xref: /freebsd/sys/kern/vfs_cluster.c (revision 1c9ca5858f45f6fe39fd85981bdb296fde48262f)
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  * $Id: vfs_cluster.c,v 1.85 1999/06/29 05:59:43 peter Exp $
37  */
38 
39 #include "opt_debug_cluster.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/proc.h>
45 #include <sys/buf.h>
46 #include <sys/vnode.h>
47 #include <sys/malloc.h>
48 #include <sys/mount.h>
49 #include <sys/resourcevar.h>
50 #include <vm/vm.h>
51 #include <vm/vm_prot.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_page.h>
54 #include <sys/sysctl.h>
55 
56 #if defined(CLUSTERDEBUG)
57 #include <sys/sysctl.h>
58 static int	rcluster= 0;
59 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
60 #endif
61 
62 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer");
63 
64 static struct cluster_save *
65 	cluster_collectbufs __P((struct vnode *vp, struct buf *last_bp));
66 static struct buf *
67 	cluster_rbuild __P((struct vnode *vp, u_quad_t filesize, daddr_t lbn,
68 			    daddr_t blkno, long size, int run, struct buf *fbp));
69 
70 static int write_behind = 1;
71 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
72 
73 extern vm_page_t	bogus_page;
74 
75 extern int cluster_pbuf_freecnt;
76 
77 /*
78  * Maximum number of blocks for read-ahead.
79  */
80 #define MAXRA 32
81 
82 /*
83  * This 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 	daddr_t blkno, origblkno;
98 	int error, num_ra;
99 	int i;
100 	int maxra, racluster;
101 	long origtotread;
102 
103 	error = 0;
104 	if (vp->v_maxio == 0)
105 		vp->v_maxio = DFLTPHYS;
106 
107 	/*
108 	 * Try to limit the amount of read-ahead by a few
109 	 * ad-hoc parameters.  This needs work!!!
110 	 */
111 	racluster = vp->v_maxio/size;
112 	maxra = 2 * racluster + (totread / size);
113 	if (maxra > MAXRA)
114 		maxra = MAXRA;
115 	if (maxra > nbuf/8)
116 		maxra = nbuf/8;
117 
118 	/*
119 	 * get the requested block
120 	 */
121 	*bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0);
122 	origblkno = lblkno;
123 	origtotread = totread;
124 
125 	/*
126 	 * if it is in the cache, then check to see if the reads have been
127 	 * sequential.  If they have, then try some read-ahead, otherwise
128 	 * back-off on prospective read-aheads.
129 	 */
130 	if (bp->b_flags & B_CACHE) {
131 		if (!seqcount) {
132 			return 0;
133 		} else if ((bp->b_flags & B_RAM) == 0) {
134 			return 0;
135 		} else {
136 			int s;
137 			struct buf *tbp;
138 			bp->b_flags &= ~B_RAM;
139 			/*
140 			 * We do the spl here so that there is no window
141 			 * between the incore and the b_usecount increment
142 			 * below.  We opt to keep the spl out of the loop
143 			 * for efficiency.
144 			 */
145 			s = splbio();
146 			for (i = 1; i < maxra; i++) {
147 
148 				if (!(tbp = incore(vp, lblkno+i))) {
149 					break;
150 				}
151 
152 				/*
153 				 * Set another read-ahead mark so we know to check
154 				 * again.
155 				 */
156 				if (((i % racluster) == (racluster - 1)) ||
157 					(i == (maxra - 1)))
158 					tbp->b_flags |= B_RAM;
159 
160 #if 0
161 				if ((tbp->b_usecount < 1) &&
162 					BUF_REFCNT(tbp) == 0 &&
163 					(tbp->b_qindex == QUEUE_LRU)) {
164 					TAILQ_REMOVE(&bufqueues[QUEUE_LRU], tbp, b_freelist);
165 					TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], tbp, b_freelist);
166 				}
167 #endif
168 			}
169 			splx(s);
170 			if (i >= maxra) {
171 				return 0;
172 			}
173 			lblkno += i;
174 		}
175 		reqbp = bp = NULL;
176 	} else {
177 		off_t firstread = bp->b_offset;
178 
179 		KASSERT(bp->b_offset != NOOFFSET,
180 		    ("cluster_read: no buffer offset"));
181 		if (firstread + totread > filesize)
182 			totread = filesize - firstread;
183 		if (totread > size) {
184 			int nblks = 0;
185 			int ncontigafter;
186 			while (totread > 0) {
187 				nblks++;
188 				totread -= size;
189 			}
190 			if (nblks == 1)
191 				goto single_block_read;
192 			if (nblks > racluster)
193 				nblks = racluster;
194 
195 	    		error = VOP_BMAP(vp, lblkno, NULL,
196 				&blkno, &ncontigafter, NULL);
197 			if (error)
198 				goto single_block_read;
199 			if (blkno == -1)
200 				goto single_block_read;
201 			if (ncontigafter == 0)
202 				goto single_block_read;
203 			if (ncontigafter + 1 < nblks)
204 				nblks = ncontigafter + 1;
205 
206 			bp = cluster_rbuild(vp, filesize, lblkno,
207 				blkno, size, nblks, bp);
208 			lblkno += (bp->b_bufsize / size);
209 		} else {
210 single_block_read:
211 			/*
212 			 * if it isn't in the cache, then get a chunk from
213 			 * disk if sequential, otherwise just get the block.
214 			 */
215 			bp->b_flags |= B_READ | B_RAM;
216 			lblkno += 1;
217 		}
218 	}
219 
220 	/*
221 	 * if we have been doing sequential I/O, then do some read-ahead
222 	 */
223 	rbp = NULL;
224 	if (seqcount && (lblkno < (origblkno + seqcount))) {
225 		/*
226 		 * we now build the read-ahead buffer if it is desirable.
227 		 */
228 		if (((u_quad_t)(lblkno + 1) * size) <= filesize &&
229 		    !(error = VOP_BMAP(vp, lblkno, NULL, &blkno, &num_ra, NULL)) &&
230 		    blkno != -1) {
231 			int nblksread;
232 			int ntoread = num_ra + 1;
233 			nblksread = (origtotread + size - 1) / size;
234 			if (seqcount < nblksread)
235 				seqcount = nblksread;
236 			if (seqcount < ntoread)
237 				ntoread = seqcount;
238 			if (num_ra) {
239 				rbp = cluster_rbuild(vp, filesize, lblkno,
240 					blkno, size, ntoread, NULL);
241 			} else {
242 				rbp = getblk(vp, lblkno, size, 0, 0);
243 				rbp->b_flags |= B_READ | B_ASYNC | B_RAM;
244 				rbp->b_blkno = blkno;
245 			}
246 		}
247 	}
248 
249 	/*
250 	 * handle the synchronous read
251 	 */
252 	if (bp) {
253 #if defined(CLUSTERDEBUG)
254 		if (rcluster)
255 			printf("S(%ld,%ld,%d) ",
256 			    (long)bp->b_lblkno, bp->b_bcount, seqcount);
257 #endif
258 		if ((bp->b_flags & B_CLUSTER) == 0)
259 			vfs_busy_pages(bp, 0);
260 		bp->b_flags &= ~(B_ERROR|B_INVAL);
261 		if (bp->b_flags & (B_ASYNC|B_CALL))
262 			BUF_KERNPROC(bp);
263 		error = VOP_STRATEGY(vp, bp);
264 		curproc->p_stats->p_ru.ru_inblock++;
265 	}
266 
267 	/*
268 	 * and if we have read-aheads, do them too
269 	 */
270 	if (rbp) {
271 		if (error) {
272 			rbp->b_flags &= ~(B_ASYNC | B_READ);
273 			brelse(rbp);
274 		} else if (rbp->b_flags & B_CACHE) {
275 			rbp->b_flags &= ~(B_ASYNC | B_READ);
276 			bqrelse(rbp);
277 		} else {
278 #if defined(CLUSTERDEBUG)
279 			if (rcluster) {
280 				if (bp)
281 					printf("A+(%ld,%ld,%ld,%d) ",
282 					    (long)rbp->b_lblkno, rbp->b_bcount,
283 					    (long)(rbp->b_lblkno - origblkno),
284 					    seqcount);
285 				else
286 					printf("A(%ld,%ld,%ld,%d) ",
287 					    (long)rbp->b_lblkno, rbp->b_bcount,
288 					    (long)(rbp->b_lblkno - origblkno),
289 					    seqcount);
290 			}
291 #endif
292 
293 			if ((rbp->b_flags & B_CLUSTER) == 0)
294 				vfs_busy_pages(rbp, 0);
295 			rbp->b_flags &= ~(B_ERROR|B_INVAL);
296 			if (rbp->b_flags & (B_ASYNC|B_CALL))
297 				BUF_KERNPROC(rbp);
298 			(void) VOP_STRATEGY(vp, rbp);
299 			curproc->p_stats->p_ru.ru_inblock++;
300 		}
301 	}
302 	if (reqbp)
303 		return (biowait(reqbp));
304 	else
305 		return (error);
306 }
307 
308 /*
309  * If blocks are contiguous on disk, use this to provide clustered
310  * read ahead.  We will read as many blocks as possible sequentially
311  * and then parcel them up into logical blocks in the buffer hash table.
312  */
313 static struct buf *
314 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
315 	struct vnode *vp;
316 	u_quad_t filesize;
317 	daddr_t lbn;
318 	daddr_t blkno;
319 	long size;
320 	int run;
321 	struct buf *fbp;
322 {
323 	struct buf *bp, *tbp;
324 	daddr_t bn;
325 	int i, inc, j;
326 
327 	KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
328 	    ("cluster_rbuild: size %ld != filesize %ld\n",
329 	    size, vp->v_mount->mnt_stat.f_iosize));
330 
331 	/*
332 	 * avoid a division
333 	 */
334 	while ((u_quad_t) size * (lbn + run) > filesize) {
335 		--run;
336 	}
337 
338 	if (fbp) {
339 		tbp = fbp;
340 		tbp->b_flags |= B_READ;
341 	} else {
342 		tbp = getblk(vp, lbn, size, 0, 0);
343 		if (tbp->b_flags & B_CACHE)
344 			return tbp;
345 		tbp->b_flags |= B_ASYNC | B_READ | B_RAM;
346 	}
347 
348 	tbp->b_blkno = blkno;
349 	if( (tbp->b_flags & B_MALLOC) ||
350 		((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
351 		return tbp;
352 
353 	bp = trypbuf(&cluster_pbuf_freecnt);
354 	if (bp == 0)
355 		return tbp;
356 
357 	bp->b_data = (char *)((vm_offset_t)bp->b_data |
358 	    ((vm_offset_t)tbp->b_data & PAGE_MASK));
359 	bp->b_flags = B_ASYNC | B_READ | B_CALL | B_CLUSTER | B_VMIO;
360 	bp->b_iodone = cluster_callback;
361 	bp->b_blkno = blkno;
362 	bp->b_lblkno = lbn;
363 	bp->b_offset = tbp->b_offset;
364 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
365 	pbgetvp(vp, bp);
366 
367 	TAILQ_INIT(&bp->b_cluster.cluster_head);
368 
369 	bp->b_bcount = 0;
370 	bp->b_bufsize = 0;
371 	bp->b_npages = 0;
372 
373 	if (vp->v_maxio == 0)
374 		vp->v_maxio = DFLTPHYS;
375 	inc = btodb(size);
376 	for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
377 		if (i != 0) {
378 			if ((bp->b_npages * PAGE_SIZE) +
379 				round_page(size) > vp->v_maxio)
380 				break;
381 
382 			if ((tbp = incore(vp, lbn + i)) != NULL) {
383 				if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
384 					break;
385 				BUF_UNLOCK(tbp);
386 
387 				for (j = 0; j < tbp->b_npages; j++)
388 					if (tbp->b_pages[j]->valid)
389 						break;
390 
391 				if (j != tbp->b_npages)
392 					break;
393 
394 				if (tbp->b_bcount != size)
395 					break;
396 			}
397 
398 			tbp = getblk(vp, lbn + i, size, 0, 0);
399 
400 			if ((tbp->b_flags & B_CACHE) ||
401 				(tbp->b_flags & B_VMIO) == 0) {
402 				bqrelse(tbp);
403 				break;
404 			}
405 
406 			for (j = 0;j < tbp->b_npages; j++)
407 				if (tbp->b_pages[j]->valid)
408 					break;
409 
410 			if (j != tbp->b_npages) {
411 				bqrelse(tbp);
412 				break;
413 			}
414 
415 			if ((fbp && (i == 1)) || (i == (run - 1)))
416 				tbp->b_flags |= B_RAM;
417 			tbp->b_flags |= B_READ | B_ASYNC;
418 			if (tbp->b_blkno == tbp->b_lblkno) {
419 				tbp->b_blkno = bn;
420 			} else if (tbp->b_blkno != bn) {
421 				brelse(tbp);
422 				break;
423 			}
424 		}
425 		/*
426 		 * XXX fbp from caller may not be B_ASYNC, but we are going
427 		 * to biodone() it in cluster_callback() anyway
428 		 */
429 		BUF_KERNPROC(tbp);
430 		TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
431 			tbp, b_cluster.cluster_entry);
432 		for (j = 0; j < tbp->b_npages; j += 1) {
433 			vm_page_t m;
434 			m = tbp->b_pages[j];
435 			vm_page_io_start(m);
436 			vm_object_pip_add(m->object, 1);
437 			if ((bp->b_npages == 0) ||
438 				(bp->b_pages[bp->b_npages-1] != m)) {
439 				bp->b_pages[bp->b_npages] = m;
440 				bp->b_npages++;
441 			}
442 			if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
443 				tbp->b_pages[j] = bogus_page;
444 		}
445 		bp->b_bcount += tbp->b_bcount;
446 		bp->b_bufsize += tbp->b_bufsize;
447 	}
448 
449 	for(j=0;j<bp->b_npages;j++) {
450 		if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
451 			VM_PAGE_BITS_ALL)
452 			bp->b_pages[j] = bogus_page;
453 	}
454 	if (bp->b_bufsize > bp->b_kvasize)
455 		panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
456 		    bp->b_bufsize, bp->b_kvasize);
457 	bp->b_kvasize = bp->b_bufsize;
458 
459 	pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
460 		(vm_page_t *)bp->b_pages, bp->b_npages);
461 	return (bp);
462 }
463 
464 /*
465  * Cleanup after a clustered read or write.
466  * This is complicated by the fact that any of the buffers might have
467  * extra memory (if there were no empty buffer headers at allocbuf time)
468  * that we will need to shift around.
469  */
470 void
471 cluster_callback(bp)
472 	struct buf *bp;
473 {
474 	struct buf *nbp, *tbp;
475 	int error = 0;
476 
477 	/*
478 	 * Must propogate errors to all the components.
479 	 */
480 	if (bp->b_flags & B_ERROR)
481 		error = bp->b_error;
482 
483 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
484 	/*
485 	 * Move memory from the large cluster buffer into the component
486 	 * buffers and mark IO as done on these.
487 	 */
488 	for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
489 		tbp; tbp = nbp) {
490 		nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
491 		if (error) {
492 			tbp->b_flags |= B_ERROR;
493 			tbp->b_error = error;
494 		} else {
495 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
496 			tbp->b_flags &= ~(B_ERROR|B_INVAL);
497 		}
498 		biodone(tbp);
499 	}
500 	relpbuf(bp, &cluster_pbuf_freecnt);
501 }
502 
503 /*
504  *	cluster_wbuild_wb:
505  *
506  *	Implement modified write build for cluster.
507  *
508  *		write_behind = 0	write behind disabled
509  *		write_behind = 1	write behind normal (default)
510  *		write_behind = 2	write behind backed-off
511  */
512 
513 static __inline int
514 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
515 {
516 	int r = 0;
517 
518 	switch(write_behind) {
519 	case 2:
520 		if (start_lbn < len)
521 			break;
522 		start_lbn -= len;
523 		/* fall through */
524 	case 1:
525 		r = cluster_wbuild(vp, size, start_lbn, len);
526 		/* fall through */
527 	default:
528 		/* fall through */
529 		break;
530 	}
531 	return(r);
532 }
533 
534 /*
535  * Do clustered write for FFS.
536  *
537  * Three cases:
538  *	1. Write is not sequential (write asynchronously)
539  *	Write is sequential:
540  *	2.	beginning of cluster - begin cluster
541  *	3.	middle of a cluster - add to cluster
542  *	4.	end of a cluster - asynchronously write cluster
543  */
544 void
545 cluster_write(bp, filesize)
546 	struct buf *bp;
547 	u_quad_t filesize;
548 {
549 	struct vnode *vp;
550 	daddr_t lbn;
551 	int maxclen, cursize;
552 	int lblocksize;
553 	int async;
554 
555 	vp = bp->b_vp;
556 	if (vp->v_maxio == 0)
557 		vp->v_maxio = DFLTPHYS;
558 	if (vp->v_type == VREG) {
559 		async = vp->v_mount->mnt_flag & MNT_ASYNC;
560 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
561 	} else {
562 		async = 0;
563 		lblocksize = bp->b_bufsize;
564 	}
565 	lbn = bp->b_lblkno;
566 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
567 
568 	/* Initialize vnode to beginning of file. */
569 	if (lbn == 0)
570 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
571 
572 	if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
573 	    (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
574 		maxclen = vp->v_maxio / lblocksize - 1;
575 		if (vp->v_clen != 0) {
576 			/*
577 			 * Next block is not sequential.
578 			 *
579 			 * If we are not writing at end of file, the process
580 			 * seeked to another point in the file since its last
581 			 * write, or we have reached our maximum cluster size,
582 			 * then push the previous cluster. Otherwise try
583 			 * reallocating to make it sequential.
584 			 */
585 			cursize = vp->v_lastw - vp->v_cstart + 1;
586 			if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
587 			    lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
588 				if (!async)
589 					cluster_wbuild(vp, lblocksize,
590 						vp->v_cstart, cursize);
591 			} else {
592 				struct buf **bpp, **endbp;
593 				struct cluster_save *buflist;
594 
595 				buflist = cluster_collectbufs(vp, bp);
596 				endbp = &buflist->bs_children
597 				    [buflist->bs_nchildren - 1];
598 				if (VOP_REALLOCBLKS(vp, buflist)) {
599 					/*
600 					 * Failed, push the previous cluster.
601 					 */
602 					for (bpp = buflist->bs_children;
603 					     bpp < endbp; bpp++)
604 						brelse(*bpp);
605 					free(buflist, M_SEGMENT);
606 					cluster_wbuild_wb(vp, lblocksize,
607 					    vp->v_cstart, cursize);
608 				} else {
609 					/*
610 					 * Succeeded, keep building cluster.
611 					 */
612 					for (bpp = buflist->bs_children;
613 					     bpp <= endbp; bpp++)
614 						bdwrite(*bpp);
615 					free(buflist, M_SEGMENT);
616 					vp->v_lastw = lbn;
617 					vp->v_lasta = bp->b_blkno;
618 					return;
619 				}
620 			}
621 		}
622 		/*
623 		 * Consider beginning a cluster. If at end of file, make
624 		 * cluster as large as possible, otherwise find size of
625 		 * existing cluster.
626 		 */
627 		if ((vp->v_type == VREG) &&
628 			((u_quad_t) bp->b_offset + lblocksize) != filesize &&
629 		    (bp->b_blkno == bp->b_lblkno) &&
630 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
631 		     bp->b_blkno == -1)) {
632 			bawrite(bp);
633 			vp->v_clen = 0;
634 			vp->v_lasta = bp->b_blkno;
635 			vp->v_cstart = lbn + 1;
636 			vp->v_lastw = lbn;
637 			return;
638 		}
639 		vp->v_clen = maxclen;
640 		if (!async && maxclen == 0) {	/* I/O not contiguous */
641 			vp->v_cstart = lbn + 1;
642 			bawrite(bp);
643 		} else {	/* Wait for rest of cluster */
644 			vp->v_cstart = lbn;
645 			bdwrite(bp);
646 		}
647 	} else if (lbn == vp->v_cstart + vp->v_clen) {
648 		/*
649 		 * At end of cluster, write it out.
650 		 */
651 		bdwrite(bp);
652 		cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
653 		vp->v_clen = 0;
654 		vp->v_cstart = lbn + 1;
655 	} else
656 		/*
657 		 * In the middle of a cluster, so just delay the I/O for now.
658 		 */
659 		bdwrite(bp);
660 	vp->v_lastw = lbn;
661 	vp->v_lasta = bp->b_blkno;
662 }
663 
664 
665 /*
666  * This is an awful lot like cluster_rbuild...wish they could be combined.
667  * The last lbn argument is the current block on which I/O is being
668  * performed.  Check to see that it doesn't fall in the middle of
669  * the current block (if last_bp == NULL).
670  */
671 int
672 cluster_wbuild(vp, size, start_lbn, len)
673 	struct vnode *vp;
674 	long size;
675 	daddr_t start_lbn;
676 	int len;
677 {
678 	struct buf *bp, *tbp;
679 	int i, j, s;
680 	int totalwritten = 0;
681 	int dbsize = btodb(size);
682 
683 	if (vp->v_maxio == 0)
684 		vp->v_maxio = DFLTPHYS;
685 	while (len > 0) {
686 		s = splbio();
687 		if (((tbp = gbincore(vp, start_lbn)) == NULL) ||
688 		  ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) ||
689 		  BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
690 			++start_lbn;
691 			--len;
692 			splx(s);
693 			continue;
694 		}
695 		bremfree(tbp);
696 		tbp->b_flags &= ~B_DONE;
697 		splx(s);
698 
699 		/*
700 		 * Extra memory in the buffer, punt on this buffer.
701 		 * XXX we could handle this in most cases, but we would
702 		 * have to push the extra memory down to after our max
703 		 * possible cluster size and then potentially pull it back
704 		 * up if the cluster was terminated prematurely--too much
705 		 * hassle.
706 		 */
707 		if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
708 		  (tbp->b_bcount != tbp->b_bufsize) ||
709 		  (tbp->b_bcount != size) ||
710 		  (len == 1) ||
711 		  ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
712 			totalwritten += tbp->b_bufsize;
713 			bawrite(tbp);
714 			++start_lbn;
715 			--len;
716 			continue;
717 		}
718 
719 		/*
720 		 * We got a pbuf to make the cluster in.
721 		 * so initialise it.
722 		 */
723 		TAILQ_INIT(&bp->b_cluster.cluster_head);
724 		bp->b_bcount = 0;
725 		bp->b_bufsize = 0;
726 		bp->b_npages = 0;
727 		if (tbp->b_wcred != NOCRED) {
728 		    bp->b_wcred = tbp->b_wcred;
729 		    crhold(bp->b_wcred);
730 		}
731 
732 		bp->b_blkno = tbp->b_blkno;
733 		bp->b_lblkno = tbp->b_lblkno;
734 		bp->b_offset = tbp->b_offset;
735 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
736 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
737 		bp->b_flags |= B_CALL | B_CLUSTER |
738 				(tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
739 		bp->b_iodone = cluster_callback;
740 		pbgetvp(vp, bp);
741 		/*
742 		 * From this location in the file, scan forward to see
743 		 * if there are buffers with adjacent data that need to
744 		 * be written as well.
745 		 */
746 		for (i = 0; i < len; ++i, ++start_lbn) {
747 			if (i != 0) { /* If not the first buffer */
748 				s = splbio();
749 				/*
750 				 * If the adjacent data is not even in core it
751 				 * can't need to be written.
752 				 */
753 				if ((tbp = gbincore(vp, start_lbn)) == NULL) {
754 					splx(s);
755 					break;
756 				}
757 
758 				/*
759 				 * If it IS in core, but has different
760 				 * characteristics, don't cluster with it.
761 				 */
762 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
763 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
764 				  != (B_DELWRI | B_CLUSTEROK |
765 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
766 				    tbp->b_wcred != bp->b_wcred ||
767 				    BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
768 					splx(s);
769 					break;
770 				}
771 
772 				/*
773 				 * Check that the combined cluster
774 				 * would make sense with regard to pages
775 				 * and would not be too large
776 				 */
777 				if ((tbp->b_bcount != size) ||
778 				  ((bp->b_blkno + (dbsize * i)) !=
779 				    tbp->b_blkno) ||
780 				  ((tbp->b_npages + bp->b_npages) >
781 				    (vp->v_maxio / PAGE_SIZE))) {
782 					BUF_UNLOCK(tbp);
783 					splx(s);
784 					break;
785 				}
786 				/*
787 				 * Ok, it's passed all the tests,
788 				 * so remove it from the free list
789 				 * and mark it busy. We will use it.
790 				 */
791 				bremfree(tbp);
792 				tbp->b_flags &= ~B_DONE;
793 				splx(s);
794 			} /* end of code for non-first buffers only */
795 			/* check for latent dependencies to be handled */
796 			if ((LIST_FIRST(&tbp->b_dep)) != NULL &&
797 			    bioops.io_start)
798 				(*bioops.io_start)(tbp);
799 			/*
800 			 * If the IO is via the VM then we do some
801 			 * special VM hackery. (yuck)
802 			 */
803 			if (tbp->b_flags & B_VMIO) {
804 				vm_page_t m;
805 
806 				if (i != 0) { /* if not first buffer */
807 					for (j = 0; j < tbp->b_npages; j += 1) {
808 						m = tbp->b_pages[j];
809 						if (m->flags & PG_BUSY)
810 							goto finishcluster;
811 					}
812 				}
813 
814 				for (j = 0; j < tbp->b_npages; j += 1) {
815 					m = tbp->b_pages[j];
816 					vm_page_io_start(m);
817 					vm_object_pip_add(m->object, 1);
818 					if ((bp->b_npages == 0) ||
819 					  (bp->b_pages[bp->b_npages - 1] != m)) {
820 						bp->b_pages[bp->b_npages] = m;
821 						bp->b_npages++;
822 					}
823 				}
824 			}
825 			bp->b_bcount += size;
826 			bp->b_bufsize += size;
827 
828 			s = splbio();
829 			bundirty(tbp);
830 			tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
831 			tbp->b_flags |= B_ASYNC;
832 			reassignbuf(tbp, tbp->b_vp);	/* put on clean list */
833 			++tbp->b_vp->v_numoutput;
834 			splx(s);
835 			BUF_KERNPROC(tbp);
836 			TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
837 				tbp, b_cluster.cluster_entry);
838 		}
839 	finishcluster:
840 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
841 			(vm_page_t *) bp->b_pages, bp->b_npages);
842 		if (bp->b_bufsize > bp->b_kvasize)
843 			panic(
844 			    "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
845 			    bp->b_bufsize, bp->b_kvasize);
846 		bp->b_kvasize = bp->b_bufsize;
847 		totalwritten += bp->b_bufsize;
848 		bp->b_dirtyoff = 0;
849 		bp->b_dirtyend = bp->b_bufsize;
850 		bawrite(bp);
851 
852 		len -= i;
853 	}
854 	return totalwritten;
855 }
856 
857 /*
858  * Collect together all the buffers in a cluster.
859  * Plus add one additional buffer.
860  */
861 static struct cluster_save *
862 cluster_collectbufs(vp, last_bp)
863 	struct vnode *vp;
864 	struct buf *last_bp;
865 {
866 	struct cluster_save *buflist;
867 	struct buf *bp;
868 	daddr_t lbn;
869 	int i, len;
870 
871 	len = vp->v_lastw - vp->v_cstart + 1;
872 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
873 	    M_SEGMENT, M_WAITOK);
874 	buflist->bs_nchildren = 0;
875 	buflist->bs_children = (struct buf **) (buflist + 1);
876 	for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
877 		(void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
878 		buflist->bs_children[i] = bp;
879 		if (bp->b_blkno == bp->b_lblkno)
880 			VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
881 				NULL, NULL);
882 	}
883 	buflist->bs_children[i] = bp = last_bp;
884 	if (bp->b_blkno == bp->b_lblkno)
885 		VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
886 			NULL, NULL);
887 	buflist->bs_nchildren = i + 1;
888 	return (buflist);
889 }
890