Keep track of the number of buffers undgoing IO, and include that number
[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 1.34 2008/05/06 00:13:54 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 #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", "cluster_save buffer");
66
67 static struct cluster_save *
68         cluster_collectbufs (struct vnode *vp, struct buf *last_bp,
69                             int lblocksize);
70 static struct buf *
71         cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
72                             off_t doffset, int size, int run, 
73                             struct buf *fbp, int doasync);
74 static void cluster_callback (struct bio *);
75
76
77 static int write_behind = 1;
78 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
79
80 extern vm_page_t        bogus_page;
81
82 extern int cluster_pbuf_freecnt;
83
84 /*
85  * Maximum number of blocks for read-ahead.
86  */
87 #define MAXRA 32
88
89 /*
90  * This replaces bread.
91  */
92 int
93 cluster_read(struct vnode *vp, off_t filesize, off_t loffset, 
94              int size, int totread, int seqcount, struct buf **bpp)
95 {
96         struct buf *bp, *rbp, *reqbp;
97         off_t origoffset;
98         off_t doffset;
99         int error;
100         int i;
101         int maxra, racluster;
102
103         error = 0;
104
105         /*
106          * Try to limit the amount of read-ahead by a few
107          * ad-hoc parameters.  This needs work!!!
108          */
109         racluster = vp->v_mount->mnt_iosize_max / size;
110         maxra = 2 * racluster + (totread / size);
111         if (maxra > MAXRA)
112                 maxra = MAXRA;
113         if (maxra > nbuf/8)
114                 maxra = nbuf/8;
115
116         /*
117          * get the requested block
118          */
119         *bpp = reqbp = bp = getblk(vp, loffset, size, 0, 0);
120         origoffset = loffset;
121
122         /*
123          * if it is in the cache, then check to see if the reads have been
124          * sequential.  If they have, then try some read-ahead, otherwise
125          * back-off on prospective read-aheads.
126          */
127         if (bp->b_flags & B_CACHE) {
128                 if (!seqcount) {
129                         return 0;
130                 } else if ((bp->b_flags & B_RAM) == 0) {
131                         return 0;
132                 } else {
133                         struct buf *tbp;
134                         bp->b_flags &= ~B_RAM;
135                         /*
136                          * We do the crit here so that there is no window
137                          * between the findblk and the b_usecount increment
138                          * below.  We opt to keep the crit out of the loop
139                          * for efficiency.
140                          */
141                         crit_enter();
142                         for (i = 1; i < maxra; i++) {
143                                 if (!(tbp = findblk(vp, loffset + i * size))) {
144                                         break;
145                                 }
146
147                                 /*
148                                  * Set another read-ahead mark so we know 
149                                  * to check again.
150                                  */
151                                 if (((i % racluster) == (racluster - 1)) ||
152                                         (i == (maxra - 1)))
153                                         tbp->b_flags |= B_RAM;
154                         }
155                         crit_exit();
156                         if (i >= maxra) {
157                                 return 0;
158                         }
159                         loffset += i * size;
160                 }
161                 reqbp = bp = NULL;
162         } else {
163                 off_t firstread = bp->b_loffset;
164                 int nblks;
165
166                 KASSERT(firstread != NOOFFSET, 
167                         ("cluster_read: no buffer offset"));
168                 if (firstread + totread > filesize)
169                         totread = (int)(filesize - firstread);
170                 nblks = totread / size;
171                 if (nblks) {
172                         int burstbytes;
173
174                         if (nblks > racluster)
175                                 nblks = racluster;
176
177                         error = VOP_BMAP(vp, loffset,
178                                          &doffset, &burstbytes, NULL);
179                         if (error)
180                                 goto single_block_read;
181                         if (doffset == NOOFFSET)
182                                 goto single_block_read;
183                         if (burstbytes < size * 2)
184                                 goto single_block_read;
185                         if (nblks > burstbytes / size)
186                                 nblks = burstbytes / size;
187
188                         bp = cluster_rbuild(vp, filesize, loffset,
189                                             doffset, size, nblks, bp, 0);
190                         loffset += bp->b_bufsize;
191                 } else {
192 single_block_read:
193                         /*
194                          * if it isn't in the cache, then get a chunk from
195                          * disk if sequential, otherwise just get the block.
196                          */
197                         bp->b_flags |= B_RAM;
198                         loffset += size;
199                 }
200         }
201
202         /*
203          * Handle the synchronous read.  This only occurs if B_CACHE was
204          * not set.  bp (and rbp) could be either a cluster bp or a normal
205          * bp depending on the what cluster_rbuild() decided to do.  If
206          * it is a cluster bp, vfs_busy_pages() has already been called.
207          */
208         if (bp) {
209 #if defined(CLUSTERDEBUG)
210                 if (rcluster)
211                         kprintf("S(%lld,%d,%d) ",
212                             bp->b_loffset, bp->b_bcount, seqcount);
213 #endif
214                 bp->b_cmd = BUF_CMD_READ;
215                 if ((bp->b_flags & B_CLUSTER) == 0)
216                         vfs_busy_pages(vp, bp);
217                 bp->b_flags &= ~(B_ERROR|B_INVAL);
218                 if ((bp->b_flags & B_ASYNC) || bp->b_bio1.bio_done != NULL)
219                         BUF_KERNPROC(bp);
220                 vn_strategy(vp, &bp->b_bio1);
221                 error = bp->b_error;
222         }
223
224         /*
225          * If we have been doing sequential I/O, then do some read-ahead.
226          */
227         rbp = NULL;
228         if (!error &&
229             seqcount &&
230             loffset < origoffset + seqcount * size &&
231             loffset + size <= filesize
232         ) {
233                 int nblksread;
234                 int ntoread;
235                 int burstbytes;
236
237                 rbp = getblk(vp, loffset, size, 0, 0);
238                 if ((rbp->b_flags & B_CACHE)) {
239                         bqrelse(rbp);
240                         goto no_read_ahead;
241                 }
242
243                 error = VOP_BMAP(vp, loffset,
244                                  &doffset, &burstbytes, NULL);
245                 if (error || doffset == NOOFFSET) {
246                         rbp->b_flags |= B_INVAL;
247                         brelse(rbp);
248                         rbp = NULL;
249                         goto no_read_ahead;
250                 }
251                 ntoread = burstbytes / size;
252                 nblksread = (totread + size - 1) / size;
253                 if (seqcount < nblksread)
254                         seqcount = nblksread;
255                 if (seqcount < ntoread)
256                         ntoread = seqcount;
257
258                 rbp->b_flags |= B_RAM;
259                 if (burstbytes) {
260                         rbp = cluster_rbuild(vp, filesize, loffset,
261                                              doffset, size, 
262                                              ntoread, rbp, 1);
263                 } else {
264                         rbp->b_bio2.bio_offset = doffset;
265                 }
266 #if defined(CLUSTERDEBUG)
267                 if (rcluster) {
268                         if (bp)
269                                 kprintf("A+(%lld,%d,%lld,%d) ",
270                                     rbp->b_loffset, rbp->b_bcount,
271                                     rbp->b_loffset - origoffset,
272                                     seqcount);
273                         else
274                                 kprintf("A(%lld,%d,%lld,%d) ",
275                                     rbp->b_loffset, rbp->b_bcount,
276                                     rbp->b_loffset - origoffset,
277                                     seqcount);
278                 }
279 #endif
280                 rbp->b_flags &= ~(B_ERROR|B_INVAL);
281                 rbp->b_flags |= B_ASYNC;
282                 rbp->b_cmd = BUF_CMD_READ;
283
284                 if ((rbp->b_flags & B_CLUSTER) == 0)
285                         vfs_busy_pages(vp, rbp);
286                 BUF_KERNPROC(rbp);                      /* B_ASYNC */
287                 vn_strategy(vp, &rbp->b_bio1);
288         }
289 no_read_ahead:
290
291         if (reqbp)
292                 return (biowait(reqbp));
293         else
294                 return (error);
295 }
296
297 /*
298  * If blocks are contiguous on disk, use this to provide clustered
299  * read ahead.  We will read as many blocks as possible sequentially
300  * and then parcel them up into logical blocks in the buffer hash table.
301  */
302 static struct buf *
303 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, 
304         off_t doffset, int size, int run, struct buf *fbp, int doasync)
305 {
306         struct buf *bp, *tbp;
307         off_t boffset;
308         int i, j;
309
310         KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
311             ("cluster_rbuild: size %d != filesize %ld\n",
312             size, vp->v_mount->mnt_stat.f_iosize));
313
314         /*
315          * avoid a division
316          */
317         while (loffset + run * size > filesize) {
318                 --run;
319         }
320
321         tbp = fbp;
322         tbp->b_bio2.bio_offset = doffset;
323         if((tbp->b_flags & B_MALLOC) ||
324             ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
325                 return tbp;
326         }
327
328         bp = trypbuf(&cluster_pbuf_freecnt);
329         if (bp == NULL)
330                 return tbp;
331
332         /*
333          * We are synthesizing a buffer out of vm_page_t's, but
334          * if the block size is not page aligned then the starting
335          * address may not be either.  Inherit the b_data offset
336          * from the original buffer.
337          */
338         bp->b_data = (char *)((vm_offset_t)bp->b_data |
339             ((vm_offset_t)tbp->b_data & PAGE_MASK));
340         bp->b_flags |= B_ASYNC | B_CLUSTER | B_VMIO;
341         bp->b_cmd = BUF_CMD_READ;
342         bp->b_bio1.bio_done = cluster_callback;
343         bp->b_bio1.bio_caller_info1.cluster_head = NULL;
344         bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
345         bp->b_loffset = loffset;
346         bp->b_bio2.bio_offset = NOOFFSET;
347         KASSERT(bp->b_loffset != NOOFFSET,
348                 ("cluster_rbuild: no buffer offset"));
349
350         bp->b_bcount = 0;
351         bp->b_bufsize = 0;
352         bp->b_xio.xio_npages = 0;
353
354         for (boffset = doffset, i = 0; i < run; ++i, boffset += size) {
355                 if (i) {
356                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
357                             round_page(size) > vp->v_mount->mnt_iosize_max) {
358                                 break;
359                         }
360
361                         /*
362                          * Shortcut some checks and try to avoid buffers that
363                          * would block in the lock.  The same checks have to
364                          * be made again after we officially get the buffer.
365                          */
366                         if ((tbp = findblk(vp, loffset + i * size)) != NULL) {
367                                 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
368                                         break;
369                                 BUF_UNLOCK(tbp);
370
371                                 for (j = 0; j < tbp->b_xio.xio_npages; j++) {
372                                         if (tbp->b_xio.xio_pages[j]->valid)
373                                                 break;
374                                 }
375                                 
376                                 if (j != tbp->b_xio.xio_npages)
377                                         break;
378         
379                                 if (tbp->b_bcount != size)
380                                         break;
381                         }
382
383                         tbp = getblk(vp, loffset + i * size, size, 0, 0);
384
385                         /*
386                          * Stop scanning if the buffer is fuly valid 
387                          * (marked B_CACHE), or locked (may be doing a
388                          * background write), or if the buffer is not
389                          * VMIO backed.  The clustering code can only deal
390                          * with VMIO-backed buffers.
391                          */
392                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
393                             (tbp->b_flags & B_VMIO) == 0 ||
394                             (LIST_FIRST(&tbp->b_dep) != NULL &&
395                              buf_checkread(tbp))
396                         ) {
397                                 bqrelse(tbp);
398                                 break;
399                         }
400
401                         /*
402                          * The buffer must be completely invalid in order to
403                          * take part in the cluster.  If it is partially valid
404                          * then we stop.
405                          */
406                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
407                                 if (tbp->b_xio.xio_pages[j]->valid)
408                                         break;
409                         }
410                         if (j != tbp->b_xio.xio_npages) {
411                                 bqrelse(tbp);
412                                 break;
413                         }
414
415                         /*
416                          * Set a read-ahead mark as appropriate
417                          */
418                         if (i == 1 || i == (run - 1))
419                                 tbp->b_flags |= B_RAM;
420
421                         /*
422                          * Set the block number if it isn't set, otherwise
423                          * if it is make sure it matches the block number we
424                          * expect.
425                          */
426                         if (tbp->b_bio2.bio_offset == NOOFFSET) {
427                                 tbp->b_bio2.bio_offset = boffset;
428                         } else if (tbp->b_bio2.bio_offset != boffset) {
429                                 brelse(tbp);
430                                 break;
431                         }
432                 }
433                 /*
434                  * The first buffer is setup async if doasync is specified.
435                  * All other buffers in the cluster are setup async.  This
436                  * way the caller can decide how to deal with the requested
437                  * buffer.
438                  */
439                 if (i || doasync)
440                         tbp->b_flags |= B_ASYNC;
441                 tbp->b_cmd = BUF_CMD_READ;
442                 BUF_KERNPROC(tbp);
443                 cluster_append(&bp->b_bio1, tbp);
444                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
445                         vm_page_t m;
446                         m = tbp->b_xio.xio_pages[j];
447                         vm_page_io_start(m);
448                         vm_object_pip_add(m->object, 1);
449                         if ((bp->b_xio.xio_npages == 0) ||
450                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
451                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
452                                 bp->b_xio.xio_npages++;
453                         }
454                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
455                                 tbp->b_xio.xio_pages[j] = bogus_page;
456                 }
457                 /*
458                  * XXX shouldn't this be += size for both, like in 
459                  * cluster_wbuild()?
460                  *
461                  * Don't inherit tbp->b_bufsize as it may be larger due to
462                  * a non-page-aligned size.  Instead just aggregate using
463                  * 'size'.
464                  */
465                 if (tbp->b_bcount != size)
466                     kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, size);
467                 if (tbp->b_bufsize != size)
468                     kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, size);
469                 bp->b_bcount += size;
470                 bp->b_bufsize += size;
471         }
472
473         /*
474          * Fully valid pages in the cluster are already good and do not need
475          * to be re-read from disk.  Replace the page with bogus_page
476          */
477         for (j = 0; j < bp->b_xio.xio_npages; j++) {
478                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
479                     VM_PAGE_BITS_ALL) {
480                         bp->b_xio.xio_pages[j] = bogus_page;
481                 }
482         }
483         if (bp->b_bufsize > bp->b_kvasize) {
484                 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
485                     bp->b_bufsize, bp->b_kvasize);
486         }
487
488         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
489                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
490         return (bp);
491 }
492
493 /*
494  * Cleanup after a clustered read or write.
495  * This is complicated by the fact that any of the buffers might have
496  * extra memory (if there were no empty buffer headers at allocbuf time)
497  * that we will need to shift around.
498  *
499  * The returned bio is &bp->b_bio1
500  */
501 void
502 cluster_callback(struct bio *bio)
503 {
504         struct buf *bp = bio->bio_buf;
505         struct buf *tbp;
506         int error = 0;
507
508         /*
509          * Must propogate errors to all the components.  A short read (EOF)
510          * is a critical error.
511          */
512         if (bp->b_flags & B_ERROR) {
513                 error = bp->b_error;
514         } else if (bp->b_bcount != bp->b_bufsize) {
515                 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
516         }
517
518         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
519         /*
520          * Move memory from the large cluster buffer into the component
521          * buffers and mark IO as done on these.  Since the memory map
522          * is the same, no actual copying is required.
523          */
524         while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
525                 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
526                 if (error) {
527                         tbp->b_flags |= B_ERROR;
528                         tbp->b_error = error;
529                 } else {
530                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
531                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
532                         /*
533                          * XXX the bdwrite()/bqrelse() issued during
534                          * cluster building clears B_RELBUF (see bqrelse()
535                          * comment).  If direct I/O was specified, we have
536                          * to restore it here to allow the buffer and VM
537                          * to be freed.
538                          */
539                         if (tbp->b_flags & B_DIRECT)
540                                 tbp->b_flags |= B_RELBUF;
541                 }
542                 biodone(&tbp->b_bio1);
543         }
544         relpbuf(bp, &cluster_pbuf_freecnt);
545 }
546
547 /*
548  *      cluster_wbuild_wb:
549  *
550  *      Implement modified write build for cluster.
551  *
552  *              write_behind = 0        write behind disabled
553  *              write_behind = 1        write behind normal (default)
554  *              write_behind = 2        write behind backed-off
555  */
556
557 static __inline int
558 cluster_wbuild_wb(struct vnode *vp, int size, off_t start_loffset, int len)
559 {
560         int r = 0;
561
562         switch(write_behind) {
563         case 2:
564                 if (start_loffset < len)
565                         break;
566                 start_loffset -= len;
567                 /* fall through */
568         case 1:
569                 r = cluster_wbuild(vp, size, start_loffset, len);
570                 /* fall through */
571         default:
572                 /* fall through */
573                 break;
574         }
575         return(r);
576 }
577
578 /*
579  * Do clustered write for FFS.
580  *
581  * Three cases:
582  *      1. Write is not sequential (write asynchronously)
583  *      Write is sequential:
584  *      2.      beginning of cluster - begin cluster
585  *      3.      middle of a cluster - add to cluster
586  *      4.      end of a cluster - asynchronously write cluster
587  */
588 void
589 cluster_write(struct buf *bp, off_t filesize, int seqcount)
590 {
591         struct vnode *vp;
592         off_t loffset;
593         int maxclen, cursize;
594         int lblocksize;
595         int async;
596
597         vp = bp->b_vp;
598         if (vp->v_type == VREG) {
599                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
600                 lblocksize = vp->v_mount->mnt_stat.f_iosize;
601         } else {
602                 async = 0;
603                 lblocksize = bp->b_bufsize;
604         }
605         loffset = bp->b_loffset;
606         KASSERT(bp->b_loffset != NOOFFSET, 
607                 ("cluster_write: no buffer offset"));
608
609         /* Initialize vnode to beginning of file. */
610         if (loffset == 0)
611                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
612
613         if (vp->v_clen == 0 || loffset != vp->v_lastw + lblocksize ||
614             bp->b_bio2.bio_offset == NOOFFSET ||
615             (bp->b_bio2.bio_offset != vp->v_lasta + lblocksize)) {
616                 maxclen = vp->v_mount->mnt_iosize_max;
617                 if (vp->v_clen != 0) {
618                         /*
619                          * Next block is not sequential.
620                          *
621                          * If we are not writing at end of file, the process
622                          * seeked to another point in the file since its last
623                          * write, or we have reached our maximum cluster size,
624                          * then push the previous cluster. Otherwise try
625                          * reallocating to make it sequential.
626                          *
627                          * Change to algorithm: only push previous cluster if
628                          * it was sequential from the point of view of the
629                          * seqcount heuristic, otherwise leave the buffer 
630                          * intact so we can potentially optimize the I/O
631                          * later on in the buf_daemon or update daemon
632                          * flush.
633                          */
634                         cursize = vp->v_lastw - vp->v_cstart + lblocksize;
635                         if (bp->b_loffset + lblocksize != filesize ||
636                             loffset != vp->v_lastw + lblocksize || vp->v_clen <= cursize) {
637                                 if (!async && seqcount > 0) {
638                                         cluster_wbuild_wb(vp, lblocksize,
639                                                 vp->v_cstart, cursize);
640                                 }
641                         } else {
642                                 struct buf **bpp, **endbp;
643                                 struct cluster_save *buflist;
644
645                                 buflist = cluster_collectbufs(vp, bp,
646                                                               lblocksize);
647                                 endbp = &buflist->bs_children
648                                     [buflist->bs_nchildren - 1];
649                                 if (VOP_REALLOCBLKS(vp, buflist)) {
650                                         /*
651                                          * Failed, push the previous cluster
652                                          * if *really* writing sequentially
653                                          * in the logical file (seqcount > 1),
654                                          * otherwise delay it in the hopes that
655                                          * the low level disk driver can
656                                          * optimize the write ordering.
657                                          */
658                                         for (bpp = buflist->bs_children;
659                                              bpp < endbp; bpp++)
660                                                 brelse(*bpp);
661                                         kfree(buflist, M_SEGMENT);
662                                         if (seqcount > 1) {
663                                                 cluster_wbuild_wb(vp, 
664                                                     lblocksize, vp->v_cstart, 
665                                                     cursize);
666                                         }
667                                 } else {
668                                         /*
669                                          * Succeeded, keep building cluster.
670                                          */
671                                         for (bpp = buflist->bs_children;
672                                              bpp <= endbp; bpp++)
673                                                 bdwrite(*bpp);
674                                         kfree(buflist, M_SEGMENT);
675                                         vp->v_lastw = loffset;
676                                         vp->v_lasta = bp->b_bio2.bio_offset;
677                                         return;
678                                 }
679                         }
680                 }
681                 /*
682                  * Consider beginning a cluster. If at end of file, make
683                  * cluster as large as possible, otherwise find size of
684                  * existing cluster.
685                  */
686                 if ((vp->v_type == VREG) &&
687                     bp->b_loffset + lblocksize != filesize &&
688                     (bp->b_bio2.bio_offset == NOOFFSET) &&
689                     (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL) ||
690                      bp->b_bio2.bio_offset == NOOFFSET)) {
691                         bawrite(bp);
692                         vp->v_clen = 0;
693                         vp->v_lasta = bp->b_bio2.bio_offset;
694                         vp->v_cstart = loffset + lblocksize;
695                         vp->v_lastw = loffset;
696                         return;
697                 }
698                 if (maxclen > lblocksize)
699                         vp->v_clen = maxclen - lblocksize;
700                 else
701                         vp->v_clen = 0;
702                 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
703                         vp->v_cstart = loffset + lblocksize;
704                         bawrite(bp);
705                 } else {        /* Wait for rest of cluster */
706                         vp->v_cstart = loffset;
707                         bdwrite(bp);
708                 }
709         } else if (loffset == vp->v_cstart + vp->v_clen) {
710                 /*
711                  * At end of cluster, write it out if seqcount tells us we
712                  * are operating sequentially, otherwise let the buf or
713                  * update daemon handle it.
714                  */
715                 bdwrite(bp);
716                 if (seqcount > 1)
717                         cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
718                                           vp->v_clen + lblocksize);
719                 vp->v_clen = 0;
720                 vp->v_cstart = loffset + lblocksize;
721         } else if (vm_page_count_severe()) {
722                 /*
723                  * We are low on memory, get it going NOW
724                  */
725                 bawrite(bp);
726         } else {
727                 /*
728                  * In the middle of a cluster, so just delay the I/O for now.
729                  */
730                 bdwrite(bp);
731         }
732         vp->v_lastw = loffset;
733         vp->v_lasta = bp->b_bio2.bio_offset;
734 }
735
736
737 /*
738  * This is an awful lot like cluster_rbuild...wish they could be combined.
739  * The last lbn argument is the current block on which I/O is being
740  * performed.  Check to see that it doesn't fall in the middle of
741  * the current block (if last_bp == NULL).
742  */
743 int
744 cluster_wbuild(struct vnode *vp, int size, off_t start_loffset, int bytes)
745 {
746         struct buf *bp, *tbp;
747         int i, j;
748         int totalwritten = 0;
749
750         while (bytes > 0) {
751                 crit_enter();
752                 /*
753                  * If the buffer is not delayed-write (i.e. dirty), or it 
754                  * is delayed-write but either locked or inval, it cannot 
755                  * partake in the clustered write.
756                  */
757                 if (((tbp = findblk(vp, start_loffset)) == NULL) ||
758                   ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
759                   (LIST_FIRST(&tbp->b_dep) != NULL && buf_checkwrite(tbp)) ||
760                   BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
761                         start_loffset += size;
762                         bytes -= size;
763                         crit_exit();
764                         continue;
765                 }
766                 bremfree(tbp);
767                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
768                 crit_exit();
769
770                 /*
771                  * Extra memory in the buffer, punt on this buffer.
772                  * XXX we could handle this in most cases, but we would
773                  * have to push the extra memory down to after our max
774                  * possible cluster size and then potentially pull it back
775                  * up if the cluster was terminated prematurely--too much
776                  * hassle.
777                  */
778                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
779                   (tbp->b_bcount != tbp->b_bufsize) ||
780                   (tbp->b_bcount != size) ||
781                   (bytes == size) ||
782                   ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
783                         totalwritten += tbp->b_bufsize;
784                         bawrite(tbp);
785                         start_loffset += size;
786                         bytes -= size;
787                         continue;
788                 }
789
790                 /*
791                  * Set up the pbuf.  Track our append point with b_bcount
792                  * and b_bufsize.  b_bufsize is not used by the device but
793                  * our caller uses it to loop clusters and we use it to
794                  * detect a premature EOF on the block device.
795                  */
796                 bp->b_bcount = 0;
797                 bp->b_bufsize = 0;
798                 bp->b_xio.xio_npages = 0;
799                 bp->b_loffset = tbp->b_loffset;
800                 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
801
802                 /*
803                  * We are synthesizing a buffer out of vm_page_t's, but
804                  * if the block size is not page aligned then the starting
805                  * address may not be either.  Inherit the b_data offset
806                  * from the original buffer.
807                  */
808                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
809                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
810                 bp->b_flags &= ~B_ERROR;
811                 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
812                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
813                 bp->b_bio1.bio_done = cluster_callback;
814                 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
815                 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
816                 /*
817                  * From this location in the file, scan forward to see
818                  * if there are buffers with adjacent data that need to
819                  * be written as well.
820                  */
821                 for (i = 0; i < bytes; (i += size), (start_loffset += size)) {
822                         if (i != 0) { /* If not the first buffer */
823                                 crit_enter();
824                                 /*
825                                  * If the adjacent data is not even in core it
826                                  * can't need to be written.
827                                  */
828                                 if ((tbp = findblk(vp, start_loffset)) == NULL) {
829                                         crit_exit();
830                                         break;
831                                 }
832
833                                 /*
834                                  * If it IS in core, but has different
835                                  * characteristics, or is locked (which
836                                  * means it could be undergoing a background
837                                  * I/O or be in a weird state), then don't
838                                  * cluster with it.
839                                  */
840                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
841                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
842                                   != (B_DELWRI | B_CLUSTEROK |
843                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
844                                     (tbp->b_flags & B_LOCKED) ||
845                   (LIST_FIRST(&tbp->b_dep) != NULL && buf_checkwrite(tbp)) ||
846                                     BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
847                                         crit_exit();
848                                         break;
849                                 }
850
851                                 /*
852                                  * Check that the combined cluster
853                                  * would make sense with regard to pages
854                                  * and would not be too large
855                                  */
856                                 if ((tbp->b_bcount != size) ||
857                                   ((bp->b_bio2.bio_offset + i) !=
858                                     tbp->b_bio2.bio_offset) ||
859                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
860                                     (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
861                                         BUF_UNLOCK(tbp);
862                                         crit_exit();
863                                         break;
864                                 }
865                                 /*
866                                  * Ok, it's passed all the tests,
867                                  * so remove it from the free list
868                                  * and mark it busy. We will use it.
869                                  */
870                                 bremfree(tbp);
871                                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
872                                 crit_exit();
873                         } /* end of code for non-first buffers only */
874
875                         /*
876                          * If the IO is via the VM then we do some
877                          * special VM hackery (yuck).  Since the buffer's
878                          * block size may not be page-aligned it is possible
879                          * for a page to be shared between two buffers.  We
880                          * have to get rid of the duplication when building
881                          * the cluster.
882                          */
883                         if (tbp->b_flags & B_VMIO) {
884                                 vm_page_t m;
885
886                                 if (i != 0) { /* if not first buffer */
887                                         for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
888                                                 m = tbp->b_xio.xio_pages[j];
889                                                 if (m->flags & PG_BUSY) {
890                                                         bqrelse(tbp);
891                                                         goto finishcluster;
892                                                 }
893                                         }
894                                 }
895                                         
896                                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
897                                         m = tbp->b_xio.xio_pages[j];
898                                         vm_page_io_start(m);
899                                         vm_object_pip_add(m->object, 1);
900                                         if ((bp->b_xio.xio_npages == 0) ||
901                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
902                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
903                                                 bp->b_xio.xio_npages++;
904                                         }
905                                 }
906                         }
907                         bp->b_bcount += size;
908                         bp->b_bufsize += size;
909
910                         crit_enter();
911                         bundirty(tbp);
912                         tbp->b_flags &= ~B_ERROR;
913                         tbp->b_flags |= B_ASYNC;
914                         tbp->b_cmd = BUF_CMD_WRITE;
915                         crit_exit();
916                         BUF_KERNPROC(tbp);
917                         cluster_append(&bp->b_bio1, tbp);
918
919                         /*
920                          * check for latent dependencies to be handled 
921                          */
922                         if (LIST_FIRST(&tbp->b_dep) != NULL)
923                                 buf_start(tbp);
924                 }
925         finishcluster:
926                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
927                         (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
928                 if (bp->b_bufsize > bp->b_kvasize) {
929                         panic(
930                             "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
931                             bp->b_bufsize, bp->b_kvasize);
932                 }
933                 totalwritten += bp->b_bufsize;
934                 bp->b_dirtyoff = 0;
935                 bp->b_dirtyend = bp->b_bufsize;
936                 bp->b_flags |= B_ASYNC;
937                 bp->b_cmd = BUF_CMD_WRITE;
938                 vfs_busy_pages(vp, bp);
939                 bp->b_runningbufspace = bp->b_bufsize;
940                 if (bp->b_runningbufspace) {
941                         runningbufspace += bp->b_runningbufspace;
942                         ++runningbufcount;
943                 }
944                 BUF_KERNPROC(bp);       /* B_ASYNC */
945                 vn_strategy(vp, &bp->b_bio1);
946
947                 bytes -= i;
948         }
949         return totalwritten;
950 }
951
952 /*
953  * Collect together all the buffers in a cluster.
954  * Plus add one additional buffer.
955  */
956 static struct cluster_save *
957 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int lblocksize)
958 {
959         struct cluster_save *buflist;
960         struct buf *bp;
961         off_t loffset;
962         int i, len;
963
964         len = (int)(vp->v_lastw - vp->v_cstart + lblocksize) / lblocksize;
965         buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
966                          M_SEGMENT, M_WAITOK);
967         buflist->bs_nchildren = 0;
968         buflist->bs_children = (struct buf **) (buflist + 1);
969         for (loffset = vp->v_cstart, i = 0; i < len; (loffset += lblocksize), i++) {
970                 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
971                 buflist->bs_children[i] = bp;
972                 if (bp->b_bio2.bio_offset == NOOFFSET) {
973                         VOP_BMAP(bp->b_vp, bp->b_loffset,
974                                  &bp->b_bio2.bio_offset, NULL, NULL);
975                 }
976         }
977         buflist->bs_children[i] = bp = last_bp;
978         if (bp->b_bio2.bio_offset == NOOFFSET) {
979                 VOP_BMAP(bp->b_vp, bp->b_loffset,
980                          &bp->b_bio2.bio_offset, NULL, NULL);
981         }
982         buflist->bs_nchildren = i + 1;
983         return (buflist);
984 }
985
986 void
987 cluster_append(struct bio *bio, struct buf *tbp)
988 {
989         tbp->b_cluster_next = NULL;
990         if (bio->bio_caller_info1.cluster_head == NULL) {
991                 bio->bio_caller_info1.cluster_head = tbp;
992                 bio->bio_caller_info2.cluster_tail = tbp;
993         } else {
994                 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
995                 bio->bio_caller_info2.cluster_tail = tbp;
996         }
997 }
998