sys/kern/vfs_cluster.c

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