sys/kern/vfs_cluster.c

   1 /*-
   2  * Copyright (c) 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  * Modifications/enhancements:
   5  *      Copyright (c) 1995 John S. Dyson.  All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. All advertising materials mentioning features or use of this software
  16  *    must display the following acknowledgement:
  17  *      This product includes software developed by the University of
  18  *      California, Berkeley and its contributors.
  19  * 4. Neither the name of the University nor the names of its contributors
  20  *    may be used to endorse or promote products derived from this software
  21  *    without specific prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  *
  35  *      @(#)vfs_cluster.c       8.7 (Berkeley) 2/13/94
  36  * $FreeBSD: src/sys/kern/vfs_cluster.c,v 1.92.2.9 2001/11/18 07:10:59 dillon Exp $
  37  * $DragonFly: src/sys/kern/vfs_cluster.c,v 1.40 2008/07/14 03:09:00 dillon Exp $
  38  */
  39
  40 #include "opt_debug_cluster.h"
  41
  42 #include <sys/param.h>
  43 #include <sys/systm.h>
  44 #include <sys/kernel.h>
  45 #include <sys/proc.h>
  46 #include <sys/buf.h>
  47 #include <sys/vnode.h>
  48 #include <sys/malloc.h>
  49 #include <sys/mount.h>
  50 #include <sys/resourcevar.h>
  51 #include <sys/vmmeter.h>
  52 #include <vm/vm.h>
  53 #include <vm/vm_object.h>
  54 #include <vm/vm_page.h>
  55 #include <sys/sysctl.h>
  56 #include <sys/buf2.h>
  57 #include <vm/vm_page2.h>
  58
  59 #include <machine/limits.h>
  60
  61 #define CLUSTERDEBUG
  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 static int max_readahead = 2 * 1024 * 1024;
  83 SYSCTL_INT(_vfs, OID_AUTO, max_readahead, CTLFLAG_RW, &max_readahead, 0, "");
  84
  85 extern vm_page_t        bogus_page;
  86
  87 extern int cluster_pbuf_freecnt;
  88
  89 /*
  90  * This replaces bread.
  91  *
  92  * filesize     - read-ahead @ blksize will not cross this boundary
  93  * loffset      - loffset for returned *bpp
  94  * blksize      - blocksize for returned *bpp and read-ahead bps
  95  * minreq       - minimum (not a hard minimum) in bytes, typically reflects
  96  *                a higher level uio resid.
  97  * maxreq       - maximum (sequential heuristic) in bytes (highet typ ~2MB)
  98  * bpp          - return buffer (*bpp) for (loffset,blksize)
  99  */
 100 int
 101 cluster_read(struct vnode *vp, off_t filesize, off_t loffset,
 102              int blksize, size_t minreq, size_t maxreq, struct buf **bpp)
 103 {
 104         struct buf *bp, *rbp, *reqbp;
 105         off_t origoffset;
 106         off_t doffset;
 107         int error;
 108         int i;
 109         int maxra;
 110         int maxrbuild;
 111
 112         error = 0;
 113
 114         /*
 115          * Calculate the desired read-ahead in blksize'd blocks (maxra).
 116          * To do this we calculate maxreq.
 117          *
 118          * maxreq typically starts out as a sequential heuristic.  If the
 119          * high level uio/resid is bigger (minreq), we pop maxreq up to
 120          * minreq.  This represents the case where random I/O is being
 121          * performed by the userland is issuing big read()'s.
 122          *
 123          * Then we limit maxreq to max_readahead to ensure it is a reasonable
 124          * value.
 125          *
 126          * Finally we must ensure that loffset + maxreq does not cross the
 127          * boundary (filesize) for the current blocksize.  If we allowed it
 128          * to cross we could end up with buffers past the boundary with the
 129          * wrong block size (HAMMER large-data areas use mixed block sizes).
 130          */
 131         if (maxreq < minreq)
 132                 maxreq = minreq;
 133         if (maxreq > max_readahead) {
 134                 maxreq = max_readahead;
 135                 if (maxreq > 16 * 1024 * 1024)
 136                         maxreq = 16 * 1024 * 1024;
 137         }
 138         if (maxreq < blksize)
 139                 maxreq = blksize;
 140         if (loffset + maxreq > filesize) {
 141                 if (loffset > filesize)
 142                         maxreq = 0;
 143                 else
 144                         maxreq = filesize - loffset;
 145         }
 146
 147         maxra = (int)(maxreq / blksize);
 148
 149         /*
 150          * Get the requested block.
 151          */
 152         *bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0);
 153         origoffset = loffset;
 154
 155         /*
 156          * Calculate the maximum cluster size for a single I/O, used
 157          * by cluster_rbuild().
 158          */
 159         maxrbuild = vmaxiosize(vp) / blksize;
 160
 161         /*
 162          * if it is in the cache, then check to see if the reads have been
 163          * sequential.  If they have, then try some read-ahead, otherwise
 164          * back-off on prospective read-aheads.
 165          */
 166         if (bp->b_flags & B_CACHE) {
 167                 /*
 168                  * Not sequential, do not do any read-ahead
 169                  */
 170                 if (maxra <= 1)
 171                         return 0;
 172
 173                 /*
 174                  * No read-ahead mark, do not do any read-ahead
 175                  * yet.
 176                  */
 177                 if ((bp->b_flags & B_RAM) == 0)
 178                         return 0;
 179
 180                 /*
 181                  * We hit a read-ahead-mark, figure out how much read-ahead
 182                  * to do (maxra) and where to start (loffset).
 183                  *
 184                  * Shortcut the scan.  Typically the way this works is that
 185                  * we've built up all the blocks inbetween except for the
 186                  * last in previous iterations, so if the second-to-last
 187                  * block is present we just skip ahead to it.
 188                  *
 189                  * This algorithm has O(1) cpu in the steady state no
 190                  * matter how large maxra is.
 191                  */
 192                 bp->b_flags &= ~B_RAM;
 193
 194                 if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST))
 195                         i = maxra - 1;
 196                 else
 197                         i = 1;
 198                 while (i < maxra) {
 199                         if (findblk(vp, loffset + i * blksize,
 200                                     FINDBLK_TEST) == NULL) {
 201                                 break;
 202                         }
 203                         ++i;
 204                 }
 205
 206                 /*
 207                  * We got everything or everything is in the cache, no
 208                  * point continuing.
 209                  */
 210                 if (i >= maxra)
 211                         return 0;
 212                 maxra -= i;
 213                 loffset += i * blksize;
 214                 reqbp = bp = NULL;
 215         } else {
 216                 off_t firstread = bp->b_loffset;
 217                 int nblks;
 218
 219                 /*
 220                  * Set-up synchronous read for bp.
 221                  */
 222                 bp->b_cmd = BUF_CMD_READ;
 223                 bp->b_bio1.bio_done = biodone_sync;
 224                 bp->b_bio1.bio_flags |= BIO_SYNC;
 225
 226                 KASSERT(firstread != NOOFFSET,
 227                         ("cluster_read: no buffer offset"));
 228
 229                 /*
 230                  * nblks is our cluster_rbuild request size, limited
 231                  * primarily by the device.
 232                  */
 233                 if ((nblks = maxra) > maxrbuild)
 234                         nblks = maxrbuild;
 235
 236                 if (nblks > 1) {
 237                         int burstbytes;
 238
 239                         error = VOP_BMAP(vp, loffset, &doffset,
 240                                          &burstbytes, NULL, BUF_CMD_READ);
 241                         if (error)
 242                                 goto single_block_read;
 243                         if (nblks > burstbytes / blksize)
 244                                 nblks = burstbytes / blksize;
 245                         if (doffset == NOOFFSET)
 246                                 goto single_block_read;
 247                         if (nblks <= 1)
 248                                 goto single_block_read;
 249
 250                         bp = cluster_rbuild(vp, filesize, loffset,
 251                                             doffset, blksize, nblks, bp);
 252                         loffset += bp->b_bufsize;
 253                         maxra -= bp->b_bufsize / blksize;
 254                 } else {
 255 single_block_read:
 256                         /*
 257                          * If it isn't in the cache, then get a chunk from
 258                          * disk if sequential, otherwise just get the block.
 259                          */
 260                         cluster_setram(bp);
 261                         loffset += blksize;
 262                         --maxra;
 263                 }
 264         }
 265
 266         /*
 267          * If B_CACHE was not set issue bp.  bp will either be an
 268          * asynchronous cluster buf or a synchronous single-buf.
 269          * If it is a single buf it will be the same as reqbp.
 270          *
 271          * NOTE: Once an async cluster buf is issued bp becomes invalid.
 272          */
 273         if (bp) {
 274 #if defined(CLUSTERDEBUG)
 275                 if (rcluster)
 276                         kprintf("S(%012jx,%d,%d)\n",
 277                             (intmax_t)bp->b_loffset, bp->b_bcount, maxra);
 278 #endif
 279                 if ((bp->b_flags & B_CLUSTER) == 0)
 280                         vfs_busy_pages(vp, bp);
 281                 bp->b_flags &= ~(B_ERROR|B_INVAL);
 282                 vn_strategy(vp, &bp->b_bio1);
 283                 error = 0;
 284                 /* bp invalid now */
 285         }
 286
 287         /*
 288          * If we have been doing sequential I/O, then do some read-ahead.
 289          * The code above us should have positioned us at the next likely
 290          * offset.
 291          *
 292          * Only mess with buffers which we can immediately lock.  HAMMER
 293          * will do device-readahead irrespective of what the blocks
 294          * represent.
 295          */
 296         while (error == 0 && maxra > 0) {
 297                 int burstbytes;
 298                 int tmp_error;
 299                 int nblks;
 300
 301                 rbp = getblk(vp, loffset, blksize,
 302                              GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
 303                 if (rbp == NULL)
 304                         goto no_read_ahead;
 305                 if ((rbp->b_flags & B_CACHE)) {
 306                         bqrelse(rbp);
 307                         goto no_read_ahead;
 308                 }
 309
 310                 /*
 311                  * An error from the read-ahead bmap has nothing to do
 312                  * with the caller's original request.
 313                  */
 314                 tmp_error = VOP_BMAP(vp, loffset, &doffset,
 315                                      &burstbytes, NULL, BUF_CMD_READ);
 316                 if (tmp_error || doffset == NOOFFSET) {
 317                         rbp->b_flags |= B_INVAL;
 318                         brelse(rbp);
 319                         rbp = NULL;
 320                         goto no_read_ahead;
 321                 }
 322                 if ((nblks = maxra) > maxrbuild)
 323                         nblks = maxrbuild;
 324                 if (nblks > burstbytes / blksize)
 325                         nblks = burstbytes / blksize;
 326
 327                 /*
 328                  * rbp: async read
 329                  */
 330                 rbp->b_cmd = BUF_CMD_READ;
 331                 /*rbp->b_flags |= B_AGE*/;
 332                 cluster_setram(rbp);
 333
 334                 if (nblks > 1) {
 335                         rbp = cluster_rbuild(vp, filesize, loffset,
 336                                              doffset, blksize,
 337                                              nblks, rbp);
 338                 } else {
 339                         rbp->b_bio2.bio_offset = doffset;
 340                 }
 341
 342 #if defined(CLUSTERDEBUG)
 343                 if (rcluster) {
 344                         if (bp) {
 345                                 kprintf("A+(%012jx,%d,%jd) "
 346                                         "doff=%012jx minr=%zd ra=%d\n",
 347                                     (intmax_t)loffset, rbp->b_bcount,
 348                                     (intmax_t)(loffset - origoffset),
 349                                     (intmax_t)doffset, minreq, maxra);
 350                         } else {
 351                                 kprintf("A-(%012jx,%d,%jd) "
 352                                         "doff=%012jx minr=%zd ra=%d\n",
 353                                     (intmax_t)rbp->b_loffset, rbp->b_bcount,
 354                                     (intmax_t)(loffset - origoffset),
 355                                     (intmax_t)doffset, minreq, maxra);
 356                         }
 357                 }
 358 #endif
 359                 rbp->b_flags &= ~(B_ERROR|B_INVAL);
 360
 361                 if ((rbp->b_flags & B_CLUSTER) == 0)
 362                         vfs_busy_pages(vp, rbp);
 363                 BUF_KERNPROC(rbp);
 364                 loffset += rbp->b_bufsize;
 365                 maxra -= rbp->b_bufsize / blksize;
 366                 vn_strategy(vp, &rbp->b_bio1);
 367                 /* rbp invalid now */
 368         }
 369
 370         /*
 371          * Wait for our original buffer to complete its I/O.  reqbp will
 372          * be NULL if the original buffer was B_CACHE.  We are returning
 373          * (*bpp) which is the same as reqbp when reqbp != NULL.
 374          */
 375 no_read_ahead:
 376         if (reqbp) {
 377                 KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
 378                 error = biowait(&reqbp->b_bio1, "clurd");
 379         }
 380         return (error);
 381 }
 382
 383 /*
 384  * If blocks are contiguous on disk, use this to provide clustered
 385  * read ahead.  We will read as many blocks as possible sequentially
 386  * and then parcel them up into logical blocks in the buffer hash table.
 387  *
 388  * This function either returns a cluster buf or it returns fbp.  fbp is
 389  * already expected to be set up as a synchronous or asynchronous request.
 390  *
 391  * If a cluster buf is returned it will always be async.
 392  */
 393 static struct buf *
 394 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
 395                int blksize, int run, struct buf *fbp)
 396 {
 397         struct buf *bp, *tbp;
 398         off_t boffset;
 399         int i, j;
 400         int maxiosize = vmaxiosize(vp);
 401
 402         /*
 403          * avoid a division
 404          */
 405         while (loffset + run * blksize > filesize) {
 406                 --run;
 407         }
 408
 409         tbp = fbp;
 410         tbp->b_bio2.bio_offset = doffset;
 411         if((tbp->b_flags & B_MALLOC) ||
 412             ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
 413                 return tbp;
 414         }
 415
 416         bp = trypbuf_kva(&cluster_pbuf_freecnt);
 417         if (bp == NULL) {
 418                 return tbp;
 419         }
 420
 421         /*
 422          * We are synthesizing a buffer out of vm_page_t's, but
 423          * if the block size is not page aligned then the starting
 424          * address may not be either.  Inherit the b_data offset
 425          * from the original buffer.
 426          */
 427         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 428             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 429         bp->b_flags |= B_CLUSTER | B_VMIO;
 430         bp->b_cmd = BUF_CMD_READ;
 431         bp->b_bio1.bio_done = cluster_callback;         /* default to async */
 432         bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 433         bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 434         bp->b_loffset = loffset;
 435         bp->b_bio2.bio_offset = doffset;
 436         KASSERT(bp->b_loffset != NOOFFSET,
 437                 ("cluster_rbuild: no buffer offset"));
 438
 439         bp->b_bcount = 0;
 440         bp->b_bufsize = 0;
 441         bp->b_xio.xio_npages = 0;
 442
 443         for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
 444                 if (i) {
 445                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
 446                             round_page(blksize) > maxiosize) {
 447                                 break;
 448                         }
 449
 450                         /*
 451                          * Shortcut some checks and try to avoid buffers that
 452                          * would block in the lock.  The same checks have to
 453                          * be made again after we officially get the buffer.
 454                          */
 455                         tbp = getblk(vp, loffset + i * blksize, blksize,
 456                                      GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
 457                         if (tbp == NULL)
 458                                 break;
 459                         for (j = 0; j < tbp->b_xio.xio_npages; j++) {
 460                                 if (tbp->b_xio.xio_pages[j]->valid)
 461                                         break;
 462                         }
 463                         if (j != tbp->b_xio.xio_npages) {
 464                                 bqrelse(tbp);
 465                                 break;
 466                         }
 467
 468                         /*
 469                          * Stop scanning if the buffer is fuly valid
 470                          * (marked B_CACHE), or locked (may be doing a
 471                          * background write), or if the buffer is not
 472                          * VMIO backed.  The clustering code can only deal
 473                          * with VMIO-backed buffers.
 474                          */
 475                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 476                             (tbp->b_flags & B_VMIO) == 0 ||
 477                             (LIST_FIRST(&tbp->b_dep) != NULL &&
 478                              buf_checkread(tbp))
 479                         ) {
 480                                 bqrelse(tbp);
 481                                 break;
 482                         }
 483
 484                         /*
 485                          * The buffer must be completely invalid in order to
 486                          * take part in the cluster.  If it is partially valid
 487                          * then we stop.
 488                          */
 489                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
 490                                 if (tbp->b_xio.xio_pages[j]->valid)
 491                                         break;
 492                         }
 493                         if (j != tbp->b_xio.xio_npages) {
 494                                 bqrelse(tbp);
 495                                 break;
 496                         }
 497
 498                         /*
 499                          * Set a read-ahead mark as appropriate
 500                          */
 501                         if (i == 1 || i == (run - 1))
 502                                 cluster_setram(tbp);
 503
 504                         /*
 505                          * Depress the priority of buffers not explicitly
 506                          * requested.
 507                          */
 508                         /* tbp->b_flags |= B_AGE; */
 509
 510                         /*
 511                          * Set the block number if it isn't set, otherwise
 512                          * if it is make sure it matches the block number we
 513                          * expect.
 514                          */
 515                         if (tbp->b_bio2.bio_offset == NOOFFSET) {
 516                                 tbp->b_bio2.bio_offset = boffset;
 517                         } else if (tbp->b_bio2.bio_offset != boffset) {
 518                                 brelse(tbp);
 519                                 break;
 520                         }
 521                 }
 522
 523                 /*
 524                  * The passed-in tbp (i == 0) will already be set up for
 525                  * async or sync operation.  All other tbp's acquire in
 526                  * our loop are set up for async operation.
 527                  */
 528                 tbp->b_cmd = BUF_CMD_READ;
 529                 BUF_KERNPROC(tbp);
 530                 cluster_append(&bp->b_bio1, tbp);
 531                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 532                         vm_page_t m;
 533                         m = tbp->b_xio.xio_pages[j];
 534                         vm_page_io_start(m);
 535                         vm_object_pip_add(m->object, 1);
 536                         if ((bp->b_xio.xio_npages == 0) ||
 537                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
 538                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 539                                 bp->b_xio.xio_npages++;
 540                         }
 541                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 542                                 tbp->b_xio.xio_pages[j] = bogus_page;
 543                 }
 544                 /*
 545                  * XXX shouldn't this be += size for both, like in
 546                  * cluster_wbuild()?
 547                  *
 548                  * Don't inherit tbp->b_bufsize as it may be larger due to
 549                  * a non-page-aligned size.  Instead just aggregate using
 550                  * 'size'.
 551                  */
 552                 if (tbp->b_bcount != blksize)
 553                     kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
 554                 if (tbp->b_bufsize != blksize)
 555                     kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
 556                 bp->b_bcount += blksize;
 557                 bp->b_bufsize += blksize;
 558         }
 559
 560         /*
 561          * Fully valid pages in the cluster are already good and do not need
 562          * to be re-read from disk.  Replace the page with bogus_page
 563          */
 564         for (j = 0; j < bp->b_xio.xio_npages; j++) {
 565                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 566                     VM_PAGE_BITS_ALL) {
 567                         bp->b_xio.xio_pages[j] = bogus_page;
 568                 }
 569         }
 570         if (bp->b_bufsize > bp->b_kvasize) {
 571                 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
 572                     bp->b_bufsize, bp->b_kvasize);
 573         }
 574         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 575                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 576         BUF_KERNPROC(bp);
 577         return (bp);
 578 }
 579
 580 /*
 581  * Cleanup after a clustered read or write.
 582  * This is complicated by the fact that any of the buffers might have
 583  * extra memory (if there were no empty buffer headers at allocbuf time)
 584  * that we will need to shift around.
 585  *
 586  * The returned bio is &bp->b_bio1
 587  */
 588 void
 589 cluster_callback(struct bio *bio)
 590 {
 591         struct buf *bp = bio->bio_buf;
 592         struct buf *tbp;
 593         int error = 0;
 594
 595         /*
 596          * Must propogate errors to all the components.  A short read (EOF)
 597          * is a critical error.
 598          */
 599         if (bp->b_flags & B_ERROR) {
 600                 error = bp->b_error;
 601         } else if (bp->b_bcount != bp->b_bufsize) {
 602                 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
 603         }
 604
 605         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
 606         /*
 607          * Move memory from the large cluster buffer into the component
 608          * buffers and mark IO as done on these.  Since the memory map
 609          * is the same, no actual copying is required.
 610          */
 611         while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
 612                 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
 613                 if (error) {
 614                         tbp->b_flags |= B_ERROR;
 615                         tbp->b_error = error;
 616                 } else {
 617                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 618                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 619                         /*
 620                          * XXX the bdwrite()/bqrelse() issued during
 621                          * cluster building clears B_RELBUF (see bqrelse()
 622                          * comment).  If direct I/O was specified, we have
 623                          * to restore it here to allow the buffer and VM
 624                          * to be freed.
 625                          */
 626                         if (tbp->b_flags & B_DIRECT)
 627                                 tbp->b_flags |= B_RELBUF;
 628                 }
 629                 biodone(&tbp->b_bio1);
 630         }
 631         relpbuf(bp, &cluster_pbuf_freecnt);
 632 }
 633
 634 /*
 635  *      cluster_wbuild_wb:
 636  *
 637  *      Implement modified write build for cluster.
 638  *
 639  *              write_behind = 0        write behind disabled
 640  *              write_behind = 1        write behind normal (default)
 641  *              write_behind = 2        write behind backed-off
 642  */
 643
 644 static __inline int
 645 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
 646 {
 647         int r = 0;
 648
 649         switch(write_behind) {
 650         case 2:
 651                 if (start_loffset < len)
 652                         break;
 653                 start_loffset -= len;
 654                 /* fall through */
 655         case 1:
 656                 r = cluster_wbuild(vp, blksize, start_loffset, len);
 657                 /* fall through */
 658         default:
 659                 /* fall through */
 660                 break;
 661         }
 662         return(r);
 663 }
 664
 665 /*
 666  * Do clustered write for FFS.
 667  *
 668  * Three cases:
 669  *      1. Write is not sequential (write asynchronously)
 670  *      Write is sequential:
 671  *      2.      beginning of cluster - begin cluster
 672  *      3.      middle of a cluster - add to cluster
 673  *      4.      end of a cluster - asynchronously write cluster
 674  */
 675 void
 676 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
 677 {
 678         struct vnode *vp;
 679         off_t loffset;
 680         int maxclen, cursize;
 681         int async;
 682
 683         vp = bp->b_vp;
 684         if (vp->v_type == VREG)
 685                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 686         else
 687                 async = 0;
 688         loffset = bp->b_loffset;
 689         KASSERT(bp->b_loffset != NOOFFSET,
 690                 ("cluster_write: no buffer offset"));
 691
 692         /* Initialize vnode to beginning of file. */
 693         if (loffset == 0)
 694                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 695
 696         if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
 697             bp->b_bio2.bio_offset == NOOFFSET ||
 698             (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
 699                 maxclen = vmaxiosize(vp);
 700                 if (vp->v_clen != 0) {
 701                         /*
 702                          * Next block is not sequential.
 703                          *
 704                          * If we are not writing at end of file, the process
 705                          * seeked to another point in the file since its last
 706                          * write, or we have reached our maximum cluster size,
 707                          * then push the previous cluster. Otherwise try
 708                          * reallocating to make it sequential.
 709                          *
 710                          * Change to algorithm: only push previous cluster if
 711                          * it was sequential from the point of view of the
 712                          * seqcount heuristic, otherwise leave the buffer
 713                          * intact so we can potentially optimize the I/O
 714                          * later on in the buf_daemon or update daemon
 715                          * flush.
 716                          */
 717                         cursize = vp->v_lastw - vp->v_cstart + blksize;
 718                         if (bp->b_loffset + blksize != filesize ||
 719                             loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
 720                                 if (!async && seqcount > 0) {
 721                                         cluster_wbuild_wb(vp, blksize,
 722                                                 vp->v_cstart, cursize);
 723                                 }
 724                         } else {
 725                                 struct buf **bpp, **endbp;
 726                                 struct cluster_save *buflist;
 727
 728                                 buflist = cluster_collectbufs(vp, bp, blksize);
 729                                 endbp = &buflist->bs_children
 730                                     [buflist->bs_nchildren - 1];
 731                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 732                                         /*
 733                                          * Failed, push the previous cluster
 734                                          * if *really* writing sequentially
 735                                          * in the logical file (seqcount > 1),
 736                                          * otherwise delay it in the hopes that
 737                                          * the low level disk driver can
 738                                          * optimize the write ordering.
 739                                          */
 740                                         for (bpp = buflist->bs_children;
 741                                              bpp < endbp; bpp++)
 742                                                 brelse(*bpp);
 743                                         kfree(buflist, M_SEGMENT);
 744                                         if (seqcount > 1) {
 745                                                 cluster_wbuild_wb(vp,
 746                                                     blksize, vp->v_cstart,
 747                                                     cursize);
 748                                         }
 749                                 } else {
 750                                         /*
 751                                          * Succeeded, keep building cluster.
 752                                          */
 753                                         for (bpp = buflist->bs_children;
 754                                              bpp <= endbp; bpp++)
 755                                                 bdwrite(*bpp);
 756                                         kfree(buflist, M_SEGMENT);
 757                                         vp->v_lastw = loffset;
 758                                         vp->v_lasta = bp->b_bio2.bio_offset;
 759                                         return;
 760                                 }
 761                         }
 762                 }
 763                 /*
 764                  * Consider beginning a cluster. If at end of file, make
 765                  * cluster as large as possible, otherwise find size of
 766                  * existing cluster.
 767                  */
 768                 if ((vp->v_type == VREG) &&
 769                     bp->b_loffset + blksize != filesize &&
 770                     (bp->b_bio2.bio_offset == NOOFFSET) &&
 771                     (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
 772                      bp->b_bio2.bio_offset == NOOFFSET)) {
 773                         bawrite(bp);
 774                         vp->v_clen = 0;
 775                         vp->v_lasta = bp->b_bio2.bio_offset;
 776                         vp->v_cstart = loffset + blksize;
 777                         vp->v_lastw = loffset;
 778                         return;
 779                 }
 780                 if (maxclen > blksize)
 781                         vp->v_clen = maxclen - blksize;
 782                 else
 783                         vp->v_clen = 0;
 784                 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
 785                         vp->v_cstart = loffset + blksize;
 786                         bawrite(bp);
 787                 } else {        /* Wait for rest of cluster */
 788                         vp->v_cstart = loffset;
 789                         bdwrite(bp);
 790                 }
 791         } else if (loffset == vp->v_cstart + vp->v_clen) {
 792                 /*
 793                  * At end of cluster, write it out if seqcount tells us we
 794                  * are operating sequentially, otherwise let the buf or
 795                  * update daemon handle it.
 796                  */
 797                 bdwrite(bp);
 798                 if (seqcount > 1)
 799                         cluster_wbuild_wb(vp, blksize, vp->v_cstart,
 800                                           vp->v_clen + blksize);
 801                 vp->v_clen = 0;
 802                 vp->v_cstart = loffset + blksize;
 803         } else if (vm_page_count_severe()) {
 804                 /*
 805                  * We are low on memory, get it going NOW
 806                  */
 807                 bawrite(bp);
 808         } else {
 809                 /*
 810                  * In the middle of a cluster, so just delay the I/O for now.
 811                  */
 812                 bdwrite(bp);
 813         }
 814         vp->v_lastw = loffset;
 815         vp->v_lasta = bp->b_bio2.bio_offset;
 816 }
 817
 818
 819 /*
 820  * This is an awful lot like cluster_rbuild...wish they could be combined.
 821  * The last lbn argument is the current block on which I/O is being
 822  * performed.  Check to see that it doesn't fall in the middle of
 823  * the current block (if last_bp == NULL).
 824  */
 825 int
 826 cluster_wbuild(struct vnode *vp, int blksize, off_t start_loffset, int bytes)
 827 {
 828         struct buf *bp, *tbp;
 829         int i, j;
 830         int totalwritten = 0;
 831         int maxiosize = vmaxiosize(vp);
 832
 833         while (bytes > 0) {
 834                 /*
 835                  * If the buffer is not delayed-write (i.e. dirty), or it
 836                  * is delayed-write but either locked or inval, it cannot
 837                  * partake in the clustered write.
 838                  */
 839                 tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK);
 840                 if (tbp == NULL ||
 841                     (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI ||
 842                     (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
 843                         if (tbp)
 844                                 BUF_UNLOCK(tbp);
 845                         start_loffset += blksize;
 846                         bytes -= blksize;
 847                         continue;
 848                 }
 849                 bremfree(tbp);
 850                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 851
 852                 /*
 853                  * Extra memory in the buffer, punt on this buffer.
 854                  * XXX we could handle this in most cases, but we would
 855                  * have to push the extra memory down to after our max
 856                  * possible cluster size and then potentially pull it back
 857                  * up if the cluster was terminated prematurely--too much
 858                  * hassle.
 859                  */
 860                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 861                     (tbp->b_bcount != tbp->b_bufsize) ||
 862                     (tbp->b_bcount != blksize) ||
 863                     (bytes == blksize) ||
 864                     ((bp = getpbuf_kva(&cluster_pbuf_freecnt)) == NULL)) {
 865                         totalwritten += tbp->b_bufsize;
 866                         bawrite(tbp);
 867                         start_loffset += blksize;
 868                         bytes -= blksize;
 869                         continue;
 870                 }
 871
 872                 /*
 873                  * Set up the pbuf.  Track our append point with b_bcount
 874                  * and b_bufsize.  b_bufsize is not used by the device but
 875                  * our caller uses it to loop clusters and we use it to
 876                  * detect a premature EOF on the block device.
 877                  */
 878                 bp->b_bcount = 0;
 879                 bp->b_bufsize = 0;
 880                 bp->b_xio.xio_npages = 0;
 881                 bp->b_loffset = tbp->b_loffset;
 882                 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
 883
 884                 /*
 885                  * We are synthesizing a buffer out of vm_page_t's, but
 886                  * if the block size is not page aligned then the starting
 887                  * address may not be either.  Inherit the b_data offset
 888                  * from the original buffer.
 889                  */
 890                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 891                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 892                 bp->b_flags &= ~B_ERROR;
 893                 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
 894                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
 895                 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 896                 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 897
 898                 /*
 899                  * From this location in the file, scan forward to see
 900                  * if there are buffers with adjacent data that need to
 901                  * be written as well.
 902                  */
 903                 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
 904                         if (i != 0) { /* If not the first buffer */
 905                                 tbp = findblk(vp, start_loffset,
 906                                               FINDBLK_NBLOCK);
 907                                 /*
 908                                  * Buffer not found or could not be locked
 909                                  * non-blocking.
 910                                  */
 911                                 if (tbp == NULL)
 912                                         break;
 913
 914                                 /*
 915                                  * If it IS in core, but has different
 916                                  * characteristics, then don't cluster
 917                                  * with it.
 918                                  */
 919                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 920                                      B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 921                                     != (B_DELWRI | B_CLUSTEROK |
 922                                      (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 923                                     (tbp->b_flags & B_LOCKED) ||
 924                                     (LIST_FIRST(&tbp->b_dep) &&
 925                                      buf_checkwrite(tbp))
 926                                 ) {
 927                                         BUF_UNLOCK(tbp);
 928                                         break;
 929                                 }
 930
 931                                 /*
 932                                  * Check that the combined cluster
 933                                  * would make sense with regard to pages
 934                                  * and would not be too large
 935                                  */
 936                                 if ((tbp->b_bcount != blksize) ||
 937                                   ((bp->b_bio2.bio_offset + i) !=
 938                                     tbp->b_bio2.bio_offset) ||
 939                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
 940                                     (maxiosize / PAGE_SIZE))) {
 941                                         BUF_UNLOCK(tbp);
 942                                         break;
 943                                 }
 944                                 /*
 945                                  * Ok, it's passed all the tests,
 946                                  * so remove it from the free list
 947                                  * and mark it busy. We will use it.
 948                                  */
 949                                 bremfree(tbp);
 950                                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 951                         } /* end of code for non-first buffers only */
 952
 953                         /*
 954                          * If the IO is via the VM then we do some
 955                          * special VM hackery (yuck).  Since the buffer's
 956                          * block size may not be page-aligned it is possible
 957                          * for a page to be shared between two buffers.  We
 958                          * have to get rid of the duplication when building
 959                          * the cluster.
 960                          */
 961                         if (tbp->b_flags & B_VMIO) {
 962                                 vm_page_t m;
 963
 964                                 if (i != 0) { /* if not first buffer */
 965                                         for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 966                                                 m = tbp->b_xio.xio_pages[j];
 967                                                 if (m->flags & PG_BUSY) {
 968                                                         bqrelse(tbp);
 969                                                         goto finishcluster;
 970                                                 }
 971                                         }
 972                                 }
 973
 974                                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 975                                         m = tbp->b_xio.xio_pages[j];
 976                                         vm_page_io_start(m);
 977                                         vm_object_pip_add(m->object, 1);
 978                                         if ((bp->b_xio.xio_npages == 0) ||
 979                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
 980                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 981                                                 bp->b_xio.xio_npages++;
 982                                         }
 983                                 }
 984                         }
 985                         bp->b_bcount += blksize;
 986                         bp->b_bufsize += blksize;
 987
 988                         bundirty(tbp);
 989                         tbp->b_flags &= ~B_ERROR;
 990                         tbp->b_cmd = BUF_CMD_WRITE;
 991                         BUF_KERNPROC(tbp);
 992                         cluster_append(&bp->b_bio1, tbp);
 993
 994                         /*
 995                          * check for latent dependencies to be handled
 996                          */
 997                         if (LIST_FIRST(&tbp->b_dep) != NULL)
 998                                 buf_start(tbp);
 999                 }
1000         finishcluster:
1001                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1002                         (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1003                 if (bp->b_bufsize > bp->b_kvasize) {
1004                         panic(
1005                             "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
1006                             bp->b_bufsize, bp->b_kvasize);
1007                 }
1008                 totalwritten += bp->b_bufsize;
1009                 bp->b_dirtyoff = 0;
1010                 bp->b_dirtyend = bp->b_bufsize;
1011                 bp->b_bio1.bio_done = cluster_callback;
1012                 bp->b_cmd = BUF_CMD_WRITE;
1013
1014                 vfs_busy_pages(vp, bp);
1015                 bp->b_runningbufspace = bp->b_bufsize;
1016                 if (bp->b_runningbufspace) {
1017                         runningbufspace += bp->b_runningbufspace;
1018                         ++runningbufcount;
1019                 }
1020                 BUF_KERNPROC(bp);
1021                 vn_strategy(vp, &bp->b_bio1);
1022
1023                 bytes -= i;
1024         }
1025         return totalwritten;
1026 }
1027
1028 /*
1029  * Collect together all the buffers in a cluster.
1030  * Plus add one additional buffer.
1031  */
1032 static struct cluster_save *
1033 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
1034 {
1035         struct cluster_save *buflist;
1036         struct buf *bp;
1037         off_t loffset;
1038         int i, len;
1039
1040         len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
1041         buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1042                          M_SEGMENT, M_WAITOK);
1043         buflist->bs_nchildren = 0;
1044         buflist->bs_children = (struct buf **) (buflist + 1);
1045         for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
1046                 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
1047                 buflist->bs_children[i] = bp;
1048                 if (bp->b_bio2.bio_offset == NOOFFSET) {
1049                         VOP_BMAP(bp->b_vp, bp->b_loffset,
1050                                  &bp->b_bio2.bio_offset,
1051                                  NULL, NULL, BUF_CMD_WRITE);
1052                 }
1053         }
1054         buflist->bs_children[i] = bp = last_bp;
1055         if (bp->b_bio2.bio_offset == NOOFFSET) {
1056                 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1057                          NULL, NULL, BUF_CMD_WRITE);
1058         }
1059         buflist->bs_nchildren = i + 1;
1060         return (buflist);
1061 }
1062
1063 void
1064 cluster_append(struct bio *bio, struct buf *tbp)
1065 {
1066         tbp->b_cluster_next = NULL;
1067         if (bio->bio_caller_info1.cluster_head == NULL) {
1068                 bio->bio_caller_info1.cluster_head = tbp;
1069                 bio->bio_caller_info2.cluster_tail = tbp;
1070         } else {
1071                 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1072                 bio->bio_caller_info2.cluster_tail = tbp;
1073         }
1074 }
1075
1076 static
1077 void
1078 cluster_setram (struct buf *bp)
1079 {
1080         bp->b_flags |= B_RAM;
1081         if (bp->b_xio.xio_npages)
1082                 vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM);
1083 }