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 1.25 2006/05/04 18:32:22 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          * If we have been doing sequential I/O, then do some read-ahead.
 204          */
 205         rbp = NULL;
 206         if (seqcount &&
 207             loffset < origoffset + seqcount * size &&
 208             loffset + size <= filesize
 209         ) {
 210                 rbp = getblk(vp, loffset, size, 0, 0);
 211                 if ((rbp->b_flags & B_CACHE) == 0) {
 212                         int nblksread;
 213                         int ntoread;
 214                         int burstbytes;
 215
 216                         error = VOP_BMAP(vp, loffset, NULL,
 217                                          &doffset, &burstbytes, NULL);
 218                         if (error || doffset == NOOFFSET) {
 219                                 brelse(rbp);
 220                                 rbp = NULL;
 221                                 goto no_read_ahead;
 222                         }
 223                         ntoread = burstbytes / size;
 224                         nblksread = (totread + size - 1) / size;
 225                         if (seqcount < nblksread)
 226                                 seqcount = nblksread;
 227                         if (seqcount < ntoread)
 228                                 ntoread = seqcount;
 229
 230                         rbp->b_flags |= B_RAM;
 231                         if (burstbytes) {
 232                                 rbp = cluster_rbuild(vp, filesize, loffset,
 233                                                      doffset, size,
 234                                                      ntoread, rbp, 1);
 235                         } else {
 236                                 rbp->b_bio2.bio_offset = doffset;
 237                         }
 238                 }
 239         }
 240 no_read_ahead:
 241
 242         /*
 243          * Handle the synchronous read.  This only occurs if B_CACHE was
 244          * not set.  bp (and rbp) could be either a cluster bp or a normal
 245          * bp depending on the what cluster_rbuild() decided to do.  If
 246          * it is a cluster bp, vfs_busy_pages() has already been called.
 247          */
 248         if (bp) {
 249 #if defined(CLUSTERDEBUG)
 250                 if (rcluster)
 251                         printf("S(%lld,%d,%d) ",
 252                             bp->b_loffset, bp->b_bcount, seqcount);
 253 #endif
 254                 bp->b_cmd = BUF_CMD_READ;
 255                 if ((bp->b_flags & B_CLUSTER) == 0)
 256                         vfs_busy_pages(vp, bp);
 257                 bp->b_flags &= ~(B_ERROR|B_INVAL);
 258                 if ((bp->b_flags & B_ASYNC) || bp->b_bio1.bio_done != NULL)
 259                         BUF_KERNPROC(bp);
 260                 vn_strategy(vp, &bp->b_bio1);
 261                 error = bp->b_error;
 262         }
 263
 264         /*
 265          * And if we have read-aheads, do them too.
 266          */
 267         if (rbp) {
 268                 if (error) {
 269                         brelse(rbp);
 270                 } else if (rbp->b_flags & B_CACHE) {
 271                         bqrelse(rbp);
 272                 } else {
 273 #if defined(CLUSTERDEBUG)
 274                         if (rcluster) {
 275                                 if (bp)
 276                                         printf("A+(%lld,%d,%lld,%d) ",
 277                                             rbp->b_loffset, rbp->b_bcount,
 278                                             rbp->b_loffset - origoffset,
 279                                             seqcount);
 280                                 else
 281                                         printf("A(%lld,%d,%lld,%d) ",
 282                                             rbp->b_loffset, rbp->b_bcount,
 283                                             rbp->b_loffset - origoffset,
 284                                             seqcount);
 285                         }
 286 #endif
 287                         rbp->b_flags &= ~(B_ERROR|B_INVAL);
 288                         rbp->b_flags |= B_ASYNC;
 289                         rbp->b_cmd = BUF_CMD_READ;
 290
 291                         if ((rbp->b_flags & B_CLUSTER) == 0)
 292                                 vfs_busy_pages(vp, rbp);
 293                         if ((rbp->b_flags & B_ASYNC) || rbp->b_bio1.bio_done != NULL)
 294                                 BUF_KERNPROC(rbp);
 295                         vn_strategy(vp, &rbp->b_bio1);
 296                 }
 297         }
 298         if (reqbp)
 299                 return (biowait(reqbp));
 300         else
 301                 return (error);
 302 }
 303
 304 /*
 305  * If blocks are contiguous on disk, use this to provide clustered
 306  * read ahead.  We will read as many blocks as possible sequentially
 307  * and then parcel them up into logical blocks in the buffer hash table.
 308  */
 309 static struct buf *
 310 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset,
 311         off_t doffset, int size, int run, struct buf *fbp, int doasync)
 312 {
 313         struct buf *bp, *tbp;
 314         off_t boffset;
 315         int i, j;
 316
 317         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 318             ("cluster_rbuild: size %d != filesize %ld\n",
 319             size, vp->v_mount->mnt_stat.f_iosize));
 320
 321         /*
 322          * avoid a division
 323          */
 324         while (loffset + run * size > filesize) {
 325                 --run;
 326         }
 327
 328         tbp = fbp;
 329         tbp->b_bio2.bio_offset = doffset;
 330         if((tbp->b_flags & B_MALLOC) ||
 331             ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
 332                 return tbp;
 333         }
 334
 335         bp = trypbuf(&cluster_pbuf_freecnt);
 336         if (bp == NULL)
 337                 return tbp;
 338
 339         /*
 340          * We are synthesizing a buffer out of vm_page_t's, but
 341          * if the block size is not page aligned then the starting
 342          * address may not be either.  Inherit the b_data offset
 343          * from the original buffer.
 344          */
 345         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 346             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 347         bp->b_flags |= B_ASYNC | B_CLUSTER | B_VMIO;
 348         bp->b_cmd = BUF_CMD_READ;
 349         bp->b_bio1.bio_done = cluster_callback;
 350         bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 351         bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 352         bp->b_loffset = loffset;
 353         bp->b_bio2.bio_offset = NOOFFSET;
 354         KASSERT(bp->b_loffset != NOOFFSET,
 355                 ("cluster_rbuild: no buffer offset"));
 356
 357         bp->b_bcount = 0;
 358         bp->b_bufsize = 0;
 359         bp->b_xio.xio_npages = 0;
 360
 361         for (boffset = doffset, i = 0; i < run; ++i, boffset += size) {
 362                 if (i) {
 363                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
 364                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 365                                 break;
 366                         }
 367
 368                         /*
 369                          * Shortcut some checks and try to avoid buffers that
 370                          * would block in the lock.  The same checks have to
 371                          * be made again after we officially get the buffer.
 372                          */
 373                         if ((tbp = findblk(vp, loffset + i * size)) != NULL) {
 374                                 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
 375                                         break;
 376                                 BUF_UNLOCK(tbp);
 377
 378                                 for (j = 0; j < tbp->b_xio.xio_npages; j++) {
 379                                         if (tbp->b_xio.xio_pages[j]->valid)
 380                                                 break;
 381                                 }
 382
 383                                 if (j != tbp->b_xio.xio_npages)
 384                                         break;
 385
 386                                 if (tbp->b_bcount != size)
 387                                         break;
 388                         }
 389
 390                         tbp = getblk(vp, loffset + i * size, size, 0, 0);
 391
 392                         /*
 393                          * Stop scanning if the buffer is fuly valid
 394                          * (marked B_CACHE), or locked (may be doing a
 395                          * background write), or if the buffer is not
 396                          * VMIO backed.  The clustering code can only deal
 397                          * with VMIO-backed buffers.
 398                          */
 399                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 400                             (tbp->b_flags & B_VMIO) == 0) {
 401                                 bqrelse(tbp);
 402                                 break;
 403                         }
 404
 405                         /*
 406                          * The buffer must be completely invalid in order to
 407                          * take part in the cluster.  If it is partially valid
 408                          * then we stop.
 409                          */
 410                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
 411                                 if (tbp->b_xio.xio_pages[j]->valid)
 412                                         break;
 413                         }
 414                         if (j != tbp->b_xio.xio_npages) {
 415                                 bqrelse(tbp);
 416                                 break;
 417                         }
 418
 419                         /*
 420                          * Set a read-ahead mark as appropriate
 421                          */
 422                         if (i == 1 || i == (run - 1))
 423                                 tbp->b_flags |= B_RAM;
 424
 425                         /*
 426                          * Set the block number if it isn't set, otherwise
 427                          * if it is make sure it matches the block number we
 428                          * expect.
 429                          */
 430                         if (tbp->b_bio2.bio_offset == NOOFFSET) {
 431                                 tbp->b_bio2.bio_offset = boffset;
 432                         } else if (tbp->b_bio2.bio_offset != boffset) {
 433                                 brelse(tbp);
 434                                 break;
 435                         }
 436                 }
 437                 /*
 438                  * The first buffer is setup async if doasync is specified.
 439                  * All other buffers in the cluster are setup async.  This
 440                  * way the caller can decide how to deal with the requested
 441                  * buffer.
 442                  */
 443                 if (i || doasync)
 444                         tbp->b_flags |= B_ASYNC;
 445                 tbp->b_cmd = BUF_CMD_READ;
 446                 BUF_KERNPROC(tbp);
 447                 cluster_append(&bp->b_bio1, tbp);
 448                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 449                         vm_page_t m;
 450                         m = tbp->b_xio.xio_pages[j];
 451                         vm_page_io_start(m);
 452                         vm_object_pip_add(m->object, 1);
 453                         if ((bp->b_xio.xio_npages == 0) ||
 454                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
 455                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 456                                 bp->b_xio.xio_npages++;
 457                         }
 458                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 459                                 tbp->b_xio.xio_pages[j] = bogus_page;
 460                 }
 461                 /*
 462                  * XXX shouldn't this be += size for both, like in
 463                  * cluster_wbuild()?
 464                  *
 465                  * Don't inherit tbp->b_bufsize as it may be larger due to
 466                  * a non-page-aligned size.  Instead just aggregate using
 467                  * 'size'.
 468                  */
 469                 if (tbp->b_bcount != size)
 470                     printf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, size);
 471                 if (tbp->b_bufsize != size)
 472                     printf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, size);
 473                 bp->b_bcount += size;
 474                 bp->b_bufsize += size;
 475         }
 476
 477         /*
 478          * Fully valid pages in the cluster are already good and do not need
 479          * to be re-read from disk.  Replace the page with bogus_page
 480          */
 481         for (j = 0; j < bp->b_xio.xio_npages; j++) {
 482                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 483                     VM_PAGE_BITS_ALL) {
 484                         bp->b_xio.xio_pages[j] = bogus_page;
 485                 }
 486         }
 487         if (bp->b_bufsize > bp->b_kvasize) {
 488                 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
 489                     bp->b_bufsize, bp->b_kvasize);
 490         }
 491
 492         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 493                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 494         return (bp);
 495 }
 496
 497 /*
 498  * Cleanup after a clustered read or write.
 499  * This is complicated by the fact that any of the buffers might have
 500  * extra memory (if there were no empty buffer headers at allocbuf time)
 501  * that we will need to shift around.
 502  *
 503  * The returned bio is &bp->b_bio1
 504  */
 505 void
 506 cluster_callback(struct bio *bio)
 507 {
 508         struct buf *bp = bio->bio_buf;
 509         struct buf *tbp;
 510         int error = 0;
 511
 512         /*
 513          * Must propogate errors to all the components.  A short read (EOF)
 514          * is a critical error.
 515          */
 516         if (bp->b_flags & B_ERROR) {
 517                 error = bp->b_error;
 518         } else if (bp->b_bcount != bp->b_bufsize) {
 519                 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
 520         }
 521
 522         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
 523         /*
 524          * Move memory from the large cluster buffer into the component
 525          * buffers and mark IO as done on these.  Since the memory map
 526          * is the same, no actual copying is required.
 527          */
 528         while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
 529                 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
 530                 if (error) {
 531                         tbp->b_flags |= B_ERROR;
 532                         tbp->b_error = error;
 533                 } else {
 534                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 535                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 536                         /*
 537                          * XXX the bdwrite()/bqrelse() issued during
 538                          * cluster building clears B_RELBUF (see bqrelse()
 539                          * comment).  If direct I/O was specified, we have
 540                          * to restore it here to allow the buffer and VM
 541                          * to be freed.
 542                          */
 543                         if (tbp->b_flags & B_DIRECT)
 544                                 tbp->b_flags |= B_RELBUF;
 545                 }
 546                 biodone(&tbp->b_bio1);
 547         }
 548         relpbuf(bp, &cluster_pbuf_freecnt);
 549 }
 550
 551 /*
 552  *      cluster_wbuild_wb:
 553  *
 554  *      Implement modified write build for cluster.
 555  *
 556  *              write_behind = 0        write behind disabled
 557  *              write_behind = 1        write behind normal (default)
 558  *              write_behind = 2        write behind backed-off
 559  */
 560
 561 static __inline int
 562 cluster_wbuild_wb(struct vnode *vp, int size, off_t start_loffset, int len)
 563 {
 564         int r = 0;
 565
 566         switch(write_behind) {
 567         case 2:
 568                 if (start_loffset < len)
 569                         break;
 570                 start_loffset -= len;
 571                 /* fall through */
 572         case 1:
 573                 r = cluster_wbuild(vp, size, start_loffset, len);
 574                 /* fall through */
 575         default:
 576                 /* fall through */
 577                 break;
 578         }
 579         return(r);
 580 }
 581
 582 /*
 583  * Do clustered write for FFS.
 584  *
 585  * Three cases:
 586  *      1. Write is not sequential (write asynchronously)
 587  *      Write is sequential:
 588  *      2.      beginning of cluster - begin cluster
 589  *      3.      middle of a cluster - add to cluster
 590  *      4.      end of a cluster - asynchronously write cluster
 591  */
 592 void
 593 cluster_write(struct buf *bp, off_t filesize, int seqcount)
 594 {
 595         struct vnode *vp;
 596         off_t loffset;
 597         int maxclen, cursize;
 598         int lblocksize;
 599         int async;
 600
 601         vp = bp->b_vp;
 602         if (vp->v_type == VREG) {
 603                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 604                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 605         } else {
 606                 async = 0;
 607                 lblocksize = bp->b_bufsize;
 608         }
 609         loffset = bp->b_loffset;
 610         KASSERT(bp->b_loffset != NOOFFSET,
 611                 ("cluster_write: no buffer offset"));
 612
 613         /* Initialize vnode to beginning of file. */
 614         if (loffset == 0)
 615                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 616
 617         if (vp->v_clen == 0 || loffset != vp->v_lastw + lblocksize ||
 618             bp->b_bio2.bio_offset == NOOFFSET ||
 619             (bp->b_bio2.bio_offset != vp->v_lasta + lblocksize)) {
 620                 maxclen = vp->v_mount->mnt_iosize_max;
 621                 if (vp->v_clen != 0) {
 622                         /*
 623                          * Next block is not sequential.
 624                          *
 625                          * If we are not writing at end of file, the process
 626                          * seeked to another point in the file since its last
 627                          * write, or we have reached our maximum cluster size,
 628                          * then push the previous cluster. Otherwise try
 629                          * reallocating to make it sequential.
 630                          *
 631                          * Change to algorithm: only push previous cluster if
 632                          * it was sequential from the point of view of the
 633                          * seqcount heuristic, otherwise leave the buffer
 634                          * intact so we can potentially optimize the I/O
 635                          * later on in the buf_daemon or update daemon
 636                          * flush.
 637                          */
 638                         cursize = vp->v_lastw - vp->v_cstart + lblocksize;
 639                         if (bp->b_loffset + lblocksize != filesize ||
 640                             loffset != vp->v_lastw + lblocksize || vp->v_clen <= cursize) {
 641                                 if (!async && seqcount > 0) {
 642                                         cluster_wbuild_wb(vp, lblocksize,
 643                                                 vp->v_cstart, cursize);
 644                                 }
 645                         } else {
 646                                 struct buf **bpp, **endbp;
 647                                 struct cluster_save *buflist;
 648
 649                                 buflist = cluster_collectbufs(vp, bp,
 650                                                               lblocksize);
 651                                 endbp = &buflist->bs_children
 652                                     [buflist->bs_nchildren - 1];
 653                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 654                                         /*
 655                                          * Failed, push the previous cluster
 656                                          * if *really* writing sequentially
 657                                          * in the logical file (seqcount > 1),
 658                                          * otherwise delay it in the hopes that
 659                                          * the low level disk driver can
 660                                          * optimize the write ordering.
 661                                          */
 662                                         for (bpp = buflist->bs_children;
 663                                              bpp < endbp; bpp++)
 664                                                 brelse(*bpp);
 665                                         free(buflist, M_SEGMENT);
 666                                         if (seqcount > 1) {
 667                                                 cluster_wbuild_wb(vp,
 668                                                     lblocksize, vp->v_cstart,
 669                                                     cursize);
 670                                         }
 671                                 } else {
 672                                         /*
 673                                          * Succeeded, keep building cluster.
 674                                          */
 675                                         for (bpp = buflist->bs_children;
 676                                              bpp <= endbp; bpp++)
 677                                                 bdwrite(*bpp);
 678                                         free(buflist, M_SEGMENT);
 679                                         vp->v_lastw = loffset;
 680                                         vp->v_lasta = bp->b_bio2.bio_offset;
 681                                         return;
 682                                 }
 683                         }
 684                 }
 685                 /*
 686                  * Consider beginning a cluster. If at end of file, make
 687                  * cluster as large as possible, otherwise find size of
 688                  * existing cluster.
 689                  */
 690                 if ((vp->v_type == VREG) &&
 691                     bp->b_loffset + lblocksize != filesize &&
 692                     (bp->b_bio2.bio_offset == NOOFFSET) &&
 693                     (VOP_BMAP(vp, loffset, NULL, &bp->b_bio2.bio_offset, &maxclen, NULL) ||
 694                      bp->b_bio2.bio_offset == NOOFFSET)) {
 695                         bawrite(bp);
 696                         vp->v_clen = 0;
 697                         vp->v_lasta = bp->b_bio2.bio_offset;
 698                         vp->v_cstart = loffset + lblocksize;
 699                         vp->v_lastw = loffset;
 700                         return;
 701                 }
 702                 if (maxclen > lblocksize)
 703                         vp->v_clen = maxclen - lblocksize;
 704                 else
 705                         vp->v_clen = 0;
 706                 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
 707                         vp->v_cstart = loffset + lblocksize;
 708                         bawrite(bp);
 709                 } else {        /* Wait for rest of cluster */
 710                         vp->v_cstart = loffset;
 711                         bdwrite(bp);
 712                 }
 713         } else if (loffset == vp->v_cstart + vp->v_clen) {
 714                 /*
 715                  * At end of cluster, write it out if seqcount tells us we
 716                  * are operating sequentially, otherwise let the buf or
 717                  * update daemon handle it.
 718                  */
 719                 bdwrite(bp);
 720                 if (seqcount > 1)
 721                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
 722                                           vp->v_clen + lblocksize);
 723                 vp->v_clen = 0;
 724                 vp->v_cstart = loffset + lblocksize;
 725         } else if (vm_page_count_severe()) {
 726                 /*
 727                  * We are low on memory, get it going NOW
 728                  */
 729                 bawrite(bp);
 730         } else {
 731                 /*
 732                  * In the middle of a cluster, so just delay the I/O for now.
 733                  */
 734                 bdwrite(bp);
 735         }
 736         vp->v_lastw = loffset;
 737         vp->v_lasta = bp->b_bio2.bio_offset;
 738 }
 739
 740
 741 /*
 742  * This is an awful lot like cluster_rbuild...wish they could be combined.
 743  * The last lbn argument is the current block on which I/O is being
 744  * performed.  Check to see that it doesn't fall in the middle of
 745  * the current block (if last_bp == NULL).
 746  */
 747 int
 748 cluster_wbuild(struct vnode *vp, int size, off_t start_loffset, int bytes)
 749 {
 750         struct buf *bp, *tbp;
 751         int i, j;
 752         int totalwritten = 0;
 753
 754         while (bytes > 0) {
 755                 crit_enter();
 756                 /*
 757                  * If the buffer is not delayed-write (i.e. dirty), or it
 758                  * is delayed-write but either locked or inval, it cannot
 759                  * partake in the clustered write.
 760                  */
 761                 if (((tbp = findblk(vp, start_loffset)) == NULL) ||
 762                   ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
 763                   BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 764                         start_loffset += size;
 765                         bytes -= size;
 766                         crit_exit();
 767                         continue;
 768                 }
 769                 bremfree(tbp);
 770                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 771                 crit_exit();
 772
 773                 /*
 774                  * Extra memory in the buffer, punt on this buffer.
 775                  * XXX we could handle this in most cases, but we would
 776                  * have to push the extra memory down to after our max
 777                  * possible cluster size and then potentially pull it back
 778                  * up if the cluster was terminated prematurely--too much
 779                  * hassle.
 780                  */
 781                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 782                   (tbp->b_bcount != tbp->b_bufsize) ||
 783                   (tbp->b_bcount != size) ||
 784                   (bytes == size) ||
 785                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 786                         totalwritten += tbp->b_bufsize;
 787                         bawrite(tbp);
 788                         start_loffset += size;
 789                         bytes -= size;
 790                         continue;
 791                 }
 792
 793                 /*
 794                  * Set up the pbuf.  Track our append point with b_bcount
 795                  * and b_bufsize.  b_bufsize is not used by the device but
 796                  * our caller uses it to loop clusters and we use it to
 797                  * detect a premature EOF on the block device.
 798                  */
 799                 bp->b_bcount = 0;
 800                 bp->b_bufsize = 0;
 801                 bp->b_xio.xio_npages = 0;
 802                 bp->b_loffset = tbp->b_loffset;
 803                 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
 804
 805                 /*
 806                  * We are synthesizing a buffer out of vm_page_t's, but
 807                  * if the block size is not page aligned then the starting
 808                  * address may not be either.  Inherit the b_data offset
 809                  * from the original buffer.
 810                  */
 811                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 812                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 813                 bp->b_flags &= ~B_ERROR;
 814                 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
 815                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN));
 816                 bp->b_bio1.bio_done = cluster_callback;
 817                 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 818                 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 819                 /*
 820                  * From this location in the file, scan forward to see
 821                  * if there are buffers with adjacent data that need to
 822                  * be written as well.
 823                  */
 824                 for (i = 0; i < bytes; (i += size), (start_loffset += size)) {
 825                         if (i != 0) { /* If not the first buffer */
 826                                 crit_enter();
 827                                 /*
 828                                  * If the adjacent data is not even in core it
 829                                  * can't need to be written.
 830                                  */
 831                                 if ((tbp = findblk(vp, start_loffset)) == NULL) {
 832                                         crit_exit();
 833                                         break;
 834                                 }
 835
 836                                 /*
 837                                  * If it IS in core, but has different
 838                                  * characteristics, or is locked (which
 839                                  * means it could be undergoing a background
 840                                  * I/O or be in a weird state), then don't
 841                                  * cluster with it.
 842                                  */
 843                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 844                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 845                                   != (B_DELWRI | B_CLUSTEROK |
 846                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 847                                     (tbp->b_flags & B_LOCKED) ||
 848                                     BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 849                                         crit_exit();
 850                                         break;
 851                                 }
 852
 853                                 /*
 854                                  * Check that the combined cluster
 855                                  * would make sense with regard to pages
 856                                  * and would not be too large
 857                                  */
 858                                 if ((tbp->b_bcount != size) ||
 859                                   ((bp->b_bio2.bio_offset + i) !=
 860                                     tbp->b_bio2.bio_offset) ||
 861                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
 862                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 863                                         BUF_UNLOCK(tbp);
 864                                         crit_exit();
 865                                         break;
 866                                 }
 867                                 /*
 868                                  * Ok, it's passed all the tests,
 869                                  * so remove it from the free list
 870                                  * and mark it busy. We will use it.
 871                                  */
 872                                 bremfree(tbp);
 873                                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 874                                 crit_exit();
 875                         } /* end of code for non-first buffers only */
 876
 877                         /*
 878                          * If the IO is via the VM then we do some
 879                          * special VM hackery (yuck).  Since the buffer's
 880                          * block size may not be page-aligned it is possible
 881                          * for a page to be shared between two buffers.  We
 882                          * have to get rid of the duplication when building
 883                          * the cluster.
 884                          */
 885                         if (tbp->b_flags & B_VMIO) {
 886                                 vm_page_t m;
 887
 888                                 if (i != 0) { /* if not first buffer */
 889                                         for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 890                                                 m = tbp->b_xio.xio_pages[j];
 891                                                 if (m->flags & PG_BUSY) {
 892                                                         bqrelse(tbp);
 893                                                         goto finishcluster;
 894                                                 }
 895                                         }
 896                                 }
 897
 898                                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 899                                         m = tbp->b_xio.xio_pages[j];
 900                                         vm_page_io_start(m);
 901                                         vm_object_pip_add(m->object, 1);
 902                                         if ((bp->b_xio.xio_npages == 0) ||
 903                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
 904                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 905                                                 bp->b_xio.xio_npages++;
 906                                         }
 907                                 }
 908                         }
 909                         bp->b_bcount += size;
 910                         bp->b_bufsize += size;
 911
 912                         crit_enter();
 913                         bundirty(tbp);
 914                         tbp->b_flags &= ~B_ERROR;
 915                         tbp->b_flags |= B_ASYNC;
 916                         tbp->b_cmd = BUF_CMD_WRITE;
 917                         crit_exit();
 918                         BUF_KERNPROC(tbp);
 919                         cluster_append(&bp->b_bio1, tbp);
 920
 921                         /*
 922                          * check for latent dependencies to be handled
 923                          */
 924                         if (LIST_FIRST(&tbp->b_dep) != NULL && bioops.io_start)
 925                                 (*bioops.io_start)(tbp);
 926
 927                 }
 928         finishcluster:
 929                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 930                         (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 931                 if (bp->b_bufsize > bp->b_kvasize) {
 932                         panic(
 933                             "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
 934                             bp->b_bufsize, bp->b_kvasize);
 935                 }
 936                 totalwritten += bp->b_bufsize;
 937                 bp->b_dirtyoff = 0;
 938                 bp->b_dirtyend = bp->b_bufsize;
 939                 bp->b_flags |= B_ASYNC;
 940                 bp->b_cmd = BUF_CMD_WRITE;
 941                 vfs_busy_pages(vp, bp);
 942                 bp->b_runningbufspace = bp->b_bufsize;
 943                 runningbufspace += bp->b_runningbufspace;
 944                 BUF_KERNPROC(bp);       /* B_ASYNC */
 945                 vn_strategy(vp, &bp->b_bio1);
 946
 947                 bytes -= i;
 948         }
 949         return totalwritten;
 950 }
 951
 952 /*
 953  * Collect together all the buffers in a cluster.
 954  * Plus add one additional buffer.
 955  */
 956 static struct cluster_save *
 957 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int lblocksize)
 958 {
 959         struct cluster_save *buflist;
 960         struct buf *bp;
 961         off_t loffset;
 962         int i, len;
 963
 964         len = (int)(vp->v_lastw - vp->v_cstart + lblocksize) / lblocksize;
 965         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
 966                          M_SEGMENT, M_WAITOK);
 967         buflist->bs_nchildren = 0;
 968         buflist->bs_children = (struct buf **) (buflist + 1);
 969         for (loffset = vp->v_cstart, i = 0; i < len; (loffset += lblocksize), i++) {
 970                 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
 971                 buflist->bs_children[i] = bp;
 972                 if (bp->b_bio2.bio_offset == NOOFFSET) {
 973                         VOP_BMAP(bp->b_vp, bp->b_loffset, NULL,
 974                                 &bp->b_bio2.bio_offset, NULL, NULL);
 975                 }
 976         }
 977         buflist->bs_children[i] = bp = last_bp;
 978         if (bp->b_bio2.bio_offset == NOOFFSET) {
 979                 VOP_BMAP(bp->b_vp, bp->b_loffset, NULL,
 980                          &bp->b_bio2.bio_offset, NULL, NULL);
 981         }
 982         buflist->bs_nchildren = i + 1;
 983         return (buflist);
 984 }
 985
 986 void
 987 cluster_append(struct bio *bio, struct buf *tbp)
 988 {
 989         tbp->b_cluster_next = NULL;
 990         if (bio->bio_caller_info1.cluster_head == NULL) {
 991                 bio->bio_caller_info1.cluster_head = tbp;
 992                 bio->bio_caller_info2.cluster_tail = tbp;
 993         } else {
 994                 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
 995                 bio->bio_caller_info2.cluster_tail = tbp;
 996         }
 997 }
 998