xref: /freebsd/sys/kern/vfs_cluster.c (revision f4b37ed0f8b307b1f3f0f630ca725d68f1dff30d)
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 buf *bp, *tbp;
314 	daddr_t bn;
315 	off_t off;
316 	long tinc, tsize;
317 	int i, inc, j, k, toff;
318 
319 	KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
320 	    ("cluster_rbuild: size %ld != f_iosize %jd\n",
321 	    size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
322 
323 	/*
324 	 * avoid a division
325 	 */
326 	while ((u_quad_t) size * (lbn + run) > filesize) {
327 		--run;
328 	}
329 
330 	if (fbp) {
331 		tbp = fbp;
332 		tbp->b_iocmd = BIO_READ;
333 	} else {
334 		tbp = getblk(vp, lbn, size, 0, 0, gbflags);
335 		if (tbp->b_flags & B_CACHE)
336 			return tbp;
337 		tbp->b_flags |= B_ASYNC | B_RAM;
338 		tbp->b_iocmd = BIO_READ;
339 	}
340 	tbp->b_blkno = blkno;
341 	if( (tbp->b_flags & B_MALLOC) ||
342 		((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
343 		return tbp;
344 
345 	bp = trypbuf(&cluster_pbuf_freecnt);
346 	if (bp == 0)
347 		return tbp;
348 
349 	/*
350 	 * We are synthesizing a buffer out of vm_page_t's, but
351 	 * if the block size is not page aligned then the starting
352 	 * address may not be either.  Inherit the b_data offset
353 	 * from the original buffer.
354 	 */
355 	bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
356 	if ((gbflags & GB_UNMAPPED) != 0) {
357 		bp->b_data = unmapped_buf;
358 	} else {
359 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
360 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
361 	}
362 	bp->b_iocmd = BIO_READ;
363 	bp->b_iodone = cluster_callback;
364 	bp->b_blkno = blkno;
365 	bp->b_lblkno = lbn;
366 	bp->b_offset = tbp->b_offset;
367 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
368 	pbgetvp(vp, bp);
369 
370 	TAILQ_INIT(&bp->b_cluster.cluster_head);
371 
372 	bp->b_bcount = 0;
373 	bp->b_bufsize = 0;
374 	bp->b_npages = 0;
375 
376 	inc = btodb(size);
377 	for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
378 		if (i == 0) {
379 			VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
380 			vfs_drain_busy_pages(tbp);
381 			vm_object_pip_add(tbp->b_bufobj->bo_object,
382 			    tbp->b_npages);
383 			for (k = 0; k < tbp->b_npages; k++)
384 				vm_page_sbusy(tbp->b_pages[k]);
385 			VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
386 		} else {
387 			if ((bp->b_npages * PAGE_SIZE) +
388 			    round_page(size) > vp->v_mount->mnt_iosize_max) {
389 				break;
390 			}
391 
392 			tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
393 			    (gbflags & GB_UNMAPPED));
394 
395 			/* Don't wait around for locked bufs. */
396 			if (tbp == NULL)
397 				break;
398 
399 			/*
400 			 * Stop scanning if the buffer is fully valid
401 			 * (marked B_CACHE), or locked (may be doing a
402 			 * background write), or if the buffer is not
403 			 * VMIO backed.  The clustering code can only deal
404 			 * with VMIO-backed buffers.  The bo lock is not
405 			 * required for the BKGRDINPROG check since it
406 			 * can not be set without the buf lock.
407 			 */
408 			if ((tbp->b_vflags & BV_BKGRDINPROG) ||
409 			    (tbp->b_flags & B_CACHE) ||
410 			    (tbp->b_flags & B_VMIO) == 0) {
411 				bqrelse(tbp);
412 				break;
413 			}
414 
415 			/*
416 			 * The buffer must be completely invalid in order to
417 			 * take part in the cluster.  If it is partially valid
418 			 * then we stop.
419 			 */
420 			off = tbp->b_offset;
421 			tsize = size;
422 			VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
423 			for (j = 0; tsize > 0; j++) {
424 				toff = off & PAGE_MASK;
425 				tinc = tsize;
426 				if (toff + tinc > PAGE_SIZE)
427 					tinc = PAGE_SIZE - toff;
428 				VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object);
429 				if ((tbp->b_pages[j]->valid &
430 				    vm_page_bits(toff, tinc)) != 0)
431 					break;
432 				if (vm_page_xbusied(tbp->b_pages[j]))
433 					break;
434 				vm_object_pip_add(tbp->b_bufobj->bo_object, 1);
435 				vm_page_sbusy(tbp->b_pages[j]);
436 				off += tinc;
437 				tsize -= tinc;
438 			}
439 			if (tsize > 0) {
440 clean_sbusy:
441 				vm_object_pip_add(tbp->b_bufobj->bo_object, -j);
442 				for (k = 0; k < j; k++)
443 					vm_page_sunbusy(tbp->b_pages[k]);
444 				VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
445 				bqrelse(tbp);
446 				break;
447 			}
448 			VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
449 
450 			/*
451 			 * Set a read-ahead mark as appropriate
452 			 */
453 			if ((fbp && (i == 1)) || (i == (run - 1)))
454 				tbp->b_flags |= B_RAM;
455 
456 			/*
457 			 * Set the buffer up for an async read (XXX should
458 			 * we do this only if we do not wind up brelse()ing?).
459 			 * Set the block number if it isn't set, otherwise
460 			 * if it is make sure it matches the block number we
461 			 * expect.
462 			 */
463 			tbp->b_flags |= B_ASYNC;
464 			tbp->b_iocmd = BIO_READ;
465 			if (tbp->b_blkno == tbp->b_lblkno) {
466 				tbp->b_blkno = bn;
467 			} else if (tbp->b_blkno != bn) {
468 				VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
469 				goto clean_sbusy;
470 			}
471 		}
472 		/*
473 		 * XXX fbp from caller may not be B_ASYNC, but we are going
474 		 * to biodone() it in cluster_callback() anyway
475 		 */
476 		BUF_KERNPROC(tbp);
477 		TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
478 			tbp, b_cluster.cluster_entry);
479 		VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
480 		for (j = 0; j < tbp->b_npages; j += 1) {
481 			vm_page_t m;
482 			m = tbp->b_pages[j];
483 			if ((bp->b_npages == 0) ||
484 			    (bp->b_pages[bp->b_npages-1] != m)) {
485 				bp->b_pages[bp->b_npages] = m;
486 				bp->b_npages++;
487 			}
488 			if (m->valid == VM_PAGE_BITS_ALL)
489 				tbp->b_pages[j] = bogus_page;
490 		}
491 		VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
492 		/*
493 		 * Don't inherit tbp->b_bufsize as it may be larger due to
494 		 * a non-page-aligned size.  Instead just aggregate using
495 		 * 'size'.
496 		 */
497 		if (tbp->b_bcount != size)
498 			printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
499 		if (tbp->b_bufsize != size)
500 			printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
501 		bp->b_bcount += size;
502 		bp->b_bufsize += size;
503 	}
504 
505 	/*
506 	 * Fully valid pages in the cluster are already good and do not need
507 	 * to be re-read from disk.  Replace the page with bogus_page
508 	 */
509 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
510 	for (j = 0; j < bp->b_npages; j++) {
511 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object);
512 		if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
513 			bp->b_pages[j] = bogus_page;
514 	}
515 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
516 	if (bp->b_bufsize > bp->b_kvasize)
517 		panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
518 		    bp->b_bufsize, bp->b_kvasize);
519 
520 	if (buf_mapped(bp)) {
521 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
522 		    (vm_page_t *)bp->b_pages, bp->b_npages);
523 	}
524 	return (bp);
525 }
526 
527 /*
528  * Cleanup after a clustered read or write.
529  * This is complicated by the fact that any of the buffers might have
530  * extra memory (if there were no empty buffer headers at allocbuf time)
531  * that we will need to shift around.
532  */
533 static void
534 cluster_callback(bp)
535 	struct buf *bp;
536 {
537 	struct buf *nbp, *tbp;
538 	int error = 0;
539 
540 	/*
541 	 * Must propogate errors to all the components.
542 	 */
543 	if (bp->b_ioflags & BIO_ERROR)
544 		error = bp->b_error;
545 
546 	if (buf_mapped(bp)) {
547 		pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
548 		    bp->b_npages);
549 	}
550 	/*
551 	 * Move memory from the large cluster buffer into the component
552 	 * buffers and mark IO as done on these.
553 	 */
554 	for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
555 		tbp; tbp = nbp) {
556 		nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
557 		if (error) {
558 			tbp->b_ioflags |= BIO_ERROR;
559 			tbp->b_error = error;
560 		} else {
561 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
562 			tbp->b_flags &= ~B_INVAL;
563 			tbp->b_ioflags &= ~BIO_ERROR;
564 			/*
565 			 * XXX the bdwrite()/bqrelse() issued during
566 			 * cluster building clears B_RELBUF (see bqrelse()
567 			 * comment).  If direct I/O was specified, we have
568 			 * to restore it here to allow the buffer and VM
569 			 * to be freed.
570 			 */
571 			if (tbp->b_flags & B_DIRECT)
572 				tbp->b_flags |= B_RELBUF;
573 		}
574 		bufdone(tbp);
575 	}
576 	pbrelvp(bp);
577 	relpbuf(bp, &cluster_pbuf_freecnt);
578 }
579 
580 /*
581  *	cluster_wbuild_wb:
582  *
583  *	Implement modified write build for cluster.
584  *
585  *		write_behind = 0	write behind disabled
586  *		write_behind = 1	write behind normal (default)
587  *		write_behind = 2	write behind backed-off
588  */
589 
590 static __inline int
591 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
592     int gbflags)
593 {
594 	int r = 0;
595 
596 	switch (write_behind) {
597 	case 2:
598 		if (start_lbn < len)
599 			break;
600 		start_lbn -= len;
601 		/* FALLTHROUGH */
602 	case 1:
603 		r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
604 		/* FALLTHROUGH */
605 	default:
606 		/* FALLTHROUGH */
607 		break;
608 	}
609 	return(r);
610 }
611 
612 /*
613  * Do clustered write for FFS.
614  *
615  * Three cases:
616  *	1. Write is not sequential (write asynchronously)
617  *	Write is sequential:
618  *	2.	beginning of cluster - begin cluster
619  *	3.	middle of a cluster - add to cluster
620  *	4.	end of a cluster - asynchronously write cluster
621  */
622 void
623 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount,
624     int gbflags)
625 {
626 	daddr_t lbn;
627 	int maxclen, cursize;
628 	int lblocksize;
629 	int async;
630 
631 	if (!unmapped_buf_allowed)
632 		gbflags &= ~GB_UNMAPPED;
633 
634 	if (vp->v_type == VREG) {
635 		async = DOINGASYNC(vp);
636 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
637 	} else {
638 		async = 0;
639 		lblocksize = bp->b_bufsize;
640 	}
641 	lbn = bp->b_lblkno;
642 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
643 
644 	/* Initialize vnode to beginning of file. */
645 	if (lbn == 0)
646 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
647 
648 	if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
649 	    (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
650 		maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
651 		if (vp->v_clen != 0) {
652 			/*
653 			 * Next block is not sequential.
654 			 *
655 			 * If we are not writing at end of file, the process
656 			 * seeked to another point in the file since its last
657 			 * write, or we have reached our maximum cluster size,
658 			 * then push the previous cluster. Otherwise try
659 			 * reallocating to make it sequential.
660 			 *
661 			 * Change to algorithm: only push previous cluster if
662 			 * it was sequential from the point of view of the
663 			 * seqcount heuristic, otherwise leave the buffer
664 			 * intact so we can potentially optimize the I/O
665 			 * later on in the buf_daemon or update daemon
666 			 * flush.
667 			 */
668 			cursize = vp->v_lastw - vp->v_cstart + 1;
669 			if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
670 			    lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
671 				if (!async && seqcount > 0) {
672 					cluster_wbuild_wb(vp, lblocksize,
673 					    vp->v_cstart, cursize, gbflags);
674 				}
675 			} else {
676 				struct buf **bpp, **endbp;
677 				struct cluster_save *buflist;
678 
679 				buflist = cluster_collectbufs(vp, bp, gbflags);
680 				endbp = &buflist->bs_children
681 				    [buflist->bs_nchildren - 1];
682 				if (VOP_REALLOCBLKS(vp, buflist)) {
683 					/*
684 					 * Failed, push the previous cluster
685 					 * if *really* writing sequentially
686 					 * in the logical file (seqcount > 1),
687 					 * otherwise delay it in the hopes that
688 					 * the low level disk driver can
689 					 * optimize the write ordering.
690 					 */
691 					for (bpp = buflist->bs_children;
692 					     bpp < endbp; bpp++)
693 						brelse(*bpp);
694 					free(buflist, M_SEGMENT);
695 					if (seqcount > 1) {
696 						cluster_wbuild_wb(vp,
697 						    lblocksize, vp->v_cstart,
698 						    cursize, gbflags);
699 					}
700 				} else {
701 					/*
702 					 * Succeeded, keep building cluster.
703 					 */
704 					for (bpp = buflist->bs_children;
705 					     bpp <= endbp; bpp++)
706 						bdwrite(*bpp);
707 					free(buflist, M_SEGMENT);
708 					vp->v_lastw = lbn;
709 					vp->v_lasta = bp->b_blkno;
710 					return;
711 				}
712 			}
713 		}
714 		/*
715 		 * Consider beginning a cluster. If at end of file, make
716 		 * cluster as large as possible, otherwise find size of
717 		 * existing cluster.
718 		 */
719 		if ((vp->v_type == VREG) &&
720 			((u_quad_t) bp->b_offset + lblocksize) != filesize &&
721 		    (bp->b_blkno == bp->b_lblkno) &&
722 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
723 		     bp->b_blkno == -1)) {
724 			bawrite(bp);
725 			vp->v_clen = 0;
726 			vp->v_lasta = bp->b_blkno;
727 			vp->v_cstart = lbn + 1;
728 			vp->v_lastw = lbn;
729 			return;
730 		}
731 		vp->v_clen = maxclen;
732 		if (!async && maxclen == 0) {	/* I/O not contiguous */
733 			vp->v_cstart = lbn + 1;
734 			bawrite(bp);
735 		} else {	/* Wait for rest of cluster */
736 			vp->v_cstart = lbn;
737 			bdwrite(bp);
738 		}
739 	} else if (lbn == vp->v_cstart + vp->v_clen) {
740 		/*
741 		 * At end of cluster, write it out if seqcount tells us we
742 		 * are operating sequentially, otherwise let the buf or
743 		 * update daemon handle it.
744 		 */
745 		bdwrite(bp);
746 		if (seqcount > 1) {
747 			cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
748 			    vp->v_clen + 1, gbflags);
749 		}
750 		vp->v_clen = 0;
751 		vp->v_cstart = lbn + 1;
752 	} else if (vm_page_count_severe()) {
753 		/*
754 		 * We are low on memory, get it going NOW
755 		 */
756 		bawrite(bp);
757 	} else {
758 		/*
759 		 * In the middle of a cluster, so just delay the I/O for now.
760 		 */
761 		bdwrite(bp);
762 	}
763 	vp->v_lastw = lbn;
764 	vp->v_lasta = bp->b_blkno;
765 }
766 
767 
768 /*
769  * This is an awful lot like cluster_rbuild...wish they could be combined.
770  * The last lbn argument is the current block on which I/O is being
771  * performed.  Check to see that it doesn't fall in the middle of
772  * the current block (if last_bp == NULL).
773  */
774 int
775 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
776     int gbflags)
777 {
778 	struct buf *bp, *tbp;
779 	struct bufobj *bo;
780 	int i, j;
781 	int totalwritten = 0;
782 	int dbsize = btodb(size);
783 
784 	if (!unmapped_buf_allowed)
785 		gbflags &= ~GB_UNMAPPED;
786 
787 	bo = &vp->v_bufobj;
788 	while (len > 0) {
789 		/*
790 		 * If the buffer is not delayed-write (i.e. dirty), or it
791 		 * is delayed-write but either locked or inval, it cannot
792 		 * partake in the clustered write.
793 		 */
794 		BO_LOCK(bo);
795 		if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
796 		    (tbp->b_vflags & BV_BKGRDINPROG)) {
797 			BO_UNLOCK(bo);
798 			++start_lbn;
799 			--len;
800 			continue;
801 		}
802 		if (BUF_LOCK(tbp,
803 		    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) {
804 			++start_lbn;
805 			--len;
806 			continue;
807 		}
808 		if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
809 			BUF_UNLOCK(tbp);
810 			++start_lbn;
811 			--len;
812 			continue;
813 		}
814 		if (tbp->b_pin_count >  0) {
815 			BUF_UNLOCK(tbp);
816 			++start_lbn;
817 			--len;
818 			continue;
819 		}
820 		bremfree(tbp);
821 		tbp->b_flags &= ~B_DONE;
822 
823 		/*
824 		 * Extra memory in the buffer, punt on this buffer.
825 		 * XXX we could handle this in most cases, but we would
826 		 * have to push the extra memory down to after our max
827 		 * possible cluster size and then potentially pull it back
828 		 * up if the cluster was terminated prematurely--too much
829 		 * hassle.
830 		 */
831 		if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
832 		     (B_CLUSTEROK | B_VMIO)) ||
833 		  (tbp->b_bcount != tbp->b_bufsize) ||
834 		  (tbp->b_bcount != size) ||
835 		  (len == 1) ||
836 		  ((bp = (vp->v_vflag & VV_MD) != 0 ?
837 		  trypbuf(&cluster_pbuf_freecnt) :
838 		  getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
839 			totalwritten += tbp->b_bufsize;
840 			bawrite(tbp);
841 			++start_lbn;
842 			--len;
843 			continue;
844 		}
845 
846 		/*
847 		 * We got a pbuf to make the cluster in.
848 		 * so initialise it.
849 		 */
850 		TAILQ_INIT(&bp->b_cluster.cluster_head);
851 		bp->b_bcount = 0;
852 		bp->b_bufsize = 0;
853 		bp->b_npages = 0;
854 		if (tbp->b_wcred != NOCRED)
855 			bp->b_wcred = crhold(tbp->b_wcred);
856 
857 		bp->b_blkno = tbp->b_blkno;
858 		bp->b_lblkno = tbp->b_lblkno;
859 		bp->b_offset = tbp->b_offset;
860 
861 		/*
862 		 * We are synthesizing a buffer out of vm_page_t's, but
863 		 * if the block size is not page aligned then the starting
864 		 * address may not be either.  Inherit the b_data offset
865 		 * from the original buffer.
866 		 */
867 		if ((gbflags & GB_UNMAPPED) == 0 ||
868 		    (tbp->b_flags & B_VMIO) == 0) {
869 			bp->b_data = (char *)((vm_offset_t)bp->b_data |
870 			    ((vm_offset_t)tbp->b_data & PAGE_MASK));
871 		} else {
872 			bp->b_data = unmapped_buf;
873 		}
874 		bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
875 		    B_NEEDCOMMIT));
876 		bp->b_iodone = cluster_callback;
877 		pbgetvp(vp, bp);
878 		/*
879 		 * From this location in the file, scan forward to see
880 		 * if there are buffers with adjacent data that need to
881 		 * be written as well.
882 		 */
883 		for (i = 0; i < len; ++i, ++start_lbn) {
884 			if (i != 0) { /* If not the first buffer */
885 				/*
886 				 * If the adjacent data is not even in core it
887 				 * can't need to be written.
888 				 */
889 				BO_LOCK(bo);
890 				if ((tbp = gbincore(bo, start_lbn)) == NULL ||
891 				    (tbp->b_vflags & BV_BKGRDINPROG)) {
892 					BO_UNLOCK(bo);
893 					break;
894 				}
895 
896 				/*
897 				 * If it IS in core, but has different
898 				 * characteristics, or is locked (which
899 				 * means it could be undergoing a background
900 				 * I/O or be in a weird state), then don't
901 				 * cluster with it.
902 				 */
903 				if (BUF_LOCK(tbp,
904 				    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
905 				    BO_LOCKPTR(bo)))
906 					break;
907 
908 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
909 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
910 				    != (B_DELWRI | B_CLUSTEROK |
911 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
912 				    tbp->b_wcred != bp->b_wcred) {
913 					BUF_UNLOCK(tbp);
914 					break;
915 				}
916 
917 				/*
918 				 * Check that the combined cluster
919 				 * would make sense with regard to pages
920 				 * and would not be too large
921 				 */
922 				if ((tbp->b_bcount != size) ||
923 				  ((bp->b_blkno + (dbsize * i)) !=
924 				    tbp->b_blkno) ||
925 				  ((tbp->b_npages + bp->b_npages) >
926 				    (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
927 					BUF_UNLOCK(tbp);
928 					break;
929 				}
930 
931 				/*
932 				 * Do not pull in pinned buffers.
933 				 */
934 				if (tbp->b_pin_count > 0) {
935 					BUF_UNLOCK(tbp);
936 					break;
937 				}
938 
939 				/*
940 				 * Ok, it's passed all the tests,
941 				 * so remove it from the free list
942 				 * and mark it busy. We will use it.
943 				 */
944 				bremfree(tbp);
945 				tbp->b_flags &= ~B_DONE;
946 			} /* end of code for non-first buffers only */
947 			/*
948 			 * If the IO is via the VM then we do some
949 			 * special VM hackery (yuck).  Since the buffer's
950 			 * block size may not be page-aligned it is possible
951 			 * for a page to be shared between two buffers.  We
952 			 * have to get rid of the duplication when building
953 			 * the cluster.
954 			 */
955 			if (tbp->b_flags & B_VMIO) {
956 				vm_page_t m;
957 
958 				VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
959 				if (i == 0) {
960 					vfs_drain_busy_pages(tbp);
961 				} else { /* if not first buffer */
962 					for (j = 0; j < tbp->b_npages; j += 1) {
963 						m = tbp->b_pages[j];
964 						if (vm_page_xbusied(m)) {
965 							VM_OBJECT_WUNLOCK(
966 							    tbp->b_object);
967 							bqrelse(tbp);
968 							goto finishcluster;
969 						}
970 					}
971 				}
972 				for (j = 0; j < tbp->b_npages; j += 1) {
973 					m = tbp->b_pages[j];
974 					vm_page_sbusy(m);
975 					vm_object_pip_add(m->object, 1);
976 					if ((bp->b_npages == 0) ||
977 					  (bp->b_pages[bp->b_npages - 1] != m)) {
978 						bp->b_pages[bp->b_npages] = m;
979 						bp->b_npages++;
980 					}
981 				}
982 				VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
983 			}
984 			bp->b_bcount += size;
985 			bp->b_bufsize += size;
986 			/*
987 			 * If any of the clustered buffers have their
988 			 * B_BARRIER flag set, transfer that request to
989 			 * the cluster.
990 			 */
991 			bp->b_flags |= (tbp->b_flags & B_BARRIER);
992 			tbp->b_flags &= ~(B_DONE | B_BARRIER);
993 			tbp->b_flags |= B_ASYNC;
994 			tbp->b_ioflags &= ~BIO_ERROR;
995 			tbp->b_iocmd = BIO_WRITE;
996 			bundirty(tbp);
997 			reassignbuf(tbp);		/* put on clean list */
998 			bufobj_wref(tbp->b_bufobj);
999 			BUF_KERNPROC(tbp);
1000 			TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
1001 				tbp, b_cluster.cluster_entry);
1002 		}
1003 	finishcluster:
1004 		if (buf_mapped(bp)) {
1005 			pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1006 			    (vm_page_t *)bp->b_pages, bp->b_npages);
1007 		}
1008 		if (bp->b_bufsize > bp->b_kvasize)
1009 			panic(
1010 			    "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
1011 			    bp->b_bufsize, bp->b_kvasize);
1012 		totalwritten += bp->b_bufsize;
1013 		bp->b_dirtyoff = 0;
1014 		bp->b_dirtyend = bp->b_bufsize;
1015 		bawrite(bp);
1016 
1017 		len -= i;
1018 	}
1019 	return totalwritten;
1020 }
1021 
1022 /*
1023  * Collect together all the buffers in a cluster.
1024  * Plus add one additional buffer.
1025  */
1026 static struct cluster_save *
1027 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags)
1028 {
1029 	struct cluster_save *buflist;
1030 	struct buf *bp;
1031 	daddr_t lbn;
1032 	int i, len;
1033 
1034 	len = vp->v_lastw - vp->v_cstart + 1;
1035 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1036 	    M_SEGMENT, M_WAITOK);
1037 	buflist->bs_nchildren = 0;
1038 	buflist->bs_children = (struct buf **) (buflist + 1);
1039 	for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1040 		(void)bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1041 		    gbflags, &bp);
1042 		buflist->bs_children[i] = bp;
1043 		if (bp->b_blkno == bp->b_lblkno)
1044 			VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1045 				NULL, NULL);
1046 	}
1047 	buflist->bs_children[i] = bp = last_bp;
1048 	if (bp->b_blkno == bp->b_lblkno)
1049 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1050 	buflist->bs_nchildren = i + 1;
1051 	return (buflist);
1052 }
1053