31dbe66243ed02647f930c0a770cc2a81dd5b5ec
[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 1.40 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 #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 blksize);
70 static struct buf *
71         cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
72                             off_t doffset, int blksize, int run, 
73                             struct buf *fbp);
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 blksize, 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 = vmaxiosize(vp) / blksize;
110         maxra = 2 * racluster + (totread / blksize);
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, blksize, 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                         /*
137                          * Set read-ahead-mark only if we can passively lock
138                          * the buffer.  Note that with these flags the bp
139                          * could very exist even though NULL is returned.
140                          */
141                         for (i = 1; i < maxra; i++) {
142                                 tbp = findblk(vp, loffset + i * blksize,
143                                               FINDBLK_NBLOCK);
144                                 if (tbp == NULL)
145                                         break;
146                                 if (((i % racluster) == (racluster - 1)) ||
147                                     (i == (maxra - 1))) {
148                                         tbp->b_flags |= B_RAM;
149                                 }
150                                 BUF_UNLOCK(tbp);
151                         }
152                         if (i >= maxra)
153                                 return 0;
154                         loffset += i * blksize;
155                 }
156                 reqbp = bp = NULL;
157         } else {
158                 off_t firstread = bp->b_loffset;
159                 int nblks;
160
161                 /*
162                  * Set-up synchronous read for bp.
163                  */
164                 bp->b_cmd = BUF_CMD_READ;
165                 bp->b_bio1.bio_done = biodone_sync;
166                 bp->b_bio1.bio_flags |= BIO_SYNC;
167
168                 KASSERT(firstread != NOOFFSET, 
169                         ("cluster_read: no buffer offset"));
170                 if (firstread + totread > filesize)
171                         totread = (int)(filesize - firstread);
172                 nblks = totread / blksize;
173                 if (nblks) {
174                         int burstbytes;
175
176                         if (nblks > racluster)
177                                 nblks = racluster;
178
179                         error = VOP_BMAP(vp, loffset, &doffset,
180                                          &burstbytes, NULL, BUF_CMD_READ);
181                         if (error)
182                                 goto single_block_read;
183                         if (doffset == NOOFFSET)
184                                 goto single_block_read;
185                         if (burstbytes < blksize * 2)
186                                 goto single_block_read;
187                         if (nblks > burstbytes / blksize)
188                                 nblks = burstbytes / blksize;
189
190                         bp = cluster_rbuild(vp, filesize, loffset,
191                                             doffset, blksize, nblks, bp);
192                         loffset += bp->b_bufsize;
193                 } else {
194 single_block_read:
195                         /*
196                          * if it isn't in the cache, then get a chunk from
197                          * disk if sequential, otherwise just get the block.
198                          */
199                         bp->b_flags |= B_RAM;
200                         loffset += blksize;
201                 }
202         }
203
204         /*
205          * If B_CACHE was not set issue bp.  bp will either be an
206          * asynchronous cluster buf or a synchronous single-buf.
207          * If it is a single buf it will be the same as reqbp.
208          *
209          * NOTE: Once an async cluster buf is issued bp becomes invalid.
210          */
211         if (bp) {
212 #if defined(CLUSTERDEBUG)
213                 if (rcluster)
214                         kprintf("S(%lld,%d,%d) ",
215                             bp->b_loffset, bp->b_bcount, seqcount);
216 #endif
217                 if ((bp->b_flags & B_CLUSTER) == 0)
218                         vfs_busy_pages(vp, bp);
219                 bp->b_flags &= ~(B_ERROR|B_INVAL);
220                 vn_strategy(vp, &bp->b_bio1);
221                 error = 0;
222                 /* bp invalid now */
223         }
224
225         /*
226          * If we have been doing sequential I/O, then do some read-ahead.
227          *
228          * Only mess with buffers which we can immediately lock.  HAMMER
229          * will do device-readahead irrespective of what the blocks
230          * represent.
231          */
232         rbp = NULL;
233         if (!error &&
234             seqcount &&
235             loffset < origoffset + seqcount * blksize &&
236             loffset + blksize <= filesize
237         ) {
238                 int nblksread;
239                 int ntoread;
240                 int burstbytes;
241                 int tmp_error;
242
243                 rbp = getblk(vp, loffset, blksize,
244                              GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
245                 if (rbp == NULL)
246                         goto no_read_ahead;
247                 if ((rbp->b_flags & B_CACHE)) {
248                         bqrelse(rbp);
249                         goto no_read_ahead;
250                 }
251
252                 /*
253                  * An error from the read-ahead bmap has nothing to do
254                  * with the caller's original request.
255                  */
256                 tmp_error = VOP_BMAP(vp, loffset, &doffset,
257                                      &burstbytes, NULL, BUF_CMD_READ);
258                 if (tmp_error || doffset == NOOFFSET) {
259                         rbp->b_flags |= B_INVAL;
260                         brelse(rbp);
261                         rbp = NULL;
262                         goto no_read_ahead;
263                 }
264                 ntoread = burstbytes / blksize;
265                 nblksread = (totread + blksize - 1) / blksize;
266                 if (seqcount < nblksread)
267                         seqcount = nblksread;
268                 if (ntoread > seqcount)
269                         ntoread = seqcount;
270
271                 /*
272                  * rbp: async read
273                  */
274                 rbp->b_cmd = BUF_CMD_READ;
275                 rbp->b_flags |= B_RAM/* | B_AGE*/;
276
277                 if (burstbytes) {
278                         rbp = cluster_rbuild(vp, filesize, loffset,
279                                              doffset, blksize, 
280                                              ntoread, rbp);
281                 } else {
282                         rbp->b_bio2.bio_offset = doffset;
283                 }
284 #if defined(CLUSTERDEBUG)
285                 if (rcluster) {
286                         if (bp)
287                                 kprintf("A+(%lld,%d,%lld,%d) ",
288                                     rbp->b_loffset, rbp->b_bcount,
289                                     rbp->b_loffset - origoffset,
290                                     seqcount);
291                         else
292                                 kprintf("A(%lld,%d,%lld,%d) ",
293                                     rbp->b_loffset, rbp->b_bcount,
294                                     rbp->b_loffset - origoffset,
295                                     seqcount);
296                 }
297 #endif
298                 rbp->b_flags &= ~(B_ERROR|B_INVAL);
299
300                 if ((rbp->b_flags & B_CLUSTER) == 0)
301                         vfs_busy_pages(vp, rbp);
302                 BUF_KERNPROC(rbp);
303                 vn_strategy(vp, &rbp->b_bio1);
304                 /* rbp invalid now */
305         }
306
307         /*
308          * Wait for our original buffer to complete its I/O.  reqbp will
309          * be NULL if the original buffer was B_CACHE.  We are returning
310          * (*bpp) which is the same as reqbp when reqbp != NULL.
311          */
312 no_read_ahead:
313         if (reqbp) {
314                 KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
315                 error = biowait(&reqbp->b_bio1, "clurd");
316         }
317         return (error);
318 }
319
320 /*
321  * If blocks are contiguous on disk, use this to provide clustered
322  * read ahead.  We will read as many blocks as possible sequentially
323  * and then parcel them up into logical blocks in the buffer hash table.
324  *
325  * This function either returns a cluster buf or it returns fbp.  fbp is
326  * already expected to be set up as a synchronous or asynchronous request.
327  *
328  * If a cluster buf is returned it will always be async.
329  */
330 static struct buf *
331 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
332                int blksize, int run, struct buf *fbp)
333 {
334         struct buf *bp, *tbp;
335         off_t boffset;
336         int i, j;
337         int maxiosize = vmaxiosize(vp);
338
339         /*
340          * avoid a division
341          */
342         while (loffset + run * blksize > filesize) {
343                 --run;
344         }
345
346         tbp = fbp;
347         tbp->b_bio2.bio_offset = doffset;
348         if((tbp->b_flags & B_MALLOC) ||
349             ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
350                 return tbp;
351         }
352
353         bp = trypbuf(&cluster_pbuf_freecnt);
354         if (bp == NULL) {
355                 return tbp;
356         }
357
358         /*
359          * We are synthesizing a buffer out of vm_page_t's, but
360          * if the block size is not page aligned then the starting
361          * address may not be either.  Inherit the b_data offset
362          * from the original buffer.
363          */
364         bp->b_data = (char *)((vm_offset_t)bp->b_data |
365             ((vm_offset_t)tbp->b_data & PAGE_MASK));
366         bp->b_flags |= B_CLUSTER | B_VMIO;
367         bp->b_cmd = BUF_CMD_READ;
368         bp->b_bio1.bio_done = cluster_callback;         /* default to async */
369         bp->b_bio1.bio_caller_info1.cluster_head = NULL;
370         bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
371         bp->b_loffset = loffset;
372         bp->b_bio2.bio_offset = doffset;
373         KASSERT(bp->b_loffset != NOOFFSET,
374                 ("cluster_rbuild: no buffer offset"));
375
376         bp->b_bcount = 0;
377         bp->b_bufsize = 0;
378         bp->b_xio.xio_npages = 0;
379
380         for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
381                 if (i) {
382                         if ((bp->b_xio.xio_npages * PAGE_SIZE) +
383                             round_page(blksize) > maxiosize) {
384                                 break;
385                         }
386
387                         /*
388                          * Shortcut some checks and try to avoid buffers that
389                          * would block in the lock.  The same checks have to
390                          * be made again after we officially get the buffer.
391                          */
392                         tbp = getblk(vp, loffset + i * blksize, blksize,
393                                      GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
394                         if (tbp == NULL)
395                                 break;
396                         for (j = 0; j < tbp->b_xio.xio_npages; j++) {
397                                 if (tbp->b_xio.xio_pages[j]->valid)
398                                         break;
399                         }
400                         if (j != tbp->b_xio.xio_npages) {
401                                 bqrelse(tbp);
402                                 break;
403                         }
404
405                         /*
406                          * Stop scanning if the buffer is fuly valid 
407                          * (marked B_CACHE), or locked (may be doing a
408                          * background write), or if the buffer is not
409                          * VMIO backed.  The clustering code can only deal
410                          * with VMIO-backed buffers.
411                          */
412                         if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
413                             (tbp->b_flags & B_VMIO) == 0 ||
414                             (LIST_FIRST(&tbp->b_dep) != NULL &&
415                              buf_checkread(tbp))
416                         ) {
417                                 bqrelse(tbp);
418                                 break;
419                         }
420
421                         /*
422                          * The buffer must be completely invalid in order to
423                          * take part in the cluster.  If it is partially valid
424                          * then we stop.
425                          */
426                         for (j = 0;j < tbp->b_xio.xio_npages; j++) {
427                                 if (tbp->b_xio.xio_pages[j]->valid)
428                                         break;
429                         }
430                         if (j != tbp->b_xio.xio_npages) {
431                                 bqrelse(tbp);
432                                 break;
433                         }
434
435                         /*
436                          * Set a read-ahead mark as appropriate
437                          */
438                         if (i == 1 || i == (run - 1))
439                                 tbp->b_flags |= B_RAM;
440
441                         /*
442                          * Depress the priority of buffers not explicitly
443                          * requested.
444                          */
445                         /* tbp->b_flags |= B_AGE; */
446
447                         /*
448                          * Set the block number if it isn't set, otherwise
449                          * if it is make sure it matches the block number we
450                          * expect.
451                          */
452                         if (tbp->b_bio2.bio_offset == NOOFFSET) {
453                                 tbp->b_bio2.bio_offset = boffset;
454                         } else if (tbp->b_bio2.bio_offset != boffset) {
455                                 brelse(tbp);
456                                 break;
457                         }
458                 }
459
460                 /*
461                  * The passed-in tbp (i == 0) will already be set up for
462                  * async or sync operation.  All other tbp's acquire in
463                  * our loop are set up for async operation.
464                  */
465                 tbp->b_cmd = BUF_CMD_READ;
466                 BUF_KERNPROC(tbp);
467                 cluster_append(&bp->b_bio1, tbp);
468                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
469                         vm_page_t m;
470                         m = tbp->b_xio.xio_pages[j];
471                         vm_page_io_start(m);
472                         vm_object_pip_add(m->object, 1);
473                         if ((bp->b_xio.xio_npages == 0) ||
474                                 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
475                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
476                                 bp->b_xio.xio_npages++;
477                         }
478                         if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
479                                 tbp->b_xio.xio_pages[j] = bogus_page;
480                 }
481                 /*
482                  * XXX shouldn't this be += size for both, like in 
483                  * cluster_wbuild()?
484                  *
485                  * Don't inherit tbp->b_bufsize as it may be larger due to
486                  * a non-page-aligned size.  Instead just aggregate using
487                  * 'size'.
488                  */
489                 if (tbp->b_bcount != blksize)
490                     kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
491                 if (tbp->b_bufsize != blksize)
492                     kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
493                 bp->b_bcount += blksize;
494                 bp->b_bufsize += blksize;
495         }
496
497         /*
498          * Fully valid pages in the cluster are already good and do not need
499          * to be re-read from disk.  Replace the page with bogus_page
500          */
501         for (j = 0; j < bp->b_xio.xio_npages; j++) {
502                 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
503                     VM_PAGE_BITS_ALL) {
504                         bp->b_xio.xio_pages[j] = bogus_page;
505                 }
506         }
507         if (bp->b_bufsize > bp->b_kvasize) {
508                 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
509                     bp->b_bufsize, bp->b_kvasize);
510         }
511         pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
512                 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
513         BUF_KERNPROC(bp);
514         return (bp);
515 }
516
517 /*
518  * Cleanup after a clustered read or write.
519  * This is complicated by the fact that any of the buffers might have
520  * extra memory (if there were no empty buffer headers at allocbuf time)
521  * that we will need to shift around.
522  *
523  * The returned bio is &bp->b_bio1
524  */
525 void
526 cluster_callback(struct bio *bio)
527 {
528         struct buf *bp = bio->bio_buf;
529         struct buf *tbp;
530         int error = 0;
531
532         /*
533          * Must propogate errors to all the components.  A short read (EOF)
534          * is a critical error.
535          */
536         if (bp->b_flags & B_ERROR) {
537                 error = bp->b_error;
538         } else if (bp->b_bcount != bp->b_bufsize) {
539                 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
540         }
541
542         pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
543         /*
544          * Move memory from the large cluster buffer into the component
545          * buffers and mark IO as done on these.  Since the memory map
546          * is the same, no actual copying is required.
547          */
548         while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
549                 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
550                 if (error) {
551                         tbp->b_flags |= B_ERROR;
552                         tbp->b_error = error;
553                 } else {
554                         tbp->b_dirtyoff = tbp->b_dirtyend = 0;
555                         tbp->b_flags &= ~(B_ERROR|B_INVAL);
556                         /*
557                          * XXX the bdwrite()/bqrelse() issued during
558                          * cluster building clears B_RELBUF (see bqrelse()
559                          * comment).  If direct I/O was specified, we have
560                          * to restore it here to allow the buffer and VM
561                          * to be freed.
562                          */
563                         if (tbp->b_flags & B_DIRECT)
564                                 tbp->b_flags |= B_RELBUF;
565                 }
566                 biodone(&tbp->b_bio1);
567         }
568         relpbuf(bp, &cluster_pbuf_freecnt);
569 }
570
571 /*
572  *      cluster_wbuild_wb:
573  *
574  *      Implement modified write build for cluster.
575  *
576  *              write_behind = 0        write behind disabled
577  *              write_behind = 1        write behind normal (default)
578  *              write_behind = 2        write behind backed-off
579  */
580
581 static __inline int
582 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
583 {
584         int r = 0;
585
586         switch(write_behind) {
587         case 2:
588                 if (start_loffset < len)
589                         break;
590                 start_loffset -= len;
591                 /* fall through */
592         case 1:
593                 r = cluster_wbuild(vp, blksize, start_loffset, len);
594                 /* fall through */
595         default:
596                 /* fall through */
597                 break;
598         }
599         return(r);
600 }
601
602 /*
603  * Do clustered write for FFS.
604  *
605  * Three cases:
606  *      1. Write is not sequential (write asynchronously)
607  *      Write is sequential:
608  *      2.      beginning of cluster - begin cluster
609  *      3.      middle of a cluster - add to cluster
610  *      4.      end of a cluster - asynchronously write cluster
611  */
612 void
613 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
614 {
615         struct vnode *vp;
616         off_t loffset;
617         int maxclen, cursize;
618         int async;
619
620         vp = bp->b_vp;
621         if (vp->v_type == VREG)
622                 async = vp->v_mount->mnt_flag & MNT_ASYNC;
623         else
624                 async = 0;
625         loffset = bp->b_loffset;
626         KASSERT(bp->b_loffset != NOOFFSET, 
627                 ("cluster_write: no buffer offset"));
628
629         /* Initialize vnode to beginning of file. */
630         if (loffset == 0)
631                 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
632
633         if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
634             bp->b_bio2.bio_offset == NOOFFSET ||
635             (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
636                 maxclen = vmaxiosize(vp);
637                 if (vp->v_clen != 0) {
638                         /*
639                          * Next block is not sequential.
640                          *
641                          * If we are not writing at end of file, the process
642                          * seeked to another point in the file since its last
643                          * write, or we have reached our maximum cluster size,
644                          * then push the previous cluster. Otherwise try
645                          * reallocating to make it sequential.
646                          *
647                          * Change to algorithm: only push previous cluster if
648                          * it was sequential from the point of view of the
649                          * seqcount heuristic, otherwise leave the buffer 
650                          * intact so we can potentially optimize the I/O
651                          * later on in the buf_daemon or update daemon
652                          * flush.
653                          */
654                         cursize = vp->v_lastw - vp->v_cstart + blksize;
655                         if (bp->b_loffset + blksize != filesize ||
656                             loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
657                                 if (!async && seqcount > 0) {
658                                         cluster_wbuild_wb(vp, blksize,
659                                                 vp->v_cstart, cursize);
660                                 }
661                         } else {
662                                 struct buf **bpp, **endbp;
663                                 struct cluster_save *buflist;
664
665                                 buflist = cluster_collectbufs(vp, bp, blksize);
666                                 endbp = &buflist->bs_children
667                                     [buflist->bs_nchildren - 1];
668                                 if (VOP_REALLOCBLKS(vp, buflist)) {
669                                         /*
670                                          * Failed, push the previous cluster
671                                          * if *really* writing sequentially
672                                          * in the logical file (seqcount > 1),
673                                          * otherwise delay it in the hopes that
674                                          * the low level disk driver can
675                                          * optimize the write ordering.
676                                          */
677                                         for (bpp = buflist->bs_children;
678                                              bpp < endbp; bpp++)
679                                                 brelse(*bpp);
680                                         kfree(buflist, M_SEGMENT);
681                                         if (seqcount > 1) {
682                                                 cluster_wbuild_wb(vp, 
683                                                     blksize, vp->v_cstart, 
684                                                     cursize);
685                                         }
686                                 } else {
687                                         /*
688                                          * Succeeded, keep building cluster.
689                                          */
690                                         for (bpp = buflist->bs_children;
691                                              bpp <= endbp; bpp++)
692                                                 bdwrite(*bpp);
693                                         kfree(buflist, M_SEGMENT);
694                                         vp->v_lastw = loffset;
695                                         vp->v_lasta = bp->b_bio2.bio_offset;
696                                         return;
697                                 }
698                         }
699                 }
700                 /*
701                  * Consider beginning a cluster. If at end of file, make
702                  * cluster as large as possible, otherwise find size of
703                  * existing cluster.
704                  */
705                 if ((vp->v_type == VREG) &&
706                     bp->b_loffset + blksize != filesize &&
707                     (bp->b_bio2.bio_offset == NOOFFSET) &&
708                     (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
709                      bp->b_bio2.bio_offset == NOOFFSET)) {
710                         bawrite(bp);
711                         vp->v_clen = 0;
712                         vp->v_lasta = bp->b_bio2.bio_offset;
713                         vp->v_cstart = loffset + blksize;
714                         vp->v_lastw = loffset;
715                         return;
716                 }
717                 if (maxclen > blksize)
718                         vp->v_clen = maxclen - blksize;
719                 else
720                         vp->v_clen = 0;
721                 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
722                         vp->v_cstart = loffset + blksize;
723                         bawrite(bp);
724                 } else {        /* Wait for rest of cluster */
725                         vp->v_cstart = loffset;
726                         bdwrite(bp);
727                 }
728         } else if (loffset == vp->v_cstart + vp->v_clen) {
729                 /*
730                  * At end of cluster, write it out if seqcount tells us we
731                  * are operating sequentially, otherwise let the buf or
732                  * update daemon handle it.
733                  */
734                 bdwrite(bp);
735                 if (seqcount > 1)
736                         cluster_wbuild_wb(vp, blksize, vp->v_cstart,
737                                           vp->v_clen + blksize);
738                 vp->v_clen = 0;
739                 vp->v_cstart = loffset + blksize;
740         } else if (vm_page_count_severe()) {
741                 /*
742                  * We are low on memory, get it going NOW
743                  */
744                 bawrite(bp);
745         } else {
746                 /*
747                  * In the middle of a cluster, so just delay the I/O for now.
748                  */
749                 bdwrite(bp);
750         }
751         vp->v_lastw = loffset;
752         vp->v_lasta = bp->b_bio2.bio_offset;
753 }
754
755
756 /*
757  * This is an awful lot like cluster_rbuild...wish they could be combined.
758  * The last lbn argument is the current block on which I/O is being
759  * performed.  Check to see that it doesn't fall in the middle of
760  * the current block (if last_bp == NULL).
761  */
762 int
763 cluster_wbuild(struct vnode *vp, int blksize, off_t start_loffset, int bytes)
764 {
765         struct buf *bp, *tbp;
766         int i, j;
767         int totalwritten = 0;
768         int maxiosize = vmaxiosize(vp);
769
770         while (bytes > 0) {
771                 /*
772                  * If the buffer is not delayed-write (i.e. dirty), or it 
773                  * is delayed-write but either locked or inval, it cannot 
774                  * partake in the clustered write.
775                  */
776                 tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK);
777                 if (tbp == NULL ||
778                     (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI ||
779                     (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
780                         if (tbp)
781                                 BUF_UNLOCK(tbp);
782                         start_loffset += blksize;
783                         bytes -= blksize;
784                         continue;
785                 }
786                 bremfree(tbp);
787                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
788
789                 /*
790                  * Extra memory in the buffer, punt on this buffer.
791                  * XXX we could handle this in most cases, but we would
792                  * have to push the extra memory down to after our max
793                  * possible cluster size and then potentially pull it back
794                  * up if the cluster was terminated prematurely--too much
795                  * hassle.
796                  */
797                 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
798                     (tbp->b_bcount != tbp->b_bufsize) ||
799                     (tbp->b_bcount != blksize) ||
800                     (bytes == blksize) ||
801                     ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
802                         totalwritten += tbp->b_bufsize;
803                         bawrite(tbp);
804                         start_loffset += blksize;
805                         bytes -= blksize;
806                         continue;
807                 }
808
809                 /*
810                  * Set up the pbuf.  Track our append point with b_bcount
811                  * and b_bufsize.  b_bufsize is not used by the device but
812                  * our caller uses it to loop clusters and we use it to
813                  * detect a premature EOF on the block device.
814                  */
815                 bp->b_bcount = 0;
816                 bp->b_bufsize = 0;
817                 bp->b_xio.xio_npages = 0;
818                 bp->b_loffset = tbp->b_loffset;
819                 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
820
821                 /*
822                  * We are synthesizing a buffer out of vm_page_t's, but
823                  * if the block size is not page aligned then the starting
824                  * address may not be either.  Inherit the b_data offset
825                  * from the original buffer.
826                  */
827                 bp->b_data = (char *)((vm_offset_t)bp->b_data |
828                     ((vm_offset_t)tbp->b_data & PAGE_MASK));
829                 bp->b_flags &= ~B_ERROR;
830                 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
831                         (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
832                 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
833                 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
834
835                 /*
836                  * From this location in the file, scan forward to see
837                  * if there are buffers with adjacent data that need to
838                  * be written as well.
839                  */
840                 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
841                         if (i != 0) { /* If not the first buffer */
842                                 tbp = findblk(vp, start_loffset,
843                                               FINDBLK_NBLOCK);
844                                 /*
845                                  * Buffer not found or could not be locked
846                                  * non-blocking.
847                                  */
848                                 if (tbp == NULL)
849                                         break;
850
851                                 /*
852                                  * If it IS in core, but has different
853                                  * characteristics, then don't cluster
854                                  * with it.
855                                  */
856                                 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
857                                      B_INVAL | B_DELWRI | B_NEEDCOMMIT))
858                                     != (B_DELWRI | B_CLUSTEROK |
859                                      (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
860                                     (tbp->b_flags & B_LOCKED) ||
861                                     (LIST_FIRST(&tbp->b_dep) &&
862                                      buf_checkwrite(tbp))
863                                 ) {
864                                         BUF_UNLOCK(tbp);
865                                         break;
866                                 }
867
868                                 /*
869                                  * Check that the combined cluster
870                                  * would make sense with regard to pages
871                                  * and would not be too large
872                                  */
873                                 if ((tbp->b_bcount != blksize) ||
874                                   ((bp->b_bio2.bio_offset + i) !=
875                                     tbp->b_bio2.bio_offset) ||
876                                   ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
877                                     (maxiosize / PAGE_SIZE))) {
878                                         BUF_UNLOCK(tbp);
879                                         break;
880                                 }
881                                 /*
882                                  * Ok, it's passed all the tests,
883                                  * so remove it from the free list
884                                  * and mark it busy. We will use it.
885                                  */
886                                 bremfree(tbp);
887                                 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
888                         } /* end of code for non-first buffers only */
889
890                         /*
891                          * If the IO is via the VM then we do some
892                          * special VM hackery (yuck).  Since the buffer's
893                          * block size may not be page-aligned it is possible
894                          * for a page to be shared between two buffers.  We
895                          * have to get rid of the duplication when building
896                          * the cluster.
897                          */
898                         if (tbp->b_flags & B_VMIO) {
899                                 vm_page_t m;
900
901                                 if (i != 0) { /* if not first buffer */
902                                         for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
903                                                 m = tbp->b_xio.xio_pages[j];
904                                                 if (m->flags & PG_BUSY) {
905                                                         bqrelse(tbp);
906                                                         goto finishcluster;
907                                                 }
908                                         }
909                                 }
910                                         
911                                 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
912                                         m = tbp->b_xio.xio_pages[j];
913                                         vm_page_io_start(m);
914                                         vm_object_pip_add(m->object, 1);
915                                         if ((bp->b_xio.xio_npages == 0) ||
916                                           (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
917                                                 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
918                                                 bp->b_xio.xio_npages++;
919                                         }
920                                 }
921                         }
922                         bp->b_bcount += blksize;
923                         bp->b_bufsize += blksize;
924
925                         bundirty(tbp);
926                         tbp->b_flags &= ~B_ERROR;
927                         tbp->b_cmd = BUF_CMD_WRITE;
928                         BUF_KERNPROC(tbp);
929                         cluster_append(&bp->b_bio1, tbp);
930
931                         /*
932                          * check for latent dependencies to be handled 
933                          */
934                         if (LIST_FIRST(&tbp->b_dep) != NULL)
935                                 buf_start(tbp);
936                 }
937         finishcluster:
938                 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
939                         (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
940                 if (bp->b_bufsize > bp->b_kvasize) {
941                         panic(
942                             "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
943                             bp->b_bufsize, bp->b_kvasize);
944                 }
945                 totalwritten += bp->b_bufsize;
946                 bp->b_dirtyoff = 0;
947                 bp->b_dirtyend = bp->b_bufsize;
948                 bp->b_bio1.bio_done = cluster_callback;
949                 bp->b_cmd = BUF_CMD_WRITE;
950
951                 vfs_busy_pages(vp, bp);
952                 bp->b_runningbufspace = bp->b_bufsize;
953                 if (bp->b_runningbufspace) {
954                         runningbufspace += bp->b_runningbufspace;
955                         ++runningbufcount;
956                 }
957                 BUF_KERNPROC(bp);
958                 vn_strategy(vp, &bp->b_bio1);
959
960                 bytes -= i;
961         }
962         return totalwritten;
963 }
964
965 /*
966  * Collect together all the buffers in a cluster.
967  * Plus add one additional buffer.
968  */
969 static struct cluster_save *
970 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
971 {
972         struct cluster_save *buflist;
973         struct buf *bp;
974         off_t loffset;
975         int i, len;
976
977         len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
978         buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
979                          M_SEGMENT, M_WAITOK);
980         buflist->bs_nchildren = 0;
981         buflist->bs_children = (struct buf **) (buflist + 1);
982         for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
983                 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
984                 buflist->bs_children[i] = bp;
985                 if (bp->b_bio2.bio_offset == NOOFFSET) {
986                         VOP_BMAP(bp->b_vp, bp->b_loffset,
987                                  &bp->b_bio2.bio_offset,
988                                  NULL, NULL, BUF_CMD_WRITE);
989                 }
990         }
991         buflist->bs_children[i] = bp = last_bp;
992         if (bp->b_bio2.bio_offset == NOOFFSET) {
993                 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
994                          NULL, NULL, BUF_CMD_WRITE);
995         }
996         buflist->bs_nchildren = i + 1;
997         return (buflist);
998 }
999
1000 void
1001 cluster_append(struct bio *bio, struct buf *tbp)
1002 {
1003         tbp->b_cluster_next = NULL;
1004         if (bio->bio_caller_info1.cluster_head == NULL) {
1005                 bio->bio_caller_info1.cluster_head = tbp;
1006                 bio->bio_caller_info2.cluster_tail = tbp;
1007         } else {
1008                 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1009                 bio->bio_caller_info2.cluster_tail = tbp;
1010         }
1011 }
1012