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