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