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
  69 static int write_behind = 1;
  70 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
  71     "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
  72
  73 static int read_max = 8;
  74 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
  75     "Cluster read-ahead max block count");
  76
  77 /* Page expended to mark partially backed buffers */
  78 extern vm_page_t        bogus_page;
  79
  80 /*
  81  * Number of physical bufs (pbufs) this subsystem is allowed.
  82  * Manipulated by vm_pager.c
  83  */
  84 extern int cluster_pbuf_freecnt;
  85
  86 /*
  87  * Read data to a buf, including read-ahead if we find this to be beneficial.
  88  * cluster_read replaces bread.
  89  */
  90 int
  91 cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp)
  92         struct vnode *vp;
  93         u_quad_t filesize;
  94         daddr_t lblkno;
  95         long size;
  96         struct ucred *cred;
  97         long totread;
  98         int seqcount;
  99         struct buf **bpp;
 100 {
 101         struct buf *bp, *rbp, *reqbp;
 102         daddr_t blkno, origblkno;
 103         int maxra, racluster;
 104         int error, ncontig;
 105         int i;
 106
 107         error = 0;
 108
 109         /*
 110          * Try to limit the amount of read-ahead by a few
 111          * ad-hoc parameters.  This needs work!!!
 112          */
 113         racluster = vp->v_mount->mnt_iosize_max / size;
 114         maxra = seqcount;
 115         maxra = min(read_max, maxra);
 116         maxra = min(nbuf/8, maxra);
 117         if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
 118                 maxra = (filesize / size) - lblkno;
 119
 120         /*
 121          * get the requested block
 122          */
 123         *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, 0);
 124         origblkno = lblkno;
 125
 126         /*
 127          * if it is in the cache, then check to see if the reads have been
 128          * sequential.  If they have, then try some read-ahead, otherwise
 129          * back-off on prospective read-aheads.
 130          */
 131         if (bp->b_flags & B_CACHE) {
 132                 if (!seqcount) {
 133                         return 0;
 134                 } else if ((bp->b_flags & B_RAM) == 0) {
 135                         return 0;
 136                 } else {
 137                         int s;
 138                         bp->b_flags &= ~B_RAM;
 139                         /*
 140                          * We do the spl here so that there is no window
 141                          * between the incore and the b_usecount increment
 142                          * below.  We opt to keep the spl out of the loop
 143                          * for efficiency.
 144                          */
 145                         s = splbio();
 146                         VI_LOCK(vp);
 147                         for (i = 1; i < maxra; i++) {
 148                                 /*
 149                                  * Stop if the buffer does not exist or it
 150                                  * is invalid (about to go away?)
 151                                  */
 152                                 rbp = gbincore(&vp->v_bufobj, lblkno+i);
 153                                 if (rbp == NULL || (rbp->b_flags & B_INVAL))
 154                                         break;
 155
 156                                 /*
 157                                  * Set another read-ahead mark so we know
 158                                  * to check again.
 159                                  */
 160                                 if (((i % racluster) == (racluster - 1)) ||
 161                                         (i == (maxra - 1)))
 162                                         rbp->b_flags |= B_RAM;
 163                         }
 164                         VI_UNLOCK(vp);
 165                         splx(s);
 166                         if (i >= maxra) {
 167                                 return 0;
 168                         }
 169                         lblkno += i;
 170                 }
 171                 reqbp = bp = NULL;
 172         /*
 173          * If it isn't in the cache, then get a chunk from
 174          * disk if sequential, otherwise just get the block.
 175          */
 176         } else {
 177                 off_t firstread = bp->b_offset;
 178                 int nblks;
 179
 180                 KASSERT(bp->b_offset != NOOFFSET,
 181                     ("cluster_read: no buffer offset"));
 182
 183                 ncontig = 0;
 184
 185                 /*
 186                  * Compute the total number of blocks that we should read
 187                  * synchronously.
 188                  */
 189                 if (firstread + totread > filesize)
 190                         totread = filesize - firstread;
 191                 nblks = howmany(totread, size);
 192                 if (nblks > racluster)
 193                         nblks = racluster;
 194
 195                 /*
 196                  * Now compute the number of contiguous blocks.
 197                  */
 198                 if (nblks > 1) {
 199                         error = VOP_BMAP(vp, lblkno, NULL,
 200                                 &blkno, &ncontig, NULL);
 201                         /*
 202                          * If this failed to map just do the original block.
 203                          */
 204                         if (error || blkno == -1)
 205                                 ncontig = 0;
 206                 }
 207
 208                 /*
 209                  * If we have contiguous data available do a cluster
 210                  * otherwise just read the requested block.
 211                  */
 212                 if (ncontig) {
 213                         /* Account for our first block. */
 214                         ncontig = min(ncontig + 1, nblks);
 215                         if (ncontig < nblks)
 216                                 nblks = ncontig;
 217                         bp = cluster_rbuild(vp, filesize, lblkno,
 218                                 blkno, size, nblks, bp);
 219                         lblkno += (bp->b_bufsize / size);
 220                 } else {
 221                         bp->b_flags |= B_RAM;
 222                         bp->b_iocmd = BIO_READ;
 223                         lblkno += 1;
 224                 }
 225         }
 226
 227         /*
 228          * handle the synchronous read so that it is available ASAP.
 229          */
 230         if (bp) {
 231                 if ((bp->b_flags & B_CLUSTER) == 0) {
 232                         vfs_busy_pages(bp, 0);
 233                 }
 234                 bp->b_flags &= ~B_INVAL;
 235                 bp->b_ioflags &= ~BIO_ERROR;
 236                 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
 237                         BUF_KERNPROC(bp);
 238                 bp->b_iooffset = dbtob(bp->b_blkno);
 239                 bstrategy(bp);
 240                 curproc->p_stats->p_ru.ru_inblock++;
 241         }
 242
 243         /*
 244          * If we have been doing sequential I/O, then do some read-ahead.
 245          */
 246         while (lblkno < (origblkno + maxra)) {
 247                 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
 248                 if (error)
 249                         break;
 250
 251                 if (blkno == -1)
 252                         break;
 253
 254                 /*
 255                  * We could throttle ncontig here by maxra but we might as
 256                  * well read the data if it is contiguous.  We're throttled
 257                  * by racluster anyway.
 258                  */
 259                 if (ncontig) {
 260                         ncontig = min(ncontig + 1, racluster);
 261                         rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
 262                                 size, ncontig, NULL);
 263                         lblkno += (rbp->b_bufsize / size);
 264                         if (rbp->b_flags & B_DELWRI) {
 265                                 bqrelse(rbp);
 266                                 continue;
 267                         }
 268                 } else {
 269                         rbp = getblk(vp, lblkno, size, 0, 0, 0);
 270                         lblkno += 1;
 271                         if (rbp->b_flags & B_DELWRI) {
 272                                 bqrelse(rbp);
 273                                 continue;
 274                         }
 275                         rbp->b_flags |= B_ASYNC | B_RAM;
 276                         rbp->b_iocmd = BIO_READ;
 277                         rbp->b_blkno = blkno;
 278                 }
 279                 if (rbp->b_flags & B_CACHE) {
 280                         rbp->b_flags &= ~B_ASYNC;
 281                         bqrelse(rbp);
 282                         continue;
 283                 }
 284                 if ((rbp->b_flags & B_CLUSTER) == 0) {
 285                         vfs_busy_pages(rbp, 0);
 286                 }
 287                 rbp->b_flags &= ~B_INVAL;
 288                 rbp->b_ioflags &= ~BIO_ERROR;
 289                 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
 290                         BUF_KERNPROC(rbp);
 291                 rbp->b_iooffset = dbtob(rbp->b_blkno);
 292                 bstrategy(rbp);
 293                 curproc->p_stats->p_ru.ru_inblock++;
 294         }
 295
 296         if (reqbp)
 297                 return (bufwait(reqbp));
 298         else
 299                 return (error);
 300 }
 301
 302 /*
 303  * If blocks are contiguous on disk, use this to provide clustered
 304  * read ahead.  We will read as many blocks as possible sequentially
 305  * and then parcel them up into logical blocks in the buffer hash table.
 306  */
 307 static struct buf *
 308 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
 309         struct vnode *vp;
 310         u_quad_t filesize;
 311         daddr_t lbn;
 312         daddr_t blkno;
 313         long size;
 314         int run;
 315         struct buf *fbp;
 316 {
 317         struct buf *bp, *tbp;
 318         daddr_t bn;
 319         int i, inc, j;
 320
 321         GIANT_REQUIRED;
 322
 323         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 324             ("cluster_rbuild: size %ld != filesize %jd\n",
 325             size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
 326
 327         /*
 328          * avoid a division
 329          */
 330         while ((u_quad_t) size * (lbn + run) > filesize) {
 331                 --run;
 332         }
 333
 334         if (fbp) {
 335                 tbp = fbp;
 336                 tbp->b_iocmd = BIO_READ;
 337         } else {
 338                 tbp = getblk(vp, lbn, size, 0, 0, 0);
 339                 if (tbp->b_flags & B_CACHE)
 340                         return tbp;
 341                 tbp->b_flags |= B_ASYNC | B_RAM;
 342                 tbp->b_iocmd = BIO_READ;
 343         }
 344
 345         tbp->b_blkno = blkno;
 346         if( (tbp->b_flags & B_MALLOC) ||
 347                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
 348                 return tbp;
 349
 350         bp = trypbuf(&cluster_pbuf_freecnt);
 351         if (bp == 0)
 352                 return tbp;
 353
 354         /*
 355          * We are synthesizing a buffer out of vm_page_t's, but
 356          * if the block size is not page aligned then the starting
 357          * address may not be either.  Inherit the b_data offset
 358          * from the original buffer.
 359          */
 360         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 361             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 362         bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
 363         bp->b_iocmd = BIO_READ;
 364         bp->b_iodone = cluster_callback;
 365         bp->b_blkno = blkno;
 366         bp->b_lblkno = lbn;
 367         bp->b_offset = tbp->b_offset;
 368         KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
 369         pbgetvp(vp, bp);
 370
 371         TAILQ_INIT(&bp->b_cluster.cluster_head);
 372
 373         bp->b_bcount = 0;
 374         bp->b_bufsize = 0;
 375         bp->b_npages = 0;
 376
 377         inc = btodb(size);
 378         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
 379                 if (i != 0) {
 380                         if ((bp->b_npages * PAGE_SIZE) +
 381                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 382                                 break;
 383                         }
 384
 385                         tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT);
 386
 387                         /* Don't wait around for locked bufs. */
 388                         if (tbp == NULL)
 389                                 break;
 390
 391                         /*
 392                          * Stop scanning if the buffer is fully valid
 393                          * (marked B_CACHE), or locked (may be doing a
 394                          * background write), or if the buffer is not
 395                          * VMIO backed.  The clustering code can only deal
 396                          * with VMIO-backed buffers.
 397                          */
 398                         VI_LOCK(vp);
 399                         if ((tbp->b_vflags & BV_BKGRDINPROG) ||
 400                             (tbp->b_flags & B_CACHE) ||
 401                             (tbp->b_flags & B_VMIO) == 0) {
 402                                 VI_UNLOCK(vp);
 403                                 bqrelse(tbp);
 404                                 break;
 405                         }
 406                         VI_UNLOCK(vp);
 407
 408                         /*
 409                          * The buffer must be completely invalid in order to
 410                          * take part in the cluster.  If it is partially valid
 411                          * then we stop.
 412                          */
 413                         VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
 414                         for (j = 0;j < tbp->b_npages; j++) {
 415                                 VM_OBJECT_LOCK_ASSERT(tbp->b_pages[j]->object,
 416                                     MA_OWNED);
 417                                 if (tbp->b_pages[j]->valid)
 418                                         break;
 419                         }
 420                         VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
 421                         if (j != tbp->b_npages) {
 422                                 bqrelse(tbp);
 423                                 break;
 424                         }
 425
 426                         /*
 427                          * Set a read-ahead mark as appropriate
 428                          */
 429                         if ((fbp && (i == 1)) || (i == (run - 1)))
 430                                 tbp->b_flags |= B_RAM;
 431
 432                         /*
 433                          * Set the buffer up for an async read (XXX should
 434                          * we do this only if we do not wind up brelse()ing?).
 435                          * Set the block number if it isn't set, otherwise
 436                          * if it is make sure it matches the block number we
 437                          * expect.
 438                          */
 439                         tbp->b_flags |= B_ASYNC;
 440                         tbp->b_iocmd = BIO_READ;
 441                         if (tbp->b_blkno == tbp->b_lblkno) {
 442                                 tbp->b_blkno = bn;
 443                         } else if (tbp->b_blkno != bn) {
 444                                 brelse(tbp);
 445                                 break;
 446                         }
 447                 }
 448                 /*
 449                  * XXX fbp from caller may not be B_ASYNC, but we are going
 450                  * to biodone() it in cluster_callback() anyway
 451                  */
 452                 BUF_KERNPROC(tbp);
 453                 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 454                         tbp, b_cluster.cluster_entry);
 455                 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
 456                 vm_page_lock_queues();
 457                 for (j = 0; j < tbp->b_npages; j += 1) {
 458                         vm_page_t m;
 459                         m = tbp->b_pages[j];
 460                         vm_page_io_start(m);
 461                         vm_object_pip_add(m->object, 1);
 462                         if ((bp->b_npages == 0) ||
 463                                 (bp->b_pages[bp->b_npages-1] != m)) {
 464                                 bp->b_pages[bp->b_npages] = m;
 465                                 bp->b_npages++;
 466                         }
 467                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 468                                 tbp->b_pages[j] = bogus_page;
 469                 }
 470                 vm_page_unlock_queues();
 471                 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
 472                 /*
 473                  * XXX shouldn't this be += size for both, like in
 474                  * cluster_wbuild()?
 475                  *
 476                  * Don't inherit tbp->b_bufsize as it may be larger due to
 477                  * a non-page-aligned size.  Instead just aggregate using
 478                  * 'size'.
 479                  */
 480                 if (tbp->b_bcount != size)
 481                         printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
 482                 if (tbp->b_bufsize != size)
 483                         printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
 484                 bp->b_bcount += size;
 485                 bp->b_bufsize += size;
 486         }
 487
 488         /*
 489          * Fully valid pages in the cluster are already good and do not need
 490          * to be re-read from disk.  Replace the page with bogus_page
 491          */
 492         VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
 493         for (j = 0; j < bp->b_npages; j++) {
 494                 VM_OBJECT_LOCK_ASSERT(bp->b_pages[j]->object, MA_OWNED);
 495                 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 496                     VM_PAGE_BITS_ALL) {
 497                         bp->b_pages[j] = bogus_page;
 498                 }
 499         }
 500         VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
 501         if (bp->b_bufsize > bp->b_kvasize)
 502                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 503                     bp->b_bufsize, bp->b_kvasize);
 504         bp->b_kvasize = bp->b_bufsize;
 505
 506         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 507                 (vm_page_t *)bp->b_pages, bp->b_npages);
 508         return (bp);
 509 }
 510
 511 /*
 512  * Cleanup after a clustered read or write.
 513  * This is complicated by the fact that any of the buffers might have
 514  * extra memory (if there were no empty buffer headers at allocbuf time)
 515  * that we will need to shift around.
 516  */
 517 void
 518 cluster_callback(bp)
 519         struct buf *bp;
 520 {
 521         struct buf *nbp, *tbp;
 522         int error = 0;
 523
 524         GIANT_REQUIRED;
 525
 526         /*
 527          * Must propogate errors to all the components.
 528          */
 529         if (bp->b_ioflags & BIO_ERROR)
 530                 error = bp->b_error;
 531
 532         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
 533         /*
 534          * Move memory from the large cluster buffer into the component
 535          * buffers and mark IO as done on these.
 536          */
 537         for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
 538                 tbp; tbp = nbp) {
 539                 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
 540                 if (error) {
 541                         tbp->b_ioflags |= BIO_ERROR;
 542                         tbp->b_error = error;
 543                 } else {
 544                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 545                         tbp->b_flags &= ~B_INVAL;
 546                         tbp->b_ioflags &= ~BIO_ERROR;
 547                         /*
 548                          * XXX the bdwrite()/bqrelse() issued during
 549                          * cluster building clears B_RELBUF (see bqrelse()
 550                          * comment).  If direct I/O was specified, we have
 551                          * to restore it here to allow the buffer and VM
 552                          * to be freed.
 553                          */
 554                         if (tbp->b_flags & B_DIRECT)
 555                                 tbp->b_flags |= B_RELBUF;
 556                 }
 557                 bufdone(tbp);
 558         }
 559         pbrelvp(bp);
 560         relpbuf(bp, &cluster_pbuf_freecnt);
 561 }
 562
 563 /*
 564  *      cluster_wbuild_wb:
 565  *
 566  *      Implement modified write build for cluster.
 567  *
 568  *              write_behind = 0        write behind disabled
 569  *              write_behind = 1        write behind normal (default)
 570  *              write_behind = 2        write behind backed-off
 571  */
 572
 573 static __inline int
 574 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
 575 {
 576         int r = 0;
 577
 578         switch(write_behind) {
 579         case 2:
 580                 if (start_lbn < len)
 581                         break;
 582                 start_lbn -= len;
 583                 /* FALLTHROUGH */
 584         case 1:
 585                 r = cluster_wbuild(vp, size, start_lbn, len);
 586                 /* FALLTHROUGH */
 587         default:
 588                 /* FALLTHROUGH */
 589                 break;
 590         }
 591         return(r);
 592 }
 593
 594 /*
 595  * Do clustered write for FFS.
 596  *
 597  * Three cases:
 598  *      1. Write is not sequential (write asynchronously)
 599  *      Write is sequential:
 600  *      2.      beginning of cluster - begin cluster
 601  *      3.      middle of a cluster - add to cluster
 602  *      4.      end of a cluster - asynchronously write cluster
 603  */
 604 void
 605 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount)
 606 {
 607         daddr_t lbn;
 608         int maxclen, cursize;
 609         int lblocksize;
 610         int async;
 611
 612         if (vp->v_type == VREG) {
 613                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 614                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 615         } else {
 616                 async = 0;
 617                 lblocksize = bp->b_bufsize;
 618         }
 619         lbn = bp->b_lblkno;
 620         KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
 621
 622         /* Initialize vnode to beginning of file. */
 623         if (lbn == 0)
 624                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 625
 626         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
 627             (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
 628                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
 629                 if (vp->v_clen != 0) {
 630                         /*
 631                          * Next block is not sequential.
 632                          *
 633                          * If we are not writing at end of file, the process
 634                          * seeked to another point in the file since its last
 635                          * write, or we have reached our maximum cluster size,
 636                          * then push the previous cluster. Otherwise try
 637                          * reallocating to make it sequential.
 638                          *
 639                          * Change to algorithm: only push previous cluster if
 640                          * it was sequential from the point of view of the
 641                          * seqcount heuristic, otherwise leave the buffer
 642                          * intact so we can potentially optimize the I/O
 643                          * later on in the buf_daemon or update daemon
 644                          * flush.
 645                          */
 646                         cursize = vp->v_lastw - vp->v_cstart + 1;
 647                         if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
 648                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
 649                                 if (!async && seqcount > 0) {
 650                                         cluster_wbuild_wb(vp, lblocksize,
 651                                                 vp->v_cstart, cursize);
 652                                 }
 653                         } else {
 654                                 struct buf **bpp, **endbp;
 655                                 struct cluster_save *buflist;
 656
 657                                 buflist = cluster_collectbufs(vp, bp);
 658                                 endbp = &buflist->bs_children
 659                                     [buflist->bs_nchildren - 1];
 660                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 661                                         /*
 662                                          * Failed, push the previous cluster
 663                                          * if *really* writing sequentially
 664                                          * in the logical file (seqcount > 1),
 665                                          * otherwise delay it in the hopes that
 666                                          * the low level disk driver can
 667                                          * optimize the write ordering.
 668                                          */
 669                                         for (bpp = buflist->bs_children;
 670                                              bpp < endbp; bpp++)
 671                                                 brelse(*bpp);
 672                                         free(buflist, M_SEGMENT);
 673                                         if (seqcount > 1) {
 674                                                 cluster_wbuild_wb(vp,
 675                                                     lblocksize, vp->v_cstart,
 676                                                     cursize);
 677                                         }
 678                                 } else {
 679                                         /*
 680                                          * Succeeded, keep building cluster.
 681                                          */
 682                                         for (bpp = buflist->bs_children;
 683                                              bpp <= endbp; bpp++)
 684                                                 bdwrite(*bpp);
 685                                         free(buflist, M_SEGMENT);
 686                                         vp->v_lastw = lbn;
 687                                         vp->v_lasta = bp->b_blkno;
 688                                         return;
 689                                 }
 690                         }
 691                 }
 692                 /*
 693                  * Consider beginning a cluster. If at end of file, make
 694                  * cluster as large as possible, otherwise find size of
 695                  * existing cluster.
 696                  */
 697                 if ((vp->v_type == VREG) &&
 698                         ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
 699                     (bp->b_blkno == bp->b_lblkno) &&
 700                     (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
 701                      bp->b_blkno == -1)) {
 702                         bawrite(bp);
 703                         vp->v_clen = 0;
 704                         vp->v_lasta = bp->b_blkno;
 705                         vp->v_cstart = lbn + 1;
 706                         vp->v_lastw = lbn;
 707                         return;
 708                 }
 709                 vp->v_clen = maxclen;
 710                 if (!async && maxclen == 0) {   /* I/O not contiguous */
 711                         vp->v_cstart = lbn + 1;
 712                         bawrite(bp);
 713                 } else {        /* Wait for rest of cluster */
 714                         vp->v_cstart = lbn;
 715                         bdwrite(bp);
 716                 }
 717         } else if (lbn == vp->v_cstart + vp->v_clen) {
 718                 /*
 719                  * At end of cluster, write it out if seqcount tells us we
 720                  * are operating sequentially, otherwise let the buf or
 721                  * update daemon handle it.
 722                  */
 723                 bdwrite(bp);
 724                 if (seqcount > 1)
 725                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
 726                 vp->v_clen = 0;
 727                 vp->v_cstart = lbn + 1;
 728         } else if (vm_page_count_severe()) {
 729                 /*
 730                  * We are low on memory, get it going NOW
 731                  */
 732                 bawrite(bp);
 733         } else {
 734                 /*
 735                  * In the middle of a cluster, so just delay the I/O for now.
 736                  */
 737                 bdwrite(bp);
 738         }
 739         vp->v_lastw = lbn;
 740         vp->v_lasta = bp->b_blkno;
 741 }
 742
 743
 744 /*
 745  * This is an awful lot like cluster_rbuild...wish they could be combined.
 746  * The last lbn argument is the current block on which I/O is being
 747  * performed.  Check to see that it doesn't fall in the middle of
 748  * the current block (if last_bp == NULL).
 749  */
 750 int
 751 cluster_wbuild(vp, size, start_lbn, len)
 752         struct vnode *vp;
 753         long size;
 754         daddr_t start_lbn;
 755         int len;
 756 {
 757         struct buf *bp, *tbp;
 758         int i, j, s;
 759         int totalwritten = 0;
 760         int dbsize = btodb(size);
 761
 762         GIANT_REQUIRED;
 763
 764         while (len > 0) {
 765                 s = splbio();
 766                 /*
 767                  * If the buffer is not delayed-write (i.e. dirty), or it
 768                  * is delayed-write but either locked or inval, it cannot
 769                  * partake in the clustered write.
 770                  */
 771                 VI_LOCK(vp);
 772                 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
 773                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 774                         VI_UNLOCK(vp);
 775                         ++start_lbn;
 776                         --len;
 777                         splx(s);
 778                         continue;
 779                 }
 780                 if (BUF_LOCK(tbp,
 781                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, VI_MTX(vp))) {
 782                         ++start_lbn;
 783                         --len;
 784                         splx(s);
 785                         continue;
 786                 }
 787                 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
 788                         BUF_UNLOCK(tbp);
 789                         ++start_lbn;
 790                         --len;
 791                         splx(s);
 792                         continue;
 793                 }
 794                 bremfree(tbp);
 795                 tbp->b_flags &= ~B_DONE;
 796                 splx(s);
 797
 798                 /*
 799                  * Extra memory in the buffer, punt on this buffer.
 800                  * XXX we could handle this in most cases, but we would
 801                  * have to push the extra memory down to after our max
 802                  * possible cluster size and then potentially pull it back
 803                  * up if the cluster was terminated prematurely--too much
 804                  * hassle.
 805                  */
 806                 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
 807                      (B_CLUSTEROK | B_VMIO)) ||
 808                   (tbp->b_bcount != tbp->b_bufsize) ||
 809                   (tbp->b_bcount != size) ||
 810                   (len == 1) ||
 811                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 812                         totalwritten += tbp->b_bufsize;
 813                         bawrite(tbp);
 814                         ++start_lbn;
 815                         --len;
 816                         continue;
 817                 }
 818
 819                 /*
 820                  * We got a pbuf to make the cluster in.
 821                  * so initialise it.
 822                  */
 823                 TAILQ_INIT(&bp->b_cluster.cluster_head);
 824                 bp->b_bcount = 0;
 825                 bp->b_bufsize = 0;
 826                 bp->b_npages = 0;
 827                 if (tbp->b_wcred != NOCRED)
 828                         bp->b_wcred = crhold(tbp->b_wcred);
 829
 830                 bp->b_blkno = tbp->b_blkno;
 831                 bp->b_lblkno = tbp->b_lblkno;
 832                 bp->b_offset = tbp->b_offset;
 833
 834                 /*
 835                  * We are synthesizing a buffer out of vm_page_t's, but
 836                  * if the block size is not page aligned then the starting
 837                  * address may not be either.  Inherit the b_data offset
 838                  * from the original buffer.
 839                  */
 840                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 841                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 842                 bp->b_flags |= B_CLUSTER |
 843                                 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
 844                 bp->b_iodone = cluster_callback;
 845                 pbgetvp(vp, bp);
 846                 /*
 847                  * From this location in the file, scan forward to see
 848                  * if there are buffers with adjacent data that need to
 849                  * be written as well.
 850                  */
 851                 for (i = 0; i < len; ++i, ++start_lbn) {
 852                         if (i != 0) { /* If not the first buffer */
 853                                 s = splbio();
 854                                 /*
 855                                  * If the adjacent data is not even in core it
 856                                  * can't need to be written.
 857                                  */
 858                                 VI_LOCK(vp);
 859                                 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
 860                                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 861                                         VI_UNLOCK(vp);
 862                                         splx(s);
 863                                         break;
 864                                 }
 865
 866                                 /*
 867                                  * If it IS in core, but has different
 868                                  * characteristics, or is locked (which
 869                                  * means it could be undergoing a background
 870                                  * I/O or be in a weird state), then don't
 871                                  * cluster with it.
 872                                  */
 873                                 if (BUF_LOCK(tbp,
 874                                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
 875                                     VI_MTX(vp))) {
 876                                         splx(s);
 877                                         break;
 878                                 }
 879
 880                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 881                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 882                                     != (B_DELWRI | B_CLUSTEROK |
 883                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 884                                     tbp->b_wcred != bp->b_wcred) {
 885                                         BUF_UNLOCK(tbp);
 886                                         splx(s);
 887                                         break;
 888                                 }
 889
 890                                 /*
 891                                  * Check that the combined cluster
 892                                  * would make sense with regard to pages
 893                                  * and would not be too large
 894                                  */
 895                                 if ((tbp->b_bcount != size) ||
 896                                   ((bp->b_blkno + (dbsize * i)) !=
 897                                     tbp->b_blkno) ||
 898                                   ((tbp->b_npages + bp->b_npages) >
 899                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 900                                         BUF_UNLOCK(tbp);
 901                                         splx(s);
 902                                         break;
 903                                 }
 904                                 /*
 905                                  * Ok, it's passed all the tests,
 906                                  * so remove it from the free list
 907                                  * and mark it busy. We will use it.
 908                                  */
 909                                 bremfree(tbp);
 910                                 tbp->b_flags &= ~B_DONE;
 911                                 splx(s);
 912                         } /* end of code for non-first buffers only */
 913                         /* check for latent dependencies to be handled */
 914                         if ((LIST_FIRST(&tbp->b_dep)) != NULL) {
 915                                 tbp->b_iocmd = BIO_WRITE;
 916                                 buf_start(tbp);
 917                         }
 918                         /*
 919                          * If the IO is via the VM then we do some
 920                          * special VM hackery (yuck).  Since the buffer's
 921                          * block size may not be page-aligned it is possible
 922                          * for a page to be shared between two buffers.  We
 923                          * have to get rid of the duplication when building
 924                          * the cluster.
 925                          */
 926                         if (tbp->b_flags & B_VMIO) {
 927                                 vm_page_t m;
 928
 929                                 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
 930                                 if (i != 0) { /* if not first buffer */
 931                                         for (j = 0; j < tbp->b_npages; j += 1) {
 932                                                 m = tbp->b_pages[j];
 933                                                 if (m->flags & PG_BUSY) {
 934                                                         VM_OBJECT_UNLOCK(
 935                                                             tbp->b_object);
 936                                                         bqrelse(tbp);
 937                                                         goto finishcluster;
 938                                                 }
 939                                         }
 940                                 }
 941                                 vm_page_lock_queues();
 942                                 for (j = 0; j < tbp->b_npages; j += 1) {
 943                                         m = tbp->b_pages[j];
 944                                         vm_page_io_start(m);
 945                                         vm_object_pip_add(m->object, 1);
 946                                         if ((bp->b_npages == 0) ||
 947                                           (bp->b_pages[bp->b_npages - 1] != m)) {
 948                                                 bp->b_pages[bp->b_npages] = m;
 949                                                 bp->b_npages++;
 950                                         }
 951                                 }
 952                                 vm_page_unlock_queues();
 953                                 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
 954                         }
 955                         bp->b_bcount += size;
 956                         bp->b_bufsize += size;
 957
 958                         s = splbio();
 959                         bundirty(tbp);
 960                         tbp->b_flags &= ~B_DONE;
 961                         tbp->b_ioflags &= ~BIO_ERROR;
 962                         tbp->b_flags |= B_ASYNC;
 963                         tbp->b_iocmd = BIO_WRITE;
 964                         reassignbuf(tbp);               /* put on clean list */
 965                         bufobj_wref(tbp->b_bufobj);
 966                         splx(s);
 967                         BUF_KERNPROC(tbp);
 968                         TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 969                                 tbp, b_cluster.cluster_entry);
 970                 }
 971         finishcluster:
 972                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 973                         (vm_page_t *) bp->b_pages, bp->b_npages);
 974                 if (bp->b_bufsize > bp->b_kvasize)
 975                         panic(
 976                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 977                             bp->b_bufsize, bp->b_kvasize);
 978                 bp->b_kvasize = bp->b_bufsize;
 979                 totalwritten += bp->b_bufsize;
 980                 bp->b_dirtyoff = 0;
 981                 bp->b_dirtyend = bp->b_bufsize;
 982                 bawrite(bp);
 983
 984                 len -= i;
 985         }
 986         return totalwritten;
 987 }
 988
 989 /*
 990  * Collect together all the buffers in a cluster.
 991  * Plus add one additional buffer.
 992  */
 993 static struct cluster_save *
 994 cluster_collectbufs(vp, last_bp)
 995         struct vnode *vp;
 996         struct buf *last_bp;
 997 {
 998         struct cluster_save *buflist;
 999         struct buf *bp;
1000         daddr_t lbn;
1001         int i, len;
1002
1003         len = vp->v_lastw - vp->v_cstart + 1;
1004         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1005             M_SEGMENT, M_WAITOK);
1006         buflist->bs_nchildren = 0;
1007         buflist->bs_children = (struct buf **) (buflist + 1);
1008         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1009                 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
1010                 buflist->bs_children[i] = bp;
1011                 if (bp->b_blkno == bp->b_lblkno)
1012                         VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1013                                 NULL, NULL);
1014         }
1015         buflist->bs_children[i] = bp = last_bp;
1016         if (bp->b_blkno == bp->b_lblkno)
1017                 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1018         buflist->bs_nchildren = i + 1;
1019         return (buflist);
1020 }