BUF/BIO work, for removing the requirement of KVA mappings for I/O
[dragonfly.git] / 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 1.11 2004/07/14 03:10:17 hmp 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 #if defined(CLUSTERDEBUG)
60 #include <sys/sysctl.h>
61 static int      rcluster= 0;
62 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
63 #endif
64
65 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer");
66
67 static struct cluster_save *
68         cluster_collectbufs (struct vnode *vp, struct buf *last_bp);
69 static struct buf *
70         cluster_rbuild (struct vnode *vp, u_quad_t filesize, daddr_t lbn,
71                             daddr_t blkno, long size, int run, struct buf *fbp);
72
73 static int write_behind = 1;
74 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
75
76 extern vm_page_t        bogus_page;
77
78 extern int cluster_pbuf_freecnt;
79
80 /*
81  * Maximum number of blocks for read-ahead.
82  */
83 #define MAXRA 32
84
85 /*
86  * This replaces bread.
87  */
88 int
89 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, 
90         long size, long totread, int seqcount, struct buf **bpp)
91 {
92         struct buf *bp, *rbp, *reqbp;
93         daddr_t blkno, origblkno;
94         int error, num_ra;
95         int i;
96         int maxra, racluster;
97         long origtotread;
98
99         error = 0;
100
101         /*
102          * Try to limit the amount of read-ahead by a few
103          * ad-hoc parameters.  This needs work!!!
104          */
105         racluster = vp->v_mount->mnt_iosize_max / size;
106         maxra = 2 * racluster + (totread / size);
107         if (maxra > MAXRA)
108                 maxra = MAXRA;
109         if (maxra > nbuf/8)
110                 maxra = nbuf/8;
111
112         /*
113          * get the requested block
114          */
115         *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0);
116         origblkno = lblkno;
117         origtotread = totread;
118
119         /*
120          * if it is in the cache, then check to see if the reads have been
121          * sequential.  If they have, then try some read-ahead, otherwise
122          * back-off on prospective read-aheads.
123          */
124         if (bp->b_flags & B_CACHE) {
125                 if (!seqcount) {
126                         return 0;
127                 } else if ((bp->b_flags & B_RAM) == 0) {
128                         return 0;
129                 } else {
130                         int s;
131                         struct buf *tbp;
132                         bp->b_flags &= ~B_RAM;
133                         /*
134                          * We do the spl here so that there is no window
135                          * between the incore and the b_usecount increment
136                          * below.  We opt to keep the spl out of the loop
137                          * for efficiency.
138                          */
139                         s = splbio();
140                         for (i = 1; i < maxra; i++) {
141
142                                 if (!(tbp = incore(vp, lblkno+i))) {
143                                         break;
144                                 }
145
146                                 /*
147                                  * Set another read-ahead mark so we know 
148                                  * to check again.
149                                  */
150                                 if (((i % racluster) == (racluster - 1)) ||
151                                         (i == (maxra - 1)))
152                                         tbp->b_flags |= B_RAM;
153                         }
154                         splx(s);
155                         if (i >= maxra) {
156                                 return 0;
157                         }
158                         lblkno += i;
159                 }
160                 reqbp = bp = NULL;
161         } else {
162                 off_t firstread = bp->b_offset;
163
164                 KASSERT(bp->b_offset != NOOFFSET,
165                     ("cluster_read: no buffer offset"));
166                 if (firstread + totread > filesize)
167                         totread = filesize - firstread;
168                 if (totread > size) {
169                         int nblks = 0;
170                         int ncontigafter;
171                         while (totread > 0) {
172                                 nblks++;
173                                 totread -= size;
174                         }
175                         if (nblks == 1)
176                                 goto single_block_read;
177                         if (nblks > racluster)
178                                 nblks = racluster;
179
180                         error = VOP_BMAP(vp, lblkno, NULL,
181                                 &blkno, &ncontigafter, NULL);
182                         if (error)
183                                 goto single_block_read;
184                         if (blkno == -1)
185                                 goto single_block_read;
186                         if (ncontigafter == 0)
187                                 goto single_block_read;
188                         if (ncontigafter + 1 < nblks)
189                                 nblks = ncontigafter + 1;
190
191                         bp = cluster_rbuild(vp, filesize, lblkno,
192                                 blkno, size, nblks, bp);
193                         lblkno += (bp->b_bufsize / size);
194                 } else {
195 single_block_read:
196                         /*
197                          * if it isn't in the cache, then get a chunk from
198                          * disk if sequential, otherwise just get the block.
199                          */
200                         bp->b_flags |= B_READ | B_RAM;
201                         lblkno += 1;
202                 }
203         }
204
205         /*
206          * if we have been doing sequential I/O, then do some read-ahead
207          */
208         rbp = NULL;
209         if (seqcount && (lblkno < (origblkno + seqcount))) {
210                 /*
211                  * we now build the read-ahead buffer if it is desirable.
212                  */
213                 if (((u_quad_t)(lblkno + 1) * size) <= filesize &&
214                     !(error = VOP_BMAP(vp, lblkno, NULL, &blkno, &num_ra, NULL)) &&
215                     blkno != -1) {
216                         int nblksread;
217                         int ntoread = num_ra + 1;
218                         nblksread = (origtotread + size - 1) / size;
219                         if (seqcount < nblksread)
220                                 seqcount = nblksread;
221                         if (seqcount < ntoread)
222                                 ntoread = seqcount;
223                         if (num_ra) {
224                                 rbp = cluster_rbuild(vp, filesize, lblkno,
225                                         blkno, size, ntoread, NULL);
226                         } else {
227                                 rbp = getblk(vp, lblkno, size, 0, 0);
228                                 rbp->b_flags |= B_READ | B_ASYNC | B_RAM;
229                                 rbp->b_blkno = blkno;
230                         }
231                 }
232         }
233
234         /*
235          * handle the synchronous read
236          */
237         if (bp) {
238 #if defined(CLUSTERDEBUG)
239                 if (rcluster)
240                         printf("S(%ld,%ld,%d) ",
241                             (long)bp->b_lblkno, bp->b_bcount, seqcount);
242 #endif
243                 if ((bp->b_flags & B_CLUSTER) == 0) {
244                         vfs_busy_pages(bp, 0);
245                 }
246                 bp->b_flags &= ~(B_ERROR|B_INVAL);
247                 if (bp->b_flags & (B_ASYNC|B_CALL))
248                         BUF_KERNPROC(bp);
249                 error = VOP_STRATEGY(vp, bp);
250         }
251
252         /*
253          * and if we have read-aheads, do them too
254          */
255         if (rbp) {
256                 if (error) {
257                         rbp->b_flags &= ~(B_ASYNC | B_READ);
258                         brelse(rbp);
259                 } else if (rbp->b_flags & B_CACHE) {
260                         rbp->b_flags &= ~(B_ASYNC | B_READ);
261                         bqrelse(rbp);
262                 } else {
263 #if defined(CLUSTERDEBUG)
264                         if (rcluster) {
265                                 if (bp)
266                                         printf("A+(%ld,%ld,%ld,%d) ",
267                                             (long)rbp->b_lblkno, rbp->b_bcount,
268                                             (long)(rbp->b_lblkno - origblkno),
269                                             seqcount);
270                                 else
271                                         printf("A(%ld,%ld,%ld,%d) ",
272                                             (long)rbp->b_lblkno, rbp->b_bcount,
273                                             (long)(rbp->b_lblkno - origblkno),
274                                             seqcount);
275                         }
276 #endif
277
278                         if ((rbp->b_flags & B_CLUSTER) == 0) {
279                                 vfs_busy_pages(rbp, 0);
280                         }
281                         rbp->b_flags &= ~(B_ERROR|B_INVAL);
282                         if (rbp->b_flags & (B_ASYNC|B_CALL))
283                                 BUF_KERNPROC(rbp);
284                         (void) VOP_STRATEGY(vp, rbp);
285                 }
286         }
287         if (reqbp)
288                 return (biowait(reqbp));
289         else
290                 return (error);
291 }
292
293 /*
294  * If blocks are contiguous on disk, use this to provide clustered
295  * read ahead.  We will read as many blocks as possible sequentially
296  * and then parcel them up into logical blocks in the buffer hash table.
297  */
298 static struct buf *
299 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn, 
300         daddr_t blkno, long size, int run, struct buf *fbp)
301 {
302         struct buf *bp, *tbp;
303         daddr_t bn;
304         int i, inc, j;
305
306         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
307             ("cluster_rbuild: size %ld != filesize %ld\n",
308             size, vp->v_mount->mnt_stat.f_iosize));
309
310         /*
311          * avoid a division
312          */
313         while ((u_quad_t) size * (lbn + run) > filesize) {
314                 --run;
315         }
316
317         if (fbp) {
318                 tbp = fbp;
319                 tbp->b_flags |= B_READ; 
320         } else {
321                 tbp = getblk(vp, lbn, size, 0, 0);
322                 if (tbp->b_flags & B_CACHE)
323                         return tbp;
324                 tbp->b_flags |= B_ASYNC | B_READ | B_RAM;
325         }
326
327         tbp->b_blkno = blkno;
328         if( (tbp->b_flags & B_MALLOC) ||
329                 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
330                 return tbp;
331
332         bp = trypbuf(&cluster_pbuf_freecnt);
333         if (bp == 0)
334                 return tbp;
335
336         /*
337          * We are synthesizing a buffer out of vm_page_t's, but
338          * if the block size is not page aligned then the starting
339          * address may not be either.  Inherit the b_data offset
340          * from the original buffer.
341          */
342         bp->b_data = (char *)((vm_offset_t)bp->b_data |
343             ((vm_offset_t)tbp->b_data & PAGE_MASK));
344         bp->b_flags = B_ASYNC | B_READ | B_CALL | B_CLUSTER | B_VMIO;
345         bp->b_iodone = cluster_callback;
346         bp->b_blkno = blkno;
347         bp->b_lblkno = lbn;
348         bp->b_offset = tbp->b_offset;
349         KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
350         pbgetvp(vp, bp);
351
352         TAILQ_INIT(&bp->b_cluster.cluster_head);
353
354         bp->b_bcount = 0;
355         bp->b_bufsize = 0;
356         bp->b_xio.xio_npages = 0;
357
358         inc = btodb(size);
359         for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
360                 if (i != 0) {
361                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
362                             round_page(size) > vp->v_mount->mnt_iosize_max) {
363                                 break;
364                         }
365
366                         /*
367                          * Shortcut some checks and try to avoid buffers that
368                          * would block in the lock.  The same checks have to
369                          * be made again after we officially get the buffer.
370                          */
371                         if ((tbp = incore(vp, lbn + i)) != NULL) {
372                                 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
373                                         break;
374                                 BUF_UNLOCK(tbp);
375
376                                 for (j = 0; j < tbp->b_xio.xio_npages; j++) {
377                                         if (tbp->b_xio.xio_pages[j]->valid)
378                                                 break;
379                                 }
380                                 
381                                 if (j != tbp->b_xio.xio_npages)
382                                         break;
383         
384                                 if (tbp->b_bcount != size)
385                                         break;
386                         }
387
388                         tbp = getblk(vp, lbn + i, size, 0, 0);
389
390                         /*
391                          * Stop scanning if the buffer is fuly valid 
392                          * (marked B_CACHE), or locked (may be doing a
393                          * background write), or if the buffer is not
394                          * VMIO backed.  The clustering code can only deal
395                          * with VMIO-backed buffers.
396                          */
397                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
398                             (tbp->b_flags & B_VMIO) == 0) {
399                                 bqrelse(tbp);
400                                 break;
401                         }
402
403                         /*
404                          * The buffer must be completely invalid in order to
405                          * take part in the cluster.  If it is partially valid
406                          * then we stop.
407                          */
408                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
409                                 if (tbp->b_xio.xio_pages[j]->valid)
410                                         break;
411                         }
412                         if (j != tbp->b_xio.xio_npages) {
413                                 bqrelse(tbp);
414                                 break;
415                         }
416
417                         /*
418                          * Set a read-ahead mark as appropriate
419                          */
420                         if ((fbp && (i == 1)) || (i == (run - 1)))
421                                 tbp->b_flags |= B_RAM;
422
423                         /*
424                          * Set the buffer up for an async read (XXX should
425                          * we do this only if we do not wind up brelse()ing?).
426                          * Set the block number if it isn't set, otherwise
427                          * if it is make sure it matches the block number we
428                          * expect.
429                          */
430                         tbp->b_flags |= B_READ | B_ASYNC;
431                         if (tbp->b_blkno == tbp->b_lblkno) {
432                                 tbp->b_blkno = bn;
433                         } else if (tbp->b_blkno != bn) {
434                                 brelse(tbp);
435                                 break;
436                         }
437                 }
438                 /*
439                  * XXX fbp from caller may not be B_ASYNC, but we are going
440                  * to biodone() it in cluster_callback() anyway
441                  */
442                 BUF_KERNPROC(tbp);
443                 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
444                         tbp, b_cluster.cluster_entry);
445                 for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
446                         vm_page_t m;
447                         m = tbp->b_xio.xio_pages[j];
448                         vm_page_io_start(m);
449                         vm_object_pip_add(m->object, 1);
450                         if ((bp->b_xio.xio_npages == 0) ||
451                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
452                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
453                                 bp->b_xio.xio_npages++;
454                         }
455                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
456                                 tbp->b_xio.xio_pages[j] = bogus_page;
457                 }
458                 /*
459                  * XXX shouldn't this be += size for both, like in 
460                  * cluster_wbuild()?
461                  *
462                  * Don't inherit tbp->b_bufsize as it may be larger due to
463                  * a non-page-aligned size.  Instead just aggregate using
464                  * 'size'.
465                  */
466                 if (tbp->b_bcount != size)
467                     printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
468                 if (tbp->b_bufsize != size)
469                     printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
470                 bp->b_bcount += size;
471                 bp->b_bufsize += size;
472         }
473
474         /*
475          * Fully valid pages in the cluster are already good and do not need
476          * to be re-read from disk.  Replace the page with bogus_page
477          */
478         for (j = 0; j < bp->b_xio.xio_npages; j++) {
479                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
480                     VM_PAGE_BITS_ALL) {
481                         bp->b_xio.xio_pages[j] = bogus_page;
482                 }
483         }
484         if (bp->b_bufsize > bp->b_kvasize)
485                 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)",
486                     bp->b_bufsize, bp->b_kvasize);
487         bp->b_kvasize = bp->b_bufsize;
488
489         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
490                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
491         return (bp);
492 }
493
494 /*
495  * Cleanup after a clustered read or write.
496  * This is complicated by the fact that any of the buffers might have
497  * extra memory (if there were no empty buffer headers at allocbuf time)
498  * that we will need to shift around.
499  */
500 void
501 cluster_callback(struct buf *bp)
502 {
503         struct buf *nbp, *tbp;
504         int error = 0;
505
506         /*
507          * Must propogate errors to all the components.
508          */
509         if (bp->b_flags & B_ERROR)
510                 error = bp->b_error;
511
512         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
513         /*
514          * Move memory from the large cluster buffer into the component
515          * buffers and mark IO as done on these.
516          */
517         for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
518                 tbp; tbp = nbp) {
519                 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
520                 if (error) {
521                         tbp->b_flags |= B_ERROR;
522                         tbp->b_error = error;
523                 } else {
524                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
525                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
526                         /*
527                          * XXX the bdwrite()/bqrelse() issued during
528                          * cluster building clears B_RELBUF (see bqrelse()
529                          * comment).  If direct I/O was specified, we have
530                          * to restore it here to allow the buffer and VM
531                          * to be freed.
532                          */
533                         if (tbp->b_flags & B_DIRECT)
534                                 tbp->b_flags |= B_RELBUF;
535                 }
536                 biodone(tbp);
537         }
538         relpbuf(bp, &cluster_pbuf_freecnt);
539 }
540
541 /*
542  *      cluster_wbuild_wb:
543  *
544  *      Implement modified write build for cluster.
545  *
546  *              write_behind = 0        write behind disabled
547  *              write_behind = 1        write behind normal (default)
548  *              write_behind = 2        write behind backed-off
549  */
550
551 static __inline int
552 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
553 {
554         int r = 0;
555
556         switch(write_behind) {
557         case 2:
558                 if (start_lbn < len)
559                         break;
560                 start_lbn -= len;
561                 /* fall through */
562         case 1:
563                 r = cluster_wbuild(vp, size, start_lbn, len);
564                 /* fall through */
565         default:
566                 /* fall through */
567                 break;
568         }
569         return(r);
570 }
571
572 /*
573  * Do clustered write for FFS.
574  *
575  * Three cases:
576  *      1. Write is not sequential (write asynchronously)
577  *      Write is sequential:
578  *      2.      beginning of cluster - begin cluster
579  *      3.      middle of a cluster - add to cluster
580  *      4.      end of a cluster - asynchronously write cluster
581  */
582 void
583 cluster_write(struct buf *bp, u_quad_t filesize, int seqcount)
584 {
585         struct vnode *vp;
586         daddr_t lbn;
587         int maxclen, cursize;
588         int lblocksize;
589         int async;
590
591         vp = bp->b_vp;
592         if (vp->v_type == VREG) {
593                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
594                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
595         } else {
596                 async = 0;
597                 lblocksize = bp->b_bufsize;
598         }
599         lbn = bp->b_lblkno;
600         KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
601
602         /* Initialize vnode to beginning of file. */
603         if (lbn == 0)
604                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
605
606         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
607             (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
608                 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
609                 if (vp->v_clen != 0) {
610                         /*
611                          * Next block is not sequential.
612                          *
613                          * If we are not writing at end of file, the process
614                          * seeked to another point in the file since its last
615                          * write, or we have reached our maximum cluster size,
616                          * then push the previous cluster. Otherwise try
617                          * reallocating to make it sequential.
618                          *
619                          * Change to algorithm: only push previous cluster if
620                          * it was sequential from the point of view of the
621                          * seqcount heuristic, otherwise leave the buffer 
622                          * intact so we can potentially optimize the I/O
623                          * later on in the buf_daemon or update daemon
624                          * flush.
625                          */
626                         cursize = vp->v_lastw - vp->v_cstart + 1;
627                         if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
628                             lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
629                                 if (!async && seqcount > 0) {
630                                         cluster_wbuild_wb(vp, lblocksize,
631                                                 vp->v_cstart, cursize);
632                                 }
633                         } else {
634                                 struct buf **bpp, **endbp;
635                                 struct cluster_save *buflist;
636
637                                 buflist = cluster_collectbufs(vp, bp);
638                                 endbp = &buflist->bs_children
639                                     [buflist->bs_nchildren - 1];
640                                 if (VOP_REALLOCBLKS(vp, buflist)) {
641                                         /*
642                                          * Failed, push the previous cluster
643                                          * if *really* writing sequentially
644                                          * in the logical file (seqcount > 1),
645                                          * otherwise delay it in the hopes that
646                                          * the low level disk driver can
647                                          * optimize the write ordering.
648                                          */
649                                         for (bpp = buflist->bs_children;
650                                              bpp < endbp; bpp++)
651                                                 brelse(*bpp);
652                                         free(buflist, M_SEGMENT);
653                                         if (seqcount > 1) {
654                                                 cluster_wbuild_wb(vp, 
655                                                     lblocksize, vp->v_cstart, 
656                                                     cursize);
657                                         }
658                                 } else {
659                                         /*
660                                          * Succeeded, keep building cluster.
661                                          */
662                                         for (bpp = buflist->bs_children;
663                                              bpp <= endbp; bpp++)
664                                                 bdwrite(*bpp);
665                                         free(buflist, M_SEGMENT);
666                                         vp->v_lastw = lbn;
667                                         vp->v_lasta = bp->b_blkno;
668                                         return;
669                                 }
670                         }
671                 }
672                 /*
673                  * Consider beginning a cluster. If at end of file, make
674                  * cluster as large as possible, otherwise find size of
675                  * existing cluster.
676                  */
677                 if ((vp->v_type == VREG) &&
678                         ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
679                     (bp->b_blkno == bp->b_lblkno) &&
680                     (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
681                      bp->b_blkno == -1)) {
682                         bawrite(bp);
683                         vp->v_clen = 0;
684                         vp->v_lasta = bp->b_blkno;
685                         vp->v_cstart = lbn + 1;
686                         vp->v_lastw = lbn;
687                         return;
688                 }
689                 vp->v_clen = maxclen;
690                 if (!async && maxclen == 0) {   /* I/O not contiguous */
691                         vp->v_cstart = lbn + 1;
692                         bawrite(bp);
693                 } else {        /* Wait for rest of cluster */
694                         vp->v_cstart = lbn;
695                         bdwrite(bp);
696                 }
697         } else if (lbn == vp->v_cstart + vp->v_clen) {
698                 /*
699                  * At end of cluster, write it out if seqcount tells us we
700                  * are operating sequentially, otherwise let the buf or
701                  * update daemon handle it.
702                  */
703                 bdwrite(bp);
704                 if (seqcount > 1)
705                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
706                 vp->v_clen = 0;
707                 vp->v_cstart = lbn + 1;
708         } else if (vm_page_count_severe()) {
709                 /*
710                  * We are low on memory, get it going NOW
711                  */
712                 bawrite(bp);
713         } else {
714                 /*
715                  * In the middle of a cluster, so just delay the I/O for now.
716                  */
717                 bdwrite(bp);
718         }
719         vp->v_lastw = lbn;
720         vp->v_lasta = bp->b_blkno;
721 }
722
723
724 /*
725  * This is an awful lot like cluster_rbuild...wish they could be combined.
726  * The last lbn argument is the current block on which I/O is being
727  * performed.  Check to see that it doesn't fall in the middle of
728  * the current block (if last_bp == NULL).
729  */
730 int
731 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len)
732 {
733         struct buf *bp, *tbp;
734         int i, j, s;
735         int totalwritten = 0;
736         int dbsize = btodb(size);
737
738         while (len > 0) {
739                 s = splbio();
740                 /*
741                  * If the buffer is not delayed-write (i.e. dirty), or it 
742                  * is delayed-write but either locked or inval, it cannot 
743                  * partake in the clustered write.
744                  */
745                 if (((tbp = gbincore(vp, start_lbn)) == NULL) ||
746                   ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
747                   BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
748                         ++start_lbn;
749                         --len;
750                         splx(s);
751                         continue;
752                 }
753                 bremfree(tbp);
754                 tbp->b_flags &= ~B_DONE;
755                 splx(s);
756
757                 /*
758                  * Extra memory in the buffer, punt on this buffer.
759                  * XXX we could handle this in most cases, but we would
760                  * have to push the extra memory down to after our max
761                  * possible cluster size and then potentially pull it back
762                  * up if the cluster was terminated prematurely--too much
763                  * hassle.
764                  */
765                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
766                   (tbp->b_bcount != tbp->b_bufsize) ||
767                   (tbp->b_bcount != size) ||
768                   (len == 1) ||
769                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
770                         totalwritten += tbp->b_bufsize;
771                         bawrite(tbp);
772                         ++start_lbn;
773                         --len;
774                         continue;
775                 }
776
777                 /*
778                  * We got a pbuf to make the cluster in.
779                  * so initialise it.
780                  */
781                 TAILQ_INIT(&bp->b_cluster.cluster_head);
782                 bp->b_bcount = 0;
783                 bp->b_bufsize = 0;
784                 bp->b_xio.xio_npages = 0;
785                 bp->b_blkno = tbp->b_blkno;
786                 bp->b_lblkno = tbp->b_lblkno;
787                 bp->b_offset = tbp->b_offset;
788
789                 /*
790                  * We are synthesizing a buffer out of vm_page_t's, but
791                  * if the block size is not page aligned then the starting
792                  * address may not be either.  Inherit the b_data offset
793                  * from the original buffer.
794                  */
795                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
796                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
797                 bp->b_flags |= B_CALL | B_CLUSTER |
798                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN));
799                 bp->b_iodone = cluster_callback;
800                 pbgetvp(vp, bp);
801                 /*
802                  * From this location in the file, scan forward to see
803                  * if there are buffers with adjacent data that need to
804                  * be written as well.
805                  */
806                 for (i = 0; i < len; ++i, ++start_lbn) {
807                         if (i != 0) { /* If not the first buffer */
808                                 s = splbio();
809                                 /*
810                                  * If the adjacent data is not even in core it
811                                  * can't need to be written.
812                                  */
813                                 if ((tbp = gbincore(vp, start_lbn)) == NULL) {
814                                         splx(s);
815                                         break;
816                                 }
817
818                                 /*
819                                  * If it IS in core, but has different
820                                  * characteristics, or is locked (which
821                                  * means it could be undergoing a background
822                                  * I/O or be in a weird state), then don't
823                                  * cluster with it.
824                                  */
825                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
826                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
827                                   != (B_DELWRI | B_CLUSTEROK |
828                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
829                                     (tbp->b_flags & B_LOCKED) ||
830                                     BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
831                                         splx(s);
832                                         break;
833                                 }
834
835                                 /*
836                                  * Check that the combined cluster
837                                  * would make sense with regard to pages
838                                  * and would not be too large
839                                  */
840                                 if ((tbp->b_bcount != size) ||
841                                   ((bp->b_blkno + (dbsize * i)) !=
842                                     tbp->b_blkno) ||
843                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
844                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
845                                         BUF_UNLOCK(tbp);
846                                         splx(s);
847                                         break;
848                                 }
849                                 /*
850                                  * Ok, it's passed all the tests,
851                                  * so remove it from the free list
852                                  * and mark it busy. We will use it.
853                                  */
854                                 bremfree(tbp);
855                                 tbp->b_flags &= ~B_DONE;
856                                 splx(s);
857                         } /* end of code for non-first buffers only */
858                         /* check for latent dependencies to be handled */
859                         if ((LIST_FIRST(&tbp->b_dep)) != NULL &&
860                             bioops.io_start)
861                                 (*bioops.io_start)(tbp);
862                         /*
863                          * If the IO is via the VM then we do some
864                          * special VM hackery (yuck).  Since the buffer's
865                          * block size may not be page-aligned it is possible
866                          * for a page to be shared between two buffers.  We
867                          * have to get rid of the duplication when building
868                          * the cluster.
869                          */
870                         if (tbp->b_flags & B_VMIO) {
871                                 vm_page_t m;
872
873                                 if (i != 0) { /* if not first buffer */
874                                         for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
875                                                 m = tbp->b_xio.xio_pages[j];
876                                                 if (m->flags & PG_BUSY) {
877                                                         bqrelse(tbp);
878                                                         goto finishcluster;
879                                                 }
880                                         }
881                                 }
882                                         
883                                 for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
884                                         m = tbp->b_xio.xio_pages[j];
885                                         vm_page_io_start(m);
886                                         vm_object_pip_add(m->object, 1);
887                                         if ((bp->b_xio.xio_npages == 0) ||
888                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
889                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
890                                                 bp->b_xio.xio_npages++;
891                                         }
892                                 }
893                         }
894                         bp->b_bcount += size;
895                         bp->b_bufsize += size;
896
897                         s = splbio();
898                         bundirty(tbp);
899                         tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
900                         tbp->b_flags |= B_ASYNC;
901                         reassignbuf(tbp, tbp->b_vp);    /* put on clean list */
902                         ++tbp->b_vp->v_numoutput;
903                         splx(s);
904                         BUF_KERNPROC(tbp);
905                         TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
906                                 tbp, b_cluster.cluster_entry);
907                 }
908         finishcluster:
909                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
910                         (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
911                 if (bp->b_bufsize > bp->b_kvasize)
912                         panic(
913                             "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
914                             bp->b_bufsize, bp->b_kvasize);
915                 bp->b_kvasize = bp->b_bufsize;
916                 totalwritten += bp->b_bufsize;
917                 bp->b_dirtyoff = 0;
918                 bp->b_dirtyend = bp->b_bufsize;
919                 bawrite(bp);
920
921                 len -= i;
922         }
923         return totalwritten;
924 }
925
926 /*
927  * Collect together all the buffers in a cluster.
928  * Plus add one additional buffer.
929  */
930 static struct cluster_save *
931 cluster_collectbufs(struct vnode *vp, struct buf *last_bp)
932 {
933         struct cluster_save *buflist;
934         struct buf *bp;
935         daddr_t lbn;
936         int i, len;
937
938         len = vp->v_lastw - vp->v_cstart + 1;
939         buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
940             M_SEGMENT, M_WAITOK);
941         buflist->bs_nchildren = 0;
942         buflist->bs_children = (struct buf **) (buflist + 1);
943         for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
944                 (void) bread(vp, lbn, last_bp->b_bcount, &bp);
945                 buflist->bs_children[i] = bp;
946                 if (bp->b_blkno == bp->b_lblkno)
947                         VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
948                                 NULL, NULL);
949         }
950         buflist->bs_children[i] = bp = last_bp;
951         if (bp->b_blkno == bp->b_lblkno)
952                 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
953                         NULL, NULL);
954         buflist->bs_nchildren = i + 1;
955         return (buflist);
956 }