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  * 3. All advertising materials mentioning features or use of this software
  16  *    must display the following acknowledgement:
  17  *      This product includes software developed by the University of
  18  *      California, Berkeley and its contributors.
  19  * 4. Neither the name of the University nor the names of its contributors
  20  *    may be used to endorse or promote products derived from this software
  21  *    without specific prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  *
  35  *      @(#)vfs_cluster.c       8.7 (Berkeley) 2/13/94
  36  * $FreeBSD: src/sys/kern/vfs_cluster.c,v 1.92.2.9 2001/11/18 07:10:59 dillon Exp $
  37  * $DragonFly: src/sys/kern/vfs_cluster.c,v 1.40 2008/07/14 03:09:00 dillon Exp $
  38  */
  39
  40 #include "opt_debug_cluster.h"
  41
  42 #include <sys/param.h>
  43 #include <sys/systm.h>
  44 #include <sys/kernel.h>
  45 #include <sys/proc.h>
  46 #include <sys/buf.h>
  47 #include <sys/vnode.h>
  48 #include <sys/malloc.h>
  49 #include <sys/mount.h>
  50 #include <sys/resourcevar.h>
  51 #include <sys/vmmeter.h>
  52 #include <vm/vm.h>
  53 #include <vm/vm_object.h>
  54 #include <vm/vm_page.h>
  55 #include <sys/sysctl.h>
  56 #include <sys/buf2.h>
  57 #include <vm/vm_page2.h>
  58
  59 #include <machine/limits.h>
  60
  61 #if defined(CLUSTERDEBUG)
  62 #include <sys/sysctl.h>
  63 static int      rcluster= 0;
  64 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
  65 #endif
  66
  67 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");
  68
  69 static struct cluster_save *
  70         cluster_collectbufs (struct vnode *vp, struct buf *last_bp,
  71                             int blksize);
  72 static struct buf *
  73         cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
  74                             off_t doffset, int blksize, int run,
  75                             struct buf *fbp);
  76 static void cluster_callback (struct bio *);
  77
  78
  79 static int write_behind = 1;
  80 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
  81
  82 extern vm_page_t        bogus_page;
  83
  84 extern int cluster_pbuf_freecnt;
  85
  86 /*
  87  * Maximum number of blocks for read-ahead.
  88  */
  89 #define MAXRA 32
  90
  91 /*
  92  * This replaces bread.
  93  */
  94 int
  95 cluster_read(struct vnode *vp, off_t filesize, off_t loffset,
  96              int blksize, size_t resid, int seqcount, struct buf **bpp)
  97 {
  98         struct buf *bp, *rbp, *reqbp;
  99         off_t origoffset;
 100         off_t doffset;
 101         int error;
 102         int i;
 103         int maxra, racluster;
 104         int totread;
 105
 106         error = 0;
 107         totread = (resid > INT_MAX) ? INT_MAX : (int)resid;
 108
 109         /*
 110          * Try to limit the amount of read-ahead by a few
 111          * ad-hoc parameters.  This needs work!!!
 112          */
 113         racluster = vmaxiosize(vp) / blksize;
 114         maxra = 2 * racluster + (totread / blksize);
 115         if (maxra > MAXRA)
 116                 maxra = MAXRA;
 117         if (maxra > nbuf/8)
 118                 maxra = nbuf/8;
 119
 120         /*
 121          * Get the requested block.
 122          */
 123         *bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0);
 124         origoffset = loffset;
 125
 126         /*
 127          * if it is in the cache, then check to see if the reads have been
 128          * sequential.  If they have, then try some read-ahead, otherwise
 129          * back-off on prospective read-aheads.
 130          */
 131         if (bp->b_flags & B_CACHE) {
 132                 if (!seqcount) {
 133                         return 0;
 134                 } else if ((bp->b_flags & B_RAM) == 0) {
 135                         return 0;
 136                 } else {
 137                         struct buf *tbp;
 138                         bp->b_flags &= ~B_RAM;
 139
 140                         /*
 141                          * Set read-ahead-mark only if we can passively lock
 142                          * the buffer.  Note that with these flags the bp
 143                          * could very exist even though NULL is returned.
 144                          */
 145                         for (i = 1; i < maxra; i++) {
 146                                 tbp = findblk(vp, loffset + i * blksize,
 147                                               FINDBLK_NBLOCK);
 148                                 if (tbp == NULL)
 149                                         break;
 150                                 if (((i % racluster) == (racluster - 1)) ||
 151                                     (i == (maxra - 1))) {
 152                                         tbp->b_flags |= B_RAM;
 153                                 }
 154                                 BUF_UNLOCK(tbp);
 155                         }
 156                         if (i >= maxra)
 157                                 return 0;
 158                         loffset += i * blksize;
 159                 }
 160                 reqbp = bp = NULL;
 161         } else {
 162                 off_t firstread = bp->b_loffset;
 163                 int nblks;
 164
 165                 /*
 166                  * Set-up synchronous read for bp.
 167                  */
 168                 bp->b_cmd = BUF_CMD_READ;
 169                 bp->b_bio1.bio_done = biodone_sync;
 170                 bp->b_bio1.bio_flags |= BIO_SYNC;
 171
 172                 KASSERT(firstread != NOOFFSET,
 173                         ("cluster_read: no buffer offset"));
 174                 if (firstread + totread > filesize)
 175                         totread = (int)(filesize - firstread);
 176                 nblks = totread / blksize;
 177                 if (nblks) {
 178                         int burstbytes;
 179
 180                         if (nblks > racluster)
 181                                 nblks = racluster;
 182
 183                         error = VOP_BMAP(vp, loffset, &doffset,
 184                                          &burstbytes, NULL, BUF_CMD_READ);
 185                         if (error)
 186                                 goto single_block_read;
 187                         if (doffset == NOOFFSET)
 188                                 goto single_block_read;
 189                         if (burstbytes < blksize * 2)
 190                                 goto single_block_read;
 191                         if (nblks > burstbytes / blksize)
 192                                 nblks = burstbytes / blksize;
 193
 194                         bp = cluster_rbuild(vp, filesize, loffset,
 195                                             doffset, blksize, nblks, bp);
 196                         loffset += bp->b_bufsize;
 197                 } else {
 198 single_block_read:
 199                         /*
 200                          * if it isn't in the cache, then get a chunk from
 201                          * disk if sequential, otherwise just get the block.
 202                          */
 203                         bp->b_flags |= B_RAM;
 204                         loffset += blksize;
 205                 }
 206         }
 207
 208         /*
 209          * If B_CACHE was not set issue bp.  bp will either be an
 210          * asynchronous cluster buf or a synchronous single-buf.
 211          * If it is a single buf it will be the same as reqbp.
 212          *
 213          * NOTE: Once an async cluster buf is issued bp becomes invalid.
 214          */
 215         if (bp) {
 216 #if defined(CLUSTERDEBUG)
 217                 if (rcluster)
 218                         kprintf("S(%lld,%d,%d) ",
 219                             bp->b_loffset, bp->b_bcount, seqcount);
 220 #endif
 221                 if ((bp->b_flags & B_CLUSTER) == 0)
 222                         vfs_busy_pages(vp, bp);
 223                 bp->b_flags &= ~(B_ERROR|B_INVAL);
 224                 vn_strategy(vp, &bp->b_bio1);
 225                 error = 0;
 226                 /* bp invalid now */
 227         }
 228
 229         /*
 230          * If we have been doing sequential I/O, then do some read-ahead.
 231          *
 232          * Only mess with buffers which we can immediately lock.  HAMMER
 233          * will do device-readahead irrespective of what the blocks
 234          * represent.
 235          */
 236         rbp = NULL;
 237         if (!error &&
 238             seqcount &&
 239             loffset < origoffset + seqcount * blksize &&
 240             loffset + blksize <= filesize
 241         ) {
 242                 int nblksread;
 243                 int ntoread;
 244                 int burstbytes;
 245                 int tmp_error;
 246
 247                 rbp = getblk(vp, loffset, blksize,
 248                              GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
 249                 if (rbp == NULL)
 250                         goto no_read_ahead;
 251                 if ((rbp->b_flags & B_CACHE)) {
 252                         bqrelse(rbp);
 253                         goto no_read_ahead;
 254                 }
 255
 256                 /*
 257                  * An error from the read-ahead bmap has nothing to do
 258                  * with the caller's original request.
 259                  */
 260                 tmp_error = VOP_BMAP(vp, loffset, &doffset,
 261                                      &burstbytes, NULL, BUF_CMD_READ);
 262                 if (tmp_error || doffset == NOOFFSET) {
 263                         rbp->b_flags |= B_INVAL;
 264                         brelse(rbp);
 265                         rbp = NULL;
 266                         goto no_read_ahead;
 267                 }
 268                 ntoread = burstbytes / blksize;
 269                 nblksread = (totread + blksize - 1) / blksize;
 270                 if (seqcount < nblksread)
 271                         seqcount = nblksread;
 272                 if (ntoread > seqcount)
 273                         ntoread = seqcount;
 274
 275                 /*
 276                  * rbp: async read
 277                  */
 278                 rbp->b_cmd = BUF_CMD_READ;
 279                 rbp->b_flags |= B_RAM/* | B_AGE*/;
 280
 281                 if (burstbytes) {
 282                         rbp = cluster_rbuild(vp, filesize, loffset,
 283                                              doffset, blksize,
 284                                              ntoread, rbp);
 285                 } else {
 286                         rbp->b_bio2.bio_offset = doffset;
 287                 }
 288 #if defined(CLUSTERDEBUG)
 289                 if (rcluster) {
 290                         if (bp)
 291                                 kprintf("A+(%lld,%d,%lld,%d) ",
 292                                     rbp->b_loffset, rbp->b_bcount,
 293                                     rbp->b_loffset - origoffset,
 294                                     seqcount);
 295                         else
 296                                 kprintf("A(%lld,%d,%lld,%d) ",
 297                                     rbp->b_loffset, rbp->b_bcount,
 298                                     rbp->b_loffset - origoffset,
 299                                     seqcount);
 300                 }
 301 #endif
 302                 rbp->b_flags &= ~(B_ERROR|B_INVAL);
 303
 304                 if ((rbp->b_flags & B_CLUSTER) == 0)
 305                         vfs_busy_pages(vp, rbp);
 306                 BUF_KERNPROC(rbp);
 307                 vn_strategy(vp, &rbp->b_bio1);
 308                 /* rbp invalid now */
 309         }
 310
 311         /*
 312          * Wait for our original buffer to complete its I/O.  reqbp will
 313          * be NULL if the original buffer was B_CACHE.  We are returning
 314          * (*bpp) which is the same as reqbp when reqbp != NULL.
 315          */
 316 no_read_ahead:
 317         if (reqbp) {
 318                 KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
 319                 error = biowait(&reqbp->b_bio1, "clurd");
 320         }
 321         return (error);
 322 }
 323
 324 /*
 325  * If blocks are contiguous on disk, use this to provide clustered
 326  * read ahead.  We will read as many blocks as possible sequentially
 327  * and then parcel them up into logical blocks in the buffer hash table.
 328  *
 329  * This function either returns a cluster buf or it returns fbp.  fbp is
 330  * already expected to be set up as a synchronous or asynchronous request.
 331  *
 332  * If a cluster buf is returned it will always be async.
 333  */
 334 static struct buf *
 335 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
 336                int blksize, int run, struct buf *fbp)
 337 {
 338         struct buf *bp, *tbp;
 339         off_t boffset;
 340         int i, j;
 341         int maxiosize = vmaxiosize(vp);
 342
 343         /*
 344          * avoid a division
 345          */
 346         while (loffset + run * blksize > filesize) {
 347                 --run;
 348         }
 349
 350         tbp = fbp;
 351         tbp->b_bio2.bio_offset = doffset;
 352         if((tbp->b_flags & B_MALLOC) ||
 353             ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
 354                 return tbp;
 355         }
 356
 357         bp = trypbuf(&cluster_pbuf_freecnt);
 358         if (bp == NULL) {
 359                 return tbp;
 360         }
 361
 362         /*
 363          * We are synthesizing a buffer out of vm_page_t's, but
 364          * if the block size is not page aligned then the starting
 365          * address may not be either.  Inherit the b_data offset
 366          * from the original buffer.
 367          */
 368         bp->b_data = (char *)((vm_offset_t)bp->b_data |
 369             ((vm_offset_t)tbp->b_data & PAGE_MASK));
 370         bp->b_flags |= B_CLUSTER | B_VMIO;
 371         bp->b_cmd = BUF_CMD_READ;
 372         bp->b_bio1.bio_done = cluster_callback;         /* default to async */
 373         bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 374         bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 375         bp->b_loffset = loffset;
 376         bp->b_bio2.bio_offset = doffset;
 377         KASSERT(bp->b_loffset != NOOFFSET,
 378                 ("cluster_rbuild: no buffer offset"));
 379
 380         bp->b_bcount = 0;
 381         bp->b_bufsize = 0;
 382         bp->b_xio.xio_npages = 0;
 383
 384         for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
 385                 if (i) {
 386                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
 387                             round_page(blksize) > maxiosize) {
 388                                 break;
 389                         }
 390
 391                         /*
 392                          * Shortcut some checks and try to avoid buffers that
 393                          * would block in the lock.  The same checks have to
 394                          * be made again after we officially get the buffer.
 395                          */
 396                         tbp = getblk(vp, loffset + i * blksize, blksize,
 397                                      GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
 398                         if (tbp == NULL)
 399                                 break;
 400                         for (j = 0; j < tbp->b_xio.xio_npages; j++) {
 401                                 if (tbp->b_xio.xio_pages[j]->valid)
 402                                         break;
 403                         }
 404                         if (j != tbp->b_xio.xio_npages) {
 405                                 bqrelse(tbp);
 406                                 break;
 407                         }
 408
 409                         /*
 410                          * Stop scanning if the buffer is fuly valid
 411                          * (marked B_CACHE), or locked (may be doing a
 412                          * background write), or if the buffer is not
 413                          * VMIO backed.  The clustering code can only deal
 414                          * with VMIO-backed buffers.
 415                          */
 416                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
 417                             (tbp->b_flags & B_VMIO) == 0 ||
 418                             (LIST_FIRST(&tbp->b_dep) != NULL &&
 419                              buf_checkread(tbp))
 420                         ) {
 421                                 bqrelse(tbp);
 422                                 break;
 423                         }
 424
 425                         /*
 426                          * The buffer must be completely invalid in order to
 427                          * take part in the cluster.  If it is partially valid
 428                          * then we stop.
 429                          */
 430                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
 431                                 if (tbp->b_xio.xio_pages[j]->valid)
 432                                         break;
 433                         }
 434                         if (j != tbp->b_xio.xio_npages) {
 435                                 bqrelse(tbp);
 436                                 break;
 437                         }
 438
 439                         /*
 440                          * Set a read-ahead mark as appropriate
 441                          */
 442                         if (i == 1 || i == (run - 1))
 443                                 tbp->b_flags |= B_RAM;
 444
 445                         /*
 446                          * Depress the priority of buffers not explicitly
 447                          * requested.
 448                          */
 449                         /* tbp->b_flags |= B_AGE; */
 450
 451                         /*
 452                          * Set the block number if it isn't set, otherwise
 453                          * if it is make sure it matches the block number we
 454                          * expect.
 455                          */
 456                         if (tbp->b_bio2.bio_offset == NOOFFSET) {
 457                                 tbp->b_bio2.bio_offset = boffset;
 458                         } else if (tbp->b_bio2.bio_offset != boffset) {
 459                                 brelse(tbp);
 460                                 break;
 461                         }
 462                 }
 463
 464                 /*
 465                  * The passed-in tbp (i == 0) will already be set up for
 466                  * async or sync operation.  All other tbp's acquire in
 467                  * our loop are set up for async operation.
 468                  */
 469                 tbp->b_cmd = BUF_CMD_READ;
 470                 BUF_KERNPROC(tbp);
 471                 cluster_append(&bp->b_bio1, tbp);
 472                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 473                         vm_page_t m;
 474                         m = tbp->b_xio.xio_pages[j];
 475                         vm_page_io_start(m);
 476                         vm_object_pip_add(m->object, 1);
 477                         if ((bp->b_xio.xio_npages == 0) ||
 478                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
 479                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 480                                 bp->b_xio.xio_npages++;
 481                         }
 482                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
 483                                 tbp->b_xio.xio_pages[j] = bogus_page;
 484                 }
 485                 /*
 486                  * XXX shouldn't this be += size for both, like in
 487                  * cluster_wbuild()?
 488                  *
 489                  * Don't inherit tbp->b_bufsize as it may be larger due to
 490                  * a non-page-aligned size.  Instead just aggregate using
 491                  * 'size'.
 492                  */
 493                 if (tbp->b_bcount != blksize)
 494                     kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
 495                 if (tbp->b_bufsize != blksize)
 496                     kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
 497                 bp->b_bcount += blksize;
 498                 bp->b_bufsize += blksize;
 499         }
 500
 501         /*
 502          * Fully valid pages in the cluster are already good and do not need
 503          * to be re-read from disk.  Replace the page with bogus_page
 504          */
 505         for (j = 0; j < bp->b_xio.xio_npages; j++) {
 506                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
 507                     VM_PAGE_BITS_ALL) {
 508                         bp->b_xio.xio_pages[j] = bogus_page;
 509                 }
 510         }
 511         if (bp->b_bufsize > bp->b_kvasize) {
 512                 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
 513                     bp->b_bufsize, bp->b_kvasize);
 514         }
 515         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 516                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 517         BUF_KERNPROC(bp);
 518         return (bp);
 519 }
 520
 521 /*
 522  * Cleanup after a clustered read or write.
 523  * This is complicated by the fact that any of the buffers might have
 524  * extra memory (if there were no empty buffer headers at allocbuf time)
 525  * that we will need to shift around.
 526  *
 527  * The returned bio is &bp->b_bio1
 528  */
 529 void
 530 cluster_callback(struct bio *bio)
 531 {
 532         struct buf *bp = bio->bio_buf;
 533         struct buf *tbp;
 534         int error = 0;
 535
 536         /*
 537          * Must propogate errors to all the components.  A short read (EOF)
 538          * is a critical error.
 539          */
 540         if (bp->b_flags & B_ERROR) {
 541                 error = bp->b_error;
 542         } else if (bp->b_bcount != bp->b_bufsize) {
 543                 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
 544         }
 545
 546         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
 547         /*
 548          * Move memory from the large cluster buffer into the component
 549          * buffers and mark IO as done on these.  Since the memory map
 550          * is the same, no actual copying is required.
 551          */
 552         while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
 553                 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
 554                 if (error) {
 555                         tbp->b_flags |= B_ERROR;
 556                         tbp->b_error = error;
 557                 } else {
 558                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
 559                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
 560                         /*
 561                          * XXX the bdwrite()/bqrelse() issued during
 562                          * cluster building clears B_RELBUF (see bqrelse()
 563                          * comment).  If direct I/O was specified, we have
 564                          * to restore it here to allow the buffer and VM
 565                          * to be freed.
 566                          */
 567                         if (tbp->b_flags & B_DIRECT)
 568                                 tbp->b_flags |= B_RELBUF;
 569                 }
 570                 biodone(&tbp->b_bio1);
 571         }
 572         relpbuf(bp, &cluster_pbuf_freecnt);
 573 }
 574
 575 /*
 576  *      cluster_wbuild_wb:
 577  *
 578  *      Implement modified write build for cluster.
 579  *
 580  *              write_behind = 0        write behind disabled
 581  *              write_behind = 1        write behind normal (default)
 582  *              write_behind = 2        write behind backed-off
 583  */
 584
 585 static __inline int
 586 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
 587 {
 588         int r = 0;
 589
 590         switch(write_behind) {
 591         case 2:
 592                 if (start_loffset < len)
 593                         break;
 594                 start_loffset -= len;
 595                 /* fall through */
 596         case 1:
 597                 r = cluster_wbuild(vp, blksize, start_loffset, len);
 598                 /* fall through */
 599         default:
 600                 /* fall through */
 601                 break;
 602         }
 603         return(r);
 604 }
 605
 606 /*
 607  * Do clustered write for FFS.
 608  *
 609  * Three cases:
 610  *      1. Write is not sequential (write asynchronously)
 611  *      Write is sequential:
 612  *      2.      beginning of cluster - begin cluster
 613  *      3.      middle of a cluster - add to cluster
 614  *      4.      end of a cluster - asynchronously write cluster
 615  */
 616 void
 617 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
 618 {
 619         struct vnode *vp;
 620         off_t loffset;
 621         int maxclen, cursize;
 622         int async;
 623
 624         vp = bp->b_vp;
 625         if (vp->v_type == VREG)
 626                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
 627         else
 628                 async = 0;
 629         loffset = bp->b_loffset;
 630         KASSERT(bp->b_loffset != NOOFFSET,
 631                 ("cluster_write: no buffer offset"));
 632
 633         /* Initialize vnode to beginning of file. */
 634         if (loffset == 0)
 635                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 636
 637         if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
 638             bp->b_bio2.bio_offset == NOOFFSET ||
 639             (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
 640                 maxclen = vmaxiosize(vp);
 641                 if (vp->v_clen != 0) {
 642                         /*
 643                          * Next block is not sequential.
 644                          *
 645                          * If we are not writing at end of file, the process
 646                          * seeked to another point in the file since its last
 647                          * write, or we have reached our maximum cluster size,
 648                          * then push the previous cluster. Otherwise try
 649                          * reallocating to make it sequential.
 650                          *
 651                          * Change to algorithm: only push previous cluster if
 652                          * it was sequential from the point of view of the
 653                          * seqcount heuristic, otherwise leave the buffer
 654                          * intact so we can potentially optimize the I/O
 655                          * later on in the buf_daemon or update daemon
 656                          * flush.
 657                          */
 658                         cursize = vp->v_lastw - vp->v_cstart + blksize;
 659                         if (bp->b_loffset + blksize != filesize ||
 660                             loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
 661                                 if (!async && seqcount > 0) {
 662                                         cluster_wbuild_wb(vp, blksize,
 663                                                 vp->v_cstart, cursize);
 664                                 }
 665                         } else {
 666                                 struct buf **bpp, **endbp;
 667                                 struct cluster_save *buflist;
 668
 669                                 buflist = cluster_collectbufs(vp, bp, blksize);
 670                                 endbp = &buflist->bs_children
 671                                     [buflist->bs_nchildren - 1];
 672                                 if (VOP_REALLOCBLKS(vp, buflist)) {
 673                                         /*
 674                                          * Failed, push the previous cluster
 675                                          * if *really* writing sequentially
 676                                          * in the logical file (seqcount > 1),
 677                                          * otherwise delay it in the hopes that
 678                                          * the low level disk driver can
 679                                          * optimize the write ordering.
 680                                          */
 681                                         for (bpp = buflist->bs_children;
 682                                              bpp < endbp; bpp++)
 683                                                 brelse(*bpp);
 684                                         kfree(buflist, M_SEGMENT);
 685                                         if (seqcount > 1) {
 686                                                 cluster_wbuild_wb(vp,
 687                                                     blksize, vp->v_cstart,
 688                                                     cursize);
 689                                         }
 690                                 } else {
 691                                         /*
 692                                          * Succeeded, keep building cluster.
 693                                          */
 694                                         for (bpp = buflist->bs_children;
 695                                              bpp <= endbp; bpp++)
 696                                                 bdwrite(*bpp);
 697                                         kfree(buflist, M_SEGMENT);
 698                                         vp->v_lastw = loffset;
 699                                         vp->v_lasta = bp->b_bio2.bio_offset;
 700                                         return;
 701                                 }
 702                         }
 703                 }
 704                 /*
 705                  * Consider beginning a cluster. If at end of file, make
 706                  * cluster as large as possible, otherwise find size of
 707                  * existing cluster.
 708                  */
 709                 if ((vp->v_type == VREG) &&
 710                     bp->b_loffset + blksize != filesize &&
 711                     (bp->b_bio2.bio_offset == NOOFFSET) &&
 712                     (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
 713                      bp->b_bio2.bio_offset == NOOFFSET)) {
 714                         bawrite(bp);
 715                         vp->v_clen = 0;
 716                         vp->v_lasta = bp->b_bio2.bio_offset;
 717                         vp->v_cstart = loffset + blksize;
 718                         vp->v_lastw = loffset;
 719                         return;
 720                 }
 721                 if (maxclen > blksize)
 722                         vp->v_clen = maxclen - blksize;
 723                 else
 724                         vp->v_clen = 0;
 725                 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
 726                         vp->v_cstart = loffset + blksize;
 727                         bawrite(bp);
 728                 } else {        /* Wait for rest of cluster */
 729                         vp->v_cstart = loffset;
 730                         bdwrite(bp);
 731                 }
 732         } else if (loffset == vp->v_cstart + vp->v_clen) {
 733                 /*
 734                  * At end of cluster, write it out if seqcount tells us we
 735                  * are operating sequentially, otherwise let the buf or
 736                  * update daemon handle it.
 737                  */
 738                 bdwrite(bp);
 739                 if (seqcount > 1)
 740                         cluster_wbuild_wb(vp, blksize, vp->v_cstart,
 741                                           vp->v_clen + blksize);
 742                 vp->v_clen = 0;
 743                 vp->v_cstart = loffset + blksize;
 744         } else if (vm_page_count_severe()) {
 745                 /*
 746                  * We are low on memory, get it going NOW
 747                  */
 748                 bawrite(bp);
 749         } else {
 750                 /*
 751                  * In the middle of a cluster, so just delay the I/O for now.
 752                  */
 753                 bdwrite(bp);
 754         }
 755         vp->v_lastw = loffset;
 756         vp->v_lasta = bp->b_bio2.bio_offset;
 757 }
 758
 759
 760 /*
 761  * This is an awful lot like cluster_rbuild...wish they could be combined.
 762  * The last lbn argument is the current block on which I/O is being
 763  * performed.  Check to see that it doesn't fall in the middle of
 764  * the current block (if last_bp == NULL).
 765  */
 766 int
 767 cluster_wbuild(struct vnode *vp, int blksize, off_t start_loffset, int bytes)
 768 {
 769         struct buf *bp, *tbp;
 770         int i, j;
 771         int totalwritten = 0;
 772         int maxiosize = vmaxiosize(vp);
 773
 774         while (bytes > 0) {
 775                 /*
 776                  * If the buffer is not delayed-write (i.e. dirty), or it
 777                  * is delayed-write but either locked or inval, it cannot
 778                  * partake in the clustered write.
 779                  */
 780                 tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK);
 781                 if (tbp == NULL ||
 782                     (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI ||
 783                     (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
 784                         if (tbp)
 785                                 BUF_UNLOCK(tbp);
 786                         start_loffset += blksize;
 787                         bytes -= blksize;
 788                         continue;
 789                 }
 790                 bremfree(tbp);
 791                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 792
 793                 /*
 794                  * Extra memory in the buffer, punt on this buffer.
 795                  * XXX we could handle this in most cases, but we would
 796                  * have to push the extra memory down to after our max
 797                  * possible cluster size and then potentially pull it back
 798                  * up if the cluster was terminated prematurely--too much
 799                  * hassle.
 800                  */
 801                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
 802                     (tbp->b_bcount != tbp->b_bufsize) ||
 803                     (tbp->b_bcount != blksize) ||
 804                     (bytes == blksize) ||
 805                     ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
 806                         totalwritten += tbp->b_bufsize;
 807                         bawrite(tbp);
 808                         start_loffset += blksize;
 809                         bytes -= blksize;
 810                         continue;
 811                 }
 812
 813                 /*
 814                  * Set up the pbuf.  Track our append point with b_bcount
 815                  * and b_bufsize.  b_bufsize is not used by the device but
 816                  * our caller uses it to loop clusters and we use it to
 817                  * detect a premature EOF on the block device.
 818                  */
 819                 bp->b_bcount = 0;
 820                 bp->b_bufsize = 0;
 821                 bp->b_xio.xio_npages = 0;
 822                 bp->b_loffset = tbp->b_loffset;
 823                 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
 824
 825                 /*
 826                  * We are synthesizing a buffer out of vm_page_t's, but
 827                  * if the block size is not page aligned then the starting
 828                  * address may not be either.  Inherit the b_data offset
 829                  * from the original buffer.
 830                  */
 831                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
 832                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
 833                 bp->b_flags &= ~B_ERROR;
 834                 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
 835                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
 836                 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
 837                 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
 838
 839                 /*
 840                  * From this location in the file, scan forward to see
 841                  * if there are buffers with adjacent data that need to
 842                  * be written as well.
 843                  */
 844                 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
 845                         if (i != 0) { /* If not the first buffer */
 846                                 tbp = findblk(vp, start_loffset,
 847                                               FINDBLK_NBLOCK);
 848                                 /*
 849                                  * Buffer not found or could not be locked
 850                                  * non-blocking.
 851                                  */
 852                                 if (tbp == NULL)
 853                                         break;
 854
 855                                 /*
 856                                  * If it IS in core, but has different
 857                                  * characteristics, then don't cluster
 858                                  * with it.
 859                                  */
 860                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
 861                                      B_INVAL | B_DELWRI | B_NEEDCOMMIT))
 862                                     != (B_DELWRI | B_CLUSTEROK |
 863                                      (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
 864                                     (tbp->b_flags & B_LOCKED) ||
 865                                     (LIST_FIRST(&tbp->b_dep) &&
 866                                      buf_checkwrite(tbp))
 867                                 ) {
 868                                         BUF_UNLOCK(tbp);
 869                                         break;
 870                                 }
 871
 872                                 /*
 873                                  * Check that the combined cluster
 874                                  * would make sense with regard to pages
 875                                  * and would not be too large
 876                                  */
 877                                 if ((tbp->b_bcount != blksize) ||
 878                                   ((bp->b_bio2.bio_offset + i) !=
 879                                     tbp->b_bio2.bio_offset) ||
 880                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
 881                                     (maxiosize / PAGE_SIZE))) {
 882                                         BUF_UNLOCK(tbp);
 883                                         break;
 884                                 }
 885                                 /*
 886                                  * Ok, it's passed all the tests,
 887                                  * so remove it from the free list
 888                                  * and mark it busy. We will use it.
 889                                  */
 890                                 bremfree(tbp);
 891                                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
 892                         } /* end of code for non-first buffers only */
 893
 894                         /*
 895                          * If the IO is via the VM then we do some
 896                          * special VM hackery (yuck).  Since the buffer's
 897                          * block size may not be page-aligned it is possible
 898                          * for a page to be shared between two buffers.  We
 899                          * have to get rid of the duplication when building
 900                          * the cluster.
 901                          */
 902                         if (tbp->b_flags & B_VMIO) {
 903                                 vm_page_t m;
 904
 905                                 if (i != 0) { /* if not first buffer */
 906                                         for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 907                                                 m = tbp->b_xio.xio_pages[j];
 908                                                 if (m->flags & PG_BUSY) {
 909                                                         bqrelse(tbp);
 910                                                         goto finishcluster;
 911                                                 }
 912                                         }
 913                                 }
 914
 915                                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
 916                                         m = tbp->b_xio.xio_pages[j];
 917                                         vm_page_io_start(m);
 918                                         vm_object_pip_add(m->object, 1);
 919                                         if ((bp->b_xio.xio_npages == 0) ||
 920                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
 921                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
 922                                                 bp->b_xio.xio_npages++;
 923                                         }
 924                                 }
 925                         }
 926                         bp->b_bcount += blksize;
 927                         bp->b_bufsize += blksize;
 928
 929                         bundirty(tbp);
 930                         tbp->b_flags &= ~B_ERROR;
 931                         tbp->b_cmd = BUF_CMD_WRITE;
 932                         BUF_KERNPROC(tbp);
 933                         cluster_append(&bp->b_bio1, tbp);
 934
 935                         /*
 936                          * check for latent dependencies to be handled
 937                          */
 938                         if (LIST_FIRST(&tbp->b_dep) != NULL)
 939                                 buf_start(tbp);
 940                 }
 941         finishcluster:
 942                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
 943                         (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
 944                 if (bp->b_bufsize > bp->b_kvasize) {
 945                         panic(
 946                             "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
 947                             bp->b_bufsize, bp->b_kvasize);
 948                 }
 949                 totalwritten += bp->b_bufsize;
 950                 bp->b_dirtyoff = 0;
 951                 bp->b_dirtyend = bp->b_bufsize;
 952                 bp->b_bio1.bio_done = cluster_callback;
 953                 bp->b_cmd = BUF_CMD_WRITE;
 954
 955                 vfs_busy_pages(vp, bp);
 956                 bp->b_runningbufspace = bp->b_bufsize;
 957                 if (bp->b_runningbufspace) {
 958                         runningbufspace += bp->b_runningbufspace;
 959                         ++runningbufcount;
 960                 }
 961                 BUF_KERNPROC(bp);
 962                 vn_strategy(vp, &bp->b_bio1);
 963
 964                 bytes -= i;
 965         }
 966         return totalwritten;
 967 }
 968
 969 /*
 970  * Collect together all the buffers in a cluster.
 971  * Plus add one additional buffer.
 972  */
 973 static struct cluster_save *
 974 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
 975 {
 976         struct cluster_save *buflist;
 977         struct buf *bp;
 978         off_t loffset;
 979         int i, len;
 980
 981         len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
 982         buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
 983                          M_SEGMENT, M_WAITOK);
 984         buflist->bs_nchildren = 0;
 985         buflist->bs_children = (struct buf **) (buflist + 1);
 986         for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
 987                 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
 988                 buflist->bs_children[i] = bp;
 989                 if (bp->b_bio2.bio_offset == NOOFFSET) {
 990                         VOP_BMAP(bp->b_vp, bp->b_loffset,
 991                                  &bp->b_bio2.bio_offset,
 992                                  NULL, NULL, BUF_CMD_WRITE);
 993                 }
 994         }
 995         buflist->bs_children[i] = bp = last_bp;
 996         if (bp->b_bio2.bio_offset == NOOFFSET) {
 997                 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
 998                          NULL, NULL, BUF_CMD_WRITE);
 999         }
1000         buflist->bs_nchildren = i + 1;
1001         return (buflist);
1002 }
1003
1004 void
1005 cluster_append(struct bio *bio, struct buf *tbp)
1006 {
1007         tbp->b_cluster_next = NULL;
1008         if (bio->bio_caller_info1.cluster_head == NULL) {
1009                 bio->bio_caller_info1.cluster_head = tbp;
1010                 bio->bio_caller_info2.cluster_tail = tbp;
1011         } else {
1012                 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1013                 bio->bio_caller_info2.cluster_tail = tbp;
1014         }
1015 }
1016