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 1.28 2006/09/05 03:48:12 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                             int lblocksize);
  70 static struct buf *
  71         cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
  72                             off_t doffset, int size, int run,
  73                             struct buf *fbp, int doasync);
  74 static void cluster_callback (struct bio *);
  75
  76
  77 static int write_behind = 1;
  78 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
  79
  80 extern vm_page_t        bogus_page;
  81
  82 extern int cluster_pbuf_freecnt;
  83
  84 /*
  85  * Maximum number of blocks for read-ahead.
  86  */
  87 #define MAXRA 32
  88
  89 /*
  90  * This replaces bread.
  91  */
  92 int
  93 cluster_read(struct vnode *vp, off_t filesize, off_t loffset,
  94              int size, int totread, int seqcount, struct buf **bpp)
  95 {
  96         struct buf *bp, *rbp, *reqbp;
  97         off_t origoffset;
  98         off_t doffset;
  99         int error;
 100         int i;
 101         int maxra, racluster;
 102
 103         error = 0;
 104
 105         /*
 106          * Try to limit the amount of read-ahead by a few
 107          * ad-hoc parameters.  This needs work!!!
 108          */
 109         racluster = vp->v_mount->mnt_iosize_max / size;
 110         maxra = 2 * racluster + (totread / size);
 111         if (maxra > MAXRA)
 112                 maxra = MAXRA;
 113         if (maxra > nbuf/8)
 114                 maxra = nbuf/8;
 115
 116         /*
 117          * get the requested block
 118          */
 119         *bpp = reqbp = bp = getblk(vp, loffset, size, 0, 0);
 120         origoffset = loffset;
 121
 122         /*
 123          * if it is in the cache, then check to see if the reads have been
 124          * sequential.  If they have, then try some read-ahead, otherwise
 125          * back-off on prospective read-aheads.
 126          */
 127         if (bp->b_flags & B_CACHE) {
 128                 if (!seqcount) {
 129                         return 0;
 130                 } else if ((bp->b_flags & B_RAM) == 0) {
 131                         return 0;
 132                 } else {
 133                         struct buf *tbp;
 134                         bp->b_flags &= ~B_RAM;
 135                         /*
 136                          * We do the crit here so that there is no window
 137                          * between the findblk and the b_usecount increment
 138                          * below.  We opt to keep the crit out of the loop
 139                          * for efficiency.
 140                          */
 141                         crit_enter();
 142                         for (i = 1; i < maxra; i++) {
 143                                 if (!(tbp = findblk(vp, loffset + i * size))) {
 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                         loffset += i * size;
 160                 }
 161                 reqbp = bp = NULL;
 162         } else {
 163                 off_t firstread = bp->b_loffset;
 164                 int nblks;
 165
 166                 KASSERT(firstread != NOOFFSET,
 167                         ("cluster_read: no buffer offset"));
 168                 if (firstread + totread > filesize)
 169                         totread = (int)(filesize - firstread);
 170                 nblks = totread / size;
 171                 if (nblks) {
 172                         int burstbytes;
 173
 174                         if (nblks > racluster)
 175                                 nblks = racluster;
 176
 177                         error = VOP_BMAP(vp, loffset, NULL,
 178                                          &doffset, &burstbytes, NULL);
 179                         if (error)
 180                                 goto single_block_read;
 181                         if (doffset == NOOFFSET)
 182                                 goto single_block_read;
 183                         if (burstbytes < size * 2)
 184                                 goto single_block_read;
 185                         if (nblks > burstbytes / size)
 186                                 nblks = burstbytes / size;
 187
 188                         bp = cluster_rbuild(vp, filesize, loffset,
 189                                             doffset, size, nblks, bp, 0);
 190                         loffset += bp->b_bufsize;
 191                 } else {
 192 single_block_read:
 193                         /*
 194                          * if it isn't in the cache, then get a chunk from
 195                          * disk if sequential, otherwise just get the block.
 196                          */
 197                         bp->b_flags |= B_RAM;
 198                         loffset += size;
 199                 }
 200         }
 201
 202         /*
 203          * Handle the synchronous read.  This only occurs if B_CACHE was
 204          * not set.  bp (and rbp) could be either a cluster bp or a normal
 205          * bp depending on the what cluster_rbuild() decided to do.  If
 206          * it is a cluster bp, vfs_busy_pages() has already been called.
 207          */
 208         if (bp) {
 209 #if defined(CLUSTERDEBUG)
 210                 if (rcluster)
 211                         printf("S(%lld,%d,%d) ",
 212                             bp->b_loffset, bp->b_bcount, seqcount);
 213 #endif
 214                 bp->b_cmd = BUF_CMD_READ;
 215                 if ((bp->b_flags & B_CLUSTER) == 0)
 216                         vfs_busy_pages(vp, bp);
 217                 bp->b_flags &= ~(B_ERROR|B_INVAL);
 218                 if ((bp->b_flags & B_ASYNC) || bp->b_bio1.bio_done != NULL)
 219                         BUF_KERNPROC(bp);
 220                 vn_strategy(vp, &bp->b_bio1);
 221                 error = bp->b_error;
 222         }
 223
 224         /*
 225          * If we have been doing sequential I/O, then do some read-ahead.
 226          */
 227         rbp = NULL;
 228         if (!error &&
 229             seqcount &&
 230             loffset < origoffset + seqcount * size &&
 231             loffset + size <= filesize
 232         ) {
 233                 int nblksread;
 234                 int ntoread;
 235                 int burstbytes;
 236
 237                 rbp = getblk(vp, loffset, size, 0, 0);
 238                 if ((rbp->b_flags & B_CACHE)) {
 239                         bqrelse(rbp);
 240                         goto no_read_ahead;
 241                 }
 242
 243                 error = VOP_BMAP(vp, loffset, NULL,
 244                                  &doffset, &burstbytes, NULL);
 245                 if (error || doffset == NOOFFSET) {
 246                         rbp->b_flags |= B_INVAL;
 247                         brelse(rbp);
 248                         rbp = NULL;
 249                         goto no_read_ahead;
 250                 }
 251                 ntoread = burstbytes / size;
 252                 nblksread = (totread + size - 1) / size;
 253                 if (seqcount < nblksread)
 254                         seqcount = nblksread;
 255                 if (seqcount < ntoread)
 256                         ntoread = seqcount;
 257
 258                 rbp->b_flags |= B_RAM;
 259                 if (burstbytes) {
 260                         rbp = cluster_rbuild(vp, filesize, loffset,
 261                                              doffset, size,
 262                                              ntoread, rbp, 1);
 263                 } else {
 264                         rbp->b_bio2.bio_offset = doffset;
 265                 }
 266 #if defined(CLUSTERDEBUG)
 267                 if (rcluster) {
 268                         if (bp)
 269                                 printf("A+(%lld,%d,%lld,%d) ",
 270                                     rbp->b_loffset, rbp->b_bcount,
 271                                     rbp->b_loffset - origoffset,
 272                                     seqcount);
 273                         else
 274                                 printf("A(%lld,%d,%lld,%d) ",
 275                                     rbp->b_loffset, rbp->b_bcount,
 276                                     rbp->b_loffset - origoffset,
 277                                     seqcount);
 278                 }
 279 #endif
 280                 rbp->b_flags &= ~(B_ERROR|B_INVAL);
 281                 rbp->b_flags |= B_ASYNC;
 282                 rbp->b_cmd = BUF_CMD_READ;
 283
 284                 if ((rbp->b_flags & B_CLUSTER) == 0)
 285                         vfs_busy_pages(vp, rbp);
 286                 BUF_KERNPROC(rbp);                      /* B_ASYNC */
 287                 vn_strategy(vp, &rbp->b_bio1);
 288         }
 289 no_read_ahead:
 290
 291         if (reqbp)
 292                 return (biowait(reqbp));
 293         else
 294                 return (error);
 295 }
 296
 297 /*
 298  * If blocks are contiguous on disk, use this to provide clustered
 299  * read ahead.  We will read as many blocks as possible sequentially
 300  * and then parcel them up into logical blocks in the buffer hash table.
 301  */
 302 static struct buf *
 303 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset,
 304         off_t doffset, int size, int run, struct buf *fbp, int doasync)
 305 {
 306         struct buf *bp, *tbp;
 307         off_t boffset;
 308         int i, j;
 309
 310         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 311             ("cluster_rbuild: size %d != filesize %ld\n",
 312             size, vp->v_mount->mnt_stat.f_iosize));
 313
 314         /*
 315          * avoid a division
 316          */
 317         while (loffset + run * size > filesize) {
 318                 --run;
 319         }
 320
 321         tbp = fbp;
 322         tbp->b_bio2.bio_offset = doffset;
 323         if((tbp->b_flags & B_MALLOC) ||
 324             ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
 325                 return tbp;
 326         }
 327
 328         bp = trypbuf(&cluster_pbuf_freecnt);
 329         if (bp == NULL)
 330                 return tbp;
 331
 332         /*
 333          * We are synthesizing a buffer out of vm_page_t's, but
 334          * if the block size is not page aligned then the starting
 335          * address may not be either.  Inherit the b_data offset
 336          * from the original buffer.
 337          */
 338         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 339             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 340         bp->b_flags |= B_ASYNC | B_CLUSTER | B_VMIO;
 341         bp->b_cmd = BUF_CMD_READ;
 342         bp->b_bio1.bio_done = cluster_callback;
 343         bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 344         bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 345         bp->b_loffset = loffset;
 346         bp->b_bio2.bio_offset = NOOFFSET;
 347         KASSERT(bp->b_loffset != NOOFFSET,
 348                 ("cluster_rbuild: no buffer offset"));
 349
 350         bp->b_bcount = 0;
 351         bp->b_bufsize = 0;
 352         bp->b_xio.xio_npages = 0;
 353
 354         for (boffset = doffset, i = 0; i < run; ++i, boffset += size) {
 355                 if (i) {
 356                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
 357                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 358                                 break;
 359                         }
 360
 361                         /*
 362                          * Shortcut some checks and try to avoid buffers that
 363                          * would block in the lock.  The same checks have to
 364                          * be made again after we officially get the buffer.
 365                          */
 366                         if ((tbp = findblk(vp, loffset + i * size)) != NULL) {
 367                                 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
 368                                         break;
 369                                 BUF_UNLOCK(tbp);
 370
 371                                 for (j = 0; j < tbp->b_xio.xio_npages; j++) {
 372                                         if (tbp->b_xio.xio_pages[j]->valid)
 373                                                 break;
 374                                 }
 375
 376                                 if (j != tbp->b_xio.xio_npages)
 377                                         break;
 378
 379                                 if (tbp->b_bcount != size)
 380                                         break;
 381                         }
 382
 383                         tbp = getblk(vp, loffset + i * size, size, 0, 0);
 384
 385                         /*
 386                          * Stop scanning if the buffer is fuly valid
 387                          * (marked B_CACHE), or locked (may be doing a
 388                          * background write), or if the buffer is not
 389                          * VMIO backed.  The clustering code can only deal
 390                          * with VMIO-backed buffers.
 391                          */
 392                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 393                             (tbp->b_flags & B_VMIO) == 0) {
 394                                 bqrelse(tbp);
 395                                 break;
 396                         }
 397
 398                         /*
 399                          * The buffer must be completely invalid in order to
 400                          * take part in the cluster.  If it is partially valid
 401                          * then we stop.
 402                          */
 403                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
 404                                 if (tbp->b_xio.xio_pages[j]->valid)
 405                                         break;
 406                         }
 407                         if (j != tbp->b_xio.xio_npages) {
 408                                 bqrelse(tbp);
 409                                 break;
 410                         }
 411
 412                         /*
 413                          * Set a read-ahead mark as appropriate
 414                          */
 415                         if (i == 1 || i == (run - 1))
 416                                 tbp->b_flags |= B_RAM;
 417
 418                         /*
 419                          * Set the block number if it isn't set, otherwise
 420                          * if it is make sure it matches the block number we
 421                          * expect.
 422                          */
 423                         if (tbp->b_bio2.bio_offset == NOOFFSET) {
 424                                 tbp->b_bio2.bio_offset = boffset;
 425                         } else if (tbp->b_bio2.bio_offset != boffset) {
 426                                 brelse(tbp);
 427                                 break;
 428                         }
 429                 }
 430                 /*
 431                  * The first buffer is setup async if doasync is specified.
 432                  * All other buffers in the cluster are setup async.  This
 433                  * way the caller can decide how to deal with the requested
 434                  * buffer.
 435                  */
 436                 if (i || doasync)
 437                         tbp->b_flags |= B_ASYNC;
 438                 tbp->b_cmd = BUF_CMD_READ;
 439                 BUF_KERNPROC(tbp);
 440                 cluster_append(&bp->b_bio1, tbp);
 441                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 442                         vm_page_t m;
 443                         m = tbp->b_xio.xio_pages[j];
 444                         vm_page_io_start(m);
 445                         vm_object_pip_add(m->object, 1);
 446                         if ((bp->b_xio.xio_npages == 0) ||
 447                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
 448                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 449                                 bp->b_xio.xio_npages++;
 450                         }
 451                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 452                                 tbp->b_xio.xio_pages[j] = bogus_page;
 453                 }
 454                 /*
 455                  * XXX shouldn't this be += size for both, like in
 456                  * cluster_wbuild()?
 457                  *
 458                  * Don't inherit tbp->b_bufsize as it may be larger due to
 459                  * a non-page-aligned size.  Instead just aggregate using
 460                  * 'size'.
 461                  */
 462                 if (tbp->b_bcount != size)
 463                     printf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, size);
 464                 if (tbp->b_bufsize != size)
 465                     printf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, size);
 466                 bp->b_bcount += size;
 467                 bp->b_bufsize += size;
 468         }
 469
 470         /*
 471          * Fully valid pages in the cluster are already good and do not need
 472          * to be re-read from disk.  Replace the page with bogus_page
 473          */
 474         for (j = 0; j < bp->b_xio.xio_npages; j++) {
 475                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 476                     VM_PAGE_BITS_ALL) {
 477                         bp->b_xio.xio_pages[j] = bogus_page;
 478                 }
 479         }
 480         if (bp->b_bufsize > bp->b_kvasize) {
 481                 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
 482                     bp->b_bufsize, bp->b_kvasize);
 483         }
 484
 485         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 486                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 487         return (bp);
 488 }
 489
 490 /*
 491  * Cleanup after a clustered read or write.
 492  * This is complicated by the fact that any of the buffers might have
 493  * extra memory (if there were no empty buffer headers at allocbuf time)
 494  * that we will need to shift around.
 495  *
 496  * The returned bio is &bp->b_bio1
 497  */
 498 void
 499 cluster_callback(struct bio *bio)
 500 {
 501         struct buf *bp = bio->bio_buf;
 502         struct buf *tbp;
 503         int error = 0;
 504
 505         /*
 506          * Must propogate errors to all the components.  A short read (EOF)
 507          * is a critical error.
 508          */
 509         if (bp->b_flags & B_ERROR) {
 510                 error = bp->b_error;
 511         } else if (bp->b_bcount != bp->b_bufsize) {
 512                 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
 513         }
 514
 515         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
 516         /*
 517          * Move memory from the large cluster buffer into the component
 518          * buffers and mark IO as done on these.  Since the memory map
 519          * is the same, no actual copying is required.
 520          */
 521         while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
 522                 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
 523                 if (error) {
 524                         tbp->b_flags |= B_ERROR;
 525                         tbp->b_error = error;
 526                 } else {
 527                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 528                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 529                         /*
 530                          * XXX the bdwrite()/bqrelse() issued during
 531                          * cluster building clears B_RELBUF (see bqrelse()
 532                          * comment).  If direct I/O was specified, we have
 533                          * to restore it here to allow the buffer and VM
 534                          * to be freed.
 535                          */
 536                         if (tbp->b_flags & B_DIRECT)
 537                                 tbp->b_flags |= B_RELBUF;
 538                 }
 539                 biodone(&tbp->b_bio1);
 540         }
 541         relpbuf(bp, &cluster_pbuf_freecnt);
 542 }
 543
 544 /*
 545  *      cluster_wbuild_wb:
 546  *
 547  *      Implement modified write build for cluster.
 548  *
 549  *              write_behind = 0        write behind disabled
 550  *              write_behind = 1        write behind normal (default)
 551  *              write_behind = 2        write behind backed-off
 552  */
 553
 554 static __inline int
 555 cluster_wbuild_wb(struct vnode *vp, int size, off_t start_loffset, int len)
 556 {
 557         int r = 0;
 558
 559         switch(write_behind) {
 560         case 2:
 561                 if (start_loffset < len)
 562                         break;
 563                 start_loffset -= len;
 564                 /* fall through */
 565         case 1:
 566                 r = cluster_wbuild(vp, size, start_loffset, len);
 567                 /* fall through */
 568         default:
 569                 /* fall through */
 570                 break;
 571         }
 572         return(r);
 573 }
 574
 575 /*
 576  * Do clustered write for FFS.
 577  *
 578  * Three cases:
 579  *      1. Write is not sequential (write asynchronously)
 580  *      Write is sequential:
 581  *      2.      beginning of cluster - begin cluster
 582  *      3.      middle of a cluster - add to cluster
 583  *      4.      end of a cluster - asynchronously write cluster
 584  */
 585 void
 586 cluster_write(struct buf *bp, off_t filesize, int seqcount)
 587 {
 588         struct vnode *vp;
 589         off_t loffset;
 590         int maxclen, cursize;
 591         int lblocksize;
 592         int async;
 593
 594         vp = bp->b_vp;
 595         if (vp->v_type == VREG) {
 596                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 597                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 598         } else {
 599                 async = 0;
 600                 lblocksize = bp->b_bufsize;
 601         }
 602         loffset = bp->b_loffset;
 603         KASSERT(bp->b_loffset != NOOFFSET,
 604                 ("cluster_write: no buffer offset"));
 605
 606         /* Initialize vnode to beginning of file. */
 607         if (loffset == 0)
 608                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 609
 610         if (vp->v_clen == 0 || loffset != vp->v_lastw + lblocksize ||
 611             bp->b_bio2.bio_offset == NOOFFSET ||
 612             (bp->b_bio2.bio_offset != vp->v_lasta + lblocksize)) {
 613                 maxclen = vp->v_mount->mnt_iosize_max;
 614                 if (vp->v_clen != 0) {
 615                         /*
 616                          * Next block is not sequential.
 617                          *
 618                          * If we are not writing at end of file, the process
 619                          * seeked to another point in the file since its last
 620                          * write, or we have reached our maximum cluster size,
 621                          * then push the previous cluster. Otherwise try
 622                          * reallocating to make it sequential.
 623                          *
 624                          * Change to algorithm: only push previous cluster if
 625                          * it was sequential from the point of view of the
 626                          * seqcount heuristic, otherwise leave the buffer
 627                          * intact so we can potentially optimize the I/O
 628                          * later on in the buf_daemon or update daemon
 629                          * flush.
 630                          */
 631                         cursize = vp->v_lastw - vp->v_cstart + lblocksize;
 632                         if (bp->b_loffset + lblocksize != filesize ||
 633                             loffset != vp->v_lastw + lblocksize || vp->v_clen <= cursize) {
 634                                 if (!async && seqcount > 0) {
 635                                         cluster_wbuild_wb(vp, lblocksize,
 636                                                 vp->v_cstart, cursize);
 637                                 }
 638                         } else {
 639                                 struct buf **bpp, **endbp;
 640                                 struct cluster_save *buflist;
 641
 642                                 buflist = cluster_collectbufs(vp, bp,
 643                                                               lblocksize);
 644                                 endbp = &buflist->bs_children
 645                                     [buflist->bs_nchildren - 1];
 646                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 647                                         /*
 648                                          * Failed, push the previous cluster
 649                                          * if *really* writing sequentially
 650                                          * in the logical file (seqcount > 1),
 651                                          * otherwise delay it in the hopes that
 652                                          * the low level disk driver can
 653                                          * optimize the write ordering.
 654                                          */
 655                                         for (bpp = buflist->bs_children;
 656                                              bpp < endbp; bpp++)
 657                                                 brelse(*bpp);
 658                                         kfree(buflist, M_SEGMENT);
 659                                         if (seqcount > 1) {
 660                                                 cluster_wbuild_wb(vp,
 661                                                     lblocksize, vp->v_cstart,
 662                                                     cursize);
 663                                         }
 664                                 } else {
 665                                         /*
 666                                          * Succeeded, keep building cluster.
 667                                          */
 668                                         for (bpp = buflist->bs_children;
 669                                              bpp <= endbp; bpp++)
 670                                                 bdwrite(*bpp);
 671                                         kfree(buflist, M_SEGMENT);
 672                                         vp->v_lastw = loffset;
 673                                         vp->v_lasta = bp->b_bio2.bio_offset;
 674                                         return;
 675                                 }
 676                         }
 677                 }
 678                 /*
 679                  * Consider beginning a cluster. If at end of file, make
 680                  * cluster as large as possible, otherwise find size of
 681                  * existing cluster.
 682                  */
 683                 if ((vp->v_type == VREG) &&
 684                     bp->b_loffset + lblocksize != filesize &&
 685                     (bp->b_bio2.bio_offset == NOOFFSET) &&
 686                     (VOP_BMAP(vp, loffset, NULL, &bp->b_bio2.bio_offset, &maxclen, NULL) ||
 687                      bp->b_bio2.bio_offset == NOOFFSET)) {
 688                         bawrite(bp);
 689                         vp->v_clen = 0;
 690                         vp->v_lasta = bp->b_bio2.bio_offset;
 691                         vp->v_cstart = loffset + lblocksize;
 692                         vp->v_lastw = loffset;
 693                         return;
 694                 }
 695                 if (maxclen > lblocksize)
 696                         vp->v_clen = maxclen - lblocksize;
 697                 else
 698                         vp->v_clen = 0;
 699                 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
 700                         vp->v_cstart = loffset + lblocksize;
 701                         bawrite(bp);
 702                 } else {        /* Wait for rest of cluster */
 703                         vp->v_cstart = loffset;
 704                         bdwrite(bp);
 705                 }
 706         } else if (loffset == vp->v_cstart + vp->v_clen) {
 707                 /*
 708                  * At end of cluster, write it out if seqcount tells us we
 709                  * are operating sequentially, otherwise let the buf or
 710                  * update daemon handle it.
 711                  */
 712                 bdwrite(bp);
 713                 if (seqcount > 1)
 714                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
 715                                           vp->v_clen + lblocksize);
 716                 vp->v_clen = 0;
 717                 vp->v_cstart = loffset + lblocksize;
 718         } else if (vm_page_count_severe()) {
 719                 /*
 720                  * We are low on memory, get it going NOW
 721                  */
 722                 bawrite(bp);
 723         } else {
 724                 /*
 725                  * In the middle of a cluster, so just delay the I/O for now.
 726                  */
 727                 bdwrite(bp);
 728         }
 729         vp->v_lastw = loffset;
 730         vp->v_lasta = bp->b_bio2.bio_offset;
 731 }
 732
 733
 734 /*
 735  * This is an awful lot like cluster_rbuild...wish they could be combined.
 736  * The last lbn argument is the current block on which I/O is being
 737  * performed.  Check to see that it doesn't fall in the middle of
 738  * the current block (if last_bp == NULL).
 739  */
 740 int
 741 cluster_wbuild(struct vnode *vp, int size, off_t start_loffset, int bytes)
 742 {
 743         struct buf *bp, *tbp;
 744         int i, j;
 745         int totalwritten = 0;
 746
 747         while (bytes > 0) {
 748                 crit_enter();
 749                 /*
 750                  * If the buffer is not delayed-write (i.e. dirty), or it
 751                  * is delayed-write but either locked or inval, it cannot
 752                  * partake in the clustered write.
 753                  */
 754                 if (((tbp = findblk(vp, start_loffset)) == NULL) ||
 755                   ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
 756                   BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 757                         start_loffset += size;
 758                         bytes -= size;
 759                         crit_exit();
 760                         continue;
 761                 }
 762                 bremfree(tbp);
 763                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 764                 crit_exit();
 765
 766                 /*
 767                  * Extra memory in the buffer, punt on this buffer.
 768                  * XXX we could handle this in most cases, but we would
 769                  * have to push the extra memory down to after our max
 770                  * possible cluster size and then potentially pull it back
 771                  * up if the cluster was terminated prematurely--too much
 772                  * hassle.
 773                  */
 774                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 775                   (tbp->b_bcount != tbp->b_bufsize) ||
 776                   (tbp->b_bcount != size) ||
 777                   (bytes == size) ||
 778                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 779                         totalwritten += tbp->b_bufsize;
 780                         bawrite(tbp);
 781                         start_loffset += size;
 782                         bytes -= size;
 783                         continue;
 784                 }
 785
 786                 /*
 787                  * Set up the pbuf.  Track our append point with b_bcount
 788                  * and b_bufsize.  b_bufsize is not used by the device but
 789                  * our caller uses it to loop clusters and we use it to
 790                  * detect a premature EOF on the block device.
 791                  */
 792                 bp->b_bcount = 0;
 793                 bp->b_bufsize = 0;
 794                 bp->b_xio.xio_npages = 0;
 795                 bp->b_loffset = tbp->b_loffset;
 796                 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
 797
 798                 /*
 799                  * We are synthesizing a buffer out of vm_page_t's, but
 800                  * if the block size is not page aligned then the starting
 801                  * address may not be either.  Inherit the b_data offset
 802                  * from the original buffer.
 803                  */
 804                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 805                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 806                 bp->b_flags &= ~B_ERROR;
 807                 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
 808                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN));
 809                 bp->b_bio1.bio_done = cluster_callback;
 810                 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 811                 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 812                 /*
 813                  * From this location in the file, scan forward to see
 814                  * if there are buffers with adjacent data that need to
 815                  * be written as well.
 816                  */
 817                 for (i = 0; i < bytes; (i += size), (start_loffset += size)) {
 818                         if (i != 0) { /* If not the first buffer */
 819                                 crit_enter();
 820                                 /*
 821                                  * If the adjacent data is not even in core it
 822                                  * can't need to be written.
 823                                  */
 824                                 if ((tbp = findblk(vp, start_loffset)) == NULL) {
 825                                         crit_exit();
 826                                         break;
 827                                 }
 828
 829                                 /*
 830                                  * If it IS in core, but has different
 831                                  * characteristics, or is locked (which
 832                                  * means it could be undergoing a background
 833                                  * I/O or be in a weird state), then don't
 834                                  * cluster with it.
 835                                  */
 836                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 837                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 838                                   != (B_DELWRI | B_CLUSTEROK |
 839                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 840                                     (tbp->b_flags & B_LOCKED) ||
 841                                     BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 842                                         crit_exit();
 843                                         break;
 844                                 }
 845
 846                                 /*
 847                                  * Check that the combined cluster
 848                                  * would make sense with regard to pages
 849                                  * and would not be too large
 850                                  */
 851                                 if ((tbp->b_bcount != size) ||
 852                                   ((bp->b_bio2.bio_offset + i) !=
 853                                     tbp->b_bio2.bio_offset) ||
 854                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
 855                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 856                                         BUF_UNLOCK(tbp);
 857                                         crit_exit();
 858                                         break;
 859                                 }
 860                                 /*
 861                                  * Ok, it's passed all the tests,
 862                                  * so remove it from the free list
 863                                  * and mark it busy. We will use it.
 864                                  */
 865                                 bremfree(tbp);
 866                                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 867                                 crit_exit();
 868                         } /* end of code for non-first buffers only */
 869
 870                         /*
 871                          * If the IO is via the VM then we do some
 872                          * special VM hackery (yuck).  Since the buffer's
 873                          * block size may not be page-aligned it is possible
 874                          * for a page to be shared between two buffers.  We
 875                          * have to get rid of the duplication when building
 876                          * the cluster.
 877                          */
 878                         if (tbp->b_flags & B_VMIO) {
 879                                 vm_page_t m;
 880
 881                                 if (i != 0) { /* if not first buffer */
 882                                         for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 883                                                 m = tbp->b_xio.xio_pages[j];
 884                                                 if (m->flags & PG_BUSY) {
 885                                                         bqrelse(tbp);
 886                                                         goto finishcluster;
 887                                                 }
 888                                         }
 889                                 }
 890
 891                                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 892                                         m = tbp->b_xio.xio_pages[j];
 893                                         vm_page_io_start(m);
 894                                         vm_object_pip_add(m->object, 1);
 895                                         if ((bp->b_xio.xio_npages == 0) ||
 896                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
 897                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 898                                                 bp->b_xio.xio_npages++;
 899                                         }
 900                                 }
 901                         }
 902                         bp->b_bcount += size;
 903                         bp->b_bufsize += size;
 904
 905                         crit_enter();
 906                         bundirty(tbp);
 907                         tbp->b_flags &= ~B_ERROR;
 908                         tbp->b_flags |= B_ASYNC;
 909                         tbp->b_cmd = BUF_CMD_WRITE;
 910                         crit_exit();
 911                         BUF_KERNPROC(tbp);
 912                         cluster_append(&bp->b_bio1, tbp);
 913
 914                         /*
 915                          * check for latent dependencies to be handled
 916                          */
 917                         if (LIST_FIRST(&tbp->b_dep) != NULL && bioops.io_start)
 918                                 (*bioops.io_start)(tbp);
 919
 920                 }
 921         finishcluster:
 922                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 923                         (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 924                 if (bp->b_bufsize > bp->b_kvasize) {
 925                         panic(
 926                             "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
 927                             bp->b_bufsize, bp->b_kvasize);
 928                 }
 929                 totalwritten += bp->b_bufsize;
 930                 bp->b_dirtyoff = 0;
 931                 bp->b_dirtyend = bp->b_bufsize;
 932                 bp->b_flags |= B_ASYNC;
 933                 bp->b_cmd = BUF_CMD_WRITE;
 934                 vfs_busy_pages(vp, bp);
 935                 bp->b_runningbufspace = bp->b_bufsize;
 936                 runningbufspace += bp->b_runningbufspace;
 937                 BUF_KERNPROC(bp);       /* B_ASYNC */
 938                 vn_strategy(vp, &bp->b_bio1);
 939
 940                 bytes -= i;
 941         }
 942         return totalwritten;
 943 }
 944
 945 /*
 946  * Collect together all the buffers in a cluster.
 947  * Plus add one additional buffer.
 948  */
 949 static struct cluster_save *
 950 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int lblocksize)
 951 {
 952         struct cluster_save *buflist;
 953         struct buf *bp;
 954         off_t loffset;
 955         int i, len;
 956
 957         len = (int)(vp->v_lastw - vp->v_cstart + lblocksize) / lblocksize;
 958         buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
 959                          M_SEGMENT, M_WAITOK);
 960         buflist->bs_nchildren = 0;
 961         buflist->bs_children = (struct buf **) (buflist + 1);
 962         for (loffset = vp->v_cstart, i = 0; i < len; (loffset += lblocksize), i++) {
 963                 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
 964                 buflist->bs_children[i] = bp;
 965                 if (bp->b_bio2.bio_offset == NOOFFSET) {
 966                         VOP_BMAP(bp->b_vp, bp->b_loffset, NULL,
 967                                 &bp->b_bio2.bio_offset, NULL, NULL);
 968                 }
 969         }
 970         buflist->bs_children[i] = bp = last_bp;
 971         if (bp->b_bio2.bio_offset == NOOFFSET) {
 972                 VOP_BMAP(bp->b_vp, bp->b_loffset, NULL,
 973                          &bp->b_bio2.bio_offset, NULL, NULL);
 974         }
 975         buflist->bs_nchildren = i + 1;
 976         return (buflist);
 977 }
 978
 979 void
 980 cluster_append(struct bio *bio, struct buf *tbp)
 981 {
 982         tbp->b_cluster_next = NULL;
 983         if (bio->bio_caller_info1.cluster_head == NULL) {
 984                 bio->bio_caller_info1.cluster_head = tbp;
 985                 bio->bio_caller_info2.cluster_tail = tbp;
 986         } else {
 987                 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
 988                 bio->bio_caller_info2.cluster_tail = tbp;
 989         }
 990 }
 991