vfs_cluster.c revision 59762
1/*- 2 * Copyright (c) 1993 3 * The Regents of the University of California. All rights reserved. 4 * Modifications/enhancements: 5 * Copyright (c) 1995 John S. Dyson. All rights reserved. 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 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. All advertising materials mentioning features or use of this software 16 * must display the following acknowledgement: 17 * This product includes software developed by the University of 18 * California, Berkeley and its contributors. 19 * 4. Neither the name of the University nor the names of its contributors 20 * may be used to endorse or promote products derived from this software 21 * without specific prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 * 35 * @(#)vfs_cluster.c 8.7 (Berkeley) 2/13/94 36 * $FreeBSD: head/sys/kern/vfs_cluster.c 59762 2000-04-29 16:25:22Z phk $ 37 */ 38 39#include "opt_debug_cluster.h" 40 41#include <sys/param.h> 42#include <sys/systm.h> 43#include <sys/kernel.h> 44#include <sys/proc.h> 45#include <sys/buf.h> 46#include <sys/vnode.h> 47#include <sys/malloc.h> 48#include <sys/mount.h> 49#include <sys/resourcevar.h> 50#include <vm/vm.h> 51#include <vm/vm_object.h> 52#include <vm/vm_page.h> 53#include <sys/sysctl.h> 54 55#if defined(CLUSTERDEBUG) 56#include <sys/sysctl.h> 57static int rcluster= 0; 58SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, ""); 59#endif 60 61static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer"); 62 63static struct cluster_save * 64 cluster_collectbufs __P((struct vnode *vp, struct buf *last_bp)); 65static struct buf * 66 cluster_rbuild __P((struct vnode *vp, u_quad_t filesize, daddr_t lbn, 67 daddr_t blkno, long size, int run, struct buf *fbp)); 68 69static int write_behind = 1; 70SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, ""); 71 72extern vm_page_t bogus_page; 73 74extern int cluster_pbuf_freecnt; 75 76/* 77 * Maximum number of blocks for read-ahead. 78 */ 79#define MAXRA 32 80 81/* 82 * This replaces bread. 83 */ 84int 85cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp) 86 struct vnode *vp; 87 u_quad_t filesize; 88 daddr_t lblkno; 89 long size; 90 struct ucred *cred; 91 long totread; 92 int seqcount; 93 struct buf **bpp; 94{ 95 struct buf *bp, *rbp, *reqbp; 96 daddr_t blkno, origblkno; 97 int error, num_ra; 98 int i; 99 int maxra, racluster; 100 long origtotread; 101 102 error = 0; 103 104 /* 105 * Try to limit the amount of read-ahead by a few 106 * ad-hoc parameters. This needs work!!! 107 */ 108 racluster = vp->v_mount->mnt_iosize_max / size; 109 maxra = 2 * racluster + (totread / size); 110 if (maxra > MAXRA) 111 maxra = MAXRA; 112 if (maxra > nbuf/8) 113 maxra = nbuf/8; 114 115 /* 116 * get the requested block 117 */ 118 *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0); 119 origblkno = lblkno; 120 origtotread = totread; 121 122 /* 123 * if it is in the cache, then check to see if the reads have been 124 * sequential. If they have, then try some read-ahead, otherwise 125 * back-off on prospective read-aheads. 126 */ 127 if (bp->b_flags & B_CACHE) { 128 if (!seqcount) { 129 return 0; 130 } else if ((bp->b_flags & B_RAM) == 0) { 131 return 0; 132 } else { 133 int s; 134 struct buf *tbp; 135 bp->b_flags &= ~B_RAM; 136 /* 137 * We do the spl here so that there is no window 138 * between the incore and the b_usecount increment 139 * below. We opt to keep the spl out of the loop 140 * for efficiency. 141 */ 142 s = splbio(); 143 for (i = 1; i < maxra; i++) { 144 145 if (!(tbp = incore(vp, lblkno+i))) { 146 break; 147 } 148 149 /* 150 * Set another read-ahead mark so we know 151 * to check again. 152 */ 153 if (((i % racluster) == (racluster - 1)) || 154 (i == (maxra - 1))) 155 tbp->b_flags |= B_RAM; 156 } 157 splx(s); 158 if (i >= maxra) { 159 return 0; 160 } 161 lblkno += i; 162 } 163 reqbp = bp = NULL; 164 } else { 165 off_t firstread = bp->b_offset; 166 167 KASSERT(bp->b_offset != NOOFFSET, 168 ("cluster_read: no buffer offset")); 169 if (firstread + totread > filesize) 170 totread = filesize - firstread; 171 if (totread > size) { 172 int nblks = 0; 173 int ncontigafter; 174 while (totread > 0) { 175 nblks++; 176 totread -= size; 177 } 178 if (nblks == 1) 179 goto single_block_read; 180 if (nblks > racluster) 181 nblks = racluster; 182 183 error = VOP_BMAP(vp, lblkno, NULL, 184 &blkno, &ncontigafter, NULL); 185 if (error) 186 goto single_block_read; 187 if (blkno == -1) 188 goto single_block_read; 189 if (ncontigafter == 0) 190 goto single_block_read; 191 if (ncontigafter + 1 < nblks) 192 nblks = ncontigafter + 1; 193 194 bp = cluster_rbuild(vp, filesize, lblkno, 195 blkno, size, nblks, bp); 196 lblkno += (bp->b_bufsize / size); 197 } else { 198single_block_read: 199 /* 200 * if it isn't in the cache, then get a chunk from 201 * disk if sequential, otherwise just get the block. 202 */ 203 bp->b_flags |= B_RAM; 204 bp->b_iocmd = BIO_READ; 205 lblkno += 1; 206 } 207 } 208 209 /* 210 * if we have been doing sequential I/O, then do some read-ahead 211 */ 212 rbp = NULL; 213 if (seqcount && (lblkno < (origblkno + seqcount))) { 214 /* 215 * we now build the read-ahead buffer if it is desirable. 216 */ 217 if (((u_quad_t)(lblkno + 1) * size) <= filesize && 218 !(error = VOP_BMAP(vp, lblkno, NULL, &blkno, &num_ra, NULL)) && 219 blkno != -1) { 220 int nblksread; 221 int ntoread = num_ra + 1; 222 nblksread = (origtotread + size - 1) / size; 223 if (seqcount < nblksread) 224 seqcount = nblksread; 225 if (seqcount < ntoread) 226 ntoread = seqcount; 227 if (num_ra) { 228 rbp = cluster_rbuild(vp, filesize, lblkno, 229 blkno, size, ntoread, NULL); 230 } else { 231 rbp = getblk(vp, lblkno, size, 0, 0); 232 rbp->b_flags |= B_ASYNC | B_RAM; 233 rbp->b_iocmd = BIO_READ; 234 rbp->b_blkno = blkno; 235 } 236 } 237 } 238 239 /* 240 * handle the synchronous read 241 */ 242 if (bp) { 243#if defined(CLUSTERDEBUG) 244 if (rcluster) 245 printf("S(%ld,%ld,%d) ", 246 (long)bp->b_lblkno, bp->b_bcount, seqcount); 247#endif 248 if ((bp->b_flags & B_CLUSTER) == 0) 249 vfs_busy_pages(bp, 0); 250 bp->b_flags &= ~B_INVAL; 251 bp->b_ioflags &= ~BIO_ERROR; 252 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL) 253 BUF_KERNPROC(bp); 254 error = VOP_STRATEGY(vp, bp); 255 curproc->p_stats->p_ru.ru_inblock++; 256 } 257 258 /* 259 * and if we have read-aheads, do them too 260 */ 261 if (rbp) { 262 if (error) { 263 rbp->b_flags &= ~B_ASYNC; 264 brelse(rbp); 265 } else if (rbp->b_flags & B_CACHE) { 266 rbp->b_flags &= ~B_ASYNC; 267 bqrelse(rbp); 268 } else { 269#if defined(CLUSTERDEBUG) 270 if (rcluster) { 271 if (bp) 272 printf("A+(%ld,%ld,%ld,%d) ", 273 (long)rbp->b_lblkno, rbp->b_bcount, 274 (long)(rbp->b_lblkno - origblkno), 275 seqcount); 276 else 277 printf("A(%ld,%ld,%ld,%d) ", 278 (long)rbp->b_lblkno, rbp->b_bcount, 279 (long)(rbp->b_lblkno - origblkno), 280 seqcount); 281 } 282#endif 283 284 if ((rbp->b_flags & B_CLUSTER) == 0) 285 vfs_busy_pages(rbp, 0); 286 rbp->b_flags &= ~B_INVAL; 287 rbp->b_ioflags &= ~BIO_ERROR; 288 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL) 289 BUF_KERNPROC(rbp); 290 (void) VOP_STRATEGY(vp, rbp); 291 curproc->p_stats->p_ru.ru_inblock++; 292 } 293 } 294 if (reqbp) 295 return (bufwait(reqbp)); 296 else 297 return (error); 298} 299 300/* 301 * If blocks are contiguous on disk, use this to provide clustered 302 * read ahead. We will read as many blocks as possible sequentially 303 * and then parcel them up into logical blocks in the buffer hash table. 304 */ 305static struct buf * 306cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp) 307 struct vnode *vp; 308 u_quad_t filesize; 309 daddr_t lbn; 310 daddr_t blkno; 311 long size; 312 int run; 313 struct buf *fbp; 314{ 315 struct buf *bp, *tbp; 316 daddr_t bn; 317 int i, inc, j; 318 319 KASSERT(size == vp->v_mount->mnt_stat.f_iosize, 320 ("cluster_rbuild: size %ld != filesize %ld\n", 321 size, vp->v_mount->mnt_stat.f_iosize)); 322 323 /* 324 * avoid a division 325 */ 326 while ((u_quad_t) size * (lbn + run) > filesize) { 327 --run; 328 } 329 330 if (fbp) { 331 tbp = fbp; 332 tbp->b_iocmd = BIO_READ; 333 } else { 334 tbp = getblk(vp, lbn, size, 0, 0); 335 if (tbp->b_flags & B_CACHE) 336 return tbp; 337 tbp->b_flags |= B_ASYNC | B_RAM; 338 tbp->b_iocmd = BIO_READ; 339 } 340 341 tbp->b_blkno = blkno; 342 if( (tbp->b_flags & B_MALLOC) || 343 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) ) 344 return tbp; 345 346 bp = trypbuf(&cluster_pbuf_freecnt); 347 if (bp == 0) 348 return tbp; 349 350 bp->b_data = (char *)((vm_offset_t)bp->b_data | 351 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 352 bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO; 353 bp->b_iocmd = BIO_READ; 354 bp->b_iodone = cluster_callback; 355 bp->b_blkno = blkno; 356 bp->b_lblkno = lbn; 357 bp->b_offset = tbp->b_offset; 358 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset")); 359 pbgetvp(vp, bp); 360 361 TAILQ_INIT(&bp->b_cluster.cluster_head); 362 363 bp->b_bcount = 0; 364 bp->b_bufsize = 0; 365 bp->b_npages = 0; 366 367 inc = btodb(size); 368 for (bn = blkno, i = 0; i < run; ++i, bn += inc) { 369 if (i != 0) { 370 if ((bp->b_npages * PAGE_SIZE) + 371 round_page(size) > vp->v_mount->mnt_iosize_max) 372 break; 373 374 if ((tbp = incore(vp, lbn + i)) != NULL) { 375 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) 376 break; 377 BUF_UNLOCK(tbp); 378 379 for (j = 0; j < tbp->b_npages; j++) 380 if (tbp->b_pages[j]->valid) 381 break; 382 383 if (j != tbp->b_npages) 384 break; 385 386 if (tbp->b_bcount != size) 387 break; 388 } 389 390 tbp = getblk(vp, lbn + i, size, 0, 0); 391 392 if ((tbp->b_flags & B_CACHE) || 393 (tbp->b_flags & B_VMIO) == 0) { 394 bqrelse(tbp); 395 break; 396 } 397 398 for (j = 0;j < tbp->b_npages; j++) 399 if (tbp->b_pages[j]->valid) 400 break; 401 402 if (j != tbp->b_npages) { 403 bqrelse(tbp); 404 break; 405 } 406 407 if ((fbp && (i == 1)) || (i == (run - 1))) 408 tbp->b_flags |= B_RAM; 409 tbp->b_flags |= B_ASYNC; 410 tbp->b_iocmd = BIO_READ; 411 if (tbp->b_blkno == tbp->b_lblkno) { 412 tbp->b_blkno = bn; 413 } else if (tbp->b_blkno != bn) { 414 brelse(tbp); 415 break; 416 } 417 } 418 /* 419 * XXX fbp from caller may not be B_ASYNC, but we are going 420 * to biodone() it in cluster_callback() anyway 421 */ 422 BUF_KERNPROC(tbp); 423 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 424 tbp, b_cluster.cluster_entry); 425 for (j = 0; j < tbp->b_npages; j += 1) { 426 vm_page_t m; 427 m = tbp->b_pages[j]; 428 vm_page_io_start(m); 429 vm_object_pip_add(m->object, 1); 430 if ((bp->b_npages == 0) || 431 (bp->b_pages[bp->b_npages-1] != m)) { 432 bp->b_pages[bp->b_npages] = m; 433 bp->b_npages++; 434 } 435 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) 436 tbp->b_pages[j] = bogus_page; 437 } 438 bp->b_bcount += tbp->b_bcount; 439 bp->b_bufsize += tbp->b_bufsize; 440 } 441 442 for(j=0;j<bp->b_npages;j++) { 443 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) == 444 VM_PAGE_BITS_ALL) 445 bp->b_pages[j] = bogus_page; 446 } 447 if (bp->b_bufsize > bp->b_kvasize) 448 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 449 bp->b_bufsize, bp->b_kvasize); 450 bp->b_kvasize = bp->b_bufsize; 451 452 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 453 (vm_page_t *)bp->b_pages, bp->b_npages); 454 return (bp); 455} 456 457/* 458 * Cleanup after a clustered read or write. 459 * This is complicated by the fact that any of the buffers might have 460 * extra memory (if there were no empty buffer headers at allocbuf time) 461 * that we will need to shift around. 462 */ 463void 464cluster_callback(bp) 465 struct buf *bp; 466{ 467 struct buf *nbp, *tbp; 468 int error = 0; 469 470 /* 471 * Must propogate errors to all the components. 472 */ 473 if (bp->b_ioflags & BIO_ERROR) 474 error = bp->b_error; 475 476 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 477 /* 478 * Move memory from the large cluster buffer into the component 479 * buffers and mark IO as done on these. 480 */ 481 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head); 482 tbp; tbp = nbp) { 483 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry); 484 if (error) { 485 tbp->b_ioflags |= BIO_ERROR; 486 tbp->b_error = error; 487 } else { 488 tbp->b_dirtyoff = tbp->b_dirtyend = 0; 489 tbp->b_flags &= ~B_INVAL; 490 tbp->b_ioflags &= ~BIO_ERROR; 491 } 492 bufdone(tbp); 493 } 494 relpbuf(bp, &cluster_pbuf_freecnt); 495} 496 497/* 498 * cluster_wbuild_wb: 499 * 500 * Implement modified write build for cluster. 501 * 502 * write_behind = 0 write behind disabled 503 * write_behind = 1 write behind normal (default) 504 * write_behind = 2 write behind backed-off 505 */ 506 507static __inline int 508cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len) 509{ 510 int r = 0; 511 512 switch(write_behind) { 513 case 2: 514 if (start_lbn < len) 515 break; 516 start_lbn -= len; 517 /* fall through */ 518 case 1: 519 r = cluster_wbuild(vp, size, start_lbn, len); 520 /* fall through */ 521 default: 522 /* fall through */ 523 break; 524 } 525 return(r); 526} 527 528/* 529 * Do clustered write for FFS. 530 * 531 * Three cases: 532 * 1. Write is not sequential (write asynchronously) 533 * Write is sequential: 534 * 2. beginning of cluster - begin cluster 535 * 3. middle of a cluster - add to cluster 536 * 4. end of a cluster - asynchronously write cluster 537 */ 538void 539cluster_write(bp, filesize, seqcount) 540 struct buf *bp; 541 u_quad_t filesize; 542 int seqcount; 543{ 544 struct vnode *vp; 545 daddr_t lbn; 546 int maxclen, cursize; 547 int lblocksize; 548 int async; 549 550 vp = bp->b_vp; 551 if (vp->v_type == VREG) { 552 async = vp->v_mount->mnt_flag & MNT_ASYNC; 553 lblocksize = vp->v_mount->mnt_stat.f_iosize; 554 } else { 555 async = 0; 556 lblocksize = bp->b_bufsize; 557 } 558 lbn = bp->b_lblkno; 559 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset")); 560 561 /* Initialize vnode to beginning of file. */ 562 if (lbn == 0) 563 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 564 565 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 || 566 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) { 567 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1; 568 if (vp->v_clen != 0) { 569 /* 570 * Next block is not sequential. 571 * 572 * If we are not writing at end of file, the process 573 * seeked to another point in the file since its last 574 * write, or we have reached our maximum cluster size, 575 * then push the previous cluster. Otherwise try 576 * reallocating to make it sequential. 577 * 578 * Change to algorithm: only push previous cluster if 579 * it was sequential from the point of view of the 580 * seqcount heuristic, otherwise leave the buffer 581 * intact so we can potentially optimize the I/O 582 * later on in the buf_daemon or update daemon 583 * flush. 584 */ 585 cursize = vp->v_lastw - vp->v_cstart + 1; 586 if (((u_quad_t) bp->b_offset + lblocksize) != filesize || 587 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) { 588 if (!async && seqcount > 0) { 589 cluster_wbuild_wb(vp, lblocksize, 590 vp->v_cstart, cursize); 591 } 592 } else { 593 struct buf **bpp, **endbp; 594 struct cluster_save *buflist; 595 596 buflist = cluster_collectbufs(vp, bp); 597 endbp = &buflist->bs_children 598 [buflist->bs_nchildren - 1]; 599 if (VOP_REALLOCBLKS(vp, buflist)) { 600 /* 601 * Failed, push the previous cluster 602 * if *really* writing sequentially 603 * in the logical file (seqcount > 1), 604 * otherwise delay it in the hopes that 605 * the low level disk driver can 606 * optimize the write ordering. 607 */ 608 for (bpp = buflist->bs_children; 609 bpp < endbp; bpp++) 610 brelse(*bpp); 611 free(buflist, M_SEGMENT); 612 if (seqcount > 1) { 613 cluster_wbuild_wb(vp, 614 lblocksize, vp->v_cstart, 615 cursize); 616 } 617 } else { 618 /* 619 * Succeeded, keep building cluster. 620 */ 621 for (bpp = buflist->bs_children; 622 bpp <= endbp; bpp++) 623 bdwrite(*bpp); 624 free(buflist, M_SEGMENT); 625 vp->v_lastw = lbn; 626 vp->v_lasta = bp->b_blkno; 627 return; 628 } 629 } 630 } 631 /* 632 * Consider beginning a cluster. If at end of file, make 633 * cluster as large as possible, otherwise find size of 634 * existing cluster. 635 */ 636 if ((vp->v_type == VREG) && 637 ((u_quad_t) bp->b_offset + lblocksize) != filesize && 638 (bp->b_blkno == bp->b_lblkno) && 639 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) || 640 bp->b_blkno == -1)) { 641 bawrite(bp); 642 vp->v_clen = 0; 643 vp->v_lasta = bp->b_blkno; 644 vp->v_cstart = lbn + 1; 645 vp->v_lastw = lbn; 646 return; 647 } 648 vp->v_clen = maxclen; 649 if (!async && maxclen == 0) { /* I/O not contiguous */ 650 vp->v_cstart = lbn + 1; 651 bawrite(bp); 652 } else { /* Wait for rest of cluster */ 653 vp->v_cstart = lbn; 654 bdwrite(bp); 655 } 656 } else if (lbn == vp->v_cstart + vp->v_clen) { 657 /* 658 * At end of cluster, write it out if seqcount tells us we 659 * are operating sequentially, otherwise let the buf or 660 * update daemon handle it. 661 */ 662 bdwrite(bp); 663 if (seqcount > 1) 664 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1); 665 vp->v_clen = 0; 666 vp->v_cstart = lbn + 1; 667 } else { 668 /* 669 * In the middle of a cluster, so just delay the I/O for now. 670 */ 671 bdwrite(bp); 672 } 673 vp->v_lastw = lbn; 674 vp->v_lasta = bp->b_blkno; 675} 676 677 678/* 679 * This is an awful lot like cluster_rbuild...wish they could be combined. 680 * The last lbn argument is the current block on which I/O is being 681 * performed. Check to see that it doesn't fall in the middle of 682 * the current block (if last_bp == NULL). 683 */ 684int 685cluster_wbuild(vp, size, start_lbn, len) 686 struct vnode *vp; 687 long size; 688 daddr_t start_lbn; 689 int len; 690{ 691 struct buf *bp, *tbp; 692 int i, j, s; 693 int totalwritten = 0; 694 int dbsize = btodb(size); 695 696 while (len > 0) { 697 s = splbio(); 698 if (((tbp = gbincore(vp, start_lbn)) == NULL) || 699 ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) || 700 BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) { 701 ++start_lbn; 702 --len; 703 splx(s); 704 continue; 705 } 706 bremfree(tbp); 707 tbp->b_flags &= ~B_DONE; 708 splx(s); 709 710 /* 711 * Extra memory in the buffer, punt on this buffer. 712 * XXX we could handle this in most cases, but we would 713 * have to push the extra memory down to after our max 714 * possible cluster size and then potentially pull it back 715 * up if the cluster was terminated prematurely--too much 716 * hassle. 717 */ 718 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) || 719 (tbp->b_bcount != tbp->b_bufsize) || 720 (tbp->b_bcount != size) || 721 (len == 1) || 722 ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) { 723 totalwritten += tbp->b_bufsize; 724 bawrite(tbp); 725 ++start_lbn; 726 --len; 727 continue; 728 } 729 730 /* 731 * We got a pbuf to make the cluster in. 732 * so initialise it. 733 */ 734 TAILQ_INIT(&bp->b_cluster.cluster_head); 735 bp->b_bcount = 0; 736 bp->b_bufsize = 0; 737 bp->b_npages = 0; 738 if (tbp->b_wcred != NOCRED) { 739 bp->b_wcred = tbp->b_wcred; 740 crhold(bp->b_wcred); 741 } 742 743 bp->b_blkno = tbp->b_blkno; 744 bp->b_lblkno = tbp->b_lblkno; 745 bp->b_offset = tbp->b_offset; 746 bp->b_data = (char *)((vm_offset_t)bp->b_data | 747 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 748 bp->b_flags |= B_CLUSTER | 749 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT)); 750 bp->b_iodone = cluster_callback; 751 pbgetvp(vp, bp); 752 /* 753 * From this location in the file, scan forward to see 754 * if there are buffers with adjacent data that need to 755 * be written as well. 756 */ 757 for (i = 0; i < len; ++i, ++start_lbn) { 758 if (i != 0) { /* If not the first buffer */ 759 s = splbio(); 760 /* 761 * If the adjacent data is not even in core it 762 * can't need to be written. 763 */ 764 if ((tbp = gbincore(vp, start_lbn)) == NULL) { 765 splx(s); 766 break; 767 } 768 769 /* 770 * If it IS in core, but has different 771 * characteristics, don't cluster with it. 772 */ 773 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | 774 B_INVAL | B_DELWRI | B_NEEDCOMMIT)) 775 != (B_DELWRI | B_CLUSTEROK | 776 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || 777 tbp->b_wcred != bp->b_wcred || 778 BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) { 779 splx(s); 780 break; 781 } 782 783 /* 784 * Check that the combined cluster 785 * would make sense with regard to pages 786 * and would not be too large 787 */ 788 if ((tbp->b_bcount != size) || 789 ((bp->b_blkno + (dbsize * i)) != 790 tbp->b_blkno) || 791 ((tbp->b_npages + bp->b_npages) > 792 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) { 793 BUF_UNLOCK(tbp); 794 splx(s); 795 break; 796 } 797 /* 798 * Ok, it's passed all the tests, 799 * so remove it from the free list 800 * and mark it busy. We will use it. 801 */ 802 bremfree(tbp); 803 tbp->b_flags &= ~B_DONE; 804 splx(s); 805 } /* end of code for non-first buffers only */ 806 /* check for latent dependencies to be handled */ 807 if ((LIST_FIRST(&tbp->b_dep)) != NULL && 808 bioops.io_start) 809 (*bioops.io_start)(tbp); 810 /* 811 * If the IO is via the VM then we do some 812 * special VM hackery. (yuck) 813 */ 814 if (tbp->b_flags & B_VMIO) { 815 vm_page_t m; 816 817 if (i != 0) { /* if not first buffer */ 818 for (j = 0; j < tbp->b_npages; j += 1) { 819 m = tbp->b_pages[j]; 820 if (m->flags & PG_BUSY) { 821 bqrelse(tbp); 822 goto finishcluster; 823 } 824 } 825 } 826 827 for (j = 0; j < tbp->b_npages; j += 1) { 828 m = tbp->b_pages[j]; 829 vm_page_io_start(m); 830 vm_object_pip_add(m->object, 1); 831 if ((bp->b_npages == 0) || 832 (bp->b_pages[bp->b_npages - 1] != m)) { 833 bp->b_pages[bp->b_npages] = m; 834 bp->b_npages++; 835 } 836 } 837 } 838 bp->b_bcount += size; 839 bp->b_bufsize += size; 840 841 s = splbio(); 842 bundirty(tbp); 843 tbp->b_flags &= ~B_DONE; 844 tbp->b_ioflags &= ~BIO_ERROR; 845 tbp->b_flags |= B_ASYNC; 846 tbp->b_iocmd = BIO_WRITE; 847 reassignbuf(tbp, tbp->b_vp); /* put on clean list */ 848 ++tbp->b_vp->v_numoutput; 849 splx(s); 850 BUF_KERNPROC(tbp); 851 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 852 tbp, b_cluster.cluster_entry); 853 } 854 finishcluster: 855 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 856 (vm_page_t *) bp->b_pages, bp->b_npages); 857 if (bp->b_bufsize > bp->b_kvasize) 858 panic( 859 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 860 bp->b_bufsize, bp->b_kvasize); 861 bp->b_kvasize = bp->b_bufsize; 862 totalwritten += bp->b_bufsize; 863 bp->b_dirtyoff = 0; 864 bp->b_dirtyend = bp->b_bufsize; 865 bawrite(bp); 866 867 len -= i; 868 } 869 return totalwritten; 870} 871 872/* 873 * Collect together all the buffers in a cluster. 874 * Plus add one additional buffer. 875 */ 876static struct cluster_save * 877cluster_collectbufs(vp, last_bp) 878 struct vnode *vp; 879 struct buf *last_bp; 880{ 881 struct cluster_save *buflist; 882 struct buf *bp; 883 daddr_t lbn; 884 int i, len; 885 886 len = vp->v_lastw - vp->v_cstart + 1; 887 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), 888 M_SEGMENT, M_WAITOK); 889 buflist->bs_nchildren = 0; 890 buflist->bs_children = (struct buf **) (buflist + 1); 891 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) { 892 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp); 893 buflist->bs_children[i] = bp; 894 if (bp->b_blkno == bp->b_lblkno) 895 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 896 NULL, NULL); 897 } 898 buflist->bs_children[i] = bp = last_bp; 899 if (bp->b_blkno == bp->b_lblkno) 900 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 901 NULL, NULL); 902 buflist->bs_nchildren = i + 1; 903 return (buflist); 904} 905