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 1.8 2003/08/26 21:09:02 rob Exp $
  38  */
  39
  40 #include "opt_debug_cluster.h"
  41
  42 #include <sys/param.h>
  43 #include <sys/systm.h>
  44 #include <sys/kernel.h>
  45 #include <sys/proc.h>
  46 #include <sys/buf.h>
  47 #include <sys/vnode.h>
  48 #include <sys/malloc.h>
  49 #include <sys/mount.h>
  50 #include <sys/resourcevar.h>
  51 #include <sys/vmmeter.h>
  52 #include <vm/vm.h>
  53 #include <vm/vm_object.h>
  54 #include <vm/vm_page.h>
  55 #include <sys/sysctl.h>
  56 #include <sys/buf2.h>
  57 #include <vm/vm_page2.h>
  58
  59 #if defined(CLUSTERDEBUG)
  60 #include <sys/sysctl.h>
  61 static int      rcluster= 0;
  62 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
  63 #endif
  64
  65 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer");
  66
  67 static struct cluster_save *
  68         cluster_collectbufs (struct vnode *vp, struct buf *last_bp);
  69 static struct buf *
  70         cluster_rbuild (struct vnode *vp, u_quad_t filesize, daddr_t lbn,
  71                             daddr_t blkno, long size, int run, struct buf *fbp);
  72
  73 static int write_behind = 1;
  74 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
  75
  76 extern vm_page_t        bogus_page;
  77
  78 extern int cluster_pbuf_freecnt;
  79
  80 /*
  81  * Maximum number of blocks for read-ahead.
  82  */
  83 #define MAXRA 32
  84
  85 /*
  86  * This replaces bread.
  87  */
  88 int
  89 cluster_read(vp, filesize, lblkno, size, totread, seqcount, bpp)
  90         struct vnode *vp;
  91         u_quad_t filesize;
  92         daddr_t lblkno;
  93         long size;
  94         long totread;
  95         int seqcount;
  96         struct buf **bpp;
  97 {
  98         struct buf *bp, *rbp, *reqbp;
  99         daddr_t blkno, origblkno;
 100         int error, num_ra;
 101         int i;
 102         int maxra, racluster;
 103         long origtotread;
 104
 105         error = 0;
 106
 107         /*
 108          * Try to limit the amount of read-ahead by a few
 109          * ad-hoc parameters.  This needs work!!!
 110          */
 111         racluster = vp->v_mount->mnt_iosize_max / size;
 112         maxra = 2 * racluster + (totread / size);
 113         if (maxra > MAXRA)
 114                 maxra = MAXRA;
 115         if (maxra > nbuf/8)
 116                 maxra = nbuf/8;
 117
 118         /*
 119          * get the requested block
 120          */
 121         *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0);
 122         origblkno = lblkno;
 123         origtotread = totread;
 124
 125         /*
 126          * if it is in the cache, then check to see if the reads have been
 127          * sequential.  If they have, then try some read-ahead, otherwise
 128          * back-off on prospective read-aheads.
 129          */
 130         if (bp->b_flags & B_CACHE) {
 131                 if (!seqcount) {
 132                         return 0;
 133                 } else if ((bp->b_flags & B_RAM) == 0) {
 134                         return 0;
 135                 } else {
 136                         int s;
 137                         struct buf *tbp;
 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                         for (i = 1; i < maxra; i++) {
 147
 148                                 if (!(tbp = incore(vp, lblkno+i))) {
 149                                         break;
 150                                 }
 151
 152                                 /*
 153                                  * Set another read-ahead mark so we know
 154                                  * to check again.
 155                                  */
 156                                 if (((i % racluster) == (racluster - 1)) ||
 157                                         (i == (maxra - 1)))
 158                                         tbp->b_flags |= B_RAM;
 159                         }
 160                         splx(s);
 161                         if (i >= maxra) {
 162                                 return 0;
 163                         }
 164                         lblkno += i;
 165                 }
 166                 reqbp = bp = NULL;
 167         } else {
 168                 off_t firstread = bp->b_offset;
 169
 170                 KASSERT(bp->b_offset != NOOFFSET,
 171                     ("cluster_read: no buffer offset"));
 172                 if (firstread + totread > filesize)
 173                         totread = filesize - firstread;
 174                 if (totread > size) {
 175                         int nblks = 0;
 176                         int ncontigafter;
 177                         while (totread > 0) {
 178                                 nblks++;
 179                                 totread -= size;
 180                         }
 181                         if (nblks == 1)
 182                                 goto single_block_read;
 183                         if (nblks > racluster)
 184                                 nblks = racluster;
 185
 186                         error = VOP_BMAP(vp, lblkno, NULL,
 187                                 &blkno, &ncontigafter, NULL);
 188                         if (error)
 189                                 goto single_block_read;
 190                         if (blkno == -1)
 191                                 goto single_block_read;
 192                         if (ncontigafter == 0)
 193                                 goto single_block_read;
 194                         if (ncontigafter + 1 < nblks)
 195                                 nblks = ncontigafter + 1;
 196
 197                         bp = cluster_rbuild(vp, filesize, lblkno,
 198                                 blkno, size, nblks, bp);
 199                         lblkno += (bp->b_bufsize / size);
 200                 } else {
 201 single_block_read:
 202                         /*
 203                          * if it isn't in the cache, then get a chunk from
 204                          * disk if sequential, otherwise just get the block.
 205                          */
 206                         bp->b_flags |= B_READ | B_RAM;
 207                         lblkno += 1;
 208                 }
 209         }
 210
 211         /*
 212          * if we have been doing sequential I/O, then do some read-ahead
 213          */
 214         rbp = NULL;
 215         if (seqcount && (lblkno < (origblkno + seqcount))) {
 216                 /*
 217                  * we now build the read-ahead buffer if it is desirable.
 218                  */
 219                 if (((u_quad_t)(lblkno + 1) * size) <= filesize &&
 220                     !(error = VOP_BMAP(vp, lblkno, NULL, &blkno, &num_ra, NULL)) &&
 221                     blkno != -1) {
 222                         int nblksread;
 223                         int ntoread = num_ra + 1;
 224                         nblksread = (origtotread + size - 1) / size;
 225                         if (seqcount < nblksread)
 226                                 seqcount = nblksread;
 227                         if (seqcount < ntoread)
 228                                 ntoread = seqcount;
 229                         if (num_ra) {
 230                                 rbp = cluster_rbuild(vp, filesize, lblkno,
 231                                         blkno, size, ntoread, NULL);
 232                         } else {
 233                                 rbp = getblk(vp, lblkno, size, 0, 0);
 234                                 rbp->b_flags |= B_READ | B_ASYNC | B_RAM;
 235                                 rbp->b_blkno = blkno;
 236                         }
 237                 }
 238         }
 239
 240         /*
 241          * handle the synchronous read
 242          */
 243         if (bp) {
 244 #if defined(CLUSTERDEBUG)
 245                 if (rcluster)
 246                         printf("S(%ld,%ld,%d) ",
 247                             (long)bp->b_lblkno, bp->b_bcount, seqcount);
 248 #endif
 249                 if ((bp->b_flags & B_CLUSTER) == 0) {
 250                         vfs_busy_pages(bp, 0);
 251                 }
 252                 bp->b_flags &= ~(B_ERROR|B_INVAL);
 253                 if (bp->b_flags & (B_ASYNC|B_CALL))
 254                         BUF_KERNPROC(bp);
 255                 error = VOP_STRATEGY(vp, bp);
 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                 }
 292         }
 293         if (reqbp)
 294                 return (biowait(reqbp));
 295         else
 296                 return (error);
 297 }
 298
 299 /*
 300  * If blocks are contiguous on disk, use this to provide clustered
 301  * read ahead.  We will read as many blocks as possible sequentially
 302  * and then parcel them up into logical blocks in the buffer hash table.
 303  */
 304 static struct buf *
 305 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
 306         struct vnode *vp;
 307         u_quad_t filesize;
 308         daddr_t lbn;
 309         daddr_t blkno;
 310         long size;
 311         int run;
 312         struct buf *fbp;
 313 {
 314         struct buf *bp, *tbp;
 315         daddr_t bn;
 316         int i, inc, j;
 317
 318         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 319             ("cluster_rbuild: size %ld != filesize %ld\n",
 320             size, vp->v_mount->mnt_stat.f_iosize));
 321
 322         /*
 323          * avoid a division
 324          */
 325         while ((u_quad_t) size * (lbn + run) > filesize) {
 326                 --run;
 327         }
 328
 329         if (fbp) {
 330                 tbp = fbp;
 331                 tbp->b_flags |= B_READ;
 332         } else {
 333                 tbp = getblk(vp, lbn, size, 0, 0);
 334                 if (tbp->b_flags & B_CACHE)
 335                         return tbp;
 336                 tbp->b_flags |= B_ASYNC | B_READ | B_RAM;
 337         }
 338
 339         tbp->b_blkno = blkno;
 340         if( (tbp->b_flags & B_MALLOC) ||
 341                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
 342                 return tbp;
 343
 344         bp = trypbuf(&cluster_pbuf_freecnt);
 345         if (bp == 0)
 346                 return tbp;
 347
 348         /*
 349          * We are synthesizing a buffer out of vm_page_t's, but
 350          * if the block size is not page aligned then the starting
 351          * address may not be either.  Inherit the b_data offset
 352          * from the original buffer.
 353          */
 354         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 355             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 356         bp->b_flags = B_ASYNC | B_READ | B_CALL | B_CLUSTER | B_VMIO;
 357         bp->b_iodone = cluster_callback;
 358         bp->b_blkno = blkno;
 359         bp->b_lblkno = lbn;
 360         bp->b_offset = tbp->b_offset;
 361         KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
 362         pbgetvp(vp, bp);
 363
 364         TAILQ_INIT(&bp->b_cluster.cluster_head);
 365
 366         bp->b_bcount = 0;
 367         bp->b_bufsize = 0;
 368         bp->b_npages = 0;
 369
 370         inc = btodb(size);
 371         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
 372                 if (i != 0) {
 373                         if ((bp->b_npages * PAGE_SIZE) +
 374                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 375                                 break;
 376                         }
 377
 378                         /*
 379                          * Shortcut some checks and try to avoid buffers that
 380                          * would block in the lock.  The same checks have to
 381                          * be made again after we officially get the buffer.
 382                          */
 383                         if ((tbp = incore(vp, lbn + i)) != NULL) {
 384                                 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
 385                                         break;
 386                                 BUF_UNLOCK(tbp);
 387
 388                                 for (j = 0; j < tbp->b_npages; j++) {
 389                                         if (tbp->b_pages[j]->valid)
 390                                                 break;
 391                                 }
 392
 393                                 if (j != tbp->b_npages)
 394                                         break;
 395
 396                                 if (tbp->b_bcount != size)
 397                                         break;
 398                         }
 399
 400                         tbp = getblk(vp, lbn + i, size, 0, 0);
 401
 402                         /*
 403                          * Stop scanning if the buffer is fuly valid
 404                          * (marked B_CACHE), or locked (may be doing a
 405                          * background write), or if the buffer is not
 406                          * VMIO backed.  The clustering code can only deal
 407                          * with VMIO-backed buffers.
 408                          */
 409                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 410                             (tbp->b_flags & B_VMIO) == 0) {
 411                                 bqrelse(tbp);
 412                                 break;
 413                         }
 414
 415                         /*
 416                          * The buffer must be completely invalid in order to
 417                          * take part in the cluster.  If it is partially valid
 418                          * then we stop.
 419                          */
 420                         for (j = 0;j < tbp->b_npages; j++) {
 421                                 if (tbp->b_pages[j]->valid)
 422                                         break;
 423                         }
 424                         if (j != tbp->b_npages) {
 425                                 bqrelse(tbp);
 426                                 break;
 427                         }
 428
 429                         /*
 430                          * Set a read-ahead mark as appropriate
 431                          */
 432                         if ((fbp && (i == 1)) || (i == (run - 1)))
 433                                 tbp->b_flags |= B_RAM;
 434
 435                         /*
 436                          * Set the buffer up for an async read (XXX should
 437                          * we do this only if we do not wind up brelse()ing?).
 438                          * Set the block number if it isn't set, otherwise
 439                          * if it is make sure it matches the block number we
 440                          * expect.
 441                          */
 442                         tbp->b_flags |= B_READ | B_ASYNC;
 443                         if (tbp->b_blkno == tbp->b_lblkno) {
 444                                 tbp->b_blkno = bn;
 445                         } else if (tbp->b_blkno != bn) {
 446                                 brelse(tbp);
 447                                 break;
 448                         }
 449                 }
 450                 /*
 451                  * XXX fbp from caller may not be B_ASYNC, but we are going
 452                  * to biodone() it in cluster_callback() anyway
 453                  */
 454                 BUF_KERNPROC(tbp);
 455                 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 456                         tbp, b_cluster.cluster_entry);
 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                 /*
 471                  * XXX shouldn't this be += size for both, like in
 472                  * cluster_wbuild()?
 473                  *
 474                  * Don't inherit tbp->b_bufsize as it may be larger due to
 475                  * a non-page-aligned size.  Instead just aggregate using
 476                  * 'size'.
 477                  */
 478                 if (tbp->b_bcount != size)
 479                     printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
 480                 if (tbp->b_bufsize != size)
 481                     printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
 482                 bp->b_bcount += size;
 483                 bp->b_bufsize += size;
 484         }
 485
 486         /*
 487          * Fully valid pages in the cluster are already good and do not need
 488          * to be re-read from disk.  Replace the page with bogus_page
 489          */
 490         for (j = 0; j < bp->b_npages; j++) {
 491                 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 492                     VM_PAGE_BITS_ALL) {
 493                         bp->b_pages[j] = bogus_page;
 494                 }
 495         }
 496         if (bp->b_bufsize > bp->b_kvasize)
 497                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 498                     bp->b_bufsize, bp->b_kvasize);
 499         bp->b_kvasize = bp->b_bufsize;
 500
 501         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 502                 (vm_page_t *)bp->b_pages, bp->b_npages);
 503         return (bp);
 504 }
 505
 506 /*
 507  * Cleanup after a clustered read or write.
 508  * This is complicated by the fact that any of the buffers might have
 509  * extra memory (if there were no empty buffer headers at allocbuf time)
 510  * that we will need to shift around.
 511  */
 512 void
 513 cluster_callback(bp)
 514         struct buf *bp;
 515 {
 516         struct buf *nbp, *tbp;
 517         int error = 0;
 518
 519         /*
 520          * Must propogate errors to all the components.
 521          */
 522         if (bp->b_flags & B_ERROR)
 523                 error = bp->b_error;
 524
 525         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
 526         /*
 527          * Move memory from the large cluster buffer into the component
 528          * buffers and mark IO as done on these.
 529          */
 530         for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
 531                 tbp; tbp = nbp) {
 532                 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
 533                 if (error) {
 534                         tbp->b_flags |= B_ERROR;
 535                         tbp->b_error = error;
 536                 } else {
 537                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 538                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 539                         /*
 540                          * XXX the bdwrite()/bqrelse() issued during
 541                          * cluster building clears B_RELBUF (see bqrelse()
 542                          * comment).  If direct I/O was specified, we have
 543                          * to restore it here to allow the buffer and VM
 544                          * to be freed.
 545                          */
 546                         if (tbp->b_flags & B_DIRECT)
 547                                 tbp->b_flags |= B_RELBUF;
 548                 }
 549                 biodone(tbp);
 550         }
 551         relpbuf(bp, &cluster_pbuf_freecnt);
 552 }
 553
 554 /*
 555  *      cluster_wbuild_wb:
 556  *
 557  *      Implement modified write build for cluster.
 558  *
 559  *              write_behind = 0        write behind disabled
 560  *              write_behind = 1        write behind normal (default)
 561  *              write_behind = 2        write behind backed-off
 562  */
 563
 564 static __inline int
 565 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
 566 {
 567         int r = 0;
 568
 569         switch(write_behind) {
 570         case 2:
 571                 if (start_lbn < len)
 572                         break;
 573                 start_lbn -= len;
 574                 /* fall through */
 575         case 1:
 576                 r = cluster_wbuild(vp, size, start_lbn, len);
 577                 /* fall through */
 578         default:
 579                 /* fall through */
 580                 break;
 581         }
 582         return(r);
 583 }
 584
 585 /*
 586  * Do clustered write for FFS.
 587  *
 588  * Three cases:
 589  *      1. Write is not sequential (write asynchronously)
 590  *      Write is sequential:
 591  *      2.      beginning of cluster - begin cluster
 592  *      3.      middle of a cluster - add to cluster
 593  *      4.      end of a cluster - asynchronously write cluster
 594  */
 595 void
 596 cluster_write(bp, filesize, seqcount)
 597         struct buf *bp;
 598         u_quad_t filesize;
 599         int seqcount;
 600 {
 601         struct vnode *vp;
 602         daddr_t lbn;
 603         int maxclen, cursize;
 604         int lblocksize;
 605         int async;
 606
 607         vp = bp->b_vp;
 608         if (vp->v_type == VREG) {
 609                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 610                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 611         } else {
 612                 async = 0;
 613                 lblocksize = bp->b_bufsize;
 614         }
 615         lbn = bp->b_lblkno;
 616         KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
 617
 618         /* Initialize vnode to beginning of file. */
 619         if (lbn == 0)
 620                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 621
 622         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
 623             (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
 624                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
 625                 if (vp->v_clen != 0) {
 626                         /*
 627                          * Next block is not sequential.
 628                          *
 629                          * If we are not writing at end of file, the process
 630                          * seeked to another point in the file since its last
 631                          * write, or we have reached our maximum cluster size,
 632                          * then push the previous cluster. Otherwise try
 633                          * reallocating to make it sequential.
 634                          *
 635                          * Change to algorithm: only push previous cluster if
 636                          * it was sequential from the point of view of the
 637                          * seqcount heuristic, otherwise leave the buffer
 638                          * intact so we can potentially optimize the I/O
 639                          * later on in the buf_daemon or update daemon
 640                          * flush.
 641                          */
 642                         cursize = vp->v_lastw - vp->v_cstart + 1;
 643                         if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
 644                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
 645                                 if (!async && seqcount > 0) {
 646                                         cluster_wbuild_wb(vp, lblocksize,
 647                                                 vp->v_cstart, cursize);
 648                                 }
 649                         } else {
 650                                 struct buf **bpp, **endbp;
 651                                 struct cluster_save *buflist;
 652
 653                                 buflist = cluster_collectbufs(vp, bp);
 654                                 endbp = &buflist->bs_children
 655                                     [buflist->bs_nchildren - 1];
 656                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 657                                         /*
 658                                          * Failed, push the previous cluster
 659                                          * if *really* writing sequentially
 660                                          * in the logical file (seqcount > 1),
 661                                          * otherwise delay it in the hopes that
 662                                          * the low level disk driver can
 663                                          * optimize the write ordering.
 664                                          */
 665                                         for (bpp = buflist->bs_children;
 666                                              bpp < endbp; bpp++)
 667                                                 brelse(*bpp);
 668                                         free(buflist, M_SEGMENT);
 669                                         if (seqcount > 1) {
 670                                                 cluster_wbuild_wb(vp,
 671                                                     lblocksize, vp->v_cstart,
 672                                                     cursize);
 673                                         }
 674                                 } else {
 675                                         /*
 676                                          * Succeeded, keep building cluster.
 677                                          */
 678                                         for (bpp = buflist->bs_children;
 679                                              bpp <= endbp; bpp++)
 680                                                 bdwrite(*bpp);
 681                                         free(buflist, M_SEGMENT);
 682                                         vp->v_lastw = lbn;
 683                                         vp->v_lasta = bp->b_blkno;
 684                                         return;
 685                                 }
 686                         }
 687                 }
 688                 /*
 689                  * Consider beginning a cluster. If at end of file, make
 690                  * cluster as large as possible, otherwise find size of
 691                  * existing cluster.
 692                  */
 693                 if ((vp->v_type == VREG) &&
 694                         ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
 695                     (bp->b_blkno == bp->b_lblkno) &&
 696                     (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
 697                      bp->b_blkno == -1)) {
 698                         bawrite(bp);
 699                         vp->v_clen = 0;
 700                         vp->v_lasta = bp->b_blkno;
 701                         vp->v_cstart = lbn + 1;
 702                         vp->v_lastw = lbn;
 703                         return;
 704                 }
 705                 vp->v_clen = maxclen;
 706                 if (!async && maxclen == 0) {   /* I/O not contiguous */
 707                         vp->v_cstart = lbn + 1;
 708                         bawrite(bp);
 709                 } else {        /* Wait for rest of cluster */
 710                         vp->v_cstart = lbn;
 711                         bdwrite(bp);
 712                 }
 713         } else if (lbn == vp->v_cstart + vp->v_clen) {
 714                 /*
 715                  * At end of cluster, write it out if seqcount tells us we
 716                  * are operating sequentially, otherwise let the buf or
 717                  * update daemon handle it.
 718                  */
 719                 bdwrite(bp);
 720                 if (seqcount > 1)
 721                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
 722                 vp->v_clen = 0;
 723                 vp->v_cstart = lbn + 1;
 724         } else if (vm_page_count_severe()) {
 725                 /*
 726                  * We are low on memory, get it going NOW
 727                  */
 728                 bawrite(bp);
 729         } else {
 730                 /*
 731                  * In the middle of a cluster, so just delay the I/O for now.
 732                  */
 733                 bdwrite(bp);
 734         }
 735         vp->v_lastw = lbn;
 736         vp->v_lasta = bp->b_blkno;
 737 }
 738
 739
 740 /*
 741  * This is an awful lot like cluster_rbuild...wish they could be combined.
 742  * The last lbn argument is the current block on which I/O is being
 743  * performed.  Check to see that it doesn't fall in the middle of
 744  * the current block (if last_bp == NULL).
 745  */
 746 int
 747 cluster_wbuild(vp, size, start_lbn, len)
 748         struct vnode *vp;
 749         long size;
 750         daddr_t start_lbn;
 751         int len;
 752 {
 753         struct buf *bp, *tbp;
 754         int i, j, s;
 755         int totalwritten = 0;
 756         int dbsize = btodb(size);
 757
 758         while (len > 0) {
 759                 s = splbio();
 760                 /*
 761                  * If the buffer is not delayed-write (i.e. dirty), or it
 762                  * is delayed-write but either locked or inval, it cannot
 763                  * partake in the clustered write.
 764                  */
 765                 if (((tbp = gbincore(vp, start_lbn)) == NULL) ||
 766                   ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
 767                   BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 768                         ++start_lbn;
 769                         --len;
 770                         splx(s);
 771                         continue;
 772                 }
 773                 bremfree(tbp);
 774                 tbp->b_flags &= ~B_DONE;
 775                 splx(s);
 776
 777                 /*
 778                  * Extra memory in the buffer, punt on this buffer.
 779                  * XXX we could handle this in most cases, but we would
 780                  * have to push the extra memory down to after our max
 781                  * possible cluster size and then potentially pull it back
 782                  * up if the cluster was terminated prematurely--too much
 783                  * hassle.
 784                  */
 785                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 786                   (tbp->b_bcount != tbp->b_bufsize) ||
 787                   (tbp->b_bcount != size) ||
 788                   (len == 1) ||
 789                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 790                         totalwritten += tbp->b_bufsize;
 791                         bawrite(tbp);
 792                         ++start_lbn;
 793                         --len;
 794                         continue;
 795                 }
 796
 797                 /*
 798                  * We got a pbuf to make the cluster in.
 799                  * so initialise it.
 800                  */
 801                 TAILQ_INIT(&bp->b_cluster.cluster_head);
 802                 bp->b_bcount = 0;
 803                 bp->b_bufsize = 0;
 804                 bp->b_npages = 0;
 805                 bp->b_blkno = tbp->b_blkno;
 806                 bp->b_lblkno = tbp->b_lblkno;
 807                 bp->b_offset = tbp->b_offset;
 808
 809                 /*
 810                  * We are synthesizing a buffer out of vm_page_t's, but
 811                  * if the block size is not page aligned then the starting
 812                  * address may not be either.  Inherit the b_data offset
 813                  * from the original buffer.
 814                  */
 815                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 816                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 817                 bp->b_flags |= B_CALL | B_CLUSTER |
 818                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN));
 819                 bp->b_iodone = cluster_callback;
 820                 pbgetvp(vp, bp);
 821                 /*
 822                  * From this location in the file, scan forward to see
 823                  * if there are buffers with adjacent data that need to
 824                  * be written as well.
 825                  */
 826                 for (i = 0; i < len; ++i, ++start_lbn) {
 827                         if (i != 0) { /* If not the first buffer */
 828                                 s = splbio();
 829                                 /*
 830                                  * If the adjacent data is not even in core it
 831                                  * can't need to be written.
 832                                  */
 833                                 if ((tbp = gbincore(vp, start_lbn)) == NULL) {
 834                                         splx(s);
 835                                         break;
 836                                 }
 837
 838                                 /*
 839                                  * If it IS in core, but has different
 840                                  * characteristics, or is locked (which
 841                                  * means it could be undergoing a background
 842                                  * I/O or be in a weird state), then don't
 843                                  * cluster with it.
 844                                  */
 845                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 846                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 847                                   != (B_DELWRI | B_CLUSTEROK |
 848                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 849                                     (tbp->b_flags & B_LOCKED) ||
 850                                     BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 851                                         splx(s);
 852                                         break;
 853                                 }
 854
 855                                 /*
 856                                  * Check that the combined cluster
 857                                  * would make sense with regard to pages
 858                                  * and would not be too large
 859                                  */
 860                                 if ((tbp->b_bcount != size) ||
 861                                   ((bp->b_blkno + (dbsize * i)) !=
 862                                     tbp->b_blkno) ||
 863                                   ((tbp->b_npages + bp->b_npages) >
 864                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 865                                         BUF_UNLOCK(tbp);
 866                                         splx(s);
 867                                         break;
 868                                 }
 869                                 /*
 870                                  * Ok, it's passed all the tests,
 871                                  * so remove it from the free list
 872                                  * and mark it busy. We will use it.
 873                                  */
 874                                 bremfree(tbp);
 875                                 tbp->b_flags &= ~B_DONE;
 876                                 splx(s);
 877                         } /* end of code for non-first buffers only */
 878                         /* check for latent dependencies to be handled */
 879                         if ((LIST_FIRST(&tbp->b_dep)) != NULL &&
 880                             bioops.io_start)
 881                                 (*bioops.io_start)(tbp);
 882                         /*
 883                          * If the IO is via the VM then we do some
 884                          * special VM hackery (yuck).  Since the buffer's
 885                          * block size may not be page-aligned it is possible
 886                          * for a page to be shared between two buffers.  We
 887                          * have to get rid of the duplication when building
 888                          * the cluster.
 889                          */
 890                         if (tbp->b_flags & B_VMIO) {
 891                                 vm_page_t m;
 892
 893                                 if (i != 0) { /* if not first buffer */
 894                                         for (j = 0; j < tbp->b_npages; j += 1) {
 895                                                 m = tbp->b_pages[j];
 896                                                 if (m->flags & PG_BUSY) {
 897                                                         bqrelse(tbp);
 898                                                         goto finishcluster;
 899                                                 }
 900                                         }
 901                                 }
 902
 903                                 for (j = 0; j < tbp->b_npages; j += 1) {
 904                                         m = tbp->b_pages[j];
 905                                         vm_page_io_start(m);
 906                                         vm_object_pip_add(m->object, 1);
 907                                         if ((bp->b_npages == 0) ||
 908                                           (bp->b_pages[bp->b_npages - 1] != m)) {
 909                                                 bp->b_pages[bp->b_npages] = m;
 910                                                 bp->b_npages++;
 911                                         }
 912                                 }
 913                         }
 914                         bp->b_bcount += size;
 915                         bp->b_bufsize += size;
 916
 917                         s = splbio();
 918                         bundirty(tbp);
 919                         tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
 920                         tbp->b_flags |= B_ASYNC;
 921                         reassignbuf(tbp, tbp->b_vp);    /* put on clean list */
 922                         ++tbp->b_vp->v_numoutput;
 923                         splx(s);
 924                         BUF_KERNPROC(tbp);
 925                         TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 926                                 tbp, b_cluster.cluster_entry);
 927                 }
 928         finishcluster:
 929                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 930                         (vm_page_t *) bp->b_pages, bp->b_npages);
 931                 if (bp->b_bufsize > bp->b_kvasize)
 932                         panic(
 933                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 934                             bp->b_bufsize, bp->b_kvasize);
 935                 bp->b_kvasize = bp->b_bufsize;
 936                 totalwritten += bp->b_bufsize;
 937                 bp->b_dirtyoff = 0;
 938                 bp->b_dirtyend = bp->b_bufsize;
 939                 bawrite(bp);
 940
 941                 len -= i;
 942         }
 943         return totalwritten;
 944 }
 945
 946 /*
 947  * Collect together all the buffers in a cluster.
 948  * Plus add one additional buffer.
 949  */
 950 static struct cluster_save *
 951 cluster_collectbufs(vp, last_bp)
 952         struct vnode *vp;
 953         struct buf *last_bp;
 954 {
 955         struct cluster_save *buflist;
 956         struct buf *bp;
 957         daddr_t lbn;
 958         int i, len;
 959
 960         len = vp->v_lastw - vp->v_cstart + 1;
 961         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
 962             M_SEGMENT, M_WAITOK);
 963         buflist->bs_nchildren = 0;
 964         buflist->bs_children = (struct buf **) (buflist + 1);
 965         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
 966                 (void) bread(vp, lbn, last_bp->b_bcount, &bp);
 967                 buflist->bs_children[i] = bp;
 968                 if (bp->b_blkno == bp->b_lblkno)
 969                         VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
 970                                 NULL, NULL);
 971         }
 972         buflist->bs_children[i] = bp = last_bp;
 973         if (bp->b_blkno == bp->b_lblkno)
 974                 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
 975                         NULL, NULL);
 976         buflist->bs_nchildren = i + 1;
 977         return (buflist);
 978 }