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  * 4. Neither the name of the University nor the names of its contributors
  16  *    may be used to endorse or promote products derived from this software
  17  *    without specific prior written permission.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  29  * SUCH DAMAGE.
  30  *
  31  *      @(#)vfs_cluster.c       8.7 (Berkeley) 2/13/94
  32  */
  33
  34 #include <sys/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include "opt_debug_cluster.h"
  38
  39 #include <sys/param.h>
  40 #include <sys/systm.h>
  41 #include <sys/kernel.h>
  42 #include <sys/proc.h>
  43 #include <sys/bio.h>
  44 #include <sys/buf.h>
  45 #include <sys/vnode.h>
  46 #include <sys/malloc.h>
  47 #include <sys/mount.h>
  48 #include <sys/resourcevar.h>
  49 #include <sys/rwlock.h>
  50 #include <sys/vmmeter.h>
  51 #include <vm/vm.h>
  52 #include <vm/vm_object.h>
  53 #include <vm/vm_page.h>
  54 #include <sys/sysctl.h>
  55
  56 #if defined(CLUSTERDEBUG)
  57 static int      rcluster= 0;
  58 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0,
  59     "Debug VFS clustering code");
  60 #endif
  61
  62 static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer");
  63
  64 static struct cluster_save *
  65         cluster_collectbufs(struct vnode *vp, struct buf *last_bp);
  66 static struct buf *
  67         cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
  68                          daddr_t blkno, long size, int run, struct buf *fbp);
  69 static void cluster_callback(struct buf *);
  70
  71 static int write_behind = 1;
  72 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
  73     "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
  74
  75 static int read_max = 64;
  76 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
  77     "Cluster read-ahead max block count");
  78
  79 /* Page expended to mark partially backed buffers */
  80 extern vm_page_t        bogus_page;
  81
  82 /*
  83  * Read data to a buf, including read-ahead if we find this to be beneficial.
  84  * cluster_read replaces bread.
  85  */
  86 int
  87 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
  88     struct ucred *cred, long totread, int seqcount, int gbflags,
  89     struct buf **bpp)
  90 {
  91         struct buf *bp, *rbp, *reqbp;
  92         struct bufobj *bo;
  93         daddr_t blkno, origblkno;
  94         int maxra, racluster;
  95         int error, ncontig;
  96         int i;
  97
  98         error = 0;
  99         bo = &vp->v_bufobj;
 100
 101         /*
 102          * Try to limit the amount of read-ahead by a few
 103          * ad-hoc parameters.  This needs work!!!
 104          */
 105         racluster = vp->v_mount->mnt_iosize_max / size;
 106         maxra = seqcount;
 107         maxra = min(read_max, maxra);
 108         maxra = min(nbuf/8, maxra);
 109         if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
 110                 maxra = (filesize / size) - lblkno;
 111
 112         /*
 113          * get the requested block
 114          */
 115         *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, 0);
 116         origblkno = lblkno;
 117
 118         /*
 119          * if it is in the cache, then check to see if the reads have been
 120          * sequential.  If they have, then try some read-ahead, otherwise
 121          * back-off on prospective read-aheads.
 122          */
 123         if (bp->b_flags & B_CACHE) {
 124                 if (!seqcount) {
 125                         return 0;
 126                 } else if ((bp->b_flags & B_RAM) == 0) {
 127                         return 0;
 128                 } else {
 129                         bp->b_flags &= ~B_RAM;
 130                         BO_LOCK(bo);
 131                         for (i = 1; i < maxra; i++) {
 132                                 /*
 133                                  * Stop if the buffer does not exist or it
 134                                  * is invalid (about to go away?)
 135                                  */
 136                                 rbp = gbincore(&vp->v_bufobj, lblkno+i);
 137                                 if (rbp == NULL || (rbp->b_flags & B_INVAL))
 138                                         break;
 139
 140                                 /*
 141                                  * Set another read-ahead mark so we know
 142                                  * to check again. (If we can lock the
 143                                  * buffer without waiting)
 144                                  */
 145                                 if ((((i % racluster) == (racluster - 1)) ||
 146                                     (i == (maxra - 1)))
 147                                     && (0 == BUF_LOCK(rbp,
 148                                         LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
 149                                         rbp->b_flags |= B_RAM;
 150                                         BUF_UNLOCK(rbp);
 151                                 }
 152                         }
 153                         BO_UNLOCK(bo);
 154                         if (i >= maxra) {
 155                                 return 0;
 156                         }
 157                         lblkno += i;
 158                 }
 159                 reqbp = bp = NULL;
 160         /*
 161          * If it isn't in the cache, then get a chunk from
 162          * disk if sequential, otherwise just get the block.
 163          */
 164         } else {
 165                 off_t firstread = bp->b_offset;
 166                 int nblks;
 167
 168                 KASSERT(bp->b_offset != NOOFFSET,
 169                     ("cluster_read: no buffer offset"));
 170
 171                 ncontig = 0;
 172
 173                 /*
 174                  * Compute the total number of blocks that we should read
 175                  * synchronously.
 176                  */
 177                 if (firstread + totread > filesize)
 178                         totread = filesize - firstread;
 179                 nblks = howmany(totread, size);
 180                 if (nblks > racluster)
 181                         nblks = racluster;
 182
 183                 /*
 184                  * Now compute the number of contiguous blocks.
 185                  */
 186                 if (nblks > 1) {
 187                         error = VOP_BMAP(vp, lblkno, NULL,
 188                                 &blkno, &ncontig, NULL);
 189                         /*
 190                          * If this failed to map just do the original block.
 191                          */
 192                         if (error || blkno == -1)
 193                                 ncontig = 0;
 194                 }
 195
 196                 /*
 197                  * If we have contiguous data available do a cluster
 198                  * otherwise just read the requested block.
 199                  */
 200                 if (ncontig) {
 201                         /* Account for our first block. */
 202                         ncontig = min(ncontig + 1, nblks);
 203                         if (ncontig < nblks)
 204                                 nblks = ncontig;
 205                         bp = cluster_rbuild(vp, filesize, lblkno,
 206                                 blkno, size, nblks, bp);
 207                         lblkno += (bp->b_bufsize / size);
 208                 } else {
 209                         bp->b_flags |= B_RAM;
 210                         bp->b_iocmd = BIO_READ;
 211                         lblkno += 1;
 212                 }
 213         }
 214
 215         /*
 216          * handle the synchronous read so that it is available ASAP.
 217          */
 218         if (bp) {
 219                 if ((bp->b_flags & B_CLUSTER) == 0) {
 220                         vfs_busy_pages(bp, 0);
 221                 }
 222                 bp->b_flags &= ~B_INVAL;
 223                 bp->b_ioflags &= ~BIO_ERROR;
 224                 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
 225                         BUF_KERNPROC(bp);
 226                 bp->b_iooffset = dbtob(bp->b_blkno);
 227                 bstrategy(bp);
 228                 curthread->td_ru.ru_inblock++;
 229         }
 230
 231         /*
 232          * If we have been doing sequential I/O, then do some read-ahead.
 233          */
 234         while (lblkno < (origblkno + maxra)) {
 235                 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
 236                 if (error)
 237                         break;
 238
 239                 if (blkno == -1)
 240                         break;
 241
 242                 /*
 243                  * We could throttle ncontig here by maxra but we might as
 244                  * well read the data if it is contiguous.  We're throttled
 245                  * by racluster anyway.
 246                  */
 247                 if (ncontig) {
 248                         ncontig = min(ncontig + 1, racluster);
 249                         rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
 250                                 size, ncontig, NULL);
 251                         lblkno += (rbp->b_bufsize / size);
 252                         if (rbp->b_flags & B_DELWRI) {
 253                                 bqrelse(rbp);
 254                                 continue;
 255                         }
 256                 } else {
 257                         rbp = getblk(vp, lblkno, size, 0, 0, 0);
 258                         lblkno += 1;
 259                         if (rbp->b_flags & B_DELWRI) {
 260                                 bqrelse(rbp);
 261                                 continue;
 262                         }
 263                         rbp->b_flags |= B_ASYNC | B_RAM;
 264                         rbp->b_iocmd = BIO_READ;
 265                         rbp->b_blkno = blkno;
 266                 }
 267                 if (rbp->b_flags & B_CACHE) {
 268                         rbp->b_flags &= ~B_ASYNC;
 269                         bqrelse(rbp);
 270                         continue;
 271                 }
 272                 if ((rbp->b_flags & B_CLUSTER) == 0) {
 273                         vfs_busy_pages(rbp, 0);
 274                 }
 275                 rbp->b_flags &= ~B_INVAL;
 276                 rbp->b_ioflags &= ~BIO_ERROR;
 277                 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
 278                         BUF_KERNPROC(rbp);
 279                 rbp->b_iooffset = dbtob(rbp->b_blkno);
 280                 bstrategy(rbp);
 281                 curthread->td_ru.ru_inblock++;
 282         }
 283
 284         if (reqbp)
 285                 return (bufwait(reqbp));
 286         else
 287                 return (error);
 288 }
 289
 290 /*
 291  * If blocks are contiguous on disk, use this to provide clustered
 292  * read ahead.  We will read as many blocks as possible sequentially
 293  * and then parcel them up into logical blocks in the buffer hash table.
 294  */
 295 static struct buf *
 296 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
 297         struct vnode *vp;
 298         u_quad_t filesize;
 299         daddr_t lbn;
 300         daddr_t blkno;
 301         long size;
 302         int run;
 303         struct buf *fbp;
 304 {
 305         struct bufobj *bo;
 306         struct buf *bp, *tbp;
 307         daddr_t bn;
 308         off_t off;
 309         long tinc, tsize;
 310         int i, inc, j, toff;
 311
 312         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 313             ("cluster_rbuild: size %ld != f_iosize %jd\n",
 314             size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
 315
 316         /*
 317          * avoid a division
 318          */
 319         while ((u_quad_t) size * (lbn + run) > filesize) {
 320                 --run;
 321         }
 322
 323         if (fbp) {
 324                 tbp = fbp;
 325                 tbp->b_iocmd = BIO_READ;
 326         } else {
 327                 tbp = getblk(vp, lbn, size, 0, 0, 0);
 328                 if (tbp->b_flags & B_CACHE)
 329                         return tbp;
 330                 tbp->b_flags |= B_ASYNC | B_RAM;
 331                 tbp->b_iocmd = BIO_READ;
 332         }
 333         tbp->b_blkno = blkno;
 334         if( (tbp->b_flags & B_MALLOC) ||
 335                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
 336                 return tbp;
 337
 338         bp = trypbuf(&cluster_pbuf_freecnt);
 339         if (bp == 0)
 340                 return tbp;
 341
 342         /*
 343          * We are synthesizing a buffer out of vm_page_t's, but
 344          * if the block size is not page aligned then the starting
 345          * address may not be either.  Inherit the b_data offset
 346          * from the original buffer.
 347          */
 348         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 349             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 350         bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
 351         bp->b_iocmd = BIO_READ;
 352         bp->b_iodone = cluster_callback;
 353         bp->b_blkno = blkno;
 354         bp->b_lblkno = lbn;
 355         bp->b_offset = tbp->b_offset;
 356         KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
 357         pbgetvp(vp, bp);
 358
 359         TAILQ_INIT(&bp->b_cluster.cluster_head);
 360
 361         bp->b_bcount = 0;
 362         bp->b_bufsize = 0;
 363         bp->b_npages = 0;
 364
 365         inc = btodb(size);
 366         bo = &vp->v_bufobj;
 367         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
 368                 if (i != 0) {
 369                         if ((bp->b_npages * PAGE_SIZE) +
 370                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 371                                 break;
 372                         }
 373
 374                         tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT);
 375
 376                         /* Don't wait around for locked bufs. */
 377                         if (tbp == NULL)
 378                                 break;
 379
 380                         /*
 381                          * Stop scanning if the buffer is fully valid
 382                          * (marked B_CACHE), or locked (may be doing a
 383                          * background write), or if the buffer is not
 384                          * VMIO backed.  The clustering code can only deal
 385                          * with VMIO-backed buffers.
 386                          */
 387                         BO_LOCK(bo);
 388                         if ((tbp->b_vflags & BV_BKGRDINPROG) ||
 389                             (tbp->b_flags & B_CACHE) ||
 390                             (tbp->b_flags & B_VMIO) == 0) {
 391                                 BO_UNLOCK(bo);
 392                                 bqrelse(tbp);
 393                                 break;
 394                         }
 395                         BO_UNLOCK(bo);
 396
 397                         /*
 398                          * The buffer must be completely invalid in order to
 399                          * take part in the cluster.  If it is partially valid
 400                          * then we stop.
 401                          */
 402                         off = tbp->b_offset;
 403                         tsize = size;
 404                         VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
 405                         for (j = 0; tsize > 0; j++) {
 406                                 toff = off & PAGE_MASK;
 407                                 tinc = tsize;
 408                                 if (toff + tinc > PAGE_SIZE)
 409                                         tinc = PAGE_SIZE - toff;
 410                                 VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object);
 411                                 if ((tbp->b_pages[j]->valid &
 412                                     vm_page_bits(toff, tinc)) != 0)
 413                                         break;
 414                                 off += tinc;
 415                                 tsize -= tinc;
 416                         }
 417                         VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
 418                         if (tsize > 0) {
 419                                 bqrelse(tbp);
 420                                 break;
 421                         }
 422
 423                         /*
 424                          * Set a read-ahead mark as appropriate
 425                          */
 426                         if ((fbp && (i == 1)) || (i == (run - 1)))
 427                                 tbp->b_flags |= B_RAM;
 428
 429                         /*
 430                          * Set the buffer up for an async read (XXX should
 431                          * we do this only if we do not wind up brelse()ing?).
 432                          * Set the block number if it isn't set, otherwise
 433                          * if it is make sure it matches the block number we
 434                          * expect.
 435                          */
 436                         tbp->b_flags |= B_ASYNC;
 437                         tbp->b_iocmd = BIO_READ;
 438                         if (tbp->b_blkno == tbp->b_lblkno) {
 439                                 tbp->b_blkno = bn;
 440                         } else if (tbp->b_blkno != bn) {
 441                                 brelse(tbp);
 442                                 break;
 443                         }
 444                 }
 445                 /*
 446                  * XXX fbp from caller may not be B_ASYNC, but we are going
 447                  * to biodone() it in cluster_callback() anyway
 448                  */
 449                 BUF_KERNPROC(tbp);
 450                 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 451                         tbp, b_cluster.cluster_entry);
 452                 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
 453                 for (j = 0; j < tbp->b_npages; j += 1) {
 454                         vm_page_t m;
 455                         m = tbp->b_pages[j];
 456                         vm_page_io_start(m);
 457                         vm_object_pip_add(m->object, 1);
 458                         if ((bp->b_npages == 0) ||
 459                                 (bp->b_pages[bp->b_npages-1] != m)) {
 460                                 bp->b_pages[bp->b_npages] = m;
 461                                 bp->b_npages++;
 462                         }
 463                         if (m->valid == VM_PAGE_BITS_ALL)
 464                                 tbp->b_pages[j] = bogus_page;
 465                 }
 466                 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
 467                 /*
 468                  * Don't inherit tbp->b_bufsize as it may be larger due to
 469                  * a non-page-aligned size.  Instead just aggregate using
 470                  * 'size'.
 471                  */
 472                 if (tbp->b_bcount != size)
 473                         printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
 474                 if (tbp->b_bufsize != size)
 475                         printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
 476                 bp->b_bcount += size;
 477                 bp->b_bufsize += size;
 478         }
 479
 480         /*
 481          * Fully valid pages in the cluster are already good and do not need
 482          * to be re-read from disk.  Replace the page with bogus_page
 483          */
 484         VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
 485         for (j = 0; j < bp->b_npages; j++) {
 486                 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object);
 487                 if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
 488                         bp->b_pages[j] = bogus_page;
 489         }
 490         VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
 491         if (bp->b_bufsize > bp->b_kvasize)
 492                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 493                     bp->b_bufsize, bp->b_kvasize);
 494         bp->b_kvasize = bp->b_bufsize;
 495
 496         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 497                 (vm_page_t *)bp->b_pages, bp->b_npages);
 498         return (bp);
 499 }
 500
 501 /*
 502  * Cleanup after a clustered read or write.
 503  * This is complicated by the fact that any of the buffers might have
 504  * extra memory (if there were no empty buffer headers at allocbuf time)
 505  * that we will need to shift around.
 506  */
 507 static void
 508 cluster_callback(bp)
 509         struct buf *bp;
 510 {
 511         struct buf *nbp, *tbp;
 512         int error = 0;
 513
 514         /*
 515          * Must propogate errors to all the components.
 516          */
 517         if (bp->b_ioflags & BIO_ERROR)
 518                 error = bp->b_error;
 519
 520         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
 521         /*
 522          * Move memory from the large cluster buffer into the component
 523          * buffers and mark IO as done on these.
 524          */
 525         for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
 526                 tbp; tbp = nbp) {
 527                 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
 528                 if (error) {
 529                         tbp->b_ioflags |= BIO_ERROR;
 530                         tbp->b_error = error;
 531                 } else {
 532                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 533                         tbp->b_flags &= ~B_INVAL;
 534                         tbp->b_ioflags &= ~BIO_ERROR;
 535                         /*
 536                          * XXX the bdwrite()/bqrelse() issued during
 537                          * cluster building clears B_RELBUF (see bqrelse()
 538                          * comment).  If direct I/O was specified, we have
 539                          * to restore it here to allow the buffer and VM
 540                          * to be freed.
 541                          */
 542                         if (tbp->b_flags & B_DIRECT)
 543                                 tbp->b_flags |= B_RELBUF;
 544                 }
 545                 bufdone(tbp);
 546         }
 547         pbrelvp(bp);
 548         relpbuf(bp, &cluster_pbuf_freecnt);
 549 }
 550
 551 /*
 552  *      cluster_wbuild_wb:
 553  *
 554  *      Implement modified write build for cluster.
 555  *
 556  *              write_behind = 0        write behind disabled
 557  *              write_behind = 1        write behind normal (default)
 558  *              write_behind = 2        write behind backed-off
 559  */
 560
 561 static __inline int
 562 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
 563 {
 564         int r = 0;
 565
 566         switch (write_behind) {
 567         case 2:
 568                 if (start_lbn < len)
 569                         break;
 570                 start_lbn -= len;
 571                 /* FALLTHROUGH */
 572         case 1:
 573                 r = cluster_wbuild(vp, size, start_lbn, len, 0);
 574                 /* FALLTHROUGH */
 575         default:
 576                 /* FALLTHROUGH */
 577                 break;
 578         }
 579         return(r);
 580 }
 581
 582 /*
 583  * Do clustered write for FFS.
 584  *
 585  * Three cases:
 586  *      1. Write is not sequential (write asynchronously)
 587  *      Write is sequential:
 588  *      2.      beginning of cluster - begin cluster
 589  *      3.      middle of a cluster - add to cluster
 590  *      4.      end of a cluster - asynchronously write cluster
 591  */
 592 void
 593 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount,
 594     int gbflags)
 595 {
 596         daddr_t lbn;
 597         int maxclen, cursize;
 598         int lblocksize;
 599         int async;
 600
 601         if (vp->v_type == VREG) {
 602                 async = DOINGASYNC(vp);
 603                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 604         } else {
 605                 async = 0;
 606                 lblocksize = bp->b_bufsize;
 607         }
 608         lbn = bp->b_lblkno;
 609         KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
 610
 611         /* Initialize vnode to beginning of file. */
 612         if (lbn == 0)
 613                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 614
 615         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
 616             (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
 617                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
 618                 if (vp->v_clen != 0) {
 619                         /*
 620                          * Next block is not sequential.
 621                          *
 622                          * If we are not writing at end of file, the process
 623                          * seeked to another point in the file since its last
 624                          * write, or we have reached our maximum cluster size,
 625                          * then push the previous cluster. Otherwise try
 626                          * reallocating to make it sequential.
 627                          *
 628                          * Change to algorithm: only push previous cluster if
 629                          * it was sequential from the point of view of the
 630                          * seqcount heuristic, otherwise leave the buffer
 631                          * intact so we can potentially optimize the I/O
 632                          * later on in the buf_daemon or update daemon
 633                          * flush.
 634                          */
 635                         cursize = vp->v_lastw - vp->v_cstart + 1;
 636                         if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
 637                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
 638                                 if (!async && seqcount > 0) {
 639                                         cluster_wbuild_wb(vp, lblocksize,
 640                                                 vp->v_cstart, cursize);
 641                                 }
 642                         } else {
 643                                 struct buf **bpp, **endbp;
 644                                 struct cluster_save *buflist;
 645
 646                                 buflist = cluster_collectbufs(vp, bp);
 647                                 endbp = &buflist->bs_children
 648                                     [buflist->bs_nchildren - 1];
 649                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 650                                         /*
 651                                          * Failed, push the previous cluster
 652                                          * if *really* writing sequentially
 653                                          * in the logical file (seqcount > 1),
 654                                          * otherwise delay it in the hopes that
 655                                          * the low level disk driver can
 656                                          * optimize the write ordering.
 657                                          */
 658                                         for (bpp = buflist->bs_children;
 659                                              bpp < endbp; bpp++)
 660                                                 brelse(*bpp);
 661                                         free(buflist, M_SEGMENT);
 662                                         if (seqcount > 1) {
 663                                                 cluster_wbuild_wb(vp,
 664                                                     lblocksize, vp->v_cstart,
 665                                                     cursize);
 666                                         }
 667                                 } else {
 668                                         /*
 669                                          * Succeeded, keep building cluster.
 670                                          */
 671                                         for (bpp = buflist->bs_children;
 672                                              bpp <= endbp; bpp++)
 673                                                 bdwrite(*bpp);
 674                                         free(buflist, M_SEGMENT);
 675                                         vp->v_lastw = lbn;
 676                                         vp->v_lasta = bp->b_blkno;
 677                                         return;
 678                                 }
 679                         }
 680                 }
 681                 /*
 682                  * Consider beginning a cluster. If at end of file, make
 683                  * cluster as large as possible, otherwise find size of
 684                  * existing cluster.
 685                  */
 686                 if ((vp->v_type == VREG) &&
 687                         ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
 688                     (bp->b_blkno == bp->b_lblkno) &&
 689                     (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
 690                      bp->b_blkno == -1)) {
 691                         bawrite(bp);
 692                         vp->v_clen = 0;
 693                         vp->v_lasta = bp->b_blkno;
 694                         vp->v_cstart = lbn + 1;
 695                         vp->v_lastw = lbn;
 696                         return;
 697                 }
 698                 vp->v_clen = maxclen;
 699                 if (!async && maxclen == 0) {   /* I/O not contiguous */
 700                         vp->v_cstart = lbn + 1;
 701                         bawrite(bp);
 702                 } else {        /* Wait for rest of cluster */
 703                         vp->v_cstart = lbn;
 704                         bdwrite(bp);
 705                 }
 706         } else if (lbn == vp->v_cstart + vp->v_clen) {
 707                 /*
 708                  * At end of cluster, write it out if seqcount tells us we
 709                  * are operating sequentially, otherwise let the buf or
 710                  * update daemon handle it.
 711                  */
 712                 bdwrite(bp);
 713                 if (seqcount > 1)
 714                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
 715                 vp->v_clen = 0;
 716                 vp->v_cstart = lbn + 1;
 717         } else if (vm_page_count_severe()) {
 718                 /*
 719                  * We are low on memory, get it going NOW
 720                  */
 721                 bawrite(bp);
 722         } else {
 723                 /*
 724                  * In the middle of a cluster, so just delay the I/O for now.
 725                  */
 726                 bdwrite(bp);
 727         }
 728         vp->v_lastw = lbn;
 729         vp->v_lasta = bp->b_blkno;
 730 }
 731
 732
 733 /*
 734  * This is an awful lot like cluster_rbuild...wish they could be combined.
 735  * The last lbn argument is the current block on which I/O is being
 736  * performed.  Check to see that it doesn't fall in the middle of
 737  * the current block (if last_bp == NULL).
 738  */
 739 int
 740 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
 741     int gbflags)
 742 {
 743         struct buf *bp, *tbp;
 744         struct bufobj *bo;
 745         int i, j;
 746         int totalwritten = 0;
 747         int dbsize = btodb(size);
 748
 749         bo = &vp->v_bufobj;
 750         while (len > 0) {
 751                 /*
 752                  * If the buffer is not delayed-write (i.e. dirty), or it
 753                  * is delayed-write but either locked or inval, it cannot
 754                  * partake in the clustered write.
 755                  */
 756                 BO_LOCK(bo);
 757                 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
 758                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 759                         BO_UNLOCK(bo);
 760                         ++start_lbn;
 761                         --len;
 762                         continue;
 763                 }
 764                 if (BUF_LOCK(tbp,
 765                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_MTX(bo))) {
 766                         ++start_lbn;
 767                         --len;
 768                         continue;
 769                 }
 770                 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
 771                         BUF_UNLOCK(tbp);
 772                         ++start_lbn;
 773                         --len;
 774                         continue;
 775                 }
 776                 if (tbp->b_pin_count >  0) {
 777                         BUF_UNLOCK(tbp);
 778                         ++start_lbn;
 779                         --len;
 780                         continue;
 781                 }
 782                 bremfree(tbp);
 783                 tbp->b_flags &= ~B_DONE;
 784
 785                 /*
 786                  * Extra memory in the buffer, punt on this buffer.
 787                  * XXX we could handle this in most cases, but we would
 788                  * have to push the extra memory down to after our max
 789                  * possible cluster size and then potentially pull it back
 790                  * up if the cluster was terminated prematurely--too much
 791                  * hassle.
 792                  */
 793                 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
 794                      (B_CLUSTEROK | B_VMIO)) ||
 795                   (tbp->b_bcount != tbp->b_bufsize) ||
 796                   (tbp->b_bcount != size) ||
 797                   (len == 1) ||
 798                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 799                         totalwritten += tbp->b_bufsize;
 800                         bawrite(tbp);
 801                         ++start_lbn;
 802                         --len;
 803                         continue;
 804                 }
 805
 806                 /*
 807                  * We got a pbuf to make the cluster in.
 808                  * so initialise it.
 809                  */
 810                 TAILQ_INIT(&bp->b_cluster.cluster_head);
 811                 bp->b_bcount = 0;
 812                 bp->b_bufsize = 0;
 813                 bp->b_npages = 0;
 814                 if (tbp->b_wcred != NOCRED)
 815                         bp->b_wcred = crhold(tbp->b_wcred);
 816
 817                 bp->b_blkno = tbp->b_blkno;
 818                 bp->b_lblkno = tbp->b_lblkno;
 819                 bp->b_offset = tbp->b_offset;
 820
 821                 /*
 822                  * We are synthesizing a buffer out of vm_page_t's, but
 823                  * if the block size is not page aligned then the starting
 824                  * address may not be either.  Inherit the b_data offset
 825                  * from the original buffer.
 826                  */
 827                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 828                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 829                 bp->b_flags |= B_CLUSTER |
 830                                 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
 831                 bp->b_iodone = cluster_callback;
 832                 pbgetvp(vp, bp);
 833                 /*
 834                  * From this location in the file, scan forward to see
 835                  * if there are buffers with adjacent data that need to
 836                  * be written as well.
 837                  */
 838                 for (i = 0; i < len; ++i, ++start_lbn) {
 839                         if (i != 0) { /* If not the first buffer */
 840                                 /*
 841                                  * If the adjacent data is not even in core it
 842                                  * can't need to be written.
 843                                  */
 844                                 BO_LOCK(bo);
 845                                 if ((tbp = gbincore(bo, start_lbn)) == NULL ||
 846                                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 847                                         BO_UNLOCK(bo);
 848                                         break;
 849                                 }
 850
 851                                 /*
 852                                  * If it IS in core, but has different
 853                                  * characteristics, or is locked (which
 854                                  * means it could be undergoing a background
 855                                  * I/O or be in a weird state), then don't
 856                                  * cluster with it.
 857                                  */
 858                                 if (BUF_LOCK(tbp,
 859                                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
 860                                     BO_MTX(bo)))
 861                                         break;
 862
 863                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 864                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 865                                     != (B_DELWRI | B_CLUSTEROK |
 866                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 867                                     tbp->b_wcred != bp->b_wcred) {
 868                                         BUF_UNLOCK(tbp);
 869                                         break;
 870                                 }
 871
 872                                 /*
 873                                  * Check that the combined cluster
 874                                  * would make sense with regard to pages
 875                                  * and would not be too large
 876                                  */
 877                                 if ((tbp->b_bcount != size) ||
 878                                   ((bp->b_blkno + (dbsize * i)) !=
 879                                     tbp->b_blkno) ||
 880                                   ((tbp->b_npages + bp->b_npages) >
 881                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 882                                         BUF_UNLOCK(tbp);
 883                                         break;
 884                                 }
 885
 886                                 /*
 887                                  * Do not pull in pinned buffers.
 888                                  */
 889                                 if (tbp->b_pin_count > 0) {
 890                                         BUF_UNLOCK(tbp);
 891                                         break;
 892                                 }
 893
 894                                 /*
 895                                  * Ok, it's passed all the tests,
 896                                  * so remove it from the free list
 897                                  * and mark it busy. We will use it.
 898                                  */
 899                                 bremfree(tbp);
 900                                 tbp->b_flags &= ~B_DONE;
 901                         } /* end of code for non-first buffers only */
 902                         /*
 903                          * If the IO is via the VM then we do some
 904                          * special VM hackery (yuck).  Since the buffer's
 905                          * block size may not be page-aligned it is possible
 906                          * for a page to be shared between two buffers.  We
 907                          * have to get rid of the duplication when building
 908                          * the cluster.
 909                          */
 910                         if (tbp->b_flags & B_VMIO) {
 911                                 vm_page_t m;
 912
 913                                 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
 914                                 if (i != 0) { /* if not first buffer */
 915                                         for (j = 0; j < tbp->b_npages; j += 1) {
 916                                                 m = tbp->b_pages[j];
 917                                                 if (m->oflags & VPO_BUSY) {
 918                                                         VM_OBJECT_WUNLOCK(
 919                                                             tbp->b_object);
 920                                                         bqrelse(tbp);
 921                                                         goto finishcluster;
 922                                                 }
 923                                         }
 924                                 }
 925                                 for (j = 0; j < tbp->b_npages; j += 1) {
 926                                         m = tbp->b_pages[j];
 927                                         vm_page_io_start(m);
 928                                         vm_object_pip_add(m->object, 1);
 929                                         if ((bp->b_npages == 0) ||
 930                                           (bp->b_pages[bp->b_npages - 1] != m)) {
 931                                                 bp->b_pages[bp->b_npages] = m;
 932                                                 bp->b_npages++;
 933                                         }
 934                                 }
 935                                 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
 936                         }
 937                         bp->b_bcount += size;
 938                         bp->b_bufsize += size;
 939                         /*
 940                          * If any of the clustered buffers have their
 941                          * B_BARRIER flag set, transfer that request to
 942                          * the cluster.
 943                          */
 944                         bp->b_flags |= (tbp->b_flags & B_BARRIER);
 945                         tbp->b_flags &= ~(B_DONE | B_BARRIER);
 946                         tbp->b_flags |= B_ASYNC;
 947                         tbp->b_ioflags &= ~BIO_ERROR;
 948                         tbp->b_iocmd = BIO_WRITE;
 949                         bundirty(tbp);
 950                         reassignbuf(tbp);               /* put on clean list */
 951                         bufobj_wref(tbp->b_bufobj);
 952                         BUF_KERNPROC(tbp);
 953                         TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 954                                 tbp, b_cluster.cluster_entry);
 955                 }
 956         finishcluster:
 957                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 958                         (vm_page_t *) bp->b_pages, bp->b_npages);
 959                 if (bp->b_bufsize > bp->b_kvasize)
 960                         panic(
 961                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 962                             bp->b_bufsize, bp->b_kvasize);
 963                 bp->b_kvasize = bp->b_bufsize;
 964                 totalwritten += bp->b_bufsize;
 965                 bp->b_dirtyoff = 0;
 966                 bp->b_dirtyend = bp->b_bufsize;
 967                 bawrite(bp);
 968
 969                 len -= i;
 970         }
 971         return totalwritten;
 972 }
 973
 974 /*
 975  * Collect together all the buffers in a cluster.
 976  * Plus add one additional buffer.
 977  */
 978 static struct cluster_save *
 979 cluster_collectbufs(vp, last_bp)
 980         struct vnode *vp;
 981         struct buf *last_bp;
 982 {
 983         struct cluster_save *buflist;
 984         struct buf *bp;
 985         daddr_t lbn;
 986         int i, len;
 987
 988         len = vp->v_lastw - vp->v_cstart + 1;
 989         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
 990             M_SEGMENT, M_WAITOK);
 991         buflist->bs_nchildren = 0;
 992         buflist->bs_children = (struct buf **) (buflist + 1);
 993         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
 994                 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
 995                 buflist->bs_children[i] = bp;
 996                 if (bp->b_blkno == bp->b_lblkno)
 997                         VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
 998                                 NULL, NULL);
 999         }
1000         buflist->bs_children[i] = bp = last_bp;
1001         if (bp->b_blkno == bp->b_lblkno)
1002                 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1003         buflist->bs_nchildren = i + 1;
1004         return (buflist);
1005 }