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 1.3 2003/06/19 01:55:06 dillon Exp $
  38  */
  39
  40 #include "opt_debug_cluster.h"
  41
  42 #include <sys/param.h>
  43 #include <sys/systm.h>
  44 #include <sys/kernel.h>
  45 #include <sys/proc.h>
  46 #include <sys/buf.h>
  47 #include <sys/vnode.h>
  48 #include <sys/malloc.h>
  49 #include <sys/mount.h>
  50 #include <sys/resourcevar.h>
  51 #include <sys/vmmeter.h>
  52 #include <vm/vm.h>
  53 #include <vm/vm_object.h>
  54 #include <vm/vm_page.h>
  55 #include <sys/sysctl.h>
  56 #include <sys/buf2.h>
  57
  58 #if defined(CLUSTERDEBUG)
  59 #include <sys/sysctl.h>
  60 static int      rcluster= 0;
  61 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
  62 #endif
  63
  64 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer");
  65
  66 static struct cluster_save *
  67         cluster_collectbufs __P((struct vnode *vp, struct buf *last_bp));
  68 static struct buf *
  69         cluster_rbuild __P((struct vnode *vp, u_quad_t filesize, daddr_t lbn,
  70                             daddr_t blkno, long size, int run, struct buf *fbp));
  71
  72 static int write_behind = 1;
  73 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
  74
  75 extern vm_page_t        bogus_page;
  76
  77 extern int cluster_pbuf_freecnt;
  78
  79 /*
  80  * Maximum number of blocks for read-ahead.
  81  */
  82 #define MAXRA 32
  83
  84 /*
  85  * This replaces bread.
  86  */
  87 int
  88 cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp)
  89         struct vnode *vp;
  90         u_quad_t filesize;
  91         daddr_t lblkno;
  92         long size;
  93         struct ucred *cred;
  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                 curproc->p_stats->p_ru.ru_inblock++;
 257         }
 258
 259         /*
 260          * and if we have read-aheads, do them too
 261          */
 262         if (rbp) {
 263                 if (error) {
 264                         rbp->b_flags &= ~(B_ASYNC | B_READ);
 265                         brelse(rbp);
 266                 } else if (rbp->b_flags & B_CACHE) {
 267                         rbp->b_flags &= ~(B_ASYNC | B_READ);
 268                         bqrelse(rbp);
 269                 } else {
 270 #if defined(CLUSTERDEBUG)
 271                         if (rcluster) {
 272                                 if (bp)
 273                                         printf("A+(%ld,%ld,%ld,%d) ",
 274                                             (long)rbp->b_lblkno, rbp->b_bcount,
 275                                             (long)(rbp->b_lblkno - origblkno),
 276                                             seqcount);
 277                                 else
 278                                         printf("A(%ld,%ld,%ld,%d) ",
 279                                             (long)rbp->b_lblkno, rbp->b_bcount,
 280                                             (long)(rbp->b_lblkno - origblkno),
 281                                             seqcount);
 282                         }
 283 #endif
 284
 285                         if ((rbp->b_flags & B_CLUSTER) == 0) {
 286                                 vfs_busy_pages(rbp, 0);
 287                         }
 288                         rbp->b_flags &= ~(B_ERROR|B_INVAL);
 289                         if (rbp->b_flags & (B_ASYNC|B_CALL))
 290                                 BUF_KERNPROC(rbp);
 291                         (void) VOP_STRATEGY(vp, rbp);
 292                         curproc->p_stats->p_ru.ru_inblock++;
 293                 }
 294         }
 295         if (reqbp)
 296                 return (biowait(reqbp));
 297         else
 298                 return (error);
 299 }
 300
 301 /*
 302  * If blocks are contiguous on disk, use this to provide clustered
 303  * read ahead.  We will read as many blocks as possible sequentially
 304  * and then parcel them up into logical blocks in the buffer hash table.
 305  */
 306 static struct buf *
 307 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
 308         struct vnode *vp;
 309         u_quad_t filesize;
 310         daddr_t lbn;
 311         daddr_t blkno;
 312         long size;
 313         int run;
 314         struct buf *fbp;
 315 {
 316         struct buf *bp, *tbp;
 317         daddr_t bn;
 318         int i, inc, j;
 319
 320         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 321             ("cluster_rbuild: size %ld != filesize %ld\n",
 322             size, vp->v_mount->mnt_stat.f_iosize));
 323
 324         /*
 325          * avoid a division
 326          */
 327         while ((u_quad_t) size * (lbn + run) > filesize) {
 328                 --run;
 329         }
 330
 331         if (fbp) {
 332                 tbp = fbp;
 333                 tbp->b_flags |= B_READ;
 334         } else {
 335                 tbp = getblk(vp, lbn, size, 0, 0);
 336                 if (tbp->b_flags & B_CACHE)
 337                         return tbp;
 338                 tbp->b_flags |= B_ASYNC | B_READ | B_RAM;
 339         }
 340
 341         tbp->b_blkno = blkno;
 342         if( (tbp->b_flags & B_MALLOC) ||
 343                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
 344                 return tbp;
 345
 346         bp = trypbuf(&cluster_pbuf_freecnt);
 347         if (bp == 0)
 348                 return tbp;
 349
 350         /*
 351          * We are synthesizing a buffer out of vm_page_t's, but
 352          * if the block size is not page aligned then the starting
 353          * address may not be either.  Inherit the b_data offset
 354          * from the original buffer.
 355          */
 356         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 357             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 358         bp->b_flags = B_ASYNC | B_READ | B_CALL | B_CLUSTER | B_VMIO;
 359         bp->b_iodone = cluster_callback;
 360         bp->b_blkno = blkno;
 361         bp->b_lblkno = lbn;
 362         bp->b_offset = tbp->b_offset;
 363         KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
 364         pbgetvp(vp, bp);
 365
 366         TAILQ_INIT(&bp->b_cluster.cluster_head);
 367
 368         bp->b_bcount = 0;
 369         bp->b_bufsize = 0;
 370         bp->b_npages = 0;
 371
 372         inc = btodb(size);
 373         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
 374                 if (i != 0) {
 375                         if ((bp->b_npages * PAGE_SIZE) +
 376                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 377                                 break;
 378                         }
 379
 380                         /*
 381                          * Shortcut some checks and try to avoid buffers that
 382                          * would block in the lock.  The same checks have to
 383                          * be made again after we officially get the buffer.
 384                          */
 385                         if ((tbp = incore(vp, lbn + i)) != NULL) {
 386                                 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
 387                                         break;
 388                                 BUF_UNLOCK(tbp);
 389
 390                                 for (j = 0; j < tbp->b_npages; j++) {
 391                                         if (tbp->b_pages[j]->valid)
 392                                                 break;
 393                                 }
 394
 395                                 if (j != tbp->b_npages)
 396                                         break;
 397
 398                                 if (tbp->b_bcount != size)
 399                                         break;
 400                         }
 401
 402                         tbp = getblk(vp, lbn + i, size, 0, 0);
 403
 404                         /*
 405                          * Stop scanning if the buffer is fuly valid
 406                          * (marked B_CACHE), or locked (may be doing a
 407                          * background write), or if the buffer is not
 408                          * VMIO backed.  The clustering code can only deal
 409                          * with VMIO-backed buffers.
 410                          */
 411                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 412                             (tbp->b_flags & B_VMIO) == 0) {
 413                                 bqrelse(tbp);
 414                                 break;
 415                         }
 416
 417                         /*
 418                          * The buffer must be completely invalid in order to
 419                          * take part in the cluster.  If it is partially valid
 420                          * then we stop.
 421                          */
 422                         for (j = 0;j < tbp->b_npages; j++) {
 423                                 if (tbp->b_pages[j]->valid)
 424                                         break;
 425                         }
 426                         if (j != tbp->b_npages) {
 427                                 bqrelse(tbp);
 428                                 break;
 429                         }
 430
 431                         /*
 432                          * Set a read-ahead mark as appropriate
 433                          */
 434                         if ((fbp && (i == 1)) || (i == (run - 1)))
 435                                 tbp->b_flags |= B_RAM;
 436
 437                         /*
 438                          * Set the buffer up for an async read (XXX should
 439                          * we do this only if we do not wind up brelse()ing?).
 440                          * Set the block number if it isn't set, otherwise
 441                          * if it is make sure it matches the block number we
 442                          * expect.
 443                          */
 444                         tbp->b_flags |= B_READ | B_ASYNC;
 445                         if (tbp->b_blkno == tbp->b_lblkno) {
 446                                 tbp->b_blkno = bn;
 447                         } else if (tbp->b_blkno != bn) {
 448                                 brelse(tbp);
 449                                 break;
 450                         }
 451                 }
 452                 /*
 453                  * XXX fbp from caller may not be B_ASYNC, but we are going
 454                  * to biodone() it in cluster_callback() anyway
 455                  */
 456                 BUF_KERNPROC(tbp);
 457                 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 458                         tbp, b_cluster.cluster_entry);
 459                 for (j = 0; j < tbp->b_npages; j += 1) {
 460                         vm_page_t m;
 461                         m = tbp->b_pages[j];
 462                         vm_page_io_start(m);
 463                         vm_object_pip_add(m->object, 1);
 464                         if ((bp->b_npages == 0) ||
 465                                 (bp->b_pages[bp->b_npages-1] != m)) {
 466                                 bp->b_pages[bp->b_npages] = m;
 467                                 bp->b_npages++;
 468                         }
 469                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 470                                 tbp->b_pages[j] = bogus_page;
 471                 }
 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         for (j = 0; j < bp->b_npages; j++) {
 493                 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 494                     VM_PAGE_BITS_ALL) {
 495                         bp->b_pages[j] = bogus_page;
 496                 }
 497         }
 498         if (bp->b_bufsize > bp->b_kvasize)
 499                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 500                     bp->b_bufsize, bp->b_kvasize);
 501         bp->b_kvasize = bp->b_bufsize;
 502
 503         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 504                 (vm_page_t *)bp->b_pages, bp->b_npages);
 505         return (bp);
 506 }
 507
 508 /*
 509  * Cleanup after a clustered read or write.
 510  * This is complicated by the fact that any of the buffers might have
 511  * extra memory (if there were no empty buffer headers at allocbuf time)
 512  * that we will need to shift around.
 513  */
 514 void
 515 cluster_callback(bp)
 516         struct buf *bp;
 517 {
 518         struct buf *nbp, *tbp;
 519         int error = 0;
 520
 521         /*
 522          * Must propogate errors to all the components.
 523          */
 524         if (bp->b_flags & B_ERROR)
 525                 error = bp->b_error;
 526
 527         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
 528         /*
 529          * Move memory from the large cluster buffer into the component
 530          * buffers and mark IO as done on these.
 531          */
 532         for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
 533                 tbp; tbp = nbp) {
 534                 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
 535                 if (error) {
 536                         tbp->b_flags |= B_ERROR;
 537                         tbp->b_error = error;
 538                 } else {
 539                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 540                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 541                         /*
 542                          * XXX the bdwrite()/bqrelse() issued during
 543                          * cluster building clears B_RELBUF (see bqrelse()
 544                          * comment).  If direct I/O was specified, we have
 545                          * to restore it here to allow the buffer and VM
 546                          * to be freed.
 547                          */
 548                         if (tbp->b_flags & B_DIRECT)
 549                                 tbp->b_flags |= B_RELBUF;
 550                 }
 551                 biodone(tbp);
 552         }
 553         relpbuf(bp, &cluster_pbuf_freecnt);
 554 }
 555
 556 /*
 557  *      cluster_wbuild_wb:
 558  *
 559  *      Implement modified write build for cluster.
 560  *
 561  *              write_behind = 0        write behind disabled
 562  *              write_behind = 1        write behind normal (default)
 563  *              write_behind = 2        write behind backed-off
 564  */
 565
 566 static __inline int
 567 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
 568 {
 569         int r = 0;
 570
 571         switch(write_behind) {
 572         case 2:
 573                 if (start_lbn < len)
 574                         break;
 575                 start_lbn -= len;
 576                 /* fall through */
 577         case 1:
 578                 r = cluster_wbuild(vp, size, start_lbn, len);
 579                 /* fall through */
 580         default:
 581                 /* fall through */
 582                 break;
 583         }
 584         return(r);
 585 }
 586
 587 /*
 588  * Do clustered write for FFS.
 589  *
 590  * Three cases:
 591  *      1. Write is not sequential (write asynchronously)
 592  *      Write is sequential:
 593  *      2.      beginning of cluster - begin cluster
 594  *      3.      middle of a cluster - add to cluster
 595  *      4.      end of a cluster - asynchronously write cluster
 596  */
 597 void
 598 cluster_write(bp, filesize, seqcount)
 599         struct buf *bp;
 600         u_quad_t filesize;
 601         int seqcount;
 602 {
 603         struct vnode *vp;
 604         daddr_t lbn;
 605         int maxclen, cursize;
 606         int lblocksize;
 607         int async;
 608
 609         vp = bp->b_vp;
 610         if (vp->v_type == VREG) {
 611                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 612                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 613         } else {
 614                 async = 0;
 615                 lblocksize = bp->b_bufsize;
 616         }
 617         lbn = bp->b_lblkno;
 618         KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
 619
 620         /* Initialize vnode to beginning of file. */
 621         if (lbn == 0)
 622                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 623
 624         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
 625             (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
 626                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
 627                 if (vp->v_clen != 0) {
 628                         /*
 629                          * Next block is not sequential.
 630                          *
 631                          * If we are not writing at end of file, the process
 632                          * seeked to another point in the file since its last
 633                          * write, or we have reached our maximum cluster size,
 634                          * then push the previous cluster. Otherwise try
 635                          * reallocating to make it sequential.
 636                          *
 637                          * Change to algorithm: only push previous cluster if
 638                          * it was sequential from the point of view of the
 639                          * seqcount heuristic, otherwise leave the buffer
 640                          * intact so we can potentially optimize the I/O
 641                          * later on in the buf_daemon or update daemon
 642                          * flush.
 643                          */
 644                         cursize = vp->v_lastw - vp->v_cstart + 1;
 645                         if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
 646                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
 647                                 if (!async && seqcount > 0) {
 648                                         cluster_wbuild_wb(vp, lblocksize,
 649                                                 vp->v_cstart, cursize);
 650                                 }
 651                         } else {
 652                                 struct buf **bpp, **endbp;
 653                                 struct cluster_save *buflist;
 654
 655                                 buflist = cluster_collectbufs(vp, bp);
 656                                 endbp = &buflist->bs_children
 657                                     [buflist->bs_nchildren - 1];
 658                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 659                                         /*
 660                                          * Failed, push the previous cluster
 661                                          * if *really* writing sequentially
 662                                          * in the logical file (seqcount > 1),
 663                                          * otherwise delay it in the hopes that
 664                                          * the low level disk driver can
 665                                          * optimize the write ordering.
 666                                          */
 667                                         for (bpp = buflist->bs_children;
 668                                              bpp < endbp; bpp++)
 669                                                 brelse(*bpp);
 670                                         free(buflist, M_SEGMENT);
 671                                         if (seqcount > 1) {
 672                                                 cluster_wbuild_wb(vp,
 673                                                     lblocksize, vp->v_cstart,
 674                                                     cursize);
 675                                         }
 676                                 } else {
 677                                         /*
 678                                          * Succeeded, keep building cluster.
 679                                          */
 680                                         for (bpp = buflist->bs_children;
 681                                              bpp <= endbp; bpp++)
 682                                                 bdwrite(*bpp);
 683                                         free(buflist, M_SEGMENT);
 684                                         vp->v_lastw = lbn;
 685                                         vp->v_lasta = bp->b_blkno;
 686                                         return;
 687                                 }
 688                         }
 689                 }
 690                 /*
 691                  * Consider beginning a cluster. If at end of file, make
 692                  * cluster as large as possible, otherwise find size of
 693                  * existing cluster.
 694                  */
 695                 if ((vp->v_type == VREG) &&
 696                         ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
 697                     (bp->b_blkno == bp->b_lblkno) &&
 698                     (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
 699                      bp->b_blkno == -1)) {
 700                         bawrite(bp);
 701                         vp->v_clen = 0;
 702                         vp->v_lasta = bp->b_blkno;
 703                         vp->v_cstart = lbn + 1;
 704                         vp->v_lastw = lbn;
 705                         return;
 706                 }
 707                 vp->v_clen = maxclen;
 708                 if (!async && maxclen == 0) {   /* I/O not contiguous */
 709                         vp->v_cstart = lbn + 1;
 710                         bawrite(bp);
 711                 } else {        /* Wait for rest of cluster */
 712                         vp->v_cstart = lbn;
 713                         bdwrite(bp);
 714                 }
 715         } else if (lbn == vp->v_cstart + vp->v_clen) {
 716                 /*
 717                  * At end of cluster, write it out if seqcount tells us we
 718                  * are operating sequentially, otherwise let the buf or
 719                  * update daemon handle it.
 720                  */
 721                 bdwrite(bp);
 722                 if (seqcount > 1)
 723                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
 724                 vp->v_clen = 0;
 725                 vp->v_cstart = lbn + 1;
 726         } else if (vm_page_count_severe()) {
 727                 /*
 728                  * We are low on memory, get it going NOW
 729                  */
 730                 bawrite(bp);
 731         } else {
 732                 /*
 733                  * In the middle of a cluster, so just delay the I/O for now.
 734                  */
 735                 bdwrite(bp);
 736         }
 737         vp->v_lastw = lbn;
 738         vp->v_lasta = bp->b_blkno;
 739 }
 740
 741
 742 /*
 743  * This is an awful lot like cluster_rbuild...wish they could be combined.
 744  * The last lbn argument is the current block on which I/O is being
 745  * performed.  Check to see that it doesn't fall in the middle of
 746  * the current block (if last_bp == NULL).
 747  */
 748 int
 749 cluster_wbuild(vp, size, start_lbn, len)
 750         struct vnode *vp;
 751         long size;
 752         daddr_t start_lbn;
 753         int len;
 754 {
 755         struct buf *bp, *tbp;
 756         int i, j, s;
 757         int totalwritten = 0;
 758         int dbsize = btodb(size);
 759
 760         while (len > 0) {
 761                 s = splbio();
 762                 /*
 763                  * If the buffer is not delayed-write (i.e. dirty), or it
 764                  * is delayed-write but either locked or inval, it cannot
 765                  * partake in the clustered write.
 766                  */
 767                 if (((tbp = gbincore(vp, start_lbn)) == NULL) ||
 768                   ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
 769                   BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 770                         ++start_lbn;
 771                         --len;
 772                         splx(s);
 773                         continue;
 774                 }
 775                 bremfree(tbp);
 776                 tbp->b_flags &= ~B_DONE;
 777                 splx(s);
 778
 779                 /*
 780                  * Extra memory in the buffer, punt on this buffer.
 781                  * XXX we could handle this in most cases, but we would
 782                  * have to push the extra memory down to after our max
 783                  * possible cluster size and then potentially pull it back
 784                  * up if the cluster was terminated prematurely--too much
 785                  * hassle.
 786                  */
 787                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 788                   (tbp->b_bcount != tbp->b_bufsize) ||
 789                   (tbp->b_bcount != size) ||
 790                   (len == 1) ||
 791                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 792                         totalwritten += tbp->b_bufsize;
 793                         bawrite(tbp);
 794                         ++start_lbn;
 795                         --len;
 796                         continue;
 797                 }
 798
 799                 /*
 800                  * We got a pbuf to make the cluster in.
 801                  * so initialise it.
 802                  */
 803                 TAILQ_INIT(&bp->b_cluster.cluster_head);
 804                 bp->b_bcount = 0;
 805                 bp->b_bufsize = 0;
 806                 bp->b_npages = 0;
 807                 if (tbp->b_wcred != NOCRED) {
 808                     bp->b_wcred = tbp->b_wcred;
 809                     crhold(bp->b_wcred);
 810                 }
 811
 812                 bp->b_blkno = tbp->b_blkno;
 813                 bp->b_lblkno = tbp->b_lblkno;
 814                 bp->b_offset = tbp->b_offset;
 815
 816                 /*
 817                  * We are synthesizing a buffer out of vm_page_t's, but
 818                  * if the block size is not page aligned then the starting
 819                  * address may not be either.  Inherit the b_data offset
 820                  * from the original buffer.
 821                  */
 822                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 823                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 824                 bp->b_flags |= B_CALL | B_CLUSTER |
 825                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN));
 826                 bp->b_iodone = cluster_callback;
 827                 pbgetvp(vp, bp);
 828                 /*
 829                  * From this location in the file, scan forward to see
 830                  * if there are buffers with adjacent data that need to
 831                  * be written as well.
 832                  */
 833                 for (i = 0; i < len; ++i, ++start_lbn) {
 834                         if (i != 0) { /* If not the first buffer */
 835                                 s = splbio();
 836                                 /*
 837                                  * If the adjacent data is not even in core it
 838                                  * can't need to be written.
 839                                  */
 840                                 if ((tbp = gbincore(vp, start_lbn)) == NULL) {
 841                                         splx(s);
 842                                         break;
 843                                 }
 844
 845                                 /*
 846                                  * If it IS in core, but has different
 847                                  * characteristics, or is locked (which
 848                                  * means it could be undergoing a background
 849                                  * I/O or be in a weird state), then don't
 850                                  * cluster with it.
 851                                  */
 852                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 853                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 854                                   != (B_DELWRI | B_CLUSTEROK |
 855                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 856                                     (tbp->b_flags & B_LOCKED) ||
 857                                     tbp->b_wcred != bp->b_wcred ||
 858                                     BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 859                                         splx(s);
 860                                         break;
 861                                 }
 862
 863                                 /*
 864                                  * Check that the combined cluster
 865                                  * would make sense with regard to pages
 866                                  * and would not be too large
 867                                  */
 868                                 if ((tbp->b_bcount != size) ||
 869                                   ((bp->b_blkno + (dbsize * i)) !=
 870                                     tbp->b_blkno) ||
 871                                   ((tbp->b_npages + bp->b_npages) >
 872                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 873                                         BUF_UNLOCK(tbp);
 874                                         splx(s);
 875                                         break;
 876                                 }
 877                                 /*
 878                                  * Ok, it's passed all the tests,
 879                                  * so remove it from the free list
 880                                  * and mark it busy. We will use it.
 881                                  */
 882                                 bremfree(tbp);
 883                                 tbp->b_flags &= ~B_DONE;
 884                                 splx(s);
 885                         } /* end of code for non-first buffers only */
 886                         /* check for latent dependencies to be handled */
 887                         if ((LIST_FIRST(&tbp->b_dep)) != NULL &&
 888                             bioops.io_start)
 889                                 (*bioops.io_start)(tbp);
 890                         /*
 891                          * If the IO is via the VM then we do some
 892                          * special VM hackery (yuck).  Since the buffer's
 893                          * block size may not be page-aligned it is possible
 894                          * for a page to be shared between two buffers.  We
 895                          * have to get rid of the duplication when building
 896                          * the cluster.
 897                          */
 898                         if (tbp->b_flags & B_VMIO) {
 899                                 vm_page_t m;
 900
 901                                 if (i != 0) { /* if not first buffer */
 902                                         for (j = 0; j < tbp->b_npages; j += 1) {
 903                                                 m = tbp->b_pages[j];
 904                                                 if (m->flags & PG_BUSY) {
 905                                                         bqrelse(tbp);
 906                                                         goto finishcluster;
 907                                                 }
 908                                         }
 909                                 }
 910
 911                                 for (j = 0; j < tbp->b_npages; j += 1) {
 912                                         m = tbp->b_pages[j];
 913                                         vm_page_io_start(m);
 914                                         vm_object_pip_add(m->object, 1);
 915                                         if ((bp->b_npages == 0) ||
 916                                           (bp->b_pages[bp->b_npages - 1] != m)) {
 917                                                 bp->b_pages[bp->b_npages] = m;
 918                                                 bp->b_npages++;
 919                                         }
 920                                 }
 921                         }
 922                         bp->b_bcount += size;
 923                         bp->b_bufsize += size;
 924
 925                         s = splbio();
 926                         bundirty(tbp);
 927                         tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
 928                         tbp->b_flags |= B_ASYNC;
 929                         reassignbuf(tbp, tbp->b_vp);    /* put on clean list */
 930                         ++tbp->b_vp->v_numoutput;
 931                         splx(s);
 932                         BUF_KERNPROC(tbp);
 933                         TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 934                                 tbp, b_cluster.cluster_entry);
 935                 }
 936         finishcluster:
 937                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 938                         (vm_page_t *) bp->b_pages, bp->b_npages);
 939                 if (bp->b_bufsize > bp->b_kvasize)
 940                         panic(
 941                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 942                             bp->b_bufsize, bp->b_kvasize);
 943                 bp->b_kvasize = bp->b_bufsize;
 944                 totalwritten += bp->b_bufsize;
 945                 bp->b_dirtyoff = 0;
 946                 bp->b_dirtyend = bp->b_bufsize;
 947                 bawrite(bp);
 948
 949                 len -= i;
 950         }
 951         return totalwritten;
 952 }
 953
 954 /*
 955  * Collect together all the buffers in a cluster.
 956  * Plus add one additional buffer.
 957  */
 958 static struct cluster_save *
 959 cluster_collectbufs(vp, last_bp)
 960         struct vnode *vp;
 961         struct buf *last_bp;
 962 {
 963         struct cluster_save *buflist;
 964         struct buf *bp;
 965         daddr_t lbn;
 966         int i, len;
 967
 968         len = vp->v_lastw - vp->v_cstart + 1;
 969         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
 970             M_SEGMENT, M_WAITOK);
 971         buflist->bs_nchildren = 0;
 972         buflist->bs_children = (struct buf **) (buflist + 1);
 973         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
 974                 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
 975                 buflist->bs_children[i] = bp;
 976                 if (bp->b_blkno == bp->b_lblkno)
 977                         VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
 978                                 NULL, NULL);
 979         }
 980         buflist->bs_children[i] = bp = last_bp;
 981         if (bp->b_blkno == bp->b_lblkno)
 982                 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
 983                         NULL, NULL);
 984         buflist->bs_nchildren = i + 1;
 985         return (buflist);
 986 }