sys/kern/vfs_cluster.c

   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: src/sys/kern/vfs_cluster.c,v 1.92.2.9 2001/11/18 07:10:59 dillon Exp $
  37  * $DragonFly: src/sys/kern/vfs_cluster.c,v 1.13 2005/06/06 15:02:28 dillon Exp $
  38  */
  39
  40 #include "opt_debug_cluster.h"
  41
  42 #include <sys/param.h>
  43 #include <sys/systm.h>
  44 #include <sys/kernel.h>
  45 #include <sys/proc.h>
  46 #include <sys/buf.h>
  47 #include <sys/vnode.h>
  48 #include <sys/malloc.h>
  49 #include <sys/mount.h>
  50 #include <sys/resourcevar.h>
  51 #include <sys/vmmeter.h>
  52 #include <vm/vm.h>
  53 #include <vm/vm_object.h>
  54 #include <vm/vm_page.h>
  55 #include <sys/sysctl.h>
  56 #include <sys/buf2.h>
  57 #include <vm/vm_page2.h>
  58
  59 #if defined(CLUSTERDEBUG)
  60 #include <sys/sysctl.h>
  61 static int      rcluster= 0;
  62 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
  63 #endif
  64
  65 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");
  66
  67 static struct cluster_save *
  68         cluster_collectbufs (struct vnode *vp, struct buf *last_bp);
  69 static struct buf *
  70         cluster_rbuild (struct vnode *vp, u_quad_t filesize, daddr_t lbn,
  71                             daddr_t blkno, long size, int run, struct buf *fbp);
  72
  73 static int write_behind = 1;
  74 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
  75
  76 extern vm_page_t        bogus_page;
  77
  78 extern int cluster_pbuf_freecnt;
  79
  80 /*
  81  * Maximum number of blocks for read-ahead.
  82  */
  83 #define MAXRA 32
  84
  85 /*
  86  * This replaces bread.
  87  */
  88 int
  89 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno,
  90         long size, long totread, int seqcount, 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
 101         /*
 102          * Try to limit the amount of read-ahead by a few
 103          * ad-hoc parameters.  This needs work!!!
 104          */
 105         racluster = vp->v_mount->mnt_iosize_max / size;
 106         maxra = 2 * racluster + (totread / size);
 107         if (maxra > MAXRA)
 108                 maxra = MAXRA;
 109         if (maxra > nbuf/8)
 110                 maxra = nbuf/8;
 111
 112         /*
 113          * get the requested block
 114          */
 115         *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0);
 116         origblkno = lblkno;
 117         origtotread = totread;
 118
 119         /*
 120          * if it is in the cache, then check to see if the reads have been
 121          * sequential.  If they have, then try some read-ahead, otherwise
 122          * back-off on prospective read-aheads.
 123          */
 124         if (bp->b_flags & B_CACHE) {
 125                 if (!seqcount) {
 126                         return 0;
 127                 } else if ((bp->b_flags & B_RAM) == 0) {
 128                         return 0;
 129                 } else {
 130                         struct buf *tbp;
 131                         bp->b_flags &= ~B_RAM;
 132                         /*
 133                          * We do the crit here so that there is no window
 134                          * between the incore and the b_usecount increment
 135                          * below.  We opt to keep the crit out of the loop
 136                          * for efficiency.
 137                          */
 138                         crit_enter();
 139                         for (i = 1; i < maxra; i++) {
 140
 141                                 if (!(tbp = incore(vp, lblkno+i))) {
 142                                         break;
 143                                 }
 144
 145                                 /*
 146                                  * Set another read-ahead mark so we know
 147                                  * to check again.
 148                                  */
 149                                 if (((i % racluster) == (racluster - 1)) ||
 150                                         (i == (maxra - 1)))
 151                                         tbp->b_flags |= B_RAM;
 152                         }
 153                         crit_exit();
 154                         if (i >= maxra) {
 155                                 return 0;
 156                         }
 157                         lblkno += i;
 158                 }
 159                 reqbp = bp = NULL;
 160         } else {
 161                 off_t firstread = bp->b_offset;
 162
 163                 KASSERT(bp->b_offset != NOOFFSET,
 164                     ("cluster_read: no buffer offset"));
 165                 if (firstread + totread > filesize)
 166                         totread = filesize - firstread;
 167                 if (totread > size) {
 168                         int nblks = 0;
 169                         int ncontigafter;
 170                         while (totread > 0) {
 171                                 nblks++;
 172                                 totread -= size;
 173                         }
 174                         if (nblks == 1)
 175                                 goto single_block_read;
 176                         if (nblks > racluster)
 177                                 nblks = racluster;
 178
 179                         error = VOP_BMAP(vp, lblkno, NULL,
 180                                 &blkno, &ncontigafter, NULL);
 181                         if (error)
 182                                 goto single_block_read;
 183                         if (blkno == -1)
 184                                 goto single_block_read;
 185                         if (ncontigafter == 0)
 186                                 goto single_block_read;
 187                         if (ncontigafter + 1 < nblks)
 188                                 nblks = ncontigafter + 1;
 189
 190                         bp = cluster_rbuild(vp, filesize, lblkno,
 191                                 blkno, size, nblks, bp);
 192                         lblkno += (bp->b_bufsize / size);
 193                 } else {
 194 single_block_read:
 195                         /*
 196                          * if it isn't in the cache, then get a chunk from
 197                          * disk if sequential, otherwise just get the block.
 198                          */
 199                         bp->b_flags |= B_READ | B_RAM;
 200                         lblkno += 1;
 201                 }
 202         }
 203
 204         /*
 205          * if we have been doing sequential I/O, then do some read-ahead
 206          */
 207         rbp = NULL;
 208         if (seqcount && (lblkno < (origblkno + seqcount))) {
 209                 /*
 210                  * we now build the read-ahead buffer if it is desirable.
 211                  */
 212                 if (((u_quad_t)(lblkno + 1) * size) <= filesize &&
 213                     !(error = VOP_BMAP(vp, lblkno, NULL, &blkno, &num_ra, NULL)) &&
 214                     blkno != -1) {
 215                         int nblksread;
 216                         int ntoread = num_ra + 1;
 217                         nblksread = (origtotread + size - 1) / size;
 218                         if (seqcount < nblksread)
 219                                 seqcount = nblksread;
 220                         if (seqcount < ntoread)
 221                                 ntoread = seqcount;
 222                         if (num_ra) {
 223                                 rbp = cluster_rbuild(vp, filesize, lblkno,
 224                                         blkno, size, ntoread, NULL);
 225                         } else {
 226                                 rbp = getblk(vp, lblkno, size, 0, 0);
 227                                 rbp->b_flags |= B_READ | B_ASYNC | B_RAM;
 228                                 rbp->b_blkno = blkno;
 229                         }
 230                 }
 231         }
 232
 233         /*
 234          * handle the synchronous read
 235          */
 236         if (bp) {
 237 #if defined(CLUSTERDEBUG)
 238                 if (rcluster)
 239                         printf("S(%ld,%ld,%d) ",
 240                             (long)bp->b_lblkno, bp->b_bcount, seqcount);
 241 #endif
 242                 if ((bp->b_flags & B_CLUSTER) == 0) {
 243                         vfs_busy_pages(bp, 0);
 244                 }
 245                 bp->b_flags &= ~(B_ERROR|B_INVAL);
 246                 if (bp->b_flags & (B_ASYNC|B_CALL))
 247                         BUF_KERNPROC(bp);
 248                 error = VOP_STRATEGY(vp, bp);
 249         }
 250
 251         /*
 252          * and if we have read-aheads, do them too
 253          */
 254         if (rbp) {
 255                 if (error) {
 256                         rbp->b_flags &= ~(B_ASYNC | B_READ);
 257                         brelse(rbp);
 258                 } else if (rbp->b_flags & B_CACHE) {
 259                         rbp->b_flags &= ~(B_ASYNC | B_READ);
 260                         bqrelse(rbp);
 261                 } else {
 262 #if defined(CLUSTERDEBUG)
 263                         if (rcluster) {
 264                                 if (bp)
 265                                         printf("A+(%ld,%ld,%ld,%d) ",
 266                                             (long)rbp->b_lblkno, rbp->b_bcount,
 267                                             (long)(rbp->b_lblkno - origblkno),
 268                                             seqcount);
 269                                 else
 270                                         printf("A(%ld,%ld,%ld,%d) ",
 271                                             (long)rbp->b_lblkno, rbp->b_bcount,
 272                                             (long)(rbp->b_lblkno - origblkno),
 273                                             seqcount);
 274                         }
 275 #endif
 276
 277                         if ((rbp->b_flags & B_CLUSTER) == 0) {
 278                                 vfs_busy_pages(rbp, 0);
 279                         }
 280                         rbp->b_flags &= ~(B_ERROR|B_INVAL);
 281                         if (rbp->b_flags & (B_ASYNC|B_CALL))
 282                                 BUF_KERNPROC(rbp);
 283                         (void) VOP_STRATEGY(vp, rbp);
 284                 }
 285         }
 286         if (reqbp)
 287                 return (biowait(reqbp));
 288         else
 289                 return (error);
 290 }
 291
 292 /*
 293  * If blocks are contiguous on disk, use this to provide clustered
 294  * read ahead.  We will read as many blocks as possible sequentially
 295  * and then parcel them up into logical blocks in the buffer hash table.
 296  */
 297 static struct buf *
 298 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
 299         daddr_t blkno, long size, int run, struct buf *fbp)
 300 {
 301         struct buf *bp, *tbp;
 302         daddr_t bn;
 303         int i, inc, j;
 304
 305         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 306             ("cluster_rbuild: size %ld != filesize %ld\n",
 307             size, vp->v_mount->mnt_stat.f_iosize));
 308
 309         /*
 310          * avoid a division
 311          */
 312         while ((u_quad_t) size * (lbn + run) > filesize) {
 313                 --run;
 314         }
 315
 316         if (fbp) {
 317                 tbp = fbp;
 318                 tbp->b_flags |= B_READ;
 319         } else {
 320                 tbp = getblk(vp, lbn, size, 0, 0);
 321                 if (tbp->b_flags & B_CACHE)
 322                         return tbp;
 323                 tbp->b_flags |= B_ASYNC | B_READ | B_RAM;
 324         }
 325
 326         tbp->b_blkno = blkno;
 327         if( (tbp->b_flags & B_MALLOC) ||
 328                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
 329                 return tbp;
 330
 331         bp = trypbuf(&cluster_pbuf_freecnt);
 332         if (bp == 0)
 333                 return tbp;
 334
 335         /*
 336          * We are synthesizing a buffer out of vm_page_t's, but
 337          * if the block size is not page aligned then the starting
 338          * address may not be either.  Inherit the b_data offset
 339          * from the original buffer.
 340          */
 341         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 342             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 343         bp->b_flags = B_ASYNC | B_READ | B_CALL | B_CLUSTER | B_VMIO;
 344         bp->b_iodone = cluster_callback;
 345         bp->b_blkno = blkno;
 346         bp->b_lblkno = lbn;
 347         bp->b_offset = tbp->b_offset;
 348         KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
 349         pbgetvp(vp, bp);
 350
 351         TAILQ_INIT(&bp->b_cluster.cluster_head);
 352
 353         bp->b_bcount = 0;
 354         bp->b_bufsize = 0;
 355         bp->b_xio.xio_npages = 0;
 356
 357         inc = btodb(size);
 358         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
 359                 if (i != 0) {
 360                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
 361                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 362                                 break;
 363                         }
 364
 365                         /*
 366                          * Shortcut some checks and try to avoid buffers that
 367                          * would block in the lock.  The same checks have to
 368                          * be made again after we officially get the buffer.
 369                          */
 370                         if ((tbp = incore(vp, lbn + i)) != NULL) {
 371                                 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
 372                                         break;
 373                                 BUF_UNLOCK(tbp);
 374
 375                                 for (j = 0; j < tbp->b_xio.xio_npages; j++) {
 376                                         if (tbp->b_xio.xio_pages[j]->valid)
 377                                                 break;
 378                                 }
 379
 380                                 if (j != tbp->b_xio.xio_npages)
 381                                         break;
 382
 383                                 if (tbp->b_bcount != size)
 384                                         break;
 385                         }
 386
 387                         tbp = getblk(vp, lbn + i, size, 0, 0);
 388
 389                         /*
 390                          * Stop scanning if the buffer is fuly valid
 391                          * (marked B_CACHE), or locked (may be doing a
 392                          * background write), or if the buffer is not
 393                          * VMIO backed.  The clustering code can only deal
 394                          * with VMIO-backed buffers.
 395                          */
 396                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 397                             (tbp->b_flags & B_VMIO) == 0) {
 398                                 bqrelse(tbp);
 399                                 break;
 400                         }
 401
 402                         /*
 403                          * The buffer must be completely invalid in order to
 404                          * take part in the cluster.  If it is partially valid
 405                          * then we stop.
 406                          */
 407                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
 408                                 if (tbp->b_xio.xio_pages[j]->valid)
 409                                         break;
 410                         }
 411                         if (j != tbp->b_xio.xio_npages) {
 412                                 bqrelse(tbp);
 413                                 break;
 414                         }
 415
 416                         /*
 417                          * Set a read-ahead mark as appropriate
 418                          */
 419                         if ((fbp && (i == 1)) || (i == (run - 1)))
 420                                 tbp->b_flags |= B_RAM;
 421
 422                         /*
 423                          * Set the buffer up for an async read (XXX should
 424                          * we do this only if we do not wind up brelse()ing?).
 425                          * Set the block number if it isn't set, otherwise
 426                          * if it is make sure it matches the block number we
 427                          * expect.
 428                          */
 429                         tbp->b_flags |= B_READ | B_ASYNC;
 430                         if (tbp->b_blkno == tbp->b_lblkno) {
 431                                 tbp->b_blkno = bn;
 432                         } else if (tbp->b_blkno != bn) {
 433                                 brelse(tbp);
 434                                 break;
 435                         }
 436                 }
 437                 /*
 438                  * XXX fbp from caller may not be B_ASYNC, but we are going
 439                  * to biodone() it in cluster_callback() anyway
 440                  */
 441                 BUF_KERNPROC(tbp);
 442                 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 443                         tbp, b_cluster.cluster_entry);
 444                 for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
 445                         vm_page_t m;
 446                         m = tbp->b_xio.xio_pages[j];
 447                         vm_page_io_start(m);
 448                         vm_object_pip_add(m->object, 1);
 449                         if ((bp->b_xio.xio_npages == 0) ||
 450                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
 451                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 452                                 bp->b_xio.xio_npages++;
 453                         }
 454                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 455                                 tbp->b_xio.xio_pages[j] = bogus_page;
 456                 }
 457                 /*
 458                  * XXX shouldn't this be += size for both, like in
 459                  * cluster_wbuild()?
 460                  *
 461                  * Don't inherit tbp->b_bufsize as it may be larger due to
 462                  * a non-page-aligned size.  Instead just aggregate using
 463                  * 'size'.
 464                  */
 465                 if (tbp->b_bcount != size)
 466                     printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
 467                 if (tbp->b_bufsize != size)
 468                     printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
 469                 bp->b_bcount += size;
 470                 bp->b_bufsize += size;
 471         }
 472
 473         /*
 474          * Fully valid pages in the cluster are already good and do not need
 475          * to be re-read from disk.  Replace the page with bogus_page
 476          */
 477         for (j = 0; j < bp->b_xio.xio_npages; j++) {
 478                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 479                     VM_PAGE_BITS_ALL) {
 480                         bp->b_xio.xio_pages[j] = bogus_page;
 481                 }
 482         }
 483         if (bp->b_bufsize > bp->b_kvasize)
 484                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)",
 485                     bp->b_bufsize, bp->b_kvasize);
 486         bp->b_kvasize = bp->b_bufsize;
 487
 488         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 489                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 490         return (bp);
 491 }
 492
 493 /*
 494  * Cleanup after a clustered read or write.
 495  * This is complicated by the fact that any of the buffers might have
 496  * extra memory (if there were no empty buffer headers at allocbuf time)
 497  * that we will need to shift around.
 498  */
 499 void
 500 cluster_callback(struct buf *bp)
 501 {
 502         struct buf *nbp, *tbp;
 503         int error = 0;
 504
 505         /*
 506          * Must propogate errors to all the components.
 507          */
 508         if (bp->b_flags & B_ERROR)
 509                 error = bp->b_error;
 510
 511         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
 512         /*
 513          * Move memory from the large cluster buffer into the component
 514          * buffers and mark IO as done on these.
 515          */
 516         for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
 517                 tbp; tbp = nbp) {
 518                 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
 519                 if (error) {
 520                         tbp->b_flags |= B_ERROR;
 521                         tbp->b_error = error;
 522                 } else {
 523                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 524                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 525                         /*
 526                          * XXX the bdwrite()/bqrelse() issued during
 527                          * cluster building clears B_RELBUF (see bqrelse()
 528                          * comment).  If direct I/O was specified, we have
 529                          * to restore it here to allow the buffer and VM
 530                          * to be freed.
 531                          */
 532                         if (tbp->b_flags & B_DIRECT)
 533                                 tbp->b_flags |= B_RELBUF;
 534                 }
 535                 biodone(tbp);
 536         }
 537         relpbuf(bp, &cluster_pbuf_freecnt);
 538 }
 539
 540 /*
 541  *      cluster_wbuild_wb:
 542  *
 543  *      Implement modified write build for cluster.
 544  *
 545  *              write_behind = 0        write behind disabled
 546  *              write_behind = 1        write behind normal (default)
 547  *              write_behind = 2        write behind backed-off
 548  */
 549
 550 static __inline int
 551 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
 552 {
 553         int r = 0;
 554
 555         switch(write_behind) {
 556         case 2:
 557                 if (start_lbn < len)
 558                         break;
 559                 start_lbn -= len;
 560                 /* fall through */
 561         case 1:
 562                 r = cluster_wbuild(vp, size, start_lbn, len);
 563                 /* fall through */
 564         default:
 565                 /* fall through */
 566                 break;
 567         }
 568         return(r);
 569 }
 570
 571 /*
 572  * Do clustered write for FFS.
 573  *
 574  * Three cases:
 575  *      1. Write is not sequential (write asynchronously)
 576  *      Write is sequential:
 577  *      2.      beginning of cluster - begin cluster
 578  *      3.      middle of a cluster - add to cluster
 579  *      4.      end of a cluster - asynchronously write cluster
 580  */
 581 void
 582 cluster_write(struct buf *bp, u_quad_t filesize, int seqcount)
 583 {
 584         struct vnode *vp;
 585         daddr_t lbn;
 586         int maxclen, cursize;
 587         int lblocksize;
 588         int async;
 589
 590         vp = bp->b_vp;
 591         if (vp->v_type == VREG) {
 592                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 593                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 594         } else {
 595                 async = 0;
 596                 lblocksize = bp->b_bufsize;
 597         }
 598         lbn = bp->b_lblkno;
 599         KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
 600
 601         /* Initialize vnode to beginning of file. */
 602         if (lbn == 0)
 603                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 604
 605         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
 606             (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
 607                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
 608                 if (vp->v_clen != 0) {
 609                         /*
 610                          * Next block is not sequential.
 611                          *
 612                          * If we are not writing at end of file, the process
 613                          * seeked to another point in the file since its last
 614                          * write, or we have reached our maximum cluster size,
 615                          * then push the previous cluster. Otherwise try
 616                          * reallocating to make it sequential.
 617                          *
 618                          * Change to algorithm: only push previous cluster if
 619                          * it was sequential from the point of view of the
 620                          * seqcount heuristic, otherwise leave the buffer
 621                          * intact so we can potentially optimize the I/O
 622                          * later on in the buf_daemon or update daemon
 623                          * flush.
 624                          */
 625                         cursize = vp->v_lastw - vp->v_cstart + 1;
 626                         if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
 627                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
 628                                 if (!async && seqcount > 0) {
 629                                         cluster_wbuild_wb(vp, lblocksize,
 630                                                 vp->v_cstart, cursize);
 631                                 }
 632                         } else {
 633                                 struct buf **bpp, **endbp;
 634                                 struct cluster_save *buflist;
 635
 636                                 buflist = cluster_collectbufs(vp, bp);
 637                                 endbp = &buflist->bs_children
 638                                     [buflist->bs_nchildren - 1];
 639                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 640                                         /*
 641                                          * Failed, push the previous cluster
 642                                          * if *really* writing sequentially
 643                                          * in the logical file (seqcount > 1),
 644                                          * otherwise delay it in the hopes that
 645                                          * the low level disk driver can
 646                                          * optimize the write ordering.
 647                                          */
 648                                         for (bpp = buflist->bs_children;
 649                                              bpp < endbp; bpp++)
 650                                                 brelse(*bpp);
 651                                         free(buflist, M_SEGMENT);
 652                                         if (seqcount > 1) {
 653                                                 cluster_wbuild_wb(vp,
 654                                                     lblocksize, vp->v_cstart,
 655                                                     cursize);
 656                                         }
 657                                 } else {
 658                                         /*
 659                                          * Succeeded, keep building cluster.
 660                                          */
 661                                         for (bpp = buflist->bs_children;
 662                                              bpp <= endbp; bpp++)
 663                                                 bdwrite(*bpp);
 664                                         free(buflist, M_SEGMENT);
 665                                         vp->v_lastw = lbn;
 666                                         vp->v_lasta = bp->b_blkno;
 667                                         return;
 668                                 }
 669                         }
 670                 }
 671                 /*
 672                  * Consider beginning a cluster. If at end of file, make
 673                  * cluster as large as possible, otherwise find size of
 674                  * existing cluster.
 675                  */
 676                 if ((vp->v_type == VREG) &&
 677                         ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
 678                     (bp->b_blkno == bp->b_lblkno) &&
 679                     (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
 680                      bp->b_blkno == -1)) {
 681                         bawrite(bp);
 682                         vp->v_clen = 0;
 683                         vp->v_lasta = bp->b_blkno;
 684                         vp->v_cstart = lbn + 1;
 685                         vp->v_lastw = lbn;
 686                         return;
 687                 }
 688                 vp->v_clen = maxclen;
 689                 if (!async && maxclen == 0) {   /* I/O not contiguous */
 690                         vp->v_cstart = lbn + 1;
 691                         bawrite(bp);
 692                 } else {        /* Wait for rest of cluster */
 693                         vp->v_cstart = lbn;
 694                         bdwrite(bp);
 695                 }
 696         } else if (lbn == vp->v_cstart + vp->v_clen) {
 697                 /*
 698                  * At end of cluster, write it out if seqcount tells us we
 699                  * are operating sequentially, otherwise let the buf or
 700                  * update daemon handle it.
 701                  */
 702                 bdwrite(bp);
 703                 if (seqcount > 1)
 704                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
 705                 vp->v_clen = 0;
 706                 vp->v_cstart = lbn + 1;
 707         } else if (vm_page_count_severe()) {
 708                 /*
 709                  * We are low on memory, get it going NOW
 710                  */
 711                 bawrite(bp);
 712         } else {
 713                 /*
 714                  * In the middle of a cluster, so just delay the I/O for now.
 715                  */
 716                 bdwrite(bp);
 717         }
 718         vp->v_lastw = lbn;
 719         vp->v_lasta = bp->b_blkno;
 720 }
 721
 722
 723 /*
 724  * This is an awful lot like cluster_rbuild...wish they could be combined.
 725  * The last lbn argument is the current block on which I/O is being
 726  * performed.  Check to see that it doesn't fall in the middle of
 727  * the current block (if last_bp == NULL).
 728  */
 729 int
 730 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len)
 731 {
 732         struct buf *bp, *tbp;
 733         int i, j;
 734         int totalwritten = 0;
 735         int dbsize = btodb(size);
 736
 737         while (len > 0) {
 738                 crit_enter();
 739                 /*
 740                  * If the buffer is not delayed-write (i.e. dirty), or it
 741                  * is delayed-write but either locked or inval, it cannot
 742                  * partake in the clustered write.
 743                  */
 744                 if (((tbp = gbincore(vp, start_lbn)) == NULL) ||
 745                   ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
 746                   BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 747                         ++start_lbn;
 748                         --len;
 749                         crit_exit();
 750                         continue;
 751                 }
 752                 bremfree(tbp);
 753                 tbp->b_flags &= ~B_DONE;
 754                 crit_exit();
 755
 756                 /*
 757                  * Extra memory in the buffer, punt on this buffer.
 758                  * XXX we could handle this in most cases, but we would
 759                  * have to push the extra memory down to after our max
 760                  * possible cluster size and then potentially pull it back
 761                  * up if the cluster was terminated prematurely--too much
 762                  * hassle.
 763                  */
 764                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 765                   (tbp->b_bcount != tbp->b_bufsize) ||
 766                   (tbp->b_bcount != size) ||
 767                   (len == 1) ||
 768                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 769                         totalwritten += tbp->b_bufsize;
 770                         bawrite(tbp);
 771                         ++start_lbn;
 772                         --len;
 773                         continue;
 774                 }
 775
 776                 /*
 777                  * We got a pbuf to make the cluster in.
 778                  * so initialise it.
 779                  */
 780                 TAILQ_INIT(&bp->b_cluster.cluster_head);
 781                 bp->b_bcount = 0;
 782                 bp->b_bufsize = 0;
 783                 bp->b_xio.xio_npages = 0;
 784                 bp->b_blkno = tbp->b_blkno;
 785                 bp->b_lblkno = tbp->b_lblkno;
 786                 bp->b_offset = tbp->b_offset;
 787
 788                 /*
 789                  * We are synthesizing a buffer out of vm_page_t's, but
 790                  * if the block size is not page aligned then the starting
 791                  * address may not be either.  Inherit the b_data offset
 792                  * from the original buffer.
 793                  */
 794                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 795                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 796                 bp->b_flags |= B_CALL | B_CLUSTER |
 797                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN));
 798                 bp->b_iodone = cluster_callback;
 799                 pbgetvp(vp, bp);
 800                 /*
 801                  * From this location in the file, scan forward to see
 802                  * if there are buffers with adjacent data that need to
 803                  * be written as well.
 804                  */
 805                 for (i = 0; i < len; ++i, ++start_lbn) {
 806                         if (i != 0) { /* If not the first buffer */
 807                                 crit_enter();
 808                                 /*
 809                                  * If the adjacent data is not even in core it
 810                                  * can't need to be written.
 811                                  */
 812                                 if ((tbp = gbincore(vp, start_lbn)) == NULL) {
 813                                         crit_exit();
 814                                         break;
 815                                 }
 816
 817                                 /*
 818                                  * If it IS in core, but has different
 819                                  * characteristics, or is locked (which
 820                                  * means it could be undergoing a background
 821                                  * I/O or be in a weird state), then don't
 822                                  * cluster with it.
 823                                  */
 824                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 825                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 826                                   != (B_DELWRI | B_CLUSTEROK |
 827                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 828                                     (tbp->b_flags & B_LOCKED) ||
 829                                     BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 830                                         crit_exit();
 831                                         break;
 832                                 }
 833
 834                                 /*
 835                                  * Check that the combined cluster
 836                                  * would make sense with regard to pages
 837                                  * and would not be too large
 838                                  */
 839                                 if ((tbp->b_bcount != size) ||
 840                                   ((bp->b_blkno + (dbsize * i)) !=
 841                                     tbp->b_blkno) ||
 842                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
 843                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 844                                         BUF_UNLOCK(tbp);
 845                                         crit_exit();
 846                                         break;
 847                                 }
 848                                 /*
 849                                  * Ok, it's passed all the tests,
 850                                  * so remove it from the free list
 851                                  * and mark it busy. We will use it.
 852                                  */
 853                                 bremfree(tbp);
 854                                 tbp->b_flags &= ~B_DONE;
 855                                 crit_exit();
 856                         } /* end of code for non-first buffers only */
 857                         /* check for latent dependencies to be handled */
 858                         if ((LIST_FIRST(&tbp->b_dep)) != NULL &&
 859                             bioops.io_start)
 860                                 (*bioops.io_start)(tbp);
 861                         /*
 862                          * If the IO is via the VM then we do some
 863                          * special VM hackery (yuck).  Since the buffer's
 864                          * block size may not be page-aligned it is possible
 865                          * for a page to be shared between two buffers.  We
 866                          * have to get rid of the duplication when building
 867                          * the cluster.
 868                          */
 869                         if (tbp->b_flags & B_VMIO) {
 870                                 vm_page_t m;
 871
 872                                 if (i != 0) { /* if not first buffer */
 873                                         for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
 874                                                 m = tbp->b_xio.xio_pages[j];
 875                                                 if (m->flags & PG_BUSY) {
 876                                                         bqrelse(tbp);
 877                                                         goto finishcluster;
 878                                                 }
 879                                         }
 880                                 }
 881
 882                                 for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
 883                                         m = tbp->b_xio.xio_pages[j];
 884                                         vm_page_io_start(m);
 885                                         vm_object_pip_add(m->object, 1);
 886                                         if ((bp->b_xio.xio_npages == 0) ||
 887                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
 888                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 889                                                 bp->b_xio.xio_npages++;
 890                                         }
 891                                 }
 892                         }
 893                         bp->b_bcount += size;
 894                         bp->b_bufsize += size;
 895
 896                         crit_enter();
 897                         bundirty(tbp);
 898                         tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
 899                         tbp->b_flags |= B_ASYNC;
 900                         reassignbuf(tbp, tbp->b_vp);    /* put on clean list */
 901                         ++tbp->b_vp->v_numoutput;
 902                         crit_exit();
 903                         BUF_KERNPROC(tbp);
 904                         TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 905                                 tbp, b_cluster.cluster_entry);
 906                 }
 907         finishcluster:
 908                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 909                         (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 910                 if (bp->b_bufsize > bp->b_kvasize)
 911                         panic(
 912                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 913                             bp->b_bufsize, bp->b_kvasize);
 914                 bp->b_kvasize = bp->b_bufsize;
 915                 totalwritten += bp->b_bufsize;
 916                 bp->b_dirtyoff = 0;
 917                 bp->b_dirtyend = bp->b_bufsize;
 918                 bawrite(bp);
 919
 920                 len -= i;
 921         }
 922         return totalwritten;
 923 }
 924
 925 /*
 926  * Collect together all the buffers in a cluster.
 927  * Plus add one additional buffer.
 928  */
 929 static struct cluster_save *
 930 cluster_collectbufs(struct vnode *vp, struct buf *last_bp)
 931 {
 932         struct cluster_save *buflist;
 933         struct buf *bp;
 934         daddr_t lbn;
 935         int i, len;
 936
 937         len = vp->v_lastw - vp->v_cstart + 1;
 938         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
 939             M_SEGMENT, M_WAITOK);
 940         buflist->bs_nchildren = 0;
 941         buflist->bs_children = (struct buf **) (buflist + 1);
 942         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
 943                 (void) bread(vp, lbn, last_bp->b_bcount, &bp);
 944                 buflist->bs_children[i] = bp;
 945                 if (bp->b_blkno == bp->b_lblkno)
 946                         VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
 947                                 NULL, NULL);
 948         }
 949         buflist->bs_children[i] = bp = last_bp;
 950         if (bp->b_blkno == bp->b_lblkno)
 951                 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
 952                         NULL, NULL);
 953         buflist->bs_nchildren = i + 1;
 954         return (buflist);
 955 }