sys/kern/vfs_cluster.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (c) 1993
   5  *      The Regents of the University of California.  All rights reserved.
   6  * Modifications/enhancements:
   7  *      Copyright (c) 1995 John S. Dyson.  All rights reserved.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  * 3. Neither the name of the University nor the names of its contributors
  18  *    may be used to endorse or promote products derived from this software
  19  *    without specific prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  31  * SUCH DAMAGE.
  32  *
  33  *      @(#)vfs_cluster.c       8.7 (Berkeley) 2/13/94
  34  */
  35
  36 #include <sys/cdefs.h>
  37 __FBSDID("$FreeBSD$");
  38
  39 #include "opt_debug_cluster.h"
  40
  41 #include <sys/param.h>
  42 #include <sys/systm.h>
  43 #include <sys/kernel.h>
  44 #include <sys/proc.h>
  45 #include <sys/bio.h>
  46 #include <sys/buf.h>
  47 #include <sys/vnode.h>
  48 #include <sys/malloc.h>
  49 #include <sys/mount.h>
  50 #include <sys/racct.h>
  51 #include <sys/resourcevar.h>
  52 #include <sys/rwlock.h>
  53 #include <sys/vmmeter.h>
  54 #include <vm/vm.h>
  55 #include <vm/vm_object.h>
  56 #include <vm/vm_page.h>
  57 #include <sys/sysctl.h>
  58
  59 #if defined(CLUSTERDEBUG)
  60 static int      rcluster= 0;
  61 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0,
  62     "Debug VFS clustering code");
  63 #endif
  64
  65 static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer");
  66 static uma_zone_t cluster_pbuf_zone;
  67
  68 static void cluster_init(void *);
  69 static struct cluster_save *cluster_collectbufs(struct vnode *vp,
  70             struct buf *last_bp, int gbflags);
  71 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize,
  72             daddr_t lbn, daddr_t blkno, long size, int run, int gbflags,
  73             struct buf *fbp);
  74 static void cluster_callback(struct buf *);
  75
  76 static int write_behind = 1;
  77 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
  78     "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
  79
  80 static int read_max = 64;
  81 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
  82     "Cluster read-ahead max block count");
  83
  84 static int read_min = 1;
  85 SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0,
  86     "Cluster read min block count");
  87
  88 SYSINIT(cluster, SI_SUB_CPU, SI_ORDER_ANY, cluster_init, NULL);
  89
  90 static void
  91 cluster_init(void *dummy)
  92 {
  93
  94         cluster_pbuf_zone = pbuf_zsecond_create("clpbuf", nswbuf / 2);
  95 }
  96
  97 /*
  98  * Read data to a buf, including read-ahead if we find this to be beneficial.
  99  * cluster_read replaces bread.
 100  */
 101 int
 102 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
 103     struct ucred *cred, long totread, int seqcount, int gbflags,
 104     struct buf **bpp)
 105 {
 106         struct buf *bp, *rbp, *reqbp;
 107         struct bufobj *bo;
 108         struct thread *td;
 109         daddr_t blkno, origblkno;
 110         int maxra, racluster;
 111         int error, ncontig;
 112         int i;
 113
 114         error = 0;
 115         td = curthread;
 116         bo = &vp->v_bufobj;
 117         if (!unmapped_buf_allowed)
 118                 gbflags &= ~GB_UNMAPPED;
 119
 120         /*
 121          * Try to limit the amount of read-ahead by a few
 122          * ad-hoc parameters.  This needs work!!!
 123          */
 124         racluster = vp->v_mount->mnt_iosize_max / size;
 125         maxra = seqcount;
 126         maxra = min(read_max, maxra);
 127         maxra = min(nbuf/8, maxra);
 128         if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
 129                 maxra = (filesize / size) - lblkno;
 130
 131         /*
 132          * get the requested block
 133          */
 134         error = getblkx(vp, lblkno, lblkno, size, 0, 0, gbflags, &bp);
 135         if (error != 0) {
 136                 *bpp = NULL;
 137                 return (error);
 138         }
 139         gbflags &= ~GB_NOSPARSE;
 140         origblkno = lblkno;
 141         *bpp = reqbp = bp;
 142
 143         /*
 144          * if it is in the cache, then check to see if the reads have been
 145          * sequential.  If they have, then try some read-ahead, otherwise
 146          * back-off on prospective read-aheads.
 147          */
 148         if (bp->b_flags & B_CACHE) {
 149                 if (!seqcount) {
 150                         return 0;
 151                 } else if ((bp->b_flags & B_RAM) == 0) {
 152                         return 0;
 153                 } else {
 154                         bp->b_flags &= ~B_RAM;
 155                         BO_RLOCK(bo);
 156                         for (i = 1; i < maxra; i++) {
 157                                 /*
 158                                  * Stop if the buffer does not exist or it
 159                                  * is invalid (about to go away?)
 160                                  */
 161                                 rbp = gbincore(&vp->v_bufobj, lblkno+i);
 162                                 if (rbp == NULL || (rbp->b_flags & B_INVAL))
 163                                         break;
 164
 165                                 /*
 166                                  * Set another read-ahead mark so we know
 167                                  * to check again. (If we can lock the
 168                                  * buffer without waiting)
 169                                  */
 170                                 if ((((i % racluster) == (racluster - 1)) ||
 171                                     (i == (maxra - 1)))
 172                                     && (0 == BUF_LOCK(rbp,
 173                                         LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
 174                                         rbp->b_flags |= B_RAM;
 175                                         BUF_UNLOCK(rbp);
 176                                 }
 177                         }
 178                         BO_RUNLOCK(bo);
 179                         if (i >= maxra) {
 180                                 return 0;
 181                         }
 182                         lblkno += i;
 183                 }
 184                 reqbp = bp = NULL;
 185         /*
 186          * If it isn't in the cache, then get a chunk from
 187          * disk if sequential, otherwise just get the block.
 188          */
 189         } else {
 190                 off_t firstread = bp->b_offset;
 191                 int nblks;
 192                 long minread;
 193
 194                 KASSERT(bp->b_offset != NOOFFSET,
 195                     ("cluster_read: no buffer offset"));
 196
 197                 ncontig = 0;
 198
 199                 /*
 200                  * Adjust totread if needed
 201                  */
 202                 minread = read_min * size;
 203                 if (minread > totread)
 204                         totread = minread;
 205
 206                 /*
 207                  * Compute the total number of blocks that we should read
 208                  * synchronously.
 209                  */
 210                 if (firstread + totread > filesize)
 211                         totread = filesize - firstread;
 212                 nblks = howmany(totread, size);
 213                 if (nblks > racluster)
 214                         nblks = racluster;
 215
 216                 /*
 217                  * Now compute the number of contiguous blocks.
 218                  */
 219                 if (nblks > 1) {
 220                         error = VOP_BMAP(vp, lblkno, NULL,
 221                                 &blkno, &ncontig, NULL);
 222                         /*
 223                          * If this failed to map just do the original block.
 224                          */
 225                         if (error || blkno == -1)
 226                                 ncontig = 0;
 227                 }
 228
 229                 /*
 230                  * If we have contiguous data available do a cluster
 231                  * otherwise just read the requested block.
 232                  */
 233                 if (ncontig) {
 234                         /* Account for our first block. */
 235                         ncontig = min(ncontig + 1, nblks);
 236                         if (ncontig < nblks)
 237                                 nblks = ncontig;
 238                         bp = cluster_rbuild(vp, filesize, lblkno,
 239                             blkno, size, nblks, gbflags, bp);
 240                         lblkno += (bp->b_bufsize / size);
 241                 } else {
 242                         bp->b_flags |= B_RAM;
 243                         bp->b_iocmd = BIO_READ;
 244                         lblkno += 1;
 245                 }
 246         }
 247
 248         /*
 249          * handle the synchronous read so that it is available ASAP.
 250          */
 251         if (bp) {
 252                 if ((bp->b_flags & B_CLUSTER) == 0) {
 253                         vfs_busy_pages(bp, 0);
 254                 }
 255                 bp->b_flags &= ~B_INVAL;
 256                 bp->b_ioflags &= ~BIO_ERROR;
 257                 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
 258                         BUF_KERNPROC(bp);
 259                 bp->b_iooffset = dbtob(bp->b_blkno);
 260                 bstrategy(bp);
 261 #ifdef RACCT
 262                 if (racct_enable) {
 263                         PROC_LOCK(td->td_proc);
 264                         racct_add_buf(td->td_proc, bp, 0);
 265                         PROC_UNLOCK(td->td_proc);
 266                 }
 267 #endif /* RACCT */
 268                 td->td_ru.ru_inblock++;
 269         }
 270
 271         /*
 272          * If we have been doing sequential I/O, then do some read-ahead.
 273          */
 274         while (lblkno < (origblkno + maxra)) {
 275                 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
 276                 if (error)
 277                         break;
 278
 279                 if (blkno == -1)
 280                         break;
 281
 282                 /*
 283                  * We could throttle ncontig here by maxra but we might as
 284                  * well read the data if it is contiguous.  We're throttled
 285                  * by racluster anyway.
 286                  */
 287                 if (ncontig) {
 288                         ncontig = min(ncontig + 1, racluster);
 289                         rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
 290                             size, ncontig, gbflags, NULL);
 291                         lblkno += (rbp->b_bufsize / size);
 292                         if (rbp->b_flags & B_DELWRI) {
 293                                 bqrelse(rbp);
 294                                 continue;
 295                         }
 296                 } else {
 297                         rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
 298                         lblkno += 1;
 299                         if (rbp->b_flags & B_DELWRI) {
 300                                 bqrelse(rbp);
 301                                 continue;
 302                         }
 303                         rbp->b_flags |= B_ASYNC | B_RAM;
 304                         rbp->b_iocmd = BIO_READ;
 305                         rbp->b_blkno = blkno;
 306                 }
 307                 if (rbp->b_flags & B_CACHE) {
 308                         rbp->b_flags &= ~B_ASYNC;
 309                         bqrelse(rbp);
 310                         continue;
 311                 }
 312                 if ((rbp->b_flags & B_CLUSTER) == 0) {
 313                         vfs_busy_pages(rbp, 0);
 314                 }
 315                 rbp->b_flags &= ~B_INVAL;
 316                 rbp->b_ioflags &= ~BIO_ERROR;
 317                 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
 318                         BUF_KERNPROC(rbp);
 319                 rbp->b_iooffset = dbtob(rbp->b_blkno);
 320                 bstrategy(rbp);
 321 #ifdef RACCT
 322                 if (racct_enable) {
 323                         PROC_LOCK(td->td_proc);
 324                         racct_add_buf(td->td_proc, rbp, 0);
 325                         PROC_UNLOCK(td->td_proc);
 326                 }
 327 #endif /* RACCT */
 328                 td->td_ru.ru_inblock++;
 329         }
 330
 331         if (reqbp) {
 332                 /*
 333                  * Like bread, always brelse() the buffer when
 334                  * returning an error.
 335                  */
 336                 error = bufwait(reqbp);
 337                 if (error != 0) {
 338                         brelse(reqbp);
 339                         *bpp = NULL;
 340                 }
 341         }
 342         return (error);
 343 }
 344
 345 /*
 346  * If blocks are contiguous on disk, use this to provide clustered
 347  * read ahead.  We will read as many blocks as possible sequentially
 348  * and then parcel them up into logical blocks in the buffer hash table.
 349  */
 350 static struct buf *
 351 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
 352     daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
 353 {
 354         struct buf *bp, *tbp;
 355         daddr_t bn;
 356         off_t off;
 357         long tinc, tsize;
 358         int i, inc, j, k, toff;
 359
 360         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
 361             ("cluster_rbuild: size %ld != f_iosize %jd\n",
 362             size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
 363
 364         /*
 365          * avoid a division
 366          */
 367         while ((u_quad_t) size * (lbn + run) > filesize) {
 368                 --run;
 369         }
 370
 371         if (fbp) {
 372                 tbp = fbp;
 373                 tbp->b_iocmd = BIO_READ;
 374         } else {
 375                 tbp = getblk(vp, lbn, size, 0, 0, gbflags);
 376                 if (tbp->b_flags & B_CACHE)
 377                         return tbp;
 378                 tbp->b_flags |= B_ASYNC | B_RAM;
 379                 tbp->b_iocmd = BIO_READ;
 380         }
 381         tbp->b_blkno = blkno;
 382         if( (tbp->b_flags & B_MALLOC) ||
 383                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
 384                 return tbp;
 385
 386         bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT);
 387         if (bp == NULL)
 388                 return tbp;
 389         MPASS((bp->b_flags & B_MAXPHYS) != 0);
 390
 391         /*
 392          * We are synthesizing a buffer out of vm_page_t's, but
 393          * if the block size is not page aligned then the starting
 394          * address may not be either.  Inherit the b_data offset
 395          * from the original buffer.
 396          */
 397         bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
 398         if ((gbflags & GB_UNMAPPED) != 0) {
 399                 bp->b_data = unmapped_buf;
 400         } else {
 401                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 402                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 403         }
 404         bp->b_iocmd = BIO_READ;
 405         bp->b_iodone = cluster_callback;
 406         bp->b_blkno = blkno;
 407         bp->b_lblkno = lbn;
 408         bp->b_offset = tbp->b_offset;
 409         KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
 410         pbgetvp(vp, bp);
 411
 412         TAILQ_INIT(&bp->b_cluster.cluster_head);
 413
 414         bp->b_bcount = 0;
 415         bp->b_bufsize = 0;
 416         bp->b_npages = 0;
 417
 418         inc = btodb(size);
 419         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
 420                 if (i == 0) {
 421                         vm_object_pip_add(tbp->b_bufobj->bo_object,
 422                             tbp->b_npages);
 423                         vfs_busy_pages_acquire(tbp);
 424                 } else {
 425                         if ((bp->b_npages * PAGE_SIZE) +
 426                             round_page(size) > vp->v_mount->mnt_iosize_max) {
 427                                 break;
 428                         }
 429
 430                         tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
 431                             (gbflags & GB_UNMAPPED));
 432
 433                         /* Don't wait around for locked bufs. */
 434                         if (tbp == NULL)
 435                                 break;
 436
 437                         /*
 438                          * Stop scanning if the buffer is fully valid
 439                          * (marked B_CACHE), or locked (may be doing a
 440                          * background write), or if the buffer is not
 441                          * VMIO backed.  The clustering code can only deal
 442                          * with VMIO-backed buffers.  The bo lock is not
 443                          * required for the BKGRDINPROG check since it
 444                          * can not be set without the buf lock.
 445                          */
 446                         if ((tbp->b_vflags & BV_BKGRDINPROG) ||
 447                             (tbp->b_flags & B_CACHE) ||
 448                             (tbp->b_flags & B_VMIO) == 0) {
 449                                 bqrelse(tbp);
 450                                 break;
 451                         }
 452
 453                         /*
 454                          * The buffer must be completely invalid in order to
 455                          * take part in the cluster.  If it is partially valid
 456                          * then we stop.
 457                          */
 458                         off = tbp->b_offset;
 459                         tsize = size;
 460                         for (j = 0; tsize > 0; j++) {
 461                                 toff = off & PAGE_MASK;
 462                                 tinc = tsize;
 463                                 if (toff + tinc > PAGE_SIZE)
 464                                         tinc = PAGE_SIZE - toff;
 465                                 if (vm_page_trysbusy(tbp->b_pages[j]) == 0)
 466                                         break;
 467                                 if ((tbp->b_pages[j]->valid &
 468                                     vm_page_bits(toff, tinc)) != 0) {
 469                                         vm_page_sunbusy(tbp->b_pages[j]);
 470                                         break;
 471                                 }
 472                                 vm_object_pip_add(tbp->b_bufobj->bo_object, 1);
 473                                 off += tinc;
 474                                 tsize -= tinc;
 475                         }
 476                         if (tsize > 0) {
 477 clean_sbusy:
 478                                 vm_object_pip_wakeupn(tbp->b_bufobj->bo_object,
 479                                     j);
 480                                 for (k = 0; k < j; k++)
 481                                         vm_page_sunbusy(tbp->b_pages[k]);
 482                                 bqrelse(tbp);
 483                                 break;
 484                         }
 485
 486                         /*
 487                          * Set a read-ahead mark as appropriate
 488                          */
 489                         if ((fbp && (i == 1)) || (i == (run - 1)))
 490                                 tbp->b_flags |= B_RAM;
 491
 492                         /*
 493                          * Set the buffer up for an async read (XXX should
 494                          * we do this only if we do not wind up brelse()ing?).
 495                          * Set the block number if it isn't set, otherwise
 496                          * if it is make sure it matches the block number we
 497                          * expect.
 498                          */
 499                         tbp->b_flags |= B_ASYNC;
 500                         tbp->b_iocmd = BIO_READ;
 501                         if (tbp->b_blkno == tbp->b_lblkno) {
 502                                 tbp->b_blkno = bn;
 503                         } else if (tbp->b_blkno != bn) {
 504                                 goto clean_sbusy;
 505                         }
 506                 }
 507                 /*
 508                  * XXX fbp from caller may not be B_ASYNC, but we are going
 509                  * to biodone() it in cluster_callback() anyway
 510                  */
 511                 BUF_KERNPROC(tbp);
 512                 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
 513                         tbp, b_cluster.cluster_entry);
 514                 for (j = 0; j < tbp->b_npages; j += 1) {
 515                         vm_page_t m;
 516
 517                         m = tbp->b_pages[j];
 518                         if ((bp->b_npages == 0) ||
 519                             (bp->b_pages[bp->b_npages-1] != m)) {
 520                                 bp->b_pages[bp->b_npages] = m;
 521                                 bp->b_npages++;
 522                         }
 523                         if (vm_page_all_valid(m))
 524                                 tbp->b_pages[j] = bogus_page;
 525                 }
 526
 527                 /*
 528                  * Don't inherit tbp->b_bufsize as it may be larger due to
 529                  * a non-page-aligned size.  Instead just aggregate using
 530                  * 'size'.
 531                  */
 532                 if (tbp->b_bcount != size)
 533                         printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
 534                 if (tbp->b_bufsize != size)
 535                         printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
 536                 bp->b_bcount += size;
 537                 bp->b_bufsize += size;
 538         }
 539
 540         /*
 541          * Fully valid pages in the cluster are already good and do not need
 542          * to be re-read from disk.  Replace the page with bogus_page
 543          */
 544         for (j = 0; j < bp->b_npages; j++) {
 545                 if (vm_page_all_valid(bp->b_pages[j]))
 546                         bp->b_pages[j] = bogus_page;
 547         }
 548         if (bp->b_bufsize > bp->b_kvasize)
 549                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
 550                     bp->b_bufsize, bp->b_kvasize);
 551
 552         if (buf_mapped(bp)) {
 553                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 554                     (vm_page_t *)bp->b_pages, bp->b_npages);
 555         }
 556         return (bp);
 557 }
 558
 559 /*
 560  * Cleanup after a clustered read or write.
 561  * This is complicated by the fact that any of the buffers might have
 562  * extra memory (if there were no empty buffer headers at allocbuf time)
 563  * that we will need to shift around.
 564  */
 565 static void
 566 cluster_callback(struct buf *bp)
 567 {
 568         struct buf *nbp, *tbp;
 569         int error = 0;
 570
 571         /*
 572          * Must propagate errors to all the components.
 573          */
 574         if (bp->b_ioflags & BIO_ERROR)
 575                 error = bp->b_error;
 576
 577         if (buf_mapped(bp)) {
 578                 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
 579                     bp->b_npages);
 580         }
 581         /*
 582          * Move memory from the large cluster buffer into the component
 583          * buffers and mark IO as done on these.
 584          */
 585         for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
 586                 tbp; tbp = nbp) {
 587                 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
 588                 if (error) {
 589                         tbp->b_ioflags |= BIO_ERROR;
 590                         tbp->b_error = error;
 591                 } else {
 592                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 593                         tbp->b_flags &= ~B_INVAL;
 594                         tbp->b_ioflags &= ~BIO_ERROR;
 595                         /*
 596                          * XXX the bdwrite()/bqrelse() issued during
 597                          * cluster building clears B_RELBUF (see bqrelse()
 598                          * comment).  If direct I/O was specified, we have
 599                          * to restore it here to allow the buffer and VM
 600                          * to be freed.
 601                          */
 602                         if (tbp->b_flags & B_DIRECT)
 603                                 tbp->b_flags |= B_RELBUF;
 604                 }
 605                 bufdone(tbp);
 606         }
 607         pbrelvp(bp);
 608         uma_zfree(cluster_pbuf_zone, bp);
 609 }
 610
 611 /*
 612  *      cluster_wbuild_wb:
 613  *
 614  *      Implement modified write build for cluster.
 615  *
 616  *              write_behind = 0        write behind disabled
 617  *              write_behind = 1        write behind normal (default)
 618  *              write_behind = 2        write behind backed-off
 619  */
 620
 621 static __inline int
 622 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
 623     int gbflags)
 624 {
 625         int r = 0;
 626
 627         switch (write_behind) {
 628         case 2:
 629                 if (start_lbn < len)
 630                         break;
 631                 start_lbn -= len;
 632                 /* FALLTHROUGH */
 633         case 1:
 634                 r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
 635                 /* FALLTHROUGH */
 636         default:
 637                 /* FALLTHROUGH */
 638                 break;
 639         }
 640         return(r);
 641 }
 642
 643 /*
 644  * Do clustered write for FFS.
 645  *
 646  * Three cases:
 647  *      1. Write is not sequential (write asynchronously)
 648  *      Write is sequential:
 649  *      2.      beginning of cluster - begin cluster
 650  *      3.      middle of a cluster - add to cluster
 651  *      4.      end of a cluster - asynchronously write cluster
 652  */
 653 void
 654 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount,
 655     int gbflags)
 656 {
 657         daddr_t lbn;
 658         int maxclen, cursize;
 659         int lblocksize;
 660         int async;
 661
 662         if (!unmapped_buf_allowed)
 663                 gbflags &= ~GB_UNMAPPED;
 664
 665         if (vp->v_type == VREG) {
 666                 async = DOINGASYNC(vp);
 667                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
 668         } else {
 669                 async = 0;
 670                 lblocksize = bp->b_bufsize;
 671         }
 672         lbn = bp->b_lblkno;
 673         KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
 674
 675         /* Initialize vnode to beginning of file. */
 676         if (lbn == 0)
 677                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 678
 679         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
 680             (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
 681                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
 682                 if (vp->v_clen != 0) {
 683                         /*
 684                          * Next block is not sequential.
 685                          *
 686                          * If we are not writing at end of file, the process
 687                          * seeked to another point in the file since its last
 688                          * write, or we have reached our maximum cluster size,
 689                          * then push the previous cluster. Otherwise try
 690                          * reallocating to make it sequential.
 691                          *
 692                          * Change to algorithm: only push previous cluster if
 693                          * it was sequential from the point of view of the
 694                          * seqcount heuristic, otherwise leave the buffer
 695                          * intact so we can potentially optimize the I/O
 696                          * later on in the buf_daemon or update daemon
 697                          * flush.
 698                          */
 699                         cursize = vp->v_lastw - vp->v_cstart + 1;
 700                         if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
 701                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
 702                                 if (!async && seqcount > 0) {
 703                                         cluster_wbuild_wb(vp, lblocksize,
 704                                             vp->v_cstart, cursize, gbflags);
 705                                 }
 706                         } else {
 707                                 struct buf **bpp, **endbp;
 708                                 struct cluster_save *buflist;
 709
 710                                 buflist = cluster_collectbufs(vp, bp, gbflags);
 711                                 if (buflist == NULL) {
 712                                         /*
 713                                          * Cluster build failed so just write
 714                                          * it now.
 715                                          */
 716                                         bawrite(bp);
 717                                         return;
 718                                 }
 719                                 endbp = &buflist->bs_children
 720                                     [buflist->bs_nchildren - 1];
 721                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 722                                         /*
 723                                          * Failed, push the previous cluster
 724                                          * if *really* writing sequentially
 725                                          * in the logical file (seqcount > 1),
 726                                          * otherwise delay it in the hopes that
 727                                          * the low level disk driver can
 728                                          * optimize the write ordering.
 729                                          */
 730                                         for (bpp = buflist->bs_children;
 731                                              bpp < endbp; bpp++)
 732                                                 brelse(*bpp);
 733                                         free(buflist, M_SEGMENT);
 734                                         if (seqcount > 1) {
 735                                                 cluster_wbuild_wb(vp,
 736                                                     lblocksize, vp->v_cstart,
 737                                                     cursize, gbflags);
 738                                         }
 739                                 } else {
 740                                         /*
 741                                          * Succeeded, keep building cluster.
 742                                          */
 743                                         for (bpp = buflist->bs_children;
 744                                              bpp <= endbp; bpp++)
 745                                                 bdwrite(*bpp);
 746                                         free(buflist, M_SEGMENT);
 747                                         vp->v_lastw = lbn;
 748                                         vp->v_lasta = bp->b_blkno;
 749                                         return;
 750                                 }
 751                         }
 752                 }
 753                 /*
 754                  * Consider beginning a cluster. If at end of file, make
 755                  * cluster as large as possible, otherwise find size of
 756                  * existing cluster.
 757                  */
 758                 if ((vp->v_type == VREG) &&
 759                         ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
 760                     (bp->b_blkno == bp->b_lblkno) &&
 761                     (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
 762                      bp->b_blkno == -1)) {
 763                         bawrite(bp);
 764                         vp->v_clen = 0;
 765                         vp->v_lasta = bp->b_blkno;
 766                         vp->v_cstart = lbn + 1;
 767                         vp->v_lastw = lbn;
 768                         return;
 769                 }
 770                 vp->v_clen = maxclen;
 771                 if (!async && maxclen == 0) {   /* I/O not contiguous */
 772                         vp->v_cstart = lbn + 1;
 773                         bawrite(bp);
 774                 } else {        /* Wait for rest of cluster */
 775                         vp->v_cstart = lbn;
 776                         bdwrite(bp);
 777                 }
 778         } else if (lbn == vp->v_cstart + vp->v_clen) {
 779                 /*
 780                  * At end of cluster, write it out if seqcount tells us we
 781                  * are operating sequentially, otherwise let the buf or
 782                  * update daemon handle it.
 783                  */
 784                 bdwrite(bp);
 785                 if (seqcount > 1) {
 786                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
 787                             vp->v_clen + 1, gbflags);
 788                 }
 789                 vp->v_clen = 0;
 790                 vp->v_cstart = lbn + 1;
 791         } else if (vm_page_count_severe()) {
 792                 /*
 793                  * We are low on memory, get it going NOW
 794                  */
 795                 bawrite(bp);
 796         } else {
 797                 /*
 798                  * In the middle of a cluster, so just delay the I/O for now.
 799                  */
 800                 bdwrite(bp);
 801         }
 802         vp->v_lastw = lbn;
 803         vp->v_lasta = bp->b_blkno;
 804 }
 805
 806 /*
 807  * This is an awful lot like cluster_rbuild...wish they could be combined.
 808  * The last lbn argument is the current block on which I/O is being
 809  * performed.  Check to see that it doesn't fall in the middle of
 810  * the current block (if last_bp == NULL).
 811  */
 812 int
 813 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
 814     int gbflags)
 815 {
 816         struct buf *bp, *tbp;
 817         struct bufobj *bo;
 818         int i, j;
 819         int totalwritten = 0;
 820         int dbsize = btodb(size);
 821
 822         if (!unmapped_buf_allowed)
 823                 gbflags &= ~GB_UNMAPPED;
 824
 825         bo = &vp->v_bufobj;
 826         while (len > 0) {
 827                 /*
 828                  * If the buffer is not delayed-write (i.e. dirty), or it
 829                  * is delayed-write but either locked or inval, it cannot
 830                  * partake in the clustered write.
 831                  */
 832                 BO_LOCK(bo);
 833                 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
 834                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 835                         BO_UNLOCK(bo);
 836                         ++start_lbn;
 837                         --len;
 838                         continue;
 839                 }
 840                 if (BUF_LOCK(tbp,
 841                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) {
 842                         ++start_lbn;
 843                         --len;
 844                         continue;
 845                 }
 846                 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
 847                         BUF_UNLOCK(tbp);
 848                         ++start_lbn;
 849                         --len;
 850                         continue;
 851                 }
 852                 bremfree(tbp);
 853                 tbp->b_flags &= ~B_DONE;
 854
 855                 /*
 856                  * Extra memory in the buffer, punt on this buffer.
 857                  * XXX we could handle this in most cases, but we would
 858                  * have to push the extra memory down to after our max
 859                  * possible cluster size and then potentially pull it back
 860                  * up if the cluster was terminated prematurely--too much
 861                  * hassle.
 862                  */
 863                 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
 864                      (B_CLUSTEROK | B_VMIO)) ||
 865                   (tbp->b_bcount != tbp->b_bufsize) ||
 866                   (tbp->b_bcount != size) ||
 867                   (len == 1) ||
 868                   ((bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT)) == NULL)) {
 869                         totalwritten += tbp->b_bufsize;
 870                         bawrite(tbp);
 871                         ++start_lbn;
 872                         --len;
 873                         continue;
 874                 }
 875                 MPASS((bp->b_flags & B_MAXPHYS) != 0);
 876
 877                 /*
 878                  * We got a pbuf to make the cluster in.
 879                  * so initialise it.
 880                  */
 881                 TAILQ_INIT(&bp->b_cluster.cluster_head);
 882                 bp->b_bcount = 0;
 883                 bp->b_bufsize = 0;
 884                 bp->b_npages = 0;
 885                 if (tbp->b_wcred != NOCRED)
 886                         bp->b_wcred = crhold(tbp->b_wcred);
 887
 888                 bp->b_blkno = tbp->b_blkno;
 889                 bp->b_lblkno = tbp->b_lblkno;
 890                 bp->b_offset = tbp->b_offset;
 891
 892                 /*
 893                  * We are synthesizing a buffer out of vm_page_t's, but
 894                  * if the block size is not page aligned then the starting
 895                  * address may not be either.  Inherit the b_data offset
 896                  * from the original buffer.
 897                  */
 898                 if ((gbflags & GB_UNMAPPED) == 0 ||
 899                     (tbp->b_flags & B_VMIO) == 0) {
 900                         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 901                             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 902                 } else {
 903                         bp->b_data = unmapped_buf;
 904                 }
 905                 bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
 906                     B_NEEDCOMMIT));
 907                 bp->b_iodone = cluster_callback;
 908                 pbgetvp(vp, bp);
 909                 /*
 910                  * From this location in the file, scan forward to see
 911                  * if there are buffers with adjacent data that need to
 912                  * be written as well.
 913                  */
 914                 for (i = 0; i < len; ++i, ++start_lbn) {
 915                         if (i != 0) { /* If not the first buffer */
 916                                 /*
 917                                  * If the adjacent data is not even in core it
 918                                  * can't need to be written.
 919                                  */
 920                                 BO_LOCK(bo);
 921                                 if ((tbp = gbincore(bo, start_lbn)) == NULL ||
 922                                     (tbp->b_vflags & BV_BKGRDINPROG)) {
 923                                         BO_UNLOCK(bo);
 924                                         break;
 925                                 }
 926
 927                                 /*
 928                                  * If it IS in core, but has different
 929                                  * characteristics, or is locked (which
 930                                  * means it could be undergoing a background
 931                                  * I/O or be in a weird state), then don't
 932                                  * cluster with it.
 933                                  */
 934                                 if (BUF_LOCK(tbp,
 935                                     LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
 936                                     BO_LOCKPTR(bo)))
 937                                         break;
 938
 939                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 940                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 941                                     != (B_DELWRI | B_CLUSTEROK |
 942                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 943                                     tbp->b_wcred != bp->b_wcred) {
 944                                         BUF_UNLOCK(tbp);
 945                                         break;
 946                                 }
 947
 948                                 /*
 949                                  * Check that the combined cluster
 950                                  * would make sense with regard to pages
 951                                  * and would not be too large
 952                                  */
 953                                 if ((tbp->b_bcount != size) ||
 954                                   ((bp->b_blkno + (dbsize * i)) !=
 955                                     tbp->b_blkno) ||
 956                                   ((tbp->b_npages + bp->b_npages) >
 957                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
 958                                         BUF_UNLOCK(tbp);
 959                                         break;
 960                                 }
 961
 962                                 /*
 963                                  * Ok, it's passed all the tests,
 964                                  * so remove it from the free list
 965                                  * and mark it busy. We will use it.
 966                                  */
 967                                 bremfree(tbp);
 968                                 tbp->b_flags &= ~B_DONE;
 969                         } /* end of code for non-first buffers only */
 970                         /*
 971                          * If the IO is via the VM then we do some
 972                          * special VM hackery (yuck).  Since the buffer's
 973                          * block size may not be page-aligned it is possible
 974                          * for a page to be shared between two buffers.  We
 975                          * have to get rid of the duplication when building
 976                          * the cluster.
 977                          */
 978                         if (tbp->b_flags & B_VMIO) {
 979                                 vm_page_t m;
 980
 981                                 if (i == 0) {
 982                                         vfs_busy_pages_acquire(tbp);
 983                                 } else { /* if not first buffer */
 984                                         for (j = 0; j < tbp->b_npages; j += 1) {
 985                                                 m = tbp->b_pages[j];
 986                                                 if (vm_page_trysbusy(m) == 0) {
 987                                                         for (j--; j >= 0; j--)
 988                                                                 vm_page_sunbusy(
 989                                                                     tbp->b_pages[j]);
 990                                                         bqrelse(tbp);
 991                                                         goto finishcluster;
 992                                                 }
 993                                         }
 994                                 }
 995                                 vm_object_pip_add(tbp->b_bufobj->bo_object,
 996                                     tbp->b_npages);
 997                                 for (j = 0; j < tbp->b_npages; j += 1) {
 998                                         m = tbp->b_pages[j];
 999                                         if ((bp->b_npages == 0) ||
1000                                           (bp->b_pages[bp->b_npages - 1] != m)) {
1001                                                 bp->b_pages[bp->b_npages] = m;
1002                                                 bp->b_npages++;
1003                                         }
1004                                 }
1005                         }
1006                         bp->b_bcount += size;
1007                         bp->b_bufsize += size;
1008                         /*
1009                          * If any of the clustered buffers have their
1010                          * B_BARRIER flag set, transfer that request to
1011                          * the cluster.
1012                          */
1013                         bp->b_flags |= (tbp->b_flags & B_BARRIER);
1014                         tbp->b_flags &= ~(B_DONE | B_BARRIER);
1015                         tbp->b_flags |= B_ASYNC;
1016                         tbp->b_ioflags &= ~BIO_ERROR;
1017                         tbp->b_iocmd = BIO_WRITE;
1018                         bundirty(tbp);
1019                         reassignbuf(tbp);               /* put on clean list */
1020                         bufobj_wref(tbp->b_bufobj);
1021                         BUF_KERNPROC(tbp);
1022                         buf_track(tbp, __func__);
1023                         TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
1024                                 tbp, b_cluster.cluster_entry);
1025                 }
1026         finishcluster:
1027                 if (buf_mapped(bp)) {
1028                         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1029                             (vm_page_t *)bp->b_pages, bp->b_npages);
1030                 }
1031                 if (bp->b_bufsize > bp->b_kvasize)
1032                         panic(
1033                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
1034                             bp->b_bufsize, bp->b_kvasize);
1035                 totalwritten += bp->b_bufsize;
1036                 bp->b_dirtyoff = 0;
1037                 bp->b_dirtyend = bp->b_bufsize;
1038                 bawrite(bp);
1039
1040                 len -= i;
1041         }
1042         return totalwritten;
1043 }
1044
1045 /*
1046  * Collect together all the buffers in a cluster.
1047  * Plus add one additional buffer.
1048  */
1049 static struct cluster_save *
1050 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags)
1051 {
1052         struct cluster_save *buflist;
1053         struct buf *bp;
1054         daddr_t lbn;
1055         int i, j, len, error;
1056
1057         len = vp->v_lastw - vp->v_cstart + 1;
1058         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1059             M_SEGMENT, M_WAITOK);
1060         buflist->bs_nchildren = 0;
1061         buflist->bs_children = (struct buf **) (buflist + 1);
1062         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1063                 error = bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1064                     gbflags, &bp);
1065                 if (error != 0) {
1066                         /*
1067                          * If read fails, release collected buffers
1068                          * and return failure.
1069                          */
1070                         for (j = 0; j < i; j++)
1071                                 brelse(buflist->bs_children[j]);
1072                         free(buflist, M_SEGMENT);
1073                         return (NULL);
1074                 }
1075                 buflist->bs_children[i] = bp;
1076                 if (bp->b_blkno == bp->b_lblkno)
1077                         VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1078                                 NULL, NULL);
1079         }
1080         buflist->bs_children[i] = bp = last_bp;
1081         if (bp->b_blkno == bp->b_lblkno)
1082                 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1083         buflist->bs_nchildren = i + 1;
1084         return (buflist);
1085 }