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