xref: /freebsd/sys/kern/vfs_cluster.c (revision 55bce0c1203e70d8b62a3dedc9235ab39660c6f4)
1 /*-
2  * Copyright (c) 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * Modifications/enhancements:
5  * 	Copyright (c) 1995 John S. Dyson.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 4. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	@(#)vfs_cluster.c	8.7 (Berkeley) 2/13/94
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include "opt_debug_cluster.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/proc.h>
43 #include <sys/bio.h>
44 #include <sys/buf.h>
45 #include <sys/vnode.h>
46 #include <sys/malloc.h>
47 #include <sys/mount.h>
48 #include <sys/resourcevar.h>
49 #include <sys/rwlock.h>
50 #include <sys/vmmeter.h>
51 #include <vm/vm.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_page.h>
54 #include <sys/sysctl.h>
55 
56 #if defined(CLUSTERDEBUG)
57 static int	rcluster= 0;
58 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0,
59     "Debug VFS clustering code");
60 #endif
61 
62 static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer");
63 
64 static struct cluster_save *cluster_collectbufs(struct vnode *vp,
65 	    struct buf *last_bp, int gbflags);
66 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize,
67 	    daddr_t lbn, daddr_t blkno, long size, int run, int gbflags,
68 	    struct buf *fbp);
69 static void cluster_callback(struct buf *);
70 
71 static int write_behind = 1;
72 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
73     "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
74 
75 static int read_max = 64;
76 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
77     "Cluster read-ahead max block count");
78 
79 static int read_min = 1;
80 SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0,
81     "Cluster read min block count");
82 
83 /* Page expended to mark partially backed buffers */
84 extern vm_page_t	bogus_page;
85 
86 /*
87  * Read data to a buf, including read-ahead if we find this to be beneficial.
88  * cluster_read replaces bread.
89  */
90 int
91 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
92     struct ucred *cred, long totread, int seqcount, int gbflags,
93     struct buf **bpp)
94 {
95 	struct buf *bp, *rbp, *reqbp;
96 	struct bufobj *bo;
97 	daddr_t blkno, origblkno;
98 	int maxra, racluster;
99 	int error, ncontig;
100 	int i;
101 
102 	error = 0;
103 	bo = &vp->v_bufobj;
104 	if (!unmapped_buf_allowed)
105 		gbflags &= ~GB_UNMAPPED;
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_mount->mnt_iosize_max / size;
112 	maxra = seqcount;
113 	maxra = min(read_max, maxra);
114 	maxra = min(nbuf/8, maxra);
115 	if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
116 		maxra = (filesize / size) - lblkno;
117 
118 	/*
119 	 * get the requested block
120 	 */
121 	*bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, gbflags);
122 	origblkno = lblkno;
123 
124 	/*
125 	 * if it is in the cache, then check to see if the reads have been
126 	 * sequential.  If they have, then try some read-ahead, otherwise
127 	 * back-off on prospective read-aheads.
128 	 */
129 	if (bp->b_flags & B_CACHE) {
130 		if (!seqcount) {
131 			return 0;
132 		} else if ((bp->b_flags & B_RAM) == 0) {
133 			return 0;
134 		} else {
135 			bp->b_flags &= ~B_RAM;
136 			BO_RLOCK(bo);
137 			for (i = 1; i < maxra; i++) {
138 				/*
139 				 * Stop if the buffer does not exist or it
140 				 * is invalid (about to go away?)
141 				 */
142 				rbp = gbincore(&vp->v_bufobj, lblkno+i);
143 				if (rbp == NULL || (rbp->b_flags & B_INVAL))
144 					break;
145 
146 				/*
147 				 * Set another read-ahead mark so we know
148 				 * to check again. (If we can lock the
149 				 * buffer without waiting)
150 				 */
151 				if ((((i % racluster) == (racluster - 1)) ||
152 				    (i == (maxra - 1)))
153 				    && (0 == BUF_LOCK(rbp,
154 					LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
155 					rbp->b_flags |= B_RAM;
156 					BUF_UNLOCK(rbp);
157 				}
158 			}
159 			BO_RUNLOCK(bo);
160 			if (i >= maxra) {
161 				return 0;
162 			}
163 			lblkno += i;
164 		}
165 		reqbp = bp = NULL;
166 	/*
167 	 * If it isn't in the cache, then get a chunk from
168 	 * disk if sequential, otherwise just get the block.
169 	 */
170 	} else {
171 		off_t firstread = bp->b_offset;
172 		int nblks;
173 		long minread;
174 
175 		KASSERT(bp->b_offset != NOOFFSET,
176 		    ("cluster_read: no buffer offset"));
177 
178 		ncontig = 0;
179 
180 		/*
181 		 * Adjust totread if needed
182 		 */
183 		minread = read_min * size;
184 		if (minread > totread)
185 			totread = minread;
186 
187 		/*
188 		 * Compute the total number of blocks that we should read
189 		 * synchronously.
190 		 */
191 		if (firstread + totread > filesize)
192 			totread = filesize - firstread;
193 		nblks = howmany(totread, size);
194 		if (nblks > racluster)
195 			nblks = racluster;
196 
197 		/*
198 		 * Now compute the number of contiguous blocks.
199 		 */
200 		if (nblks > 1) {
201 	    		error = VOP_BMAP(vp, lblkno, NULL,
202 				&blkno, &ncontig, NULL);
203 			/*
204 			 * If this failed to map just do the original block.
205 			 */
206 			if (error || blkno == -1)
207 				ncontig = 0;
208 		}
209 
210 		/*
211 		 * If we have contiguous data available do a cluster
212 		 * otherwise just read the requested block.
213 		 */
214 		if (ncontig) {
215 			/* Account for our first block. */
216 			ncontig = min(ncontig + 1, nblks);
217 			if (ncontig < nblks)
218 				nblks = ncontig;
219 			bp = cluster_rbuild(vp, filesize, lblkno,
220 			    blkno, size, nblks, gbflags, bp);
221 			lblkno += (bp->b_bufsize / size);
222 		} else {
223 			bp->b_flags |= B_RAM;
224 			bp->b_iocmd = BIO_READ;
225 			lblkno += 1;
226 		}
227 	}
228 
229 	/*
230 	 * handle the synchronous read so that it is available ASAP.
231 	 */
232 	if (bp) {
233 		if ((bp->b_flags & B_CLUSTER) == 0) {
234 			vfs_busy_pages(bp, 0);
235 		}
236 		bp->b_flags &= ~B_INVAL;
237 		bp->b_ioflags &= ~BIO_ERROR;
238 		if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
239 			BUF_KERNPROC(bp);
240 		bp->b_iooffset = dbtob(bp->b_blkno);
241 		bstrategy(bp);
242 		curthread->td_ru.ru_inblock++;
243 	}
244 
245 	/*
246 	 * If we have been doing sequential I/O, then do some read-ahead.
247 	 */
248 	while (lblkno < (origblkno + maxra)) {
249 		error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
250 		if (error)
251 			break;
252 
253 		if (blkno == -1)
254 			break;
255 
256 		/*
257 		 * We could throttle ncontig here by maxra but we might as
258 		 * well read the data if it is contiguous.  We're throttled
259 		 * by racluster anyway.
260 		 */
261 		if (ncontig) {
262 			ncontig = min(ncontig + 1, racluster);
263 			rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
264 			    size, ncontig, gbflags, NULL);
265 			lblkno += (rbp->b_bufsize / size);
266 			if (rbp->b_flags & B_DELWRI) {
267 				bqrelse(rbp);
268 				continue;
269 			}
270 		} else {
271 			rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
272 			lblkno += 1;
273 			if (rbp->b_flags & B_DELWRI) {
274 				bqrelse(rbp);
275 				continue;
276 			}
277 			rbp->b_flags |= B_ASYNC | B_RAM;
278 			rbp->b_iocmd = BIO_READ;
279 			rbp->b_blkno = blkno;
280 		}
281 		if (rbp->b_flags & B_CACHE) {
282 			rbp->b_flags &= ~B_ASYNC;
283 			bqrelse(rbp);
284 			continue;
285 		}
286 		if ((rbp->b_flags & B_CLUSTER) == 0) {
287 			vfs_busy_pages(rbp, 0);
288 		}
289 		rbp->b_flags &= ~B_INVAL;
290 		rbp->b_ioflags &= ~BIO_ERROR;
291 		if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
292 			BUF_KERNPROC(rbp);
293 		rbp->b_iooffset = dbtob(rbp->b_blkno);
294 		bstrategy(rbp);
295 		curthread->td_ru.ru_inblock++;
296 	}
297 
298 	if (reqbp)
299 		return (bufwait(reqbp));
300 	else
301 		return (error);
302 }
303 
304 /*
305  * If blocks are contiguous on disk, use this to provide clustered
306  * read ahead.  We will read as many blocks as possible sequentially
307  * and then parcel them up into logical blocks in the buffer hash table.
308  */
309 static struct buf *
310 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
311     daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
312 {
313 	struct bufobj *bo;
314 	struct buf *bp, *tbp;
315 	daddr_t bn;
316 	off_t off;
317 	long tinc, tsize;
318 	int i, inc, j, toff;
319 
320 	KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
321 	    ("cluster_rbuild: size %ld != f_iosize %jd\n",
322 	    size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
323 
324 	/*
325 	 * avoid a division
326 	 */
327 	while ((u_quad_t) size * (lbn + run) > filesize) {
328 		--run;
329 	}
330 
331 	if (fbp) {
332 		tbp = fbp;
333 		tbp->b_iocmd = BIO_READ;
334 	} else {
335 		tbp = getblk(vp, lbn, size, 0, 0, gbflags);
336 		if (tbp->b_flags & B_CACHE)
337 			return tbp;
338 		tbp->b_flags |= B_ASYNC | B_RAM;
339 		tbp->b_iocmd = BIO_READ;
340 	}
341 	tbp->b_blkno = blkno;
342 	if( (tbp->b_flags & B_MALLOC) ||
343 		((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
344 		return tbp;
345 
346 	bp = trypbuf(&cluster_pbuf_freecnt);
347 	if (bp == 0)
348 		return tbp;
349 
350 	/*
351 	 * We are synthesizing a buffer out of vm_page_t's, but
352 	 * if the block size is not page aligned then the starting
353 	 * address may not be either.  Inherit the b_data offset
354 	 * from the original buffer.
355 	 */
356 	bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
357 	if ((gbflags & GB_UNMAPPED) != 0) {
358 		bp->b_flags |= B_UNMAPPED;
359 		bp->b_data = unmapped_buf;
360 	} else {
361 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
362 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
363 	}
364 	bp->b_iocmd = BIO_READ;
365 	bp->b_iodone = cluster_callback;
366 	bp->b_blkno = blkno;
367 	bp->b_lblkno = lbn;
368 	bp->b_offset = tbp->b_offset;
369 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
370 	pbgetvp(vp, bp);
371 
372 	TAILQ_INIT(&bp->b_cluster.cluster_head);
373 
374 	bp->b_bcount = 0;
375 	bp->b_bufsize = 0;
376 	bp->b_npages = 0;
377 
378 	inc = btodb(size);
379 	bo = &vp->v_bufobj;
380 	for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
381 		if (i != 0) {
382 			if ((bp->b_npages * PAGE_SIZE) +
383 			    round_page(size) > vp->v_mount->mnt_iosize_max) {
384 				break;
385 			}
386 
387 			tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
388 			    (gbflags & GB_UNMAPPED));
389 
390 			/* Don't wait around for locked bufs. */
391 			if (tbp == NULL)
392 				break;
393 
394 			/*
395 			 * Stop scanning if the buffer is fully valid
396 			 * (marked B_CACHE), or locked (may be doing a
397 			 * background write), or if the buffer is not
398 			 * VMIO backed.  The clustering code can only deal
399 			 * with VMIO-backed buffers.  The bo lock is not
400 			 * required for the BKGRDINPROG check since it
401 			 * can not be set without the buf lock.
402 			 */
403 			if ((tbp->b_vflags & BV_BKGRDINPROG) ||
404 			    (tbp->b_flags & B_CACHE) ||
405 			    (tbp->b_flags & B_VMIO) == 0) {
406 				bqrelse(tbp);
407 				break;
408 			}
409 
410 			/*
411 			 * The buffer must be completely invalid in order to
412 			 * take part in the cluster.  If it is partially valid
413 			 * then we stop.
414 			 */
415 			off = tbp->b_offset;
416 			tsize = size;
417 			VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
418 			for (j = 0; tsize > 0; j++) {
419 				toff = off & PAGE_MASK;
420 				tinc = tsize;
421 				if (toff + tinc > PAGE_SIZE)
422 					tinc = PAGE_SIZE - toff;
423 				VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object);
424 				if ((tbp->b_pages[j]->valid &
425 				    vm_page_bits(toff, tinc)) != 0)
426 					break;
427 				off += tinc;
428 				tsize -= tinc;
429 			}
430 			VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
431 			if (tsize > 0) {
432 				bqrelse(tbp);
433 				break;
434 			}
435 
436 			/*
437 			 * Set a read-ahead mark as appropriate
438 			 */
439 			if ((fbp && (i == 1)) || (i == (run - 1)))
440 				tbp->b_flags |= B_RAM;
441 
442 			/*
443 			 * Set the buffer up for an async read (XXX should
444 			 * we do this only if we do not wind up brelse()ing?).
445 			 * Set the block number if it isn't set, otherwise
446 			 * if it is make sure it matches the block number we
447 			 * expect.
448 			 */
449 			tbp->b_flags |= B_ASYNC;
450 			tbp->b_iocmd = BIO_READ;
451 			if (tbp->b_blkno == tbp->b_lblkno) {
452 				tbp->b_blkno = bn;
453 			} else if (tbp->b_blkno != bn) {
454 				brelse(tbp);
455 				break;
456 			}
457 		}
458 		/*
459 		 * XXX fbp from caller may not be B_ASYNC, but we are going
460 		 * to biodone() it in cluster_callback() anyway
461 		 */
462 		BUF_KERNPROC(tbp);
463 		TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
464 			tbp, b_cluster.cluster_entry);
465 		VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
466 		for (j = 0; j < tbp->b_npages; j += 1) {
467 			vm_page_t m;
468 			m = tbp->b_pages[j];
469 			vm_page_io_start(m);
470 			vm_object_pip_add(m->object, 1);
471 			if ((bp->b_npages == 0) ||
472 				(bp->b_pages[bp->b_npages-1] != m)) {
473 				bp->b_pages[bp->b_npages] = m;
474 				bp->b_npages++;
475 			}
476 			if (m->valid == VM_PAGE_BITS_ALL)
477 				tbp->b_pages[j] = bogus_page;
478 		}
479 		VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
480 		/*
481 		 * Don't inherit tbp->b_bufsize as it may be larger due to
482 		 * a non-page-aligned size.  Instead just aggregate using
483 		 * 'size'.
484 		 */
485 		if (tbp->b_bcount != size)
486 			printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
487 		if (tbp->b_bufsize != size)
488 			printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
489 		bp->b_bcount += size;
490 		bp->b_bufsize += size;
491 	}
492 
493 	/*
494 	 * Fully valid pages in the cluster are already good and do not need
495 	 * to be re-read from disk.  Replace the page with bogus_page
496 	 */
497 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
498 	for (j = 0; j < bp->b_npages; j++) {
499 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object);
500 		if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
501 			bp->b_pages[j] = bogus_page;
502 	}
503 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
504 	if (bp->b_bufsize > bp->b_kvasize)
505 		panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
506 		    bp->b_bufsize, bp->b_kvasize);
507 	bp->b_kvasize = bp->b_bufsize;
508 
509 	if ((bp->b_flags & B_UNMAPPED) == 0) {
510 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
511 		    (vm_page_t *)bp->b_pages, bp->b_npages);
512 	}
513 	return (bp);
514 }
515 
516 /*
517  * Cleanup after a clustered read or write.
518  * This is complicated by the fact that any of the buffers might have
519  * extra memory (if there were no empty buffer headers at allocbuf time)
520  * that we will need to shift around.
521  */
522 static void
523 cluster_callback(bp)
524 	struct buf *bp;
525 {
526 	struct buf *nbp, *tbp;
527 	int error = 0;
528 
529 	/*
530 	 * Must propogate errors to all the components.
531 	 */
532 	if (bp->b_ioflags & BIO_ERROR)
533 		error = bp->b_error;
534 
535 	if ((bp->b_flags & B_UNMAPPED) == 0) {
536 		pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
537 		    bp->b_npages);
538 	}
539 	/*
540 	 * Move memory from the large cluster buffer into the component
541 	 * buffers and mark IO as done on these.
542 	 */
543 	for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
544 		tbp; tbp = nbp) {
545 		nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
546 		if (error) {
547 			tbp->b_ioflags |= BIO_ERROR;
548 			tbp->b_error = error;
549 		} else {
550 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
551 			tbp->b_flags &= ~B_INVAL;
552 			tbp->b_ioflags &= ~BIO_ERROR;
553 			/*
554 			 * XXX the bdwrite()/bqrelse() issued during
555 			 * cluster building clears B_RELBUF (see bqrelse()
556 			 * comment).  If direct I/O was specified, we have
557 			 * to restore it here to allow the buffer and VM
558 			 * to be freed.
559 			 */
560 			if (tbp->b_flags & B_DIRECT)
561 				tbp->b_flags |= B_RELBUF;
562 		}
563 		bufdone(tbp);
564 	}
565 	pbrelvp(bp);
566 	relpbuf(bp, &cluster_pbuf_freecnt);
567 }
568 
569 /*
570  *	cluster_wbuild_wb:
571  *
572  *	Implement modified write build for cluster.
573  *
574  *		write_behind = 0	write behind disabled
575  *		write_behind = 1	write behind normal (default)
576  *		write_behind = 2	write behind backed-off
577  */
578 
579 static __inline int
580 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
581     int gbflags)
582 {
583 	int r = 0;
584 
585 	switch (write_behind) {
586 	case 2:
587 		if (start_lbn < len)
588 			break;
589 		start_lbn -= len;
590 		/* FALLTHROUGH */
591 	case 1:
592 		r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
593 		/* FALLTHROUGH */
594 	default:
595 		/* FALLTHROUGH */
596 		break;
597 	}
598 	return(r);
599 }
600 
601 /*
602  * Do clustered write for FFS.
603  *
604  * Three cases:
605  *	1. Write is not sequential (write asynchronously)
606  *	Write is sequential:
607  *	2.	beginning of cluster - begin cluster
608  *	3.	middle of a cluster - add to cluster
609  *	4.	end of a cluster - asynchronously write cluster
610  */
611 void
612 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount,
613     int gbflags)
614 {
615 	daddr_t lbn;
616 	int maxclen, cursize;
617 	int lblocksize;
618 	int async;
619 
620 	if (!unmapped_buf_allowed)
621 		gbflags &= ~GB_UNMAPPED;
622 
623 	if (vp->v_type == VREG) {
624 		async = DOINGASYNC(vp);
625 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
626 	} else {
627 		async = 0;
628 		lblocksize = bp->b_bufsize;
629 	}
630 	lbn = bp->b_lblkno;
631 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
632 
633 	/* Initialize vnode to beginning of file. */
634 	if (lbn == 0)
635 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
636 
637 	if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
638 	    (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
639 		maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
640 		if (vp->v_clen != 0) {
641 			/*
642 			 * Next block is not sequential.
643 			 *
644 			 * If we are not writing at end of file, the process
645 			 * seeked to another point in the file since its last
646 			 * write, or we have reached our maximum cluster size,
647 			 * then push the previous cluster. Otherwise try
648 			 * reallocating to make it sequential.
649 			 *
650 			 * Change to algorithm: only push previous cluster if
651 			 * it was sequential from the point of view of the
652 			 * seqcount heuristic, otherwise leave the buffer
653 			 * intact so we can potentially optimize the I/O
654 			 * later on in the buf_daemon or update daemon
655 			 * flush.
656 			 */
657 			cursize = vp->v_lastw - vp->v_cstart + 1;
658 			if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
659 			    lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
660 				if (!async && seqcount > 0) {
661 					cluster_wbuild_wb(vp, lblocksize,
662 					    vp->v_cstart, cursize, gbflags);
663 				}
664 			} else {
665 				struct buf **bpp, **endbp;
666 				struct cluster_save *buflist;
667 
668 				buflist = cluster_collectbufs(vp, bp, gbflags);
669 				endbp = &buflist->bs_children
670 				    [buflist->bs_nchildren - 1];
671 				if (VOP_REALLOCBLKS(vp, buflist)) {
672 					/*
673 					 * Failed, push the previous cluster
674 					 * if *really* writing sequentially
675 					 * in the logical file (seqcount > 1),
676 					 * otherwise delay it in the hopes that
677 					 * the low level disk driver can
678 					 * optimize the write ordering.
679 					 */
680 					for (bpp = buflist->bs_children;
681 					     bpp < endbp; bpp++)
682 						brelse(*bpp);
683 					free(buflist, M_SEGMENT);
684 					if (seqcount > 1) {
685 						cluster_wbuild_wb(vp,
686 						    lblocksize, vp->v_cstart,
687 						    cursize, gbflags);
688 					}
689 				} else {
690 					/*
691 					 * Succeeded, keep building cluster.
692 					 */
693 					for (bpp = buflist->bs_children;
694 					     bpp <= endbp; bpp++)
695 						bdwrite(*bpp);
696 					free(buflist, M_SEGMENT);
697 					vp->v_lastw = lbn;
698 					vp->v_lasta = bp->b_blkno;
699 					return;
700 				}
701 			}
702 		}
703 		/*
704 		 * Consider beginning a cluster. If at end of file, make
705 		 * cluster as large as possible, otherwise find size of
706 		 * existing cluster.
707 		 */
708 		if ((vp->v_type == VREG) &&
709 			((u_quad_t) bp->b_offset + lblocksize) != filesize &&
710 		    (bp->b_blkno == bp->b_lblkno) &&
711 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
712 		     bp->b_blkno == -1)) {
713 			bawrite(bp);
714 			vp->v_clen = 0;
715 			vp->v_lasta = bp->b_blkno;
716 			vp->v_cstart = lbn + 1;
717 			vp->v_lastw = lbn;
718 			return;
719 		}
720 		vp->v_clen = maxclen;
721 		if (!async && maxclen == 0) {	/* I/O not contiguous */
722 			vp->v_cstart = lbn + 1;
723 			bawrite(bp);
724 		} else {	/* Wait for rest of cluster */
725 			vp->v_cstart = lbn;
726 			bdwrite(bp);
727 		}
728 	} else if (lbn == vp->v_cstart + vp->v_clen) {
729 		/*
730 		 * At end of cluster, write it out if seqcount tells us we
731 		 * are operating sequentially, otherwise let the buf or
732 		 * update daemon handle it.
733 		 */
734 		bdwrite(bp);
735 		if (seqcount > 1) {
736 			cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
737 			    vp->v_clen + 1, gbflags);
738 		}
739 		vp->v_clen = 0;
740 		vp->v_cstart = lbn + 1;
741 	} else if (vm_page_count_severe()) {
742 		/*
743 		 * We are low on memory, get it going NOW
744 		 */
745 		bawrite(bp);
746 	} else {
747 		/*
748 		 * In the middle of a cluster, so just delay the I/O for now.
749 		 */
750 		bdwrite(bp);
751 	}
752 	vp->v_lastw = lbn;
753 	vp->v_lasta = bp->b_blkno;
754 }
755 
756 
757 /*
758  * This is an awful lot like cluster_rbuild...wish they could be combined.
759  * The last lbn argument is the current block on which I/O is being
760  * performed.  Check to see that it doesn't fall in the middle of
761  * the current block (if last_bp == NULL).
762  */
763 int
764 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
765     int gbflags)
766 {
767 	struct buf *bp, *tbp;
768 	struct bufobj *bo;
769 	int i, j;
770 	int totalwritten = 0;
771 	int dbsize = btodb(size);
772 
773 	if (!unmapped_buf_allowed)
774 		gbflags &= ~GB_UNMAPPED;
775 
776 	bo = &vp->v_bufobj;
777 	while (len > 0) {
778 		/*
779 		 * If the buffer is not delayed-write (i.e. dirty), or it
780 		 * is delayed-write but either locked or inval, it cannot
781 		 * partake in the clustered write.
782 		 */
783 		BO_LOCK(bo);
784 		if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
785 		    (tbp->b_vflags & BV_BKGRDINPROG)) {
786 			BO_UNLOCK(bo);
787 			++start_lbn;
788 			--len;
789 			continue;
790 		}
791 		if (BUF_LOCK(tbp,
792 		    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) {
793 			++start_lbn;
794 			--len;
795 			continue;
796 		}
797 		if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
798 			BUF_UNLOCK(tbp);
799 			++start_lbn;
800 			--len;
801 			continue;
802 		}
803 		if (tbp->b_pin_count >  0) {
804 			BUF_UNLOCK(tbp);
805 			++start_lbn;
806 			--len;
807 			continue;
808 		}
809 		bremfree(tbp);
810 		tbp->b_flags &= ~B_DONE;
811 
812 		/*
813 		 * Extra memory in the buffer, punt on this buffer.
814 		 * XXX we could handle this in most cases, but we would
815 		 * have to push the extra memory down to after our max
816 		 * possible cluster size and then potentially pull it back
817 		 * up if the cluster was terminated prematurely--too much
818 		 * hassle.
819 		 */
820 		if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
821 		     (B_CLUSTEROK | B_VMIO)) ||
822 		  (tbp->b_bcount != tbp->b_bufsize) ||
823 		  (tbp->b_bcount != size) ||
824 		  (len == 1) ||
825 		  ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
826 			totalwritten += tbp->b_bufsize;
827 			bawrite(tbp);
828 			++start_lbn;
829 			--len;
830 			continue;
831 		}
832 
833 		/*
834 		 * We got a pbuf to make the cluster in.
835 		 * so initialise it.
836 		 */
837 		TAILQ_INIT(&bp->b_cluster.cluster_head);
838 		bp->b_bcount = 0;
839 		bp->b_bufsize = 0;
840 		bp->b_npages = 0;
841 		if (tbp->b_wcred != NOCRED)
842 			bp->b_wcred = crhold(tbp->b_wcred);
843 
844 		bp->b_blkno = tbp->b_blkno;
845 		bp->b_lblkno = tbp->b_lblkno;
846 		bp->b_offset = tbp->b_offset;
847 
848 		/*
849 		 * We are synthesizing a buffer out of vm_page_t's, but
850 		 * if the block size is not page aligned then the starting
851 		 * address may not be either.  Inherit the b_data offset
852 		 * from the original buffer.
853 		 */
854 		if ((gbflags & GB_UNMAPPED) == 0 ||
855 		    (tbp->b_flags & B_VMIO) == 0) {
856 			bp->b_data = (char *)((vm_offset_t)bp->b_data |
857 			    ((vm_offset_t)tbp->b_data & PAGE_MASK));
858 		} else {
859 			bp->b_flags |= B_UNMAPPED;
860 			bp->b_data = unmapped_buf;
861 		}
862 		bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
863 		    B_NEEDCOMMIT));
864 		bp->b_iodone = cluster_callback;
865 		pbgetvp(vp, bp);
866 		/*
867 		 * From this location in the file, scan forward to see
868 		 * if there are buffers with adjacent data that need to
869 		 * be written as well.
870 		 */
871 		for (i = 0; i < len; ++i, ++start_lbn) {
872 			if (i != 0) { /* If not the first buffer */
873 				/*
874 				 * If the adjacent data is not even in core it
875 				 * can't need to be written.
876 				 */
877 				BO_LOCK(bo);
878 				if ((tbp = gbincore(bo, start_lbn)) == NULL ||
879 				    (tbp->b_vflags & BV_BKGRDINPROG)) {
880 					BO_UNLOCK(bo);
881 					break;
882 				}
883 
884 				/*
885 				 * If it IS in core, but has different
886 				 * characteristics, or is locked (which
887 				 * means it could be undergoing a background
888 				 * I/O or be in a weird state), then don't
889 				 * cluster with it.
890 				 */
891 				if (BUF_LOCK(tbp,
892 				    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
893 				    BO_LOCKPTR(bo)))
894 					break;
895 
896 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
897 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
898 				    != (B_DELWRI | B_CLUSTEROK |
899 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
900 				    tbp->b_wcred != bp->b_wcred) {
901 					BUF_UNLOCK(tbp);
902 					break;
903 				}
904 
905 				/*
906 				 * Check that the combined cluster
907 				 * would make sense with regard to pages
908 				 * and would not be too large
909 				 */
910 				if ((tbp->b_bcount != size) ||
911 				  ((bp->b_blkno + (dbsize * i)) !=
912 				    tbp->b_blkno) ||
913 				  ((tbp->b_npages + bp->b_npages) >
914 				    (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
915 					BUF_UNLOCK(tbp);
916 					break;
917 				}
918 
919 				/*
920 				 * Do not pull in pinned buffers.
921 				 */
922 				if (tbp->b_pin_count > 0) {
923 					BUF_UNLOCK(tbp);
924 					break;
925 				}
926 
927 				/*
928 				 * Ok, it's passed all the tests,
929 				 * so remove it from the free list
930 				 * and mark it busy. We will use it.
931 				 */
932 				bremfree(tbp);
933 				tbp->b_flags &= ~B_DONE;
934 			} /* end of code for non-first buffers only */
935 			/*
936 			 * If the IO is via the VM then we do some
937 			 * special VM hackery (yuck).  Since the buffer's
938 			 * block size may not be page-aligned it is possible
939 			 * for a page to be shared between two buffers.  We
940 			 * have to get rid of the duplication when building
941 			 * the cluster.
942 			 */
943 			if (tbp->b_flags & B_VMIO) {
944 				vm_page_t m;
945 
946 				VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
947 				if (i != 0) { /* if not first buffer */
948 					for (j = 0; j < tbp->b_npages; j += 1) {
949 						m = tbp->b_pages[j];
950 						if (m->oflags & VPO_BUSY) {
951 							VM_OBJECT_WUNLOCK(
952 							    tbp->b_object);
953 							bqrelse(tbp);
954 							goto finishcluster;
955 						}
956 					}
957 				}
958 				for (j = 0; j < tbp->b_npages; j += 1) {
959 					m = tbp->b_pages[j];
960 					vm_page_io_start(m);
961 					vm_object_pip_add(m->object, 1);
962 					if ((bp->b_npages == 0) ||
963 					  (bp->b_pages[bp->b_npages - 1] != m)) {
964 						bp->b_pages[bp->b_npages] = m;
965 						bp->b_npages++;
966 					}
967 				}
968 				VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
969 			}
970 			bp->b_bcount += size;
971 			bp->b_bufsize += size;
972 			/*
973 			 * If any of the clustered buffers have their
974 			 * B_BARRIER flag set, transfer that request to
975 			 * the cluster.
976 			 */
977 			bp->b_flags |= (tbp->b_flags & B_BARRIER);
978 			tbp->b_flags &= ~(B_DONE | B_BARRIER);
979 			tbp->b_flags |= B_ASYNC;
980 			tbp->b_ioflags &= ~BIO_ERROR;
981 			tbp->b_iocmd = BIO_WRITE;
982 			bundirty(tbp);
983 			reassignbuf(tbp);		/* put on clean list */
984 			bufobj_wref(tbp->b_bufobj);
985 			BUF_KERNPROC(tbp);
986 			TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
987 				tbp, b_cluster.cluster_entry);
988 		}
989 	finishcluster:
990 		if ((bp->b_flags & B_UNMAPPED) == 0) {
991 			pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
992 			    (vm_page_t *)bp->b_pages, bp->b_npages);
993 		}
994 		if (bp->b_bufsize > bp->b_kvasize)
995 			panic(
996 			    "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
997 			    bp->b_bufsize, bp->b_kvasize);
998 		bp->b_kvasize = bp->b_bufsize;
999 		totalwritten += bp->b_bufsize;
1000 		bp->b_dirtyoff = 0;
1001 		bp->b_dirtyend = bp->b_bufsize;
1002 		bawrite(bp);
1003 
1004 		len -= i;
1005 	}
1006 	return totalwritten;
1007 }
1008 
1009 /*
1010  * Collect together all the buffers in a cluster.
1011  * Plus add one additional buffer.
1012  */
1013 static struct cluster_save *
1014 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags)
1015 {
1016 	struct cluster_save *buflist;
1017 	struct buf *bp;
1018 	daddr_t lbn;
1019 	int i, len;
1020 
1021 	len = vp->v_lastw - vp->v_cstart + 1;
1022 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1023 	    M_SEGMENT, M_WAITOK);
1024 	buflist->bs_nchildren = 0;
1025 	buflist->bs_children = (struct buf **) (buflist + 1);
1026 	for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1027 		(void)bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1028 		    gbflags, &bp);
1029 		buflist->bs_children[i] = bp;
1030 		if (bp->b_blkno == bp->b_lblkno)
1031 			VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1032 				NULL, NULL);
1033 	}
1034 	buflist->bs_children[i] = bp = last_bp;
1035 	if (bp->b_blkno == bp->b_lblkno)
1036 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1037 	buflist->bs_nchildren = i + 1;
1038 	return (buflist);
1039 }
1040