xref: /freebsd/sys/kern/vfs_cluster.c (revision 4cf49a43559ed9fdad601bdcccd2c55963008675)
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  * $FreeBSD$
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 
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 	 */
119 	*bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0);
120 	origblkno = lblkno;
121 	origtotread = totread;
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 			int s;
135 			struct buf *tbp;
136 			bp->b_flags &= ~B_RAM;
137 			/*
138 			 * We do the spl here so that there is no window
139 			 * between the incore and the b_usecount increment
140 			 * below.  We opt to keep the spl out of the loop
141 			 * for efficiency.
142 			 */
143 			s = splbio();
144 			for (i = 1; i < maxra; i++) {
145 
146 				if (!(tbp = incore(vp, lblkno+i))) {
147 					break;
148 				}
149 
150 				/*
151 				 * Set another read-ahead mark so we know
152 				 * to check again.
153 				 */
154 				if (((i % racluster) == (racluster - 1)) ||
155 					(i == (maxra - 1)))
156 					tbp->b_flags |= B_RAM;
157 			}
158 			splx(s);
159 			if (i >= maxra) {
160 				return 0;
161 			}
162 			lblkno += i;
163 		}
164 		reqbp = bp = NULL;
165 	} else {
166 		off_t firstread = bp->b_offset;
167 
168 		KASSERT(bp->b_offset != NOOFFSET,
169 		    ("cluster_read: no buffer offset"));
170 		if (firstread + totread > filesize)
171 			totread = filesize - firstread;
172 		if (totread > size) {
173 			int nblks = 0;
174 			int ncontigafter;
175 			while (totread > 0) {
176 				nblks++;
177 				totread -= size;
178 			}
179 			if (nblks == 1)
180 				goto single_block_read;
181 			if (nblks > racluster)
182 				nblks = racluster;
183 
184 	    		error = VOP_BMAP(vp, lblkno, NULL,
185 				&blkno, &ncontigafter, NULL);
186 			if (error)
187 				goto single_block_read;
188 			if (blkno == -1)
189 				goto single_block_read;
190 			if (ncontigafter == 0)
191 				goto single_block_read;
192 			if (ncontigafter + 1 < nblks)
193 				nblks = ncontigafter + 1;
194 
195 			bp = cluster_rbuild(vp, filesize, lblkno,
196 				blkno, size, nblks, bp);
197 			lblkno += (bp->b_bufsize / size);
198 		} else {
199 single_block_read:
200 			/*
201 			 * if it isn't in the cache, then get a chunk from
202 			 * disk if sequential, otherwise just get the block.
203 			 */
204 			bp->b_flags |= B_READ | B_RAM;
205 			lblkno += 1;
206 		}
207 	}
208 
209 	/*
210 	 * if we have been doing sequential I/O, then do some read-ahead
211 	 */
212 	rbp = NULL;
213 	if (seqcount && (lblkno < (origblkno + seqcount))) {
214 		/*
215 		 * we now build the read-ahead buffer if it is desirable.
216 		 */
217 		if (((u_quad_t)(lblkno + 1) * size) <= filesize &&
218 		    !(error = VOP_BMAP(vp, lblkno, NULL, &blkno, &num_ra, NULL)) &&
219 		    blkno != -1) {
220 			int nblksread;
221 			int ntoread = num_ra + 1;
222 			nblksread = (origtotread + size - 1) / size;
223 			if (seqcount < nblksread)
224 				seqcount = nblksread;
225 			if (seqcount < ntoread)
226 				ntoread = seqcount;
227 			if (num_ra) {
228 				rbp = cluster_rbuild(vp, filesize, lblkno,
229 					blkno, size, ntoread, NULL);
230 			} else {
231 				rbp = getblk(vp, lblkno, size, 0, 0);
232 				rbp->b_flags |= B_READ | B_ASYNC | B_RAM;
233 				rbp->b_blkno = blkno;
234 			}
235 		}
236 	}
237 
238 	/*
239 	 * handle the synchronous read
240 	 */
241 	if (bp) {
242 #if defined(CLUSTERDEBUG)
243 		if (rcluster)
244 			printf("S(%ld,%ld,%d) ",
245 			    (long)bp->b_lblkno, bp->b_bcount, seqcount);
246 #endif
247 		if ((bp->b_flags & B_CLUSTER) == 0)
248 			vfs_busy_pages(bp, 0);
249 		bp->b_flags &= ~(B_ERROR|B_INVAL);
250 		if (bp->b_flags & (B_ASYNC|B_CALL))
251 			BUF_KERNPROC(bp);
252 		error = VOP_STRATEGY(vp, bp);
253 		curproc->p_stats->p_ru.ru_inblock++;
254 	}
255 
256 	/*
257 	 * and if we have read-aheads, do them too
258 	 */
259 	if (rbp) {
260 		if (error) {
261 			rbp->b_flags &= ~(B_ASYNC | B_READ);
262 			brelse(rbp);
263 		} else if (rbp->b_flags & B_CACHE) {
264 			rbp->b_flags &= ~(B_ASYNC | B_READ);
265 			bqrelse(rbp);
266 		} else {
267 #if defined(CLUSTERDEBUG)
268 			if (rcluster) {
269 				if (bp)
270 					printf("A+(%ld,%ld,%ld,%d) ",
271 					    (long)rbp->b_lblkno, rbp->b_bcount,
272 					    (long)(rbp->b_lblkno - origblkno),
273 					    seqcount);
274 				else
275 					printf("A(%ld,%ld,%ld,%d) ",
276 					    (long)rbp->b_lblkno, rbp->b_bcount,
277 					    (long)(rbp->b_lblkno - origblkno),
278 					    seqcount);
279 			}
280 #endif
281 
282 			if ((rbp->b_flags & B_CLUSTER) == 0)
283 				vfs_busy_pages(rbp, 0);
284 			rbp->b_flags &= ~(B_ERROR|B_INVAL);
285 			if (rbp->b_flags & (B_ASYNC|B_CALL))
286 				BUF_KERNPROC(rbp);
287 			(void) VOP_STRATEGY(vp, rbp);
288 			curproc->p_stats->p_ru.ru_inblock++;
289 		}
290 	}
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  */
302 static struct buf *
303 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
304 	struct vnode *vp;
305 	u_quad_t filesize;
306 	daddr_t lbn;
307 	daddr_t blkno;
308 	long size;
309 	int run;
310 	struct buf *fbp;
311 {
312 	struct buf *bp, *tbp;
313 	daddr_t bn;
314 	int i, inc, j;
315 
316 	KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
317 	    ("cluster_rbuild: size %ld != filesize %ld\n",
318 	    size, vp->v_mount->mnt_stat.f_iosize));
319 
320 	/*
321 	 * avoid a division
322 	 */
323 	while ((u_quad_t) size * (lbn + run) > filesize) {
324 		--run;
325 	}
326 
327 	if (fbp) {
328 		tbp = fbp;
329 		tbp->b_flags |= B_READ;
330 	} else {
331 		tbp = getblk(vp, lbn, size, 0, 0);
332 		if (tbp->b_flags & B_CACHE)
333 			return tbp;
334 		tbp->b_flags |= B_ASYNC | B_READ | B_RAM;
335 	}
336 
337 	tbp->b_blkno = blkno;
338 	if( (tbp->b_flags & B_MALLOC) ||
339 		((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
340 		return tbp;
341 
342 	bp = trypbuf(&cluster_pbuf_freecnt);
343 	if (bp == 0)
344 		return tbp;
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_CALL | B_CLUSTER | B_VMIO;
349 	bp->b_iodone = cluster_callback;
350 	bp->b_blkno = blkno;
351 	bp->b_lblkno = lbn;
352 	bp->b_offset = tbp->b_offset;
353 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
354 	pbgetvp(vp, bp);
355 
356 	TAILQ_INIT(&bp->b_cluster.cluster_head);
357 
358 	bp->b_bcount = 0;
359 	bp->b_bufsize = 0;
360 	bp->b_npages = 0;
361 
362 	inc = btodb(size);
363 	for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
364 		if (i != 0) {
365 			if ((bp->b_npages * PAGE_SIZE) +
366 				round_page(size) > vp->v_mount->mnt_iosize_max)
367 				break;
368 
369 			if ((tbp = incore(vp, lbn + i)) != NULL) {
370 				if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
371 					break;
372 				BUF_UNLOCK(tbp);
373 
374 				for (j = 0; j < tbp->b_npages; j++)
375 					if (tbp->b_pages[j]->valid)
376 						break;
377 
378 				if (j != tbp->b_npages)
379 					break;
380 
381 				if (tbp->b_bcount != size)
382 					break;
383 			}
384 
385 			tbp = getblk(vp, lbn + i, size, 0, 0);
386 
387 			if ((tbp->b_flags & B_CACHE) ||
388 				(tbp->b_flags & B_VMIO) == 0) {
389 				bqrelse(tbp);
390 				break;
391 			}
392 
393 			for (j = 0;j < tbp->b_npages; j++)
394 				if (tbp->b_pages[j]->valid)
395 					break;
396 
397 			if (j != tbp->b_npages) {
398 				bqrelse(tbp);
399 				break;
400 			}
401 
402 			if ((fbp && (i == 1)) || (i == (run - 1)))
403 				tbp->b_flags |= B_RAM;
404 			tbp->b_flags |= B_READ | B_ASYNC;
405 			if (tbp->b_blkno == tbp->b_lblkno) {
406 				tbp->b_blkno = bn;
407 			} else if (tbp->b_blkno != bn) {
408 				brelse(tbp);
409 				break;
410 			}
411 		}
412 		/*
413 		 * XXX fbp from caller may not be B_ASYNC, but we are going
414 		 * to biodone() it in cluster_callback() anyway
415 		 */
416 		BUF_KERNPROC(tbp);
417 		TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
418 			tbp, b_cluster.cluster_entry);
419 		for (j = 0; j < tbp->b_npages; j += 1) {
420 			vm_page_t m;
421 			m = tbp->b_pages[j];
422 			vm_page_io_start(m);
423 			vm_object_pip_add(m->object, 1);
424 			if ((bp->b_npages == 0) ||
425 				(bp->b_pages[bp->b_npages-1] != m)) {
426 				bp->b_pages[bp->b_npages] = m;
427 				bp->b_npages++;
428 			}
429 			if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
430 				tbp->b_pages[j] = bogus_page;
431 		}
432 		bp->b_bcount += tbp->b_bcount;
433 		bp->b_bufsize += tbp->b_bufsize;
434 	}
435 
436 	for(j=0;j<bp->b_npages;j++) {
437 		if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
438 			VM_PAGE_BITS_ALL)
439 			bp->b_pages[j] = bogus_page;
440 	}
441 	if (bp->b_bufsize > bp->b_kvasize)
442 		panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
443 		    bp->b_bufsize, bp->b_kvasize);
444 	bp->b_kvasize = bp->b_bufsize;
445 
446 	pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
447 		(vm_page_t *)bp->b_pages, bp->b_npages);
448 	return (bp);
449 }
450 
451 /*
452  * Cleanup after a clustered read or write.
453  * This is complicated by the fact that any of the buffers might have
454  * extra memory (if there were no empty buffer headers at allocbuf time)
455  * that we will need to shift around.
456  */
457 void
458 cluster_callback(bp)
459 	struct buf *bp;
460 {
461 	struct buf *nbp, *tbp;
462 	int error = 0;
463 
464 	/*
465 	 * Must propogate errors to all the components.
466 	 */
467 	if (bp->b_flags & B_ERROR)
468 		error = bp->b_error;
469 
470 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
471 	/*
472 	 * Move memory from the large cluster buffer into the component
473 	 * buffers and mark IO as done on these.
474 	 */
475 	for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
476 		tbp; tbp = nbp) {
477 		nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
478 		if (error) {
479 			tbp->b_flags |= B_ERROR;
480 			tbp->b_error = error;
481 		} else {
482 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
483 			tbp->b_flags &= ~(B_ERROR|B_INVAL);
484 		}
485 		biodone(tbp);
486 	}
487 	relpbuf(bp, &cluster_pbuf_freecnt);
488 }
489 
490 /*
491  *	cluster_wbuild_wb:
492  *
493  *	Implement modified write build for cluster.
494  *
495  *		write_behind = 0	write behind disabled
496  *		write_behind = 1	write behind normal (default)
497  *		write_behind = 2	write behind backed-off
498  */
499 
500 static __inline int
501 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
502 {
503 	int r = 0;
504 
505 	switch(write_behind) {
506 	case 2:
507 		if (start_lbn < len)
508 			break;
509 		start_lbn -= len;
510 		/* fall through */
511 	case 1:
512 		r = cluster_wbuild(vp, size, start_lbn, len);
513 		/* fall through */
514 	default:
515 		/* fall through */
516 		break;
517 	}
518 	return(r);
519 }
520 
521 /*
522  * Do clustered write for FFS.
523  *
524  * Three cases:
525  *	1. Write is not sequential (write asynchronously)
526  *	Write is sequential:
527  *	2.	beginning of cluster - begin cluster
528  *	3.	middle of a cluster - add to cluster
529  *	4.	end of a cluster - asynchronously write cluster
530  */
531 void
532 cluster_write(bp, filesize)
533 	struct buf *bp;
534 	u_quad_t filesize;
535 {
536 	struct vnode *vp;
537 	daddr_t lbn;
538 	int maxclen, cursize;
539 	int lblocksize;
540 	int async;
541 
542 	vp = bp->b_vp;
543 	if (vp->v_type == VREG) {
544 		async = vp->v_mount->mnt_flag & MNT_ASYNC;
545 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
546 	} else {
547 		async = 0;
548 		lblocksize = bp->b_bufsize;
549 	}
550 	lbn = bp->b_lblkno;
551 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
552 
553 	/* Initialize vnode to beginning of file. */
554 	if (lbn == 0)
555 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
556 
557 	if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
558 	    (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
559 		maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
560 		if (vp->v_clen != 0) {
561 			/*
562 			 * Next block is not sequential.
563 			 *
564 			 * If we are not writing at end of file, the process
565 			 * seeked to another point in the file since its last
566 			 * write, or we have reached our maximum cluster size,
567 			 * then push the previous cluster. Otherwise try
568 			 * reallocating to make it sequential.
569 			 */
570 			cursize = vp->v_lastw - vp->v_cstart + 1;
571 			if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
572 			    lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
573 				if (!async)
574 					cluster_wbuild_wb(vp, lblocksize,
575 						vp->v_cstart, cursize);
576 			} else {
577 				struct buf **bpp, **endbp;
578 				struct cluster_save *buflist;
579 
580 				buflist = cluster_collectbufs(vp, bp);
581 				endbp = &buflist->bs_children
582 				    [buflist->bs_nchildren - 1];
583 				if (VOP_REALLOCBLKS(vp, buflist)) {
584 					/*
585 					 * Failed, push the previous cluster.
586 					 */
587 					for (bpp = buflist->bs_children;
588 					     bpp < endbp; bpp++)
589 						brelse(*bpp);
590 					free(buflist, M_SEGMENT);
591 					cluster_wbuild_wb(vp, lblocksize,
592 					    vp->v_cstart, cursize);
593 				} else {
594 					/*
595 					 * Succeeded, keep building cluster.
596 					 */
597 					for (bpp = buflist->bs_children;
598 					     bpp <= endbp; bpp++)
599 						bdwrite(*bpp);
600 					free(buflist, M_SEGMENT);
601 					vp->v_lastw = lbn;
602 					vp->v_lasta = bp->b_blkno;
603 					return;
604 				}
605 			}
606 		}
607 		/*
608 		 * Consider beginning a cluster. If at end of file, make
609 		 * cluster as large as possible, otherwise find size of
610 		 * existing cluster.
611 		 */
612 		if ((vp->v_type == VREG) &&
613 			((u_quad_t) bp->b_offset + lblocksize) != filesize &&
614 		    (bp->b_blkno == bp->b_lblkno) &&
615 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
616 		     bp->b_blkno == -1)) {
617 			bawrite(bp);
618 			vp->v_clen = 0;
619 			vp->v_lasta = bp->b_blkno;
620 			vp->v_cstart = lbn + 1;
621 			vp->v_lastw = lbn;
622 			return;
623 		}
624 		vp->v_clen = maxclen;
625 		if (!async && maxclen == 0) {	/* I/O not contiguous */
626 			vp->v_cstart = lbn + 1;
627 			bawrite(bp);
628 		} else {	/* Wait for rest of cluster */
629 			vp->v_cstart = lbn;
630 			bdwrite(bp);
631 		}
632 	} else if (lbn == vp->v_cstart + vp->v_clen) {
633 		/*
634 		 * At end of cluster, write it out.
635 		 */
636 		bdwrite(bp);
637 		cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
638 		vp->v_clen = 0;
639 		vp->v_cstart = lbn + 1;
640 	} else
641 		/*
642 		 * In the middle of a cluster, so just delay the I/O for now.
643 		 */
644 		bdwrite(bp);
645 	vp->v_lastw = lbn;
646 	vp->v_lasta = bp->b_blkno;
647 }
648 
649 
650 /*
651  * This is an awful lot like cluster_rbuild...wish they could be combined.
652  * The last lbn argument is the current block on which I/O is being
653  * performed.  Check to see that it doesn't fall in the middle of
654  * the current block (if last_bp == NULL).
655  */
656 int
657 cluster_wbuild(vp, size, start_lbn, len)
658 	struct vnode *vp;
659 	long size;
660 	daddr_t start_lbn;
661 	int len;
662 {
663 	struct buf *bp, *tbp;
664 	int i, j, s;
665 	int totalwritten = 0;
666 	int dbsize = btodb(size);
667 
668 	while (len > 0) {
669 		s = splbio();
670 		if (((tbp = gbincore(vp, start_lbn)) == NULL) ||
671 		  ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) ||
672 		  BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
673 			++start_lbn;
674 			--len;
675 			splx(s);
676 			continue;
677 		}
678 		bremfree(tbp);
679 		tbp->b_flags &= ~B_DONE;
680 		splx(s);
681 
682 		/*
683 		 * Extra memory in the buffer, punt on this buffer.
684 		 * XXX we could handle this in most cases, but we would
685 		 * have to push the extra memory down to after our max
686 		 * possible cluster size and then potentially pull it back
687 		 * up if the cluster was terminated prematurely--too much
688 		 * hassle.
689 		 */
690 		if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
691 		  (tbp->b_bcount != tbp->b_bufsize) ||
692 		  (tbp->b_bcount != size) ||
693 		  (len == 1) ||
694 		  ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
695 			totalwritten += tbp->b_bufsize;
696 			bawrite(tbp);
697 			++start_lbn;
698 			--len;
699 			continue;
700 		}
701 
702 		/*
703 		 * We got a pbuf to make the cluster in.
704 		 * so initialise it.
705 		 */
706 		TAILQ_INIT(&bp->b_cluster.cluster_head);
707 		bp->b_bcount = 0;
708 		bp->b_bufsize = 0;
709 		bp->b_npages = 0;
710 		if (tbp->b_wcred != NOCRED) {
711 		    bp->b_wcred = tbp->b_wcred;
712 		    crhold(bp->b_wcred);
713 		}
714 
715 		bp->b_blkno = tbp->b_blkno;
716 		bp->b_lblkno = tbp->b_lblkno;
717 		bp->b_offset = tbp->b_offset;
718 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
719 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
720 		bp->b_flags |= B_CALL | B_CLUSTER |
721 				(tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
722 		bp->b_iodone = cluster_callback;
723 		pbgetvp(vp, bp);
724 		/*
725 		 * From this location in the file, scan forward to see
726 		 * if there are buffers with adjacent data that need to
727 		 * be written as well.
728 		 */
729 		for (i = 0; i < len; ++i, ++start_lbn) {
730 			if (i != 0) { /* If not the first buffer */
731 				s = splbio();
732 				/*
733 				 * If the adjacent data is not even in core it
734 				 * can't need to be written.
735 				 */
736 				if ((tbp = gbincore(vp, start_lbn)) == NULL) {
737 					splx(s);
738 					break;
739 				}
740 
741 				/*
742 				 * If it IS in core, but has different
743 				 * characteristics, don't cluster with it.
744 				 */
745 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
746 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
747 				  != (B_DELWRI | B_CLUSTEROK |
748 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
749 				    tbp->b_wcred != bp->b_wcred ||
750 				    BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
751 					splx(s);
752 					break;
753 				}
754 
755 				/*
756 				 * Check that the combined cluster
757 				 * would make sense with regard to pages
758 				 * and would not be too large
759 				 */
760 				if ((tbp->b_bcount != size) ||
761 				  ((bp->b_blkno + (dbsize * i)) !=
762 				    tbp->b_blkno) ||
763 				  ((tbp->b_npages + bp->b_npages) >
764 				    (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
765 					BUF_UNLOCK(tbp);
766 					splx(s);
767 					break;
768 				}
769 				/*
770 				 * Ok, it's passed all the tests,
771 				 * so remove it from the free list
772 				 * and mark it busy. We will use it.
773 				 */
774 				bremfree(tbp);
775 				tbp->b_flags &= ~B_DONE;
776 				splx(s);
777 			} /* end of code for non-first buffers only */
778 			/* check for latent dependencies to be handled */
779 			if ((LIST_FIRST(&tbp->b_dep)) != NULL &&
780 			    bioops.io_start)
781 				(*bioops.io_start)(tbp);
782 			/*
783 			 * If the IO is via the VM then we do some
784 			 * special VM hackery. (yuck)
785 			 */
786 			if (tbp->b_flags & B_VMIO) {
787 				vm_page_t m;
788 
789 				if (i != 0) { /* if not first buffer */
790 					for (j = 0; j < tbp->b_npages; j += 1) {
791 						m = tbp->b_pages[j];
792 						if (m->flags & PG_BUSY) {
793 							bqrelse(tbp);
794 							goto finishcluster;
795 						}
796 					}
797 				}
798 
799 				for (j = 0; j < tbp->b_npages; j += 1) {
800 					m = tbp->b_pages[j];
801 					vm_page_io_start(m);
802 					vm_object_pip_add(m->object, 1);
803 					if ((bp->b_npages == 0) ||
804 					  (bp->b_pages[bp->b_npages - 1] != m)) {
805 						bp->b_pages[bp->b_npages] = m;
806 						bp->b_npages++;
807 					}
808 				}
809 			}
810 			bp->b_bcount += size;
811 			bp->b_bufsize += size;
812 
813 			s = splbio();
814 			bundirty(tbp);
815 			tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
816 			tbp->b_flags |= B_ASYNC;
817 			reassignbuf(tbp, tbp->b_vp);	/* put on clean list */
818 			++tbp->b_vp->v_numoutput;
819 			splx(s);
820 			BUF_KERNPROC(tbp);
821 			TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
822 				tbp, b_cluster.cluster_entry);
823 		}
824 	finishcluster:
825 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
826 			(vm_page_t *) bp->b_pages, bp->b_npages);
827 		if (bp->b_bufsize > bp->b_kvasize)
828 			panic(
829 			    "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
830 			    bp->b_bufsize, bp->b_kvasize);
831 		bp->b_kvasize = bp->b_bufsize;
832 		totalwritten += bp->b_bufsize;
833 		bp->b_dirtyoff = 0;
834 		bp->b_dirtyend = bp->b_bufsize;
835 		bawrite(bp);
836 
837 		len -= i;
838 	}
839 	return totalwritten;
840 }
841 
842 /*
843  * Collect together all the buffers in a cluster.
844  * Plus add one additional buffer.
845  */
846 static struct cluster_save *
847 cluster_collectbufs(vp, last_bp)
848 	struct vnode *vp;
849 	struct buf *last_bp;
850 {
851 	struct cluster_save *buflist;
852 	struct buf *bp;
853 	daddr_t lbn;
854 	int i, len;
855 
856 	len = vp->v_lastw - vp->v_cstart + 1;
857 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
858 	    M_SEGMENT, M_WAITOK);
859 	buflist->bs_nchildren = 0;
860 	buflist->bs_children = (struct buf **) (buflist + 1);
861 	for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
862 		(void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
863 		buflist->bs_children[i] = bp;
864 		if (bp->b_blkno == bp->b_lblkno)
865 			VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
866 				NULL, NULL);
867 	}
868 	buflist->bs_children[i] = bp = last_bp;
869 	if (bp->b_blkno == bp->b_lblkno)
870 		VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
871 			NULL, NULL);
872 	buflist->bs_nchildren = i + 1;
873 	return (buflist);
874 }
875