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 * $Id: vfs_cluster.c,v 1.79 1999/01/27 21:49:58 dillon Exp $ 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_prot.h> 52 #include <vm/vm_object.h> 53 #include <vm/vm_page.h> 54 55 #if defined(CLUSTERDEBUG) 56 #include <sys/sysctl.h> 57 static int rcluster= 0; 58 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, ""); 59 #endif 60 61 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer"); 62 63 static struct cluster_save * 64 cluster_collectbufs __P((struct vnode *vp, struct buf *last_bp)); 65 static 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 69 extern vm_page_t bogus_page; 70 71 extern int cluster_pbuf_freecnt; 72 73 /* 74 * Maximum number of blocks for read-ahead. 75 */ 76 #define MAXRA 32 77 78 /* 79 * This replaces bread. 80 */ 81 int 82 cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp) 83 struct vnode *vp; 84 u_quad_t filesize; 85 daddr_t lblkno; 86 long size; 87 struct ucred *cred; 88 long totread; 89 int seqcount; 90 struct buf **bpp; 91 { 92 struct buf *bp, *rbp, *reqbp; 93 daddr_t blkno, origblkno; 94 int error, num_ra; 95 int i; 96 int maxra, racluster; 97 long origtotread; 98 99 error = 0; 100 if (vp->v_maxio == 0) 101 vp->v_maxio = DFLTPHYS; 102 103 /* 104 * Try to limit the amount of read-ahead by a few 105 * ad-hoc parameters. This needs work!!! 106 */ 107 racluster = vp->v_maxio/size; 108 maxra = 2 * racluster + (totread / size); 109 if (maxra > MAXRA) 110 maxra = MAXRA; 111 if (maxra > nbuf/8) 112 maxra = nbuf/8; 113 114 /* 115 * get the requested block 116 */ 117 *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0); 118 origblkno = lblkno; 119 origtotread = totread; 120 121 /* 122 * if it is in the cache, then check to see if the reads have been 123 * sequential. If they have, then try some read-ahead, otherwise 124 * back-off on prospective read-aheads. 125 */ 126 if (bp->b_flags & B_CACHE) { 127 if (!seqcount) { 128 return 0; 129 } else if ((bp->b_flags & B_RAM) == 0) { 130 return 0; 131 } else { 132 int s; 133 struct buf *tbp; 134 bp->b_flags &= ~B_RAM; 135 /* 136 * We do the spl here so that there is no window 137 * between the incore and the b_usecount increment 138 * below. We opt to keep the spl out of the loop 139 * for efficiency. 140 */ 141 s = splbio(); 142 for(i=1;i<maxra;i++) { 143 144 if (!(tbp = incore(vp, lblkno+i))) { 145 break; 146 } 147 148 /* 149 * Set another read-ahead mark so we know to check 150 * again. 151 */ 152 if (((i % racluster) == (racluster - 1)) || 153 (i == (maxra - 1))) 154 tbp->b_flags |= B_RAM; 155 156 if ((tbp->b_usecount < 1) && 157 ((tbp->b_flags & B_BUSY) == 0) && 158 (tbp->b_qindex == QUEUE_LRU)) { 159 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], tbp, b_freelist); 160 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], tbp, b_freelist); 161 } 162 } 163 splx(s); 164 if (i >= maxra) { 165 return 0; 166 } 167 lblkno += i; 168 } 169 reqbp = bp = NULL; 170 } else { 171 off_t firstread = bp->b_offset; 172 173 KASSERT(bp->b_offset != NOOFFSET, 174 ("cluster_read: no buffer offset")); 175 if (firstread + totread > filesize) 176 totread = filesize - firstread; 177 if (totread > size) { 178 int nblks = 0; 179 int ncontigafter; 180 while (totread > 0) { 181 nblks++; 182 totread -= size; 183 } 184 if (nblks == 1) 185 goto single_block_read; 186 if (nblks > racluster) 187 nblks = racluster; 188 189 error = VOP_BMAP(vp, lblkno, NULL, 190 &blkno, &ncontigafter, NULL); 191 if (error) 192 goto single_block_read; 193 if (blkno == -1) 194 goto single_block_read; 195 if (ncontigafter == 0) 196 goto single_block_read; 197 if (ncontigafter + 1 < nblks) 198 nblks = ncontigafter + 1; 199 200 bp = cluster_rbuild(vp, filesize, lblkno, 201 blkno, size, nblks, bp); 202 lblkno += (bp->b_bufsize / size); 203 } else { 204 single_block_read: 205 /* 206 * if it isn't in the cache, then get a chunk from 207 * disk if sequential, otherwise just get the block. 208 */ 209 bp->b_flags |= B_READ | B_RAM; 210 lblkno += 1; 211 } 212 } 213 214 /* 215 * if we have been doing sequential I/O, then do some read-ahead 216 */ 217 rbp = NULL; 218 if (seqcount && (lblkno < (origblkno + seqcount))) { 219 /* 220 * we now build the read-ahead buffer if it is desirable. 221 */ 222 if (((u_quad_t)(lblkno + 1) * size) <= filesize && 223 !(error = VOP_BMAP(vp, lblkno, NULL, &blkno, &num_ra, NULL)) && 224 blkno != -1) { 225 int nblksread; 226 int ntoread = num_ra + 1; 227 nblksread = (origtotread + size - 1) / size; 228 if (seqcount < nblksread) 229 seqcount = nblksread; 230 if (seqcount < ntoread) 231 ntoread = seqcount; 232 if (num_ra) { 233 rbp = cluster_rbuild(vp, filesize, lblkno, 234 blkno, size, ntoread, NULL); 235 } else { 236 rbp = getblk(vp, lblkno, size, 0, 0); 237 rbp->b_flags |= B_READ | B_ASYNC | B_RAM; 238 rbp->b_blkno = blkno; 239 } 240 } 241 } 242 243 /* 244 * handle the synchronous read 245 */ 246 if (bp) { 247 #if defined(CLUSTERDEBUG) 248 if (rcluster) 249 printf("S(%ld,%ld,%d) ", 250 (long)bp->b_lblkno, bp->b_bcount, seqcount); 251 #endif 252 if ((bp->b_flags & B_CLUSTER) == 0) 253 vfs_busy_pages(bp, 0); 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 | B_READ); 264 brelse(rbp); 265 } else if (rbp->b_flags & B_CACHE) { 266 rbp->b_flags &= ~(B_ASYNC | B_READ); 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 (void) VOP_STRATEGY(vp, rbp); 287 curproc->p_stats->p_ru.ru_inblock++; 288 } 289 } 290 if (reqbp) 291 return (biowait(reqbp)); 292 else 293 return (error); 294 } 295 296 /* 297 * If blocks are contiguous on disk, use this to provide clustered 298 * read ahead. We will read as many blocks as possible sequentially 299 * and then parcel them up into logical blocks in the buffer hash table. 300 */ 301 static struct buf * 302 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp) 303 struct vnode *vp; 304 u_quad_t filesize; 305 daddr_t lbn; 306 daddr_t blkno; 307 long size; 308 int run; 309 struct buf *fbp; 310 { 311 struct buf *bp, *tbp; 312 daddr_t bn; 313 int i, inc, j; 314 315 KASSERT(size == vp->v_mount->mnt_stat.f_iosize, 316 ("cluster_rbuild: size %ld != filesize %ld\n", 317 size, vp->v_mount->mnt_stat.f_iosize)); 318 319 /* 320 * avoid a division 321 */ 322 while ((u_quad_t) size * (lbn + run) > filesize) { 323 --run; 324 } 325 326 if (fbp) { 327 tbp = fbp; 328 tbp->b_flags |= B_READ; 329 } else { 330 tbp = getblk(vp, lbn, size, 0, 0); 331 if (tbp->b_flags & B_CACHE) 332 return tbp; 333 tbp->b_flags |= B_ASYNC | B_READ | B_RAM; 334 } 335 336 tbp->b_blkno = blkno; 337 if( (tbp->b_flags & B_MALLOC) || 338 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) ) 339 return tbp; 340 341 bp = trypbuf(&cluster_pbuf_freecnt); 342 if (bp == 0) 343 return tbp; 344 345 bp->b_data = (char *)((vm_offset_t)bp->b_data | 346 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 347 bp->b_flags = B_ASYNC | B_READ | B_CALL | B_BUSY | B_CLUSTER | B_VMIO; 348 bp->b_iodone = cluster_callback; 349 bp->b_blkno = blkno; 350 bp->b_lblkno = lbn; 351 bp->b_offset = tbp->b_offset; 352 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset")); 353 pbgetvp(vp, bp); 354 355 TAILQ_INIT(&bp->b_cluster.cluster_head); 356 357 bp->b_bcount = 0; 358 bp->b_bufsize = 0; 359 bp->b_npages = 0; 360 361 if (vp->v_maxio == 0) 362 vp->v_maxio = DFLTPHYS; 363 inc = btodb(size); 364 for (bn = blkno, i = 0; i < run; ++i, bn += inc) { 365 if (i != 0) { 366 if ((bp->b_npages * PAGE_SIZE) + 367 round_page(size) > vp->v_maxio) 368 break; 369 370 if ((tbp = incore(vp, lbn + i)) != NULL) { 371 if (tbp->b_flags & B_BUSY) 372 break; 373 374 for (j = 0; j < tbp->b_npages; j++) 375 if (tbp->b_pages[j]->valid) 376 break; 377 378 if (j != tbp->b_npages) 379 break; 380 381 if (tbp->b_bcount != size) 382 break; 383 } 384 385 tbp = getblk(vp, lbn + i, size, 0, 0); 386 387 if ((tbp->b_flags & B_CACHE) || 388 (tbp->b_flags & B_VMIO) == 0) { 389 bqrelse(tbp); 390 break; 391 } 392 393 for (j = 0;j < tbp->b_npages; j++) 394 if (tbp->b_pages[j]->valid) 395 break; 396 397 if (j != tbp->b_npages) { 398 bqrelse(tbp); 399 break; 400 } 401 402 if ((fbp && (i == 1)) || (i == (run - 1))) 403 tbp->b_flags |= B_RAM; 404 tbp->b_flags |= B_READ | B_ASYNC; 405 if (tbp->b_blkno == tbp->b_lblkno) { 406 tbp->b_blkno = bn; 407 } else if (tbp->b_blkno != bn) { 408 brelse(tbp); 409 break; 410 } 411 } 412 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 413 tbp, b_cluster.cluster_entry); 414 for (j = 0; j < tbp->b_npages; j += 1) { 415 vm_page_t m; 416 m = tbp->b_pages[j]; 417 vm_page_io_start(m); 418 vm_object_pip_add(m->object, 1); 419 if ((bp->b_npages == 0) || 420 (bp->b_pages[bp->b_npages-1] != m)) { 421 bp->b_pages[bp->b_npages] = m; 422 bp->b_npages++; 423 } 424 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) 425 tbp->b_pages[j] = bogus_page; 426 } 427 bp->b_bcount += tbp->b_bcount; 428 bp->b_bufsize += tbp->b_bufsize; 429 } 430 431 for(j=0;j<bp->b_npages;j++) { 432 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) == 433 VM_PAGE_BITS_ALL) 434 bp->b_pages[j] = bogus_page; 435 } 436 if (bp->b_bufsize > bp->b_kvasize) 437 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 438 bp->b_bufsize, bp->b_kvasize); 439 bp->b_kvasize = bp->b_bufsize; 440 441 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 442 (vm_page_t *)bp->b_pages, bp->b_npages); 443 return (bp); 444 } 445 446 /* 447 * Cleanup after a clustered read or write. 448 * This is complicated by the fact that any of the buffers might have 449 * extra memory (if there were no empty buffer headers at allocbuf time) 450 * that we will need to shift around. 451 */ 452 void 453 cluster_callback(bp) 454 struct buf *bp; 455 { 456 struct buf *nbp, *tbp; 457 int error = 0; 458 459 /* 460 * Must propogate errors to all the components. 461 */ 462 if (bp->b_flags & B_ERROR) 463 error = bp->b_error; 464 465 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 466 /* 467 * Move memory from the large cluster buffer into the component 468 * buffers and mark IO as done on these. 469 */ 470 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head); 471 tbp; tbp = nbp) { 472 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry); 473 if (error) { 474 tbp->b_flags |= B_ERROR; 475 tbp->b_error = error; 476 } else 477 tbp->b_dirtyoff = tbp->b_dirtyend = 0; 478 biodone(tbp); 479 } 480 relpbuf(bp, &cluster_pbuf_freecnt); 481 } 482 483 /* 484 * Do clustered write for FFS. 485 * 486 * Three cases: 487 * 1. Write is not sequential (write asynchronously) 488 * Write is sequential: 489 * 2. beginning of cluster - begin cluster 490 * 3. middle of a cluster - add to cluster 491 * 4. end of a cluster - asynchronously write cluster 492 */ 493 void 494 cluster_write(bp, filesize) 495 struct buf *bp; 496 u_quad_t filesize; 497 { 498 struct vnode *vp; 499 daddr_t lbn; 500 int maxclen, cursize; 501 int lblocksize; 502 int async; 503 504 vp = bp->b_vp; 505 if (vp->v_maxio == 0) 506 vp->v_maxio = DFLTPHYS; 507 if (vp->v_type == VREG) { 508 async = vp->v_mount->mnt_flag & MNT_ASYNC; 509 lblocksize = vp->v_mount->mnt_stat.f_iosize; 510 } else { 511 async = 0; 512 lblocksize = bp->b_bufsize; 513 } 514 lbn = bp->b_lblkno; 515 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset")); 516 517 /* Initialize vnode to beginning of file. */ 518 if (lbn == 0) 519 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 520 521 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 || 522 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) { 523 maxclen = vp->v_maxio / lblocksize - 1; 524 if (vp->v_clen != 0) { 525 /* 526 * Next block is not sequential. 527 * 528 * If we are not writing at end of file, the process 529 * seeked to another point in the file since its last 530 * write, or we have reached our maximum cluster size, 531 * then push the previous cluster. Otherwise try 532 * reallocating to make it sequential. 533 */ 534 cursize = vp->v_lastw - vp->v_cstart + 1; 535 if (((u_quad_t) bp->b_offset + lblocksize) != filesize || 536 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) { 537 if (!async) 538 cluster_wbuild(vp, lblocksize, 539 vp->v_cstart, cursize); 540 } else { 541 struct buf **bpp, **endbp; 542 struct cluster_save *buflist; 543 544 buflist = cluster_collectbufs(vp, bp); 545 endbp = &buflist->bs_children 546 [buflist->bs_nchildren - 1]; 547 if (VOP_REALLOCBLKS(vp, buflist)) { 548 /* 549 * Failed, push the previous cluster. 550 */ 551 for (bpp = buflist->bs_children; 552 bpp < endbp; bpp++) 553 brelse(*bpp); 554 free(buflist, M_SEGMENT); 555 cluster_wbuild(vp, lblocksize, 556 vp->v_cstart, cursize); 557 } else { 558 /* 559 * Succeeded, keep building cluster. 560 */ 561 for (bpp = buflist->bs_children; 562 bpp <= endbp; bpp++) 563 bdwrite(*bpp); 564 free(buflist, M_SEGMENT); 565 vp->v_lastw = lbn; 566 vp->v_lasta = bp->b_blkno; 567 return; 568 } 569 } 570 } 571 /* 572 * Consider beginning a cluster. If at end of file, make 573 * cluster as large as possible, otherwise find size of 574 * existing cluster. 575 */ 576 if ((vp->v_type == VREG) && 577 ((u_quad_t) bp->b_offset + lblocksize) != filesize && 578 (bp->b_blkno == bp->b_lblkno) && 579 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) || 580 bp->b_blkno == -1)) { 581 bawrite(bp); 582 vp->v_clen = 0; 583 vp->v_lasta = bp->b_blkno; 584 vp->v_cstart = lbn + 1; 585 vp->v_lastw = lbn; 586 return; 587 } 588 vp->v_clen = maxclen; 589 if (!async && maxclen == 0) { /* I/O not contiguous */ 590 vp->v_cstart = lbn + 1; 591 bawrite(bp); 592 } else { /* Wait for rest of cluster */ 593 vp->v_cstart = lbn; 594 bdwrite(bp); 595 } 596 } else if (lbn == vp->v_cstart + vp->v_clen) { 597 /* 598 * At end of cluster, write it out. 599 */ 600 bdwrite(bp); 601 cluster_wbuild(vp, lblocksize, vp->v_cstart, vp->v_clen + 1); 602 vp->v_clen = 0; 603 vp->v_cstart = lbn + 1; 604 } else 605 /* 606 * In the middle of a cluster, so just delay the I/O for now. 607 */ 608 bdwrite(bp); 609 vp->v_lastw = lbn; 610 vp->v_lasta = bp->b_blkno; 611 } 612 613 614 /* 615 * This is an awful lot like cluster_rbuild...wish they could be combined. 616 * The last lbn argument is the current block on which I/O is being 617 * performed. Check to see that it doesn't fall in the middle of 618 * the current block (if last_bp == NULL). 619 */ 620 int 621 cluster_wbuild(vp, size, start_lbn, len) 622 struct vnode *vp; 623 long size; 624 daddr_t start_lbn; 625 int len; 626 { 627 struct buf *bp, *tbp; 628 int i, j, s; 629 int totalwritten = 0; 630 int dbsize = btodb(size); 631 632 if (vp->v_maxio == 0) 633 vp->v_maxio = DFLTPHYS; 634 while (len > 0) { 635 s = splbio(); 636 if (((tbp = gbincore(vp, start_lbn)) == NULL) || 637 ((tbp->b_flags & (B_INVAL|B_BUSY|B_DELWRI)) != B_DELWRI)) { 638 ++start_lbn; 639 --len; 640 splx(s); 641 continue; 642 } 643 bremfree(tbp); 644 tbp->b_flags |= B_BUSY; 645 tbp->b_flags &= ~B_DONE; 646 splx(s); 647 648 /* 649 * Extra memory in the buffer, punt on this buffer. XXX we could 650 * handle this in most cases, but we would have to push the extra 651 * memory down to after our max possible cluster size and then 652 * potentially pull it back up if the cluster was terminated 653 * prematurely--too much hassle. 654 */ 655 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) || 656 (tbp->b_bcount != tbp->b_bufsize) || 657 (tbp->b_bcount != size) || 658 (len == 1) || 659 ((bp = trypbuf(&cluster_pbuf_freecnt)) == NULL)) { 660 totalwritten += tbp->b_bufsize; 661 bawrite(tbp); 662 ++start_lbn; 663 --len; 664 continue; 665 } 666 667 /* 668 * We got a pbuf to make the cluster in. 669 * so initialise it. 670 */ 671 TAILQ_INIT(&bp->b_cluster.cluster_head); 672 bp->b_bcount = 0; 673 bp->b_bufsize = 0; 674 bp->b_npages = 0; 675 if (tbp->b_wcred != NOCRED) { 676 bp->b_wcred = tbp->b_wcred; 677 crhold(bp->b_wcred); 678 } 679 680 bp->b_blkno = tbp->b_blkno; 681 bp->b_lblkno = tbp->b_lblkno; 682 bp->b_offset = tbp->b_offset; 683 bp->b_data = (char *)((vm_offset_t)bp->b_data | 684 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 685 bp->b_flags |= B_CALL | B_BUSY | B_CLUSTER | 686 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT)); 687 bp->b_iodone = cluster_callback; 688 pbgetvp(vp, bp); 689 /* 690 * From this location in the file, scan forward to see 691 * if there are buffers with adjacent data that need to 692 * be written as well. 693 */ 694 for (i = 0; i < len; ++i, ++start_lbn) { 695 if (i != 0) { /* If not the first buffer */ 696 s = splbio(); 697 /* 698 * If the adjacent data is not even in core it 699 * can't need to be written. 700 */ 701 if ((tbp = gbincore(vp, start_lbn)) == NULL) { 702 splx(s); 703 break; 704 } 705 706 /* 707 * If it IS in core, but has different 708 * characteristics, don't cluster with it. 709 */ 710 if ((tbp->b_flags & 711 (B_VMIO | B_CLUSTEROK | B_INVAL | B_BUSY | 712 B_DELWRI | B_NEEDCOMMIT)) 713 != (B_DELWRI | B_CLUSTEROK | 714 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT)))) { 715 splx(s); 716 break; 717 } 718 719 if (tbp->b_wcred != bp->b_wcred) { 720 splx(s); 721 break; 722 } 723 724 /* 725 * Check that the combined cluster 726 * would make sense with regard to pages 727 * and would not be too large 728 */ 729 if ((tbp->b_bcount != size) || 730 ((bp->b_blkno + (dbsize * i)) != 731 tbp->b_blkno) || 732 ((tbp->b_npages + bp->b_npages) > 733 (vp->v_maxio / PAGE_SIZE))) { 734 splx(s); 735 break; 736 } 737 /* 738 * Ok, it's passed all the tests, 739 * so remove it from the free list 740 * and mark it busy. We will use it. 741 */ 742 bremfree(tbp); 743 tbp->b_flags |= B_BUSY; 744 tbp->b_flags &= ~B_DONE; 745 splx(s); 746 } /* end of code for non-first buffers only */ 747 /* check for latent dependencies to be handled */ 748 if ((LIST_FIRST(&tbp->b_dep)) != NULL && 749 bioops.io_start) 750 (*bioops.io_start)(tbp); 751 /* 752 * If the IO is via the VM then we do some 753 * special VM hackery. (yuck) 754 */ 755 if (tbp->b_flags & B_VMIO) { 756 vm_page_t m; 757 758 if (i != 0) { /* if not first buffer */ 759 for (j = 0; j < tbp->b_npages; j += 1) { 760 m = tbp->b_pages[j]; 761 if (m->flags & PG_BUSY) 762 goto finishcluster; 763 } 764 } 765 766 for (j = 0; j < tbp->b_npages; j += 1) { 767 m = tbp->b_pages[j]; 768 vm_page_io_start(m); 769 vm_object_pip_add(m->object, 1); 770 if ((bp->b_npages == 0) || 771 (bp->b_pages[bp->b_npages - 1] != m)) { 772 bp->b_pages[bp->b_npages] = m; 773 bp->b_npages++; 774 } 775 } 776 } 777 bp->b_bcount += size; 778 bp->b_bufsize += size; 779 780 s = splbio(); 781 bundirty(tbp); 782 tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR); 783 tbp->b_flags |= B_ASYNC; 784 reassignbuf(tbp, tbp->b_vp); /* put on clean list */ 785 ++tbp->b_vp->v_numoutput; 786 splx(s); 787 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 788 tbp, b_cluster.cluster_entry); 789 } 790 finishcluster: 791 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 792 (vm_page_t *) bp->b_pages, bp->b_npages); 793 if (bp->b_bufsize > bp->b_kvasize) 794 panic( 795 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 796 bp->b_bufsize, bp->b_kvasize); 797 bp->b_kvasize = bp->b_bufsize; 798 totalwritten += bp->b_bufsize; 799 bp->b_dirtyoff = 0; 800 bp->b_dirtyend = bp->b_bufsize; 801 bawrite(bp); 802 803 len -= i; 804 } 805 return totalwritten; 806 } 807 808 /* 809 * Collect together all the buffers in a cluster. 810 * Plus add one additional buffer. 811 */ 812 static struct cluster_save * 813 cluster_collectbufs(vp, last_bp) 814 struct vnode *vp; 815 struct buf *last_bp; 816 { 817 struct cluster_save *buflist; 818 struct buf *bp; 819 daddr_t lbn; 820 int i, len; 821 822 len = vp->v_lastw - vp->v_cstart + 1; 823 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), 824 M_SEGMENT, M_WAITOK); 825 buflist->bs_nchildren = 0; 826 buflist->bs_children = (struct buf **) (buflist + 1); 827 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) { 828 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp); 829 buflist->bs_children[i] = bp; 830 if (bp->b_blkno == bp->b_lblkno) 831 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 832 NULL, NULL); 833 } 834 buflist->bs_children[i] = bp = last_bp; 835 if (bp->b_blkno == bp->b_lblkno) 836 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 837 NULL, NULL); 838 buflist->bs_nchildren = i + 1; 839 return (buflist); 840 } 841