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 1.17 2006/03/05 18:38:34 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 #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", "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 static void cluster_callback (struct bio *);
  73
  74
  75 static int write_behind = 1;
  76 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
  77
  78 extern vm_page_t        bogus_page;
  79
  80 extern int cluster_pbuf_freecnt;
  81
  82 /*
  83  * Maximum number of blocks for read-ahead.
  84  */
  85 #define MAXRA 32
  86
  87 /*
  88  * This replaces bread.
  89  */
  90 int
  91 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno,
  92         long size, long totread, int seqcount, struct buf **bpp)
  93 {
  94         struct buf *bp, *rbp, *reqbp;
  95         daddr_t blkno, origblkno;
  96         int error, num_ra;
  97         int i;
  98         int maxra, racluster;
  99         long origtotread;
 100
 101         error = 0;
 102
 103         /*
 104          * Try to limit the amount of read-ahead by a few
 105          * ad-hoc parameters.  This needs work!!!
 106          */
 107         racluster = vp->v_mount->mnt_iosize_max / size;
 108         maxra = 2 * racluster + (totread / size);
 109         if (maxra > MAXRA)
 110                 maxra = MAXRA;
 111         if (maxra > nbuf/8)
 112                 maxra = nbuf/8;
 113
 114         /*
 115          * get the requested block
 116          */
 117         *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0);
 118         origblkno = lblkno;
 119         origtotread = totread;
 120
 121         /*
 122          * if it is in the cache, then check to see if the reads have been
 123          * sequential.  If they have, then try some read-ahead, otherwise
 124          * back-off on prospective read-aheads.
 125          */
 126         if (bp->b_flags & B_CACHE) {
 127                 if (!seqcount) {
 128                         return 0;
 129                 } else if ((bp->b_flags & B_RAM) == 0) {
 130                         return 0;
 131                 } else {
 132                         struct buf *tbp;
 133                         bp->b_flags &= ~B_RAM;
 134                         /*
 135                          * We do the crit here so that there is no window
 136                          * between the findblk and the b_usecount increment
 137                          * below.  We opt to keep the crit out of the loop
 138                          * for efficiency.
 139                          */
 140                         crit_enter();
 141                         for (i = 1; i < maxra; i++) {
 142
 143                                 if (!(tbp = findblk(vp, lblkno+i))) {
 144                                         break;
 145                                 }
 146
 147                                 /*
 148                                  * Set another read-ahead mark so we know
 149                                  * to check again.
 150                                  */
 151                                 if (((i % racluster) == (racluster - 1)) ||
 152                                         (i == (maxra - 1)))
 153                                         tbp->b_flags |= B_RAM;
 154                         }
 155                         crit_exit();
 156                         if (i >= maxra) {
 157                                 return 0;
 158                         }
 159                         lblkno += i;
 160                 }
 161                 reqbp = bp = NULL;
 162         } else {
 163                 off_t firstread = bp->b_loffset;
 164
 165                 KASSERT(firstread != NOOFFSET,
 166                         ("cluster_read: no buffer offset"));
 167                 if (firstread + totread > filesize)
 168                         totread = filesize - firstread;
 169                 if (totread > size) {
 170                         int nblks = 0;
 171                         int ncontigafter;
 172                         while (totread > 0) {
 173                                 nblks++;
 174                                 totread -= size;
 175                         }
 176                         if (nblks == 1)
 177                                 goto single_block_read;
 178                         if (nblks > racluster)
 179                                 nblks = racluster;
 180
 181                         error = VOP_BMAP(vp, lblkno, NULL,
 182                                 &blkno, &ncontigafter, NULL);
 183                         if (error)
 184                                 goto single_block_read;
 185                         if (blkno == -1)
 186                                 goto single_block_read;
 187                         if (ncontigafter == 0)
 188                                 goto single_block_read;
 189                         if (ncontigafter + 1 < nblks)
 190                                 nblks = ncontigafter + 1;
 191
 192                         bp = cluster_rbuild(vp, filesize, lblkno,
 193                                 blkno, size, nblks, bp);
 194                         lblkno += (bp->b_bufsize / size);
 195                 } else {
 196 single_block_read:
 197                         /*
 198                          * if it isn't in the cache, then get a chunk from
 199                          * disk if sequential, otherwise just get the block.
 200                          */
 201                         bp->b_flags |= B_READ | B_RAM;
 202                         lblkno += 1;
 203                 }
 204         }
 205
 206         /*
 207          * if we have been doing sequential I/O, then do some read-ahead
 208          */
 209         rbp = NULL;
 210         if (seqcount && (lblkno < (origblkno + seqcount))) {
 211                 /*
 212                  * we now build the read-ahead buffer if it is desirable.
 213                  */
 214                 if (((u_quad_t)(lblkno + 1) * size) <= filesize &&
 215                     !(error = VOP_BMAP(vp, lblkno, NULL, &blkno, &num_ra, NULL)) &&
 216                     blkno != -1) {
 217                         int nblksread;
 218                         int ntoread = num_ra + 1;
 219                         nblksread = (origtotread + size - 1) / size;
 220                         if (seqcount < nblksread)
 221                                 seqcount = nblksread;
 222                         if (seqcount < ntoread)
 223                                 ntoread = seqcount;
 224                         if (num_ra) {
 225                                 rbp = cluster_rbuild(vp, filesize, lblkno,
 226                                         blkno, size, ntoread, NULL);
 227                         } else {
 228                                 rbp = getblk(vp, lblkno, size, 0, 0);
 229                                 rbp->b_flags |= B_READ | B_ASYNC | B_RAM;
 230                                 rbp->b_bio2.bio_blkno = blkno;
 231                         }
 232                 }
 233         }
 234
 235         /*
 236          * handle the synchronous read
 237          */
 238         if (bp) {
 239 #if defined(CLUSTERDEBUG)
 240                 if (rcluster)
 241                         printf("S(%ld,%ld,%d) ",
 242                             (long)bp->b_lblkno, bp->b_bcount, seqcount);
 243 #endif
 244                 if ((bp->b_flags & B_CLUSTER) == 0) {
 245                         vfs_busy_pages(bp, 0);
 246                 }
 247                 bp->b_flags &= ~(B_ERROR|B_INVAL);
 248                 if ((bp->b_flags & B_ASYNC) || bp->b_bio1.bio_done != NULL)
 249                         BUF_KERNPROC(bp);
 250                 vn_strategy(vp, &bp->b_bio1);
 251                 error = bp->b_error;
 252         }
 253
 254         /*
 255          * and if we have read-aheads, do them too
 256          */
 257         if (rbp) {
 258                 if (error) {
 259                         rbp->b_flags &= ~(B_ASYNC | B_READ);
 260                         brelse(rbp);
 261                 } else if (rbp->b_flags & B_CACHE) {
 262                         rbp->b_flags &= ~(B_ASYNC | B_READ);
 263                         bqrelse(rbp);
 264                 } else {
 265 #if defined(CLUSTERDEBUG)
 266                         if (rcluster) {
 267                                 if (bp)
 268                                         printf("A+(%ld,%ld,%ld,%d) ",
 269                                             (long)rbp->b_lblkno, rbp->b_bcount,
 270                                             (long)(rbp->b_lblkno - origblkno),
 271                                             seqcount);
 272                                 else
 273                                         printf("A(%ld,%ld,%ld,%d) ",
 274                                             (long)rbp->b_lblkno, rbp->b_bcount,
 275                                             (long)(rbp->b_lblkno - origblkno),
 276                                             seqcount);
 277                         }
 278 #endif
 279
 280                         if ((rbp->b_flags & B_CLUSTER) == 0) {
 281                                 vfs_busy_pages(rbp, 0);
 282                         }
 283                         rbp->b_flags &= ~(B_ERROR|B_INVAL);
 284                         if ((rbp->b_flags & B_ASYNC) || rbp->b_bio1.bio_done != NULL)
 285                                 BUF_KERNPROC(rbp);
 286                         vn_strategy(vp, &rbp->b_bio1);
 287                 }
 288         }
 289         if (reqbp)
 290                 return (biowait(reqbp));
 291         else
 292                 return (error);
 293 }
 294
 295 /*
 296  * If blocks are contiguous on disk, use this to provide clustered
 297  * read ahead.  We will read as many blocks as possible sequentially
 298  * and then parcel them up into logical blocks in the buffer hash table.
 299  */
 300 static struct buf *
 301 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
 302         daddr_t blkno, long size, int run, struct buf *fbp)
 303 {
 304         struct buf *bp, *tbp;
 305         daddr_t bn;
 306         int i, inc, j;
 307
 308         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 309             ("cluster_rbuild: size %ld != filesize %ld\n",
 310             size, vp->v_mount->mnt_stat.f_iosize));
 311
 312         /*
 313          * avoid a division
 314          */
 315         while ((u_quad_t) size * (lbn + run) > filesize) {
 316                 --run;
 317         }
 318
 319         if (fbp) {
 320                 tbp = fbp;
 321                 tbp->b_flags |= B_READ;
 322         } else {
 323                 tbp = getblk(vp, lbn, size, 0, 0);
 324                 if (tbp->b_flags & B_CACHE)
 325                         return tbp;
 326                 tbp->b_flags |= B_ASYNC | B_READ | B_RAM;
 327         }
 328
 329         tbp->b_bio2.bio_blkno = blkno;
 330         if( (tbp->b_flags & B_MALLOC) ||
 331                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
 332                 return tbp;
 333
 334         bp = trypbuf(&cluster_pbuf_freecnt);
 335         if (bp == 0)
 336                 return tbp;
 337
 338         /*
 339          * We are synthesizing a buffer out of vm_page_t's, but
 340          * if the block size is not page aligned then the starting
 341          * address may not be either.  Inherit the b_data offset
 342          * from the original buffer.
 343          */
 344         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 345             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 346         bp->b_flags = B_ASYNC | B_READ | B_CLUSTER | B_VMIO;
 347         bp->b_bio1.bio_done = cluster_callback;
 348         bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 349         bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 350         bp->b_lblkno = lbn;
 351         bp->b_loffset = tbp->b_loffset;
 352         bp->b_bio2.bio_blkno = (daddr_t)-1;
 353         KASSERT(bp->b_loffset != NOOFFSET,
 354                 ("cluster_rbuild: no buffer offset"));
 355         pbgetvp(vp, bp);
 356
 357         bp->b_bcount = 0;
 358         bp->b_bufsize = 0;
 359         bp->b_xio.xio_npages = 0;
 360
 361         inc = btodb(size);
 362         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
 363                 if (i != 0) {
 364                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
 365                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 366                                 break;
 367                         }
 368
 369                         /*
 370                          * Shortcut some checks and try to avoid buffers that
 371                          * would block in the lock.  The same checks have to
 372                          * be made again after we officially get the buffer.
 373                          */
 374                         if ((tbp = findblk(vp, lbn + i)) != NULL) {
 375                                 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
 376                                         break;
 377                                 BUF_UNLOCK(tbp);
 378
 379                                 for (j = 0; j < tbp->b_xio.xio_npages; j++) {
 380                                         if (tbp->b_xio.xio_pages[j]->valid)
 381                                                 break;
 382                                 }
 383
 384                                 if (j != tbp->b_xio.xio_npages)
 385                                         break;
 386
 387                                 if (tbp->b_bcount != size)
 388                                         break;
 389                         }
 390
 391                         tbp = getblk(vp, lbn + i, size, 0, 0);
 392
 393                         /*
 394                          * Stop scanning if the buffer is fuly valid
 395                          * (marked B_CACHE), or locked (may be doing a
 396                          * background write), or if the buffer is not
 397                          * VMIO backed.  The clustering code can only deal
 398                          * with VMIO-backed buffers.
 399                          */
 400                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 401                             (tbp->b_flags & B_VMIO) == 0) {
 402                                 bqrelse(tbp);
 403                                 break;
 404                         }
 405
 406                         /*
 407                          * The buffer must be completely invalid in order to
 408                          * take part in the cluster.  If it is partially valid
 409                          * then we stop.
 410                          */
 411                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
 412                                 if (tbp->b_xio.xio_pages[j]->valid)
 413                                         break;
 414                         }
 415                         if (j != tbp->b_xio.xio_npages) {
 416                                 bqrelse(tbp);
 417                                 break;
 418                         }
 419
 420                         /*
 421                          * Set a read-ahead mark as appropriate
 422                          */
 423                         if ((fbp && (i == 1)) || (i == (run - 1)))
 424                                 tbp->b_flags |= B_RAM;
 425
 426                         /*
 427                          * Set the buffer up for an async read (XXX should
 428                          * we do this only if we do not wind up brelse()ing?).
 429                          * Set the block number if it isn't set, otherwise
 430                          * if it is make sure it matches the block number we
 431                          * expect.
 432                          */
 433                         tbp->b_flags |= B_READ | B_ASYNC;
 434                         if (tbp->b_bio2.bio_blkno == (daddr_t)-1) {
 435                                 tbp->b_bio2.bio_blkno = bn;
 436                         } else if (tbp->b_bio2.bio_blkno != bn) {
 437                                 brelse(tbp);
 438                                 break;
 439                         }
 440                 }
 441                 /*
 442                  * XXX fbp from caller may not be B_ASYNC, but we are going
 443                  * to biodone() it in cluster_callback() anyway
 444                  */
 445                 BUF_KERNPROC(tbp);
 446                 cluster_append(&bp->b_bio1, tbp);
 447                 for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
 448                         vm_page_t m;
 449                         m = tbp->b_xio.xio_pages[j];
 450                         vm_page_io_start(m);
 451                         vm_object_pip_add(m->object, 1);
 452                         if ((bp->b_xio.xio_npages == 0) ||
 453                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
 454                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 455                                 bp->b_xio.xio_npages++;
 456                         }
 457                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 458                                 tbp->b_xio.xio_pages[j] = bogus_page;
 459                 }
 460                 /*
 461                  * XXX shouldn't this be += size for both, like in
 462                  * cluster_wbuild()?
 463                  *
 464                  * Don't inherit tbp->b_bufsize as it may be larger due to
 465                  * a non-page-aligned size.  Instead just aggregate using
 466                  * 'size'.
 467                  */
 468                 if (tbp->b_bcount != size)
 469                     printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
 470                 if (tbp->b_bufsize != size)
 471                     printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
 472                 bp->b_bcount += size;
 473                 bp->b_bufsize += size;
 474         }
 475
 476         /*
 477          * Fully valid pages in the cluster are already good and do not need
 478          * to be re-read from disk.  Replace the page with bogus_page
 479          */
 480         for (j = 0; j < bp->b_xio.xio_npages; j++) {
 481                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 482                     VM_PAGE_BITS_ALL) {
 483                         bp->b_xio.xio_pages[j] = bogus_page;
 484                 }
 485         }
 486         if (bp->b_bufsize > bp->b_kvasize)
 487                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)",
 488                     bp->b_bufsize, bp->b_kvasize);
 489         bp->b_kvasize = bp->b_bufsize;
 490
 491         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 492                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 493         return (bp);
 494 }
 495
 496 /*
 497  * Cleanup after a clustered read or write.
 498  * This is complicated by the fact that any of the buffers might have
 499  * extra memory (if there were no empty buffer headers at allocbuf time)
 500  * that we will need to shift around.
 501  *
 502  * The returned bio is &bp->b_bio1
 503  */
 504 void
 505 cluster_callback(struct bio *bio)
 506 {
 507         struct buf *bp = bio->bio_buf;
 508         struct buf *tbp;
 509         int error = 0;
 510
 511         /*
 512          * Must propogate errors to all the components.
 513          */
 514         if (bp->b_flags & B_ERROR)
 515                 error = bp->b_error;
 516
 517         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
 518         /*
 519          * Move memory from the large cluster buffer into the component
 520          * buffers and mark IO as done on these.  Since the memory map
 521          * is the same, no actual copying is required.
 522          */
 523         while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
 524                 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
 525                 if (error) {
 526                         tbp->b_flags |= B_ERROR;
 527                         tbp->b_error = error;
 528                 } else {
 529                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 530                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 531                         /*
 532                          * XXX the bdwrite()/bqrelse() issued during
 533                          * cluster building clears B_RELBUF (see bqrelse()
 534                          * comment).  If direct I/O was specified, we have
 535                          * to restore it here to allow the buffer and VM
 536                          * to be freed.
 537                          */
 538                         if (tbp->b_flags & B_DIRECT)
 539                                 tbp->b_flags |= B_RELBUF;
 540                 }
 541                 biodone(&tbp->b_bio1);
 542         }
 543         relpbuf(bp, &cluster_pbuf_freecnt);
 544 }
 545
 546 /*
 547  *      cluster_wbuild_wb:
 548  *
 549  *      Implement modified write build for cluster.
 550  *
 551  *              write_behind = 0        write behind disabled
 552  *              write_behind = 1        write behind normal (default)
 553  *              write_behind = 2        write behind backed-off
 554  */
 555
 556 static __inline int
 557 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
 558 {
 559         int r = 0;
 560
 561         switch(write_behind) {
 562         case 2:
 563                 if (start_lbn < len)
 564                         break;
 565                 start_lbn -= len;
 566                 /* fall through */
 567         case 1:
 568                 r = cluster_wbuild(vp, size, start_lbn, len);
 569                 /* fall through */
 570         default:
 571                 /* fall through */
 572                 break;
 573         }
 574         return(r);
 575 }
 576
 577 /*
 578  * Do clustered write for FFS.
 579  *
 580  * Three cases:
 581  *      1. Write is not sequential (write asynchronously)
 582  *      Write is sequential:
 583  *      2.      beginning of cluster - begin cluster
 584  *      3.      middle of a cluster - add to cluster
 585  *      4.      end of a cluster - asynchronously write cluster
 586  */
 587 void
 588 cluster_write(struct buf *bp, u_quad_t filesize, int seqcount)
 589 {
 590         struct vnode *vp;
 591         daddr_t lbn;
 592         int maxclen, cursize;
 593         int lblocksize;
 594         int async;
 595
 596         vp = bp->b_vp;
 597         if (vp->v_type == VREG) {
 598                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 599                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 600         } else {
 601                 async = 0;
 602                 lblocksize = bp->b_bufsize;
 603         }
 604         lbn = bp->b_lblkno;
 605         KASSERT(bp->b_loffset != NOOFFSET,
 606                 ("cluster_write: no buffer offset"));
 607
 608         /* Initialize vnode to beginning of file. */
 609         if (lbn == 0)
 610                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 611
 612         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
 613             bp->b_bio2.bio_blkno == (daddr_t)-1 ||
 614             (bp->b_bio2.bio_blkno != vp->v_lasta + btodb(lblocksize))) {
 615                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
 616                 if (vp->v_clen != 0) {
 617                         /*
 618                          * Next block is not sequential.
 619                          *
 620                          * If we are not writing at end of file, the process
 621                          * seeked to another point in the file since its last
 622                          * write, or we have reached our maximum cluster size,
 623                          * then push the previous cluster. Otherwise try
 624                          * reallocating to make it sequential.
 625                          *
 626                          * Change to algorithm: only push previous cluster if
 627                          * it was sequential from the point of view of the
 628                          * seqcount heuristic, otherwise leave the buffer
 629                          * intact so we can potentially optimize the I/O
 630                          * later on in the buf_daemon or update daemon
 631                          * flush.
 632                          */
 633                         cursize = vp->v_lastw - vp->v_cstart + 1;
 634                         if (((u_quad_t) bp->b_loffset + lblocksize) != filesize ||
 635                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
 636                                 if (!async && seqcount > 0) {
 637                                         cluster_wbuild_wb(vp, lblocksize,
 638                                                 vp->v_cstart, cursize);
 639                                 }
 640                         } else {
 641                                 struct buf **bpp, **endbp;
 642                                 struct cluster_save *buflist;
 643
 644                                 buflist = cluster_collectbufs(vp, bp);
 645                                 endbp = &buflist->bs_children
 646                                     [buflist->bs_nchildren - 1];
 647                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 648                                         /*
 649                                          * Failed, push the previous cluster
 650                                          * if *really* writing sequentially
 651                                          * in the logical file (seqcount > 1),
 652                                          * otherwise delay it in the hopes that
 653                                          * the low level disk driver can
 654                                          * optimize the write ordering.
 655                                          */
 656                                         for (bpp = buflist->bs_children;
 657                                              bpp < endbp; bpp++)
 658                                                 brelse(*bpp);
 659                                         free(buflist, M_SEGMENT);
 660                                         if (seqcount > 1) {
 661                                                 cluster_wbuild_wb(vp,
 662                                                     lblocksize, vp->v_cstart,
 663                                                     cursize);
 664                                         }
 665                                 } else {
 666                                         /*
 667                                          * Succeeded, keep building cluster.
 668                                          */
 669                                         for (bpp = buflist->bs_children;
 670                                              bpp <= endbp; bpp++)
 671                                                 bdwrite(*bpp);
 672                                         free(buflist, M_SEGMENT);
 673                                         vp->v_lastw = lbn;
 674                                         vp->v_lasta = bp->b_bio2.bio_blkno;
 675                                         return;
 676                                 }
 677                         }
 678                 }
 679                 /*
 680                  * Consider beginning a cluster. If at end of file, make
 681                  * cluster as large as possible, otherwise find size of
 682                  * existing cluster.
 683                  */
 684                 if ((vp->v_type == VREG) &&
 685                         ((u_quad_t) bp->b_loffset + lblocksize) != filesize &&
 686                     (bp->b_bio2.bio_blkno == (daddr_t)-1) &&
 687                     (VOP_BMAP(vp, lbn, NULL, &bp->b_bio2.bio_blkno, &maxclen, NULL) ||
 688                      bp->b_bio2.bio_blkno == (daddr_t)-1)) {
 689                         bawrite(bp);
 690                         vp->v_clen = 0;
 691                         vp->v_lasta = bp->b_bio2.bio_blkno;
 692                         vp->v_cstart = lbn + 1;
 693                         vp->v_lastw = lbn;
 694                         return;
 695                 }
 696                 vp->v_clen = maxclen;
 697                 if (!async && maxclen == 0) {   /* I/O not contiguous */
 698                         vp->v_cstart = lbn + 1;
 699                         bawrite(bp);
 700                 } else {        /* Wait for rest of cluster */
 701                         vp->v_cstart = lbn;
 702                         bdwrite(bp);
 703                 }
 704         } else if (lbn == vp->v_cstart + vp->v_clen) {
 705                 /*
 706                  * At end of cluster, write it out if seqcount tells us we
 707                  * are operating sequentially, otherwise let the buf or
 708                  * update daemon handle it.
 709                  */
 710                 bdwrite(bp);
 711                 if (seqcount > 1)
 712                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
 713                 vp->v_clen = 0;
 714                 vp->v_cstart = lbn + 1;
 715         } else if (vm_page_count_severe()) {
 716                 /*
 717                  * We are low on memory, get it going NOW
 718                  */
 719                 bawrite(bp);
 720         } else {
 721                 /*
 722                  * In the middle of a cluster, so just delay the I/O for now.
 723                  */
 724                 bdwrite(bp);
 725         }
 726         vp->v_lastw = lbn;
 727         vp->v_lasta = bp->b_bio2.bio_blkno;
 728 }
 729
 730
 731 /*
 732  * This is an awful lot like cluster_rbuild...wish they could be combined.
 733  * The last lbn argument is the current block on which I/O is being
 734  * performed.  Check to see that it doesn't fall in the middle of
 735  * the current block (if last_bp == NULL).
 736  */
 737 int
 738 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len)
 739 {
 740         struct buf *bp, *tbp;
 741         int i, j;
 742         int totalwritten = 0;
 743         int dbsize = btodb(size);
 744
 745         while (len > 0) {
 746                 crit_enter();
 747                 /*
 748                  * If the buffer is not delayed-write (i.e. dirty), or it
 749                  * is delayed-write but either locked or inval, it cannot
 750                  * partake in the clustered write.
 751                  */
 752                 if (((tbp = findblk(vp, start_lbn)) == NULL) ||
 753                   ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
 754                   BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 755                         ++start_lbn;
 756                         --len;
 757                         crit_exit();
 758                         continue;
 759                 }
 760                 bremfree(tbp);
 761                 tbp->b_flags &= ~B_DONE;
 762                 crit_exit();
 763
 764                 /*
 765                  * Extra memory in the buffer, punt on this buffer.
 766                  * XXX we could handle this in most cases, but we would
 767                  * have to push the extra memory down to after our max
 768                  * possible cluster size and then potentially pull it back
 769                  * up if the cluster was terminated prematurely--too much
 770                  * hassle.
 771                  */
 772                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 773                   (tbp->b_bcount != tbp->b_bufsize) ||
 774                   (tbp->b_bcount != size) ||
 775                   (len == 1) ||
 776                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 777                         totalwritten += tbp->b_bufsize;
 778                         bawrite(tbp);
 779                         ++start_lbn;
 780                         --len;
 781                         continue;
 782                 }
 783
 784                 /*
 785                  * We got a pbuf to make the cluster in.
 786                  * so initialise it.
 787                  */
 788                 bp->b_bcount = 0;
 789                 bp->b_bufsize = 0;
 790                 bp->b_xio.xio_npages = 0;
 791                 bp->b_lblkno = tbp->b_lblkno;
 792                 bp->b_loffset = tbp->b_loffset;
 793                 bp->b_bio2.bio_blkno = tbp->b_bio2.bio_blkno;
 794
 795                 /*
 796                  * We are synthesizing a buffer out of vm_page_t's, but
 797                  * if the block size is not page aligned then the starting
 798                  * address may not be either.  Inherit the b_data offset
 799                  * from the original buffer.
 800                  */
 801                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 802                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 803                 bp->b_flags |= B_CLUSTER |
 804                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN));
 805                 bp->b_bio1.bio_done = cluster_callback;
 806                 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 807                 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 808                 pbgetvp(vp, bp);
 809                 /*
 810                  * From this location in the file, scan forward to see
 811                  * if there are buffers with adjacent data that need to
 812                  * be written as well.
 813                  */
 814                 for (i = 0; i < len; ++i, ++start_lbn) {
 815                         if (i != 0) { /* If not the first buffer */
 816                                 crit_enter();
 817                                 /*
 818                                  * If the adjacent data is not even in core it
 819                                  * can't need to be written.
 820                                  */
 821                                 if ((tbp = findblk(vp, start_lbn)) == NULL) {
 822                                         crit_exit();
 823                                         break;
 824                                 }
 825
 826                                 /*
 827                                  * If it IS in core, but has different
 828                                  * characteristics, or is locked (which
 829                                  * means it could be undergoing a background
 830                                  * I/O or be in a weird state), then don't
 831                                  * cluster with it.
 832                                  */
 833                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 834                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 835                                   != (B_DELWRI | B_CLUSTEROK |
 836                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 837                                     (tbp->b_flags & B_LOCKED) ||
 838                                     BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 839                                         crit_exit();
 840                                         break;
 841                                 }
 842
 843                                 /*
 844                                  * Check that the combined cluster
 845                                  * would make sense with regard to pages
 846                                  * and would not be too large
 847                                  */
 848                                 if ((tbp->b_bcount != size) ||
 849                                   ((bp->b_bio2.bio_blkno + (dbsize * i)) !=
 850                                     tbp->b_bio2.bio_blkno) ||
 851                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
 852                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 853                                         BUF_UNLOCK(tbp);
 854                                         crit_exit();
 855                                         break;
 856                                 }
 857                                 /*
 858                                  * Ok, it's passed all the tests,
 859                                  * so remove it from the free list
 860                                  * and mark it busy. We will use it.
 861                                  */
 862                                 bremfree(tbp);
 863                                 tbp->b_flags &= ~B_DONE;
 864                                 crit_exit();
 865                         } /* end of code for non-first buffers only */
 866
 867                         /*
 868                          * If the IO is via the VM then we do some
 869                          * special VM hackery (yuck).  Since the buffer's
 870                          * block size may not be page-aligned it is possible
 871                          * for a page to be shared between two buffers.  We
 872                          * have to get rid of the duplication when building
 873                          * the cluster.
 874                          */
 875                         if (tbp->b_flags & B_VMIO) {
 876                                 vm_page_t m;
 877
 878                                 if (i != 0) { /* if not first buffer */
 879                                         for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
 880                                                 m = tbp->b_xio.xio_pages[j];
 881                                                 if (m->flags & PG_BUSY) {
 882                                                         bqrelse(tbp);
 883                                                         goto finishcluster;
 884                                                 }
 885                                         }
 886                                 }
 887
 888                                 for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
 889                                         m = tbp->b_xio.xio_pages[j];
 890                                         vm_page_io_start(m);
 891                                         vm_object_pip_add(m->object, 1);
 892                                         if ((bp->b_xio.xio_npages == 0) ||
 893                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
 894                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 895                                                 bp->b_xio.xio_npages++;
 896                                         }
 897                                 }
 898                         }
 899                         bp->b_bcount += size;
 900                         bp->b_bufsize += size;
 901
 902                         crit_enter();
 903                         bundirty(tbp);
 904                         tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
 905                         tbp->b_flags |= B_ASYNC;
 906                         crit_exit();
 907                         BUF_KERNPROC(tbp);
 908                         cluster_append(&bp->b_bio1, tbp);
 909
 910                         /*
 911                          * check for latent dependencies to be handled
 912                          */
 913                         if (LIST_FIRST(&tbp->b_dep) != NULL && bioops.io_start)
 914                                 (*bioops.io_start)(tbp);
 915
 916                 }
 917         finishcluster:
 918                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 919                         (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 920                 if (bp->b_bufsize > bp->b_kvasize)
 921                         panic(
 922                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 923                             bp->b_bufsize, bp->b_kvasize);
 924                 bp->b_kvasize = bp->b_bufsize;
 925                 totalwritten += bp->b_bufsize;
 926                 bp->b_dirtyoff = 0;
 927                 bp->b_dirtyend = bp->b_bufsize;
 928                 bawrite(bp);
 929
 930                 len -= i;
 931         }
 932         return totalwritten;
 933 }
 934
 935 /*
 936  * Collect together all the buffers in a cluster.
 937  * Plus add one additional buffer.
 938  */
 939 static struct cluster_save *
 940 cluster_collectbufs(struct vnode *vp, struct buf *last_bp)
 941 {
 942         struct cluster_save *buflist;
 943         struct buf *bp;
 944         daddr_t lbn;
 945         int i, len;
 946
 947         len = vp->v_lastw - vp->v_cstart + 1;
 948         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
 949             M_SEGMENT, M_WAITOK);
 950         buflist->bs_nchildren = 0;
 951         buflist->bs_children = (struct buf **) (buflist + 1);
 952         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
 953                 (void) bread(vp, lbn, last_bp->b_bcount, &bp);
 954                 buflist->bs_children[i] = bp;
 955                 if (bp->b_bio2.bio_blkno == (daddr_t)-1)
 956                         VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_bio2.bio_blkno,
 957                                 NULL, NULL);
 958         }
 959         buflist->bs_children[i] = bp = last_bp;
 960         if (bp->b_bio2.bio_blkno == (daddr_t)-1)
 961                 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_bio2.bio_blkno,
 962                         NULL, NULL);
 963         buflist->bs_nchildren = i + 1;
 964         return (buflist);
 965 }
 966
 967 void
 968 cluster_append(struct bio *bio, struct buf *tbp)
 969 {
 970         tbp->b_cluster_next = NULL;
 971         if (bio->bio_caller_info1.cluster_head == NULL) {
 972                 bio->bio_caller_info1.cluster_head = tbp;
 973                 bio->bio_caller_info2.cluster_tail = tbp;
 974         } else {
 975                 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
 976                 bio->bio_caller_info2.cluster_tail = tbp;
 977         }
 978 }
 979