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 static int cluster_wbuild(struct vnode *vp, struct buf **bpp, int blksize,
  80                             off_t start_loffset, int bytes);
  81
  82 static int write_behind = 1;
  83 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
  84     "Cluster write-behind setting");
  85 static int max_readahead = 2 * 1024 * 1024;
  86 SYSCTL_INT(_vfs, OID_AUTO, max_readahead, CTLFLAG_RW, &max_readahead, 0,
  87     "Limit in bytes for desired cluster read-ahead");
  88
  89 extern vm_page_t        bogus_page;
  90
  91 extern int cluster_pbuf_freecnt;
  92
  93 /*
  94  * This replaces bread.
  95  *
  96  * filesize     - read-ahead @ blksize will not cross this boundary
  97  * loffset      - loffset for returned *bpp
  98  * blksize      - blocksize for returned *bpp and read-ahead bps
  99  * minreq       - minimum (not a hard minimum) in bytes, typically reflects
 100  *                a higher level uio resid.
 101  * maxreq       - maximum (sequential heuristic) in bytes (highet typ ~2MB)
 102  * bpp          - return buffer (*bpp) for (loffset,blksize)
 103  */
 104 int
 105 cluster_readx(struct vnode *vp, off_t filesize, off_t loffset,
 106              int blksize, size_t minreq, size_t maxreq, struct buf **bpp)
 107 {
 108         struct buf *bp, *rbp, *reqbp;
 109         off_t origoffset;
 110         off_t doffset;
 111         int error;
 112         int i;
 113         int maxra;
 114         int maxrbuild;
 115
 116         error = 0;
 117
 118         /*
 119          * Calculate the desired read-ahead in blksize'd blocks (maxra).
 120          * To do this we calculate maxreq.
 121          *
 122          * maxreq typically starts out as a sequential heuristic.  If the
 123          * high level uio/resid is bigger (minreq), we pop maxreq up to
 124          * minreq.  This represents the case where random I/O is being
 125          * performed by the userland is issuing big read()'s.
 126          *
 127          * Then we limit maxreq to max_readahead to ensure it is a reasonable
 128          * value.
 129          *
 130          * Finally we must ensure that (loffset + maxreq) does not cross the
 131          * boundary (filesize) for the current blocksize.  If we allowed it
 132          * to cross we could end up with buffers past the boundary with the
 133          * wrong block size (HAMMER large-data areas use mixed block sizes).
 134          * minreq is also absolutely limited to filesize.
 135          */
 136         if (maxreq < minreq)
 137                 maxreq = minreq;
 138         /* minreq not used beyond this point */
 139
 140         if (maxreq > max_readahead) {
 141                 maxreq = max_readahead;
 142                 if (maxreq > 16 * 1024 * 1024)
 143                         maxreq = 16 * 1024 * 1024;
 144         }
 145         if (maxreq < blksize)
 146                 maxreq = blksize;
 147         if (loffset + maxreq > filesize) {
 148                 if (loffset > filesize)
 149                         maxreq = 0;
 150                 else
 151                         maxreq = filesize - loffset;
 152         }
 153
 154         maxra = (int)(maxreq / blksize);
 155
 156         /*
 157          * Get the requested block.
 158          */
 159         if (*bpp)
 160                 reqbp = bp = *bpp;
 161         else
 162                 *bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0);
 163         origoffset = loffset;
 164
 165         /*
 166          * Calculate the maximum cluster size for a single I/O, used
 167          * by cluster_rbuild().
 168          */
 169         maxrbuild = vmaxiosize(vp) / blksize;
 170
 171         /*
 172          * if it is in the cache, then check to see if the reads have been
 173          * sequential.  If they have, then try some read-ahead, otherwise
 174          * back-off on prospective read-aheads.
 175          */
 176         if (bp->b_flags & B_CACHE) {
 177                 /*
 178                  * Not sequential, do not do any read-ahead
 179                  */
 180                 if (maxra <= 1)
 181                         return 0;
 182
 183                 /*
 184                  * No read-ahead mark, do not do any read-ahead
 185                  * yet.
 186                  */
 187                 if ((bp->b_flags & B_RAM) == 0)
 188                         return 0;
 189
 190                 /*
 191                  * We hit a read-ahead-mark, figure out how much read-ahead
 192                  * to do (maxra) and where to start (loffset).
 193                  *
 194                  * Shortcut the scan.  Typically the way this works is that
 195                  * we've built up all the blocks inbetween except for the
 196                  * last in previous iterations, so if the second-to-last
 197                  * block is present we just skip ahead to it.
 198                  *
 199                  * This algorithm has O(1) cpu in the steady state no
 200                  * matter how large maxra is.
 201                  */
 202                 bp->b_flags &= ~B_RAM;
 203
 204                 if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST))
 205                         i = maxra - 1;
 206                 else
 207                         i = 1;
 208                 while (i < maxra) {
 209                         if (findblk(vp, loffset + i * blksize,
 210                                     FINDBLK_TEST) == NULL) {
 211                                 break;
 212                         }
 213                         ++i;
 214                 }
 215
 216                 /*
 217                  * We got everything or everything is in the cache, no
 218                  * point continuing.
 219                  */
 220                 if (i >= maxra)
 221                         return 0;
 222                 maxra -= i;
 223                 loffset += i * blksize;
 224                 reqbp = bp = NULL;
 225         } else {
 226                 __debugvar off_t firstread = bp->b_loffset;
 227                 int nblks;
 228
 229                 /*
 230                  * Set-up synchronous read for bp.
 231                  */
 232                 bp->b_cmd = BUF_CMD_READ;
 233                 bp->b_bio1.bio_done = biodone_sync;
 234                 bp->b_bio1.bio_flags |= BIO_SYNC;
 235
 236                 KASSERT(firstread != NOOFFSET,
 237                         ("cluster_read: no buffer offset"));
 238
 239                 /*
 240                  * nblks is our cluster_rbuild request size, limited
 241                  * primarily by the device.
 242                  */
 243                 if ((nblks = maxra) > maxrbuild)
 244                         nblks = maxrbuild;
 245
 246                 if (nblks > 1) {
 247                         int burstbytes;
 248
 249                         error = VOP_BMAP(vp, loffset, &doffset,
 250                                          &burstbytes, NULL, BUF_CMD_READ);
 251                         if (error)
 252                                 goto single_block_read;
 253                         if (nblks > burstbytes / blksize)
 254                                 nblks = burstbytes / blksize;
 255                         if (doffset == NOOFFSET)
 256                                 goto single_block_read;
 257                         if (nblks <= 1)
 258                                 goto single_block_read;
 259
 260                         bp = cluster_rbuild(vp, filesize, loffset,
 261                                             doffset, blksize, nblks, bp);
 262                         loffset += bp->b_bufsize;
 263                         maxra -= bp->b_bufsize / blksize;
 264                 } else {
 265 single_block_read:
 266                         /*
 267                          * If it isn't in the cache, then get a chunk from
 268                          * disk if sequential, otherwise just get the block.
 269                          */
 270                         cluster_setram(bp);
 271                         loffset += blksize;
 272                         --maxra;
 273                 }
 274         }
 275
 276         /*
 277          * If B_CACHE was not set issue bp.  bp will either be an
 278          * asynchronous cluster buf or a synchronous single-buf.
 279          * If it is a single buf it will be the same as reqbp.
 280          *
 281          * NOTE: Once an async cluster buf is issued bp becomes invalid.
 282          */
 283         if (bp) {
 284 #if defined(CLUSTERDEBUG)
 285                 if (rcluster)
 286                         kprintf("S(%012jx,%d,%d)\n",
 287                             (intmax_t)bp->b_loffset, bp->b_bcount, maxra);
 288 #endif
 289                 if ((bp->b_flags & B_CLUSTER) == 0)
 290                         vfs_busy_pages(vp, bp);
 291                 bp->b_flags &= ~(B_ERROR|B_INVAL);
 292                 vn_strategy(vp, &bp->b_bio1);
 293                 error = 0;
 294                 /* bp invalid now */
 295         }
 296
 297         /*
 298          * If we have been doing sequential I/O, then do some read-ahead.
 299          * The code above us should have positioned us at the next likely
 300          * offset.
 301          *
 302          * Only mess with buffers which we can immediately lock.  HAMMER
 303          * will do device-readahead irrespective of what the blocks
 304          * represent.
 305          */
 306         while (error == 0 && maxra > 0) {
 307                 int burstbytes;
 308                 int tmp_error;
 309                 int nblks;
 310
 311                 rbp = getblk(vp, loffset, blksize,
 312                              GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
 313                 if (rbp == NULL)
 314                         goto no_read_ahead;
 315                 if ((rbp->b_flags & B_CACHE)) {
 316                         bqrelse(rbp);
 317                         goto no_read_ahead;
 318                 }
 319
 320                 /*
 321                  * An error from the read-ahead bmap has nothing to do
 322                  * with the caller's original request.
 323                  */
 324                 tmp_error = VOP_BMAP(vp, loffset, &doffset,
 325                                      &burstbytes, NULL, BUF_CMD_READ);
 326                 if (tmp_error || doffset == NOOFFSET) {
 327                         rbp->b_flags |= B_INVAL;
 328                         brelse(rbp);
 329                         rbp = NULL;
 330                         goto no_read_ahead;
 331                 }
 332                 if ((nblks = maxra) > maxrbuild)
 333                         nblks = maxrbuild;
 334                 if (nblks > burstbytes / blksize)
 335                         nblks = burstbytes / blksize;
 336
 337                 /*
 338                  * rbp: async read
 339                  */
 340                 rbp->b_cmd = BUF_CMD_READ;
 341                 /*rbp->b_flags |= B_AGE*/;
 342                 cluster_setram(rbp);
 343
 344                 if (nblks > 1) {
 345                         rbp = cluster_rbuild(vp, filesize, loffset,
 346                                              doffset, blksize,
 347                                              nblks, rbp);
 348                 } else {
 349                         rbp->b_bio2.bio_offset = doffset;
 350                 }
 351
 352 #if defined(CLUSTERDEBUG)
 353                 if (rcluster) {
 354                         if (bp) {
 355                                 kprintf("A+(%012jx,%d,%jd) "
 356                                         "doff=%012jx minr=%zd ra=%d\n",
 357                                     (intmax_t)loffset, rbp->b_bcount,
 358                                     (intmax_t)(loffset - origoffset),
 359                                     (intmax_t)doffset, minreq, maxra);
 360                         } else {
 361                                 kprintf("A-(%012jx,%d,%jd) "
 362                                         "doff=%012jx minr=%zd ra=%d\n",
 363                                     (intmax_t)rbp->b_loffset, rbp->b_bcount,
 364                                     (intmax_t)(loffset - origoffset),
 365                                     (intmax_t)doffset, minreq, maxra);
 366                         }
 367                 }
 368 #endif
 369                 rbp->b_flags &= ~(B_ERROR|B_INVAL);
 370
 371                 if ((rbp->b_flags & B_CLUSTER) == 0)
 372                         vfs_busy_pages(vp, rbp);
 373                 BUF_KERNPROC(rbp);
 374                 loffset += rbp->b_bufsize;
 375                 maxra -= rbp->b_bufsize / blksize;
 376                 vn_strategy(vp, &rbp->b_bio1);
 377                 /* rbp invalid now */
 378         }
 379
 380         /*
 381          * Wait for our original buffer to complete its I/O.  reqbp will
 382          * be NULL if the original buffer was B_CACHE.  We are returning
 383          * (*bpp) which is the same as reqbp when reqbp != NULL.
 384          */
 385 no_read_ahead:
 386         if (reqbp) {
 387                 KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
 388                 error = biowait(&reqbp->b_bio1, "clurd");
 389         }
 390         return (error);
 391 }
 392
 393 /*
 394  * If blocks are contiguous on disk, use this to provide clustered
 395  * read ahead.  We will read as many blocks as possible sequentially
 396  * and then parcel them up into logical blocks in the buffer hash table.
 397  *
 398  * This function either returns a cluster buf or it returns fbp.  fbp is
 399  * already expected to be set up as a synchronous or asynchronous request.
 400  *
 401  * If a cluster buf is returned it will always be async.
 402  */
 403 static struct buf *
 404 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
 405                int blksize, int run, struct buf *fbp)
 406 {
 407         struct buf *bp, *tbp;
 408         off_t boffset;
 409         int i, j;
 410         int maxiosize = vmaxiosize(vp);
 411
 412         /*
 413          * avoid a division
 414          */
 415         while (loffset + run * blksize > filesize) {
 416                 --run;
 417         }
 418
 419         tbp = fbp;
 420         tbp->b_bio2.bio_offset = doffset;
 421         if((tbp->b_flags & B_MALLOC) ||
 422             ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
 423                 return tbp;
 424         }
 425
 426         bp = trypbuf_kva(&cluster_pbuf_freecnt);
 427         if (bp == NULL) {
 428                 return tbp;
 429         }
 430
 431         /*
 432          * We are synthesizing a buffer out of vm_page_t's, but
 433          * if the block size is not page aligned then the starting
 434          * address may not be either.  Inherit the b_data offset
 435          * from the original buffer.
 436          */
 437         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 438             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 439         bp->b_flags |= B_CLUSTER | B_VMIO;
 440         bp->b_cmd = BUF_CMD_READ;
 441         bp->b_bio1.bio_done = cluster_callback;         /* default to async */
 442         bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 443         bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 444         bp->b_loffset = loffset;
 445         bp->b_bio2.bio_offset = doffset;
 446         KASSERT(bp->b_loffset != NOOFFSET,
 447                 ("cluster_rbuild: no buffer offset"));
 448
 449         bp->b_bcount = 0;
 450         bp->b_bufsize = 0;
 451         bp->b_xio.xio_npages = 0;
 452
 453         for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
 454                 if (i) {
 455                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
 456                             round_page(blksize) > maxiosize) {
 457                                 break;
 458                         }
 459
 460                         /*
 461                          * Shortcut some checks and try to avoid buffers that
 462                          * would block in the lock.  The same checks have to
 463                          * be made again after we officially get the buffer.
 464                          */
 465                         tbp = getblk(vp, loffset + i * blksize, blksize,
 466                                      GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
 467                         if (tbp == NULL)
 468                                 break;
 469                         for (j = 0; j < tbp->b_xio.xio_npages; j++) {
 470                                 if (tbp->b_xio.xio_pages[j]->valid)
 471                                         break;
 472                         }
 473                         if (j != tbp->b_xio.xio_npages) {
 474                                 bqrelse(tbp);
 475                                 break;
 476                         }
 477
 478                         /*
 479                          * Stop scanning if the buffer is fuly valid
 480                          * (marked B_CACHE), or locked (may be doing a
 481                          * background write), or if the buffer is not
 482                          * VMIO backed.  The clustering code can only deal
 483                          * with VMIO-backed buffers.
 484                          */
 485                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 486                             (tbp->b_flags & B_VMIO) == 0 ||
 487                             (LIST_FIRST(&tbp->b_dep) != NULL &&
 488                              buf_checkread(tbp))
 489                         ) {
 490                                 bqrelse(tbp);
 491                                 break;
 492                         }
 493
 494                         /*
 495                          * The buffer must be completely invalid in order to
 496                          * take part in the cluster.  If it is partially valid
 497                          * then we stop.
 498                          */
 499                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
 500                                 if (tbp->b_xio.xio_pages[j]->valid)
 501                                         break;
 502                         }
 503                         if (j != tbp->b_xio.xio_npages) {
 504                                 bqrelse(tbp);
 505                                 break;
 506                         }
 507
 508                         /*
 509                          * Set a read-ahead mark as appropriate
 510                          */
 511                         if (i == 1 || i == (run - 1))
 512                                 cluster_setram(tbp);
 513
 514                         /*
 515                          * Depress the priority of buffers not explicitly
 516                          * requested.
 517                          */
 518                         /* tbp->b_flags |= B_AGE; */
 519
 520                         /*
 521                          * Set the block number if it isn't set, otherwise
 522                          * if it is make sure it matches the block number we
 523                          * expect.
 524                          */
 525                         if (tbp->b_bio2.bio_offset == NOOFFSET) {
 526                                 tbp->b_bio2.bio_offset = boffset;
 527                         } else if (tbp->b_bio2.bio_offset != boffset) {
 528                                 brelse(tbp);
 529                                 break;
 530                         }
 531                 }
 532
 533                 /*
 534                  * The passed-in tbp (i == 0) will already be set up for
 535                  * async or sync operation.  All other tbp's acquire in
 536                  * our loop are set up for async operation.
 537                  */
 538                 tbp->b_cmd = BUF_CMD_READ;
 539                 BUF_KERNPROC(tbp);
 540                 cluster_append(&bp->b_bio1, tbp);
 541                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 542                         vm_page_t m;
 543
 544                         m = tbp->b_xio.xio_pages[j];
 545                         vm_page_busy_wait(m, FALSE, "clurpg");
 546                         vm_page_io_start(m);
 547                         vm_page_wakeup(m);
 548                         vm_object_pip_add(m->object, 1);
 549                         if ((bp->b_xio.xio_npages == 0) ||
 550                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
 551                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 552                                 bp->b_xio.xio_npages++;
 553                         }
 554                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 555                                 tbp->b_xio.xio_pages[j] = bogus_page;
 556                 }
 557                 /*
 558                  * XXX shouldn't this be += size for both, like in
 559                  * cluster_wbuild()?
 560                  *
 561                  * Don't inherit tbp->b_bufsize as it may be larger due to
 562                  * a non-page-aligned size.  Instead just aggregate using
 563                  * 'size'.
 564                  */
 565                 if (tbp->b_bcount != blksize)
 566                     kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
 567                 if (tbp->b_bufsize != blksize)
 568                     kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
 569                 bp->b_bcount += blksize;
 570                 bp->b_bufsize += blksize;
 571         }
 572
 573         /*
 574          * Fully valid pages in the cluster are already good and do not need
 575          * to be re-read from disk.  Replace the page with bogus_page
 576          */
 577         for (j = 0; j < bp->b_xio.xio_npages; j++) {
 578                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 579                     VM_PAGE_BITS_ALL) {
 580                         bp->b_xio.xio_pages[j] = bogus_page;
 581                 }
 582         }
 583         if (bp->b_bufsize > bp->b_kvasize) {
 584                 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
 585                     bp->b_bufsize, bp->b_kvasize);
 586         }
 587         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 588                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 589         BUF_KERNPROC(bp);
 590         return (bp);
 591 }
 592
 593 /*
 594  * Cleanup after a clustered read or write.
 595  * This is complicated by the fact that any of the buffers might have
 596  * extra memory (if there were no empty buffer headers at allocbuf time)
 597  * that we will need to shift around.
 598  *
 599  * The returned bio is &bp->b_bio1
 600  */
 601 void
 602 cluster_callback(struct bio *bio)
 603 {
 604         struct buf *bp = bio->bio_buf;
 605         struct buf *tbp;
 606         int error = 0;
 607
 608         /*
 609          * Must propogate errors to all the components.  A short read (EOF)
 610          * is a critical error.
 611          */
 612         if (bp->b_flags & B_ERROR) {
 613                 error = bp->b_error;
 614         } else if (bp->b_bcount != bp->b_bufsize) {
 615                 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
 616         }
 617
 618         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
 619         /*
 620          * Move memory from the large cluster buffer into the component
 621          * buffers and mark IO as done on these.  Since the memory map
 622          * is the same, no actual copying is required.
 623          */
 624         while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
 625                 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
 626                 if (error) {
 627                         tbp->b_flags |= B_ERROR | B_IODEBUG;
 628                         tbp->b_error = error;
 629                 } else {
 630                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 631                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 632                         tbp->b_flags |= B_IODEBUG;
 633                         /*
 634                          * XXX the bdwrite()/bqrelse() issued during
 635                          * cluster building clears B_RELBUF (see bqrelse()
 636                          * comment).  If direct I/O was specified, we have
 637                          * to restore it here to allow the buffer and VM
 638                          * to be freed.
 639                          */
 640                         if (tbp->b_flags & B_DIRECT)
 641                                 tbp->b_flags |= B_RELBUF;
 642                 }
 643                 biodone(&tbp->b_bio1);
 644         }
 645         relpbuf(bp, &cluster_pbuf_freecnt);
 646 }
 647
 648 /*
 649  *      cluster_wbuild_wb:
 650  *
 651  *      Implement modified write build for cluster.
 652  *
 653  *              write_behind = 0        write behind disabled
 654  *              write_behind = 1        write behind normal (default)
 655  *              write_behind = 2        write behind backed-off
 656  */
 657
 658 static __inline int
 659 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
 660 {
 661         int r = 0;
 662
 663         switch(write_behind) {
 664         case 2:
 665                 if (start_loffset < len)
 666                         break;
 667                 start_loffset -= len;
 668                 /* fall through */
 669         case 1:
 670                 r = cluster_wbuild(vp, NULL, blksize, start_loffset, len);
 671                 /* fall through */
 672         default:
 673                 /* fall through */
 674                 break;
 675         }
 676         return(r);
 677 }
 678
 679 /*
 680  * Do clustered write for FFS.
 681  *
 682  * Three cases:
 683  *      1. Write is not sequential (write asynchronously)
 684  *      Write is sequential:
 685  *      2.      beginning of cluster - begin cluster
 686  *      3.      middle of a cluster - add to cluster
 687  *      4.      end of a cluster - asynchronously write cluster
 688  */
 689 void
 690 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
 691 {
 692         struct vnode *vp;
 693         off_t loffset;
 694         int maxclen, cursize;
 695         int async;
 696
 697         vp = bp->b_vp;
 698         if (vp->v_type == VREG)
 699                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 700         else
 701                 async = 0;
 702         loffset = bp->b_loffset;
 703         KASSERT(bp->b_loffset != NOOFFSET,
 704                 ("cluster_write: no buffer offset"));
 705
 706         /* Initialize vnode to beginning of file. */
 707         if (loffset == 0)
 708                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 709
 710         if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
 711             bp->b_bio2.bio_offset == NOOFFSET ||
 712             (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
 713                 maxclen = vmaxiosize(vp);
 714                 if (vp->v_clen != 0) {
 715                         /*
 716                          * Next block is not sequential.
 717                          *
 718                          * If we are not writing at end of file, the process
 719                          * seeked to another point in the file since its last
 720                          * write, or we have reached our maximum cluster size,
 721                          * then push the previous cluster. Otherwise try
 722                          * reallocating to make it sequential.
 723                          *
 724                          * Change to algorithm: only push previous cluster if
 725                          * it was sequential from the point of view of the
 726                          * seqcount heuristic, otherwise leave the buffer
 727                          * intact so we can potentially optimize the I/O
 728                          * later on in the buf_daemon or update daemon
 729                          * flush.
 730                          */
 731                         cursize = vp->v_lastw - vp->v_cstart + blksize;
 732                         if (bp->b_loffset + blksize < filesize ||
 733                             loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
 734                                 if (!async && seqcount > 0) {
 735                                         cluster_wbuild_wb(vp, blksize,
 736                                                 vp->v_cstart, cursize);
 737                                 }
 738                         } else {
 739                                 struct buf **bpp, **endbp;
 740                                 struct cluster_save *buflist;
 741
 742                                 buflist = cluster_collectbufs(vp, bp, blksize);
 743                                 endbp = &buflist->bs_children
 744                                     [buflist->bs_nchildren - 1];
 745                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 746                                         /*
 747                                          * Failed, push the previous cluster
 748                                          * if *really* writing sequentially
 749                                          * in the logical file (seqcount > 1),
 750                                          * otherwise delay it in the hopes that
 751                                          * the low level disk driver can
 752                                          * optimize the write ordering.
 753                                          */
 754                                         for (bpp = buflist->bs_children;
 755                                              bpp < endbp; bpp++)
 756                                                 brelse(*bpp);
 757                                         kfree(buflist, M_SEGMENT);
 758                                         if (seqcount > 1) {
 759                                                 cluster_wbuild_wb(vp,
 760                                                     blksize, vp->v_cstart,
 761                                                     cursize);
 762                                         }
 763                                 } else {
 764                                         /*
 765                                          * Succeeded, keep building cluster.
 766                                          */
 767                                         for (bpp = buflist->bs_children;
 768                                              bpp <= endbp; bpp++)
 769                                                 bdwrite(*bpp);
 770                                         kfree(buflist, M_SEGMENT);
 771                                         vp->v_lastw = loffset;
 772                                         vp->v_lasta = bp->b_bio2.bio_offset;
 773                                         return;
 774                                 }
 775                         }
 776                 }
 777                 /*
 778                  * Consider beginning a cluster. If at end of file, make
 779                  * cluster as large as possible, otherwise find size of
 780                  * existing cluster.
 781                  */
 782                 if ((vp->v_type == VREG) &&
 783                     bp->b_loffset + blksize < filesize &&
 784                     (bp->b_bio2.bio_offset == NOOFFSET) &&
 785                     (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
 786                      bp->b_bio2.bio_offset == NOOFFSET)) {
 787                         bawrite(bp);
 788                         vp->v_clen = 0;
 789                         vp->v_lasta = bp->b_bio2.bio_offset;
 790                         vp->v_cstart = loffset + blksize;
 791                         vp->v_lastw = loffset;
 792                         return;
 793                 }
 794                 if (maxclen > blksize)
 795                         vp->v_clen = maxclen - blksize;
 796                 else
 797                         vp->v_clen = 0;
 798                 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
 799                         vp->v_cstart = loffset + blksize;
 800                         bawrite(bp);
 801                 } else {        /* Wait for rest of cluster */
 802                         vp->v_cstart = loffset;
 803                         bdwrite(bp);
 804                 }
 805         } else if (loffset == vp->v_cstart + vp->v_clen) {
 806                 /*
 807                  * At end of cluster, write it out if seqcount tells us we
 808                  * are operating sequentially, otherwise let the buf or
 809                  * update daemon handle it.
 810                  */
 811                 bdwrite(bp);
 812                 if (seqcount > 1)
 813                         cluster_wbuild_wb(vp, blksize, vp->v_cstart,
 814                                           vp->v_clen + blksize);
 815                 vp->v_clen = 0;
 816                 vp->v_cstart = loffset + blksize;
 817         } else if (vm_page_count_severe()) {
 818                 /*
 819                  * We are low on memory, get it going NOW
 820                  */
 821                 bawrite(bp);
 822         } else {
 823                 /*
 824                  * In the middle of a cluster, so just delay the I/O for now.
 825                  */
 826                 bdwrite(bp);
 827         }
 828         vp->v_lastw = loffset;
 829         vp->v_lasta = bp->b_bio2.bio_offset;
 830 }
 831
 832 /*
 833  * This is the clustered version of bawrite().  It works similarly to
 834  * cluster_write() except I/O on the buffer is guaranteed to occur.
 835  */
 836 int
 837 cluster_awrite(struct buf *bp)
 838 {
 839         int total;
 840
 841         /*
 842          * Don't bother if it isn't clusterable.
 843          */
 844         if ((bp->b_flags & B_CLUSTEROK) == 0 ||
 845             bp->b_vp == NULL ||
 846             (bp->b_vp->v_flag & VOBJBUF) == 0) {
 847                 total = bp->b_bufsize;
 848                 bawrite(bp);
 849                 return (total);
 850         }
 851
 852         total = cluster_wbuild(bp->b_vp, &bp, bp->b_bufsize,
 853                                bp->b_loffset, vmaxiosize(bp->b_vp));
 854         if (bp)
 855                 bawrite(bp);
 856
 857         return total;
 858 }
 859
 860 /*
 861  * This is an awful lot like cluster_rbuild...wish they could be combined.
 862  * The last lbn argument is the current block on which I/O is being
 863  * performed.  Check to see that it doesn't fall in the middle of
 864  * the current block (if last_bp == NULL).
 865  *
 866  * cluster_wbuild() normally does not guarantee anything.  If bpp is
 867  * non-NULL and cluster_wbuild() is able to incorporate it into the
 868  * I/O it will set *bpp to NULL, otherwise it will leave it alone and
 869  * the caller must dispose of *bpp.
 870  */
 871 static int
 872 cluster_wbuild(struct vnode *vp, struct buf **bpp,
 873                int blksize, off_t start_loffset, int bytes)
 874 {
 875         struct buf *bp, *tbp;
 876         int i, j;
 877         int totalwritten = 0;
 878         int must_initiate;
 879         int maxiosize = vmaxiosize(vp);
 880
 881         while (bytes > 0) {
 882                 /*
 883                  * If the buffer matches the passed locked & removed buffer
 884                  * we used the passed buffer (which might not be B_DELWRI).
 885                  *
 886                  * Otherwise locate the buffer and determine if it is
 887                  * compatible.
 888                  */
 889                 if (bpp && (*bpp)->b_loffset == start_loffset) {
 890                         tbp = *bpp;
 891                         *bpp = NULL;
 892                         bpp = NULL;
 893                 } else {
 894                         tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK);
 895                         if (tbp == NULL ||
 896                             (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) !=
 897                              B_DELWRI ||
 898                             (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
 899                                 if (tbp)
 900                                         BUF_UNLOCK(tbp);
 901                                 start_loffset += blksize;
 902                                 bytes -= blksize;
 903                                 continue;
 904                         }
 905                         bremfree(tbp);
 906                 }
 907                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 908
 909                 /*
 910                  * Extra memory in the buffer, punt on this buffer.
 911                  * XXX we could handle this in most cases, but we would
 912                  * have to push the extra memory down to after our max
 913                  * possible cluster size and then potentially pull it back
 914                  * up if the cluster was terminated prematurely--too much
 915                  * hassle.
 916                  */
 917                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 918                     (tbp->b_bcount != tbp->b_bufsize) ||
 919                     (tbp->b_bcount != blksize) ||
 920                     (bytes == blksize) ||
 921                     ((bp = getpbuf_kva(&cluster_pbuf_freecnt)) == NULL)) {
 922                         totalwritten += tbp->b_bufsize;
 923                         bawrite(tbp);
 924                         start_loffset += blksize;
 925                         bytes -= blksize;
 926                         continue;
 927                 }
 928
 929                 /*
 930                  * Set up the pbuf.  Track our append point with b_bcount
 931                  * and b_bufsize.  b_bufsize is not used by the device but
 932                  * our caller uses it to loop clusters and we use it to
 933                  * detect a premature EOF on the block device.
 934                  */
 935                 bp->b_bcount = 0;
 936                 bp->b_bufsize = 0;
 937                 bp->b_xio.xio_npages = 0;
 938                 bp->b_loffset = tbp->b_loffset;
 939                 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
 940
 941                 /*
 942                  * We are synthesizing a buffer out of vm_page_t's, but
 943                  * if the block size is not page aligned then the starting
 944                  * address may not be either.  Inherit the b_data offset
 945                  * from the original buffer.
 946                  */
 947                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 948                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 949                 bp->b_flags &= ~B_ERROR;
 950                 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
 951                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
 952                 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 953                 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 954
 955                 /*
 956                  * From this location in the file, scan forward to see
 957                  * if there are buffers with adjacent data that need to
 958                  * be written as well.
 959                  *
 960                  * IO *must* be initiated on index 0 at this point
 961                  * (particularly when called from cluster_awrite()).
 962                  */
 963                 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
 964                         if (i == 0) {
 965                                 must_initiate = 1;
 966                         } else {
 967                                 /*
 968                                  * Not first buffer.
 969                                  */
 970                                 must_initiate = 0;
 971                                 tbp = findblk(vp, start_loffset,
 972                                               FINDBLK_NBLOCK);
 973                                 /*
 974                                  * Buffer not found or could not be locked
 975                                  * non-blocking.
 976                                  */
 977                                 if (tbp == NULL)
 978                                         break;
 979
 980                                 /*
 981                                  * If it IS in core, but has different
 982                                  * characteristics, then don't cluster
 983                                  * with it.
 984                                  */
 985                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 986                                      B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 987                                     != (B_DELWRI | B_CLUSTEROK |
 988                                      (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 989                                     (tbp->b_flags & B_LOCKED)
 990                                 ) {
 991                                         BUF_UNLOCK(tbp);
 992                                         break;
 993                                 }
 994
 995                                 /*
 996                                  * Check that the combined cluster
 997                                  * would make sense with regard to pages
 998                                  * and would not be too large
 999                                  *
1000                                  * WARNING! buf_checkwrite() must be the last
1001                                  *          check made.  If it returns 0 then
1002                                  *          we must initiate the I/O.
1003                                  */
1004                                 if ((tbp->b_bcount != blksize) ||
1005                                   ((bp->b_bio2.bio_offset + i) !=
1006                                     tbp->b_bio2.bio_offset) ||
1007                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
1008                                     (maxiosize / PAGE_SIZE)) ||
1009                                   (LIST_FIRST(&tbp->b_dep) &&
1010                                    buf_checkwrite(tbp))
1011                                 ) {
1012                                         BUF_UNLOCK(tbp);
1013                                         break;
1014                                 }
1015                                 if (LIST_FIRST(&tbp->b_dep))
1016                                         must_initiate = 1;
1017                                 /*
1018                                  * Ok, it's passed all the tests,
1019                                  * so remove it from the free list
1020                                  * and mark it busy. We will use it.
1021                                  */
1022                                 bremfree(tbp);
1023                                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
1024                         }
1025
1026                         /*
1027                          * If the IO is via the VM then we do some
1028                          * special VM hackery (yuck).  Since the buffer's
1029                          * block size may not be page-aligned it is possible
1030                          * for a page to be shared between two buffers.  We
1031                          * have to get rid of the duplication when building
1032                          * the cluster.
1033                          */
1034                         if (tbp->b_flags & B_VMIO) {
1035                                 vm_page_t m;
1036
1037                                 /*
1038                                  * Try to avoid deadlocks with the VM system.
1039                                  * However, we cannot abort the I/O if
1040                                  * must_initiate is non-zero.
1041                                  */
1042                                 if (must_initiate == 0) {
1043                                         for (j = 0;
1044                                              j < tbp->b_xio.xio_npages;
1045                                              ++j) {
1046                                                 m = tbp->b_xio.xio_pages[j];
1047                                                 if (m->flags & PG_BUSY) {
1048                                                         bqrelse(tbp);
1049                                                         goto finishcluster;
1050                                                 }
1051                                         }
1052                                 }
1053
1054                                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
1055                                         m = tbp->b_xio.xio_pages[j];
1056                                         vm_page_busy_wait(m, FALSE, "clurpg");
1057                                         vm_page_io_start(m);
1058                                         vm_page_wakeup(m);
1059                                         vm_object_pip_add(m->object, 1);
1060                                         if ((bp->b_xio.xio_npages == 0) ||
1061                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
1062                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
1063                                                 bp->b_xio.xio_npages++;
1064                                         }
1065                                 }
1066                         }
1067                         bp->b_bcount += blksize;
1068                         bp->b_bufsize += blksize;
1069
1070                         bundirty(tbp);
1071                         tbp->b_flags &= ~B_ERROR;
1072                         tbp->b_cmd = BUF_CMD_WRITE;
1073                         BUF_KERNPROC(tbp);
1074                         cluster_append(&bp->b_bio1, tbp);
1075
1076                         /*
1077                          * check for latent dependencies to be handled
1078                          */
1079                         if (LIST_FIRST(&tbp->b_dep) != NULL)
1080                                 buf_start(tbp);
1081                 }
1082         finishcluster:
1083                 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1084                             (vm_page_t *)bp->b_xio.xio_pages,
1085                             bp->b_xio.xio_npages);
1086                 if (bp->b_bufsize > bp->b_kvasize) {
1087                         panic("cluster_wbuild: b_bufsize(%d) "
1088                               "> b_kvasize(%d)\n",
1089                               bp->b_bufsize, bp->b_kvasize);
1090                 }
1091                 totalwritten += bp->b_bufsize;
1092                 bp->b_dirtyoff = 0;
1093                 bp->b_dirtyend = bp->b_bufsize;
1094                 bp->b_bio1.bio_done = cluster_callback;
1095                 bp->b_cmd = BUF_CMD_WRITE;
1096
1097                 vfs_busy_pages(vp, bp);
1098                 bsetrunningbufspace(bp, bp->b_bufsize);
1099                 BUF_KERNPROC(bp);
1100                 vn_strategy(vp, &bp->b_bio1);
1101
1102                 bytes -= i;
1103         }
1104         return totalwritten;
1105 }
1106
1107 /*
1108  * Collect together all the buffers in a cluster.
1109  * Plus add one additional buffer.
1110  */
1111 static struct cluster_save *
1112 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
1113 {
1114         struct cluster_save *buflist;
1115         struct buf *bp;
1116         off_t loffset;
1117         int i, len;
1118
1119         len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
1120         buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1121                          M_SEGMENT, M_WAITOK);
1122         buflist->bs_nchildren = 0;
1123         buflist->bs_children = (struct buf **) (buflist + 1);
1124         for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
1125                 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
1126                 buflist->bs_children[i] = bp;
1127                 if (bp->b_bio2.bio_offset == NOOFFSET) {
1128                         VOP_BMAP(bp->b_vp, bp->b_loffset,
1129                                  &bp->b_bio2.bio_offset,
1130                                  NULL, NULL, BUF_CMD_WRITE);
1131                 }
1132         }
1133         buflist->bs_children[i] = bp = last_bp;
1134         if (bp->b_bio2.bio_offset == NOOFFSET) {
1135                 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1136                          NULL, NULL, BUF_CMD_WRITE);
1137         }
1138         buflist->bs_nchildren = i + 1;
1139         return (buflist);
1140 }
1141
1142 void
1143 cluster_append(struct bio *bio, struct buf *tbp)
1144 {
1145         tbp->b_cluster_next = NULL;
1146         if (bio->bio_caller_info1.cluster_head == NULL) {
1147                 bio->bio_caller_info1.cluster_head = tbp;
1148                 bio->bio_caller_info2.cluster_tail = tbp;
1149         } else {
1150                 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1151                 bio->bio_caller_info2.cluster_tail = tbp;
1152         }
1153 }
1154
1155 static
1156 void
1157 cluster_setram (struct buf *bp)
1158 {
1159         bp->b_flags |= B_RAM;
1160         if (bp->b_xio.xio_npages)
1161                 vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM);
1162 }