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 *cluster_collectbufs(struct vnode *vp,
  65             struct buf *last_bp, int gbflags);
  66 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize,
  67             daddr_t lbn, daddr_t blkno, long size, int run, int gbflags,
  68             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         if (!unmapped_buf_allowed)
 101                 gbflags &= ~GB_UNMAPPED;
 102
 103         /*
 104          * Try to limit the amount of read-ahead by a few
 105          * ad-hoc parameters.  This needs work!!!
 106          */
 107         racluster = vp->v_mount->mnt_iosize_max / size;
 108         maxra = seqcount;
 109         maxra = min(read_max, maxra);
 110         maxra = min(nbuf/8, maxra);
 111         if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
 112                 maxra = (filesize / size) - lblkno;
 113
 114         /*
 115          * get the requested block
 116          */
 117         *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, gbflags);
 118         origblkno = lblkno;
 119
 120         /*
 121          * if it is in the cache, then check to see if the reads have been
 122          * sequential.  If they have, then try some read-ahead, otherwise
 123          * back-off on prospective read-aheads.
 124          */
 125         if (bp->b_flags & B_CACHE) {
 126                 if (!seqcount) {
 127                         return 0;
 128                 } else if ((bp->b_flags & B_RAM) == 0) {
 129                         return 0;
 130                 } else {
 131                         bp->b_flags &= ~B_RAM;
 132                         BO_LOCK(bo);
 133                         for (i = 1; i < maxra; i++) {
 134                                 /*
 135                                  * Stop if the buffer does not exist or it
 136                                  * is invalid (about to go away?)
 137                                  */
 138                                 rbp = gbincore(&vp->v_bufobj, lblkno+i);
 139                                 if (rbp == NULL || (rbp->b_flags & B_INVAL))
 140                                         break;
 141
 142                                 /*
 143                                  * Set another read-ahead mark so we know
 144                                  * to check again. (If we can lock the
 145                                  * buffer without waiting)
 146                                  */
 147                                 if ((((i % racluster) == (racluster - 1)) ||
 148                                     (i == (maxra - 1)))
 149                                     && (0 == BUF_LOCK(rbp,
 150                                         LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
 151                                         rbp->b_flags |= B_RAM;
 152                                         BUF_UNLOCK(rbp);
 153                                 }
 154                         }
 155                         BO_UNLOCK(bo);
 156                         if (i >= maxra) {
 157                                 return 0;
 158                         }
 159                         lblkno += i;
 160                 }
 161                 reqbp = bp = NULL;
 162         /*
 163          * If it isn't in the cache, then get a chunk from
 164          * disk if sequential, otherwise just get the block.
 165          */
 166         } else {
 167                 off_t firstread = bp->b_offset;
 168                 int nblks;
 169
 170                 KASSERT(bp->b_offset != NOOFFSET,
 171                     ("cluster_read: no buffer offset"));
 172
 173                 ncontig = 0;
 174
 175                 /*
 176                  * Compute the total number of blocks that we should read
 177                  * synchronously.
 178                  */
 179                 if (firstread + totread > filesize)
 180                         totread = filesize - firstread;
 181                 nblks = howmany(totread, size);
 182                 if (nblks > racluster)
 183                         nblks = racluster;
 184
 185                 /*
 186                  * Now compute the number of contiguous blocks.
 187                  */
 188                 if (nblks > 1) {
 189                         error = VOP_BMAP(vp, lblkno, NULL,
 190                                 &blkno, &ncontig, NULL);
 191                         /*
 192                          * If this failed to map just do the original block.
 193                          */
 194                         if (error || blkno == -1)
 195                                 ncontig = 0;
 196                 }
 197
 198                 /*
 199                  * If we have contiguous data available do a cluster
 200                  * otherwise just read the requested block.
 201                  */
 202                 if (ncontig) {
 203                         /* Account for our first block. */
 204                         ncontig = min(ncontig + 1, nblks);
 205                         if (ncontig < nblks)
 206                                 nblks = ncontig;
 207                         bp = cluster_rbuild(vp, filesize, lblkno,
 208                             blkno, size, nblks, gbflags, bp);
 209                         lblkno += (bp->b_bufsize / size);
 210                 } else {
 211                         bp->b_flags |= B_RAM;
 212                         bp->b_iocmd = BIO_READ;
 213                         lblkno += 1;
 214                 }
 215         }
 216
 217         /*
 218          * handle the synchronous read so that it is available ASAP.
 219          */
 220         if (bp) {
 221                 if ((bp->b_flags & B_CLUSTER) == 0) {
 222                         vfs_busy_pages(bp, 0);
 223                 }
 224                 bp->b_flags &= ~B_INVAL;
 225                 bp->b_ioflags &= ~BIO_ERROR;
 226                 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
 227                         BUF_KERNPROC(bp);
 228                 bp->b_iooffset = dbtob(bp->b_blkno);
 229                 bstrategy(bp);
 230                 curthread->td_ru.ru_inblock++;
 231         }
 232
 233         /*
 234          * If we have been doing sequential I/O, then do some read-ahead.
 235          */
 236         while (lblkno < (origblkno + maxra)) {
 237                 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
 238                 if (error)
 239                         break;
 240
 241                 if (blkno == -1)
 242                         break;
 243
 244                 /*
 245                  * We could throttle ncontig here by maxra but we might as
 246                  * well read the data if it is contiguous.  We're throttled
 247                  * by racluster anyway.
 248                  */
 249                 if (ncontig) {
 250                         ncontig = min(ncontig + 1, racluster);
 251                         rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
 252                             size, ncontig, gbflags, NULL);
 253                         lblkno += (rbp->b_bufsize / size);
 254                         if (rbp->b_flags & B_DELWRI) {
 255                                 bqrelse(rbp);
 256                                 continue;
 257                         }
 258                 } else {
 259                         rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
 260                         lblkno += 1;
 261                         if (rbp->b_flags & B_DELWRI) {
 262                                 bqrelse(rbp);
 263                                 continue;
 264                         }
 265                         rbp->b_flags |= B_ASYNC | B_RAM;
 266                         rbp->b_iocmd = BIO_READ;
 267                         rbp->b_blkno = blkno;
 268                 }
 269                 if (rbp->b_flags & B_CACHE) {
 270                         rbp->b_flags &= ~B_ASYNC;
 271                         bqrelse(rbp);
 272                         continue;
 273                 }
 274                 if ((rbp->b_flags & B_CLUSTER) == 0) {
 275                         vfs_busy_pages(rbp, 0);
 276                 }
 277                 rbp->b_flags &= ~B_INVAL;
 278                 rbp->b_ioflags &= ~BIO_ERROR;
 279                 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
 280                         BUF_KERNPROC(rbp);
 281                 rbp->b_iooffset = dbtob(rbp->b_blkno);
 282                 bstrategy(rbp);
 283                 curthread->td_ru.ru_inblock++;
 284         }
 285
 286         if (reqbp)
 287                 return (bufwait(reqbp));
 288         else
 289                 return (error);
 290 }
 291
 292 /*
 293  * If blocks are contiguous on disk, use this to provide clustered
 294  * read ahead.  We will read as many blocks as possible sequentially
 295  * and then parcel them up into logical blocks in the buffer hash table.
 296  */
 297 static struct buf *
 298 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
 299     daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
 300 {
 301         struct bufobj *bo;
 302         struct buf *bp, *tbp;
 303         daddr_t bn;
 304         off_t off;
 305         long tinc, tsize;
 306         int i, inc, j, toff;
 307
 308         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 309             ("cluster_rbuild: size %ld != f_iosize %jd\n",
 310             size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
 311
 312         /*
 313          * avoid a division
 314          */
 315         while ((u_quad_t) size * (lbn + run) > filesize) {
 316                 --run;
 317         }
 318
 319         if (fbp) {
 320                 tbp = fbp;
 321                 tbp->b_iocmd = BIO_READ;
 322         } else {
 323                 tbp = getblk(vp, lbn, size, 0, 0, gbflags);
 324                 if (tbp->b_flags & B_CACHE)
 325                         return tbp;
 326                 tbp->b_flags |= B_ASYNC | B_RAM;
 327                 tbp->b_iocmd = BIO_READ;
 328         }
 329         tbp->b_blkno = blkno;
 330         if( (tbp->b_flags & B_MALLOC) ||
 331                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
 332                 return tbp;
 333
 334         bp = trypbuf(&cluster_pbuf_freecnt);
 335         if (bp == 0)
 336                 return tbp;
 337
 338         /*
 339          * We are synthesizing a buffer out of vm_page_t's, but
 340          * if the block size is not page aligned then the starting
 341          * address may not be either.  Inherit the b_data offset
 342          * from the original buffer.
 343          */
 344         bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
 345         if ((gbflags & GB_UNMAPPED) != 0) {
 346                 bp->b_flags |= B_UNMAPPED;
 347                 bp->b_data = unmapped_buf;
 348         } else {
 349                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 350                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 351         }
 352         bp->b_iocmd = BIO_READ;
 353         bp->b_iodone = cluster_callback;
 354         bp->b_blkno = blkno;
 355         bp->b_lblkno = lbn;
 356         bp->b_offset = tbp->b_offset;
 357         KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
 358         pbgetvp(vp, bp);
 359
 360         TAILQ_INIT(&bp->b_cluster.cluster_head);
 361
 362         bp->b_bcount = 0;
 363         bp->b_bufsize = 0;
 364         bp->b_npages = 0;
 365
 366         inc = btodb(size);
 367         bo = &vp->v_bufobj;
 368         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
 369                 if (i != 0) {
 370                         if ((bp->b_npages * PAGE_SIZE) +
 371                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 372                                 break;
 373                         }
 374
 375                         tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
 376                             (gbflags & GB_UNMAPPED));
 377
 378                         /* Don't wait around for locked bufs. */
 379                         if (tbp == NULL)
 380                                 break;
 381
 382                         /*
 383                          * Stop scanning if the buffer is fully valid
 384                          * (marked B_CACHE), or locked (may be doing a
 385                          * background write), or if the buffer is not
 386                          * VMIO backed.  The clustering code can only deal
 387                          * with VMIO-backed buffers.
 388                          */
 389                         BO_LOCK(bo);
 390                         if ((tbp->b_vflags & BV_BKGRDINPROG) ||
 391                             (tbp->b_flags & B_CACHE) ||
 392                             (tbp->b_flags & B_VMIO) == 0) {
 393                                 BO_UNLOCK(bo);
 394                                 bqrelse(tbp);
 395                                 break;
 396                         }
 397                         BO_UNLOCK(bo);
 398
 399                         /*
 400                          * The buffer must be completely invalid in order to
 401                          * take part in the cluster.  If it is partially valid
 402                          * then we stop.
 403                          */
 404                         off = tbp->b_offset;
 405                         tsize = size;
 406                         VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
 407                         for (j = 0; tsize > 0; j++) {
 408                                 toff = off & PAGE_MASK;
 409                                 tinc = tsize;
 410                                 if (toff + tinc > PAGE_SIZE)
 411                                         tinc = PAGE_SIZE - toff;
 412                                 VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object);
 413                                 if ((tbp->b_pages[j]->valid &
 414                                     vm_page_bits(toff, tinc)) != 0)
 415                                         break;
 416                                 off += tinc;
 417                                 tsize -= tinc;
 418                         }
 419                         VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
 420                         if (tsize > 0) {
 421                                 bqrelse(tbp);
 422                                 break;
 423                         }
 424
 425                         /*
 426                          * Set a read-ahead mark as appropriate
 427                          */
 428                         if ((fbp && (i == 1)) || (i == (run - 1)))
 429                                 tbp->b_flags |= B_RAM;
 430
 431                         /*
 432                          * Set the buffer up for an async read (XXX should
 433                          * we do this only if we do not wind up brelse()ing?).
 434                          * Set the block number if it isn't set, otherwise
 435                          * if it is make sure it matches the block number we
 436                          * expect.
 437                          */
 438                         tbp->b_flags |= B_ASYNC;
 439                         tbp->b_iocmd = BIO_READ;
 440                         if (tbp->b_blkno == tbp->b_lblkno) {
 441                                 tbp->b_blkno = bn;
 442                         } else if (tbp->b_blkno != bn) {
 443                                 brelse(tbp);
 444                                 break;
 445                         }
 446                 }
 447                 /*
 448                  * XXX fbp from caller may not be B_ASYNC, but we are going
 449                  * to biodone() it in cluster_callback() anyway
 450                  */
 451                 BUF_KERNPROC(tbp);
 452                 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 453                         tbp, b_cluster.cluster_entry);
 454                 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
 455                 for (j = 0; j < tbp->b_npages; j += 1) {
 456                         vm_page_t m;
 457                         m = tbp->b_pages[j];
 458                         vm_page_io_start(m);
 459                         vm_object_pip_add(m->object, 1);
 460                         if ((bp->b_npages == 0) ||
 461                                 (bp->b_pages[bp->b_npages-1] != m)) {
 462                                 bp->b_pages[bp->b_npages] = m;
 463                                 bp->b_npages++;
 464                         }
 465                         if (m->valid == VM_PAGE_BITS_ALL)
 466                                 tbp->b_pages[j] = bogus_page;
 467                 }
 468                 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
 469                 /*
 470                  * Don't inherit tbp->b_bufsize as it may be larger due to
 471                  * a non-page-aligned size.  Instead just aggregate using
 472                  * 'size'.
 473                  */
 474                 if (tbp->b_bcount != size)
 475                         printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
 476                 if (tbp->b_bufsize != size)
 477                         printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
 478                 bp->b_bcount += size;
 479                 bp->b_bufsize += size;
 480         }
 481
 482         /*
 483          * Fully valid pages in the cluster are already good and do not need
 484          * to be re-read from disk.  Replace the page with bogus_page
 485          */
 486         VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
 487         for (j = 0; j < bp->b_npages; j++) {
 488                 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object);
 489                 if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
 490                         bp->b_pages[j] = bogus_page;
 491         }
 492         VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
 493         if (bp->b_bufsize > bp->b_kvasize)
 494                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 495                     bp->b_bufsize, bp->b_kvasize);
 496         bp->b_kvasize = bp->b_bufsize;
 497
 498         if ((bp->b_flags & B_UNMAPPED) == 0) {
 499                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 500                     (vm_page_t *)bp->b_pages, bp->b_npages);
 501         }
 502         return (bp);
 503 }
 504
 505 /*
 506  * Cleanup after a clustered read or write.
 507  * This is complicated by the fact that any of the buffers might have
 508  * extra memory (if there were no empty buffer headers at allocbuf time)
 509  * that we will need to shift around.
 510  */
 511 static void
 512 cluster_callback(bp)
 513         struct buf *bp;
 514 {
 515         struct buf *nbp, *tbp;
 516         int error = 0;
 517
 518         /*
 519          * Must propogate errors to all the components.
 520          */
 521         if (bp->b_ioflags & BIO_ERROR)
 522                 error = bp->b_error;
 523
 524         if ((bp->b_flags & B_UNMAPPED) == 0) {
 525                 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
 526                     bp->b_npages);
 527         }
 528         /*
 529          * Move memory from the large cluster buffer into the component
 530          * buffers and mark IO as done on these.
 531          */
 532         for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
 533                 tbp; tbp = nbp) {
 534                 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
 535                 if (error) {
 536                         tbp->b_ioflags |= BIO_ERROR;
 537                         tbp->b_error = error;
 538                 } else {
 539                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 540                         tbp->b_flags &= ~B_INVAL;
 541                         tbp->b_ioflags &= ~BIO_ERROR;
 542                         /*
 543                          * XXX the bdwrite()/bqrelse() issued during
 544                          * cluster building clears B_RELBUF (see bqrelse()
 545                          * comment).  If direct I/O was specified, we have
 546                          * to restore it here to allow the buffer and VM
 547                          * to be freed.
 548                          */
 549                         if (tbp->b_flags & B_DIRECT)
 550                                 tbp->b_flags |= B_RELBUF;
 551                 }
 552                 bufdone(tbp);
 553         }
 554         pbrelvp(bp);
 555         relpbuf(bp, &cluster_pbuf_freecnt);
 556 }
 557
 558 /*
 559  *      cluster_wbuild_wb:
 560  *
 561  *      Implement modified write build for cluster.
 562  *
 563  *              write_behind = 0        write behind disabled
 564  *              write_behind = 1        write behind normal (default)
 565  *              write_behind = 2        write behind backed-off
 566  */
 567
 568 static __inline int
 569 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
 570     int gbflags)
 571 {
 572         int r = 0;
 573
 574         switch (write_behind) {
 575         case 2:
 576                 if (start_lbn < len)
 577                         break;
 578                 start_lbn -= len;
 579                 /* FALLTHROUGH */
 580         case 1:
 581                 r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
 582                 /* FALLTHROUGH */
 583         default:
 584                 /* FALLTHROUGH */
 585                 break;
 586         }
 587         return(r);
 588 }
 589
 590 /*
 591  * Do clustered write for FFS.
 592  *
 593  * Three cases:
 594  *      1. Write is not sequential (write asynchronously)
 595  *      Write is sequential:
 596  *      2.      beginning of cluster - begin cluster
 597  *      3.      middle of a cluster - add to cluster
 598  *      4.      end of a cluster - asynchronously write cluster
 599  */
 600 void
 601 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount,
 602     int gbflags)
 603 {
 604         daddr_t lbn;
 605         int maxclen, cursize;
 606         int lblocksize;
 607         int async;
 608
 609         if (!unmapped_buf_allowed)
 610                 gbflags &= ~GB_UNMAPPED;
 611
 612         if (vp->v_type == VREG) {
 613                 async = DOINGASYNC(vp);
 614                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 615         } else {
 616                 async = 0;
 617                 lblocksize = bp->b_bufsize;
 618         }
 619         lbn = bp->b_lblkno;
 620         KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
 621
 622         /* Initialize vnode to beginning of file. */
 623         if (lbn == 0)
 624                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 625
 626         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
 627             (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
 628                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
 629                 if (vp->v_clen != 0) {
 630                         /*
 631                          * Next block is not sequential.
 632                          *
 633                          * If we are not writing at end of file, the process
 634                          * seeked to another point in the file since its last
 635                          * write, or we have reached our maximum cluster size,
 636                          * then push the previous cluster. Otherwise try
 637                          * reallocating to make it sequential.
 638                          *
 639                          * Change to algorithm: only push previous cluster if
 640                          * it was sequential from the point of view of the
 641                          * seqcount heuristic, otherwise leave the buffer
 642                          * intact so we can potentially optimize the I/O
 643                          * later on in the buf_daemon or update daemon
 644                          * flush.
 645                          */
 646                         cursize = vp->v_lastw - vp->v_cstart + 1;
 647                         if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
 648                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
 649                                 if (!async && seqcount > 0) {
 650                                         cluster_wbuild_wb(vp, lblocksize,
 651                                             vp->v_cstart, cursize, gbflags);
 652                                 }
 653                         } else {
 654                                 struct buf **bpp, **endbp;
 655                                 struct cluster_save *buflist;
 656
 657                                 buflist = cluster_collectbufs(vp, bp, gbflags);
 658                                 endbp = &buflist->bs_children
 659                                     [buflist->bs_nchildren - 1];
 660                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 661                                         /*
 662                                          * Failed, push the previous cluster
 663                                          * if *really* writing sequentially
 664                                          * in the logical file (seqcount > 1),
 665                                          * otherwise delay it in the hopes that
 666                                          * the low level disk driver can
 667                                          * optimize the write ordering.
 668                                          */
 669                                         for (bpp = buflist->bs_children;
 670                                              bpp < endbp; bpp++)
 671                                                 brelse(*bpp);
 672                                         free(buflist, M_SEGMENT);
 673                                         if (seqcount > 1) {
 674                                                 cluster_wbuild_wb(vp,
 675                                                     lblocksize, vp->v_cstart,
 676                                                     cursize, gbflags);
 677                                         }
 678                                 } else {
 679                                         /*
 680                                          * Succeeded, keep building cluster.
 681                                          */
 682                                         for (bpp = buflist->bs_children;
 683                                              bpp <= endbp; bpp++)
 684                                                 bdwrite(*bpp);
 685                                         free(buflist, M_SEGMENT);
 686                                         vp->v_lastw = lbn;
 687                                         vp->v_lasta = bp->b_blkno;
 688                                         return;
 689                                 }
 690                         }
 691                 }
 692                 /*
 693                  * Consider beginning a cluster. If at end of file, make
 694                  * cluster as large as possible, otherwise find size of
 695                  * existing cluster.
 696                  */
 697                 if ((vp->v_type == VREG) &&
 698                         ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
 699                     (bp->b_blkno == bp->b_lblkno) &&
 700                     (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
 701                      bp->b_blkno == -1)) {
 702                         bawrite(bp);
 703                         vp->v_clen = 0;
 704                         vp->v_lasta = bp->b_blkno;
 705                         vp->v_cstart = lbn + 1;
 706                         vp->v_lastw = lbn;
 707                         return;
 708                 }
 709                 vp->v_clen = maxclen;
 710                 if (!async && maxclen == 0) {   /* I/O not contiguous */
 711                         vp->v_cstart = lbn + 1;
 712                         bawrite(bp);
 713                 } else {        /* Wait for rest of cluster */
 714                         vp->v_cstart = lbn;
 715                         bdwrite(bp);
 716                 }
 717         } else if (lbn == vp->v_cstart + vp->v_clen) {
 718                 /*
 719                  * At end of cluster, write it out if seqcount tells us we
 720                  * are operating sequentially, otherwise let the buf or
 721                  * update daemon handle it.
 722                  */
 723                 bdwrite(bp);
 724                 if (seqcount > 1) {
 725                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
 726                             vp->v_clen + 1, gbflags);
 727                 }
 728                 vp->v_clen = 0;
 729                 vp->v_cstart = lbn + 1;
 730         } else if (vm_page_count_severe()) {
 731                 /*
 732                  * We are low on memory, get it going NOW
 733                  */
 734                 bawrite(bp);
 735         } else {
 736                 /*
 737                  * In the middle of a cluster, so just delay the I/O for now.
 738                  */
 739                 bdwrite(bp);
 740         }
 741         vp->v_lastw = lbn;
 742         vp->v_lasta = bp->b_blkno;
 743 }
 744
 745
 746 /*
 747  * This is an awful lot like cluster_rbuild...wish they could be combined.
 748  * The last lbn argument is the current block on which I/O is being
 749  * performed.  Check to see that it doesn't fall in the middle of
 750  * the current block (if last_bp == NULL).
 751  */
 752 int
 753 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
 754     int gbflags)
 755 {
 756         struct buf *bp, *tbp;
 757         struct bufobj *bo;
 758         int i, j;
 759         int totalwritten = 0;
 760         int dbsize = btodb(size);
 761
 762         if (!unmapped_buf_allowed)
 763                 gbflags &= ~GB_UNMAPPED;
 764
 765         bo = &vp->v_bufobj;
 766         while (len > 0) {
 767                 /*
 768                  * If the buffer is not delayed-write (i.e. dirty), or it
 769                  * is delayed-write but either locked or inval, it cannot
 770                  * partake in the clustered write.
 771                  */
 772                 BO_LOCK(bo);
 773                 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
 774                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 775                         BO_UNLOCK(bo);
 776                         ++start_lbn;
 777                         --len;
 778                         continue;
 779                 }
 780                 if (BUF_LOCK(tbp,
 781                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_MTX(bo))) {
 782                         ++start_lbn;
 783                         --len;
 784                         continue;
 785                 }
 786                 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
 787                         BUF_UNLOCK(tbp);
 788                         ++start_lbn;
 789                         --len;
 790                         continue;
 791                 }
 792                 if (tbp->b_pin_count >  0) {
 793                         BUF_UNLOCK(tbp);
 794                         ++start_lbn;
 795                         --len;
 796                         continue;
 797                 }
 798                 bremfree(tbp);
 799                 tbp->b_flags &= ~B_DONE;
 800
 801                 /*
 802                  * Extra memory in the buffer, punt on this buffer.
 803                  * XXX we could handle this in most cases, but we would
 804                  * have to push the extra memory down to after our max
 805                  * possible cluster size and then potentially pull it back
 806                  * up if the cluster was terminated prematurely--too much
 807                  * hassle.
 808                  */
 809                 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
 810                      (B_CLUSTEROK | B_VMIO)) ||
 811                   (tbp->b_bcount != tbp->b_bufsize) ||
 812                   (tbp->b_bcount != size) ||
 813                   (len == 1) ||
 814                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 815                         totalwritten += tbp->b_bufsize;
 816                         bawrite(tbp);
 817                         ++start_lbn;
 818                         --len;
 819                         continue;
 820                 }
 821
 822                 /*
 823                  * We got a pbuf to make the cluster in.
 824                  * so initialise it.
 825                  */
 826                 TAILQ_INIT(&bp->b_cluster.cluster_head);
 827                 bp->b_bcount = 0;
 828                 bp->b_bufsize = 0;
 829                 bp->b_npages = 0;
 830                 if (tbp->b_wcred != NOCRED)
 831                         bp->b_wcred = crhold(tbp->b_wcred);
 832
 833                 bp->b_blkno = tbp->b_blkno;
 834                 bp->b_lblkno = tbp->b_lblkno;
 835                 bp->b_offset = tbp->b_offset;
 836
 837                 /*
 838                  * We are synthesizing a buffer out of vm_page_t's, but
 839                  * if the block size is not page aligned then the starting
 840                  * address may not be either.  Inherit the b_data offset
 841                  * from the original buffer.
 842                  */
 843                 if ((gbflags & GB_UNMAPPED) == 0 ||
 844                     (tbp->b_flags & B_VMIO) == 0) {
 845                         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 846                             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 847                 } else {
 848                         bp->b_flags |= B_UNMAPPED;
 849                         bp->b_data = unmapped_buf;
 850                 }
 851                 bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
 852                     B_NEEDCOMMIT));
 853                 bp->b_iodone = cluster_callback;
 854                 pbgetvp(vp, bp);
 855                 /*
 856                  * From this location in the file, scan forward to see
 857                  * if there are buffers with adjacent data that need to
 858                  * be written as well.
 859                  */
 860                 for (i = 0; i < len; ++i, ++start_lbn) {
 861                         if (i != 0) { /* If not the first buffer */
 862                                 /*
 863                                  * If the adjacent data is not even in core it
 864                                  * can't need to be written.
 865                                  */
 866                                 BO_LOCK(bo);
 867                                 if ((tbp = gbincore(bo, start_lbn)) == NULL ||
 868                                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 869                                         BO_UNLOCK(bo);
 870                                         break;
 871                                 }
 872
 873                                 /*
 874                                  * If it IS in core, but has different
 875                                  * characteristics, or is locked (which
 876                                  * means it could be undergoing a background
 877                                  * I/O or be in a weird state), then don't
 878                                  * cluster with it.
 879                                  */
 880                                 if (BUF_LOCK(tbp,
 881                                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
 882                                     BO_MTX(bo)))
 883                                         break;
 884
 885                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 886                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 887                                     != (B_DELWRI | B_CLUSTEROK |
 888                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 889                                     tbp->b_wcred != bp->b_wcred) {
 890                                         BUF_UNLOCK(tbp);
 891                                         break;
 892                                 }
 893
 894                                 /*
 895                                  * Check that the combined cluster
 896                                  * would make sense with regard to pages
 897                                  * and would not be too large
 898                                  */
 899                                 if ((tbp->b_bcount != size) ||
 900                                   ((bp->b_blkno + (dbsize * i)) !=
 901                                     tbp->b_blkno) ||
 902                                   ((tbp->b_npages + bp->b_npages) >
 903                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 904                                         BUF_UNLOCK(tbp);
 905                                         break;
 906                                 }
 907
 908                                 /*
 909                                  * Do not pull in pinned buffers.
 910                                  */
 911                                 if (tbp->b_pin_count > 0) {
 912                                         BUF_UNLOCK(tbp);
 913                                         break;
 914                                 }
 915
 916                                 /*
 917                                  * Ok, it's passed all the tests,
 918                                  * so remove it from the free list
 919                                  * and mark it busy. We will use it.
 920                                  */
 921                                 bremfree(tbp);
 922                                 tbp->b_flags &= ~B_DONE;
 923                         } /* end of code for non-first buffers only */
 924                         /*
 925                          * If the IO is via the VM then we do some
 926                          * special VM hackery (yuck).  Since the buffer's
 927                          * block size may not be page-aligned it is possible
 928                          * for a page to be shared between two buffers.  We
 929                          * have to get rid of the duplication when building
 930                          * the cluster.
 931                          */
 932                         if (tbp->b_flags & B_VMIO) {
 933                                 vm_page_t m;
 934
 935                                 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
 936                                 if (i != 0) { /* if not first buffer */
 937                                         for (j = 0; j < tbp->b_npages; j += 1) {
 938                                                 m = tbp->b_pages[j];
 939                                                 if (m->oflags & VPO_BUSY) {
 940                                                         VM_OBJECT_WUNLOCK(
 941                                                             tbp->b_object);
 942                                                         bqrelse(tbp);
 943                                                         goto finishcluster;
 944                                                 }
 945                                         }
 946                                 }
 947                                 for (j = 0; j < tbp->b_npages; j += 1) {
 948                                         m = tbp->b_pages[j];
 949                                         vm_page_io_start(m);
 950                                         vm_object_pip_add(m->object, 1);
 951                                         if ((bp->b_npages == 0) ||
 952                                           (bp->b_pages[bp->b_npages - 1] != m)) {
 953                                                 bp->b_pages[bp->b_npages] = m;
 954                                                 bp->b_npages++;
 955                                         }
 956                                 }
 957                                 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
 958                         }
 959                         bp->b_bcount += size;
 960                         bp->b_bufsize += size;
 961                         /*
 962                          * If any of the clustered buffers have their
 963                          * B_BARRIER flag set, transfer that request to
 964                          * the cluster.
 965                          */
 966                         bp->b_flags |= (tbp->b_flags & B_BARRIER);
 967                         tbp->b_flags &= ~(B_DONE | B_BARRIER);
 968                         tbp->b_flags |= B_ASYNC;
 969                         tbp->b_ioflags &= ~BIO_ERROR;
 970                         tbp->b_iocmd = BIO_WRITE;
 971                         bundirty(tbp);
 972                         reassignbuf(tbp);               /* put on clean list */
 973                         bufobj_wref(tbp->b_bufobj);
 974                         BUF_KERNPROC(tbp);
 975                         TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 976                                 tbp, b_cluster.cluster_entry);
 977                 }
 978         finishcluster:
 979                 if ((bp->b_flags & B_UNMAPPED) == 0) {
 980                         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 981                             (vm_page_t *)bp->b_pages, bp->b_npages);
 982                 }
 983                 if (bp->b_bufsize > bp->b_kvasize)
 984                         panic(
 985                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 986                             bp->b_bufsize, bp->b_kvasize);
 987                 bp->b_kvasize = bp->b_bufsize;
 988                 totalwritten += bp->b_bufsize;
 989                 bp->b_dirtyoff = 0;
 990                 bp->b_dirtyend = bp->b_bufsize;
 991                 bawrite(bp);
 992
 993                 len -= i;
 994         }
 995         return totalwritten;
 996 }
 997
 998 /*
 999  * Collect together all the buffers in a cluster.
1000  * Plus add one additional buffer.
1001  */
1002 static struct cluster_save *
1003 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags)
1004 {
1005         struct cluster_save *buflist;
1006         struct buf *bp;
1007         daddr_t lbn;
1008         int i, len;
1009
1010         len = vp->v_lastw - vp->v_cstart + 1;
1011         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1012             M_SEGMENT, M_WAITOK);
1013         buflist->bs_nchildren = 0;
1014         buflist->bs_children = (struct buf **) (buflist + 1);
1015         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1016                 (void)bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1017                     gbflags, &bp);
1018                 buflist->bs_children[i] = bp;
1019                 if (bp->b_blkno == bp->b_lblkno)
1020                         VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1021                                 NULL, NULL);
1022         }
1023         buflist->bs_children[i] = bp = last_bp;
1024         if (bp->b_blkno == bp->b_lblkno)
1025                 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1026         buflist->bs_nchildren = i + 1;
1027         return (buflist);
1028 }