xref: /freebsd/sys/kern/vfs_cluster.c (revision 2546665afcaf0d53dc2c7058fee96354b3680f5a)
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/vmmeter.h>
50 #include <vm/vm.h>
51 #include <vm/vm_object.h>
52 #include <vm/vm_page.h>
53 #include <sys/sysctl.h>
54 
55 #if defined(CLUSTERDEBUG)
56 static int	rcluster= 0;
57 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0,
58     "Debug VFS clustering code");
59 #endif
60 
61 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer");
62 
63 static struct cluster_save *
64 	cluster_collectbufs(struct vnode *vp, struct buf *last_bp);
65 static struct buf *
66 	cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
67 			 daddr_t blkno, long size, int run, struct buf *fbp);
68 
69 static int write_behind = 1;
70 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
71     "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
72 
73 static int read_max = 8;
74 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
75     "Cluster read-ahead max block count");
76 
77 /* Page expended to mark partially backed buffers */
78 extern vm_page_t	bogus_page;
79 
80 /*
81  * Number of physical bufs (pbufs) this subsystem is allowed.
82  * Manipulated by vm_pager.c
83  */
84 extern int cluster_pbuf_freecnt;
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(vp, filesize, lblkno, size, cred, totread, seqcount, bpp)
92 	struct vnode *vp;
93 	u_quad_t filesize;
94 	daddr_t lblkno;
95 	long size;
96 	struct ucred *cred;
97 	long totread;
98 	int seqcount;
99 	struct buf **bpp;
100 {
101 	struct buf *bp, *rbp, *reqbp;
102 	daddr_t blkno, origblkno;
103 	int maxra, racluster;
104 	int error, ncontig;
105 	int i;
106 
107 	error = 0;
108 
109 	/*
110 	 * Try to limit the amount of read-ahead by a few
111 	 * ad-hoc parameters.  This needs work!!!
112 	 */
113 	racluster = vp->v_mount->mnt_iosize_max / size;
114 	maxra = seqcount;
115 	maxra = min(read_max, maxra);
116 	maxra = min(nbuf/8, maxra);
117 	if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
118 		maxra = (filesize / size) - lblkno;
119 
120 	/*
121 	 * get the requested block
122 	 */
123 	*bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, 0);
124 	origblkno = lblkno;
125 
126 	/*
127 	 * if it is in the cache, then check to see if the reads have been
128 	 * sequential.  If they have, then try some read-ahead, otherwise
129 	 * back-off on prospective read-aheads.
130 	 */
131 	if (bp->b_flags & B_CACHE) {
132 		if (!seqcount) {
133 			return 0;
134 		} else if ((bp->b_flags & B_RAM) == 0) {
135 			return 0;
136 		} else {
137 			int s;
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 			VI_LOCK(vp);
147 			for (i = 1; i < maxra; i++) {
148 				/*
149 				 * Stop if the buffer does not exist or it
150 				 * is invalid (about to go away?)
151 				 */
152 				rbp = gbincore(vp, lblkno+i);
153 				if (rbp == NULL || (rbp->b_flags & B_INVAL))
154 					break;
155 
156 				/*
157 				 * Set another read-ahead mark so we know
158 				 * to check again.
159 				 */
160 				if (((i % racluster) == (racluster - 1)) ||
161 					(i == (maxra - 1)))
162 					rbp->b_flags |= B_RAM;
163 			}
164 			VI_UNLOCK(vp);
165 			splx(s);
166 			if (i >= maxra) {
167 				return 0;
168 			}
169 			lblkno += i;
170 		}
171 		reqbp = bp = NULL;
172 	/*
173 	 * If it isn't in the cache, then get a chunk from
174 	 * disk if sequential, otherwise just get the block.
175 	 */
176 	} else {
177 		off_t firstread = bp->b_offset;
178 		int nblks;
179 
180 		KASSERT(bp->b_offset != NOOFFSET,
181 		    ("cluster_read: no buffer offset"));
182 
183 		ncontig = 0;
184 
185 		/*
186 		 * Compute the total number of blocks that we should read
187 		 * synchronously.
188 		 */
189 		if (firstread + totread > filesize)
190 			totread = filesize - firstread;
191 		nblks = howmany(totread, size);
192 		if (nblks > racluster)
193 			nblks = racluster;
194 
195 		/*
196 		 * Now compute the number of contiguous blocks.
197 		 */
198 		if (nblks > 1) {
199 	    		error = VOP_BMAP(vp, lblkno, NULL,
200 				&blkno, &ncontig, NULL);
201 			/*
202 			 * If this failed to map just do the original block.
203 			 */
204 			if (error || blkno == -1)
205 				ncontig = 0;
206 		}
207 
208 		/*
209 		 * If we have contiguous data available do a cluster
210 		 * otherwise just read the requested block.
211 		 */
212 		if (ncontig) {
213 			/* Account for our first block. */
214 			ncontig = min(ncontig + 1, nblks);
215 			if (ncontig < nblks)
216 				nblks = ncontig;
217 			bp = cluster_rbuild(vp, filesize, lblkno,
218 				blkno, size, nblks, bp);
219 			lblkno += (bp->b_bufsize / size);
220 		} else {
221 			bp->b_flags |= B_RAM;
222 			bp->b_iocmd = BIO_READ;
223 			lblkno += 1;
224 		}
225 	}
226 
227 	/*
228 	 * handle the synchronous read so that it is available ASAP.
229 	 */
230 	if (bp) {
231 		if ((bp->b_flags & B_CLUSTER) == 0) {
232 			vfs_busy_pages(bp, 0);
233 		}
234 		bp->b_flags &= ~B_INVAL;
235 		bp->b_ioflags &= ~BIO_ERROR;
236 		if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
237 			BUF_KERNPROC(bp);
238 		bp->b_iooffset = dbtob(bp->b_blkno);
239 		error = VOP_STRATEGY(vp, bp);
240 		curproc->p_stats->p_ru.ru_inblock++;
241 		if (error)
242 			return (error);
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, 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, 0);
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 		(void) VOP_STRATEGY(vp, rbp);
295 		curproc->p_stats->p_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(vp, filesize, lbn, blkno, size, run, fbp)
311 	struct vnode *vp;
312 	u_quad_t filesize;
313 	daddr_t lbn;
314 	daddr_t blkno;
315 	long size;
316 	int run;
317 	struct buf *fbp;
318 {
319 	struct buf *bp, *tbp;
320 	daddr_t bn;
321 	int i, inc, j;
322 
323 	GIANT_REQUIRED;
324 
325 	KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
326 	    ("cluster_rbuild: size %ld != filesize %jd\n",
327 	    size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
328 
329 	/*
330 	 * avoid a division
331 	 */
332 	while ((u_quad_t) size * (lbn + run) > filesize) {
333 		--run;
334 	}
335 
336 	if (fbp) {
337 		tbp = fbp;
338 		tbp->b_iocmd = BIO_READ;
339 	} else {
340 		tbp = getblk(vp, lbn, size, 0, 0, 0);
341 		if (tbp->b_flags & B_CACHE)
342 			return tbp;
343 		tbp->b_flags |= B_ASYNC | B_RAM;
344 		tbp->b_iocmd = BIO_READ;
345 	}
346 
347 	tbp->b_blkno = blkno;
348 	if( (tbp->b_flags & B_MALLOC) ||
349 		((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
350 		return tbp;
351 
352 	bp = trypbuf(&cluster_pbuf_freecnt);
353 	if (bp == 0)
354 		return tbp;
355 
356 	/*
357 	 * We are synthesizing a buffer out of vm_page_t's, but
358 	 * if the block size is not page aligned then the starting
359 	 * address may not be either.  Inherit the b_data offset
360 	 * from the original buffer.
361 	 */
362 	bp->b_data = (char *)((vm_offset_t)bp->b_data |
363 	    ((vm_offset_t)tbp->b_data & PAGE_MASK));
364 	bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
365 	bp->b_iocmd = BIO_READ;
366 	bp->b_iodone = cluster_callback;
367 	bp->b_blkno = blkno;
368 	bp->b_lblkno = lbn;
369 	bp->b_offset = tbp->b_offset;
370 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
371 	pbgetvp(vp, bp);
372 
373 	TAILQ_INIT(&bp->b_cluster.cluster_head);
374 
375 	bp->b_bcount = 0;
376 	bp->b_bufsize = 0;
377 	bp->b_npages = 0;
378 
379 	inc = btodb(size);
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 
389 			/* Don't wait around for locked bufs. */
390 			if (tbp == NULL)
391 				break;
392 
393 			/*
394 			 * Stop scanning if the buffer is fully valid
395 			 * (marked B_CACHE), or locked (may be doing a
396 			 * background write), or if the buffer is not
397 			 * VMIO backed.  The clustering code can only deal
398 			 * with VMIO-backed buffers.
399 			 */
400 			VI_LOCK(bp->b_vp);
401 			if ((tbp->b_vflags & BV_BKGRDINPROG) ||
402 			    (tbp->b_flags & B_CACHE) ||
403 			    (tbp->b_flags & B_VMIO) == 0) {
404 				VI_UNLOCK(bp->b_vp);
405 				bqrelse(tbp);
406 				break;
407 			}
408 			VI_UNLOCK(bp->b_vp);
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 			VM_OBJECT_LOCK(tbp->b_object);
416 			for (j = 0;j < tbp->b_npages; j++) {
417 				VM_OBJECT_LOCK_ASSERT(tbp->b_pages[j]->object,
418 				    MA_OWNED);
419 				if (tbp->b_pages[j]->valid)
420 					break;
421 			}
422 			VM_OBJECT_UNLOCK(tbp->b_object);
423 			if (j != tbp->b_npages) {
424 				bqrelse(tbp);
425 				break;
426 			}
427 
428 			/*
429 			 * Set a read-ahead mark as appropriate
430 			 */
431 			if ((fbp && (i == 1)) || (i == (run - 1)))
432 				tbp->b_flags |= B_RAM;
433 
434 			/*
435 			 * Set the buffer up for an async read (XXX should
436 			 * we do this only if we do not wind up brelse()ing?).
437 			 * Set the block number if it isn't set, otherwise
438 			 * if it is make sure it matches the block number we
439 			 * expect.
440 			 */
441 			tbp->b_flags |= B_ASYNC;
442 			tbp->b_iocmd = BIO_READ;
443 			if (tbp->b_blkno == tbp->b_lblkno) {
444 				tbp->b_blkno = bn;
445 			} else if (tbp->b_blkno != bn) {
446 				brelse(tbp);
447 				break;
448 			}
449 		}
450 		/*
451 		 * XXX fbp from caller may not be B_ASYNC, but we are going
452 		 * to biodone() it in cluster_callback() anyway
453 		 */
454 		BUF_KERNPROC(tbp);
455 		TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
456 			tbp, b_cluster.cluster_entry);
457 		VM_OBJECT_LOCK(tbp->b_object);
458 		vm_page_lock_queues();
459 		for (j = 0; j < tbp->b_npages; j += 1) {
460 			vm_page_t m;
461 			m = tbp->b_pages[j];
462 			vm_page_io_start(m);
463 			vm_object_pip_add(m->object, 1);
464 			if ((bp->b_npages == 0) ||
465 				(bp->b_pages[bp->b_npages-1] != m)) {
466 				bp->b_pages[bp->b_npages] = m;
467 				bp->b_npages++;
468 			}
469 			if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
470 				tbp->b_pages[j] = bogus_page;
471 		}
472 		vm_page_unlock_queues();
473 		VM_OBJECT_UNLOCK(tbp->b_object);
474 		/*
475 		 * XXX shouldn't this be += size for both, like in
476 		 * cluster_wbuild()?
477 		 *
478 		 * Don't inherit tbp->b_bufsize as it may be larger due to
479 		 * a non-page-aligned size.  Instead just aggregate using
480 		 * 'size'.
481 		 */
482 		if (tbp->b_bcount != size)
483 			printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
484 		if (tbp->b_bufsize != size)
485 			printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
486 		bp->b_bcount += size;
487 		bp->b_bufsize += size;
488 	}
489 
490 	/*
491 	 * Fully valid pages in the cluster are already good and do not need
492 	 * to be re-read from disk.  Replace the page with bogus_page
493 	 */
494 	VM_OBJECT_LOCK(bp->b_object);
495 	for (j = 0; j < bp->b_npages; j++) {
496 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[j]->object, MA_OWNED);
497 		if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
498 		    VM_PAGE_BITS_ALL) {
499 			bp->b_pages[j] = bogus_page;
500 		}
501 	}
502 	VM_OBJECT_UNLOCK(bp->b_object);
503 	if (bp->b_bufsize > bp->b_kvasize)
504 		panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
505 		    bp->b_bufsize, bp->b_kvasize);
506 	bp->b_kvasize = bp->b_bufsize;
507 
508 	pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
509 		(vm_page_t *)bp->b_pages, bp->b_npages);
510 	return (bp);
511 }
512 
513 /*
514  * Cleanup after a clustered read or write.
515  * This is complicated by the fact that any of the buffers might have
516  * extra memory (if there were no empty buffer headers at allocbuf time)
517  * that we will need to shift around.
518  */
519 void
520 cluster_callback(bp)
521 	struct buf *bp;
522 {
523 	struct buf *nbp, *tbp;
524 	int error = 0;
525 
526 	GIANT_REQUIRED;
527 
528 	/*
529 	 * Must propogate errors to all the components.
530 	 */
531 	if (bp->b_ioflags & BIO_ERROR)
532 		error = bp->b_error;
533 
534 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
535 	/*
536 	 * Move memory from the large cluster buffer into the component
537 	 * buffers and mark IO as done on these.
538 	 */
539 	for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
540 		tbp; tbp = nbp) {
541 		nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
542 		if (error) {
543 			tbp->b_ioflags |= BIO_ERROR;
544 			tbp->b_error = error;
545 		} else {
546 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
547 			tbp->b_flags &= ~B_INVAL;
548 			tbp->b_ioflags &= ~BIO_ERROR;
549 			/*
550 			 * XXX the bdwrite()/bqrelse() issued during
551 			 * cluster building clears B_RELBUF (see bqrelse()
552 			 * comment).  If direct I/O was specified, we have
553 			 * to restore it here to allow the buffer and VM
554 			 * to be freed.
555 			 */
556 			if (tbp->b_flags & B_DIRECT)
557 				tbp->b_flags |= B_RELBUF;
558 		}
559 		bufdone(tbp);
560 	}
561 	relpbuf(bp, &cluster_pbuf_freecnt);
562 }
563 
564 /*
565  *	cluster_wbuild_wb:
566  *
567  *	Implement modified write build for cluster.
568  *
569  *		write_behind = 0	write behind disabled
570  *		write_behind = 1	write behind normal (default)
571  *		write_behind = 2	write behind backed-off
572  */
573 
574 static __inline int
575 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
576 {
577 	int r = 0;
578 
579 	switch(write_behind) {
580 	case 2:
581 		if (start_lbn < len)
582 			break;
583 		start_lbn -= len;
584 		/* FALLTHROUGH */
585 	case 1:
586 		r = cluster_wbuild(vp, size, start_lbn, len);
587 		/* FALLTHROUGH */
588 	default:
589 		/* FALLTHROUGH */
590 		break;
591 	}
592 	return(r);
593 }
594 
595 /*
596  * Do clustered write for FFS.
597  *
598  * Three cases:
599  *	1. Write is not sequential (write asynchronously)
600  *	Write is sequential:
601  *	2.	beginning of cluster - begin cluster
602  *	3.	middle of a cluster - add to cluster
603  *	4.	end of a cluster - asynchronously write cluster
604  */
605 void
606 cluster_write(bp, filesize, seqcount)
607 	struct buf *bp;
608 	u_quad_t filesize;
609 	int seqcount;
610 {
611 	struct vnode *vp;
612 	daddr_t lbn;
613 	int maxclen, cursize;
614 	int lblocksize;
615 	int async;
616 
617 	vp = bp->b_vp;
618 	if (vp->v_type == VREG) {
619 		async = vp->v_mount->mnt_flag & MNT_ASYNC;
620 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
621 	} else {
622 		async = 0;
623 		lblocksize = bp->b_bufsize;
624 	}
625 	lbn = bp->b_lblkno;
626 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
627 
628 	/* Initialize vnode to beginning of file. */
629 	if (lbn == 0)
630 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
631 
632 	if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
633 	    (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
634 		maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
635 		if (vp->v_clen != 0) {
636 			/*
637 			 * Next block is not sequential.
638 			 *
639 			 * If we are not writing at end of file, the process
640 			 * seeked to another point in the file since its last
641 			 * write, or we have reached our maximum cluster size,
642 			 * then push the previous cluster. Otherwise try
643 			 * reallocating to make it sequential.
644 			 *
645 			 * Change to algorithm: only push previous cluster if
646 			 * it was sequential from the point of view of the
647 			 * seqcount heuristic, otherwise leave the buffer
648 			 * intact so we can potentially optimize the I/O
649 			 * later on in the buf_daemon or update daemon
650 			 * flush.
651 			 */
652 			cursize = vp->v_lastw - vp->v_cstart + 1;
653 			if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
654 			    lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
655 				if (!async && seqcount > 0) {
656 					cluster_wbuild_wb(vp, lblocksize,
657 						vp->v_cstart, cursize);
658 				}
659 			} else {
660 				struct buf **bpp, **endbp;
661 				struct cluster_save *buflist;
662 
663 				buflist = cluster_collectbufs(vp, bp);
664 				endbp = &buflist->bs_children
665 				    [buflist->bs_nchildren - 1];
666 				if (VOP_REALLOCBLKS(vp, buflist)) {
667 					/*
668 					 * Failed, push the previous cluster
669 					 * if *really* writing sequentially
670 					 * in the logical file (seqcount > 1),
671 					 * otherwise delay it in the hopes that
672 					 * the low level disk driver can
673 					 * optimize the write ordering.
674 					 */
675 					for (bpp = buflist->bs_children;
676 					     bpp < endbp; bpp++)
677 						brelse(*bpp);
678 					free(buflist, M_SEGMENT);
679 					if (seqcount > 1) {
680 						cluster_wbuild_wb(vp,
681 						    lblocksize, vp->v_cstart,
682 						    cursize);
683 					}
684 				} else {
685 					/*
686 					 * Succeeded, keep building cluster.
687 					 */
688 					for (bpp = buflist->bs_children;
689 					     bpp <= endbp; bpp++)
690 						bdwrite(*bpp);
691 					free(buflist, M_SEGMENT);
692 					vp->v_lastw = lbn;
693 					vp->v_lasta = bp->b_blkno;
694 					return;
695 				}
696 			}
697 		}
698 		/*
699 		 * Consider beginning a cluster. If at end of file, make
700 		 * cluster as large as possible, otherwise find size of
701 		 * existing cluster.
702 		 */
703 		if ((vp->v_type == VREG) &&
704 			((u_quad_t) bp->b_offset + lblocksize) != filesize &&
705 		    (bp->b_blkno == bp->b_lblkno) &&
706 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
707 		     bp->b_blkno == -1)) {
708 			bawrite(bp);
709 			vp->v_clen = 0;
710 			vp->v_lasta = bp->b_blkno;
711 			vp->v_cstart = lbn + 1;
712 			vp->v_lastw = lbn;
713 			return;
714 		}
715 		vp->v_clen = maxclen;
716 		if (!async && maxclen == 0) {	/* I/O not contiguous */
717 			vp->v_cstart = lbn + 1;
718 			bawrite(bp);
719 		} else {	/* Wait for rest of cluster */
720 			vp->v_cstart = lbn;
721 			bdwrite(bp);
722 		}
723 	} else if (lbn == vp->v_cstart + vp->v_clen) {
724 		/*
725 		 * At end of cluster, write it out if seqcount tells us we
726 		 * are operating sequentially, otherwise let the buf or
727 		 * update daemon handle it.
728 		 */
729 		bdwrite(bp);
730 		if (seqcount > 1)
731 			cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
732 		vp->v_clen = 0;
733 		vp->v_cstart = lbn + 1;
734 	} else if (vm_page_count_severe()) {
735 		/*
736 		 * We are low on memory, get it going NOW
737 		 */
738 		bawrite(bp);
739 	} else {
740 		/*
741 		 * In the middle of a cluster, so just delay the I/O for now.
742 		 */
743 		bdwrite(bp);
744 	}
745 	vp->v_lastw = lbn;
746 	vp->v_lasta = bp->b_blkno;
747 }
748 
749 
750 /*
751  * This is an awful lot like cluster_rbuild...wish they could be combined.
752  * The last lbn argument is the current block on which I/O is being
753  * performed.  Check to see that it doesn't fall in the middle of
754  * the current block (if last_bp == NULL).
755  */
756 int
757 cluster_wbuild(vp, size, start_lbn, len)
758 	struct vnode *vp;
759 	long size;
760 	daddr_t start_lbn;
761 	int len;
762 {
763 	struct buf *bp, *tbp;
764 	int i, j, s;
765 	int totalwritten = 0;
766 	int dbsize = btodb(size);
767 
768 	GIANT_REQUIRED;
769 
770 	while (len > 0) {
771 		s = splbio();
772 		/*
773 		 * If the buffer is not delayed-write (i.e. dirty), or it
774 		 * is delayed-write but either locked or inval, it cannot
775 		 * partake in the clustered write.
776 		 */
777 		VI_LOCK(vp);
778 		if ((tbp = gbincore(vp, start_lbn)) == NULL ||
779 		    (tbp->b_vflags & BV_BKGRDINPROG)) {
780 			VI_UNLOCK(vp);
781 			++start_lbn;
782 			--len;
783 			splx(s);
784 			continue;
785 		}
786 		if (BUF_LOCK(tbp,
787 		    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, VI_MTX(vp))) {
788 			++start_lbn;
789 			--len;
790 			splx(s);
791 			continue;
792 		}
793 		if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
794 			BUF_UNLOCK(tbp);
795 			++start_lbn;
796 			--len;
797 			splx(s);
798 			continue;
799 		}
800 		bremfree(tbp);
801 		tbp->b_flags &= ~B_DONE;
802 		splx(s);
803 
804 		/*
805 		 * Extra memory in the buffer, punt on this buffer.
806 		 * XXX we could handle this in most cases, but we would
807 		 * have to push the extra memory down to after our max
808 		 * possible cluster size and then potentially pull it back
809 		 * up if the cluster was terminated prematurely--too much
810 		 * hassle.
811 		 */
812 		if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
813 		     (B_CLUSTEROK | B_VMIO)) ||
814 		  (tbp->b_bcount != tbp->b_bufsize) ||
815 		  (tbp->b_bcount != size) ||
816 		  (len == 1) ||
817 		  ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
818 			totalwritten += tbp->b_bufsize;
819 			bawrite(tbp);
820 			++start_lbn;
821 			--len;
822 			continue;
823 		}
824 
825 		/*
826 		 * We got a pbuf to make the cluster in.
827 		 * so initialise it.
828 		 */
829 		TAILQ_INIT(&bp->b_cluster.cluster_head);
830 		bp->b_bcount = 0;
831 		bp->b_magic = tbp->b_magic;
832 		bp->b_op = tbp->b_op;
833 		bp->b_bufsize = 0;
834 		bp->b_npages = 0;
835 		if (tbp->b_wcred != NOCRED)
836 			bp->b_wcred = crhold(tbp->b_wcred);
837 
838 		bp->b_blkno = tbp->b_blkno;
839 		bp->b_lblkno = tbp->b_lblkno;
840 		bp->b_offset = tbp->b_offset;
841 
842 		/*
843 		 * We are synthesizing a buffer out of vm_page_t's, but
844 		 * if the block size is not page aligned then the starting
845 		 * address may not be either.  Inherit the b_data offset
846 		 * from the original buffer.
847 		 */
848 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
849 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
850 		bp->b_flags |= B_CLUSTER |
851 				(tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
852 		bp->b_iodone = cluster_callback;
853 		pbgetvp(vp, bp);
854 		/*
855 		 * From this location in the file, scan forward to see
856 		 * if there are buffers with adjacent data that need to
857 		 * be written as well.
858 		 */
859 		for (i = 0; i < len; ++i, ++start_lbn) {
860 			if (i != 0) { /* If not the first buffer */
861 				s = splbio();
862 				/*
863 				 * If the adjacent data is not even in core it
864 				 * can't need to be written.
865 				 */
866 				VI_LOCK(vp);
867 				if ((tbp = gbincore(vp, start_lbn)) == NULL ||
868 				    (tbp->b_vflags & BV_BKGRDINPROG)) {
869 					VI_UNLOCK(vp);
870 					splx(s);
871 					break;
872 				}
873 
874 				/*
875 				 * If it IS in core, but has different
876 				 * characteristics, or is locked (which
877 				 * means it could be undergoing a background
878 				 * I/O or be in a weird state), then don't
879 				 * cluster with it.
880 				 */
881 				if (BUF_LOCK(tbp,
882 				    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
883 				    VI_MTX(vp))) {
884 					splx(s);
885 					break;
886 				}
887 
888 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
889 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
890 				    != (B_DELWRI | B_CLUSTEROK |
891 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
892 				    tbp->b_wcred != bp->b_wcred) {
893 					BUF_UNLOCK(tbp);
894 					splx(s);
895 					break;
896 				}
897 
898 				/*
899 				 * Check that the combined cluster
900 				 * would make sense with regard to pages
901 				 * and would not be too large
902 				 */
903 				if ((tbp->b_bcount != size) ||
904 				  ((bp->b_blkno + (dbsize * i)) !=
905 				    tbp->b_blkno) ||
906 				  ((tbp->b_npages + bp->b_npages) >
907 				    (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
908 					BUF_UNLOCK(tbp);
909 					splx(s);
910 					break;
911 				}
912 				/*
913 				 * Ok, it's passed all the tests,
914 				 * so remove it from the free list
915 				 * and mark it busy. We will use it.
916 				 */
917 				bremfree(tbp);
918 				tbp->b_flags &= ~B_DONE;
919 				splx(s);
920 			} /* end of code for non-first buffers only */
921 			/* check for latent dependencies to be handled */
922 			if ((LIST_FIRST(&tbp->b_dep)) != NULL) {
923 				tbp->b_iocmd = BIO_WRITE;
924 				buf_start(tbp);
925 			}
926 			/*
927 			 * If the IO is via the VM then we do some
928 			 * special VM hackery (yuck).  Since the buffer's
929 			 * block size may not be page-aligned it is possible
930 			 * for a page to be shared between two buffers.  We
931 			 * have to get rid of the duplication when building
932 			 * the cluster.
933 			 */
934 			if (tbp->b_flags & B_VMIO) {
935 				vm_page_t m;
936 
937 				if (i != 0) { /* if not first buffer */
938 					for (j = 0; j < tbp->b_npages; j += 1) {
939 						m = tbp->b_pages[j];
940 						if (m->flags & PG_BUSY) {
941 							bqrelse(tbp);
942 							goto finishcluster;
943 						}
944 					}
945 				}
946 				if (tbp->b_object != NULL)
947 					VM_OBJECT_LOCK(tbp->b_object);
948 				vm_page_lock_queues();
949 				for (j = 0; j < tbp->b_npages; j += 1) {
950 					m = tbp->b_pages[j];
951 					vm_page_io_start(m);
952 					vm_object_pip_add(m->object, 1);
953 					if ((bp->b_npages == 0) ||
954 					  (bp->b_pages[bp->b_npages - 1] != m)) {
955 						bp->b_pages[bp->b_npages] = m;
956 						bp->b_npages++;
957 					}
958 				}
959 				vm_page_unlock_queues();
960 				if (tbp->b_object != NULL)
961 					VM_OBJECT_UNLOCK(tbp->b_object);
962 			}
963 			bp->b_bcount += size;
964 			bp->b_bufsize += size;
965 
966 			s = splbio();
967 			bundirty(tbp);
968 			tbp->b_flags &= ~B_DONE;
969 			tbp->b_ioflags &= ~BIO_ERROR;
970 			tbp->b_flags |= B_ASYNC;
971 			tbp->b_iocmd = BIO_WRITE;
972 			reassignbuf(tbp);		/* put on clean list */
973 			VI_LOCK(tbp->b_vp);
974 			++tbp->b_vp->v_numoutput;
975 			VI_UNLOCK(tbp->b_vp);
976 			splx(s);
977 			BUF_KERNPROC(tbp);
978 			TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
979 				tbp, b_cluster.cluster_entry);
980 		}
981 	finishcluster:
982 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
983 			(vm_page_t *) bp->b_pages, bp->b_npages);
984 		if (bp->b_bufsize > bp->b_kvasize)
985 			panic(
986 			    "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
987 			    bp->b_bufsize, bp->b_kvasize);
988 		bp->b_kvasize = bp->b_bufsize;
989 		totalwritten += bp->b_bufsize;
990 		bp->b_dirtyoff = 0;
991 		bp->b_dirtyend = bp->b_bufsize;
992 		bawrite(bp);
993 
994 		len -= i;
995 	}
996 	return totalwritten;
997 }
998 
999 /*
1000  * Collect together all the buffers in a cluster.
1001  * Plus add one additional buffer.
1002  */
1003 static struct cluster_save *
1004 cluster_collectbufs(vp, last_bp)
1005 	struct vnode *vp;
1006 	struct buf *last_bp;
1007 {
1008 	struct cluster_save *buflist;
1009 	struct buf *bp;
1010 	daddr_t lbn;
1011 	int i, len;
1012 
1013 	len = vp->v_lastw - vp->v_cstart + 1;
1014 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1015 	    M_SEGMENT, M_WAITOK);
1016 	buflist->bs_nchildren = 0;
1017 	buflist->bs_children = (struct buf **) (buflist + 1);
1018 	for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1019 		(void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
1020 		buflist->bs_children[i] = bp;
1021 		if (bp->b_blkno == bp->b_lblkno)
1022 			VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
1023 				NULL, NULL);
1024 	}
1025 	buflist->bs_children[i] = bp = last_bp;
1026 	if (bp->b_blkno == bp->b_lblkno)
1027 		VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
1028 			NULL, NULL);
1029 	buflist->bs_nchildren = i + 1;
1030 	return (buflist);
1031 }
1032