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