible to 1139 * the user. Clear B_ERROR so the buffer is not re-dirtied 1140 * by the kernel and ref the io so it doesn't get thrown 1141 * away. 1142 */ 1143 if (bp->b_flags & B_ERROR) { 1144 lwkt_gettoken(&hmp->fs_token); 1145 hammer_critical_error(hmp, NULL, bp->b_error, 1146 "while flushing meta-data"); 1147 lwkt_reltoken(&hmp->fs_token); 1148 1149 switch(io->type) { 1150 case HAMMER_IOTYPE_UNDO_BUFFER: 1151 break; 1152 default: 1153 if (io->ioerror == 0) { 1154 io->ioerror = 1; 1155 hammer_ref(&io->lock); 1156 } 1157 break; 1158 } 1159 bp->b_flags &= ~B_ERROR; 1160 bundirty(bp); 1161 #if 0 1162 hammer_io_set_modlist(io); 1163 io->modified = 1; 1164 #endif 1165 } 1166 hammer_stats_disk_write += io->bytes; 1167 atomic_add_long(&hammer_count_io_running_write, -io->bytes); 1168 atomic_add_long(&hmp->io_running_space, -io->bytes); 1169 KKASSERT(hmp->io_running_space >= 0); 1170 io->running = 0; 1171 1172 /* 1173 * Remove from iorun list and wakeup any multi-io waiter(s). 1174 */ 1175 if (TAILQ_FIRST(&hmp->iorun_list) == io) { 1176 ionext = TAILQ_NEXT(io, iorun_entry); 1177 if (ionext && ionext->type == HAMMER_IOTYPE_DUMMY) 1178 wakeup(ionext); 1179 } 1180 TAILQ_REMOVE(&hmp->iorun_list, io, iorun_entry); 1181 } else { 1182 hammer_stats_disk_read += io->bytes; 1183 } 1184 1185 if (io->waiting) { 1186 io->waiting = 0; 1187 wakeup(io); 1188 } 1189 1190 /* 1191 * If B_LOCKED is set someone wanted to deallocate the bp at some 1192 * point, try to do it now. The operation will fail if there are 1193 * refs or if hammer_io_deallocate() is unable to gain the 1194 * interlock. 1195 */ 1196 if (bp->b_flags & B_LOCKED) { 1197 atomic_add_int(&hammer_count_io_locked, -1); 1198 bp->b_flags &= ~B_LOCKED; 1199 hammer_io_deallocate(bp); 1200 /* structure may be dead now */ 1201 } 1202 lwkt_reltoken(&hmp->io_token); 1203 } 1204 1205 /* 1206 * Callback from kernel when it wishes to deallocate a passively 1207 * associated structure. This mostly occurs with clean buffers 1208 * but it may be possible for a holding structure to be marked dirty 1209 * while its buffer is passively associated. The caller owns the bp. 1210 * 1211 * If we cannot disassociate we set B_LOCKED to prevent the buffer 1212 * from getting reused. 1213 * 1214 * WARNING: Because this can be called directly by getnewbuf we cannot 1215 * recurse into the tree. If a bp cannot be immediately disassociated 1216 * our only recourse is to set B_LOCKED. 1217 * 1218 * WARNING: This may be called from an interrupt via hammer_io_complete() 1219 * 1220 * bioops callback - hold io_token 1221 */ 1222 static void 1223 hammer_io_deallocate(struct buf *bp) 1224 { 1225 hammer_io_t io = hammer_buf_peek_io(bp); 1226 hammer_mount_t hmp; 1227 1228 hmp = io->hmp; 1229 1230 lwkt_gettoken(&hmp->io_token); 1231 1232 KKASSERT((bp->b_flags & B_LOCKED) == 0 && io->running == 0); 1233 if (hammer_try_interlock_norefs(&io->lock) == 0) { 1234 /* 1235 * We cannot safely disassociate a bp from a referenced 1236 * or interlocked HAMMER structure. 1237 */ 1238 bp->b_flags |= B_LOCKED; 1239 atomic_add_int(&hammer_count_io_locked, 1); 1240 } else if (io->modified) { 1241 /* 1242 * It is not legal to disassociate a modified buffer. This 1243 * case really shouldn't ever occur. 1244 */ 1245 bp->b_flags |= B_LOCKED; 1246 atomic_add_int(&hammer_count_io_locked, 1); 1247 hammer_put_interlock(&io->lock, 0); 1248 } else { 1249 /* 1250 * Disassociate the BP. If the io has no refs left we 1251 * have to add it to the loose list. The kernel has 1252 * locked the buffer and therefore our io must be 1253 * in a released state. 1254 */ 1255 hammer_io_disassociate(io); 1256 if (io->type != HAMMER_IOTYPE_VOLUME) { 1257 KKASSERT(io->bp == NULL); 1258 KKASSERT(io->mod_root == NULL); 1259 io->mod_root = &hmp->lose_root; 1260 if (RB_INSERT(hammer_mod_rb_tree, io->mod_root, io)) { 1261 hpanic("duplicate entry @ %d:%015jx", 1262 io->volume->vol_no, io->offset); 1263 /* NOT REACHED */ 1264 } 1265 } 1266 hammer_put_interlock(&io->lock, 1); 1267 } 1268 lwkt_reltoken(&hmp->io_token); 1269 } 1270 1271 /* 1272 * bioops callback - hold io_token 1273 */ 1274 static int 1275 hammer_io_fsync(struct vnode *vp) 1276 { 1277 /* nothing to do, so io_token not needed */ 1278 return(0); 1279 } 1280 1281 /* 1282 * NOTE: will not be called unless we tell the kernel about the 1283 * bioops. Unused... we use the mount's VFS_SYNC instead. 1284 * 1285 * bioops callback - hold io_token 1286 */ 1287 static int 1288 hammer_io_sync(struct mount *mp) 1289 { 1290 /* nothing to do, so io_token not needed */ 1291 return(0); 1292 } 1293 1294 /* 1295 * bioops callback - hold io_token 1296 */ 1297 static void 1298 hammer_io_movedeps(struct buf *bp1, struct buf *bp2) 1299 { 1300 /* nothing to do, so io_token not needed */ 1301 } 1302 1303 /* 1304 * I/O pre-check for reading and writing. HAMMER only uses this for 1305 * B_CACHE buffers so checkread just shouldn't happen, but if it does 1306 * allow it. 1307 * 1308 * Writing is a different case. We don't want the kernel to try to write 1309 * out a buffer that HAMMER may be modifying passively or which has a 1310 * dependancy. In addition, kernel-demanded writes can only proceed for 1311 * certain types of buffers (i.e. UNDO and DATA types). Other dirty 1312 * buffer types can only be explicitly written by the flusher. 1313 * 1314 * checkwrite will only be called for bdwrite()n buffers. If we return 1315 * success the kernel is guaranteed to initiate the buffer write. 1316 * 1317 * bioops callback - hold io_token 1318 */ 1319 static int 1320 hammer_io_checkread(struct buf *bp) 1321 { 1322 /* nothing to do, so io_token not needed */ 1323 return(0); 1324 } 1325 1326 /* 1327 * The kernel is asking us whether it can write out a dirty buffer or not. 1328 * 1329 * bioops callback - hold io_token 1330 */ 1331 static int 1332 hammer_io_checkwrite(struct buf *bp) 1333 { 1334 hammer_io_t io = hammer_buf_peek_io(bp); 1335 hammer_mount_t hmp = io->hmp; 1336 1337 /* 1338 * This shouldn't happen under normal operation. 1339 */ 1340 lwkt_gettoken(&hmp->io_token); 1341 if (io->type == HAMMER_IOTYPE_VOLUME || 1342 io->type == HAMMER_IOTYPE_META_BUFFER) { 1343 if (!panicstr) 1344 hpanic("illegal buffer"); 1345 if ((bp->b_flags & B_LOCKED) == 0) { 1346 bp->b_flags |= B_LOCKED; 1347 atomic_add_int(&hammer_count_io_locked, 1); 1348 } 1349 lwkt_reltoken(&hmp->io_token); 1350 return(1); 1351 } 1352 1353 /* 1354 * We have to be able to interlock the IO to safely modify any 1355 * of its fields without holding the fs_token. If we can't lock 1356 * it then we are racing someone. 1357 * 1358 * Our ownership of the bp lock prevents the io from being ripped 1359 * out from under us. 1360 */ 1361 if (hammer_try_interlock_norefs(&io->lock) == 0) { 1362 bp->b_flags |= B_LOCKED; 1363 atomic_add_int(&hammer_count_io_locked, 1); 1364 lwkt_reltoken(&hmp->io_token); 1365 return(1); 1366 } 1367 1368 /* 1369 * The modified bit must be cleared prior to the initiation of 1370 * any IO (returning 0 initiates the IO). Because this is a 1371 * normal data buffer hammer_io_clear_modify() runs through a 1372 * simple degenerate case. 1373 * 1374 * Return 0 will cause the kernel to initiate the IO, and we 1375 * must normally clear the modified bit before we begin. If 1376 * the io has modify_refs we do not clear the modified bit, 1377 * otherwise we may miss changes. 1378 * 1379 * Only data and undo buffers can reach here. These buffers do 1380 * not have terminal crc functions but we temporarily reference 1381 * the IO anyway, just in case. 1382 */ 1383 if (io->modify_refs == 0 && io->modified) { 1384 hammer_ref(&io->lock); 1385 hammer_io_clear_modify(io, 0); 1386 hammer_rel(&io->lock); 1387 } else if (io->modified) { 1388 KKASSERT(io->type == HAMMER_IOTYPE_DATA_BUFFER); 1389 } 1390 1391 /* 1392 * The kernel is going to start the IO, set io->running. 1393 */ 1394 KKASSERT(io->running == 0); 1395 io->running = 1; 1396 atomic_add_long(&io->hmp->io_running_space, io->bytes); 1397 atomic_add_long(&hammer_count_io_running_write, io->bytes); 1398 TAILQ_INSERT_TAIL(&io->hmp->iorun_list, io, iorun_entry); 1399 1400 hammer_put_interlock(&io->lock, 1); 1401 lwkt_reltoken(&hmp->io_token); 1402 1403 return(0); 1404 } 1405 1406 /* 1407 * Return non-zero if we wish to delay the kernel's attempt to flush 1408 * this buffer to disk. 1409 * 1410 * bioops callback - hold io_token 1411 */ 1412 static int 1413 hammer_io_countdeps(struct buf *bp, int n) 1414 { 1415 /* nothing to do, so io_token not needed */ 1416 return(0); 1417 } 1418 1419 static struct bio_ops hammer_bioops = { 1420 .io_start = hammer_io_start, 1421 .io_complete = hammer_io_complete, 1422 .io_deallocate = hammer_io_deallocate, 1423 .io_fsync = hammer_io_fsync, 1424 .io_sync = hammer_io_sync, 1425 .io_movedeps = hammer_io_movedeps, 1426 .io_countdeps = hammer_io_countdeps, 1427 .io_checkread = hammer_io_checkread, 1428 .io_checkwrite = hammer_io_checkwrite, 1429 }; 1430 1431 /************************************************************************ 1432 * DIRECT IO OPS * 1433 ************************************************************************ 1434 * 1435 * These functions operate directly on the buffer cache buffer associated 1436 * with a front-end vnode rather then a back-end device vnode. 1437 */ 1438 1439 /* 1440 * Read a buffer associated with a front-end vnode directly from the 1441 * disk media. The bio may be issued asynchronously. If leaf is non-NULL 1442 * we validate the CRC. 1443 * 1444 * We must check for the presence of a HAMMER buffer to handle the case 1445 * where the reblocker has rewritten the data (which it does via the HAMMER 1446 * buffer system, not via the high-level vnode buffer cache), but not yet 1447 * committed the buffer to the media. 1448 */ 1449 int 1450 hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio, 1451 hammer_btree_leaf_elm_t leaf) 1452 { 1453 hammer_off_t buf_offset; 1454 hammer_off_t zone2_offset; 1455 hammer_volume_t volume; 1456 struct buf *bp; 1457 struct bio *nbio; 1458 int vol_no; 1459 int error; 1460 1461 buf_offset = bio->bio_offset; 1462 KKASSERT(hammer_is_zone_large_data(buf_offset)); 1463 1464 /* 1465 * The buffer cache may have an aliased buffer (the reblocker can 1466 * write them). If it does we have to sync any dirty data before 1467 * we can build our direct-read. This is a non-critical code path. 1468 */ 1469 bp = bio->bio_buf; 1470 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize); 1471 1472 /* 1473 * Resolve to a zone-2 offset. The conversion just requires 1474 * munging the top 4 bits but we want to abstract it anyway 1475 * so the blockmap code can verify the zone assignment. 1476 */ 1477 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1478 if (error) 1479 goto done; 1480 KKASSERT(hammer_is_zone_raw_buffer(zone2_offset)); 1481 1482 /* 1483 * Resolve volume and raw-offset for 3rd level bio. The 1484 * offset will be specific to the volume. 1485 */ 1486 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1487 volume = hammer_get_volume(hmp, vol_no, &error); 1488 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1489 error = EIO; 1490 1491 if (error == 0) { 1492 /* 1493 * 3rd level bio (the caller has already pushed once) 1494 */ 1495 nbio = push_bio(bio); 1496 nbio->bio_offset = hammer_xlate_to_phys(volume->ondisk, 1497 zone2_offset); 1498 hammer_stats_disk_read += bp->b_bufsize; 1499 vn_strategy(volume->devvp, nbio); 1500 } 1501 hammer_rel_volume(volume, 0); 1502 done: 1503 if (error) { 1504 hdkprintf("failed @ %016jx\n", (intmax_t)zone2_offset); 1505 bp->b_error = error; 1506 bp->b_flags |= B_ERROR; 1507 biodone(bio); 1508 } 1509 return(error); 1510 } 1511 1512 /* 1513 * This works similarly to hammer_io_direct_read() except instead of 1514 * directly reading from the device into the bio we instead indirectly 1515 * read through the device's buffer cache and then copy the data into 1516 * the bio. 1517 * 1518 * If leaf is non-NULL and validation is enabled, the CRC will be checked. 1519 * 1520 * This routine also executes asynchronously. It allows hammer strategy 1521 * calls to operate asynchronously when in double_buffer mode (in addition 1522 * to operating asynchronously when in normal mode). 1523 */ 1524 int 1525 hammer_io_indirect_read(hammer_mount_t hmp, struct bio *bio, 1526 hammer_btree_leaf_elm_t leaf) 1527 { 1528 hammer_off_t buf_offset; 1529 hammer_off_t zone2_offset; 1530 hammer_volume_t volume; 1531 struct buf *bp; 1532 int vol_no; 1533 int error; 1534 1535 buf_offset = bio->bio_offset; 1536 KKASSERT(hammer_is_zone_large_data(buf_offset)); 1537 1538 /* 1539 * The buffer cache may have an aliased buffer (the reblocker can 1540 * write them). If it does we have to sync any dirty data before 1541 * we can build our direct-read. This is a non-critical code path. 1542 */ 1543 bp = bio->bio_buf; 1544 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize); 1545 1546 /* 1547 * Resolve to a zone-2 offset. The conversion just requires 1548 * munging the top 4 bits but we want to abstract it anyway 1549 * so the blockmap code can verify the zone assignment. 1550 */ 1551 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1552 if (error) 1553 goto done; 1554 KKASSERT(hammer_is_zone_raw_buffer(zone2_offset)); 1555 1556 /* 1557 * Resolve volume and raw-offset for 3rd level bio. The 1558 * offset will be specific to the volume. 1559 */ 1560 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1561 volume = hammer_get_volume(hmp, vol_no, &error); 1562 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1563 error = EIO; 1564 1565 if (error == 0) { 1566 /* 1567 * Convert to the raw volume->devvp offset and acquire 1568 * the buf, issuing async I/O if necessary. 1569 */ 1570 hammer_off_t limit; 1571 int hce; 1572 1573 buf_offset = hammer_xlate_to_phys(volume->ondisk, zone2_offset); 1574 1575 if (leaf && hammer_verify_data) { 1576 bio->bio_caller_info1.uvalue32 = leaf->data_crc; 1577 bio->bio_caller_info2.index = 1; 1578 } else { 1579 bio->bio_caller_info2.index = 0; 1580 } 1581 bio->bio_caller_info3.ptr = hmp; 1582 1583 hce = hammer_cluster_enable; 1584 if (hce > 0) { 1585 limit = HAMMER_BIGBLOCK_DOALIGN(zone2_offset); 1586 limit -= zone2_offset; 1587 cluster_readcb(volume->devvp, limit, buf_offset, 1588 bp->b_bufsize, 1589 B_NOTMETA, 1590 HAMMER_CLUSTER_SIZE, 1591 HAMMER_CLUSTER_SIZE * hce, 1592 hammer_indirect_callback, 1593 bio); 1594 } else { 1595 breadcb(volume->devvp, buf_offset, bp->b_bufsize, 1596 B_NOTMETA, 1597 hammer_indirect_callback, bio); 1598 } 1599 } 1600 hammer_rel_volume(volume, 0); 1601 done: 1602 if (error) { 1603 hdkprintf("failed @ %016jx\n", (intmax_t)zone2_offset); 1604 bp->b_error = error; 1605 bp->b_flags |= B_ERROR; 1606 biodone(bio); 1607 } 1608 return(error); 1609 } 1610 1611 /* 1612 * Indirect callback on completion. bio/bp specify the device-backed 1613 * buffer. bio->bio_caller_info1.ptr holds obio. 1614 * 1615 * obio/obp is the original regular file buffer. obio->bio_caller_info* 1616 * contains the crc specification. 1617 * 1618 * We are responsible for calling bpdone() and bqrelse() on bio/bp, and 1619 * for calling biodone() on obio. 1620 */ 1621 static void 1622 hammer_indirect_callback(struct bio *bio) 1623 { 1624 struct buf *bp = bio->bio_buf; 1625 struct buf *obp; 1626 struct bio *obio; 1627 hammer_mount_t hmp; 1628 1629 /* 1630 * If BIO_DONE is already set the device buffer was already 1631 * fully valid (B_CACHE). If it is not set then I/O was issued 1632 * and we have to run I/O completion as the last bio. 1633 * 1634 * Nobody is waiting for our device I/O to complete, we are 1635 * responsible for bqrelse()ing it which means we also have to do 1636 * the equivalent of biowait() and clear BIO_DONE (which breadcb() 1637 * may have set). 1638 * 1639 * Any preexisting device buffer should match the requested size, 1640 * but due to big-block recycling and other factors there is some 1641 * fragility there, so we assert that the device buffer covers 1642 * the request. 1643 */ 1644 if ((bio->bio_flags & BIO_DONE) == 0) 1645 bpdone(bp, 0); 1646 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC); 1647 1648 obio = bio->bio_caller_info1.ptr; 1649 obp = obio->bio_buf; 1650 hmp = obio->bio_caller_info3.ptr; 1651 1652 if (bp->b_flags & B_ERROR) { 1653 /* 1654 * Error from block device 1655 */ 1656 obp->b_flags |= B_ERROR; 1657 obp->b_error = bp->b_error; 1658 } else if (obio->bio_caller_info2.index && 1659 obio->bio_caller_info1.uvalue32 != 1660 hammer_datacrc(hmp->version, 1661 bp->b_data, obp->b_bufsize) && 1662 obio->bio_caller_info1.uvalue32 != 1663 hammer_datacrc(HAMMER_VOL_VERSION_SIX, 1664 bp->b_data, obp->b_bufsize)) { 1665 /* 1666 * CRC error. First check against current hammer version, 1667 * then back-off and check against version 6 (the original 1668 * crc). 1669 */ 1670 obp->b_flags |= B_ERROR; 1671 obp->b_error = EIO; 1672 } else { 1673 /* 1674 * Everything is ok 1675 */ 1676 KKASSERT(bp->b_bufsize >= obp->b_bufsize); 1677 bcopy(bp->b_data, obp->b_data, obp->b_bufsize); 1678 obp->b_resid = 0; 1679 obp->b_flags |= B_AGE; 1680 } 1681 biodone(obio); 1682 bqrelse(bp); 1683 } 1684 1685 /* 1686 * Write a buffer associated with a front-end vnode directly to the 1687 * disk media. The bio may be issued asynchronously. 1688 * 1689 * The BIO is associated with the specified record and RECG_DIRECT_IO 1690 * is set. The recorded is added to its object. 1691 */ 1692 int 1693 hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio, 1694 hammer_record_t record) 1695 { 1696 hammer_btree_leaf_elm_t leaf = &record->leaf; 1697 hammer_off_t buf_offset; 1698 hammer_off_t zone2_offset; 1699 hammer_volume_t volume; 1700 hammer_buffer_t buffer; 1701 struct buf *bp; 1702 struct bio *nbio; 1703 char *ptr; 1704 int vol_no; 1705 int error; 1706 1707 buf_offset = leaf->data_offset; 1708 1709 KKASSERT(hammer_is_zone_record(buf_offset)); 1710 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE); 1711 1712 /* 1713 * Issue or execute the I/O. The new memory record must replace 1714 * the old one before the I/O completes, otherwise a reaquisition of 1715 * the buffer will load the old media data instead of the new. 1716 */ 1717 if ((buf_offset & HAMMER_BUFMASK) == 0 && 1718 leaf->data_len >= HAMMER_BUFSIZE) { 1719 /* 1720 * We are using the vnode's bio to write directly to the 1721 * media, any hammer_buffer at the same zone-X offset will 1722 * now have stale data. 1723 */ 1724 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error); 1725 vol_no = HAMMER_VOL_DECODE(zone2_offset); 1726 volume = hammer_get_volume(hmp, vol_no, &error); 1727 1728 if (error == 0 && zone2_offset >= volume->maxbuf_off) 1729 error = EIO; 1730 if (error == 0) { 1731 bp = bio->bio_buf; 1732 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0); 1733 1734 /* 1735 * Second level bio - cached zone2 offset. 1736 * 1737 * (We can put our bio_done function in either the 1738 * 2nd or 3rd level). 1739 */ 1740 nbio = push_bio(bio); 1741 nbio->bio_offset = zone2_offset; 1742 nbio->bio_done = hammer_io_direct_write_complete; 1743 nbio->bio_caller_info1.ptr = record; 1744 record->zone2_offset = zone2_offset; 1745 record->gflags |= HAMMER_RECG_DIRECT_IO | 1746 HAMMER_RECG_DIRECT_INVAL; 1747 1748 /* 1749 * Third level bio - raw offset specific to the 1750 * correct volume. 1751 */ 1752 nbio = push_bio(nbio); 1753 nbio->bio_offset = hammer_xlate_to_phys(volume->ondisk, 1754 zone2_offset); 1755 hammer_stats_disk_write += bp->b_bufsize; 1756 hammer_ip_replace_bulk(hmp, record); 1757 vn_strategy(volume->devvp, nbio); 1758 hammer_io_flush_mark(volume); 1759 } 1760 hammer_rel_volume(volume, 0); 1761 } else { 1762 /* 1763 * Must fit in a standard HAMMER buffer. In this case all 1764 * consumers use the HAMMER buffer system and RECG_DIRECT_IO 1765 * does not need to be set-up. 1766 */ 1767 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0); 1768 buffer = NULL; 1769 ptr = hammer_bread(hmp, buf_offset, &error, &buffer); 1770 if (error == 0) { 1771 bp = bio->bio_buf; 1772 bp->b_flags |= B_AGE; 1773 hammer_io_modify(&buffer->io, 1); 1774 bcopy(bp->b_data, ptr, leaf->data_len); 1775 hammer_io_modify_done(&buffer->io); 1776 hammer_rel_buffer(buffer, 0); 1777 bp->b_resid = 0; 1778 hammer_ip_replace_bulk(hmp, record); 1779 biodone(bio); 1780 } 1781 } 1782 if (error) { 1783 /* 1784 * Major suckage occured. Also note: The record was 1785 * never added to the tree so we do not have to worry 1786 * about the backend. 1787 */ 1788 hdkprintf("failed @ %016jx\n", (intmax_t)leaf->data_offset); 1789 bp = bio->bio_buf; 1790 bp->b_resid = 0; 1791 bp->b_error = EIO; 1792 bp->b_flags |= B_ERROR; 1793 biodone(bio); 1794 record->flags |= HAMMER_RECF_DELETED_FE; 1795 hammer_rel_mem_record(record); 1796 } 1797 return(error); 1798 } 1799 1800 /* 1801 * On completion of the BIO this callback must disconnect 1802 * it from the hammer_record and chain to the previous bio. 1803 * 1804 * An I/O error forces the mount to read-only. Data buffers 1805 * are not B_LOCKED like meta-data buffers are, so we have to 1806 * throw the buffer away to prevent the kernel from retrying. 1807 * 1808 * NOTE: MPSAFE callback, only modify fields we have explicit 1809 * access to (the bp and the record->gflags). 1810 */ 1811 static 1812 void 1813 hammer_io_direct_write_complete(struct bio *nbio) 1814 { 1815 struct bio *obio; 1816 struct buf *bp; 1817 hammer_record_t record; 1818 hammer_mount_t hmp; 1819 1820 record = nbio->bio_caller_info1.ptr; 1821 KKASSERT(record != NULL); 1822 hmp = record->ip->hmp; 1823 1824 lwkt_gettoken(&hmp->io_token); 1825 1826 bp = nbio->bio_buf; 1827 obio = pop_bio(nbio); 1828 if (bp->b_flags & B_ERROR) { 1829 lwkt_gettoken(&hmp->fs_token); 1830 hammer_critical_error(hmp, record->ip, bp->b_error, 1831 "while writing bulk data"); 1832 lwkt_reltoken(&hmp->fs_token); 1833 bp->b_flags |= B_INVAL; 1834 } 1835 1836 KKASSERT(record->gflags & HAMMER_RECG_DIRECT_IO); 1837 if (record->gflags & HAMMER_RECG_DIRECT_WAIT) { 1838 record->gflags &= ~(HAMMER_RECG_DIRECT_IO | 1839 HAMMER_RECG_DIRECT_WAIT); 1840 /* record can disappear once DIRECT_IO flag is cleared */ 1841 wakeup(&record->flags); 1842 } else { 1843 record->gflags &= ~HAMMER_RECG_DIRECT_IO; 1844 /* record can disappear once DIRECT_IO flag is cleared */ 1845 } 1846 1847 lwkt_reltoken(&hmp->io_token); 1848 1849 biodone(obio); 1850 } 1851 1852 1853 /* 1854 * This is called before a record is either committed to the B-Tree 1855 * or destroyed, to resolve any associated direct-IO. 1856 * 1857 * (1) We must wait for any direct-IO related to the record to complete. 1858 * 1859 * (2) We must remove any buffer cache aliases for data accessed via 1860 * leaf->data_offset or zone2_offset so non-direct-IO consumers 1861 * (the mirroring and reblocking code) do not see stale data. 1862 */ 1863 void 1864 hammer_io_direct_wait(hammer_record_t record) 1865 { 1866 hammer_mount_t hmp = record->ip->hmp; 1867 1868 /* 1869 * Wait for I/O to complete 1870 */ 1871 if (record->gflags & HAMMER_RECG_DIRECT_IO) { 1872 lwkt_gettoken(&hmp->io_token); 1873 while (record->gflags & HAMMER_RECG_DIRECT_IO) { 1874 record->gflags |= HAMMER_RECG_DIRECT_WAIT; 1875 tsleep(&record->flags, 0, "hmdiow", 0); 1876 } 1877 lwkt_reltoken(&hmp->io_token); 1878 } 1879 1880 /* 1881 * Invalidate any related buffer cache aliases associated with the 1882 * backing device. This is needed because the buffer cache buffer 1883 * for file data is associated with the file vnode, not the backing 1884 * device vnode. 1885 * 1886 * XXX I do not think this case can occur any more now that 1887 * reservations ensure that all such buffers are removed before 1888 * an area can be reused. 1889 */ 1890 if (record->gflags & HAMMER_RECG_DIRECT_INVAL) { 1891 KKASSERT(record->leaf.data_offset); 1892 hammer_del_buffers(hmp, record->leaf.data_offset, 1893 record->zone2_offset, record->leaf.data_len, 1894 1); 1895 record->gflags &= ~HAMMER_RECG_DIRECT_INVAL; 1896 } 1897 } 1898 1899 /* 1900 * This is called to remove the second-level cached zone-2 offset from 1901 * frontend buffer cache buffers, now stale due to a data relocation. 1902 * These offsets are generated by cluster_read() via VOP_BMAP, or directly 1903 * by hammer_vop_strategy_read(). 1904 * 1905 * This is rather nasty because here we have something like the reblocker 1906 * scanning the raw B-Tree with no held references on anything, really, 1907 * other then a shared lock on the B-Tree node, and we have to access the 1908 * frontend's buffer cache to check for and clean out the association. 1909 * Specifically, if the reblocker is moving data on the disk, these cached 1910 * offsets will become invalid. 1911 * 1912 * Only data record types associated with the large-data zone are subject 1913 * to direct-io and need to be checked. 1914 * 1915 */ 1916 void 1917 hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf) 1918 { 1919 struct hammer_inode_info iinfo; 1920 int zone; 1921 1922 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA) 1923 return; 1924 zone = HAMMER_ZONE_DECODE(leaf->data_offset); 1925 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX) 1926 return; 1927 iinfo.obj_id = leaf->base.obj_id; 1928 iinfo.obj_asof = 0; /* unused */ 1929 iinfo.obj_localization = leaf->base.localization & 1930 HAMMER_LOCALIZE_PSEUDOFS_MASK; 1931 iinfo.u.leaf = leaf; 1932 hammer_scan_inode_snapshots(hmp, &iinfo, 1933 hammer_io_direct_uncache_callback, 1934 leaf); 1935 } 1936 1937 static int 1938 hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data) 1939 { 1940 hammer_inode_info_t iinfo = data; 1941 hammer_off_t file_offset; 1942 struct vnode *vp; 1943 struct buf *bp; 1944 int blksize; 1945 1946 if (ip->vp == NULL) 1947 return(0); 1948 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len; 1949 blksize = iinfo->u.leaf->data_len; 1950 KKASSERT((blksize & HAMMER_BUFMASK) == 0); 1951 1952 /* 1953 * Warning: FINDBLK_TEST return stable storage but not stable 1954 * contents. It happens to be ok in this case. 1955 */ 1956 hammer_ref(&ip->lock); 1957 if (hammer_get_vnode(ip, &vp) == 0) { 1958 if ((bp = findblk(ip->vp, file_offset, FINDBLK_TEST)) != NULL && 1959 bp->b_bio2.bio_offset != NOOFFSET) { 1960 bp = getblk(ip->vp, file_offset, blksize, 0, 0); 1961 bp->b_bio2.bio_offset = NOOFFSET; 1962 brelse(bp); 1963 } 1964 vput(vp); 1965 } 1966 hammer_rel_inode(ip, 0); 1967 return(0); 1968 } 1969 1970 1971 /* 1972 * This function is called when writes may have occured on the volume, 1973 * indicating that the device may be holding cached writes. 1974 */ 1975 static __inline void 1976 hammer_io_flush_mark(hammer_volume_t volume) 1977 { 1978 atomic_set_int(&volume->vol_flags, HAMMER_VOLF_NEEDFLUSH); 1979 } 1980 1981 /* 1982 * This function ensures that the device has flushed any cached writes out. 1983 */ 1984 void 1985 hammer_io_flush_sync(hammer_mount_t hmp) 1986 { 1987 hammer_volume_t volume; 1988 struct buf *bp_base = NULL; 1989 struct buf *bp; 1990 1991 RB_FOREACH(volume, hammer_vol_rb_tree, &hmp->rb_vols_root) { 1992 if (volume->vol_flags & HAMMER_VOLF_NEEDFLUSH) { 1993 atomic_clear_int(&volume->vol_flags, 1994 HAMMER_VOLF_NEEDFLUSH); 1995 bp = getpbuf(NULL); 1996 bp->b_bio1.bio_offset = 0; 1997 bp->b_bufsize = 0; 1998 bp->b_bcount = 0; 1999 bp->b_cmd = BUF_CMD_FLUSH; 2000 bp->b_bio1.bio_caller_info1.cluster_head = bp_base; 2001 bp->b_bio1.bio_done = biodone_sync; 2002 bp->b_bio1.bio_flags |= BIO_SYNC; 2003 bp_base = bp; 2004 vn_strategy(volume->devvp, &bp->b_bio1); 2005 } 2006 } 2007 while ((bp = bp_base) != NULL) { 2008 bp_base = bp->b_bio1.bio_caller_info1.cluster_head; 2009 biowait(&bp->b_bio1, "hmrFLS"); 2010 relpbuf(bp, NULL); 2011 } 2012 } 2013 2014 /* 2015 * Limit the amount of backlog which we allow to build up 2016 */ 2017 void 2018 hammer_io_limit_backlog(hammer_mount_t hmp) 2019 { 2020 waitrunningbufspace(); 2021 } 2022