xref: /dflybsd-src/sys/kern/vfs_cluster.c (revision 564dee328e7789515768448325c8268f2e00093f)
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 1.18 2006/03/10 17:51:54 dillon Exp $
38  */
39 
40 #include "opt_debug_cluster.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/proc.h>
46 #include <sys/buf.h>
47 #include <sys/vnode.h>
48 #include <sys/malloc.h>
49 #include <sys/mount.h>
50 #include <sys/resourcevar.h>
51 #include <sys/vmmeter.h>
52 #include <vm/vm.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_page.h>
55 #include <sys/sysctl.h>
56 #include <sys/buf2.h>
57 #include <vm/vm_page2.h>
58 
59 #if defined(CLUSTERDEBUG)
60 #include <sys/sysctl.h>
61 static int	rcluster= 0;
62 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
63 #endif
64 
65 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");
66 
67 static struct cluster_save *
68 	cluster_collectbufs (struct vnode *vp, struct buf *last_bp);
69 static struct buf *
70 	cluster_rbuild (struct vnode *vp, u_quad_t filesize, daddr_t lbn,
71 			    daddr_t blkno, long size, int run, struct buf *fbp);
72 static void cluster_callback (struct bio *);
73 
74 
75 static int write_behind = 1;
76 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
77 
78 extern vm_page_t	bogus_page;
79 
80 extern int cluster_pbuf_freecnt;
81 
82 /*
83  * Maximum number of blocks for read-ahead.
84  */
85 #define MAXRA 32
86 
87 /*
88  * This replaces bread.
89  */
90 int
91 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno,
92 	long size, long totread, int seqcount, struct buf **bpp)
93 {
94 	struct buf *bp, *rbp, *reqbp;
95 	daddr_t blkno, origblkno;
96 	int error, num_ra;
97 	int i;
98 	int maxra, racluster;
99 	long origtotread;
100 
101 	error = 0;
102 
103 	/*
104 	 * Try to limit the amount of read-ahead by a few
105 	 * ad-hoc parameters.  This needs work!!!
106 	 */
107 	racluster = vp->v_mount->mnt_iosize_max / size;
108 	maxra = 2 * racluster + (totread / size);
109 	if (maxra > MAXRA)
110 		maxra = MAXRA;
111 	if (maxra > nbuf/8)
112 		maxra = nbuf/8;
113 
114 	/*
115 	 * get the requested block
116 	 */
117 	*bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0);
118 	origblkno = lblkno;
119 	origtotread = totread;
120 
121 	/*
122 	 * if it is in the cache, then check to see if the reads have been
123 	 * sequential.  If they have, then try some read-ahead, otherwise
124 	 * back-off on prospective read-aheads.
125 	 */
126 	if (bp->b_flags & B_CACHE) {
127 		if (!seqcount) {
128 			return 0;
129 		} else if ((bp->b_flags & B_RAM) == 0) {
130 			return 0;
131 		} else {
132 			struct buf *tbp;
133 			bp->b_flags &= ~B_RAM;
134 			/*
135 			 * We do the crit here so that there is no window
136 			 * between the findblk and the b_usecount increment
137 			 * below.  We opt to keep the crit out of the loop
138 			 * for efficiency.
139 			 */
140 			crit_enter();
141 			for (i = 1; i < maxra; i++) {
142 
143 				if (!(tbp = findblk(vp, lblkno+i))) {
144 					break;
145 				}
146 
147 				/*
148 				 * Set another read-ahead mark so we know
149 				 * to check again.
150 				 */
151 				if (((i % racluster) == (racluster - 1)) ||
152 					(i == (maxra - 1)))
153 					tbp->b_flags |= B_RAM;
154 			}
155 			crit_exit();
156 			if (i >= maxra) {
157 				return 0;
158 			}
159 			lblkno += i;
160 		}
161 		reqbp = bp = NULL;
162 	} else {
163 		off_t firstread = bp->b_loffset;
164 
165 		KASSERT(firstread != NOOFFSET,
166 			("cluster_read: no buffer offset"));
167 		if (firstread + totread > filesize)
168 			totread = filesize - firstread;
169 		if (totread > size) {
170 			int nblks = 0;
171 			int ncontigafter;
172 			while (totread > 0) {
173 				nblks++;
174 				totread -= size;
175 			}
176 			if (nblks == 1)
177 				goto single_block_read;
178 			if (nblks > racluster)
179 				nblks = racluster;
180 
181 	    		error = VOP_BMAP(vp, lblkno, NULL,
182 					 &blkno, &ncontigafter, NULL);
183 			if (error)
184 				goto single_block_read;
185 			if (blkno == -1)
186 				goto single_block_read;
187 			if (ncontigafter == 0)
188 				goto single_block_read;
189 			if (ncontigafter + 1 < nblks)
190 				nblks = ncontigafter + 1;
191 
192 			bp = cluster_rbuild(vp, filesize, lblkno,
193 					    blkno, size, nblks, bp);
194 			lblkno += (bp->b_bufsize / size);
195 		} else {
196 single_block_read:
197 			/*
198 			 * if it isn't in the cache, then get a chunk from
199 			 * disk if sequential, otherwise just get the block.
200 			 */
201 			bp->b_flags |= B_READ | B_RAM;
202 			lblkno += 1;
203 		}
204 	}
205 
206 	/*
207 	 * If we have been doing sequential I/O, then do some read-ahead.
208 	 */
209 	rbp = NULL;
210 	if (seqcount &&
211 	    lblkno < origblkno + seqcount &&
212 	    (u_quad_t)(lblkno + 1) * size <= filesize
213 	) {
214 		rbp = getblk(vp, lblkno, size, 0, 0);
215 		if ((rbp->b_flags & B_CACHE) == 0) {
216 			int nblksread;
217 			int ntoread;
218 
219 			error = VOP_BMAP(vp, lblkno, NULL,
220 					 &blkno, &num_ra, NULL);
221 			if (error || blkno == (daddr_t)-1) {
222 				rbp->b_flags &= ~(B_ASYNC | B_READ);
223 				brelse(rbp);
224 				rbp = NULL;
225 				goto no_read_ahead;
226 			}
227 			ntoread = num_ra + 1;
228 			nblksread = (origtotread + size - 1) / size;
229 			if (seqcount < nblksread)
230 				seqcount = nblksread;
231 			if (seqcount < ntoread)
232 				ntoread = seqcount;
233 
234 			rbp->b_flags |= B_READ | B_ASYNC | B_RAM;
235 			if (num_ra) {
236 				rbp = cluster_rbuild(vp, filesize, lblkno,
237 						     blkno, size, ntoread, rbp);
238 			} else {
239 				rbp->b_bio2.bio_blkno = blkno;
240 			}
241 		}
242 	}
243 no_read_ahead:
244 
245 	/*
246 	 * Handle the synchronous read
247 	 */
248 	if (bp) {
249 #if defined(CLUSTERDEBUG)
250 		if (rcluster)
251 			printf("S(%ld,%ld,%d) ",
252 			    (long)bp->b_lblkno, bp->b_bcount, seqcount);
253 #endif
254 		if ((bp->b_flags & B_CLUSTER) == 0) {
255 			vfs_busy_pages(bp, 0);
256 		}
257 		bp->b_flags &= ~(B_ERROR|B_INVAL);
258 		if ((bp->b_flags & B_ASYNC) || bp->b_bio1.bio_done != NULL)
259 			BUF_KERNPROC(bp);
260 		vn_strategy(vp, &bp->b_bio1);
261 		error = bp->b_error;
262 	}
263 
264 	/*
265 	 * And if we have read-aheads, do them too
266 	 */
267 	if (rbp) {
268 		if (error) {
269 			rbp->b_flags &= ~(B_ASYNC | B_READ);
270 			brelse(rbp);
271 		} else if (rbp->b_flags & B_CACHE) {
272 			rbp->b_flags &= ~(B_ASYNC | B_READ);
273 			bqrelse(rbp);
274 		} else {
275 #if defined(CLUSTERDEBUG)
276 			if (rcluster) {
277 				if (bp)
278 					printf("A+(%ld,%ld,%ld,%d) ",
279 					    (long)rbp->b_lblkno, rbp->b_bcount,
280 					    (long)(rbp->b_lblkno - origblkno),
281 					    seqcount);
282 				else
283 					printf("A(%ld,%ld,%ld,%d) ",
284 					    (long)rbp->b_lblkno, rbp->b_bcount,
285 					    (long)(rbp->b_lblkno - origblkno),
286 					    seqcount);
287 			}
288 #endif
289 
290 			if ((rbp->b_flags & B_CLUSTER) == 0) {
291 				vfs_busy_pages(rbp, 0);
292 			}
293 			rbp->b_flags &= ~(B_ERROR|B_INVAL);
294 			if ((rbp->b_flags & B_ASYNC) || rbp->b_bio1.bio_done != NULL)
295 				BUF_KERNPROC(rbp);
296 			vn_strategy(vp, &rbp->b_bio1);
297 		}
298 	}
299 	if (reqbp)
300 		return (biowait(reqbp));
301 	else
302 		return (error);
303 }
304 
305 /*
306  * If blocks are contiguous on disk, use this to provide clustered
307  * read ahead.  We will read as many blocks as possible sequentially
308  * and then parcel them up into logical blocks in the buffer hash table.
309  */
310 static struct buf *
311 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
312 	daddr_t blkno, long size, int run, struct buf *fbp)
313 {
314 	struct buf *bp, *tbp;
315 	daddr_t bn;
316 	int i, inc, j;
317 
318 	KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
319 	    ("cluster_rbuild: size %ld != filesize %ld\n",
320 	    size, vp->v_mount->mnt_stat.f_iosize));
321 
322 	/*
323 	 * avoid a division
324 	 */
325 	while ((u_quad_t) size * (lbn + run) > filesize) {
326 		--run;
327 	}
328 
329 	tbp = fbp;
330 	tbp->b_flags |= B_READ;
331 	tbp->b_bio2.bio_blkno = blkno;
332 	if( (tbp->b_flags & B_MALLOC) ||
333 		((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
334 		return tbp;
335 
336 	bp = trypbuf(&cluster_pbuf_freecnt);
337 	if (bp == 0)
338 		return tbp;
339 
340 	/*
341 	 * We are synthesizing a buffer out of vm_page_t's, but
342 	 * if the block size is not page aligned then the starting
343 	 * address may not be either.  Inherit the b_data offset
344 	 * from the original buffer.
345 	 */
346 	bp->b_data = (char *)((vm_offset_t)bp->b_data |
347 	    ((vm_offset_t)tbp->b_data & PAGE_MASK));
348 	bp->b_flags = B_ASYNC | B_READ | B_CLUSTER | B_VMIO;
349 	bp->b_bio1.bio_done = cluster_callback;
350 	bp->b_bio1.bio_caller_info1.cluster_head = NULL;
351 	bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
352 	bp->b_lblkno = lbn;
353 	bp->b_loffset = tbp->b_loffset;
354 	bp->b_bio2.bio_blkno = (daddr_t)-1;
355 	KASSERT(bp->b_loffset != NOOFFSET,
356 		("cluster_rbuild: no buffer offset"));
357 	pbgetvp(vp, bp);
358 
359 	bp->b_bcount = 0;
360 	bp->b_bufsize = 0;
361 	bp->b_xio.xio_npages = 0;
362 
363 	inc = btodb(size);
364 	for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
365 		if (i != 0) {
366 			if ((bp->b_xio.xio_npages * PAGE_SIZE) +
367 			    round_page(size) > vp->v_mount->mnt_iosize_max) {
368 				break;
369 			}
370 
371 			/*
372 			 * Shortcut some checks and try to avoid buffers that
373 			 * would block in the lock.  The same checks have to
374 			 * be made again after we officially get the buffer.
375 			 */
376 			if ((tbp = findblk(vp, lbn + i)) != NULL) {
377 				if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
378 					break;
379 				BUF_UNLOCK(tbp);
380 
381 				for (j = 0; j < tbp->b_xio.xio_npages; j++) {
382 					if (tbp->b_xio.xio_pages[j]->valid)
383 						break;
384 				}
385 
386 				if (j != tbp->b_xio.xio_npages)
387 					break;
388 
389 				if (tbp->b_bcount != size)
390 					break;
391 			}
392 
393 			tbp = getblk(vp, lbn + i, size, 0, 0);
394 
395 			/*
396 			 * Stop scanning if the buffer is fuly valid
397 			 * (marked B_CACHE), or locked (may be doing a
398 			 * background write), or if the buffer is not
399 			 * VMIO backed.  The clustering code can only deal
400 			 * with VMIO-backed buffers.
401 			 */
402 			if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
403 			    (tbp->b_flags & B_VMIO) == 0) {
404 				bqrelse(tbp);
405 				break;
406 			}
407 
408 			/*
409 			 * The buffer must be completely invalid in order to
410 			 * take part in the cluster.  If it is partially valid
411 			 * then we stop.
412 			 */
413 			for (j = 0;j < tbp->b_xio.xio_npages; j++) {
414 				if (tbp->b_xio.xio_pages[j]->valid)
415 					break;
416 			}
417 			if (j != tbp->b_xio.xio_npages) {
418 				bqrelse(tbp);
419 				break;
420 			}
421 
422 			/*
423 			 * Set a read-ahead mark as appropriate
424 			 */
425 			if (i == 1 || i == (run - 1))
426 				tbp->b_flags |= B_RAM;
427 
428 			/*
429 			 * Set the buffer up for an async read (XXX should
430 			 * we do this only if we do not wind up brelse()ing?).
431 			 * Set the block number if it isn't set, otherwise
432 			 * if it is make sure it matches the block number we
433 			 * expect.
434 			 */
435 			tbp->b_flags |= B_READ | B_ASYNC;
436 			if (tbp->b_bio2.bio_blkno == (daddr_t)-1) {
437 				tbp->b_bio2.bio_blkno = bn;
438 			} else if (tbp->b_bio2.bio_blkno != bn) {
439 				brelse(tbp);
440 				break;
441 			}
442 		}
443 		/*
444 		 * XXX fbp from caller may not be B_ASYNC, but we are going
445 		 * to biodone() it in cluster_callback() anyway
446 		 */
447 		BUF_KERNPROC(tbp);
448 		cluster_append(&bp->b_bio1, tbp);
449 		for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
450 			vm_page_t m;
451 			m = tbp->b_xio.xio_pages[j];
452 			vm_page_io_start(m);
453 			vm_object_pip_add(m->object, 1);
454 			if ((bp->b_xio.xio_npages == 0) ||
455 				(bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
456 				bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
457 				bp->b_xio.xio_npages++;
458 			}
459 			if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
460 				tbp->b_xio.xio_pages[j] = bogus_page;
461 		}
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 	for (j = 0; j < bp->b_xio.xio_npages; j++) {
483 		if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
484 		    VM_PAGE_BITS_ALL) {
485 			bp->b_xio.xio_pages[j] = bogus_page;
486 		}
487 	}
488 	if (bp->b_bufsize > bp->b_kvasize)
489 		panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)",
490 		    bp->b_bufsize, bp->b_kvasize);
491 	bp->b_kvasize = bp->b_bufsize;
492 
493 	pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
494 		(vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
495 	return (bp);
496 }
497 
498 /*
499  * Cleanup after a clustered read or write.
500  * This is complicated by the fact that any of the buffers might have
501  * extra memory (if there were no empty buffer headers at allocbuf time)
502  * that we will need to shift around.
503  *
504  * The returned bio is &bp->b_bio1
505  */
506 void
507 cluster_callback(struct bio *bio)
508 {
509 	struct buf *bp = bio->bio_buf;
510 	struct buf *tbp;
511 	int error = 0;
512 
513 	/*
514 	 * Must propogate errors to all the components.
515 	 */
516 	if (bp->b_flags & B_ERROR)
517 		error = bp->b_error;
518 
519 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
520 	/*
521 	 * Move memory from the large cluster buffer into the component
522 	 * buffers and mark IO as done on these.  Since the memory map
523 	 * is the same, no actual copying is required.
524 	 */
525 	while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
526 		bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
527 		if (error) {
528 			tbp->b_flags |= B_ERROR;
529 			tbp->b_error = error;
530 		} else {
531 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
532 			tbp->b_flags &= ~(B_ERROR|B_INVAL);
533 			/*
534 			 * XXX the bdwrite()/bqrelse() issued during
535 			 * cluster building clears B_RELBUF (see bqrelse()
536 			 * comment).  If direct I/O was specified, we have
537 			 * to restore it here to allow the buffer and VM
538 			 * to be freed.
539 			 */
540 			if (tbp->b_flags & B_DIRECT)
541 				tbp->b_flags |= B_RELBUF;
542 		}
543 		biodone(&tbp->b_bio1);
544 	}
545 	relpbuf(bp, &cluster_pbuf_freecnt);
546 }
547 
548 /*
549  *	cluster_wbuild_wb:
550  *
551  *	Implement modified write build for cluster.
552  *
553  *		write_behind = 0	write behind disabled
554  *		write_behind = 1	write behind normal (default)
555  *		write_behind = 2	write behind backed-off
556  */
557 
558 static __inline int
559 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
560 {
561 	int r = 0;
562 
563 	switch(write_behind) {
564 	case 2:
565 		if (start_lbn < len)
566 			break;
567 		start_lbn -= len;
568 		/* fall through */
569 	case 1:
570 		r = cluster_wbuild(vp, size, start_lbn, len);
571 		/* fall through */
572 	default:
573 		/* fall through */
574 		break;
575 	}
576 	return(r);
577 }
578 
579 /*
580  * Do clustered write for FFS.
581  *
582  * Three cases:
583  *	1. Write is not sequential (write asynchronously)
584  *	Write is sequential:
585  *	2.	beginning of cluster - begin cluster
586  *	3.	middle of a cluster - add to cluster
587  *	4.	end of a cluster - asynchronously write cluster
588  */
589 void
590 cluster_write(struct buf *bp, u_quad_t filesize, int seqcount)
591 {
592 	struct vnode *vp;
593 	daddr_t lbn;
594 	int maxclen, cursize;
595 	int lblocksize;
596 	int async;
597 
598 	vp = bp->b_vp;
599 	if (vp->v_type == VREG) {
600 		async = vp->v_mount->mnt_flag & MNT_ASYNC;
601 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
602 	} else {
603 		async = 0;
604 		lblocksize = bp->b_bufsize;
605 	}
606 	lbn = bp->b_lblkno;
607 	KASSERT(bp->b_loffset != NOOFFSET,
608 		("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_bio2.bio_blkno == (daddr_t)-1 ||
616 	    (bp->b_bio2.bio_blkno != vp->v_lasta + btodb(lblocksize))) {
617 		maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
618 		if (vp->v_clen != 0) {
619 			/*
620 			 * Next block is not sequential.
621 			 *
622 			 * If we are not writing at end of file, the process
623 			 * seeked to another point in the file since its last
624 			 * write, or we have reached our maximum cluster size,
625 			 * then push the previous cluster. Otherwise try
626 			 * reallocating to make it sequential.
627 			 *
628 			 * Change to algorithm: only push previous cluster if
629 			 * it was sequential from the point of view of the
630 			 * seqcount heuristic, otherwise leave the buffer
631 			 * intact so we can potentially optimize the I/O
632 			 * later on in the buf_daemon or update daemon
633 			 * flush.
634 			 */
635 			cursize = vp->v_lastw - vp->v_cstart + 1;
636 			if (((u_quad_t) bp->b_loffset + lblocksize) != filesize ||
637 			    lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
638 				if (!async && seqcount > 0) {
639 					cluster_wbuild_wb(vp, lblocksize,
640 						vp->v_cstart, cursize);
641 				}
642 			} else {
643 				struct buf **bpp, **endbp;
644 				struct cluster_save *buflist;
645 
646 				buflist = cluster_collectbufs(vp, bp);
647 				endbp = &buflist->bs_children
648 				    [buflist->bs_nchildren - 1];
649 				if (VOP_REALLOCBLKS(vp, buflist)) {
650 					/*
651 					 * Failed, push the previous cluster
652 					 * if *really* writing sequentially
653 					 * in the logical file (seqcount > 1),
654 					 * otherwise delay it in the hopes that
655 					 * the low level disk driver can
656 					 * optimize the write ordering.
657 					 */
658 					for (bpp = buflist->bs_children;
659 					     bpp < endbp; bpp++)
660 						brelse(*bpp);
661 					free(buflist, M_SEGMENT);
662 					if (seqcount > 1) {
663 						cluster_wbuild_wb(vp,
664 						    lblocksize, vp->v_cstart,
665 						    cursize);
666 					}
667 				} else {
668 					/*
669 					 * Succeeded, keep building cluster.
670 					 */
671 					for (bpp = buflist->bs_children;
672 					     bpp <= endbp; bpp++)
673 						bdwrite(*bpp);
674 					free(buflist, M_SEGMENT);
675 					vp->v_lastw = lbn;
676 					vp->v_lasta = bp->b_bio2.bio_blkno;
677 					return;
678 				}
679 			}
680 		}
681 		/*
682 		 * Consider beginning a cluster. If at end of file, make
683 		 * cluster as large as possible, otherwise find size of
684 		 * existing cluster.
685 		 */
686 		if ((vp->v_type == VREG) &&
687 			((u_quad_t) bp->b_loffset + lblocksize) != filesize &&
688 		    (bp->b_bio2.bio_blkno == (daddr_t)-1) &&
689 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_bio2.bio_blkno, &maxclen, NULL) ||
690 		     bp->b_bio2.bio_blkno == (daddr_t)-1)) {
691 			bawrite(bp);
692 			vp->v_clen = 0;
693 			vp->v_lasta = bp->b_bio2.bio_blkno;
694 			vp->v_cstart = lbn + 1;
695 			vp->v_lastw = lbn;
696 			return;
697 		}
698 		vp->v_clen = maxclen;
699 		if (!async && maxclen == 0) {	/* I/O not contiguous */
700 			vp->v_cstart = lbn + 1;
701 			bawrite(bp);
702 		} else {	/* Wait for rest of cluster */
703 			vp->v_cstart = lbn;
704 			bdwrite(bp);
705 		}
706 	} else if (lbn == vp->v_cstart + vp->v_clen) {
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)
714 			cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
715 		vp->v_clen = 0;
716 		vp->v_cstart = lbn + 1;
717 	} else if (vm_page_count_severe()) {
718 		/*
719 		 * We are low on memory, get it going NOW
720 		 */
721 		bawrite(bp);
722 	} else {
723 		/*
724 		 * In the middle of a cluster, so just delay the I/O for now.
725 		 */
726 		bdwrite(bp);
727 	}
728 	vp->v_lastw = lbn;
729 	vp->v_lasta = bp->b_bio2.bio_blkno;
730 }
731 
732 
733 /*
734  * This is an awful lot like cluster_rbuild...wish they could be combined.
735  * The last lbn argument is the current block on which I/O is being
736  * performed.  Check to see that it doesn't fall in the middle of
737  * the current block (if last_bp == NULL).
738  */
739 int
740 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len)
741 {
742 	struct buf *bp, *tbp;
743 	int i, j;
744 	int totalwritten = 0;
745 	int dbsize = btodb(size);
746 
747 	while (len > 0) {
748 		crit_enter();
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 		 */
754 		if (((tbp = findblk(vp, start_lbn)) == NULL) ||
755 		  ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
756 		  BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
757 			++start_lbn;
758 			--len;
759 			crit_exit();
760 			continue;
761 		}
762 		bremfree(tbp);
763 		tbp->b_flags &= ~B_DONE;
764 		crit_exit();
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) ||
777 		  (len == 1) ||
778 		  ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
779 			totalwritten += tbp->b_bufsize;
780 			bawrite(tbp);
781 			++start_lbn;
782 			--len;
783 			continue;
784 		}
785 
786 		/*
787 		 * We got a pbuf to make the cluster in.
788 		 * so initialise it.
789 		 */
790 		bp->b_bcount = 0;
791 		bp->b_bufsize = 0;
792 		bp->b_xio.xio_npages = 0;
793 		bp->b_lblkno = tbp->b_lblkno;
794 		bp->b_loffset = tbp->b_loffset;
795 		bp->b_bio2.bio_blkno = tbp->b_bio2.bio_blkno;
796 
797 		/*
798 		 * We are synthesizing a buffer out of vm_page_t's, but
799 		 * if the block size is not page aligned then the starting
800 		 * address may not be either.  Inherit the b_data offset
801 		 * from the original buffer.
802 		 */
803 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
804 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
805 		bp->b_flags |= B_CLUSTER |
806 			(tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN));
807 		bp->b_bio1.bio_done = cluster_callback;
808 		bp->b_bio1.bio_caller_info1.cluster_head = NULL;
809 		bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
810 		pbgetvp(vp, bp);
811 		/*
812 		 * From this location in the file, scan forward to see
813 		 * if there are buffers with adjacent data that need to
814 		 * be written as well.
815 		 */
816 		for (i = 0; i < len; ++i, ++start_lbn) {
817 			if (i != 0) { /* If not the first buffer */
818 				crit_enter();
819 				/*
820 				 * If the adjacent data is not even in core it
821 				 * can't need to be written.
822 				 */
823 				if ((tbp = findblk(vp, start_lbn)) == NULL) {
824 					crit_exit();
825 					break;
826 				}
827 
828 				/*
829 				 * If it IS in core, but has different
830 				 * characteristics, or is locked (which
831 				 * means it could be undergoing a background
832 				 * I/O or be in a weird state), then don't
833 				 * cluster with it.
834 				 */
835 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
836 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
837 				  != (B_DELWRI | B_CLUSTEROK |
838 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
839 				    (tbp->b_flags & B_LOCKED) ||
840 				    BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
841 					crit_exit();
842 					break;
843 				}
844 
845 				/*
846 				 * Check that the combined cluster
847 				 * would make sense with regard to pages
848 				 * and would not be too large
849 				 */
850 				if ((tbp->b_bcount != size) ||
851 				  ((bp->b_bio2.bio_blkno + (dbsize * i)) !=
852 				    tbp->b_bio2.bio_blkno) ||
853 				  ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
854 				    (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
855 					BUF_UNLOCK(tbp);
856 					crit_exit();
857 					break;
858 				}
859 				/*
860 				 * Ok, it's passed all the tests,
861 				 * so remove it from the free list
862 				 * and mark it busy. We will use it.
863 				 */
864 				bremfree(tbp);
865 				tbp->b_flags &= ~B_DONE;
866 				crit_exit();
867 			} /* end of code for non-first buffers only */
868 
869 			/*
870 			 * If the IO is via the VM then we do some
871 			 * special VM hackery (yuck).  Since the buffer's
872 			 * block size may not be page-aligned it is possible
873 			 * for a page to be shared between two buffers.  We
874 			 * have to get rid of the duplication when building
875 			 * the cluster.
876 			 */
877 			if (tbp->b_flags & B_VMIO) {
878 				vm_page_t m;
879 
880 				if (i != 0) { /* if not first buffer */
881 					for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
882 						m = tbp->b_xio.xio_pages[j];
883 						if (m->flags & PG_BUSY) {
884 							bqrelse(tbp);
885 							goto finishcluster;
886 						}
887 					}
888 				}
889 
890 				for (j = 0; j < tbp->b_xio.xio_npages; j += 1) {
891 					m = tbp->b_xio.xio_pages[j];
892 					vm_page_io_start(m);
893 					vm_object_pip_add(m->object, 1);
894 					if ((bp->b_xio.xio_npages == 0) ||
895 					  (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
896 						bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
897 						bp->b_xio.xio_npages++;
898 					}
899 				}
900 			}
901 			bp->b_bcount += size;
902 			bp->b_bufsize += size;
903 
904 			crit_enter();
905 			bundirty(tbp);
906 			tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
907 			tbp->b_flags |= B_ASYNC;
908 			crit_exit();
909 			BUF_KERNPROC(tbp);
910 			cluster_append(&bp->b_bio1, tbp);
911 
912 			/*
913 			 * check for latent dependencies to be handled
914 			 */
915 			if (LIST_FIRST(&tbp->b_dep) != NULL && bioops.io_start)
916 				(*bioops.io_start)(tbp);
917 
918 		}
919 	finishcluster:
920 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
921 			(vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
922 		if (bp->b_bufsize > bp->b_kvasize)
923 			panic(
924 			    "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
925 			    bp->b_bufsize, bp->b_kvasize);
926 		bp->b_kvasize = bp->b_bufsize;
927 		totalwritten += bp->b_bufsize;
928 		bp->b_dirtyoff = 0;
929 		bp->b_dirtyend = bp->b_bufsize;
930 		bawrite(bp);
931 
932 		len -= i;
933 	}
934 	return totalwritten;
935 }
936 
937 /*
938  * Collect together all the buffers in a cluster.
939  * Plus add one additional buffer.
940  */
941 static struct cluster_save *
942 cluster_collectbufs(struct vnode *vp, struct buf *last_bp)
943 {
944 	struct cluster_save *buflist;
945 	struct buf *bp;
946 	daddr_t lbn;
947 	int i, len;
948 
949 	len = vp->v_lastw - vp->v_cstart + 1;
950 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
951 	    M_SEGMENT, M_WAITOK);
952 	buflist->bs_nchildren = 0;
953 	buflist->bs_children = (struct buf **) (buflist + 1);
954 	for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
955 		(void) bread(vp, lbn, last_bp->b_bcount, &bp);
956 		buflist->bs_children[i] = bp;
957 		if (bp->b_bio2.bio_blkno == (daddr_t)-1)
958 			VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_bio2.bio_blkno,
959 				NULL, NULL);
960 	}
961 	buflist->bs_children[i] = bp = last_bp;
962 	if (bp->b_bio2.bio_blkno == (daddr_t)-1)
963 		VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_bio2.bio_blkno,
964 			NULL, NULL);
965 	buflist->bs_nchildren = i + 1;
966 	return (buflist);
967 }
968 
969 void
970 cluster_append(struct bio *bio, struct buf *tbp)
971 {
972 	tbp->b_cluster_next = NULL;
973 	if (bio->bio_caller_info1.cluster_head == NULL) {
974 		bio->bio_caller_info1.cluster_head = tbp;
975 		bio->bio_caller_info2.cluster_tail = tbp;
976 	} else {
977 		bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
978 		bio->bio_caller_info2.cluster_tail = tbp;
979 	}
980 }
981 
982