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 1.40 2008/07/14 03:09:00 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 blksize);
  70 static struct buf *
  71         cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
  72                             off_t doffset, int blksize, 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 blksize, 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 = vmaxiosize(vp) / blksize;
 110         maxra = 2 * racluster + (totread / blksize);
 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, blksize, 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 * blksize))) {
 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 * blksize;
 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 / blksize;
 171                 if (nblks) {
 172                         int burstbytes;
 173
 174                         if (nblks > racluster)
 175                                 nblks = racluster;
 176
 177                         error = VOP_BMAP(vp, loffset, &doffset,
 178                                          &burstbytes, NULL, BUF_CMD_READ);
 179                         if (error)
 180                                 goto single_block_read;
 181                         if (doffset == NOOFFSET)
 182                                 goto single_block_read;
 183                         if (burstbytes < blksize * 2)
 184                                 goto single_block_read;
 185                         if (nblks > burstbytes / blksize)
 186                                 nblks = burstbytes / blksize;
 187
 188                         bp = cluster_rbuild(vp, filesize, loffset,
 189                                             doffset, blksize, 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 += blksize;
 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                         kprintf("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                 if (bp->b_flags & B_ERROR) {
 222                         if ((error = bp->b_error) == 0)
 223                                 error = EIO;
 224                 } else {
 225                         error = 0;
 226                 }
 227         }
 228
 229         /*
 230          * If we have been doing sequential I/O, then do some read-ahead.
 231          *
 232          * Only mess with buffers which we can immediately lock.  HAMMER
 233          * will do device-readahead irrespective of what the blocks
 234          * represent.
 235          */
 236         rbp = NULL;
 237         if (!error &&
 238             seqcount &&
 239             loffset < origoffset + seqcount * blksize &&
 240             loffset + blksize <= filesize
 241         ) {
 242                 int nblksread;
 243                 int ntoread;
 244                 int burstbytes;
 245                 int tmp_error;
 246
 247                 rbp = getblk(vp, loffset, blksize,
 248                              GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
 249                 if (rbp == NULL)
 250                         goto no_read_ahead;
 251                 if ((rbp->b_flags & B_CACHE)) {
 252                         bqrelse(rbp);
 253                         goto no_read_ahead;
 254                 }
 255
 256                 /*
 257                  * An error from the read-ahead bmap has nothing to do
 258                  * with the caller's original request.
 259                  */
 260                 tmp_error = VOP_BMAP(vp, loffset, &doffset,
 261                                      &burstbytes, NULL, BUF_CMD_READ);
 262                 if (tmp_error || doffset == NOOFFSET) {
 263                         rbp->b_flags |= B_INVAL;
 264                         brelse(rbp);
 265                         rbp = NULL;
 266                         goto no_read_ahead;
 267                 }
 268                 ntoread = burstbytes / blksize;
 269                 nblksread = (totread + blksize - 1) / blksize;
 270                 if (seqcount < nblksread)
 271                         seqcount = nblksread;
 272                 if (ntoread > seqcount)
 273                         ntoread = seqcount;
 274
 275                 rbp->b_flags |= B_RAM/* | B_AGE*/;
 276                 if (burstbytes) {
 277                         rbp = cluster_rbuild(vp, filesize, loffset,
 278                                              doffset, blksize,
 279                                              ntoread, rbp, 1);
 280                 } else {
 281                         rbp->b_bio2.bio_offset = doffset;
 282                 }
 283 #if defined(CLUSTERDEBUG)
 284                 if (rcluster) {
 285                         if (bp)
 286                                 kprintf("A+(%lld,%d,%lld,%d) ",
 287                                     rbp->b_loffset, rbp->b_bcount,
 288                                     rbp->b_loffset - origoffset,
 289                                     seqcount);
 290                         else
 291                                 kprintf("A(%lld,%d,%lld,%d) ",
 292                                     rbp->b_loffset, rbp->b_bcount,
 293                                     rbp->b_loffset - origoffset,
 294                                     seqcount);
 295                 }
 296 #endif
 297                 rbp->b_flags &= ~(B_ERROR|B_INVAL);
 298                 rbp->b_flags |= B_ASYNC;
 299                 rbp->b_cmd = BUF_CMD_READ;
 300
 301                 if ((rbp->b_flags & B_CLUSTER) == 0)
 302                         vfs_busy_pages(vp, rbp);
 303                 BUF_KERNPROC(rbp);                      /* B_ASYNC */
 304                 vn_strategy(vp, &rbp->b_bio1);
 305         }
 306 no_read_ahead:
 307
 308         if (reqbp)
 309                 return (biowait(reqbp));
 310         else
 311                 return (error);
 312 }
 313
 314 /*
 315  * If blocks are contiguous on disk, use this to provide clustered
 316  * read ahead.  We will read as many blocks as possible sequentially
 317  * and then parcel them up into logical blocks in the buffer hash table.
 318  */
 319 static struct buf *
 320 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset,
 321         off_t doffset, int blksize, int run, struct buf *fbp, int doasync)
 322 {
 323         struct buf *bp, *tbp;
 324         off_t boffset;
 325         int i, j;
 326         int maxiosize = vmaxiosize(vp);
 327
 328         /*
 329          * avoid a division
 330          */
 331         while (loffset + run * blksize > filesize) {
 332                 --run;
 333         }
 334
 335         tbp = fbp;
 336         tbp->b_bio2.bio_offset = doffset;
 337         if((tbp->b_flags & B_MALLOC) ||
 338             ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
 339                 return tbp;
 340         }
 341
 342         bp = trypbuf(&cluster_pbuf_freecnt);
 343         if (bp == NULL)
 344                 return tbp;
 345
 346         /*
 347          * We are synthesizing a buffer out of vm_page_t's, but
 348          * if the block size is not page aligned then the starting
 349          * address may not be either.  Inherit the b_data offset
 350          * from the original buffer.
 351          */
 352         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 353             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 354         bp->b_flags |= B_ASYNC | B_CLUSTER | B_VMIO;
 355         bp->b_cmd = BUF_CMD_READ;
 356         bp->b_bio1.bio_done = cluster_callback;
 357         bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 358         bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 359         bp->b_loffset = loffset;
 360         bp->b_bio2.bio_offset = doffset;
 361         KASSERT(bp->b_loffset != NOOFFSET,
 362                 ("cluster_rbuild: no buffer offset"));
 363
 364         bp->b_bcount = 0;
 365         bp->b_bufsize = 0;
 366         bp->b_xio.xio_npages = 0;
 367
 368         for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
 369                 if (i) {
 370                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
 371                             round_page(blksize) > maxiosize) {
 372                                 break;
 373                         }
 374
 375                         /*
 376                          * Shortcut some checks and try to avoid buffers that
 377                          * would block in the lock.  The same checks have to
 378                          * be made again after we officially get the buffer.
 379                          */
 380                         tbp = getblk(vp, loffset + i * blksize, blksize,
 381                                      GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
 382                         if (tbp == NULL)
 383                                 break;
 384                         for (j = 0; j < tbp->b_xio.xio_npages; j++) {
 385                                 if (tbp->b_xio.xio_pages[j]->valid)
 386                                         break;
 387                         }
 388                         if (j != tbp->b_xio.xio_npages) {
 389                                 bqrelse(tbp);
 390                                 break;
 391                         }
 392
 393                         /*
 394                          * Stop scanning if the buffer is fuly valid
 395                          * (marked B_CACHE), or locked (may be doing a
 396                          * background write), or if the buffer is not
 397                          * VMIO backed.  The clustering code can only deal
 398                          * with VMIO-backed buffers.
 399                          */
 400                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 401                             (tbp->b_flags & B_VMIO) == 0 ||
 402                             (LIST_FIRST(&tbp->b_dep) != NULL &&
 403                              buf_checkread(tbp))
 404                         ) {
 405                                 bqrelse(tbp);
 406                                 break;
 407                         }
 408
 409                         /*
 410                          * The buffer must be completely invalid in order to
 411                          * take part in the cluster.  If it is partially valid
 412                          * then we stop.
 413                          */
 414                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
 415                                 if (tbp->b_xio.xio_pages[j]->valid)
 416                                         break;
 417                         }
 418                         if (j != tbp->b_xio.xio_npages) {
 419                                 bqrelse(tbp);
 420                                 break;
 421                         }
 422
 423                         /*
 424                          * Set a read-ahead mark as appropriate
 425                          */
 426                         if (i == 1 || i == (run - 1))
 427                                 tbp->b_flags |= B_RAM;
 428
 429                         /*
 430                          * Depress the priority of buffers not explicitly
 431                          * requested.
 432                          */
 433                         /* tbp->b_flags |= B_AGE; */
 434
 435                         /*
 436                          * Set the block number if it isn't set, otherwise
 437                          * if it is make sure it matches the block number we
 438                          * expect.
 439                          */
 440                         if (tbp->b_bio2.bio_offset == NOOFFSET) {
 441                                 tbp->b_bio2.bio_offset = boffset;
 442                         } else if (tbp->b_bio2.bio_offset != boffset) {
 443                                 brelse(tbp);
 444                                 break;
 445                         }
 446                 }
 447                 /*
 448                  * The first buffer is setup async if doasync is specified.
 449                  * All other buffers in the cluster are setup async.  This
 450                  * way the caller can decide how to deal with the requested
 451                  * buffer.
 452                  */
 453                 if (i || doasync)
 454                         tbp->b_flags |= B_ASYNC;
 455                 tbp->b_cmd = BUF_CMD_READ;
 456                 BUF_KERNPROC(tbp);
 457                 cluster_append(&bp->b_bio1, tbp);
 458                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 459                         vm_page_t m;
 460                         m = tbp->b_xio.xio_pages[j];
 461                         vm_page_io_start(m);
 462                         vm_object_pip_add(m->object, 1);
 463                         if ((bp->b_xio.xio_npages == 0) ||
 464                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
 465                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 466                                 bp->b_xio.xio_npages++;
 467                         }
 468                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 469                                 tbp->b_xio.xio_pages[j] = bogus_page;
 470                 }
 471                 /*
 472                  * XXX shouldn't this be += size for both, like in
 473                  * cluster_wbuild()?
 474                  *
 475                  * Don't inherit tbp->b_bufsize as it may be larger due to
 476                  * a non-page-aligned size.  Instead just aggregate using
 477                  * 'size'.
 478                  */
 479                 if (tbp->b_bcount != blksize)
 480                     kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
 481                 if (tbp->b_bufsize != blksize)
 482                     kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
 483                 bp->b_bcount += blksize;
 484                 bp->b_bufsize += blksize;
 485         }
 486
 487         /*
 488          * Fully valid pages in the cluster are already good and do not need
 489          * to be re-read from disk.  Replace the page with bogus_page
 490          */
 491         for (j = 0; j < bp->b_xio.xio_npages; j++) {
 492                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 493                     VM_PAGE_BITS_ALL) {
 494                         bp->b_xio.xio_pages[j] = bogus_page;
 495                 }
 496         }
 497         if (bp->b_bufsize > bp->b_kvasize) {
 498                 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
 499                     bp->b_bufsize, bp->b_kvasize);
 500         }
 501
 502         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 503                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 504         return (bp);
 505 }
 506
 507 /*
 508  * Cleanup after a clustered read or write.
 509  * This is complicated by the fact that any of the buffers might have
 510  * extra memory (if there were no empty buffer headers at allocbuf time)
 511  * that we will need to shift around.
 512  *
 513  * The returned bio is &bp->b_bio1
 514  */
 515 void
 516 cluster_callback(struct bio *bio)
 517 {
 518         struct buf *bp = bio->bio_buf;
 519         struct buf *tbp;
 520         int error = 0;
 521
 522         /*
 523          * Must propogate errors to all the components.  A short read (EOF)
 524          * is a critical error.
 525          */
 526         if (bp->b_flags & B_ERROR) {
 527                 error = bp->b_error;
 528         } else if (bp->b_bcount != bp->b_bufsize) {
 529                 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
 530         }
 531
 532         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
 533         /*
 534          * Move memory from the large cluster buffer into the component
 535          * buffers and mark IO as done on these.  Since the memory map
 536          * is the same, no actual copying is required.
 537          */
 538         while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
 539                 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
 540                 if (error) {
 541                         tbp->b_flags |= B_ERROR;
 542                         tbp->b_error = error;
 543                 } else {
 544                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 545                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 546                         /*
 547                          * XXX the bdwrite()/bqrelse() issued during
 548                          * cluster building clears B_RELBUF (see bqrelse()
 549                          * comment).  If direct I/O was specified, we have
 550                          * to restore it here to allow the buffer and VM
 551                          * to be freed.
 552                          */
 553                         if (tbp->b_flags & B_DIRECT)
 554                                 tbp->b_flags |= B_RELBUF;
 555                 }
 556                 biodone(&tbp->b_bio1);
 557         }
 558         relpbuf(bp, &cluster_pbuf_freecnt);
 559 }
 560
 561 /*
 562  *      cluster_wbuild_wb:
 563  *
 564  *      Implement modified write build for cluster.
 565  *
 566  *              write_behind = 0        write behind disabled
 567  *              write_behind = 1        write behind normal (default)
 568  *              write_behind = 2        write behind backed-off
 569  */
 570
 571 static __inline int
 572 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
 573 {
 574         int r = 0;
 575
 576         switch(write_behind) {
 577         case 2:
 578                 if (start_loffset < len)
 579                         break;
 580                 start_loffset -= len;
 581                 /* fall through */
 582         case 1:
 583                 r = cluster_wbuild(vp, blksize, start_loffset, len);
 584                 /* fall through */
 585         default:
 586                 /* fall through */
 587                 break;
 588         }
 589         return(r);
 590 }
 591
 592 /*
 593  * Do clustered write for FFS.
 594  *
 595  * Three cases:
 596  *      1. Write is not sequential (write asynchronously)
 597  *      Write is sequential:
 598  *      2.      beginning of cluster - begin cluster
 599  *      3.      middle of a cluster - add to cluster
 600  *      4.      end of a cluster - asynchronously write cluster
 601  */
 602 void
 603 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
 604 {
 605         struct vnode *vp;
 606         off_t loffset;
 607         int maxclen, cursize;
 608         int async;
 609
 610         vp = bp->b_vp;
 611         if (vp->v_type == VREG)
 612                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 613         else
 614                 async = 0;
 615         loffset = bp->b_loffset;
 616         KASSERT(bp->b_loffset != NOOFFSET,
 617                 ("cluster_write: no buffer offset"));
 618
 619         /* Initialize vnode to beginning of file. */
 620         if (loffset == 0)
 621                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 622
 623         if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
 624             bp->b_bio2.bio_offset == NOOFFSET ||
 625             (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
 626                 maxclen = vmaxiosize(vp);
 627                 if (vp->v_clen != 0) {
 628                         /*
 629                          * Next block is not sequential.
 630                          *
 631                          * If we are not writing at end of file, the process
 632                          * seeked to another point in the file since its last
 633                          * write, or we have reached our maximum cluster size,
 634                          * then push the previous cluster. Otherwise try
 635                          * reallocating to make it sequential.
 636                          *
 637                          * Change to algorithm: only push previous cluster if
 638                          * it was sequential from the point of view of the
 639                          * seqcount heuristic, otherwise leave the buffer
 640                          * intact so we can potentially optimize the I/O
 641                          * later on in the buf_daemon or update daemon
 642                          * flush.
 643                          */
 644                         cursize = vp->v_lastw - vp->v_cstart + blksize;
 645                         if (bp->b_loffset + blksize != filesize ||
 646                             loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
 647                                 if (!async && seqcount > 0) {
 648                                         cluster_wbuild_wb(vp, blksize,
 649                                                 vp->v_cstart, cursize);
 650                                 }
 651                         } else {
 652                                 struct buf **bpp, **endbp;
 653                                 struct cluster_save *buflist;
 654
 655                                 buflist = cluster_collectbufs(vp, bp, blksize);
 656                                 endbp = &buflist->bs_children
 657                                     [buflist->bs_nchildren - 1];
 658                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 659                                         /*
 660                                          * Failed, push the previous cluster
 661                                          * if *really* writing sequentially
 662                                          * in the logical file (seqcount > 1),
 663                                          * otherwise delay it in the hopes that
 664                                          * the low level disk driver can
 665                                          * optimize the write ordering.
 666                                          */
 667                                         for (bpp = buflist->bs_children;
 668                                              bpp < endbp; bpp++)
 669                                                 brelse(*bpp);
 670                                         kfree(buflist, M_SEGMENT);
 671                                         if (seqcount > 1) {
 672                                                 cluster_wbuild_wb(vp,
 673                                                     blksize, vp->v_cstart,
 674                                                     cursize);
 675                                         }
 676                                 } else {
 677                                         /*
 678                                          * Succeeded, keep building cluster.
 679                                          */
 680                                         for (bpp = buflist->bs_children;
 681                                              bpp <= endbp; bpp++)
 682                                                 bdwrite(*bpp);
 683                                         kfree(buflist, M_SEGMENT);
 684                                         vp->v_lastw = loffset;
 685                                         vp->v_lasta = bp->b_bio2.bio_offset;
 686                                         return;
 687                                 }
 688                         }
 689                 }
 690                 /*
 691                  * Consider beginning a cluster. If at end of file, make
 692                  * cluster as large as possible, otherwise find size of
 693                  * existing cluster.
 694                  */
 695                 if ((vp->v_type == VREG) &&
 696                     bp->b_loffset + blksize != filesize &&
 697                     (bp->b_bio2.bio_offset == NOOFFSET) &&
 698                     (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
 699                      bp->b_bio2.bio_offset == NOOFFSET)) {
 700                         bawrite(bp);
 701                         vp->v_clen = 0;
 702                         vp->v_lasta = bp->b_bio2.bio_offset;
 703                         vp->v_cstart = loffset + blksize;
 704                         vp->v_lastw = loffset;
 705                         return;
 706                 }
 707                 if (maxclen > blksize)
 708                         vp->v_clen = maxclen - blksize;
 709                 else
 710                         vp->v_clen = 0;
 711                 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
 712                         vp->v_cstart = loffset + blksize;
 713                         bawrite(bp);
 714                 } else {        /* Wait for rest of cluster */
 715                         vp->v_cstart = loffset;
 716                         bdwrite(bp);
 717                 }
 718         } else if (loffset == vp->v_cstart + vp->v_clen) {
 719                 /*
 720                  * At end of cluster, write it out if seqcount tells us we
 721                  * are operating sequentially, otherwise let the buf or
 722                  * update daemon handle it.
 723                  */
 724                 bdwrite(bp);
 725                 if (seqcount > 1)
 726                         cluster_wbuild_wb(vp, blksize, vp->v_cstart,
 727                                           vp->v_clen + blksize);
 728                 vp->v_clen = 0;
 729                 vp->v_cstart = loffset + blksize;
 730         } else if (vm_page_count_severe()) {
 731                 /*
 732                  * We are low on memory, get it going NOW
 733                  */
 734                 bawrite(bp);
 735         } else {
 736                 /*
 737                  * In the middle of a cluster, so just delay the I/O for now.
 738                  */
 739                 bdwrite(bp);
 740         }
 741         vp->v_lastw = loffset;
 742         vp->v_lasta = bp->b_bio2.bio_offset;
 743 }
 744
 745
 746 /*
 747  * This is an awful lot like cluster_rbuild...wish they could be combined.
 748  * The last lbn argument is the current block on which I/O is being
 749  * performed.  Check to see that it doesn't fall in the middle of
 750  * the current block (if last_bp == NULL).
 751  */
 752 int
 753 cluster_wbuild(struct vnode *vp, int blksize, off_t start_loffset, int bytes)
 754 {
 755         struct buf *bp, *tbp;
 756         int i, j;
 757         int totalwritten = 0;
 758         int maxiosize = vmaxiosize(vp);
 759
 760         while (bytes > 0) {
 761                 crit_enter();
 762                 /*
 763                  * If the buffer is not delayed-write (i.e. dirty), or it
 764                  * is delayed-write but either locked or inval, it cannot
 765                  * partake in the clustered write.
 766                  */
 767                 if (((tbp = findblk(vp, start_loffset)) == NULL) ||
 768                   ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
 769                   (LIST_FIRST(&tbp->b_dep) != NULL && buf_checkwrite(tbp)) ||
 770                   BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 771                         start_loffset += blksize;
 772                         bytes -= blksize;
 773                         crit_exit();
 774                         continue;
 775                 }
 776                 bremfree(tbp);
 777                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 778                 crit_exit();
 779
 780                 /*
 781                  * Extra memory in the buffer, punt on this buffer.
 782                  * XXX we could handle this in most cases, but we would
 783                  * have to push the extra memory down to after our max
 784                  * possible cluster size and then potentially pull it back
 785                  * up if the cluster was terminated prematurely--too much
 786                  * hassle.
 787                  */
 788                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 789                   (tbp->b_bcount != tbp->b_bufsize) ||
 790                   (tbp->b_bcount != blksize) ||
 791                   (bytes == blksize) ||
 792                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 793                         totalwritten += tbp->b_bufsize;
 794                         bawrite(tbp);
 795                         start_loffset += blksize;
 796                         bytes -= blksize;
 797                         continue;
 798                 }
 799
 800                 /*
 801                  * Set up the pbuf.  Track our append point with b_bcount
 802                  * and b_bufsize.  b_bufsize is not used by the device but
 803                  * our caller uses it to loop clusters and we use it to
 804                  * detect a premature EOF on the block device.
 805                  */
 806                 bp->b_bcount = 0;
 807                 bp->b_bufsize = 0;
 808                 bp->b_xio.xio_npages = 0;
 809                 bp->b_loffset = tbp->b_loffset;
 810                 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
 811
 812                 /*
 813                  * We are synthesizing a buffer out of vm_page_t's, but
 814                  * if the block size is not page aligned then the starting
 815                  * address may not be either.  Inherit the b_data offset
 816                  * from the original buffer.
 817                  */
 818                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 819                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 820                 bp->b_flags &= ~B_ERROR;
 821                 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
 822                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
 823                 bp->b_bio1.bio_done = cluster_callback;
 824                 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 825                 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 826                 /*
 827                  * From this location in the file, scan forward to see
 828                  * if there are buffers with adjacent data that need to
 829                  * be written as well.
 830                  */
 831                 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
 832                         if (i != 0) { /* If not the first buffer */
 833                                 crit_enter();
 834                                 /*
 835                                  * If the adjacent data is not even in core it
 836                                  * can't need to be written.
 837                                  */
 838                                 if ((tbp = findblk(vp, start_loffset)) == NULL) {
 839                                         crit_exit();
 840                                         break;
 841                                 }
 842
 843                                 /*
 844                                  * If it IS in core, but has different
 845                                  * characteristics, or is locked (which
 846                                  * means it could be undergoing a background
 847                                  * I/O or be in a weird state), then don't
 848                                  * cluster with it.
 849                                  */
 850                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 851                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 852                                   != (B_DELWRI | B_CLUSTEROK |
 853                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 854                                     (tbp->b_flags & B_LOCKED) ||
 855                   (LIST_FIRST(&tbp->b_dep) != NULL && buf_checkwrite(tbp)) ||
 856                                     BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
 857                                         crit_exit();
 858                                         break;
 859                                 }
 860
 861                                 /*
 862                                  * Check that the combined cluster
 863                                  * would make sense with regard to pages
 864                                  * and would not be too large
 865                                  */
 866                                 if ((tbp->b_bcount != blksize) ||
 867                                   ((bp->b_bio2.bio_offset + i) !=
 868                                     tbp->b_bio2.bio_offset) ||
 869                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
 870                                     (maxiosize / PAGE_SIZE))) {
 871                                         BUF_UNLOCK(tbp);
 872                                         crit_exit();
 873                                         break;
 874                                 }
 875                                 /*
 876                                  * Ok, it's passed all the tests,
 877                                  * so remove it from the free list
 878                                  * and mark it busy. We will use it.
 879                                  */
 880                                 bremfree(tbp);
 881                                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 882                                 crit_exit();
 883                         } /* end of code for non-first buffers only */
 884
 885                         /*
 886                          * If the IO is via the VM then we do some
 887                          * special VM hackery (yuck).  Since the buffer's
 888                          * block size may not be page-aligned it is possible
 889                          * for a page to be shared between two buffers.  We
 890                          * have to get rid of the duplication when building
 891                          * the cluster.
 892                          */
 893                         if (tbp->b_flags & B_VMIO) {
 894                                 vm_page_t m;
 895
 896                                 if (i != 0) { /* if not first buffer */
 897                                         for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 898                                                 m = tbp->b_xio.xio_pages[j];
 899                                                 if (m->flags & PG_BUSY) {
 900                                                         bqrelse(tbp);
 901                                                         goto finishcluster;
 902                                                 }
 903                                         }
 904                                 }
 905
 906                                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 907                                         m = tbp->b_xio.xio_pages[j];
 908                                         vm_page_io_start(m);
 909                                         vm_object_pip_add(m->object, 1);
 910                                         if ((bp->b_xio.xio_npages == 0) ||
 911                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
 912                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 913                                                 bp->b_xio.xio_npages++;
 914                                         }
 915                                 }
 916                         }
 917                         bp->b_bcount += blksize;
 918                         bp->b_bufsize += blksize;
 919
 920                         crit_enter();
 921                         bundirty(tbp);
 922                         tbp->b_flags &= ~B_ERROR;
 923                         tbp->b_flags |= B_ASYNC;
 924                         tbp->b_cmd = BUF_CMD_WRITE;
 925                         crit_exit();
 926                         BUF_KERNPROC(tbp);
 927                         cluster_append(&bp->b_bio1, tbp);
 928
 929                         /*
 930                          * check for latent dependencies to be handled
 931                          */
 932                         if (LIST_FIRST(&tbp->b_dep) != NULL)
 933                                 buf_start(tbp);
 934                 }
 935         finishcluster:
 936                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 937                         (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 938                 if (bp->b_bufsize > bp->b_kvasize) {
 939                         panic(
 940                             "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
 941                             bp->b_bufsize, bp->b_kvasize);
 942                 }
 943                 totalwritten += bp->b_bufsize;
 944                 bp->b_dirtyoff = 0;
 945                 bp->b_dirtyend = bp->b_bufsize;
 946                 bp->b_flags |= B_ASYNC;
 947                 bp->b_cmd = BUF_CMD_WRITE;
 948                 vfs_busy_pages(vp, bp);
 949                 bp->b_runningbufspace = bp->b_bufsize;
 950                 if (bp->b_runningbufspace) {
 951                         runningbufspace += bp->b_runningbufspace;
 952                         ++runningbufcount;
 953                 }
 954                 BUF_KERNPROC(bp);       /* B_ASYNC */
 955                 vn_strategy(vp, &bp->b_bio1);
 956
 957                 bytes -= i;
 958         }
 959         return totalwritten;
 960 }
 961
 962 /*
 963  * Collect together all the buffers in a cluster.
 964  * Plus add one additional buffer.
 965  */
 966 static struct cluster_save *
 967 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
 968 {
 969         struct cluster_save *buflist;
 970         struct buf *bp;
 971         off_t loffset;
 972         int i, len;
 973
 974         len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
 975         buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
 976                          M_SEGMENT, M_WAITOK);
 977         buflist->bs_nchildren = 0;
 978         buflist->bs_children = (struct buf **) (buflist + 1);
 979         for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
 980                 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
 981                 buflist->bs_children[i] = bp;
 982                 if (bp->b_bio2.bio_offset == NOOFFSET) {
 983                         VOP_BMAP(bp->b_vp, bp->b_loffset,
 984                                  &bp->b_bio2.bio_offset,
 985                                  NULL, NULL, BUF_CMD_WRITE);
 986                 }
 987         }
 988         buflist->bs_children[i] = bp = last_bp;
 989         if (bp->b_bio2.bio_offset == NOOFFSET) {
 990                 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
 991                          NULL, NULL, BUF_CMD_WRITE);
 992         }
 993         buflist->bs_nchildren = i + 1;
 994         return (buflist);
 995 }
 996
 997 void
 998 cluster_append(struct bio *bio, struct buf *tbp)
 999 {
1000         tbp->b_cluster_next = NULL;
1001         if (bio->bio_caller_info1.cluster_head == NULL) {
1002                 bio->bio_caller_info1.cluster_head = tbp;
1003                 bio->bio_caller_info2.cluster_tail = tbp;
1004         } else {
1005                 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1006                 bio->bio_caller_info2.cluster_tail = tbp;
1007         }
1008 }
1009