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/vmmeter.h>
  50 #include <vm/vm.h>
  51 #include <vm/vm_object.h>
  52 #include <vm/vm_page.h>
  53 #include <sys/sysctl.h>
  54
  55 #if defined(CLUSTERDEBUG)
  56 static int      rcluster= 0;
  57 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0,
  58     "Debug VFS clustering code");
  59 #endif
  60
  61 static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer");
  62
  63 static struct cluster_save *cluster_collectbufs(struct vnode *vp,
  64             struct buf *last_bp, int gbflags);
  65 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize,
  66             daddr_t lbn, daddr_t blkno, long size, int run, int gbflags,
  67             struct buf *fbp);
  68 static void cluster_callback(struct buf *);
  69
  70 static int write_behind = 1;
  71 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
  72     "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
  73
  74 static int read_max = 64;
  75 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
  76     "Cluster read-ahead max block count");
  77
  78 static int read_min = 1;
  79 SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0,
  80     "Cluster read min block count");
  81
  82 /* Page expended to mark partially backed buffers */
  83 extern vm_page_t        bogus_page;
  84
  85 /*
  86  * Read data to a buf, including read-ahead if we find this to be beneficial.
  87  * cluster_read replaces bread.
  88  */
  89 int
  90 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
  91     struct ucred *cred, long totread, int seqcount, struct buf **bpp)
  92 {
  93
  94         return (cluster_read_gb(vp, filesize, lblkno, size, cred, totread,
  95             seqcount, 0, bpp));
  96 }
  97
  98 int
  99 cluster_read_gb(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
 100     struct ucred *cred, long totread, int seqcount, int gbflags,
 101     struct buf **bpp)
 102 {
 103         struct buf *bp, *rbp, *reqbp;
 104         struct bufobj *bo;
 105         daddr_t blkno, origblkno;
 106         int maxra, racluster;
 107         int error, ncontig;
 108         int i;
 109
 110         error = 0;
 111         bo = &vp->v_bufobj;
 112         if (!unmapped_buf_allowed)
 113                 gbflags &= ~GB_UNMAPPED;
 114
 115         /*
 116          * Try to limit the amount of read-ahead by a few
 117          * ad-hoc parameters.  This needs work!!!
 118          */
 119         racluster = vp->v_mount->mnt_iosize_max / size;
 120         maxra = seqcount;
 121         maxra = min(read_max, maxra);
 122         maxra = min(nbuf/8, maxra);
 123         if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
 124                 maxra = (filesize / size) - lblkno;
 125
 126         /*
 127          * get the requested block
 128          */
 129         *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, gbflags);
 130         origblkno = lblkno;
 131
 132         /*
 133          * if it is in the cache, then check to see if the reads have been
 134          * sequential.  If they have, then try some read-ahead, otherwise
 135          * back-off on prospective read-aheads.
 136          */
 137         if (bp->b_flags & B_CACHE) {
 138                 if (!seqcount) {
 139                         return 0;
 140                 } else if ((bp->b_flags & B_RAM) == 0) {
 141                         return 0;
 142                 } else {
 143                         bp->b_flags &= ~B_RAM;
 144                         BO_LOCK(bo);
 145                         for (i = 1; i < maxra; i++) {
 146                                 /*
 147                                  * Stop if the buffer does not exist or it
 148                                  * is invalid (about to go away?)
 149                                  */
 150                                 rbp = gbincore(&vp->v_bufobj, lblkno+i);
 151                                 if (rbp == NULL || (rbp->b_flags & B_INVAL))
 152                                         break;
 153
 154                                 /*
 155                                  * Set another read-ahead mark so we know
 156                                  * to check again. (If we can lock the
 157                                  * buffer without waiting)
 158                                  */
 159                                 if ((((i % racluster) == (racluster - 1)) ||
 160                                     (i == (maxra - 1)))
 161                                     && (0 == BUF_LOCK(rbp,
 162                                         LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
 163                                         rbp->b_flags |= B_RAM;
 164                                         BUF_UNLOCK(rbp);
 165                                 }
 166                         }
 167                         BO_UNLOCK(bo);
 168                         if (i >= maxra) {
 169                                 return 0;
 170                         }
 171                         lblkno += i;
 172                 }
 173                 reqbp = bp = NULL;
 174         /*
 175          * If it isn't in the cache, then get a chunk from
 176          * disk if sequential, otherwise just get the block.
 177          */
 178         } else {
 179                 off_t firstread = bp->b_offset;
 180                 int nblks;
 181                 long minread;
 182
 183                 KASSERT(bp->b_offset != NOOFFSET,
 184                     ("cluster_read: no buffer offset"));
 185
 186                 ncontig = 0;
 187
 188                 /*
 189                  * Adjust totread if needed
 190                  */
 191                 minread = read_min * size;
 192                 if (minread > totread)
 193                         totread = minread;
 194
 195                 /*
 196                  * Compute the total number of blocks that we should read
 197                  * synchronously.
 198                  */
 199                 if (firstread + totread > filesize)
 200                         totread = filesize - firstread;
 201                 nblks = howmany(totread, size);
 202                 if (nblks > racluster)
 203                         nblks = racluster;
 204
 205                 /*
 206                  * Now compute the number of contiguous blocks.
 207                  */
 208                 if (nblks > 1) {
 209                         error = VOP_BMAP(vp, lblkno, NULL,
 210                                 &blkno, &ncontig, NULL);
 211                         /*
 212                          * If this failed to map just do the original block.
 213                          */
 214                         if (error || blkno == -1)
 215                                 ncontig = 0;
 216                 }
 217
 218                 /*
 219                  * If we have contiguous data available do a cluster
 220                  * otherwise just read the requested block.
 221                  */
 222                 if (ncontig) {
 223                         /* Account for our first block. */
 224                         ncontig = min(ncontig + 1, nblks);
 225                         if (ncontig < nblks)
 226                                 nblks = ncontig;
 227                         bp = cluster_rbuild(vp, filesize, lblkno,
 228                             blkno, size, nblks, gbflags, bp);
 229                         lblkno += (bp->b_bufsize / size);
 230                 } else {
 231                         bp->b_flags |= B_RAM;
 232                         bp->b_iocmd = BIO_READ;
 233                         lblkno += 1;
 234                 }
 235         }
 236
 237         /*
 238          * handle the synchronous read so that it is available ASAP.
 239          */
 240         if (bp) {
 241                 if ((bp->b_flags & B_CLUSTER) == 0) {
 242                         vfs_busy_pages(bp, 0);
 243                 }
 244                 bp->b_flags &= ~B_INVAL;
 245                 bp->b_ioflags &= ~BIO_ERROR;
 246                 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
 247                         BUF_KERNPROC(bp);
 248                 bp->b_iooffset = dbtob(bp->b_blkno);
 249                 bstrategy(bp);
 250                 curthread->td_ru.ru_inblock++;
 251         }
 252
 253         /*
 254          * If we have been doing sequential I/O, then do some read-ahead.
 255          */
 256         while (lblkno < (origblkno + maxra)) {
 257                 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
 258                 if (error)
 259                         break;
 260
 261                 if (blkno == -1)
 262                         break;
 263
 264                 /*
 265                  * We could throttle ncontig here by maxra but we might as
 266                  * well read the data if it is contiguous.  We're throttled
 267                  * by racluster anyway.
 268                  */
 269                 if (ncontig) {
 270                         ncontig = min(ncontig + 1, racluster);
 271                         rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
 272                             size, ncontig, gbflags, NULL);
 273                         lblkno += (rbp->b_bufsize / size);
 274                         if (rbp->b_flags & B_DELWRI) {
 275                                 bqrelse(rbp);
 276                                 continue;
 277                         }
 278                 } else {
 279                         rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
 280                         lblkno += 1;
 281                         if (rbp->b_flags & B_DELWRI) {
 282                                 bqrelse(rbp);
 283                                 continue;
 284                         }
 285                         rbp->b_flags |= B_ASYNC | B_RAM;
 286                         rbp->b_iocmd = BIO_READ;
 287                         rbp->b_blkno = blkno;
 288                 }
 289                 if (rbp->b_flags & B_CACHE) {
 290                         rbp->b_flags &= ~B_ASYNC;
 291                         bqrelse(rbp);
 292                         continue;
 293                 }
 294                 if ((rbp->b_flags & B_CLUSTER) == 0) {
 295                         vfs_busy_pages(rbp, 0);
 296                 }
 297                 rbp->b_flags &= ~B_INVAL;
 298                 rbp->b_ioflags &= ~BIO_ERROR;
 299                 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
 300                         BUF_KERNPROC(rbp);
 301                 rbp->b_iooffset = dbtob(rbp->b_blkno);
 302                 bstrategy(rbp);
 303                 curthread->td_ru.ru_inblock++;
 304         }
 305
 306         if (reqbp)
 307                 return (bufwait(reqbp));
 308         else
 309                 return (error);
 310 }
 311
 312 /*
 313  * If blocks are contiguous on disk, use this to provide clustered
 314  * read ahead.  We will read as many blocks as possible sequentially
 315  * and then parcel them up into logical blocks in the buffer hash table.
 316  */
 317 static struct buf *
 318 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
 319     daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
 320 {
 321         struct bufobj *bo;
 322         struct buf *bp, *tbp;
 323         daddr_t bn;
 324         off_t off;
 325         long tinc, tsize;
 326         int i, inc, j, toff;
 327
 328         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 329             ("cluster_rbuild: size %ld != filesize %jd\n",
 330             size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
 331
 332         /*
 333          * avoid a division
 334          */
 335         while ((u_quad_t) size * (lbn + run) > filesize) {
 336                 --run;
 337         }
 338
 339         if (fbp) {
 340                 tbp = fbp;
 341                 tbp->b_iocmd = BIO_READ;
 342         } else {
 343                 tbp = getblk(vp, lbn, size, 0, 0, gbflags);
 344                 if (tbp->b_flags & B_CACHE)
 345                         return tbp;
 346                 tbp->b_flags |= B_ASYNC | B_RAM;
 347                 tbp->b_iocmd = BIO_READ;
 348         }
 349         tbp->b_blkno = blkno;
 350         if( (tbp->b_flags & B_MALLOC) ||
 351                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
 352                 return tbp;
 353
 354         bp = trypbuf(&cluster_pbuf_freecnt);
 355         if (bp == 0)
 356                 return tbp;
 357
 358         /*
 359          * We are synthesizing a buffer out of vm_page_t's, but
 360          * if the block size is not page aligned then the starting
 361          * address may not be either.  Inherit the b_data offset
 362          * from the original buffer.
 363          */
 364         bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
 365         if ((gbflags & GB_UNMAPPED) != 0) {
 366                 bp->b_flags |= B_UNMAPPED;
 367                 bp->b_data = unmapped_buf;
 368         } else {
 369                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 370                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 371         }
 372         bp->b_iocmd = BIO_READ;
 373         bp->b_iodone = cluster_callback;
 374         bp->b_blkno = blkno;
 375         bp->b_lblkno = lbn;
 376         bp->b_offset = tbp->b_offset;
 377         KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
 378         pbgetvp(vp, bp);
 379
 380         TAILQ_INIT(&bp->b_cluster.cluster_head);
 381
 382         bp->b_bcount = 0;
 383         bp->b_bufsize = 0;
 384         bp->b_npages = 0;
 385
 386         inc = btodb(size);
 387         bo = &vp->v_bufobj;
 388         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
 389                 if (i != 0) {
 390                         if ((bp->b_npages * PAGE_SIZE) +
 391                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 392                                 break;
 393                         }
 394
 395                         tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
 396                             (gbflags & GB_UNMAPPED));
 397
 398                         /* Don't wait around for locked bufs. */
 399                         if (tbp == NULL)
 400                                 break;
 401
 402                         /*
 403                          * Stop scanning if the buffer is fully valid
 404                          * (marked B_CACHE), or locked (may be doing a
 405                          * background write), or if the buffer is not
 406                          * VMIO backed.  The clustering code can only deal
 407                          * with VMIO-backed buffers.
 408                          */
 409                         BO_LOCK(bo);
 410                         if ((tbp->b_vflags & BV_BKGRDINPROG) ||
 411                             (tbp->b_flags & B_CACHE) ||
 412                             (tbp->b_flags & B_VMIO) == 0) {
 413                                 BO_UNLOCK(bo);
 414                                 bqrelse(tbp);
 415                                 break;
 416                         }
 417                         BO_UNLOCK(bo);
 418
 419                         /*
 420                          * The buffer must be completely invalid in order to
 421                          * take part in the cluster.  If it is partially valid
 422                          * then we stop.
 423                          */
 424                         off = tbp->b_offset;
 425                         tsize = size;
 426                         VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
 427                         for (j = 0; tsize > 0; j++) {
 428                                 toff = off & PAGE_MASK;
 429                                 tinc = tsize;
 430                                 if (toff + tinc > PAGE_SIZE)
 431                                         tinc = PAGE_SIZE - toff;
 432                                 VM_OBJECT_LOCK_ASSERT(tbp->b_pages[j]->object,
 433                                     MA_OWNED);
 434                                 if ((tbp->b_pages[j]->valid &
 435                                     vm_page_bits(toff, tinc)) != 0)
 436                                         break;
 437                                 off += tinc;
 438                                 tsize -= tinc;
 439                         }
 440                         VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
 441                         if (tsize > 0) {
 442                                 bqrelse(tbp);
 443                                 break;
 444                         }
 445
 446                         /*
 447                          * Set a read-ahead mark as appropriate
 448                          */
 449                         if ((fbp && (i == 1)) || (i == (run - 1)))
 450                                 tbp->b_flags |= B_RAM;
 451
 452                         /*
 453                          * Set the buffer up for an async read (XXX should
 454                          * we do this only if we do not wind up brelse()ing?).
 455                          * Set the block number if it isn't set, otherwise
 456                          * if it is make sure it matches the block number we
 457                          * expect.
 458                          */
 459                         tbp->b_flags |= B_ASYNC;
 460                         tbp->b_iocmd = BIO_READ;
 461                         if (tbp->b_blkno == tbp->b_lblkno) {
 462                                 tbp->b_blkno = bn;
 463                         } else if (tbp->b_blkno != bn) {
 464                                 brelse(tbp);
 465                                 break;
 466                         }
 467                 }
 468                 /*
 469                  * XXX fbp from caller may not be B_ASYNC, but we are going
 470                  * to biodone() it in cluster_callback() anyway
 471                  */
 472                 BUF_KERNPROC(tbp);
 473                 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 474                         tbp, b_cluster.cluster_entry);
 475                 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
 476                 for (j = 0; j < tbp->b_npages; j += 1) {
 477                         vm_page_t m;
 478                         m = tbp->b_pages[j];
 479                         vm_page_io_start(m);
 480                         vm_object_pip_add(m->object, 1);
 481                         if ((bp->b_npages == 0) ||
 482                                 (bp->b_pages[bp->b_npages-1] != m)) {
 483                                 bp->b_pages[bp->b_npages] = m;
 484                                 bp->b_npages++;
 485                         }
 486                         if (m->valid == VM_PAGE_BITS_ALL)
 487                                 tbp->b_pages[j] = bogus_page;
 488                 }
 489                 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
 490                 /*
 491                  * Don't inherit tbp->b_bufsize as it may be larger due to
 492                  * a non-page-aligned size.  Instead just aggregate using
 493                  * 'size'.
 494                  */
 495                 if (tbp->b_bcount != size)
 496                         printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
 497                 if (tbp->b_bufsize != size)
 498                         printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
 499                 bp->b_bcount += size;
 500                 bp->b_bufsize += size;
 501         }
 502
 503         /*
 504          * Fully valid pages in the cluster are already good and do not need
 505          * to be re-read from disk.  Replace the page with bogus_page
 506          */
 507         VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
 508         for (j = 0; j < bp->b_npages; j++) {
 509                 VM_OBJECT_LOCK_ASSERT(bp->b_pages[j]->object, MA_OWNED);
 510                 if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
 511                         bp->b_pages[j] = bogus_page;
 512         }
 513         VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
 514         if (bp->b_bufsize > bp->b_kvasize)
 515                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 516                     bp->b_bufsize, bp->b_kvasize);
 517         bp->b_kvasize = bp->b_bufsize;
 518
 519         if ((bp->b_flags & B_UNMAPPED) == 0) {
 520                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 521                     (vm_page_t *)bp->b_pages, bp->b_npages);
 522         }
 523         return (bp);
 524 }
 525
 526 /*
 527  * Cleanup after a clustered read or write.
 528  * This is complicated by the fact that any of the buffers might have
 529  * extra memory (if there were no empty buffer headers at allocbuf time)
 530  * that we will need to shift around.
 531  */
 532 static void
 533 cluster_callback(bp)
 534         struct buf *bp;
 535 {
 536         struct buf *nbp, *tbp;
 537         int error = 0;
 538
 539         /*
 540          * Must propogate errors to all the components.
 541          */
 542         if (bp->b_ioflags & BIO_ERROR)
 543                 error = bp->b_error;
 544
 545         if ((bp->b_flags & B_UNMAPPED) == 0) {
 546                 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
 547                     bp->b_npages);
 548         }
 549         /*
 550          * Move memory from the large cluster buffer into the component
 551          * buffers and mark IO as done on these.
 552          */
 553         for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
 554                 tbp; tbp = nbp) {
 555                 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
 556                 if (error) {
 557                         tbp->b_ioflags |= BIO_ERROR;
 558                         tbp->b_error = error;
 559                 } else {
 560                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 561                         tbp->b_flags &= ~B_INVAL;
 562                         tbp->b_ioflags &= ~BIO_ERROR;
 563                         /*
 564                          * XXX the bdwrite()/bqrelse() issued during
 565                          * cluster building clears B_RELBUF (see bqrelse()
 566                          * comment).  If direct I/O was specified, we have
 567                          * to restore it here to allow the buffer and VM
 568                          * to be freed.
 569                          */
 570                         if (tbp->b_flags & B_DIRECT)
 571                                 tbp->b_flags |= B_RELBUF;
 572                 }
 573                 bufdone(tbp);
 574         }
 575         pbrelvp(bp);
 576         relpbuf(bp, &cluster_pbuf_freecnt);
 577 }
 578
 579 /*
 580  *      cluster_wbuild_wb:
 581  *
 582  *      Implement modified write build for cluster.
 583  *
 584  *              write_behind = 0        write behind disabled
 585  *              write_behind = 1        write behind normal (default)
 586  *              write_behind = 2        write behind backed-off
 587  */
 588
 589 static __inline int
 590 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
 591     int gbflags)
 592 {
 593         int r = 0;
 594
 595         switch (write_behind) {
 596         case 2:
 597                 if (start_lbn < len)
 598                         break;
 599                 start_lbn -= len;
 600                 /* FALLTHROUGH */
 601         case 1:
 602                 r = cluster_wbuild_gb(vp, size, start_lbn, len, gbflags);
 603                 /* FALLTHROUGH */
 604         default:
 605                 /* FALLTHROUGH */
 606                 break;
 607         }
 608         return(r);
 609 }
 610
 611 /*
 612  * Do clustered write for FFS.
 613  *
 614  * Three cases:
 615  *      1. Write is not sequential (write asynchronously)
 616  *      Write is sequential:
 617  *      2.      beginning of cluster - begin cluster
 618  *      3.      middle of a cluster - add to cluster
 619  *      4.      end of a cluster - asynchronously write cluster
 620  */
 621 void
 622 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount)
 623 {
 624
 625         cluster_write_gb(vp, bp, filesize, seqcount, 0);
 626 }
 627
 628 void
 629 cluster_write_gb(struct vnode *vp, struct buf *bp, u_quad_t filesize,
 630     int seqcount, int gbflags)
 631 {
 632         daddr_t lbn;
 633         int maxclen, cursize;
 634         int lblocksize;
 635         int async;
 636
 637         if (!unmapped_buf_allowed)
 638                 gbflags &= ~GB_UNMAPPED;
 639
 640         if (vp->v_type == VREG) {
 641                 async = DOINGASYNC(vp);
 642                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 643         } else {
 644                 async = 0;
 645                 lblocksize = bp->b_bufsize;
 646         }
 647         lbn = bp->b_lblkno;
 648         KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
 649
 650         /* Initialize vnode to beginning of file. */
 651         if (lbn == 0)
 652                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 653
 654         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
 655             (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
 656                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
 657                 if (vp->v_clen != 0) {
 658                         /*
 659                          * Next block is not sequential.
 660                          *
 661                          * If we are not writing at end of file, the process
 662                          * seeked to another point in the file since its last
 663                          * write, or we have reached our maximum cluster size,
 664                          * then push the previous cluster. Otherwise try
 665                          * reallocating to make it sequential.
 666                          *
 667                          * Change to algorithm: only push previous cluster if
 668                          * it was sequential from the point of view of the
 669                          * seqcount heuristic, otherwise leave the buffer
 670                          * intact so we can potentially optimize the I/O
 671                          * later on in the buf_daemon or update daemon
 672                          * flush.
 673                          */
 674                         cursize = vp->v_lastw - vp->v_cstart + 1;
 675                         if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
 676                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
 677                                 if (!async && seqcount > 0) {
 678                                         cluster_wbuild_wb(vp, lblocksize,
 679                                             vp->v_cstart, cursize, gbflags);
 680                                 }
 681                         } else {
 682                                 struct buf **bpp, **endbp;
 683                                 struct cluster_save *buflist;
 684
 685                                 buflist = cluster_collectbufs(vp, bp, gbflags);
 686                                 endbp = &buflist->bs_children
 687                                     [buflist->bs_nchildren - 1];
 688                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 689                                         /*
 690                                          * Failed, push the previous cluster
 691                                          * if *really* writing sequentially
 692                                          * in the logical file (seqcount > 1),
 693                                          * otherwise delay it in the hopes that
 694                                          * the low level disk driver can
 695                                          * optimize the write ordering.
 696                                          */
 697                                         for (bpp = buflist->bs_children;
 698                                              bpp < endbp; bpp++)
 699                                                 brelse(*bpp);
 700                                         free(buflist, M_SEGMENT);
 701                                         if (seqcount > 1) {
 702                                                 cluster_wbuild_wb(vp,
 703                                                     lblocksize, vp->v_cstart,
 704                                                     cursize, gbflags);
 705                                         }
 706                                 } else {
 707                                         /*
 708                                          * Succeeded, keep building cluster.
 709                                          */
 710                                         for (bpp = buflist->bs_children;
 711                                              bpp <= endbp; bpp++)
 712                                                 bdwrite(*bpp);
 713                                         free(buflist, M_SEGMENT);
 714                                         vp->v_lastw = lbn;
 715                                         vp->v_lasta = bp->b_blkno;
 716                                         return;
 717                                 }
 718                         }
 719                 }
 720                 /*
 721                  * Consider beginning a cluster. If at end of file, make
 722                  * cluster as large as possible, otherwise find size of
 723                  * existing cluster.
 724                  */
 725                 if ((vp->v_type == VREG) &&
 726                         ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
 727                     (bp->b_blkno == bp->b_lblkno) &&
 728                     (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
 729                      bp->b_blkno == -1)) {
 730                         bawrite(bp);
 731                         vp->v_clen = 0;
 732                         vp->v_lasta = bp->b_blkno;
 733                         vp->v_cstart = lbn + 1;
 734                         vp->v_lastw = lbn;
 735                         return;
 736                 }
 737                 vp->v_clen = maxclen;
 738                 if (!async && maxclen == 0) {   /* I/O not contiguous */
 739                         vp->v_cstart = lbn + 1;
 740                         bawrite(bp);
 741                 } else {        /* Wait for rest of cluster */
 742                         vp->v_cstart = lbn;
 743                         bdwrite(bp);
 744                 }
 745         } else if (lbn == vp->v_cstart + vp->v_clen) {
 746                 /*
 747                  * At end of cluster, write it out if seqcount tells us we
 748                  * are operating sequentially, otherwise let the buf or
 749                  * update daemon handle it.
 750                  */
 751                 bdwrite(bp);
 752                 if (seqcount > 1) {
 753                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
 754                             vp->v_clen + 1, gbflags);
 755                 }
 756                 vp->v_clen = 0;
 757                 vp->v_cstart = lbn + 1;
 758         } else if (vm_page_count_severe()) {
 759                 /*
 760                  * We are low on memory, get it going NOW
 761                  */
 762                 bawrite(bp);
 763         } else {
 764                 /*
 765                  * In the middle of a cluster, so just delay the I/O for now.
 766                  */
 767                 bdwrite(bp);
 768         }
 769         vp->v_lastw = lbn;
 770         vp->v_lasta = bp->b_blkno;
 771 }
 772
 773
 774 /*
 775  * This is an awful lot like cluster_rbuild...wish they could be combined.
 776  * The last lbn argument is the current block on which I/O is being
 777  * performed.  Check to see that it doesn't fall in the middle of
 778  * the current block (if last_bp == NULL).
 779  */
 780 int
 781 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len)
 782 {
 783
 784         return (cluster_wbuild_gb(vp, size, start_lbn, len, 0));
 785 }
 786
 787 int
 788 cluster_wbuild_gb(struct vnode *vp, long size, daddr_t start_lbn, int len,
 789     int gbflags)
 790 {
 791         struct buf *bp, *tbp;
 792         struct bufobj *bo;
 793         int i, j;
 794         int totalwritten = 0;
 795         int dbsize = btodb(size);
 796
 797         if (!unmapped_buf_allowed)
 798                 gbflags &= ~GB_UNMAPPED;
 799
 800         bo = &vp->v_bufobj;
 801         while (len > 0) {
 802                 /*
 803                  * If the buffer is not delayed-write (i.e. dirty), or it
 804                  * is delayed-write but either locked or inval, it cannot
 805                  * partake in the clustered write.
 806                  */
 807                 BO_LOCK(bo);
 808                 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
 809                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 810                         BO_UNLOCK(bo);
 811                         ++start_lbn;
 812                         --len;
 813                         continue;
 814                 }
 815                 if (BUF_LOCK(tbp,
 816                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_MTX(bo))) {
 817                         ++start_lbn;
 818                         --len;
 819                         continue;
 820                 }
 821                 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
 822                         BUF_UNLOCK(tbp);
 823                         ++start_lbn;
 824                         --len;
 825                         continue;
 826                 }
 827                 if (tbp->b_pin_count >  0) {
 828                         BUF_UNLOCK(tbp);
 829                         ++start_lbn;
 830                         --len;
 831                         continue;
 832                 }
 833                 bremfree(tbp);
 834                 tbp->b_flags &= ~B_DONE;
 835
 836                 /*
 837                  * Extra memory in the buffer, punt on this buffer.
 838                  * XXX we could handle this in most cases, but we would
 839                  * have to push the extra memory down to after our max
 840                  * possible cluster size and then potentially pull it back
 841                  * up if the cluster was terminated prematurely--too much
 842                  * hassle.
 843                  */
 844                 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
 845                      (B_CLUSTEROK | B_VMIO)) ||
 846                   (tbp->b_bcount != tbp->b_bufsize) ||
 847                   (tbp->b_bcount != size) ||
 848                   (len == 1) ||
 849                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 850                         totalwritten += tbp->b_bufsize;
 851                         bawrite(tbp);
 852                         ++start_lbn;
 853                         --len;
 854                         continue;
 855                 }
 856
 857                 /*
 858                  * We got a pbuf to make the cluster in.
 859                  * so initialise it.
 860                  */
 861                 TAILQ_INIT(&bp->b_cluster.cluster_head);
 862                 bp->b_bcount = 0;
 863                 bp->b_bufsize = 0;
 864                 bp->b_npages = 0;
 865                 if (tbp->b_wcred != NOCRED)
 866                         bp->b_wcred = crhold(tbp->b_wcred);
 867
 868                 bp->b_blkno = tbp->b_blkno;
 869                 bp->b_lblkno = tbp->b_lblkno;
 870                 bp->b_offset = tbp->b_offset;
 871
 872                 /*
 873                  * We are synthesizing a buffer out of vm_page_t's, but
 874                  * if the block size is not page aligned then the starting
 875                  * address may not be either.  Inherit the b_data offset
 876                  * from the original buffer.
 877                  */
 878                 if ((gbflags & GB_UNMAPPED) == 0 ||
 879                     (tbp->b_flags & B_VMIO) == 0) {
 880                         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 881                             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 882                 } else {
 883                         bp->b_flags |= B_UNMAPPED;
 884                         bp->b_data = unmapped_buf;
 885                 }
 886                 bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
 887                     B_NEEDCOMMIT));
 888                 bp->b_iodone = cluster_callback;
 889                 pbgetvp(vp, bp);
 890                 /*
 891                  * From this location in the file, scan forward to see
 892                  * if there are buffers with adjacent data that need to
 893                  * be written as well.
 894                  */
 895                 for (i = 0; i < len; ++i, ++start_lbn) {
 896                         if (i != 0) { /* If not the first buffer */
 897                                 /*
 898                                  * If the adjacent data is not even in core it
 899                                  * can't need to be written.
 900                                  */
 901                                 BO_LOCK(bo);
 902                                 if ((tbp = gbincore(bo, start_lbn)) == NULL ||
 903                                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 904                                         BO_UNLOCK(bo);
 905                                         break;
 906                                 }
 907
 908                                 /*
 909                                  * If it IS in core, but has different
 910                                  * characteristics, or is locked (which
 911                                  * means it could be undergoing a background
 912                                  * I/O or be in a weird state), then don't
 913                                  * cluster with it.
 914                                  */
 915                                 if (BUF_LOCK(tbp,
 916                                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
 917                                     BO_MTX(bo)))
 918                                         break;
 919
 920                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 921                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 922                                     != (B_DELWRI | B_CLUSTEROK |
 923                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 924                                     tbp->b_wcred != bp->b_wcred) {
 925                                         BUF_UNLOCK(tbp);
 926                                         break;
 927                                 }
 928
 929                                 /*
 930                                  * Check that the combined cluster
 931                                  * would make sense with regard to pages
 932                                  * and would not be too large
 933                                  */
 934                                 if ((tbp->b_bcount != size) ||
 935                                   ((bp->b_blkno + (dbsize * i)) !=
 936                                     tbp->b_blkno) ||
 937                                   ((tbp->b_npages + bp->b_npages) >
 938                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 939                                         BUF_UNLOCK(tbp);
 940                                         break;
 941                                 }
 942
 943                                 /*
 944                                  * Do not pull in pinned buffers.
 945                                  */
 946                                 if (tbp->b_pin_count > 0) {
 947                                         BUF_UNLOCK(tbp);
 948                                         break;
 949                                 }
 950
 951                                 /*
 952                                  * Ok, it's passed all the tests,
 953                                  * so remove it from the free list
 954                                  * and mark it busy. We will use it.
 955                                  */
 956                                 bremfree(tbp);
 957                                 tbp->b_flags &= ~B_DONE;
 958                         } /* end of code for non-first buffers only */
 959                         /*
 960                          * If the IO is via the VM then we do some
 961                          * special VM hackery (yuck).  Since the buffer's
 962                          * block size may not be page-aligned it is possible
 963                          * for a page to be shared between two buffers.  We
 964                          * have to get rid of the duplication when building
 965                          * the cluster.
 966                          */
 967                         if (tbp->b_flags & B_VMIO) {
 968                                 vm_page_t m;
 969
 970                                 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
 971                                 if (i != 0) { /* if not first buffer */
 972                                         for (j = 0; j < tbp->b_npages; j += 1) {
 973                                                 m = tbp->b_pages[j];
 974                                                 if (m->oflags & VPO_BUSY) {
 975                                                         VM_OBJECT_UNLOCK(
 976                                                             tbp->b_object);
 977                                                         bqrelse(tbp);
 978                                                         goto finishcluster;
 979                                                 }
 980                                         }
 981                                 }
 982                                 for (j = 0; j < tbp->b_npages; j += 1) {
 983                                         m = tbp->b_pages[j];
 984                                         vm_page_io_start(m);
 985                                         vm_object_pip_add(m->object, 1);
 986                                         if ((bp->b_npages == 0) ||
 987                                           (bp->b_pages[bp->b_npages - 1] != m)) {
 988                                                 bp->b_pages[bp->b_npages] = m;
 989                                                 bp->b_npages++;
 990                                         }
 991                                 }
 992                                 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
 993                         }
 994                         bp->b_bcount += size;
 995                         bp->b_bufsize += size;
 996                         /*
 997                          * If any of the clustered buffers have their
 998                          * B_BARRIER flag set, transfer that request to
 999                          * the cluster.
1000                          */
1001                         bp->b_flags |= (tbp->b_flags & B_BARRIER);
1002                         tbp->b_flags &= ~(B_DONE | B_BARRIER);
1003                         tbp->b_flags |= B_ASYNC;
1004                         tbp->b_ioflags &= ~BIO_ERROR;
1005                         tbp->b_iocmd = BIO_WRITE;
1006                         bundirty(tbp);
1007                         reassignbuf(tbp);               /* put on clean list */
1008                         bufobj_wref(tbp->b_bufobj);
1009                         BUF_KERNPROC(tbp);
1010                         TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
1011                                 tbp, b_cluster.cluster_entry);
1012                 }
1013         finishcluster:
1014                 if ((bp->b_flags & B_UNMAPPED) == 0) {
1015                         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1016                             (vm_page_t *)bp->b_pages, bp->b_npages);
1017                 }
1018                 if (bp->b_bufsize > bp->b_kvasize)
1019                         panic(
1020                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
1021                             bp->b_bufsize, bp->b_kvasize);
1022                 bp->b_kvasize = bp->b_bufsize;
1023                 totalwritten += bp->b_bufsize;
1024                 bp->b_dirtyoff = 0;
1025                 bp->b_dirtyend = bp->b_bufsize;
1026                 bawrite(bp);
1027
1028                 len -= i;
1029         }
1030         return totalwritten;
1031 }
1032
1033 /*
1034  * Collect together all the buffers in a cluster.
1035  * Plus add one additional buffer.
1036  */
1037 static struct cluster_save *
1038 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags)
1039 {
1040         struct cluster_save *buflist;
1041         struct buf *bp;
1042         daddr_t lbn;
1043         int i, len;
1044
1045         len = vp->v_lastw - vp->v_cstart + 1;
1046         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1047             M_SEGMENT, M_WAITOK);
1048         buflist->bs_nchildren = 0;
1049         buflist->bs_children = (struct buf **) (buflist + 1);
1050         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1051                 (void)bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1052                     gbflags, &bp);
1053                 buflist->bs_children[i] = bp;
1054                 if (bp->b_blkno == bp->b_lblkno)
1055                         VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1056                                 NULL, NULL);
1057         }
1058         buflist->bs_children[i] = bp = last_bp;
1059         if (bp->b_blkno == bp->b_lblkno)
1060                 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1061         buflist->bs_nchildren = i + 1;
1062         return (buflist);
1063 }