Fix two bugs in cluster_read:
[dragonfly.git] / sys / kern / vfs_cluster.c
CommitLineData
984263bc
MD
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 $
bfda7080 37 * $DragonFly: src/sys/kern/vfs_cluster.c,v 1.26 2006/07/14 19:23:39 corecode Exp $
984263bc
MD
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>
3020e3be 56#include <sys/buf2.h>
12e4aaff 57#include <vm/vm_page2.h>
984263bc
MD
58
59#if defined(CLUSTERDEBUG)
60#include <sys/sysctl.h>
61static int rcluster= 0;
62SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
63#endif
64
d1cd9d97 65static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");
984263bc
MD
66
67static struct cluster_save *
54078292
MD
68 cluster_collectbufs (struct vnode *vp, struct buf *last_bp,
69 int lblocksize);
984263bc 70static struct buf *
54078292 71 cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
10f3fee5
MD
72 off_t doffset, int size, int run,
73 struct buf *fbp, int doasync);
81b5c339
MD
74static void cluster_callback (struct bio *);
75
984263bc
MD
76
77static int write_behind = 1;
78SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
79
80extern vm_page_t bogus_page;
81
82extern 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 */
92int
54078292
MD
93cluster_read(struct vnode *vp, off_t filesize, off_t loffset,
94 int size, int totread, int seqcount, struct buf **bpp)
984263bc
MD
95{
96 struct buf *bp, *rbp, *reqbp;
54078292
MD
97 off_t origoffset;
98 off_t doffset;
99 int error;
984263bc
MD
100 int i;
101 int maxra, racluster;
984263bc
MD
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 */
54078292
MD
119 *bpp = reqbp = bp = getblk(vp, loffset, size, 0, 0);
120 origoffset = loffset;
984263bc
MD
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 {
984263bc
MD
133 struct buf *tbp;
134 bp->b_flags &= ~B_RAM;
135 /*
e43a034f 136 * We do the crit here so that there is no window
1f1ea522 137 * between the findblk and the b_usecount increment
e43a034f 138 * below. We opt to keep the crit out of the loop
984263bc
MD
139 * for efficiency.
140 */
e43a034f 141 crit_enter();
984263bc 142 for (i = 1; i < maxra; i++) {
54078292 143 if (!(tbp = findblk(vp, loffset + i * size))) {
984263bc
MD
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 }
e43a034f 155 crit_exit();
984263bc
MD
156 if (i >= maxra) {
157 return 0;
158 }
54078292 159 loffset += i * size;
984263bc
MD
160 }
161 reqbp = bp = NULL;
162 } else {
81b5c339 163 off_t firstread = bp->b_loffset;
54078292 164 int nblks;
984263bc 165
81b5c339
MD
166 KASSERT(firstread != NOOFFSET,
167 ("cluster_read: no buffer offset"));
984263bc 168 if (firstread + totread > filesize)
54078292
MD
169 totread = (int)(filesize - firstread);
170 nblks = totread / size;
171 if (nblks) {
172 int burstbytes;
173
984263bc
MD
174 if (nblks > racluster)
175 nblks = racluster;
176
54078292
MD
177 error = VOP_BMAP(vp, loffset, NULL,
178 &doffset, &burstbytes, NULL);
984263bc
MD
179 if (error)
180 goto single_block_read;
54078292 181 if (doffset == NOOFFSET)
984263bc 182 goto single_block_read;
54078292 183 if (burstbytes < size * 2)
984263bc 184 goto single_block_read;
54078292
MD
185 if (nblks > burstbytes / size)
186 nblks = burstbytes / size;
984263bc 187
54078292 188 bp = cluster_rbuild(vp, filesize, loffset,
10f3fee5 189 doffset, size, nblks, bp, 0);
54078292 190 loffset += bp->b_bufsize;
984263bc
MD
191 } else {
192single_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 */
10f3fee5 197 bp->b_flags |= B_RAM;
54078292 198 loffset += size;
984263bc
MD
199 }
200 }
201
984263bc 202 /*
a8f169e2 203 * Handle the synchronous read. This only occurs if B_CACHE was
10f3fee5
MD
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.
984263bc
MD
207 */
208 if (bp) {
209#if defined(CLUSTERDEBUG)
210 if (rcluster)
591bdbe9 211 printf("S(%lld,%d,%d) ",
54078292 212 bp->b_loffset, bp->b_bcount, seqcount);
984263bc 213#endif
10f3fee5
MD
214 bp->b_cmd = BUF_CMD_READ;
215 if ((bp->b_flags & B_CLUSTER) == 0)
216 vfs_busy_pages(vp, bp);
984263bc 217 bp->b_flags &= ~(B_ERROR|B_INVAL);
81b5c339 218 if ((bp->b_flags & B_ASYNC) || bp->b_bio1.bio_done != NULL)
984263bc 219 BUF_KERNPROC(bp);
81b5c339
MD
220 vn_strategy(vp, &bp->b_bio1);
221 error = bp->b_error;
984263bc
MD
222 }
223
224 /*
bfda7080 225 * If we have been doing sequential I/O, then do some read-ahead.
984263bc 226 */
bfda7080
SS
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)) {
984263bc 239 bqrelse(rbp);
bfda7080
SS
240 goto no_read_ahead;
241 }
242
243 error = VOP_BMAP(vp, loffset, NULL,
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);
984263bc 263 } else {
bfda7080
SS
264 rbp->b_bio2.bio_offset = doffset;
265 }
984263bc 266#if defined(CLUSTERDEBUG)
bfda7080
SS
267 if (rcluster) {
268 if (bp)
269 printf("A+(%lld,%d,%lld,%d) ",
270 rbp->b_loffset, rbp->b_bcount,
271 rbp->b_loffset - origoffset,
272 seqcount);
273 else
274 printf("A(%lld,%d,%lld,%d) ",
275 rbp->b_loffset, rbp->b_bcount,
276 rbp->b_loffset - origoffset,
277 seqcount);
278 }
984263bc 279#endif
bfda7080
SS
280 rbp->b_flags &= ~(B_ERROR|B_INVAL);
281 rbp->b_flags |= B_ASYNC;
282 rbp->b_cmd = BUF_CMD_READ;
10f3fee5 283
bfda7080
SS
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);
984263bc 288 }
bfda7080
SS
289no_read_ahead:
290
984263bc
MD
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 */
302static struct buf *
54078292 303cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset,
10f3fee5 304 off_t doffset, int size, int run, struct buf *fbp, int doasync)
984263bc
MD
305{
306 struct buf *bp, *tbp;
54078292
MD
307 off_t boffset;
308 int i, j;
984263bc
MD
309
310 KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
54078292 311 ("cluster_rbuild: size %d != filesize %ld\n",
984263bc
MD
312 size, vp->v_mount->mnt_stat.f_iosize));
313
314 /*
315 * avoid a division
316 */
54078292 317 while (loffset + run * size > filesize) {
984263bc
MD
318 --run;
319 }
320
6260e485 321 tbp = fbp;
54078292 322 tbp->b_bio2.bio_offset = doffset;
10f3fee5
MD
323 if((tbp->b_flags & B_MALLOC) ||
324 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
984263bc 325 return tbp;
10f3fee5 326 }
984263bc
MD
327
328 bp = trypbuf(&cluster_pbuf_freecnt);
a8f169e2 329 if (bp == NULL)
984263bc
MD
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));
10f3fee5
MD
340 bp->b_flags |= B_ASYNC | B_CLUSTER | B_VMIO;
341 bp->b_cmd = BUF_CMD_READ;
81b5c339
MD
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;
54078292
MD
345 bp->b_loffset = loffset;
346 bp->b_bio2.bio_offset = NOOFFSET;
81b5c339
MD
347 KASSERT(bp->b_loffset != NOOFFSET,
348 ("cluster_rbuild: no buffer offset"));
984263bc 349
984263bc
MD
350 bp->b_bcount = 0;
351 bp->b_bufsize = 0;
54f51aeb 352 bp->b_xio.xio_npages = 0;
984263bc 353
54078292 354 for (boffset = doffset, i = 0; i < run; ++i, boffset += size) {
10f3fee5 355 if (i) {
54f51aeb 356 if ((bp->b_xio.xio_npages * PAGE_SIZE) +
984263bc
MD
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 */
54078292 366 if ((tbp = findblk(vp, loffset + i * size)) != NULL) {
984263bc
MD
367 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
368 break;
369 BUF_UNLOCK(tbp);
370
54f51aeb
HP
371 for (j = 0; j < tbp->b_xio.xio_npages; j++) {
372 if (tbp->b_xio.xio_pages[j]->valid)
984263bc
MD
373 break;
374 }
375
54f51aeb 376 if (j != tbp->b_xio.xio_npages)
984263bc
MD
377 break;
378
379 if (tbp->b_bcount != size)
380 break;
381 }
382
54078292 383 tbp = getblk(vp, loffset + i * size, size, 0, 0);
984263bc
MD
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 bqrelse(tbp);
395 break;
396 }
397
398 /*
399 * The buffer must be completely invalid in order to
400 * take part in the cluster. If it is partially valid
401 * then we stop.
402 */
54f51aeb
HP
403 for (j = 0;j < tbp->b_xio.xio_npages; j++) {
404 if (tbp->b_xio.xio_pages[j]->valid)
984263bc
MD
405 break;
406 }
54f51aeb 407 if (j != tbp->b_xio.xio_npages) {
984263bc
MD
408 bqrelse(tbp);
409 break;
410 }
411
412 /*
413 * Set a read-ahead mark as appropriate
414 */
6260e485 415 if (i == 1 || i == (run - 1))
984263bc
MD
416 tbp->b_flags |= B_RAM;
417
418 /*
984263bc
MD
419 * Set the block number if it isn't set, otherwise
420 * if it is make sure it matches the block number we
421 * expect.
422 */
54078292
MD
423 if (tbp->b_bio2.bio_offset == NOOFFSET) {
424 tbp->b_bio2.bio_offset = boffset;
425 } else if (tbp->b_bio2.bio_offset != boffset) {
984263bc
MD
426 brelse(tbp);
427 break;
428 }
429 }
430 /*
10f3fee5
MD
431 * The first buffer is setup async if doasync is specified.
432 * All other buffers in the cluster are setup async. This
433 * way the caller can decide how to deal with the requested
434 * buffer.
984263bc 435 */
10f3fee5
MD
436 if (i || doasync)
437 tbp->b_flags |= B_ASYNC;
438 tbp->b_cmd = BUF_CMD_READ;
984263bc 439 BUF_KERNPROC(tbp);
81b5c339 440 cluster_append(&bp->b_bio1, tbp);
54078292 441 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
984263bc 442 vm_page_t m;
54f51aeb 443 m = tbp->b_xio.xio_pages[j];
984263bc
MD
444 vm_page_io_start(m);
445 vm_object_pip_add(m->object, 1);
54f51aeb
HP
446 if ((bp->b_xio.xio_npages == 0) ||
447 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
448 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
449 bp->b_xio.xio_npages++;
984263bc
MD
450 }
451 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
54f51aeb 452 tbp->b_xio.xio_pages[j] = bogus_page;
984263bc
MD
453 }
454 /*
455 * XXX shouldn't this be += size for both, like in
456 * cluster_wbuild()?
457 *
458 * Don't inherit tbp->b_bufsize as it may be larger due to
459 * a non-page-aligned size. Instead just aggregate using
460 * 'size'.
461 */
462 if (tbp->b_bcount != size)
54078292 463 printf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, size);
984263bc 464 if (tbp->b_bufsize != size)
54078292 465 printf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, size);
984263bc
MD
466 bp->b_bcount += size;
467 bp->b_bufsize += size;
468 }
469
470 /*
471 * Fully valid pages in the cluster are already good and do not need
472 * to be re-read from disk. Replace the page with bogus_page
473 */
54f51aeb
HP
474 for (j = 0; j < bp->b_xio.xio_npages; j++) {
475 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
984263bc 476 VM_PAGE_BITS_ALL) {
54f51aeb 477 bp->b_xio.xio_pages[j] = bogus_page;
984263bc
MD
478 }
479 }
312dcd01 480 if (bp->b_bufsize > bp->b_kvasize) {
54078292 481 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
984263bc 482 bp->b_bufsize, bp->b_kvasize);
312dcd01 483 }
984263bc
MD
484
485 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
54f51aeb 486 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
984263bc
MD
487 return (bp);
488}
489
490/*
491 * Cleanup after a clustered read or write.
492 * This is complicated by the fact that any of the buffers might have
493 * extra memory (if there were no empty buffer headers at allocbuf time)
494 * that we will need to shift around.
81b5c339
MD
495 *
496 * The returned bio is &bp->b_bio1
984263bc
MD
497 */
498void
81b5c339 499cluster_callback(struct bio *bio)
984263bc 500{
81b5c339
MD
501 struct buf *bp = bio->bio_buf;
502 struct buf *tbp;
984263bc
MD
503 int error = 0;
504
505 /*
9a71d53f
MD
506 * Must propogate errors to all the components. A short read (EOF)
507 * is a critical error.
984263bc 508 */
9a71d53f 509 if (bp->b_flags & B_ERROR) {
984263bc 510 error = bp->b_error;
9a71d53f
MD
511 } else if (bp->b_bcount != bp->b_bufsize) {
512 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
513 }
984263bc 514
54f51aeb 515 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
984263bc
MD
516 /*
517 * Move memory from the large cluster buffer into the component
81b5c339
MD
518 * buffers and mark IO as done on these. Since the memory map
519 * is the same, no actual copying is required.
984263bc 520 */
81b5c339
MD
521 while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
522 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
984263bc
MD
523 if (error) {
524 tbp->b_flags |= B_ERROR;
525 tbp->b_error = error;
526 } else {
527 tbp->b_dirtyoff = tbp->b_dirtyend = 0;
528 tbp->b_flags &= ~(B_ERROR|B_INVAL);
529 /*
530 * XXX the bdwrite()/bqrelse() issued during
531 * cluster building clears B_RELBUF (see bqrelse()
532 * comment). If direct I/O was specified, we have
533 * to restore it here to allow the buffer and VM
534 * to be freed.
535 */
536 if (tbp->b_flags & B_DIRECT)
537 tbp->b_flags |= B_RELBUF;
538 }
81b5c339 539 biodone(&tbp->b_bio1);
984263bc
MD
540 }
541 relpbuf(bp, &cluster_pbuf_freecnt);
542}
543
544/*
545 * cluster_wbuild_wb:
546 *
547 * Implement modified write build for cluster.
548 *
549 * write_behind = 0 write behind disabled
550 * write_behind = 1 write behind normal (default)
551 * write_behind = 2 write behind backed-off
552 */
553
554static __inline int
54078292 555cluster_wbuild_wb(struct vnode *vp, int size, off_t start_loffset, int len)
984263bc
MD
556{
557 int r = 0;
558
559 switch(write_behind) {
560 case 2:
54078292 561 if (start_loffset < len)
984263bc 562 break;
54078292 563 start_loffset -= len;
984263bc
MD
564 /* fall through */
565 case 1:
54078292 566 r = cluster_wbuild(vp, size, start_loffset, len);
984263bc
MD
567 /* fall through */
568 default:
569 /* fall through */
570 break;
571 }
572 return(r);
573}
574
575/*
576 * Do clustered write for FFS.
577 *
578 * Three cases:
579 * 1. Write is not sequential (write asynchronously)
580 * Write is sequential:
581 * 2. beginning of cluster - begin cluster
582 * 3. middle of a cluster - add to cluster
583 * 4. end of a cluster - asynchronously write cluster
584 */
585void
54078292 586cluster_write(struct buf *bp, off_t filesize, int seqcount)
984263bc
MD
587{
588 struct vnode *vp;
54078292 589 off_t loffset;
984263bc
MD
590 int maxclen, cursize;
591 int lblocksize;
592 int async;
593
594 vp = bp->b_vp;
595 if (vp->v_type == VREG) {
596 async = vp->v_mount->mnt_flag & MNT_ASYNC;
597 lblocksize = vp->v_mount->mnt_stat.f_iosize;
598 } else {
599 async = 0;
600 lblocksize = bp->b_bufsize;
601 }
54078292 602 loffset = bp->b_loffset;
81b5c339
MD
603 KASSERT(bp->b_loffset != NOOFFSET,
604 ("cluster_write: no buffer offset"));
984263bc
MD
605
606 /* Initialize vnode to beginning of file. */
54078292 607 if (loffset == 0)
984263bc
MD
608 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
609
54078292
MD
610 if (vp->v_clen == 0 || loffset != vp->v_lastw + lblocksize ||
611 bp->b_bio2.bio_offset == NOOFFSET ||
612 (bp->b_bio2.bio_offset != vp->v_lasta + lblocksize)) {
613 maxclen = vp->v_mount->mnt_iosize_max;
984263bc
MD
614 if (vp->v_clen != 0) {
615 /*
616 * Next block is not sequential.
617 *
618 * If we are not writing at end of file, the process
619 * seeked to another point in the file since its last
620 * write, or we have reached our maximum cluster size,
621 * then push the previous cluster. Otherwise try
622 * reallocating to make it sequential.
623 *
624 * Change to algorithm: only push previous cluster if
625 * it was sequential from the point of view of the
626 * seqcount heuristic, otherwise leave the buffer
627 * intact so we can potentially optimize the I/O
628 * later on in the buf_daemon or update daemon
629 * flush.
630 */
54078292
MD
631 cursize = vp->v_lastw - vp->v_cstart + lblocksize;
632 if (bp->b_loffset + lblocksize != filesize ||
633 loffset != vp->v_lastw + lblocksize || vp->v_clen <= cursize) {
984263bc
MD
634 if (!async && seqcount > 0) {
635 cluster_wbuild_wb(vp, lblocksize,
636 vp->v_cstart, cursize);
637 }
638 } else {
639 struct buf **bpp, **endbp;
640 struct cluster_save *buflist;
641
54078292
MD
642 buflist = cluster_collectbufs(vp, bp,
643 lblocksize);
984263bc
MD
644 endbp = &buflist->bs_children
645 [buflist->bs_nchildren - 1];
646 if (VOP_REALLOCBLKS(vp, buflist)) {
647 /*
648 * Failed, push the previous cluster
649 * if *really* writing sequentially
650 * in the logical file (seqcount > 1),
651 * otherwise delay it in the hopes that
652 * the low level disk driver can
653 * optimize the write ordering.
654 */
655 for (bpp = buflist->bs_children;
656 bpp < endbp; bpp++)
657 brelse(*bpp);
658 free(buflist, M_SEGMENT);
659 if (seqcount > 1) {
660 cluster_wbuild_wb(vp,
661 lblocksize, vp->v_cstart,
662 cursize);
663 }
664 } else {
665 /*
666 * Succeeded, keep building cluster.
667 */
668 for (bpp = buflist->bs_children;
669 bpp <= endbp; bpp++)
670 bdwrite(*bpp);
671 free(buflist, M_SEGMENT);
54078292
MD
672 vp->v_lastw = loffset;
673 vp->v_lasta = bp->b_bio2.bio_offset;
984263bc
MD
674 return;
675 }
676 }
677 }
678 /*
679 * Consider beginning a cluster. If at end of file, make
680 * cluster as large as possible, otherwise find size of
681 * existing cluster.
682 */
683 if ((vp->v_type == VREG) &&
54078292
MD
684 bp->b_loffset + lblocksize != filesize &&
685 (bp->b_bio2.bio_offset == NOOFFSET) &&
686 (VOP_BMAP(vp, loffset, NULL, &bp->b_bio2.bio_offset, &maxclen, NULL) ||
687 bp->b_bio2.bio_offset == NOOFFSET)) {
984263bc
MD
688 bawrite(bp);
689 vp->v_clen = 0;
54078292
MD
690 vp->v_lasta = bp->b_bio2.bio_offset;
691 vp->v_cstart = loffset + lblocksize;
692 vp->v_lastw = loffset;
984263bc
MD
693 return;
694 }
54078292
MD
695 if (maxclen > lblocksize)
696 vp->v_clen = maxclen - lblocksize;
697 else
698 vp->v_clen = 0;
699 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
700 vp->v_cstart = loffset + lblocksize;
984263bc
MD
701 bawrite(bp);
702 } else { /* Wait for rest of cluster */
54078292 703 vp->v_cstart = loffset;
984263bc
MD
704 bdwrite(bp);
705 }
54078292 706 } else if (loffset == vp->v_cstart + vp->v_clen) {
984263bc
MD
707 /*
708 * At end of cluster, write it out if seqcount tells us we
709 * are operating sequentially, otherwise let the buf or
710 * update daemon handle it.
711 */
712 bdwrite(bp);
713 if (seqcount > 1)
54078292
MD
714 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
715 vp->v_clen + lblocksize);
984263bc 716 vp->v_clen = 0;
54078292 717 vp->v_cstart = loffset + lblocksize;
984263bc
MD
718 } else if (vm_page_count_severe()) {
719 /*
720 * We are low on memory, get it going NOW
721 */
722 bawrite(bp);
723 } else {
724 /*
725 * In the middle of a cluster, so just delay the I/O for now.
726 */
727 bdwrite(bp);
728 }
54078292
MD
729 vp->v_lastw = loffset;
730 vp->v_lasta = bp->b_bio2.bio_offset;
984263bc
MD
731}
732
733
734/*
735 * This is an awful lot like cluster_rbuild...wish they could be combined.
736 * The last lbn argument is the current block on which I/O is being
737 * performed. Check to see that it doesn't fall in the middle of
738 * the current block (if last_bp == NULL).
739 */
740int
54078292 741cluster_wbuild(struct vnode *vp, int size, off_t start_loffset, int bytes)
984263bc
MD
742{
743 struct buf *bp, *tbp;
e43a034f 744 int i, j;
984263bc 745 int totalwritten = 0;
984263bc 746
54078292 747 while (bytes > 0) {
e43a034f 748 crit_enter();
984263bc
MD
749 /*
750 * If the buffer is not delayed-write (i.e. dirty), or it
751 * is delayed-write but either locked or inval, it cannot
752 * partake in the clustered write.
753 */
54078292 754 if (((tbp = findblk(vp, start_loffset)) == NULL) ||
984263bc
MD
755 ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
756 BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
54078292
MD
757 start_loffset += size;
758 bytes -= size;
e43a034f 759 crit_exit();
984263bc
MD
760 continue;
761 }
762 bremfree(tbp);
10f3fee5 763 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
e43a034f 764 crit_exit();
984263bc
MD
765
766 /*
767 * Extra memory in the buffer, punt on this buffer.
768 * XXX we could handle this in most cases, but we would
769 * have to push the extra memory down to after our max
770 * possible cluster size and then potentially pull it back
771 * up if the cluster was terminated prematurely--too much
772 * hassle.
773 */
774 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
775 (tbp->b_bcount != tbp->b_bufsize) ||
776 (tbp->b_bcount != size) ||
54078292 777 (bytes == size) ||
984263bc
MD
778 ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
779 totalwritten += tbp->b_bufsize;
780 bawrite(tbp);
54078292
MD
781 start_loffset += size;
782 bytes -= size;
984263bc
MD
783 continue;
784 }
785
786 /*
9a71d53f
MD
787 * Set up the pbuf. Track our append point with b_bcount
788 * and b_bufsize. b_bufsize is not used by the device but
789 * our caller uses it to loop clusters and we use it to
790 * detect a premature EOF on the block device.
984263bc 791 */
984263bc
MD
792 bp->b_bcount = 0;
793 bp->b_bufsize = 0;
54f51aeb 794 bp->b_xio.xio_npages = 0;
81b5c339 795 bp->b_loffset = tbp->b_loffset;
54078292 796 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
984263bc
MD
797
798 /*
799 * We are synthesizing a buffer out of vm_page_t's, but
800 * if the block size is not page aligned then the starting
801 * address may not be either. Inherit the b_data offset
802 * from the original buffer.
803 */
804 bp->b_data = (char *)((vm_offset_t)bp->b_data |
805 ((vm_offset_t)tbp->b_data & PAGE_MASK));
10f3fee5 806 bp->b_flags &= ~B_ERROR;
4414f2c9 807 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
984263bc 808 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN));
81b5c339
MD
809 bp->b_bio1.bio_done = cluster_callback;
810 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
811 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
984263bc
MD
812 /*
813 * From this location in the file, scan forward to see
814 * if there are buffers with adjacent data that need to
815 * be written as well.
816 */
54078292 817 for (i = 0; i < bytes; (i += size), (start_loffset += size)) {
984263bc 818 if (i != 0) { /* If not the first buffer */
e43a034f 819 crit_enter();
984263bc
MD
820 /*
821 * If the adjacent data is not even in core it
822 * can't need to be written.
823 */
54078292 824 if ((tbp = findblk(vp, start_loffset)) == NULL) {
e43a034f 825 crit_exit();
984263bc
MD
826 break;
827 }
828
829 /*
830 * If it IS in core, but has different
831 * characteristics, or is locked (which
832 * means it could be undergoing a background
833 * I/O or be in a weird state), then don't
834 * cluster with it.
835 */
836 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
837 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
838 != (B_DELWRI | B_CLUSTEROK |
839 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
840 (tbp->b_flags & B_LOCKED) ||
984263bc 841 BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
e43a034f 842 crit_exit();
984263bc
MD
843 break;
844 }
845
846 /*
847 * Check that the combined cluster
848 * would make sense with regard to pages
849 * and would not be too large
850 */
851 if ((tbp->b_bcount != size) ||
54078292
MD
852 ((bp->b_bio2.bio_offset + i) !=
853 tbp->b_bio2.bio_offset) ||
54f51aeb 854 ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
984263bc
MD
855 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
856 BUF_UNLOCK(tbp);
e43a034f 857 crit_exit();
984263bc
MD
858 break;
859 }
860 /*
861 * Ok, it's passed all the tests,
862 * so remove it from the free list
863 * and mark it busy. We will use it.
864 */
865 bremfree(tbp);
10f3fee5 866 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
e43a034f 867 crit_exit();
984263bc 868 } /* end of code for non-first buffers only */
81b5c339 869
984263bc
MD
870 /*
871 * If the IO is via the VM then we do some
872 * special VM hackery (yuck). Since the buffer's
873 * block size may not be page-aligned it is possible
874 * for a page to be shared between two buffers. We
875 * have to get rid of the duplication when building
876 * the cluster.
877 */
878 if (tbp->b_flags & B_VMIO) {
879 vm_page_t m;
880
881 if (i != 0) { /* if not first buffer */
54078292 882 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
54f51aeb 883 m = tbp->b_xio.xio_pages[j];
984263bc
MD
884 if (m->flags & PG_BUSY) {
885 bqrelse(tbp);
886 goto finishcluster;
887 }
888 }
889 }
890
54078292 891 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
54f51aeb 892 m = tbp->b_xio.xio_pages[j];
984263bc
MD
893 vm_page_io_start(m);
894 vm_object_pip_add(m->object, 1);
54f51aeb
HP
895 if ((bp->b_xio.xio_npages == 0) ||
896 (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
897 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
898 bp->b_xio.xio_npages++;
984263bc
MD
899 }
900 }
901 }
902 bp->b_bcount += size;
903 bp->b_bufsize += size;
904
e43a034f 905 crit_enter();
984263bc 906 bundirty(tbp);
10f3fee5 907 tbp->b_flags &= ~B_ERROR;
984263bc 908 tbp->b_flags |= B_ASYNC;
10f3fee5 909 tbp->b_cmd = BUF_CMD_WRITE;
e43a034f 910 crit_exit();
984263bc 911 BUF_KERNPROC(tbp);
81b5c339 912 cluster_append(&bp->b_bio1, tbp);
2aee763b
MD
913
914 /*
915 * check for latent dependencies to be handled
916 */
917 if (LIST_FIRST(&tbp->b_dep) != NULL && bioops.io_start)
918 (*bioops.io_start)(tbp);
919
984263bc
MD
920 }
921 finishcluster:
922 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
54f51aeb 923 (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
312dcd01 924 if (bp->b_bufsize > bp->b_kvasize) {
984263bc 925 panic(
54078292 926 "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
984263bc 927 bp->b_bufsize, bp->b_kvasize);
312dcd01 928 }
984263bc
MD
929 totalwritten += bp->b_bufsize;
930 bp->b_dirtyoff = 0;
931 bp->b_dirtyend = bp->b_bufsize;
10f3fee5
MD
932 bp->b_flags |= B_ASYNC;
933 bp->b_cmd = BUF_CMD_WRITE;
934 vfs_busy_pages(vp, bp);
a8f169e2
MD
935 bp->b_runningbufspace = bp->b_bufsize;
936 runningbufspace += bp->b_runningbufspace;
937 BUF_KERNPROC(bp); /* B_ASYNC */
938 vn_strategy(vp, &bp->b_bio1);
984263bc 939
54078292 940 bytes -= i;
984263bc
MD
941 }
942 return totalwritten;
943}
944
945/*
946 * Collect together all the buffers in a cluster.
947 * Plus add one additional buffer.
948 */
949static struct cluster_save *
54078292 950cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int lblocksize)
984263bc
MD
951{
952 struct cluster_save *buflist;
953 struct buf *bp;
54078292 954 off_t loffset;
984263bc
MD
955 int i, len;
956
54078292 957 len = (int)(vp->v_lastw - vp->v_cstart + lblocksize) / lblocksize;
984263bc 958 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
54078292 959 M_SEGMENT, M_WAITOK);
984263bc
MD
960 buflist->bs_nchildren = 0;
961 buflist->bs_children = (struct buf **) (buflist + 1);
54078292
MD
962 for (loffset = vp->v_cstart, i = 0; i < len; (loffset += lblocksize), i++) {
963 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
984263bc 964 buflist->bs_children[i] = bp;
54078292
MD
965 if (bp->b_bio2.bio_offset == NOOFFSET) {
966 VOP_BMAP(bp->b_vp, bp->b_loffset, NULL,
967 &bp->b_bio2.bio_offset, NULL, NULL);
968 }
984263bc
MD
969 }
970 buflist->bs_children[i] = bp = last_bp;
54078292
MD
971 if (bp->b_bio2.bio_offset == NOOFFSET) {
972 VOP_BMAP(bp->b_vp, bp->b_loffset, NULL,
973 &bp->b_bio2.bio_offset, NULL, NULL);
974 }
984263bc
MD
975 buflist->bs_nchildren = i + 1;
976 return (buflist);
977}
81b5c339
MD
978
979void
980cluster_append(struct bio *bio, struct buf *tbp)
981{
982 tbp->b_cluster_next = NULL;
983 if (bio->bio_caller_info1.cluster_head == NULL) {
984 bio->bio_caller_info1.cluster_head = tbp;
985 bio->bio_caller_info2.cluster_tail = tbp;
986 } else {
987 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
988 bio->bio_caller_info2.cluster_tail = tbp;
989 }
990}
991