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