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
  57 #include <sys/buf2.h>
  58 #include <vm/vm_page2.h>
  59
  60 #include <machine/limits.h>
  61
  62 #if defined(CLUSTERDEBUG)
  63 #include <sys/sysctl.h>
  64 static int      rcluster= 0;
  65 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
  66 #endif
  67
  68 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");
  69
  70 static struct cluster_save *
  71         cluster_collectbufs (struct vnode *vp, struct buf *last_bp,
  72                             int blksize);
  73 static struct buf *
  74         cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
  75                             off_t doffset, int blksize, int run,
  76                             struct buf *fbp);
  77 static void cluster_callback (struct bio *);
  78 static void cluster_setram (struct buf *);
  79
  80 static int write_behind = 1;
  81 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
  82     "Cluster write-behind setting");
  83 static int max_readahead = 2 * 1024 * 1024;
  84 SYSCTL_INT(_vfs, OID_AUTO, max_readahead, CTLFLAG_RW, &max_readahead, 0,
  85     "Limit in bytes for desired cluster read-ahead");
  86
  87 extern vm_page_t        bogus_page;
  88
  89 extern int cluster_pbuf_freecnt;
  90
  91 /*
  92  * This replaces bread.
  93  *
  94  * filesize     - read-ahead @ blksize will not cross this boundary
  95  * loffset      - loffset for returned *bpp
  96  * blksize      - blocksize for returned *bpp and read-ahead bps
  97  * minreq       - minimum (not a hard minimum) in bytes, typically reflects
  98  *                a higher level uio resid.
  99  * maxreq       - maximum (sequential heuristic) in bytes (highet typ ~2MB)
 100  * bpp          - return buffer (*bpp) for (loffset,blksize)
 101  */
 102 int
 103 cluster_readx(struct vnode *vp, off_t filesize, off_t loffset,
 104              int blksize, size_t minreq, size_t maxreq, struct buf **bpp)
 105 {
 106         struct buf *bp, *rbp, *reqbp;
 107         off_t origoffset;
 108         off_t doffset;
 109         int error;
 110         int i;
 111         int maxra;
 112         int maxrbuild;
 113
 114         error = 0;
 115
 116         /*
 117          * Calculate the desired read-ahead in blksize'd blocks (maxra).
 118          * To do this we calculate maxreq.
 119          *
 120          * maxreq typically starts out as a sequential heuristic.  If the
 121          * high level uio/resid is bigger (minreq), we pop maxreq up to
 122          * minreq.  This represents the case where random I/O is being
 123          * performed by the userland is issuing big read()'s.
 124          *
 125          * Then we limit maxreq to max_readahead to ensure it is a reasonable
 126          * value.
 127          *
 128          * Finally we must ensure that (loffset + maxreq) does not cross the
 129          * boundary (filesize) for the current blocksize.  If we allowed it
 130          * to cross we could end up with buffers past the boundary with the
 131          * wrong block size (HAMMER large-data areas use mixed block sizes).
 132          * minreq is also absolutely limited to filesize.
 133          */
 134         if (maxreq < minreq)
 135                 maxreq = minreq;
 136         /* minreq not used beyond this point */
 137
 138         if (maxreq > max_readahead) {
 139                 maxreq = max_readahead;
 140                 if (maxreq > 16 * 1024 * 1024)
 141                         maxreq = 16 * 1024 * 1024;
 142         }
 143         if (maxreq < blksize)
 144                 maxreq = blksize;
 145         if (loffset + maxreq > filesize) {
 146                 if (loffset > filesize)
 147                         maxreq = 0;
 148                 else
 149                         maxreq = filesize - loffset;
 150         }
 151
 152         maxra = (int)(maxreq / blksize);
 153
 154         /*
 155          * Get the requested block.
 156          */
 157         if (*bpp)
 158                 reqbp = bp = *bpp;
 159         else
 160                 *bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0);
 161         origoffset = loffset;
 162
 163         /*
 164          * Calculate the maximum cluster size for a single I/O, used
 165          * by cluster_rbuild().
 166          */
 167         maxrbuild = vmaxiosize(vp) / blksize;
 168
 169         /*
 170          * if it is in the cache, then check to see if the reads have been
 171          * sequential.  If they have, then try some read-ahead, otherwise
 172          * back-off on prospective read-aheads.
 173          */
 174         if (bp->b_flags & B_CACHE) {
 175                 /*
 176                  * Not sequential, do not do any read-ahead
 177                  */
 178                 if (maxra <= 1)
 179                         return 0;
 180
 181                 /*
 182                  * No read-ahead mark, do not do any read-ahead
 183                  * yet.
 184                  */
 185                 if ((bp->b_flags & B_RAM) == 0)
 186                         return 0;
 187
 188                 /*
 189                  * We hit a read-ahead-mark, figure out how much read-ahead
 190                  * to do (maxra) and where to start (loffset).
 191                  *
 192                  * Shortcut the scan.  Typically the way this works is that
 193                  * we've built up all the blocks inbetween except for the
 194                  * last in previous iterations, so if the second-to-last
 195                  * block is present we just skip ahead to it.
 196                  *
 197                  * This algorithm has O(1) cpu in the steady state no
 198                  * matter how large maxra is.
 199                  */
 200                 bp->b_flags &= ~B_RAM;
 201
 202                 if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST))
 203                         i = maxra - 1;
 204                 else
 205                         i = 1;
 206                 while (i < maxra) {
 207                         if (findblk(vp, loffset + i * blksize,
 208                                     FINDBLK_TEST) == NULL) {
 209                                 break;
 210                         }
 211                         ++i;
 212                 }
 213
 214                 /*
 215                  * We got everything or everything is in the cache, no
 216                  * point continuing.
 217                  */
 218                 if (i >= maxra)
 219                         return 0;
 220
 221                 /*
 222                  * Calculate where to start the read-ahead and how much
 223                  * to do.  Generally speaking we want to read-ahead by
 224                  * (maxra) when we've found a read-ahead mark.  We do
 225                  * not want to reduce maxra here as it will cause
 226                  * successive read-ahead I/O's to be smaller and smaller.
 227                  *
 228                  * However, we have to make sure we don't break the
 229                  * filesize limitation for the clustered operation.
 230                  */
 231                 loffset += i * blksize;
 232                 reqbp = bp = NULL;
 233
 234                 if (loffset >= filesize)
 235                         return 0;
 236                 if (loffset + maxra * blksize > filesize) {
 237                         maxreq = filesize - loffset;
 238                         maxra = (int)(maxreq / blksize);
 239                 }
 240         } else {
 241                 __debugvar off_t firstread = bp->b_loffset;
 242                 int nblks;
 243
 244                 /*
 245                  * Set-up synchronous read for bp.
 246                  */
 247                 bp->b_cmd = BUF_CMD_READ;
 248                 bp->b_bio1.bio_done = biodone_sync;
 249                 bp->b_bio1.bio_flags |= BIO_SYNC;
 250
 251                 KASSERT(firstread != NOOFFSET,
 252                         ("cluster_read: no buffer offset"));
 253
 254                 /*
 255                  * nblks is our cluster_rbuild request size, limited
 256                  * primarily by the device.
 257                  */
 258                 if ((nblks = maxra) > maxrbuild)
 259                         nblks = maxrbuild;
 260
 261                 if (nblks > 1) {
 262                         int burstbytes;
 263
 264                         error = VOP_BMAP(vp, loffset, &doffset,
 265                                          &burstbytes, NULL, BUF_CMD_READ);
 266                         if (error)
 267                                 goto single_block_read;
 268                         if (nblks > burstbytes / blksize)
 269                                 nblks = burstbytes / blksize;
 270                         if (doffset == NOOFFSET)
 271                                 goto single_block_read;
 272                         if (nblks <= 1)
 273                                 goto single_block_read;
 274
 275                         bp = cluster_rbuild(vp, filesize, loffset,
 276                                             doffset, blksize, nblks, bp);
 277                         loffset += bp->b_bufsize;
 278                         maxra -= bp->b_bufsize / blksize;
 279                 } else {
 280 single_block_read:
 281                         /*
 282                          * If it isn't in the cache, then get a chunk from
 283                          * disk if sequential, otherwise just get the block.
 284                          */
 285                         cluster_setram(bp);
 286                         loffset += blksize;
 287                         --maxra;
 288                 }
 289         }
 290
 291         /*
 292          * If B_CACHE was not set issue bp.  bp will either be an
 293          * asynchronous cluster buf or a synchronous single-buf.
 294          * If it is a single buf it will be the same as reqbp.
 295          *
 296          * NOTE: Once an async cluster buf is issued bp becomes invalid.
 297          */
 298         if (bp) {
 299 #if defined(CLUSTERDEBUG)
 300                 if (rcluster)
 301                         kprintf("S(%012jx,%d,%d)\n",
 302                             (intmax_t)bp->b_loffset, bp->b_bcount, maxra);
 303 #endif
 304                 if ((bp->b_flags & B_CLUSTER) == 0)
 305                         vfs_busy_pages(vp, bp);
 306                 bp->b_flags &= ~(B_ERROR|B_INVAL);
 307                 vn_strategy(vp, &bp->b_bio1);
 308                 error = 0;
 309                 /* bp invalid now */
 310         }
 311
 312         /*
 313          * If we have been doing sequential I/O, then do some read-ahead.
 314          * The code above us should have positioned us at the next likely
 315          * offset.
 316          *
 317          * Only mess with buffers which we can immediately lock.  HAMMER
 318          * will do device-readahead irrespective of what the blocks
 319          * represent.
 320          */
 321         while (error == 0 && maxra > 0) {
 322                 int burstbytes;
 323                 int tmp_error;
 324                 int nblks;
 325
 326                 rbp = getblk(vp, loffset, blksize,
 327                              GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
 328                 if (rbp == NULL)
 329                         goto no_read_ahead;
 330                 if ((rbp->b_flags & B_CACHE)) {
 331                         bqrelse(rbp);
 332                         goto no_read_ahead;
 333                 }
 334
 335                 /*
 336                  * An error from the read-ahead bmap has nothing to do
 337                  * with the caller's original request.
 338                  */
 339                 tmp_error = VOP_BMAP(vp, loffset, &doffset,
 340                                      &burstbytes, NULL, BUF_CMD_READ);
 341                 if (tmp_error || doffset == NOOFFSET) {
 342                         rbp->b_flags |= B_INVAL;
 343                         brelse(rbp);
 344                         rbp = NULL;
 345                         goto no_read_ahead;
 346                 }
 347                 if ((nblks = maxra) > maxrbuild)
 348                         nblks = maxrbuild;
 349                 if (nblks > burstbytes / blksize)
 350                         nblks = burstbytes / blksize;
 351
 352                 /*
 353                  * rbp: async read
 354                  */
 355                 rbp->b_cmd = BUF_CMD_READ;
 356                 /*rbp->b_flags |= B_AGE*/;
 357                 cluster_setram(rbp);
 358
 359                 if (nblks > 1) {
 360                         rbp = cluster_rbuild(vp, filesize, loffset,
 361                                              doffset, blksize,
 362                                              nblks, rbp);
 363                 } else {
 364                         rbp->b_bio2.bio_offset = doffset;
 365                 }
 366
 367 #if defined(CLUSTERDEBUG)
 368                 if (rcluster) {
 369                         if (bp) {
 370                                 kprintf("A+(%012jx,%d,%jd) "
 371                                         "doff=%012jx minr=%zd ra=%d\n",
 372                                     (intmax_t)loffset, rbp->b_bcount,
 373                                     (intmax_t)(loffset - origoffset),
 374                                     (intmax_t)doffset, minreq, maxra);
 375                         } else {
 376                                 kprintf("A-(%012jx,%d,%jd) "
 377                                         "doff=%012jx minr=%zd ra=%d\n",
 378                                     (intmax_t)rbp->b_loffset, rbp->b_bcount,
 379                                     (intmax_t)(loffset - origoffset),
 380                                     (intmax_t)doffset, minreq, maxra);
 381                         }
 382                 }
 383 #endif
 384                 rbp->b_flags &= ~(B_ERROR|B_INVAL);
 385
 386                 if ((rbp->b_flags & B_CLUSTER) == 0)
 387                         vfs_busy_pages(vp, rbp);
 388                 BUF_KERNPROC(rbp);
 389                 loffset += rbp->b_bufsize;
 390                 maxra -= rbp->b_bufsize / blksize;
 391                 vn_strategy(vp, &rbp->b_bio1);
 392                 /* rbp invalid now */
 393         }
 394
 395         /*
 396          * Wait for our original buffer to complete its I/O.  reqbp will
 397          * be NULL if the original buffer was B_CACHE.  We are returning
 398          * (*bpp) which is the same as reqbp when reqbp != NULL.
 399          */
 400 no_read_ahead:
 401         if (reqbp) {
 402                 KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
 403                 error = biowait(&reqbp->b_bio1, "clurd");
 404         }
 405         return (error);
 406 }
 407
 408 /*
 409  * If blocks are contiguous on disk, use this to provide clustered
 410  * read ahead.  We will read as many blocks as possible sequentially
 411  * and then parcel them up into logical blocks in the buffer hash table.
 412  *
 413  * This function either returns a cluster buf or it returns fbp.  fbp is
 414  * already expected to be set up as a synchronous or asynchronous request.
 415  *
 416  * If a cluster buf is returned it will always be async.
 417  */
 418 static struct buf *
 419 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
 420                int blksize, int run, struct buf *fbp)
 421 {
 422         struct buf *bp, *tbp;
 423         off_t boffset;
 424         int i, j;
 425         int maxiosize = vmaxiosize(vp);
 426
 427         /*
 428          * avoid a division
 429          */
 430         while (loffset + run * blksize > filesize) {
 431                 --run;
 432         }
 433
 434         tbp = fbp;
 435         tbp->b_bio2.bio_offset = doffset;
 436         if((tbp->b_flags & B_MALLOC) ||
 437             ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
 438                 return tbp;
 439         }
 440
 441         bp = trypbuf_kva(&cluster_pbuf_freecnt);
 442         if (bp == NULL) {
 443                 return tbp;
 444         }
 445
 446         /*
 447          * We are synthesizing a buffer out of vm_page_t's, but
 448          * if the block size is not page aligned then the starting
 449          * address may not be either.  Inherit the b_data offset
 450          * from the original buffer.
 451          */
 452         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 453             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 454         bp->b_flags |= B_CLUSTER | B_VMIO;
 455         bp->b_cmd = BUF_CMD_READ;
 456         bp->b_bio1.bio_done = cluster_callback;         /* default to async */
 457         bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 458         bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 459         bp->b_loffset = loffset;
 460         bp->b_bio2.bio_offset = doffset;
 461         KASSERT(bp->b_loffset != NOOFFSET,
 462                 ("cluster_rbuild: no buffer offset"));
 463
 464         bp->b_bcount = 0;
 465         bp->b_bufsize = 0;
 466         bp->b_xio.xio_npages = 0;
 467
 468         for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
 469                 if (i) {
 470                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
 471                             round_page(blksize) > maxiosize) {
 472                                 break;
 473                         }
 474
 475                         /*
 476                          * Shortcut some checks and try to avoid buffers that
 477                          * would block in the lock.  The same checks have to
 478                          * be made again after we officially get the buffer.
 479                          */
 480                         tbp = getblk(vp, loffset + i * blksize, blksize,
 481                                      GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
 482                         if (tbp == NULL)
 483                                 break;
 484                         for (j = 0; j < tbp->b_xio.xio_npages; j++) {
 485                                 if (tbp->b_xio.xio_pages[j]->valid)
 486                                         break;
 487                         }
 488                         if (j != tbp->b_xio.xio_npages) {
 489                                 bqrelse(tbp);
 490                                 break;
 491                         }
 492
 493                         /*
 494                          * Stop scanning if the buffer is fuly valid
 495                          * (marked B_CACHE), or locked (may be doing a
 496                          * background write), or if the buffer is not
 497                          * VMIO backed.  The clustering code can only deal
 498                          * with VMIO-backed buffers.
 499                          */
 500                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 501                             (tbp->b_flags & B_VMIO) == 0 ||
 502                             (LIST_FIRST(&tbp->b_dep) != NULL &&
 503                              buf_checkread(tbp))
 504                         ) {
 505                                 bqrelse(tbp);
 506                                 break;
 507                         }
 508
 509                         /*
 510                          * The buffer must be completely invalid in order to
 511                          * take part in the cluster.  If it is partially valid
 512                          * then we stop.
 513                          */
 514                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
 515                                 if (tbp->b_xio.xio_pages[j]->valid)
 516                                         break;
 517                         }
 518                         if (j != tbp->b_xio.xio_npages) {
 519                                 bqrelse(tbp);
 520                                 break;
 521                         }
 522
 523                         /*
 524                          * Set a read-ahead mark as appropriate.  Always
 525                          * set the read-ahead mark at (run - 1).  It is
 526                          * unclear why we were also setting it at i == 1.
 527                          */
 528                         if (/*i == 1 ||*/ i == (run - 1))
 529                                 cluster_setram(tbp);
 530
 531                         /*
 532                          * Depress the priority of buffers not explicitly
 533                          * requested.
 534                          */
 535                         /* tbp->b_flags |= B_AGE; */
 536
 537                         /*
 538                          * Set the block number if it isn't set, otherwise
 539                          * if it is make sure it matches the block number we
 540                          * expect.
 541                          */
 542                         if (tbp->b_bio2.bio_offset == NOOFFSET) {
 543                                 tbp->b_bio2.bio_offset = boffset;
 544                         } else if (tbp->b_bio2.bio_offset != boffset) {
 545                                 brelse(tbp);
 546                                 break;
 547                         }
 548                 }
 549
 550                 /*
 551                  * The passed-in tbp (i == 0) will already be set up for
 552                  * async or sync operation.  All other tbp's acquire in
 553                  * our loop are set up for async operation.
 554                  */
 555                 tbp->b_cmd = BUF_CMD_READ;
 556                 BUF_KERNPROC(tbp);
 557                 cluster_append(&bp->b_bio1, tbp);
 558                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 559                         vm_page_t m;
 560
 561                         m = tbp->b_xio.xio_pages[j];
 562                         vm_page_busy_wait(m, FALSE, "clurpg");
 563                         vm_page_io_start(m);
 564                         vm_page_wakeup(m);
 565                         vm_object_pip_add(m->object, 1);
 566                         if ((bp->b_xio.xio_npages == 0) ||
 567                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
 568                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 569                                 bp->b_xio.xio_npages++;
 570                         }
 571                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 572                                 tbp->b_xio.xio_pages[j] = bogus_page;
 573                 }
 574                 /*
 575                  * XXX shouldn't this be += size for both, like in
 576                  * cluster_wbuild()?
 577                  *
 578                  * Don't inherit tbp->b_bufsize as it may be larger due to
 579                  * a non-page-aligned size.  Instead just aggregate using
 580                  * 'size'.
 581                  */
 582                 if (tbp->b_bcount != blksize)
 583                     kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
 584                 if (tbp->b_bufsize != blksize)
 585                     kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
 586                 bp->b_bcount += blksize;
 587                 bp->b_bufsize += blksize;
 588         }
 589
 590         /*
 591          * Fully valid pages in the cluster are already good and do not need
 592          * to be re-read from disk.  Replace the page with bogus_page
 593          */
 594         for (j = 0; j < bp->b_xio.xio_npages; j++) {
 595                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 596                     VM_PAGE_BITS_ALL) {
 597                         bp->b_xio.xio_pages[j] = bogus_page;
 598                 }
 599         }
 600         if (bp->b_bufsize > bp->b_kvasize) {
 601                 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
 602                     bp->b_bufsize, bp->b_kvasize);
 603         }
 604         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 605                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 606         BUF_KERNPROC(bp);
 607         return (bp);
 608 }
 609
 610 /*
 611  * Cleanup after a clustered read or write.
 612  * This is complicated by the fact that any of the buffers might have
 613  * extra memory (if there were no empty buffer headers at allocbuf time)
 614  * that we will need to shift around.
 615  *
 616  * The returned bio is &bp->b_bio1
 617  */
 618 void
 619 cluster_callback(struct bio *bio)
 620 {
 621         struct buf *bp = bio->bio_buf;
 622         struct buf *tbp;
 623         int error = 0;
 624
 625         /*
 626          * Must propogate errors to all the components.  A short read (EOF)
 627          * is a critical error.
 628          */
 629         if (bp->b_flags & B_ERROR) {
 630                 error = bp->b_error;
 631         } else if (bp->b_bcount != bp->b_bufsize) {
 632                 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
 633         }
 634
 635         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
 636         /*
 637          * Move memory from the large cluster buffer into the component
 638          * buffers and mark IO as done on these.  Since the memory map
 639          * is the same, no actual copying is required.
 640          */
 641         while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
 642                 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
 643                 if (error) {
 644                         tbp->b_flags |= B_ERROR | B_IODEBUG;
 645                         tbp->b_error = error;
 646                 } else {
 647                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 648                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 649                         tbp->b_flags |= B_IODEBUG;
 650                         /*
 651                          * XXX the bdwrite()/bqrelse() issued during
 652                          * cluster building clears B_RELBUF (see bqrelse()
 653                          * comment).  If direct I/O was specified, we have
 654                          * to restore it here to allow the buffer and VM
 655                          * to be freed.
 656                          */
 657                         if (tbp->b_flags & B_DIRECT)
 658                                 tbp->b_flags |= B_RELBUF;
 659                 }
 660                 biodone(&tbp->b_bio1);
 661         }
 662         relpbuf(bp, &cluster_pbuf_freecnt);
 663 }
 664
 665 /*
 666  *      cluster_wbuild_wb:
 667  *
 668  *      Implement modified write build for cluster.
 669  *
 670  *              write_behind = 0        write behind disabled
 671  *              write_behind = 1        write behind normal (default)
 672  *              write_behind = 2        write behind backed-off
 673  */
 674
 675 static __inline int
 676 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
 677 {
 678         int r = 0;
 679
 680         switch(write_behind) {
 681         case 2:
 682                 if (start_loffset < len)
 683                         break;
 684                 start_loffset -= len;
 685                 /* fall through */
 686         case 1:
 687                 r = cluster_wbuild(vp, blksize, start_loffset, len);
 688                 /* fall through */
 689         default:
 690                 /* fall through */
 691                 break;
 692         }
 693         return(r);
 694 }
 695
 696 /*
 697  * Do clustered write for FFS.
 698  *
 699  * Three cases:
 700  *      1. Write is not sequential (write asynchronously)
 701  *      Write is sequential:
 702  *      2.      beginning of cluster - begin cluster
 703  *      3.      middle of a cluster - add to cluster
 704  *      4.      end of a cluster - asynchronously write cluster
 705  */
 706 void
 707 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
 708 {
 709         struct vnode *vp;
 710         off_t loffset;
 711         int maxclen, cursize;
 712         int async;
 713
 714         vp = bp->b_vp;
 715         if (vp->v_type == VREG)
 716                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 717         else
 718                 async = 0;
 719         loffset = bp->b_loffset;
 720         KASSERT(bp->b_loffset != NOOFFSET,
 721                 ("cluster_write: no buffer offset"));
 722
 723         /* Initialize vnode to beginning of file. */
 724         if (loffset == 0)
 725                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 726
 727         if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
 728             bp->b_bio2.bio_offset == NOOFFSET ||
 729             (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
 730                 maxclen = vmaxiosize(vp);
 731                 if (vp->v_clen != 0) {
 732                         /*
 733                          * Next block is not sequential.
 734                          *
 735                          * If we are not writing at end of file, the process
 736                          * seeked to another point in the file since its last
 737                          * write, or we have reached our maximum cluster size,
 738                          * then push the previous cluster. Otherwise try
 739                          * reallocating to make it sequential.
 740                          *
 741                          * Change to algorithm: only push previous cluster if
 742                          * it was sequential from the point of view of the
 743                          * seqcount heuristic, otherwise leave the buffer
 744                          * intact so we can potentially optimize the I/O
 745                          * later on in the buf_daemon or update daemon
 746                          * flush.
 747                          */
 748                         cursize = vp->v_lastw - vp->v_cstart + blksize;
 749                         if (bp->b_loffset + blksize != filesize ||
 750                             loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
 751                                 if (!async && seqcount > 0) {
 752                                         cluster_wbuild_wb(vp, blksize,
 753                                                 vp->v_cstart, cursize);
 754                                 }
 755                         } else {
 756                                 struct buf **bpp, **endbp;
 757                                 struct cluster_save *buflist;
 758
 759                                 buflist = cluster_collectbufs(vp, bp, blksize);
 760                                 endbp = &buflist->bs_children
 761                                     [buflist->bs_nchildren - 1];
 762                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 763                                         /*
 764                                          * Failed, push the previous cluster
 765                                          * if *really* writing sequentially
 766                                          * in the logical file (seqcount > 1),
 767                                          * otherwise delay it in the hopes that
 768                                          * the low level disk driver can
 769                                          * optimize the write ordering.
 770                                          */
 771                                         for (bpp = buflist->bs_children;
 772                                              bpp < endbp; bpp++)
 773                                                 brelse(*bpp);
 774                                         kfree(buflist, M_SEGMENT);
 775                                         if (seqcount > 1) {
 776                                                 cluster_wbuild_wb(vp,
 777                                                     blksize, vp->v_cstart,
 778                                                     cursize);
 779                                         }
 780                                 } else {
 781                                         /*
 782                                          * Succeeded, keep building cluster.
 783                                          */
 784                                         for (bpp = buflist->bs_children;
 785                                              bpp <= endbp; bpp++)
 786                                                 bdwrite(*bpp);
 787                                         kfree(buflist, M_SEGMENT);
 788                                         vp->v_lastw = loffset;
 789                                         vp->v_lasta = bp->b_bio2.bio_offset;
 790                                         return;
 791                                 }
 792                         }
 793                 }
 794                 /*
 795                  * Consider beginning a cluster. If at end of file, make
 796                  * cluster as large as possible, otherwise find size of
 797                  * existing cluster.
 798                  */
 799                 if ((vp->v_type == VREG) &&
 800                     bp->b_loffset + blksize != filesize &&
 801                     (bp->b_bio2.bio_offset == NOOFFSET) &&
 802                     (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
 803                      bp->b_bio2.bio_offset == NOOFFSET)) {
 804                         bawrite(bp);
 805                         vp->v_clen = 0;
 806                         vp->v_lasta = bp->b_bio2.bio_offset;
 807                         vp->v_cstart = loffset + blksize;
 808                         vp->v_lastw = loffset;
 809                         return;
 810                 }
 811                 if (maxclen > blksize)
 812                         vp->v_clen = maxclen - blksize;
 813                 else
 814                         vp->v_clen = 0;
 815                 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
 816                         vp->v_cstart = loffset + blksize;
 817                         bawrite(bp);
 818                 } else {        /* Wait for rest of cluster */
 819                         vp->v_cstart = loffset;
 820                         bdwrite(bp);
 821                 }
 822         } else if (loffset == vp->v_cstart + vp->v_clen) {
 823                 /*
 824                  * At end of cluster, write it out if seqcount tells us we
 825                  * are operating sequentially, otherwise let the buf or
 826                  * update daemon handle it.
 827                  */
 828                 bdwrite(bp);
 829                 if (seqcount > 1)
 830                         cluster_wbuild_wb(vp, blksize, vp->v_cstart,
 831                                           vp->v_clen + blksize);
 832                 vp->v_clen = 0;
 833                 vp->v_cstart = loffset + blksize;
 834         } else if (vm_page_count_severe()) {
 835                 /*
 836                  * We are low on memory, get it going NOW
 837                  */
 838                 bawrite(bp);
 839         } else {
 840                 /*
 841                  * In the middle of a cluster, so just delay the I/O for now.
 842                  */
 843                 bdwrite(bp);
 844         }
 845         vp->v_lastw = loffset;
 846         vp->v_lasta = bp->b_bio2.bio_offset;
 847 }
 848
 849
 850 /*
 851  * This is an awful lot like cluster_rbuild...wish they could be combined.
 852  * The last lbn argument is the current block on which I/O is being
 853  * performed.  Check to see that it doesn't fall in the middle of
 854  * the current block (if last_bp == NULL).
 855  */
 856 int
 857 cluster_wbuild(struct vnode *vp, int blksize, off_t start_loffset, int bytes)
 858 {
 859         struct buf *bp, *tbp;
 860         int i, j;
 861         int totalwritten = 0;
 862         int maxiosize = vmaxiosize(vp);
 863
 864         while (bytes > 0) {
 865                 /*
 866                  * If the buffer is not delayed-write (i.e. dirty), or it
 867                  * is delayed-write but either locked or inval, it cannot
 868                  * partake in the clustered write.
 869                  */
 870                 tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK);
 871                 if (tbp == NULL ||
 872                     (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI ||
 873                     (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
 874                         if (tbp)
 875                                 BUF_UNLOCK(tbp);
 876                         start_loffset += blksize;
 877                         bytes -= blksize;
 878                         continue;
 879                 }
 880                 bremfree(tbp);
 881                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 882
 883                 /*
 884                  * Extra memory in the buffer, punt on this buffer.
 885                  * XXX we could handle this in most cases, but we would
 886                  * have to push the extra memory down to after our max
 887                  * possible cluster size and then potentially pull it back
 888                  * up if the cluster was terminated prematurely--too much
 889                  * hassle.
 890                  */
 891                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 892                     (tbp->b_bcount != tbp->b_bufsize) ||
 893                     (tbp->b_bcount != blksize) ||
 894                     (bytes == blksize) ||
 895                     ((bp = getpbuf_kva(&cluster_pbuf_freecnt)) == NULL)) {
 896                         totalwritten += tbp->b_bufsize;
 897                         bawrite(tbp);
 898                         start_loffset += blksize;
 899                         bytes -= blksize;
 900                         continue;
 901                 }
 902
 903                 /*
 904                  * Set up the pbuf.  Track our append point with b_bcount
 905                  * and b_bufsize.  b_bufsize is not used by the device but
 906                  * our caller uses it to loop clusters and we use it to
 907                  * detect a premature EOF on the block device.
 908                  */
 909                 bp->b_bcount = 0;
 910                 bp->b_bufsize = 0;
 911                 bp->b_xio.xio_npages = 0;
 912                 bp->b_loffset = tbp->b_loffset;
 913                 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
 914
 915                 /*
 916                  * We are synthesizing a buffer out of vm_page_t's, but
 917                  * if the block size is not page aligned then the starting
 918                  * address may not be either.  Inherit the b_data offset
 919                  * from the original buffer.
 920                  */
 921                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 922                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 923                 bp->b_flags &= ~B_ERROR;
 924                 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
 925                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
 926                 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 927                 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 928
 929                 /*
 930                  * From this location in the file, scan forward to see
 931                  * if there are buffers with adjacent data that need to
 932                  * be written as well.
 933                  */
 934                 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
 935                         if (i != 0) { /* If not the first buffer */
 936                                 tbp = findblk(vp, start_loffset,
 937                                               FINDBLK_NBLOCK);
 938                                 /*
 939                                  * Buffer not found or could not be locked
 940                                  * non-blocking.
 941                                  */
 942                                 if (tbp == NULL)
 943                                         break;
 944
 945                                 /*
 946                                  * If it IS in core, but has different
 947                                  * characteristics, then don't cluster
 948                                  * with it.
 949                                  */
 950                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 951                                      B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 952                                     != (B_DELWRI | B_CLUSTEROK |
 953                                      (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 954                                     (tbp->b_flags & B_LOCKED) ||
 955                                     (LIST_FIRST(&tbp->b_dep) &&
 956                                      buf_checkwrite(tbp))
 957                                 ) {
 958                                         BUF_UNLOCK(tbp);
 959                                         break;
 960                                 }
 961
 962                                 /*
 963                                  * Check that the combined cluster
 964                                  * would make sense with regard to pages
 965                                  * and would not be too large
 966                                  */
 967                                 if ((tbp->b_bcount != blksize) ||
 968                                   ((bp->b_bio2.bio_offset + i) !=
 969                                     tbp->b_bio2.bio_offset) ||
 970                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
 971                                     (maxiosize / PAGE_SIZE))) {
 972                                         BUF_UNLOCK(tbp);
 973                                         break;
 974                                 }
 975                                 /*
 976                                  * Ok, it's passed all the tests,
 977                                  * so remove it from the free list
 978                                  * and mark it busy. We will use it.
 979                                  */
 980                                 bremfree(tbp);
 981                                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 982                         } /* end of code for non-first buffers only */
 983
 984                         /*
 985                          * If the IO is via the VM then we do some
 986                          * special VM hackery (yuck).  Since the buffer's
 987                          * block size may not be page-aligned it is possible
 988                          * for a page to be shared between two buffers.  We
 989                          * have to get rid of the duplication when building
 990                          * the cluster.
 991                          */
 992                         if (tbp->b_flags & B_VMIO) {
 993                                 vm_page_t m;
 994
 995                                 if (i != 0) { /* if not first buffer */
 996                                         for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 997                                                 m = tbp->b_xio.xio_pages[j];
 998                                                 if (m->flags & PG_BUSY) {
 999                                                         bqrelse(tbp);
1000                                                         goto finishcluster;
1001                                                 }
1002                                         }
1003                                 }
1004
1005                                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
1006                                         m = tbp->b_xio.xio_pages[j];
1007                                         vm_page_busy_wait(m, FALSE, "clurpg");
1008                                         vm_page_io_start(m);
1009                                         vm_page_wakeup(m);
1010                                         vm_object_pip_add(m->object, 1);
1011                                         if ((bp->b_xio.xio_npages == 0) ||
1012                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
1013                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
1014                                                 bp->b_xio.xio_npages++;
1015                                         }
1016                                 }
1017                         }
1018                         bp->b_bcount += blksize;
1019                         bp->b_bufsize += blksize;
1020
1021                         bundirty(tbp);
1022                         tbp->b_flags &= ~B_ERROR;
1023                         tbp->b_cmd = BUF_CMD_WRITE;
1024                         BUF_KERNPROC(tbp);
1025                         cluster_append(&bp->b_bio1, tbp);
1026
1027                         /*
1028                          * check for latent dependencies to be handled
1029                          */
1030                         if (LIST_FIRST(&tbp->b_dep) != NULL)
1031                                 buf_start(tbp);
1032                 }
1033         finishcluster:
1034                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1035                         (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1036                 if (bp->b_bufsize > bp->b_kvasize) {
1037                         panic(
1038                             "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
1039                             bp->b_bufsize, bp->b_kvasize);
1040                 }
1041                 totalwritten += bp->b_bufsize;
1042                 bp->b_dirtyoff = 0;
1043                 bp->b_dirtyend = bp->b_bufsize;
1044                 bp->b_bio1.bio_done = cluster_callback;
1045                 bp->b_cmd = BUF_CMD_WRITE;
1046
1047                 vfs_busy_pages(vp, bp);
1048                 bsetrunningbufspace(bp, bp->b_bufsize);
1049                 BUF_KERNPROC(bp);
1050                 vn_strategy(vp, &bp->b_bio1);
1051
1052                 bytes -= i;
1053         }
1054         return totalwritten;
1055 }
1056
1057 /*
1058  * Collect together all the buffers in a cluster.
1059  * Plus add one additional buffer.
1060  */
1061 static struct cluster_save *
1062 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
1063 {
1064         struct cluster_save *buflist;
1065         struct buf *bp;
1066         off_t loffset;
1067         int i, len;
1068
1069         len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
1070         buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1071                          M_SEGMENT, M_WAITOK);
1072         buflist->bs_nchildren = 0;
1073         buflist->bs_children = (struct buf **) (buflist + 1);
1074         for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
1075                 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
1076                 buflist->bs_children[i] = bp;
1077                 if (bp->b_bio2.bio_offset == NOOFFSET) {
1078                         VOP_BMAP(bp->b_vp, bp->b_loffset,
1079                                  &bp->b_bio2.bio_offset,
1080                                  NULL, NULL, BUF_CMD_WRITE);
1081                 }
1082         }
1083         buflist->bs_children[i] = bp = last_bp;
1084         if (bp->b_bio2.bio_offset == NOOFFSET) {
1085                 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1086                          NULL, NULL, BUF_CMD_WRITE);
1087         }
1088         buflist->bs_nchildren = i + 1;
1089         return (buflist);
1090 }
1091
1092 void
1093 cluster_append(struct bio *bio, struct buf *tbp)
1094 {
1095         tbp->b_cluster_next = NULL;
1096         if (bio->bio_caller_info1.cluster_head == NULL) {
1097                 bio->bio_caller_info1.cluster_head = tbp;
1098                 bio->bio_caller_info2.cluster_tail = tbp;
1099         } else {
1100                 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1101                 bio->bio_caller_info2.cluster_tail = tbp;
1102         }
1103 }
1104
1105 static
1106 void
1107 cluster_setram (struct buf *bp)
1108 {
1109         bp->b_flags |= B_RAM;
1110         if (bp->b_xio.xio_npages)
1111                 vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM);
1112 }