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  * 4. Neither the name of the University nor the names of its contributors
  16  *    may be used to endorse or promote products derived from this software
  17  *    without specific prior written permission.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  29  * SUCH DAMAGE.
  30  *
  31  *      @(#)vfs_cluster.c       8.7 (Berkeley) 2/13/94
  32  */
  33
  34 #include <sys/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include "opt_debug_cluster.h"
  38
  39 #include <sys/param.h>
  40 #include <sys/systm.h>
  41 #include <sys/kernel.h>
  42 #include <sys/proc.h>
  43 #include <sys/bio.h>
  44 #include <sys/buf.h>
  45 #include <sys/vnode.h>
  46 #include <sys/malloc.h>
  47 #include <sys/mount.h>
  48 #include <sys/resourcevar.h>
  49 #include <sys/vmmeter.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 static int      rcluster= 0;
  57 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0,
  58     "Debug VFS clustering code");
  59 #endif
  60
  61 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer");
  62
  63 static struct cluster_save *
  64         cluster_collectbufs(struct vnode *vp, struct buf *last_bp);
  65 static struct buf *
  66         cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
  67                          daddr_t blkno, long size, int run, struct buf *fbp);
  68 static void cluster_callback(struct buf *);
  69
  70 static int write_behind = 1;
  71 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
  72     "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
  73
  74 static int read_max = 8;
  75 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
  76     "Cluster read-ahead max block count");
  77
  78 /* Page expended to mark partially backed buffers */
  79 extern vm_page_t        bogus_page;
  80
  81 /*
  82  * Read data to a buf, including read-ahead if we find this to be beneficial.
  83  * cluster_read replaces bread.
  84  */
  85 int
  86 cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp)
  87         struct vnode *vp;
  88         u_quad_t filesize;
  89         daddr_t lblkno;
  90         long size;
  91         struct ucred *cred;
  92         long totread;
  93         int seqcount;
  94         struct buf **bpp;
  95 {
  96         struct buf *bp, *rbp, *reqbp;
  97         daddr_t blkno, origblkno;
  98         int maxra, racluster;
  99         int error, ncontig;
 100         int i;
 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 = seqcount;
 110         maxra = min(read_max, maxra);
 111         maxra = min(nbuf/8, maxra);
 112         if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
 113                 maxra = (filesize / size) - lblkno;
 114
 115         /*
 116          * get the requested block
 117          */
 118         *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, 0);
 119         origblkno = lblkno;
 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                         bp->b_flags &= ~B_RAM;
 133                         VI_LOCK(vp);
 134                         for (i = 1; i < maxra; i++) {
 135                                 /*
 136                                  * Stop if the buffer does not exist or it
 137                                  * is invalid (about to go away?)
 138                                  */
 139                                 rbp = gbincore(&vp->v_bufobj, lblkno+i);
 140                                 if (rbp == NULL || (rbp->b_flags & B_INVAL))
 141                                         break;
 142
 143                                 /*
 144                                  * Set another read-ahead mark so we know
 145                                  * to check again.
 146                                  */
 147                                 if (((i % racluster) == (racluster - 1)) ||
 148                                         (i == (maxra - 1)))
 149                                         rbp->b_flags |= B_RAM;
 150                         }
 151                         VI_UNLOCK(vp);
 152                         if (i >= maxra) {
 153                                 return 0;
 154                         }
 155                         lblkno += i;
 156                 }
 157                 reqbp = bp = NULL;
 158         /*
 159          * If it isn't in the cache, then get a chunk from
 160          * disk if sequential, otherwise just get the block.
 161          */
 162         } else {
 163                 off_t firstread = bp->b_offset;
 164                 int nblks;
 165
 166                 KASSERT(bp->b_offset != NOOFFSET,
 167                     ("cluster_read: no buffer offset"));
 168
 169                 ncontig = 0;
 170
 171                 /*
 172                  * Compute the total number of blocks that we should read
 173                  * synchronously.
 174                  */
 175                 if (firstread + totread > filesize)
 176                         totread = filesize - firstread;
 177                 nblks = howmany(totread, size);
 178                 if (nblks > racluster)
 179                         nblks = racluster;
 180
 181                 /*
 182                  * Now compute the number of contiguous blocks.
 183                  */
 184                 if (nblks > 1) {
 185                         error = VOP_BMAP(vp, lblkno, NULL,
 186                                 &blkno, &ncontig, NULL);
 187                         /*
 188                          * If this failed to map just do the original block.
 189                          */
 190                         if (error || blkno == -1)
 191                                 ncontig = 0;
 192                 }
 193
 194                 /*
 195                  * If we have contiguous data available do a cluster
 196                  * otherwise just read the requested block.
 197                  */
 198                 if (ncontig) {
 199                         /* Account for our first block. */
 200                         ncontig = min(ncontig + 1, nblks);
 201                         if (ncontig < nblks)
 202                                 nblks = ncontig;
 203                         bp = cluster_rbuild(vp, filesize, lblkno,
 204                                 blkno, size, nblks, bp);
 205                         lblkno += (bp->b_bufsize / size);
 206                 } else {
 207                         bp->b_flags |= B_RAM;
 208                         bp->b_iocmd = BIO_READ;
 209                         lblkno += 1;
 210                 }
 211         }
 212
 213         /*
 214          * handle the synchronous read so that it is available ASAP.
 215          */
 216         if (bp) {
 217                 if ((bp->b_flags & B_CLUSTER) == 0) {
 218                         vfs_busy_pages(bp, 0);
 219                 }
 220                 bp->b_flags &= ~B_INVAL;
 221                 bp->b_ioflags &= ~BIO_ERROR;
 222                 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
 223                         BUF_KERNPROC(bp);
 224                 bp->b_iooffset = dbtob(bp->b_blkno);
 225                 bstrategy(bp);
 226                 curproc->p_stats->p_ru.ru_inblock++;
 227         }
 228
 229         /*
 230          * If we have been doing sequential I/O, then do some read-ahead.
 231          */
 232         while (lblkno < (origblkno + maxra)) {
 233                 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
 234                 if (error)
 235                         break;
 236
 237                 if (blkno == -1)
 238                         break;
 239
 240                 /*
 241                  * We could throttle ncontig here by maxra but we might as
 242                  * well read the data if it is contiguous.  We're throttled
 243                  * by racluster anyway.
 244                  */
 245                 if (ncontig) {
 246                         ncontig = min(ncontig + 1, racluster);
 247                         rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
 248                                 size, ncontig, NULL);
 249                         lblkno += (rbp->b_bufsize / size);
 250                         if (rbp->b_flags & B_DELWRI) {
 251                                 bqrelse(rbp);
 252                                 continue;
 253                         }
 254                 } else {
 255                         rbp = getblk(vp, lblkno, size, 0, 0, 0);
 256                         lblkno += 1;
 257                         if (rbp->b_flags & B_DELWRI) {
 258                                 bqrelse(rbp);
 259                                 continue;
 260                         }
 261                         rbp->b_flags |= B_ASYNC | B_RAM;
 262                         rbp->b_iocmd = BIO_READ;
 263                         rbp->b_blkno = blkno;
 264                 }
 265                 if (rbp->b_flags & B_CACHE) {
 266                         rbp->b_flags &= ~B_ASYNC;
 267                         bqrelse(rbp);
 268                         continue;
 269                 }
 270                 if ((rbp->b_flags & B_CLUSTER) == 0) {
 271                         vfs_busy_pages(rbp, 0);
 272                 }
 273                 rbp->b_flags &= ~B_INVAL;
 274                 rbp->b_ioflags &= ~BIO_ERROR;
 275                 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
 276                         BUF_KERNPROC(rbp);
 277                 rbp->b_iooffset = dbtob(rbp->b_blkno);
 278                 bstrategy(rbp);
 279                 curproc->p_stats->p_ru.ru_inblock++;
 280         }
 281
 282         if (reqbp)
 283                 return (bufwait(reqbp));
 284         else
 285                 return (error);
 286 }
 287
 288 /*
 289  * If blocks are contiguous on disk, use this to provide clustered
 290  * read ahead.  We will read as many blocks as possible sequentially
 291  * and then parcel them up into logical blocks in the buffer hash table.
 292  */
 293 static struct buf *
 294 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
 295         struct vnode *vp;
 296         u_quad_t filesize;
 297         daddr_t lbn;
 298         daddr_t blkno;
 299         long size;
 300         int run;
 301         struct buf *fbp;
 302 {
 303         struct buf *bp, *tbp;
 304         daddr_t bn;
 305         int i, inc, j;
 306
 307         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 308             ("cluster_rbuild: size %ld != filesize %jd\n",
 309             size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
 310
 311         /*
 312          * avoid a division
 313          */
 314         while ((u_quad_t) size * (lbn + run) > filesize) {
 315                 --run;
 316         }
 317
 318         if (fbp) {
 319                 tbp = fbp;
 320                 tbp->b_iocmd = BIO_READ;
 321         } else {
 322                 tbp = getblk(vp, lbn, size, 0, 0, 0);
 323                 if (tbp->b_flags & B_CACHE)
 324                         return tbp;
 325                 tbp->b_flags |= B_ASYNC | B_RAM;
 326                 tbp->b_iocmd = BIO_READ;
 327         }
 328
 329         tbp->b_blkno = blkno;
 330         if( (tbp->b_flags & B_MALLOC) ||
 331                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
 332                 return tbp;
 333
 334         bp = trypbuf(&cluster_pbuf_freecnt);
 335         if (bp == 0)
 336                 return tbp;
 337
 338         /*
 339          * We are synthesizing a buffer out of vm_page_t's, but
 340          * if the block size is not page aligned then the starting
 341          * address may not be either.  Inherit the b_data offset
 342          * from the original buffer.
 343          */
 344         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 345             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 346         bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
 347         bp->b_iocmd = BIO_READ;
 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         inc = btodb(size);
 362         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
 363                 if (i != 0) {
 364                         if ((bp->b_npages * PAGE_SIZE) +
 365                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 366                                 break;
 367                         }
 368
 369                         tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT);
 370
 371                         /* Don't wait around for locked bufs. */
 372                         if (tbp == NULL)
 373                                 break;
 374
 375                         /*
 376                          * Stop scanning if the buffer is fully valid
 377                          * (marked B_CACHE), or locked (may be doing a
 378                          * background write), or if the buffer is not
 379                          * VMIO backed.  The clustering code can only deal
 380                          * with VMIO-backed buffers.
 381                          */
 382                         VI_LOCK(vp);
 383                         if ((tbp->b_vflags & BV_BKGRDINPROG) ||
 384                             (tbp->b_flags & B_CACHE) ||
 385                             (tbp->b_flags & B_VMIO) == 0) {
 386                                 VI_UNLOCK(vp);
 387                                 bqrelse(tbp);
 388                                 break;
 389                         }
 390                         VI_UNLOCK(vp);
 391
 392                         /*
 393                          * The buffer must be completely invalid in order to
 394                          * take part in the cluster.  If it is partially valid
 395                          * then we stop.
 396                          */
 397                         VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
 398                         for (j = 0;j < tbp->b_npages; j++) {
 399                                 VM_OBJECT_LOCK_ASSERT(tbp->b_pages[j]->object,
 400                                     MA_OWNED);
 401                                 if (tbp->b_pages[j]->valid)
 402                                         break;
 403                         }
 404                         VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
 405                         if (j != tbp->b_npages) {
 406                                 bqrelse(tbp);
 407                                 break;
 408                         }
 409
 410                         /*
 411                          * Set a read-ahead mark as appropriate
 412                          */
 413                         if ((fbp && (i == 1)) || (i == (run - 1)))
 414                                 tbp->b_flags |= B_RAM;
 415
 416                         /*
 417                          * Set the buffer up for an async read (XXX should
 418                          * we do this only if we do not wind up brelse()ing?).
 419                          * Set the block number if it isn't set, otherwise
 420                          * if it is make sure it matches the block number we
 421                          * expect.
 422                          */
 423                         tbp->b_flags |= B_ASYNC;
 424                         tbp->b_iocmd = BIO_READ;
 425                         if (tbp->b_blkno == tbp->b_lblkno) {
 426                                 tbp->b_blkno = bn;
 427                         } else if (tbp->b_blkno != bn) {
 428                                 brelse(tbp);
 429                                 break;
 430                         }
 431                 }
 432                 /*
 433                  * XXX fbp from caller may not be B_ASYNC, but we are going
 434                  * to biodone() it in cluster_callback() anyway
 435                  */
 436                 BUF_KERNPROC(tbp);
 437                 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 438                         tbp, b_cluster.cluster_entry);
 439                 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
 440                 for (j = 0; j < tbp->b_npages; j += 1) {
 441                         vm_page_t m;
 442                         m = tbp->b_pages[j];
 443                         vm_page_io_start(m);
 444                         vm_object_pip_add(m->object, 1);
 445                         if ((bp->b_npages == 0) ||
 446                                 (bp->b_pages[bp->b_npages-1] != m)) {
 447                                 bp->b_pages[bp->b_npages] = m;
 448                                 bp->b_npages++;
 449                         }
 450                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 451                                 tbp->b_pages[j] = bogus_page;
 452                 }
 453                 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
 454                 /*
 455                  * XXX shouldn't this be += size for both, like in
 456                  * cluster_wbuild()?
 457                  *
 458                  * Don't inherit tbp->b_bufsize as it may be larger due to
 459                  * a non-page-aligned size.  Instead just aggregate using
 460                  * 'size'.
 461                  */
 462                 if (tbp->b_bcount != size)
 463                         printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
 464                 if (tbp->b_bufsize != size)
 465                         printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
 466                 bp->b_bcount += size;
 467                 bp->b_bufsize += size;
 468         }
 469
 470         /*
 471          * Fully valid pages in the cluster are already good and do not need
 472          * to be re-read from disk.  Replace the page with bogus_page
 473          */
 474         VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
 475         for (j = 0; j < bp->b_npages; j++) {
 476                 VM_OBJECT_LOCK_ASSERT(bp->b_pages[j]->object, MA_OWNED);
 477                 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 478                     VM_PAGE_BITS_ALL) {
 479                         bp->b_pages[j] = bogus_page;
 480                 }
 481         }
 482         VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
 483         if (bp->b_bufsize > bp->b_kvasize)
 484                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 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_pages, bp->b_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 static void
 500 cluster_callback(bp)
 501         struct buf *bp;
 502 {
 503         struct buf *nbp, *tbp;
 504         int error = 0;
 505
 506         /*
 507          * Must propogate errors to all the components.
 508          */
 509         if (bp->b_ioflags & BIO_ERROR)
 510                 error = bp->b_error;
 511
 512         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
 513         /*
 514          * Move memory from the large cluster buffer into the component
 515          * buffers and mark IO as done on these.
 516          */
 517         for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
 518                 tbp; tbp = nbp) {
 519                 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
 520                 if (error) {
 521                         tbp->b_ioflags |= BIO_ERROR;
 522                         tbp->b_error = error;
 523                 } else {
 524                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 525                         tbp->b_flags &= ~B_INVAL;
 526                         tbp->b_ioflags &= ~BIO_ERROR;
 527                         /*
 528                          * XXX the bdwrite()/bqrelse() issued during
 529                          * cluster building clears B_RELBUF (see bqrelse()
 530                          * comment).  If direct I/O was specified, we have
 531                          * to restore it here to allow the buffer and VM
 532                          * to be freed.
 533                          */
 534                         if (tbp->b_flags & B_DIRECT)
 535                                 tbp->b_flags |= B_RELBUF;
 536                 }
 537                 bufdone(tbp);
 538         }
 539         pbrelvp(bp);
 540         relpbuf(bp, &cluster_pbuf_freecnt);
 541 }
 542
 543 /*
 544  *      cluster_wbuild_wb:
 545  *
 546  *      Implement modified write build for cluster.
 547  *
 548  *              write_behind = 0        write behind disabled
 549  *              write_behind = 1        write behind normal (default)
 550  *              write_behind = 2        write behind backed-off
 551  */
 552
 553 static __inline int
 554 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
 555 {
 556         int r = 0;
 557
 558         switch(write_behind) {
 559         case 2:
 560                 if (start_lbn < len)
 561                         break;
 562                 start_lbn -= len;
 563                 /* FALLTHROUGH */
 564         case 1:
 565                 r = cluster_wbuild(vp, size, start_lbn, len);
 566                 /* FALLTHROUGH */
 567         default:
 568                 /* FALLTHROUGH */
 569                 break;
 570         }
 571         return(r);
 572 }
 573
 574 /*
 575  * Do clustered write for FFS.
 576  *
 577  * Three cases:
 578  *      1. Write is not sequential (write asynchronously)
 579  *      Write is sequential:
 580  *      2.      beginning of cluster - begin cluster
 581  *      3.      middle of a cluster - add to cluster
 582  *      4.      end of a cluster - asynchronously write cluster
 583  */
 584 void
 585 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount)
 586 {
 587         daddr_t lbn;
 588         int maxclen, cursize;
 589         int lblocksize;
 590         int async;
 591
 592         if (vp->v_type == VREG) {
 593                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 594                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 595         } else {
 596                 async = 0;
 597                 lblocksize = bp->b_bufsize;
 598         }
 599         lbn = bp->b_lblkno;
 600         KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
 601
 602         /* Initialize vnode to beginning of file. */
 603         if (lbn == 0)
 604                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 605
 606         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
 607             (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
 608                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
 609                 if (vp->v_clen != 0) {
 610                         /*
 611                          * Next block is not sequential.
 612                          *
 613                          * If we are not writing at end of file, the process
 614                          * seeked to another point in the file since its last
 615                          * write, or we have reached our maximum cluster size,
 616                          * then push the previous cluster. Otherwise try
 617                          * reallocating to make it sequential.
 618                          *
 619                          * Change to algorithm: only push previous cluster if
 620                          * it was sequential from the point of view of the
 621                          * seqcount heuristic, otherwise leave the buffer
 622                          * intact so we can potentially optimize the I/O
 623                          * later on in the buf_daemon or update daemon
 624                          * flush.
 625                          */
 626                         cursize = vp->v_lastw - vp->v_cstart + 1;
 627                         if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
 628                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
 629                                 if (!async && seqcount > 0) {
 630                                         cluster_wbuild_wb(vp, lblocksize,
 631                                                 vp->v_cstart, cursize);
 632                                 }
 633                         } else {
 634                                 struct buf **bpp, **endbp;
 635                                 struct cluster_save *buflist;
 636
 637                                 buflist = cluster_collectbufs(vp, bp);
 638                                 endbp = &buflist->bs_children
 639                                     [buflist->bs_nchildren - 1];
 640                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 641                                         /*
 642                                          * Failed, push the previous cluster
 643                                          * if *really* writing sequentially
 644                                          * in the logical file (seqcount > 1),
 645                                          * otherwise delay it in the hopes that
 646                                          * the low level disk driver can
 647                                          * optimize the write ordering.
 648                                          */
 649                                         for (bpp = buflist->bs_children;
 650                                              bpp < endbp; bpp++)
 651                                                 brelse(*bpp);
 652                                         free(buflist, M_SEGMENT);
 653                                         if (seqcount > 1) {
 654                                                 cluster_wbuild_wb(vp,
 655                                                     lblocksize, vp->v_cstart,
 656                                                     cursize);
 657                                         }
 658                                 } else {
 659                                         /*
 660                                          * Succeeded, keep building cluster.
 661                                          */
 662                                         for (bpp = buflist->bs_children;
 663                                              bpp <= endbp; bpp++)
 664                                                 bdwrite(*bpp);
 665                                         free(buflist, M_SEGMENT);
 666                                         vp->v_lastw = lbn;
 667                                         vp->v_lasta = bp->b_blkno;
 668                                         return;
 669                                 }
 670                         }
 671                 }
 672                 /*
 673                  * Consider beginning a cluster. If at end of file, make
 674                  * cluster as large as possible, otherwise find size of
 675                  * existing cluster.
 676                  */
 677                 if ((vp->v_type == VREG) &&
 678                         ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
 679                     (bp->b_blkno == bp->b_lblkno) &&
 680                     (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
 681                      bp->b_blkno == -1)) {
 682                         bawrite(bp);
 683                         vp->v_clen = 0;
 684                         vp->v_lasta = bp->b_blkno;
 685                         vp->v_cstart = lbn + 1;
 686                         vp->v_lastw = lbn;
 687                         return;
 688                 }
 689                 vp->v_clen = maxclen;
 690                 if (!async && maxclen == 0) {   /* I/O not contiguous */
 691                         vp->v_cstart = lbn + 1;
 692                         bawrite(bp);
 693                 } else {        /* Wait for rest of cluster */
 694                         vp->v_cstart = lbn;
 695                         bdwrite(bp);
 696                 }
 697         } else if (lbn == vp->v_cstart + vp->v_clen) {
 698                 /*
 699                  * At end of cluster, write it out if seqcount tells us we
 700                  * are operating sequentially, otherwise let the buf or
 701                  * update daemon handle it.
 702                  */
 703                 bdwrite(bp);
 704                 if (seqcount > 1)
 705                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
 706                 vp->v_clen = 0;
 707                 vp->v_cstart = lbn + 1;
 708         } else if (vm_page_count_severe()) {
 709                 /*
 710                  * We are low on memory, get it going NOW
 711                  */
 712                 bawrite(bp);
 713         } else {
 714                 /*
 715                  * In the middle of a cluster, so just delay the I/O for now.
 716                  */
 717                 bdwrite(bp);
 718         }
 719         vp->v_lastw = lbn;
 720         vp->v_lasta = bp->b_blkno;
 721 }
 722
 723
 724 /*
 725  * This is an awful lot like cluster_rbuild...wish they could be combined.
 726  * The last lbn argument is the current block on which I/O is being
 727  * performed.  Check to see that it doesn't fall in the middle of
 728  * the current block (if last_bp == NULL).
 729  */
 730 int
 731 cluster_wbuild(vp, size, start_lbn, len)
 732         struct vnode *vp;
 733         long size;
 734         daddr_t start_lbn;
 735         int len;
 736 {
 737         struct buf *bp, *tbp;
 738         int i, j;
 739         int totalwritten = 0;
 740         int dbsize = btodb(size);
 741
 742         while (len > 0) {
 743                 /*
 744                  * If the buffer is not delayed-write (i.e. dirty), or it
 745                  * is delayed-write but either locked or inval, it cannot
 746                  * partake in the clustered write.
 747                  */
 748                 VI_LOCK(vp);
 749                 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
 750                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 751                         VI_UNLOCK(vp);
 752                         ++start_lbn;
 753                         --len;
 754                         continue;
 755                 }
 756                 if (BUF_LOCK(tbp,
 757                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, VI_MTX(vp))) {
 758                         ++start_lbn;
 759                         --len;
 760                         continue;
 761                 }
 762                 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
 763                         BUF_UNLOCK(tbp);
 764                         ++start_lbn;
 765                         --len;
 766                         continue;
 767                 }
 768                 bremfree(tbp);
 769                 tbp->b_flags &= ~B_DONE;
 770
 771                 /*
 772                  * Extra memory in the buffer, punt on this buffer.
 773                  * XXX we could handle this in most cases, but we would
 774                  * have to push the extra memory down to after our max
 775                  * possible cluster size and then potentially pull it back
 776                  * up if the cluster was terminated prematurely--too much
 777                  * hassle.
 778                  */
 779                 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
 780                      (B_CLUSTEROK | B_VMIO)) ||
 781                   (tbp->b_bcount != tbp->b_bufsize) ||
 782                   (tbp->b_bcount != size) ||
 783                   (len == 1) ||
 784                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 785                         totalwritten += tbp->b_bufsize;
 786                         bawrite(tbp);
 787                         ++start_lbn;
 788                         --len;
 789                         continue;
 790                 }
 791
 792                 /*
 793                  * We got a pbuf to make the cluster in.
 794                  * so initialise it.
 795                  */
 796                 TAILQ_INIT(&bp->b_cluster.cluster_head);
 797                 bp->b_bcount = 0;
 798                 bp->b_bufsize = 0;
 799                 bp->b_npages = 0;
 800                 if (tbp->b_wcred != NOCRED)
 801                         bp->b_wcred = crhold(tbp->b_wcred);
 802
 803                 bp->b_blkno = tbp->b_blkno;
 804                 bp->b_lblkno = tbp->b_lblkno;
 805                 bp->b_offset = tbp->b_offset;
 806
 807                 /*
 808                  * We are synthesizing a buffer out of vm_page_t's, but
 809                  * if the block size is not page aligned then the starting
 810                  * address may not be either.  Inherit the b_data offset
 811                  * from the original buffer.
 812                  */
 813                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 814                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 815                 bp->b_flags |= B_CLUSTER |
 816                                 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
 817                 bp->b_iodone = cluster_callback;
 818                 pbgetvp(vp, bp);
 819                 /*
 820                  * From this location in the file, scan forward to see
 821                  * if there are buffers with adjacent data that need to
 822                  * be written as well.
 823                  */
 824                 for (i = 0; i < len; ++i, ++start_lbn) {
 825                         if (i != 0) { /* If not the first buffer */
 826                                 /*
 827                                  * If the adjacent data is not even in core it
 828                                  * can't need to be written.
 829                                  */
 830                                 VI_LOCK(vp);
 831                                 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
 832                                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 833                                         VI_UNLOCK(vp);
 834                                         break;
 835                                 }
 836
 837                                 /*
 838                                  * If it IS in core, but has different
 839                                  * characteristics, or is locked (which
 840                                  * means it could be undergoing a background
 841                                  * I/O or be in a weird state), then don't
 842                                  * cluster with it.
 843                                  */
 844                                 if (BUF_LOCK(tbp,
 845                                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
 846                                     VI_MTX(vp)))
 847                                         break;
 848
 849                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 850                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 851                                     != (B_DELWRI | B_CLUSTEROK |
 852                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 853                                     tbp->b_wcred != bp->b_wcred) {
 854                                         BUF_UNLOCK(tbp);
 855                                         break;
 856                                 }
 857
 858                                 /*
 859                                  * Check that the combined cluster
 860                                  * would make sense with regard to pages
 861                                  * and would not be too large
 862                                  */
 863                                 if ((tbp->b_bcount != size) ||
 864                                   ((bp->b_blkno + (dbsize * i)) !=
 865                                     tbp->b_blkno) ||
 866                                   ((tbp->b_npages + bp->b_npages) >
 867                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 868                                         BUF_UNLOCK(tbp);
 869                                         break;
 870                                 }
 871                                 /*
 872                                  * Ok, it's passed all the tests,
 873                                  * so remove it from the free list
 874                                  * and mark it busy. We will use it.
 875                                  */
 876                                 bremfree(tbp);
 877                                 tbp->b_flags &= ~B_DONE;
 878                         } /* end of code for non-first buffers only */
 879                         /* check for latent dependencies to be handled */
 880                         if ((LIST_FIRST(&tbp->b_dep)) != NULL) {
 881                                 tbp->b_iocmd = BIO_WRITE;
 882                                 buf_start(tbp);
 883                         }
 884                         /*
 885                          * If the IO is via the VM then we do some
 886                          * special VM hackery (yuck).  Since the buffer's
 887                          * block size may not be page-aligned it is possible
 888                          * for a page to be shared between two buffers.  We
 889                          * have to get rid of the duplication when building
 890                          * the cluster.
 891                          */
 892                         if (tbp->b_flags & B_VMIO) {
 893                                 vm_page_t m;
 894
 895                                 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
 896                                 if (i != 0) { /* if not first buffer */
 897                                         for (j = 0; j < tbp->b_npages; j += 1) {
 898                                                 m = tbp->b_pages[j];
 899                                                 if (m->flags & PG_BUSY) {
 900                                                         VM_OBJECT_UNLOCK(
 901                                                             tbp->b_object);
 902                                                         bqrelse(tbp);
 903                                                         goto finishcluster;
 904                                                 }
 905                                         }
 906                                 }
 907                                 for (j = 0; j < tbp->b_npages; j += 1) {
 908                                         m = tbp->b_pages[j];
 909                                         vm_page_io_start(m);
 910                                         vm_object_pip_add(m->object, 1);
 911                                         if ((bp->b_npages == 0) ||
 912                                           (bp->b_pages[bp->b_npages - 1] != m)) {
 913                                                 bp->b_pages[bp->b_npages] = m;
 914                                                 bp->b_npages++;
 915                                         }
 916                                 }
 917                                 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
 918                         }
 919                         bp->b_bcount += size;
 920                         bp->b_bufsize += size;
 921                         bundirty(tbp);
 922                         tbp->b_flags &= ~B_DONE;
 923                         tbp->b_ioflags &= ~BIO_ERROR;
 924                         tbp->b_flags |= B_ASYNC;
 925                         tbp->b_iocmd = BIO_WRITE;
 926                         reassignbuf(tbp);               /* put on clean list */
 927                         bufobj_wref(tbp->b_bufobj);
 928                         BUF_KERNPROC(tbp);
 929                         TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 930                                 tbp, b_cluster.cluster_entry);
 931                 }
 932         finishcluster:
 933                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 934                         (vm_page_t *) bp->b_pages, bp->b_npages);
 935                 if (bp->b_bufsize > bp->b_kvasize)
 936                         panic(
 937                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 938                             bp->b_bufsize, bp->b_kvasize);
 939                 bp->b_kvasize = bp->b_bufsize;
 940                 totalwritten += bp->b_bufsize;
 941                 bp->b_dirtyoff = 0;
 942                 bp->b_dirtyend = bp->b_bufsize;
 943                 bawrite(bp);
 944
 945                 len -= i;
 946         }
 947         return totalwritten;
 948 }
 949
 950 /*
 951  * Collect together all the buffers in a cluster.
 952  * Plus add one additional buffer.
 953  */
 954 static struct cluster_save *
 955 cluster_collectbufs(vp, last_bp)
 956         struct vnode *vp;
 957         struct buf *last_bp;
 958 {
 959         struct cluster_save *buflist;
 960         struct buf *bp;
 961         daddr_t lbn;
 962         int i, len;
 963
 964         len = vp->v_lastw - vp->v_cstart + 1;
 965         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
 966             M_SEGMENT, M_WAITOK);
 967         buflist->bs_nchildren = 0;
 968         buflist->bs_children = (struct buf **) (buflist + 1);
 969         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
 970                 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
 971                 buflist->bs_children[i] = bp;
 972                 if (bp->b_blkno == bp->b_lblkno)
 973                         VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
 974                                 NULL, NULL);
 975         }
 976         buflist->bs_children[i] = bp = last_bp;
 977         if (bp->b_blkno == bp->b_lblkno)
 978                 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
 979         buflist->bs_nchildren = i + 1;
 980         return (buflist);
 981 }