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 1.2 2003/06/17 04:28:42 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
  57 #if defined(CLUSTERDEBUG)
  58 #include <sys/sysctl.h>
  59 static int      rcluster= 0;
  60 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
  61 #endif
  62
  63 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer");
  64
  65 static struct cluster_save *
  66         cluster_collectbufs __P((struct vnode *vp, struct buf *last_bp));
  67 static struct buf *
  68         cluster_rbuild __P((struct vnode *vp, u_quad_t filesize, daddr_t lbn,
  69                             daddr_t blkno, long size, int run, struct buf *fbp));
  70
  71 static int write_behind = 1;
  72 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
  73
  74 extern vm_page_t        bogus_page;
  75
  76 extern int cluster_pbuf_freecnt;
  77
  78 /*
  79  * Maximum number of blocks for read-ahead.
  80  */
  81 #define MAXRA 32
  82
  83 /*
  84  * This replaces bread.
  85  */
  86 int
  87 cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp)
  88         struct vnode *vp;
  89         u_quad_t filesize;
  90         daddr_t lblkno;
  91         long size;
  92         struct ucred *cred;
  93         long totread;
  94         int seqcount;
  95         struct buf **bpp;
  96 {
  97         struct buf *bp, *rbp, *reqbp;
  98         daddr_t blkno, origblkno;
  99         int error, num_ra;
 100         int i;
 101         int maxra, racluster;
 102         long origtotread;
 103
 104         error = 0;
 105
 106         /*
 107          * Try to limit the amount of read-ahead by a few
 108          * ad-hoc parameters.  This needs work!!!
 109          */
 110         racluster = vp->v_mount->mnt_iosize_max / size;
 111         maxra = 2 * racluster + (totread / size);
 112         if (maxra > MAXRA)
 113                 maxra = MAXRA;
 114         if (maxra > nbuf/8)
 115                 maxra = nbuf/8;
 116
 117         /*
 118          * get the requested block
 119          */
 120         *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0);
 121         origblkno = lblkno;
 122         origtotread = totread;
 123
 124         /*
 125          * if it is in the cache, then check to see if the reads have been
 126          * sequential.  If they have, then try some read-ahead, otherwise
 127          * back-off on prospective read-aheads.
 128          */
 129         if (bp->b_flags & B_CACHE) {
 130                 if (!seqcount) {
 131                         return 0;
 132                 } else if ((bp->b_flags & B_RAM) == 0) {
 133                         return 0;
 134                 } else {
 135                         int s;
 136                         struct buf *tbp;
 137                         bp->b_flags &= ~B_RAM;
 138                         /*
 139                          * We do the spl here so that there is no window
 140                          * between the incore and the b_usecount increment
 141                          * below.  We opt to keep the spl out of the loop
 142                          * for efficiency.
 143                          */
 144                         s = splbio();
 145                         for (i = 1; i < maxra; i++) {
 146
 147                                 if (!(tbp = incore(vp, lblkno+i))) {
 148                                         break;
 149                                 }
 150
 151                                 /*
 152                                  * Set another read-ahead mark so we know
 153                                  * to check again.
 154                                  */
 155                                 if (((i % racluster) == (racluster - 1)) ||
 156                                         (i == (maxra - 1)))
 157                                         tbp->b_flags |= B_RAM;
 158                         }
 159                         splx(s);
 160                         if (i >= maxra) {
 161                                 return 0;
 162                         }
 163                         lblkno += i;
 164                 }
 165                 reqbp = bp = NULL;
 166         } else {
 167                 off_t firstread = bp->b_offset;
 168
 169                 KASSERT(bp->b_offset != NOOFFSET,
 170                     ("cluster_read: no buffer offset"));
 171                 if (firstread + totread > filesize)
 172                         totread = filesize - firstread;
 173                 if (totread > size) {
 174                         int nblks = 0;
 175                         int ncontigafter;
 176                         while (totread > 0) {
 177                                 nblks++;
 178                                 totread -= size;
 179                         }
 180                         if (nblks == 1)
 181                                 goto single_block_read;
 182                         if (nblks > racluster)
 183                                 nblks = racluster;
 184
 185                         error = VOP_BMAP(vp, lblkno, NULL,
 186                                 &blkno, &ncontigafter, NULL);
 187                         if (error)
 188                                 goto single_block_read;
 189                         if (blkno == -1)
 190                                 goto single_block_read;
 191                         if (ncontigafter == 0)
 192                                 goto single_block_read;
 193                         if (ncontigafter + 1 < nblks)
 194                                 nblks = ncontigafter + 1;
 195
 196                         bp = cluster_rbuild(vp, filesize, lblkno,
 197                                 blkno, size, nblks, bp);
 198                         lblkno += (bp->b_bufsize / size);
 199                 } else {
 200 single_block_read:
 201                         /*
 202                          * if it isn't in the cache, then get a chunk from
 203                          * disk if sequential, otherwise just get the block.
 204                          */
 205                         bp->b_flags |= B_READ | B_RAM;
 206                         lblkno += 1;
 207                 }
 208         }
 209
 210         /*
 211          * if we have been doing sequential I/O, then do some read-ahead
 212          */
 213         rbp = NULL;
 214         if (seqcount && (lblkno < (origblkno + seqcount))) {
 215                 /*
 216                  * we now build the read-ahead buffer if it is desirable.
 217                  */
 218                 if (((u_quad_t)(lblkno + 1) * size) <= filesize &&
 219                     !(error = VOP_BMAP(vp, lblkno, NULL, &blkno, &num_ra, NULL)) &&
 220                     blkno != -1) {
 221                         int nblksread;
 222                         int ntoread = num_ra + 1;
 223                         nblksread = (origtotread + size - 1) / size;
 224                         if (seqcount < nblksread)
 225                                 seqcount = nblksread;
 226                         if (seqcount < ntoread)
 227                                 ntoread = seqcount;
 228                         if (num_ra) {
 229                                 rbp = cluster_rbuild(vp, filesize, lblkno,
 230                                         blkno, size, ntoread, NULL);
 231                         } else {
 232                                 rbp = getblk(vp, lblkno, size, 0, 0);
 233                                 rbp->b_flags |= B_READ | B_ASYNC | B_RAM;
 234                                 rbp->b_blkno = blkno;
 235                         }
 236                 }
 237         }
 238
 239         /*
 240          * handle the synchronous read
 241          */
 242         if (bp) {
 243 #if defined(CLUSTERDEBUG)
 244                 if (rcluster)
 245                         printf("S(%ld,%ld,%d) ",
 246                             (long)bp->b_lblkno, bp->b_bcount, seqcount);
 247 #endif
 248                 if ((bp->b_flags & B_CLUSTER) == 0) {
 249                         vfs_busy_pages(bp, 0);
 250                 }
 251                 bp->b_flags &= ~(B_ERROR|B_INVAL);
 252                 if (bp->b_flags & (B_ASYNC|B_CALL))
 253                         BUF_KERNPROC(bp);
 254                 error = VOP_STRATEGY(vp, bp);
 255                 curproc->p_stats->p_ru.ru_inblock++;
 256         }
 257
 258         /*
 259          * and if we have read-aheads, do them too
 260          */
 261         if (rbp) {
 262                 if (error) {
 263                         rbp->b_flags &= ~(B_ASYNC | B_READ);
 264                         brelse(rbp);
 265                 } else if (rbp->b_flags & B_CACHE) {
 266                         rbp->b_flags &= ~(B_ASYNC | B_READ);
 267                         bqrelse(rbp);
 268                 } else {
 269 #if defined(CLUSTERDEBUG)
 270                         if (rcluster) {
 271                                 if (bp)
 272                                         printf("A+(%ld,%ld,%ld,%d) ",
 273                                             (long)rbp->b_lblkno, rbp->b_bcount,
 274                                             (long)(rbp->b_lblkno - origblkno),
 275                                             seqcount);
 276                                 else
 277                                         printf("A(%ld,%ld,%ld,%d) ",
 278                                             (long)rbp->b_lblkno, rbp->b_bcount,
 279                                             (long)(rbp->b_lblkno - origblkno),
 280                                             seqcount);
 281                         }
 282 #endif
 283
 284                         if ((rbp->b_flags & B_CLUSTER) == 0) {
 285                                 vfs_busy_pages(rbp, 0);
 286                         }
 287                         rbp->b_flags &= ~(B_ERROR|B_INVAL);
 288                         if (rbp->b_flags & (B_ASYNC|B_CALL))
 289                                 BUF_KERNPROC(rbp);
 290                         (void) VOP_STRATEGY(vp, rbp);
 291                         curproc->p_stats->p_ru.ru_inblock++;
 292                 }
 293         }
 294         if (reqbp)
 295                 return (biowait(reqbp));
 296         else
 297                 return (error);
 298 }
 299
 300 /*
 301  * If blocks are contiguous on disk, use this to provide clustered
 302  * read ahead.  We will read as many blocks as possible sequentially
 303  * and then parcel them up into logical blocks in the buffer hash table.
 304  */
 305 static struct buf *
 306 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
 307         struct vnode *vp;
 308         u_quad_t filesize;
 309         daddr_t lbn;
 310         daddr_t blkno;
 311         long size;
 312         int run;
 313         struct buf *fbp;
 314 {
 315         struct buf *bp, *tbp;
 316         daddr_t bn;
 317         int i, inc, j;
 318
 319         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 320             ("cluster_rbuild: size %ld != filesize %ld\n",
 321             size, vp->v_mount->mnt_stat.f_iosize));
 322
 323         /*
 324          * avoid a division
 325          */
 326         while ((u_quad_t) size * (lbn + run) > filesize) {
 327                 --run;
 328         }
 329
 330         if (fbp) {
 331                 tbp = fbp;
 332                 tbp->b_flags |= B_READ;
 333         } else {
 334                 tbp = getblk(vp, lbn, size, 0, 0);
 335                 if (tbp->b_flags & B_CACHE)
 336                         return tbp;
 337                 tbp->b_flags |= B_ASYNC | B_READ | B_RAM;
 338         }
 339
 340         tbp->b_blkno = blkno;
 341         if( (tbp->b_flags & B_MALLOC) ||
 342                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
 343                 return tbp;
 344
 345         bp = trypbuf(&cluster_pbuf_freecnt);
 346         if (bp == 0)
 347                 return tbp;
 348
 349         /*
 350          * We are synthesizing a buffer out of vm_page_t's, but
 351          * if the block size is not page aligned then the starting
 352          * address may not be either.  Inherit the b_data offset
 353          * from the original buffer.
 354          */
 355         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 356             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 357         bp->b_flags = B_ASYNC | B_READ | B_CALL | B_CLUSTER | B_VMIO;
 358         bp->b_iodone = cluster_callback;
 359         bp->b_blkno = blkno;
 360         bp->b_lblkno = lbn;
 361         bp->b_offset = tbp->b_offset;
 362         KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
 363         pbgetvp(vp, bp);
 364
 365         TAILQ_INIT(&bp->b_cluster.cluster_head);
 366
 367         bp->b_bcount = 0;
 368         bp->b_bufsize = 0;
 369         bp->b_npages = 0;
 370
 371         inc = btodb(size);
 372         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
 373                 if (i != 0) {
 374                         if ((bp->b_npages * PAGE_SIZE) +
 375                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 376                                 break;
 377                         }
 378
 379                         /*
 380                          * Shortcut some checks and try to avoid buffers that
 381                          * would block in the lock.  The same checks have to
 382                          * be made again after we officially get the buffer.
 383                          */
 384                         if ((tbp = incore(vp, lbn + i)) != NULL) {
 385                                 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
 386                                         break;
 387                                 BUF_UNLOCK(tbp);
 388
 389                                 for (j = 0; j < tbp->b_npages; j++) {
 390                                         if (tbp->b_pages[j]->valid)
 391                                                 break;
 392                                 }
 393
 394                                 if (j != tbp->b_npages)
 395                                         break;
 396
 397                                 if (tbp->b_bcount != size)
 398                                         break;
 399                         }
 400
 401                         tbp = getblk(vp, lbn + i, size, 0, 0);
 402
 403                         /*
 404                          * Stop scanning if the buffer is fuly valid
 405                          * (marked B_CACHE), or locked (may be doing a
 406                          * background write), or if the buffer is not
 407                          * VMIO backed.  The clustering code can only deal
 408                          * with VMIO-backed buffers.
 409                          */
 410                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 411                             (tbp->b_flags & B_VMIO) == 0) {
 412                                 bqrelse(tbp);
 413                                 break;
 414                         }
 415
 416                         /*
 417                          * The buffer must be completely invalid in order to
 418                          * take part in the cluster.  If it is partially valid
 419                          * then we stop.
 420                          */
 421                         for (j = 0;j < tbp->b_npages; j++) {
 422                                 if (tbp->b_pages[j]->valid)
 423                                         break;
 424                         }
 425                         if (j != tbp->b_npages) {
 426                                 bqrelse(tbp);
 427                                 break;
 428                         }
 429
 430                         /*
 431                          * Set a read-ahead mark as appropriate
 432                          */
 433                         if ((fbp && (i == 1)) || (i == (run - 1)))
 434                                 tbp->b_flags |= B_RAM;
 435
 436                         /*
 437                          * Set the buffer up for an async read (XXX should
 438                          * we do this only if we do not wind up brelse()ing?).
 439                          * Set the block number if it isn't set, otherwise
 440                          * if it is make sure it matches the block number we
 441                          * expect.
 442                          */
 443                         tbp->b_flags |= B_READ | B_ASYNC;
 444                         if (tbp->b_blkno == tbp->b_lblkno) {
 445                                 tbp->b_blkno = bn;
 446                         } else if (tbp->b_blkno != bn) {
 447                                 brelse(tbp);
 448                                 break;
 449                         }
 450                 }
 451                 /*
 452                  * XXX fbp from caller may not be B_ASYNC, but we are going
 453                  * to biodone() it in cluster_callback() anyway
 454                  */
 455                 BUF_KERNPROC(tbp);
 456                 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 457                         tbp, b_cluster.cluster_entry);
 458                 for (j = 0; j < tbp->b_npages; j += 1) {
 459                         vm_page_t m;
 460                         m = tbp->b_pages[j];
 461                         vm_page_io_start(m);
 462                         vm_object_pip_add(m->object, 1);
 463                         if ((bp->b_npages == 0) ||
 464                                 (bp->b_pages[bp->b_npages-1] != m)) {
 465                                 bp->b_pages[bp->b_npages] = m;
 466                                 bp->b_npages++;
 467                         }
 468                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 469                                 tbp->b_pages[j] = bogus_page;
 470                 }
 471                 /*
 472                  * XXX shouldn't this be += size for both, like in
 473                  * cluster_wbuild()?
 474                  *
 475                  * Don't inherit tbp->b_bufsize as it may be larger due to
 476                  * a non-page-aligned size.  Instead just aggregate using
 477                  * 'size'.
 478                  */
 479                 if (tbp->b_bcount != size)
 480                     printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
 481                 if (tbp->b_bufsize != size)
 482                     printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
 483                 bp->b_bcount += size;
 484                 bp->b_bufsize += size;
 485         }
 486
 487         /*
 488          * Fully valid pages in the cluster are already good and do not need
 489          * to be re-read from disk.  Replace the page with bogus_page
 490          */
 491         for (j = 0; j < bp->b_npages; j++) {
 492                 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 493                     VM_PAGE_BITS_ALL) {
 494                         bp->b_pages[j] = bogus_page;
 495                 }
 496         }
 497         if (bp->b_bufsize > bp->b_kvasize)
 498                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 499                     bp->b_bufsize, bp->b_kvasize);
 500         bp->b_kvasize = bp->b_bufsize;
 501
 502         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 503                 (vm_page_t *)bp->b_pages, bp->b_npages);
 504         return (bp);
 505 }
 506
 507 /*
 508  * Cleanup after a clustered read or write.
 509  * This is complicated by the fact that any of the buffers might have
 510  * extra memory (if there were no empty buffer headers at allocbuf time)
 511  * that we will need to shift around.
 512  */
 513 void
 514 cluster_callback(bp)
 515         struct buf *bp;
 516 {
 517         struct buf *nbp, *tbp;
 518         int error = 0;
 519
 520         /*
 521          * Must propogate errors to all the components.
 522          */
 523         if (bp->b_flags & B_ERROR)
 524                 error = bp->b_error;
 525
 526         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
 527         /*
 528          * Move memory from the large cluster buffer into the component
 529          * buffers and mark IO as done on these.
 530          */
 531         for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
 532                 tbp; tbp = nbp) {
 533                 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
 534                 if (error) {
 535                         tbp->b_flags |= B_ERROR;
 536                         tbp->b_error = error;
 537                 } else {
 538                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 539                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 540                         /*
 541                          * XXX the bdwrite()/bqrelse() issued during
 542                          * cluster building clears B_RELBUF (see bqrelse()
 543                          * comment).  If direct I/O was specified, we have
 544                          * to restore it here to allow the buffer and VM
 545                          * to be freed.
 546                          */
 547                         if (tbp->b_flags & B_DIRECT)
 548                                 tbp->b_flags |= B_RELBUF;
 549                 }
 550                 biodone(tbp);
 551         }
 552         relpbuf(bp, &cluster_pbuf_freecnt);
 553 }
 554
 555 /*
 556  *      cluster_wbuild_wb:
 557  *
 558  *      Implement modified write build for cluster.
 559  *
 560  *              write_behind = 0        write behind disabled
 561  *              write_behind = 1        write behind normal (default)
 562  *              write_behind = 2        write behind backed-off
 563  */
 564
 565 static __inline int
 566 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
 567 {
 568         int r = 0;
 569
 570         switch(write_behind) {
 571         case 2:
 572                 if (start_lbn < len)
 573                         break;
 574                 start_lbn -= len;
 575                 /* fall through */
 576         case 1:
 577                 r = cluster_wbuild(vp, size, start_lbn, len);
 578                 /* fall through */
 579         default:
 580                 /* fall through */
 581                 break;
 582         }
 583         return(r);
 584 }
 585
 586 /*
 587  * Do clustered write for FFS.
 588  *
 589  * Three cases:
 590  *      1. Write is not sequential (write asynchronously)
 591  *      Write is sequential:
 592  *      2.      beginning of cluster - begin cluster
 593  *      3.      middle of a cluster - add to cluster
 594  *      4.      end of a cluster - asynchronously write cluster
 595  */
 596 void
 597 cluster_write(bp, filesize, seqcount)
 598         struct buf *bp;
 599         u_quad_t filesize;
 600         int seqcount;
 601 {
 602         struct vnode *vp;
 603         daddr_t lbn;
 604         int maxclen, cursize;
 605         int lblocksize;
 606         int async;
 607
 608         vp = bp->b_vp;
 609         if (vp->v_type == VREG) {
 610                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 611                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 612         } else {
 613                 async = 0;
 614                 lblocksize = bp->b_bufsize;
 615         }
 616         lbn = bp->b_lblkno;
 617         KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
 618
 619         /* Initialize vnode to beginning of file. */
 620         if (lbn == 0)
 621                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 622
 623         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
 624             (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
 625                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
 626                 if (vp->v_clen != 0) {
 627                         /*
 628                          * Next block is not sequential.
 629                          *
 630                          * If we are not writing at end of file, the process
 631                          * seeked to another point in the file since its last
 632                          * write, or we have reached our maximum cluster size,
 633                          * then push the previous cluster. Otherwise try
 634                          * reallocating to make it sequential.
 635                          *
 636                          * Change to algorithm: only push previous cluster if
 637                          * it was sequential from the point of view of the
 638                          * seqcount heuristic, otherwise leave the buffer
 639                          * intact so we can potentially optimize the I/O
 640                          * later on in the buf_daemon or update daemon
 641                          * flush.
 642                          */
 643                         cursize = vp->v_lastw - vp->v_cstart + 1;
 644                         if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
 645                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
 646                                 if (!async && seqcount > 0) {
 647                                         cluster_wbuild_wb(vp, lblocksize,
 648                                                 vp->v_cstart, cursize);
 649                                 }
 650                         } else {
 651                                 struct buf **bpp, **endbp;
 652                                 struct cluster_save *buflist;
 653
 654                                 buflist = cluster_collectbufs(vp, bp);
 655                                 endbp = &buflist->bs_children
 656                                     [buflist->bs_nchildren - 1];
 657                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 658                                         /*
 659                                          * Failed, push the previous cluster
 660                                          * if *really* writing sequentially
 661                                          * in the logical file (seqcount > 1),
 662                                          * otherwise delay it in the hopes that
 663                                          * the low level disk driver can
 664                                          * optimize the write ordering.
 665                                          */
 666                                         for (bpp = buflist->bs_children;
 667                                              bpp < endbp; bpp++)
 668                                                 brelse(*bpp);
 669                                         free(buflist, M_SEGMENT);
 670                                         if (seqcount > 1) {
 671                                                 cluster_wbuild_wb(vp,
 672                                                     lblocksize, vp->v_cstart,
 673                                                     cursize);
 674                                         }
 675                                 } else {
 676                                         /*
 677                                          * Succeeded, keep building cluster.
 678                                          */
 679                                         for (bpp = buflist->bs_children;
 680                                              bpp <= endbp; bpp++)
 681                                                 bdwrite(*bpp);
 682                                         free(buflist, M_SEGMENT);
 683                                         vp->v_lastw = lbn;
 684                                         vp->v_lasta = bp->b_blkno;
 685                                         return;
 686                                 }
 687                         }
 688                 }
 689                 /*
 690                  * Consider beginning a cluster. If at end of file, make
 691                  * cluster as large as possible, otherwise find size of
 692                  * existing cluster.
 693                  */
 694                 if ((vp->v_type == VREG) &&
 695                         ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
 696                     (bp->b_blkno == bp->b_lblkno) &&
 697                     (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
 698                      bp->b_blkno == -1)) {
 699                         bawrite(bp);
 700                         vp->v_clen = 0;
 701                         vp->v_lasta = bp->b_blkno;
 702                         vp->v_cstart = lbn + 1;
 703                         vp->v_lastw = lbn;
 704                         return;
 705                 }
 706                 vp->v_clen = maxclen;
 707                 if (!async && maxclen == 0) {   /* I/O not contiguous */
 708                         vp->v_cstart = lbn + 1;
 709                         bawrite(bp);
 710                 } else {        /* Wait for rest of cluster */
 711                         vp->v_cstart = lbn;
 712                         bdwrite(bp);
 713                 }
 714         } else if (lbn == vp->v_cstart + vp->v_clen) {
 715                 /*
 716                  * At end of cluster, write it out if seqcount tells us we
 717                  * are operating sequentially, otherwise let the buf or
 718                  * update daemon handle it.
 719                  */
 720                 bdwrite(bp);
 721                 if (seqcount > 1)
 722                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
 723                 vp->v_clen = 0;
 724                 vp->v_cstart = lbn + 1;
 725         } else if (vm_page_count_severe()) {
 726                 /*
 727                  * We are low on memory, get it going NOW
 728                  */
 729                 bawrite(bp);
 730         } else {
 731                 /*
 732                  * In the middle of a cluster, so just delay the I/O for now.
 733                  */
 734                 bdwrite(bp);
 735         }
 736         vp->v_lastw = lbn;
 737         vp->v_lasta = bp->b_blkno;
 738 }
 739
 740
 741 /*
 742  * This is an awful lot like cluster_rbuild...wish they could be combined.
 743  * The last lbn argument is the current block on which I/O is being
 744  * performed.  Check to see that it doesn't fall in the middle of
 745  * the current block (if last_bp == NULL).
 746  */
 747 int
 748 cluster_wbuild(vp, size, start_lbn, len)
 749         struct vnode *vp;
 750         long size;
 751         daddr_t start_lbn;
 752         int len;
 753 {
 754         struct buf *bp, *tbp;
 755         int i, j, s;
 756         int totalwritten = 0;
 757         int dbsize = btodb(size);
 758
 759         while (len > 0) {
 760                 s = splbio();
 761                 /*
 762                  * If the buffer is not delayed-write (i.e. dirty), or it
 763                  * is delayed-write but either locked or inval, it cannot
 764                  * partake in the clustered write.
 765                  */
 766                 if (((tbp = gbincore(vp, start_lbn)) == NULL) ||
 767                   ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
 768                   BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 769                         ++start_lbn;
 770                         --len;
 771                         splx(s);
 772                         continue;
 773                 }
 774                 bremfree(tbp);
 775                 tbp->b_flags &= ~B_DONE;
 776                 splx(s);
 777
 778                 /*
 779                  * Extra memory in the buffer, punt on this buffer.
 780                  * XXX we could handle this in most cases, but we would
 781                  * have to push the extra memory down to after our max
 782                  * possible cluster size and then potentially pull it back
 783                  * up if the cluster was terminated prematurely--too much
 784                  * hassle.
 785                  */
 786                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 787                   (tbp->b_bcount != tbp->b_bufsize) ||
 788                   (tbp->b_bcount != size) ||
 789                   (len == 1) ||
 790                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 791                         totalwritten += tbp->b_bufsize;
 792                         bawrite(tbp);
 793                         ++start_lbn;
 794                         --len;
 795                         continue;
 796                 }
 797
 798                 /*
 799                  * We got a pbuf to make the cluster in.
 800                  * so initialise it.
 801                  */
 802                 TAILQ_INIT(&bp->b_cluster.cluster_head);
 803                 bp->b_bcount = 0;
 804                 bp->b_bufsize = 0;
 805                 bp->b_npages = 0;
 806                 if (tbp->b_wcred != NOCRED) {
 807                     bp->b_wcred = tbp->b_wcred;
 808                     crhold(bp->b_wcred);
 809                 }
 810
 811                 bp->b_blkno = tbp->b_blkno;
 812                 bp->b_lblkno = tbp->b_lblkno;
 813                 bp->b_offset = tbp->b_offset;
 814
 815                 /*
 816                  * We are synthesizing a buffer out of vm_page_t's, but
 817                  * if the block size is not page aligned then the starting
 818                  * address may not be either.  Inherit the b_data offset
 819                  * from the original buffer.
 820                  */
 821                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 822                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 823                 bp->b_flags |= B_CALL | B_CLUSTER |
 824                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN));
 825                 bp->b_iodone = cluster_callback;
 826                 pbgetvp(vp, bp);
 827                 /*
 828                  * From this location in the file, scan forward to see
 829                  * if there are buffers with adjacent data that need to
 830                  * be written as well.
 831                  */
 832                 for (i = 0; i < len; ++i, ++start_lbn) {
 833                         if (i != 0) { /* If not the first buffer */
 834                                 s = splbio();
 835                                 /*
 836                                  * If the adjacent data is not even in core it
 837                                  * can't need to be written.
 838                                  */
 839                                 if ((tbp = gbincore(vp, start_lbn)) == NULL) {
 840                                         splx(s);
 841                                         break;
 842                                 }
 843
 844                                 /*
 845                                  * If it IS in core, but has different
 846                                  * characteristics, or is locked (which
 847                                  * means it could be undergoing a background
 848                                  * I/O or be in a weird state), then don't
 849                                  * cluster with it.
 850                                  */
 851                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 852                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 853                                   != (B_DELWRI | B_CLUSTEROK |
 854                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 855                                     (tbp->b_flags & B_LOCKED) ||
 856                                     tbp->b_wcred != bp->b_wcred ||
 857                                     BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 858                                         splx(s);
 859                                         break;
 860                                 }
 861
 862                                 /*
 863                                  * Check that the combined cluster
 864                                  * would make sense with regard to pages
 865                                  * and would not be too large
 866                                  */
 867                                 if ((tbp->b_bcount != size) ||
 868                                   ((bp->b_blkno + (dbsize * i)) !=
 869                                     tbp->b_blkno) ||
 870                                   ((tbp->b_npages + bp->b_npages) >
 871                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 872                                         BUF_UNLOCK(tbp);
 873                                         splx(s);
 874                                         break;
 875                                 }
 876                                 /*
 877                                  * Ok, it's passed all the tests,
 878                                  * so remove it from the free list
 879                                  * and mark it busy. We will use it.
 880                                  */
 881                                 bremfree(tbp);
 882                                 tbp->b_flags &= ~B_DONE;
 883                                 splx(s);
 884                         } /* end of code for non-first buffers only */
 885                         /* check for latent dependencies to be handled */
 886                         if ((LIST_FIRST(&tbp->b_dep)) != NULL &&
 887                             bioops.io_start)
 888                                 (*bioops.io_start)(tbp);
 889                         /*
 890                          * If the IO is via the VM then we do some
 891                          * special VM hackery (yuck).  Since the buffer's
 892                          * block size may not be page-aligned it is possible
 893                          * for a page to be shared between two buffers.  We
 894                          * have to get rid of the duplication when building
 895                          * the cluster.
 896                          */
 897                         if (tbp->b_flags & B_VMIO) {
 898                                 vm_page_t m;
 899
 900                                 if (i != 0) { /* if not first buffer */
 901                                         for (j = 0; j < tbp->b_npages; j += 1) {
 902                                                 m = tbp->b_pages[j];
 903                                                 if (m->flags & PG_BUSY) {
 904                                                         bqrelse(tbp);
 905                                                         goto finishcluster;
 906                                                 }
 907                                         }
 908                                 }
 909
 910                                 for (j = 0; j < tbp->b_npages; j += 1) {
 911                                         m = tbp->b_pages[j];
 912                                         vm_page_io_start(m);
 913                                         vm_object_pip_add(m->object, 1);
 914                                         if ((bp->b_npages == 0) ||
 915                                           (bp->b_pages[bp->b_npages - 1] != m)) {
 916                                                 bp->b_pages[bp->b_npages] = m;
 917                                                 bp->b_npages++;
 918                                         }
 919                                 }
 920                         }
 921                         bp->b_bcount += size;
 922                         bp->b_bufsize += size;
 923
 924                         s = splbio();
 925                         bundirty(tbp);
 926                         tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
 927                         tbp->b_flags |= B_ASYNC;
 928                         reassignbuf(tbp, tbp->b_vp);    /* put on clean list */
 929                         ++tbp->b_vp->v_numoutput;
 930                         splx(s);
 931                         BUF_KERNPROC(tbp);
 932                         TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 933                                 tbp, b_cluster.cluster_entry);
 934                 }
 935         finishcluster:
 936                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 937                         (vm_page_t *) bp->b_pages, bp->b_npages);
 938                 if (bp->b_bufsize > bp->b_kvasize)
 939                         panic(
 940                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 941                             bp->b_bufsize, bp->b_kvasize);
 942                 bp->b_kvasize = bp->b_bufsize;
 943                 totalwritten += bp->b_bufsize;
 944                 bp->b_dirtyoff = 0;
 945                 bp->b_dirtyend = bp->b_bufsize;
 946                 bawrite(bp);
 947
 948                 len -= i;
 949         }
 950         return totalwritten;
 951 }
 952
 953 /*
 954  * Collect together all the buffers in a cluster.
 955  * Plus add one additional buffer.
 956  */
 957 static struct cluster_save *
 958 cluster_collectbufs(vp, last_bp)
 959         struct vnode *vp;
 960         struct buf *last_bp;
 961 {
 962         struct cluster_save *buflist;
 963         struct buf *bp;
 964         daddr_t lbn;
 965         int i, len;
 966
 967         len = vp->v_lastw - vp->v_cstart + 1;
 968         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
 969             M_SEGMENT, M_WAITOK);
 970         buflist->bs_nchildren = 0;
 971         buflist->bs_children = (struct buf **) (buflist + 1);
 972         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
 973                 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
 974                 buflist->bs_children[i] = bp;
 975                 if (bp->b_blkno == bp->b_lblkno)
 976                         VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
 977                                 NULL, NULL);
 978         }
 979         buflist->bs_children[i] = bp = last_bp;
 980         if (bp->b_blkno == bp->b_lblkno)
 981                 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
 982                         NULL, NULL);
 983         buflist->bs_nchildren = i + 1;
 984         return (buflist);
 985 }