?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Ignore machine-check MSRs
[freebsd-src/fkvm-freebsd.git] / sys / kern / vfs_cluster.c
bloba74c2729034045d05f8468631d27a8be92c7ae40
1 /*-
2 * Copyright (c) 1993
3 * The Regents of the University of California. All rights reserved.
4 * Modifications/enhancements:
5 * Copyright (c) 1995 John S. Dyson. All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 4. Neither the name of the University nor the names of its contributors
16 * may be used to endorse or promote products derived from this software
17 * without specific prior written permission.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * SUCH DAMAGE.
30 *
31 * @(#)vfs_cluster.c 8.7 (Berkeley) 2/13/94
32 */
33
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36
37 #include "opt_debug_cluster.h"
38
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/proc.h>
43 #include <sys/bio.h>
44 #include <sys/buf.h>
45 #include <sys/vnode.h>
46 #include <sys/malloc.h>
47 #include <sys/mount.h>
48 #include <sys/resourcevar.h>
49 #include <sys/vmmeter.h>
50 #include <vm/vm.h>
51 #include <vm/vm_object.h>
52 #include <vm/vm_page.h>
53 #include <sys/sysctl.h>
54
55 #if defined(CLUSTERDEBUG)
56 static int rcluster= 0;
57 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0,
58 "Debug VFS clustering code");
59 #endif
60
61 static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer");
62
63 static struct cluster_save *
64 cluster_collectbufs(struct vnode *vp, struct buf *last_bp);
65 static struct buf *
66 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
67 daddr_t blkno, long size, int run, struct buf *fbp);
68 static void cluster_callback(struct buf *);
69
70 static int write_behind = 1;
71 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
72 "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
73
74 static int read_max = 8;
75 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
76 "Cluster read-ahead max block count");
77
78 /* Page expended to mark partially backed buffers */
79 extern vm_page_t bogus_page;
80
81 /*
82 * Read data to a buf, including read-ahead if we find this to be beneficial.
83 * cluster_read replaces bread.
84 */
85 int
86 cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp)
87 struct vnode *vp;
88 u_quad_t filesize;
89 daddr_t lblkno;
90 long size;
91 struct ucred *cred;
92 long totread;
93 int seqcount;
94 struct buf **bpp;
95 {
96 struct buf *bp, *rbp, *reqbp;
97 struct bufobj *bo;
98 daddr_t blkno, origblkno;
99 int maxra, racluster;
100 int error, ncontig;
101 int i;
102
103 error = 0;
104 bo = &vp->v_bufobj;
105
106 /*
107 * Try to limit the amount of read-ahead by a few
108 * ad-hoc parameters. This needs work!!!
109 */
110 racluster = vp->v_mount->mnt_iosize_max / size;
111 maxra = seqcount;
112 maxra = min(read_max, maxra);
113 maxra = min(nbuf/8, maxra);
114 if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
115 maxra = (filesize / size) - lblkno;
116
117 /*
118 * get the requested block
119 */
120 *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, 0);
121 origblkno = lblkno;
122
123 /*
124 * if it is in the cache, then check to see if the reads have been
125 * sequential. If they have, then try some read-ahead, otherwise
126 * back-off on prospective read-aheads.
127 */
128 if (bp->b_flags & B_CACHE) {
129 if (!seqcount) {
130 return 0;
131 } else if ((bp->b_flags & B_RAM) == 0) {
132 return 0;
133 } else {
134 bp->b_flags &= ~B_RAM;
135 BO_LOCK(bo);
136 for (i = 1; i < maxra; i++) {
137 /*
138 * Stop if the buffer does not exist or it
139 * is invalid (about to go away?)
140 */
141 rbp = gbincore(&vp->v_bufobj, lblkno+i);
142 if (rbp == NULL || (rbp->b_flags & B_INVAL))
143 break;
144
145 /*
146 * Set another read-ahead mark so we know
147 * to check again. (If we can lock the
148 * buffer without waiting)
149 */
150 if ((((i % racluster) == (racluster - 1)) ||
151 (i == (maxra - 1)))
152 && (0 == BUF_LOCK(rbp,
153 LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
154 rbp->b_flags |= B_RAM;
155 BUF_UNLOCK(rbp);
156 }
157 }
158 BO_UNLOCK(bo);
159 if (i >= maxra) {
160 return 0;
161 }
162 lblkno += i;
163 }
164 reqbp = bp = NULL;
165 /*
166 * If it isn't in the cache, then get a chunk from
167 * disk if sequential, otherwise just get the block.
168 */
169 } else {
170 off_t firstread = bp->b_offset;
171 int nblks;
172
173 KASSERT(bp->b_offset != NOOFFSET,
174 ("cluster_read: no buffer offset"));
175
176 ncontig = 0;
177
178 /*
179 * Compute the total number of blocks that we should read
180 * synchronously.
181 */
182 if (firstread + totread > filesize)
183 totread = filesize - firstread;
184 nblks = howmany(totread, size);
185 if (nblks > racluster)
186 nblks = racluster;
187
188 /*
189 * Now compute the number of contiguous blocks.
190 */
191 if (nblks > 1) {
192 error = VOP_BMAP(vp, lblkno, NULL,
193 &blkno, &ncontig, NULL);
194 /*
195 * If this failed to map just do the original block.
196 */
197 if (error || blkno == -1)
198 ncontig = 0;
199 }
200
201 /*
202 * If we have contiguous data available do a cluster
203 * otherwise just read the requested block.
204 */
205 if (ncontig) {
206 /* Account for our first block. */
207 ncontig = min(ncontig + 1, nblks);
208 if (ncontig < nblks)
209 nblks = ncontig;
210 bp = cluster_rbuild(vp, filesize, lblkno,
211 blkno, size, nblks, bp);
212 lblkno += (bp->b_bufsize / size);
213 } else {
214 bp->b_flags |= B_RAM;
215 bp->b_iocmd = BIO_READ;
216 lblkno += 1;
217 }
218 }
219
220 /*
221 * handle the synchronous read so that it is available ASAP.
222 */
223 if (bp) {
224 if ((bp->b_flags & B_CLUSTER) == 0) {
225 vfs_busy_pages(bp, 0);
226 }
227 bp->b_flags &= ~B_INVAL;
228 bp->b_ioflags &= ~BIO_ERROR;
229 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
230 BUF_KERNPROC(bp);
231 bp->b_iooffset = dbtob(bp->b_blkno);
232 bstrategy(bp);
233 curthread->td_ru.ru_inblock++;
234 }
235
236 /*
237 * If we have been doing sequential I/O, then do some read-ahead.
238 */
239 while (lblkno < (origblkno + maxra)) {
240 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
241 if (error)
242 break;
243
244 if (blkno == -1)
245 break;
246
247 /*
248 * We could throttle ncontig here by maxra but we might as
249 * well read the data if it is contiguous. We're throttled
250 * by racluster anyway.
251 */
252 if (ncontig) {
253 ncontig = min(ncontig + 1, racluster);
254 rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
255 size, ncontig, NULL);
256 lblkno += (rbp->b_bufsize / size);
257 if (rbp->b_flags & B_DELWRI) {
258 bqrelse(rbp);
259 continue;
260 }
261 } else {
262 rbp = getblk(vp, lblkno, size, 0, 0, 0);
263 lblkno += 1;
264 if (rbp->b_flags & B_DELWRI) {
265 bqrelse(rbp);
266 continue;
267 }
268 rbp->b_flags |= B_ASYNC | B_RAM;
269 rbp->b_iocmd = BIO_READ;
270 rbp->b_blkno = blkno;
271 }
272 if (rbp->b_flags & B_CACHE) {
273 rbp->b_flags &= ~B_ASYNC;
274 bqrelse(rbp);
275 continue;
276 }
277 if ((rbp->b_flags & B_CLUSTER) == 0) {
278 vfs_busy_pages(rbp, 0);
279 }
280 rbp->b_flags &= ~B_INVAL;
281 rbp->b_ioflags &= ~BIO_ERROR;
282 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
283 BUF_KERNPROC(rbp);
284 rbp->b_iooffset = dbtob(rbp->b_blkno);
285 bstrategy(rbp);
286 curthread->td_ru.ru_inblock++;
287 }
288
289 if (reqbp)
290 return (bufwait(reqbp));
291 else
292 return (error);
293 }
294
295 /*
296 * If blocks are contiguous on disk, use this to provide clustered
297 * read ahead. We will read as many blocks as possible sequentially
298 * and then parcel them up into logical blocks in the buffer hash table.
299 */
300 static struct buf *
301 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
302 struct vnode *vp;
303 u_quad_t filesize;
304 daddr_t lbn;
305 daddr_t blkno;
306 long size;
307 int run;
308 struct buf *fbp;
309 {
310 struct bufobj *bo;
311 struct buf *bp, *tbp;
312 daddr_t bn;
313 int i, inc, j;
314
315 KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
316 ("cluster_rbuild: size %ld != filesize %jd\n",
317 size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
318
319 /*
320 * avoid a division
321 */
322 while ((u_quad_t) size * (lbn + run) > filesize) {
323 --run;
324 }
325
326 if (fbp) {
327 tbp = fbp;
328 tbp->b_iocmd = BIO_READ;
329 } else {
330 tbp = getblk(vp, lbn, size, 0, 0, 0);
331 if (tbp->b_flags & B_CACHE)
332 return tbp;
333 tbp->b_flags |= B_ASYNC | B_RAM;
334 tbp->b_iocmd = BIO_READ;
335 }
336 tbp->b_blkno = blkno;
337 if( (tbp->b_flags & B_MALLOC) ||
338 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
339 return tbp;
340
341 bp = trypbuf(&cluster_pbuf_freecnt);
342 if (bp == 0)
343 return tbp;
344
345 /*
346 * We are synthesizing a buffer out of vm_page_t's, but
347 * if the block size is not page aligned then the starting
348 * address may not be either. Inherit the b_data offset
349 * from the original buffer.
350 */
351 bp->b_data = (char *)((vm_offset_t)bp->b_data |
352 ((vm_offset_t)tbp->b_data & PAGE_MASK));
353 bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
354 bp->b_iocmd = BIO_READ;
355 bp->b_iodone = cluster_callback;
356 bp->b_blkno = blkno;
357 bp->b_lblkno = lbn;
358 bp->b_offset = tbp->b_offset;
359 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
360 pbgetvp(vp, bp);
361
362 TAILQ_INIT(&bp->b_cluster.cluster_head);
363
364 bp->b_bcount = 0;
365 bp->b_bufsize = 0;
366 bp->b_npages = 0;
367
368 inc = btodb(size);
369 bo = &vp->v_bufobj;
370 for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
371 if (i != 0) {
372 if ((bp->b_npages * PAGE_SIZE) +
373 round_page(size) > vp->v_mount->mnt_iosize_max) {
374 break;
375 }
376
377 tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT);
378
379 /* Don't wait around for locked bufs. */
380 if (tbp == NULL)
381 break;
382
383 /*
384 * Stop scanning if the buffer is fully valid
385 * (marked B_CACHE), or locked (may be doing a
386 * background write), or if the buffer is not
387 * VMIO backed. The clustering code can only deal
388 * with VMIO-backed buffers.
389 */
390 BO_LOCK(bo);
391 if ((tbp->b_vflags & BV_BKGRDINPROG) ||
392 (tbp->b_flags & B_CACHE) ||
393 (tbp->b_flags & B_VMIO) == 0) {
394 BO_UNLOCK(bo);
395 bqrelse(tbp);
396 break;
397 }
398 BO_UNLOCK(bo);
399
400 /*
401 * The buffer must be completely invalid in order to
402 * take part in the cluster. If it is partially valid
403 * then we stop.
404 */
405 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
406 for (j = 0;j < tbp->b_npages; j++) {
407 VM_OBJECT_LOCK_ASSERT(tbp->b_pages[j]->object,
408 MA_OWNED);
409 if (tbp->b_pages[j]->valid)
410 break;
411 }
412 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
413 if (j != tbp->b_npages) {
414 bqrelse(tbp);
415 break;
416 }
417
418 /*
419 * Set a read-ahead mark as appropriate
420 */
421 if ((fbp && (i == 1)) || (i == (run - 1)))
422 tbp->b_flags |= B_RAM;
423
424 /*
425 * Set the buffer up for an async read (XXX should
426 * we do this only if we do not wind up brelse()ing?).
427 * Set the block number if it isn't set, otherwise
428 * if it is make sure it matches the block number we
429 * expect.
430 */
431 tbp->b_flags |= B_ASYNC;
432 tbp->b_iocmd = BIO_READ;
433 if (tbp->b_blkno == tbp->b_lblkno) {
434 tbp->b_blkno = bn;
435 } else if (tbp->b_blkno != bn) {
436 brelse(tbp);
437 break;
438 }
439 }
440 /*
441 * XXX fbp from caller may not be B_ASYNC, but we are going
442 * to biodone() it in cluster_callback() anyway
443 */
444 BUF_KERNPROC(tbp);
445 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
446 tbp, b_cluster.cluster_entry);
447 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
448 for (j = 0; j < tbp->b_npages; j += 1) {
449 vm_page_t m;
450 m = tbp->b_pages[j];
451 vm_page_io_start(m);
452 vm_object_pip_add(m->object, 1);
453 if ((bp->b_npages == 0) ||
454 (bp->b_pages[bp->b_npages-1] != m)) {
455 bp->b_pages[bp->b_npages] = m;
456 bp->b_npages++;
457 }
458 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
459 tbp->b_pages[j] = bogus_page;
460 }
461 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
462 /*
463 * XXX shouldn't this be += size for both, like in
464 * cluster_wbuild()?
465 *
466 * Don't inherit tbp->b_bufsize as it may be larger due to
467 * a non-page-aligned size. Instead just aggregate using
468 * 'size'.
469 */
470 if (tbp->b_bcount != size)
471 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
472 if (tbp->b_bufsize != size)
473 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
474 bp->b_bcount += size;
475 bp->b_bufsize += size;
476 }
477
478 /*
479 * Fully valid pages in the cluster are already good and do not need
480 * to be re-read from disk. Replace the page with bogus_page
481 */
482 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
483 for (j = 0; j < bp->b_npages; j++) {
484 VM_OBJECT_LOCK_ASSERT(bp->b_pages[j]->object, MA_OWNED);
485 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
486 VM_PAGE_BITS_ALL) {
487 bp->b_pages[j] = bogus_page;
488 }
489 }
490 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
491 if (bp->b_bufsize > bp->b_kvasize)
492 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
493 bp->b_bufsize, bp->b_kvasize);
494 bp->b_kvasize = bp->b_bufsize;
495
496 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
497 (vm_page_t *)bp->b_pages, bp->b_npages);
498 return (bp);
499 }
500
501 /*
502 * Cleanup after a clustered read or write.
503 * This is complicated by the fact that any of the buffers might have
504 * extra memory (if there were no empty buffer headers at allocbuf time)
505 * that we will need to shift around.
506 */
507 static void
508 cluster_callback(bp)
509 struct buf *bp;
510 {
511 struct buf *nbp, *tbp;
512 int error = 0;
513
514 /*
515 * Must propogate errors to all the components.
516 */
517 if (bp->b_ioflags & BIO_ERROR)
518 error = bp->b_error;
519
520 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
521 /*
522 * Move memory from the large cluster buffer into the component
523 * buffers and mark IO as done on these.
524 */
525 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
526 tbp; tbp = nbp) {
527 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
528 if (error) {
529 tbp->b_ioflags |= BIO_ERROR;
530 tbp->b_error = error;
531 } else {
532 tbp->b_dirtyoff = tbp->b_dirtyend = 0;
533 tbp->b_flags &= ~B_INVAL;
534 tbp->b_ioflags &= ~BIO_ERROR;
535 /*
536 * XXX the bdwrite()/bqrelse() issued during
537 * cluster building clears B_RELBUF (see bqrelse()
538 * comment). If direct I/O was specified, we have
539 * to restore it here to allow the buffer and VM
540 * to be freed.
541 */
542 if (tbp->b_flags & B_DIRECT)
543 tbp->b_flags |= B_RELBUF;
544 }
545 bufdone(tbp);
546 }
547 pbrelvp(bp);
548 relpbuf(bp, &cluster_pbuf_freecnt);
549 }
550
551 /*
552 * cluster_wbuild_wb:
553 *
554 * Implement modified write build for cluster.
555 *
556 * write_behind = 0 write behind disabled
557 * write_behind = 1 write behind normal (default)
558 * write_behind = 2 write behind backed-off
559 */
560
561 static __inline int
562 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
563 {
564 int r = 0;
565
566 switch(write_behind) {
567 case 2:
568 if (start_lbn < len)
569 break;
570 start_lbn -= len;
571 /* FALLTHROUGH */
572 case 1:
573 r = cluster_wbuild(vp, size, start_lbn, len);
574 /* FALLTHROUGH */
575 default:
576 /* FALLTHROUGH */
577 break;
578 }
579 return(r);
580 }
581
582 /*
583 * Do clustered write for FFS.
584 *
585 * Three cases:
586 * 1. Write is not sequential (write asynchronously)
587 * Write is sequential:
588 * 2. beginning of cluster - begin cluster
589 * 3. middle of a cluster - add to cluster
590 * 4. end of a cluster - asynchronously write cluster
591 */
592 void
593 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount)
594 {
595 daddr_t lbn;
596 int maxclen, cursize;
597 int lblocksize;
598 int async;
599
600 if (vp->v_type == VREG) {
601 async = vp->v_mount->mnt_kern_flag & MNTK_ASYNC;
602 lblocksize = vp->v_mount->mnt_stat.f_iosize;
603 } else {
604 async = 0;
605 lblocksize = bp->b_bufsize;
606 }
607 lbn = bp->b_lblkno;
608 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
609
610 /* Initialize vnode to beginning of file. */
611 if (lbn == 0)
612 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
613
614 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
615 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
616 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
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 + 1;
635 if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
636 lbn != vp->v_lastw + 1 || 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 endbp = &buflist->bs_children
647 [buflist->bs_nchildren - 1];
648 if (VOP_REALLOCBLKS(vp, buflist)) {
649 /*
650 * Failed, push the previous cluster
651 * if *really* writing sequentially
652 * in the logical file (seqcount > 1),
653 * otherwise delay it in the hopes that
654 * the low level disk driver can
655 * optimize the write ordering.
656 */
657 for (bpp = buflist->bs_children;
658 bpp < endbp; bpp++)
659 brelse(*bpp);
660 free(buflist, M_SEGMENT);
661 if (seqcount > 1) {
662 cluster_wbuild_wb(vp,
663 lblocksize, vp->v_cstart,
664 cursize);
665 }
666 } else {
667 /*
668 * Succeeded, keep building cluster.
669 */
670 for (bpp = buflist->bs_children;
671 bpp <= endbp; bpp++)
672 bdwrite(*bpp);
673 free(buflist, M_SEGMENT);
674 vp->v_lastw = lbn;
675 vp->v_lasta = bp->b_blkno;
676 return;
677 }
678 }
679 }
680 /*
681 * Consider beginning a cluster. If at end of file, make
682 * cluster as large as possible, otherwise find size of
683 * existing cluster.
684 */
685 if ((vp->v_type == VREG) &&
686 ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
687 (bp->b_blkno == bp->b_lblkno) &&
688 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
689 bp->b_blkno == -1)) {
690 bawrite(bp);
691 vp->v_clen = 0;
692 vp->v_lasta = bp->b_blkno;
693 vp->v_cstart = lbn + 1;
694 vp->v_lastw = lbn;
695 return;
696 }
697 vp->v_clen = maxclen;
698 if (!async && maxclen == 0) { /* I/O not contiguous */
699 vp->v_cstart = lbn + 1;
700 bawrite(bp);
701 } else { /* Wait for rest of cluster */
702 vp->v_cstart = lbn;
703 bdwrite(bp);
704 }
705 } else if (lbn == vp->v_cstart + vp->v_clen) {
706 /*
707 * At end of cluster, write it out if seqcount tells us we
708 * are operating sequentially, otherwise let the buf or
709 * update daemon handle it.
710 */
711 bdwrite(bp);
712 if (seqcount > 1)
713 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
714 vp->v_clen = 0;
715 vp->v_cstart = lbn + 1;
716 } else if (vm_page_count_severe()) {
717 /*
718 * We are low on memory, get it going NOW
719 */
720 bawrite(bp);
721 } else {
722 /*
723 * In the middle of a cluster, so just delay the I/O for now.
724 */
725 bdwrite(bp);
726 }
727 vp->v_lastw = lbn;
728 vp->v_lasta = bp->b_blkno;
729 }
730
731
732 /*
733 * This is an awful lot like cluster_rbuild...wish they could be combined.
734 * The last lbn argument is the current block on which I/O is being
735 * performed. Check to see that it doesn't fall in the middle of
736 * the current block (if last_bp == NULL).
737 */
738 int
739 cluster_wbuild(vp, size, start_lbn, len)
740 struct vnode *vp;
741 long size;
742 daddr_t start_lbn;
743 int len;
744 {
745 struct buf *bp, *tbp;
746 struct bufobj *bo;
747 int i, j;
748 int totalwritten = 0;
749 int dbsize = btodb(size);
750
751 bo = &vp->v_bufobj;
752 while (len > 0) {
753 /*
754 * If the buffer is not delayed-write (i.e. dirty), or it
755 * is delayed-write but either locked or inval, it cannot
756 * partake in the clustered write.
757 */
758 BO_LOCK(bo);
759 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
760 (tbp->b_vflags & BV_BKGRDINPROG)) {
761 BO_UNLOCK(bo);
762 ++start_lbn;
763 --len;
764 continue;
765 }
766 if (BUF_LOCK(tbp,
767 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_MTX(bo))) {
768 ++start_lbn;
769 --len;
770 continue;
771 }
772 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
773 BUF_UNLOCK(tbp);
774 ++start_lbn;
775 --len;
776 continue;
777 }
778 if (tbp->b_pin_count > 0) {
779 BUF_UNLOCK(tbp);
780 ++start_lbn;
781 --len;
782 continue;
783 }
784 bremfree(tbp);
785 tbp->b_flags &= ~B_DONE;
786
787 /*
788 * Extra memory in the buffer, punt on this buffer.
789 * XXX we could handle this in most cases, but we would
790 * have to push the extra memory down to after our max
791 * possible cluster size and then potentially pull it back
792 * up if the cluster was terminated prematurely--too much
793 * hassle.
794 */
795 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
796 (B_CLUSTEROK | B_VMIO)) ||
797 (tbp->b_bcount != tbp->b_bufsize) ||
798 (tbp->b_bcount != size) ||
799 (len == 1) ||
800 ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
801 totalwritten += tbp->b_bufsize;
802 bawrite(tbp);
803 ++start_lbn;
804 --len;
805 continue;
806 }
807
808 /*
809 * We got a pbuf to make the cluster in.
810 * so initialise it.
811 */
812 TAILQ_INIT(&bp->b_cluster.cluster_head);
813 bp->b_bcount = 0;
814 bp->b_bufsize = 0;
815 bp->b_npages = 0;
816 if (tbp->b_wcred != NOCRED)
817 bp->b_wcred = crhold(tbp->b_wcred);
818
819 bp->b_blkno = tbp->b_blkno;
820 bp->b_lblkno = tbp->b_lblkno;
821 bp->b_offset = tbp->b_offset;
822
823 /*
824 * We are synthesizing a buffer out of vm_page_t's, but
825 * if the block size is not page aligned then the starting
826 * address may not be either. Inherit the b_data offset
827 * from the original buffer.
828 */
829 bp->b_data = (char *)((vm_offset_t)bp->b_data |
830 ((vm_offset_t)tbp->b_data & PAGE_MASK));
831 bp->b_flags |= B_CLUSTER |
832 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
833 bp->b_iodone = cluster_callback;
834 pbgetvp(vp, bp);
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 < len; ++i, ++start_lbn) {
841 if (i != 0) { /* If not the first buffer */
842 /*
843 * If the adjacent data is not even in core it
844 * can't need to be written.
845 */
846 BO_LOCK(bo);
847 if ((tbp = gbincore(bo, start_lbn)) == NULL ||
848 (tbp->b_vflags & BV_BKGRDINPROG)) {
849 BO_UNLOCK(bo);
850 break;
851 }
852
853 /*
854 * If it IS in core, but has different
855 * characteristics, or is locked (which
856 * means it could be undergoing a background
857 * I/O or be in a weird state), then don't
858 * cluster with it.
859 */
860 if (BUF_LOCK(tbp,
861 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
862 BO_MTX(bo)))
863 break;
864
865 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
866 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
867 != (B_DELWRI | B_CLUSTEROK |
868 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
869 tbp->b_wcred != bp->b_wcred) {
870 BUF_UNLOCK(tbp);
871 break;
872 }
873
874 /*
875 * Check that the combined cluster
876 * would make sense with regard to pages
877 * and would not be too large
878 */
879 if ((tbp->b_bcount != size) ||
880 ((bp->b_blkno + (dbsize * i)) !=
881 tbp->b_blkno) ||
882 ((tbp->b_npages + bp->b_npages) >
883 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
884 BUF_UNLOCK(tbp);
885 break;
886 }
887
888 /*
889 * Do not pull in pinned buffers.
890 */
891 if (tbp->b_pin_count > 0) {
892 BUF_UNLOCK(tbp);
893 break;
894 }
895
896 /*
897 * Ok, it's passed all the tests,
898 * so remove it from the free list
899 * and mark it busy. We will use it.
900 */
901 bremfree(tbp);
902 tbp->b_flags &= ~B_DONE;
903 } /* end of code for non-first buffers only */
904 /*
905 * If the IO is via the VM then we do some
906 * special VM hackery (yuck). Since the buffer's
907 * block size may not be page-aligned it is possible
908 * for a page to be shared between two buffers. We
909 * have to get rid of the duplication when building
910 * the cluster.
911 */
912 if (tbp->b_flags & B_VMIO) {
913 vm_page_t m;
914
915 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
916 if (i != 0) { /* if not first buffer */
917 for (j = 0; j < tbp->b_npages; j += 1) {
918 m = tbp->b_pages[j];
919 if (m->oflags & VPO_BUSY) {
920 VM_OBJECT_UNLOCK(
921 tbp->b_object);
922 bqrelse(tbp);
923 goto finishcluster;
924 }
925 }
926 }
927 for (j = 0; j < tbp->b_npages; j += 1) {
928 m = tbp->b_pages[j];
929 vm_page_io_start(m);
930 vm_object_pip_add(m->object, 1);
931 if ((bp->b_npages == 0) ||
932 (bp->b_pages[bp->b_npages - 1] != m)) {
933 bp->b_pages[bp->b_npages] = m;
934 bp->b_npages++;
935 }
936 }
937 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
938 }
939 bp->b_bcount += size;
940 bp->b_bufsize += size;
941 bundirty(tbp);
942 tbp->b_flags &= ~B_DONE;
943 tbp->b_ioflags &= ~BIO_ERROR;
944 tbp->b_flags |= B_ASYNC;
945 tbp->b_iocmd = BIO_WRITE;
946 reassignbuf(tbp); /* put on clean list */
947 bufobj_wref(tbp->b_bufobj);
948 BUF_KERNPROC(tbp);
949 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
950 tbp, b_cluster.cluster_entry);
951 }
952 finishcluster:
953 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
954 (vm_page_t *) bp->b_pages, bp->b_npages);
955 if (bp->b_bufsize > bp->b_kvasize)
956 panic(
957 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
958 bp->b_bufsize, bp->b_kvasize);
959 bp->b_kvasize = bp->b_bufsize;
960 totalwritten += bp->b_bufsize;
961 bp->b_dirtyoff = 0;
962 bp->b_dirtyend = bp->b_bufsize;
963 bawrite(bp);
964
965 len -= i;
966 }
967 return totalwritten;
968 }
969
970 /*
971 * Collect together all the buffers in a cluster.
972 * Plus add one additional buffer.
973 */
974 static struct cluster_save *
975 cluster_collectbufs(vp, last_bp)
976 struct vnode *vp;
977 struct buf *last_bp;
978 {
979 struct cluster_save *buflist;
980 struct buf *bp;
981 daddr_t lbn;
982 int i, len;
983
984 len = vp->v_lastw - vp->v_cstart + 1;
985 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
986 M_SEGMENT, M_WAITOK);
987 buflist->bs_nchildren = 0;
988 buflist->bs_children = (struct buf **) (buflist + 1);
989 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
990 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
991 buflist->bs_children[i] = bp;
992 if (bp->b_blkno == bp->b_lblkno)
993 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
994 NULL, NULL);
995 }
996 buflist->bs_children[i] = bp = last_bp;
997 if (bp->b_blkno == bp->b_lblkno)
998 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
999 buflist->bs_nchildren = i + 1;
1000 return (buflist);
1001 }
FreeBSD kvm: FreeBSD sources