xref: /freebsd/sys/kern/vfs_cluster.c (revision fcb560670601b2a4d87bb31d7531c8dcc37ee71b)
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, k, 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 			VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
383 			vfs_drain_busy_pages(tbp);
384 			vm_object_pip_add(tbp->b_bufobj->bo_object,
385 			    tbp->b_npages);
386 			for (k = 0; k < tbp->b_npages; k++)
387 				vm_page_sbusy(tbp->b_pages[k]);
388 			VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
389 		} else {
390 			if ((bp->b_npages * PAGE_SIZE) +
391 			    round_page(size) > vp->v_mount->mnt_iosize_max) {
392 				break;
393 			}
394 
395 			tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
396 			    (gbflags & GB_UNMAPPED));
397 
398 			/* Don't wait around for locked bufs. */
399 			if (tbp == NULL)
400 				break;
401 
402 			/*
403 			 * Stop scanning if the buffer is fully valid
404 			 * (marked B_CACHE), or locked (may be doing a
405 			 * background write), or if the buffer is not
406 			 * VMIO backed.  The clustering code can only deal
407 			 * with VMIO-backed buffers.  The bo lock is not
408 			 * required for the BKGRDINPROG check since it
409 			 * can not be set without the buf lock.
410 			 */
411 			if ((tbp->b_vflags & BV_BKGRDINPROG) ||
412 			    (tbp->b_flags & B_CACHE) ||
413 			    (tbp->b_flags & B_VMIO) == 0) {
414 				bqrelse(tbp);
415 				break;
416 			}
417 
418 			/*
419 			 * The buffer must be completely invalid in order to
420 			 * take part in the cluster.  If it is partially valid
421 			 * then we stop.
422 			 */
423 			off = tbp->b_offset;
424 			tsize = size;
425 			VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
426 			for (j = 0; tsize > 0; j++) {
427 				toff = off & PAGE_MASK;
428 				tinc = tsize;
429 				if (toff + tinc > PAGE_SIZE)
430 					tinc = PAGE_SIZE - toff;
431 				VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object);
432 				if ((tbp->b_pages[j]->valid &
433 				    vm_page_bits(toff, tinc)) != 0)
434 					break;
435 				if (vm_page_xbusied(tbp->b_pages[j]))
436 					break;
437 				vm_object_pip_add(tbp->b_bufobj->bo_object, 1);
438 				vm_page_sbusy(tbp->b_pages[j]);
439 				off += tinc;
440 				tsize -= tinc;
441 			}
442 			if (tsize > 0) {
443 clean_sbusy:
444 				vm_object_pip_add(tbp->b_bufobj->bo_object, -j);
445 				for (k = 0; k < j; k++)
446 					vm_page_sunbusy(tbp->b_pages[k]);
447 				VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
448 				bqrelse(tbp);
449 				break;
450 			}
451 			VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
452 
453 			/*
454 			 * Set a read-ahead mark as appropriate
455 			 */
456 			if ((fbp && (i == 1)) || (i == (run - 1)))
457 				tbp->b_flags |= B_RAM;
458 
459 			/*
460 			 * Set the buffer up for an async read (XXX should
461 			 * we do this only if we do not wind up brelse()ing?).
462 			 * Set the block number if it isn't set, otherwise
463 			 * if it is make sure it matches the block number we
464 			 * expect.
465 			 */
466 			tbp->b_flags |= B_ASYNC;
467 			tbp->b_iocmd = BIO_READ;
468 			if (tbp->b_blkno == tbp->b_lblkno) {
469 				tbp->b_blkno = bn;
470 			} else if (tbp->b_blkno != bn) {
471 				VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
472 				goto clean_sbusy;
473 			}
474 		}
475 		/*
476 		 * XXX fbp from caller may not be B_ASYNC, but we are going
477 		 * to biodone() it in cluster_callback() anyway
478 		 */
479 		BUF_KERNPROC(tbp);
480 		TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
481 			tbp, b_cluster.cluster_entry);
482 		VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
483 		for (j = 0; j < tbp->b_npages; j += 1) {
484 			vm_page_t m;
485 			m = tbp->b_pages[j];
486 			if ((bp->b_npages == 0) ||
487 			    (bp->b_pages[bp->b_npages-1] != m)) {
488 				bp->b_pages[bp->b_npages] = m;
489 				bp->b_npages++;
490 			}
491 			if (m->valid == VM_PAGE_BITS_ALL)
492 				tbp->b_pages[j] = bogus_page;
493 		}
494 		VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
495 		/*
496 		 * Don't inherit tbp->b_bufsize as it may be larger due to
497 		 * a non-page-aligned size.  Instead just aggregate using
498 		 * 'size'.
499 		 */
500 		if (tbp->b_bcount != size)
501 			printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
502 		if (tbp->b_bufsize != size)
503 			printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
504 		bp->b_bcount += size;
505 		bp->b_bufsize += size;
506 	}
507 
508 	/*
509 	 * Fully valid pages in the cluster are already good and do not need
510 	 * to be re-read from disk.  Replace the page with bogus_page
511 	 */
512 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
513 	for (j = 0; j < bp->b_npages; j++) {
514 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object);
515 		if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
516 			bp->b_pages[j] = bogus_page;
517 	}
518 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
519 	if (bp->b_bufsize > bp->b_kvasize)
520 		panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
521 		    bp->b_bufsize, bp->b_kvasize);
522 	bp->b_kvasize = bp->b_bufsize;
523 
524 	if ((bp->b_flags & B_UNMAPPED) == 0) {
525 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
526 		    (vm_page_t *)bp->b_pages, bp->b_npages);
527 	}
528 	return (bp);
529 }
530 
531 /*
532  * Cleanup after a clustered read or write.
533  * This is complicated by the fact that any of the buffers might have
534  * extra memory (if there were no empty buffer headers at allocbuf time)
535  * that we will need to shift around.
536  */
537 static void
538 cluster_callback(bp)
539 	struct buf *bp;
540 {
541 	struct buf *nbp, *tbp;
542 	int error = 0;
543 
544 	/*
545 	 * Must propogate errors to all the components.
546 	 */
547 	if (bp->b_ioflags & BIO_ERROR)
548 		error = bp->b_error;
549 
550 	if ((bp->b_flags & B_UNMAPPED) == 0) {
551 		pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
552 		    bp->b_npages);
553 	}
554 	/*
555 	 * Move memory from the large cluster buffer into the component
556 	 * buffers and mark IO as done on these.
557 	 */
558 	for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
559 		tbp; tbp = nbp) {
560 		nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
561 		if (error) {
562 			tbp->b_ioflags |= BIO_ERROR;
563 			tbp->b_error = error;
564 		} else {
565 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
566 			tbp->b_flags &= ~B_INVAL;
567 			tbp->b_ioflags &= ~BIO_ERROR;
568 			/*
569 			 * XXX the bdwrite()/bqrelse() issued during
570 			 * cluster building clears B_RELBUF (see bqrelse()
571 			 * comment).  If direct I/O was specified, we have
572 			 * to restore it here to allow the buffer and VM
573 			 * to be freed.
574 			 */
575 			if (tbp->b_flags & B_DIRECT)
576 				tbp->b_flags |= B_RELBUF;
577 		}
578 		bufdone(tbp);
579 	}
580 	pbrelvp(bp);
581 	relpbuf(bp, &cluster_pbuf_freecnt);
582 }
583 
584 /*
585  *	cluster_wbuild_wb:
586  *
587  *	Implement modified write build for cluster.
588  *
589  *		write_behind = 0	write behind disabled
590  *		write_behind = 1	write behind normal (default)
591  *		write_behind = 2	write behind backed-off
592  */
593 
594 static __inline int
595 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
596     int gbflags)
597 {
598 	int r = 0;
599 
600 	switch (write_behind) {
601 	case 2:
602 		if (start_lbn < len)
603 			break;
604 		start_lbn -= len;
605 		/* FALLTHROUGH */
606 	case 1:
607 		r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
608 		/* FALLTHROUGH */
609 	default:
610 		/* FALLTHROUGH */
611 		break;
612 	}
613 	return(r);
614 }
615 
616 /*
617  * Do clustered write for FFS.
618  *
619  * Three cases:
620  *	1. Write is not sequential (write asynchronously)
621  *	Write is sequential:
622  *	2.	beginning of cluster - begin cluster
623  *	3.	middle of a cluster - add to cluster
624  *	4.	end of a cluster - asynchronously write cluster
625  */
626 void
627 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount,
628     int gbflags)
629 {
630 	daddr_t lbn;
631 	int maxclen, cursize;
632 	int lblocksize;
633 	int async;
634 
635 	if (!unmapped_buf_allowed)
636 		gbflags &= ~GB_UNMAPPED;
637 
638 	if (vp->v_type == VREG) {
639 		async = DOINGASYNC(vp);
640 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
641 	} else {
642 		async = 0;
643 		lblocksize = bp->b_bufsize;
644 	}
645 	lbn = bp->b_lblkno;
646 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
647 
648 	/* Initialize vnode to beginning of file. */
649 	if (lbn == 0)
650 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
651 
652 	if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
653 	    (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
654 		maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
655 		if (vp->v_clen != 0) {
656 			/*
657 			 * Next block is not sequential.
658 			 *
659 			 * If we are not writing at end of file, the process
660 			 * seeked to another point in the file since its last
661 			 * write, or we have reached our maximum cluster size,
662 			 * then push the previous cluster. Otherwise try
663 			 * reallocating to make it sequential.
664 			 *
665 			 * Change to algorithm: only push previous cluster if
666 			 * it was sequential from the point of view of the
667 			 * seqcount heuristic, otherwise leave the buffer
668 			 * intact so we can potentially optimize the I/O
669 			 * later on in the buf_daemon or update daemon
670 			 * flush.
671 			 */
672 			cursize = vp->v_lastw - vp->v_cstart + 1;
673 			if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
674 			    lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
675 				if (!async && seqcount > 0) {
676 					cluster_wbuild_wb(vp, lblocksize,
677 					    vp->v_cstart, cursize, gbflags);
678 				}
679 			} else {
680 				struct buf **bpp, **endbp;
681 				struct cluster_save *buflist;
682 
683 				buflist = cluster_collectbufs(vp, bp, gbflags);
684 				endbp = &buflist->bs_children
685 				    [buflist->bs_nchildren - 1];
686 				if (VOP_REALLOCBLKS(vp, buflist)) {
687 					/*
688 					 * Failed, push the previous cluster
689 					 * if *really* writing sequentially
690 					 * in the logical file (seqcount > 1),
691 					 * otherwise delay it in the hopes that
692 					 * the low level disk driver can
693 					 * optimize the write ordering.
694 					 */
695 					for (bpp = buflist->bs_children;
696 					     bpp < endbp; bpp++)
697 						brelse(*bpp);
698 					free(buflist, M_SEGMENT);
699 					if (seqcount > 1) {
700 						cluster_wbuild_wb(vp,
701 						    lblocksize, vp->v_cstart,
702 						    cursize, gbflags);
703 					}
704 				} else {
705 					/*
706 					 * Succeeded, keep building cluster.
707 					 */
708 					for (bpp = buflist->bs_children;
709 					     bpp <= endbp; bpp++)
710 						bdwrite(*bpp);
711 					free(buflist, M_SEGMENT);
712 					vp->v_lastw = lbn;
713 					vp->v_lasta = bp->b_blkno;
714 					return;
715 				}
716 			}
717 		}
718 		/*
719 		 * Consider beginning a cluster. If at end of file, make
720 		 * cluster as large as possible, otherwise find size of
721 		 * existing cluster.
722 		 */
723 		if ((vp->v_type == VREG) &&
724 			((u_quad_t) bp->b_offset + lblocksize) != filesize &&
725 		    (bp->b_blkno == bp->b_lblkno) &&
726 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
727 		     bp->b_blkno == -1)) {
728 			bawrite(bp);
729 			vp->v_clen = 0;
730 			vp->v_lasta = bp->b_blkno;
731 			vp->v_cstart = lbn + 1;
732 			vp->v_lastw = lbn;
733 			return;
734 		}
735 		vp->v_clen = maxclen;
736 		if (!async && maxclen == 0) {	/* I/O not contiguous */
737 			vp->v_cstart = lbn + 1;
738 			bawrite(bp);
739 		} else {	/* Wait for rest of cluster */
740 			vp->v_cstart = lbn;
741 			bdwrite(bp);
742 		}
743 	} else if (lbn == vp->v_cstart + vp->v_clen) {
744 		/*
745 		 * At end of cluster, write it out if seqcount tells us we
746 		 * are operating sequentially, otherwise let the buf or
747 		 * update daemon handle it.
748 		 */
749 		bdwrite(bp);
750 		if (seqcount > 1) {
751 			cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
752 			    vp->v_clen + 1, gbflags);
753 		}
754 		vp->v_clen = 0;
755 		vp->v_cstart = lbn + 1;
756 	} else if (vm_page_count_severe()) {
757 		/*
758 		 * We are low on memory, get it going NOW
759 		 */
760 		bawrite(bp);
761 	} else {
762 		/*
763 		 * In the middle of a cluster, so just delay the I/O for now.
764 		 */
765 		bdwrite(bp);
766 	}
767 	vp->v_lastw = lbn;
768 	vp->v_lasta = bp->b_blkno;
769 }
770 
771 
772 /*
773  * This is an awful lot like cluster_rbuild...wish they could be combined.
774  * The last lbn argument is the current block on which I/O is being
775  * performed.  Check to see that it doesn't fall in the middle of
776  * the current block (if last_bp == NULL).
777  */
778 int
779 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
780     int gbflags)
781 {
782 	struct buf *bp, *tbp;
783 	struct bufobj *bo;
784 	int i, j;
785 	int totalwritten = 0;
786 	int dbsize = btodb(size);
787 
788 	if (!unmapped_buf_allowed)
789 		gbflags &= ~GB_UNMAPPED;
790 
791 	bo = &vp->v_bufobj;
792 	while (len > 0) {
793 		/*
794 		 * If the buffer is not delayed-write (i.e. dirty), or it
795 		 * is delayed-write but either locked or inval, it cannot
796 		 * partake in the clustered write.
797 		 */
798 		BO_LOCK(bo);
799 		if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
800 		    (tbp->b_vflags & BV_BKGRDINPROG)) {
801 			BO_UNLOCK(bo);
802 			++start_lbn;
803 			--len;
804 			continue;
805 		}
806 		if (BUF_LOCK(tbp,
807 		    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) {
808 			++start_lbn;
809 			--len;
810 			continue;
811 		}
812 		if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
813 			BUF_UNLOCK(tbp);
814 			++start_lbn;
815 			--len;
816 			continue;
817 		}
818 		if (tbp->b_pin_count >  0) {
819 			BUF_UNLOCK(tbp);
820 			++start_lbn;
821 			--len;
822 			continue;
823 		}
824 		bremfree(tbp);
825 		tbp->b_flags &= ~B_DONE;
826 
827 		/*
828 		 * Extra memory in the buffer, punt on this buffer.
829 		 * XXX we could handle this in most cases, but we would
830 		 * have to push the extra memory down to after our max
831 		 * possible cluster size and then potentially pull it back
832 		 * up if the cluster was terminated prematurely--too much
833 		 * hassle.
834 		 */
835 		if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
836 		     (B_CLUSTEROK | B_VMIO)) ||
837 		  (tbp->b_bcount != tbp->b_bufsize) ||
838 		  (tbp->b_bcount != size) ||
839 		  (len == 1) ||
840 		  ((bp = (vp->v_vflag & VV_MD) != 0 ?
841 		  trypbuf(&cluster_pbuf_freecnt) :
842 		  getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
843 			totalwritten += tbp->b_bufsize;
844 			bawrite(tbp);
845 			++start_lbn;
846 			--len;
847 			continue;
848 		}
849 
850 		/*
851 		 * We got a pbuf to make the cluster in.
852 		 * so initialise it.
853 		 */
854 		TAILQ_INIT(&bp->b_cluster.cluster_head);
855 		bp->b_bcount = 0;
856 		bp->b_bufsize = 0;
857 		bp->b_npages = 0;
858 		if (tbp->b_wcred != NOCRED)
859 			bp->b_wcred = crhold(tbp->b_wcred);
860 
861 		bp->b_blkno = tbp->b_blkno;
862 		bp->b_lblkno = tbp->b_lblkno;
863 		bp->b_offset = tbp->b_offset;
864 
865 		/*
866 		 * We are synthesizing a buffer out of vm_page_t's, but
867 		 * if the block size is not page aligned then the starting
868 		 * address may not be either.  Inherit the b_data offset
869 		 * from the original buffer.
870 		 */
871 		if ((gbflags & GB_UNMAPPED) == 0 ||
872 		    (tbp->b_flags & B_VMIO) == 0) {
873 			bp->b_data = (char *)((vm_offset_t)bp->b_data |
874 			    ((vm_offset_t)tbp->b_data & PAGE_MASK));
875 		} else {
876 			bp->b_flags |= B_UNMAPPED;
877 			bp->b_data = unmapped_buf;
878 		}
879 		bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
880 		    B_NEEDCOMMIT));
881 		bp->b_iodone = cluster_callback;
882 		pbgetvp(vp, bp);
883 		/*
884 		 * From this location in the file, scan forward to see
885 		 * if there are buffers with adjacent data that need to
886 		 * be written as well.
887 		 */
888 		for (i = 0; i < len; ++i, ++start_lbn) {
889 			if (i != 0) { /* If not the first buffer */
890 				/*
891 				 * If the adjacent data is not even in core it
892 				 * can't need to be written.
893 				 */
894 				BO_LOCK(bo);
895 				if ((tbp = gbincore(bo, start_lbn)) == NULL ||
896 				    (tbp->b_vflags & BV_BKGRDINPROG)) {
897 					BO_UNLOCK(bo);
898 					break;
899 				}
900 
901 				/*
902 				 * If it IS in core, but has different
903 				 * characteristics, or is locked (which
904 				 * means it could be undergoing a background
905 				 * I/O or be in a weird state), then don't
906 				 * cluster with it.
907 				 */
908 				if (BUF_LOCK(tbp,
909 				    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
910 				    BO_LOCKPTR(bo)))
911 					break;
912 
913 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
914 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
915 				    != (B_DELWRI | B_CLUSTEROK |
916 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
917 				    tbp->b_wcred != bp->b_wcred) {
918 					BUF_UNLOCK(tbp);
919 					break;
920 				}
921 
922 				/*
923 				 * Check that the combined cluster
924 				 * would make sense with regard to pages
925 				 * and would not be too large
926 				 */
927 				if ((tbp->b_bcount != size) ||
928 				  ((bp->b_blkno + (dbsize * i)) !=
929 				    tbp->b_blkno) ||
930 				  ((tbp->b_npages + bp->b_npages) >
931 				    (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
932 					BUF_UNLOCK(tbp);
933 					break;
934 				}
935 
936 				/*
937 				 * Do not pull in pinned buffers.
938 				 */
939 				if (tbp->b_pin_count > 0) {
940 					BUF_UNLOCK(tbp);
941 					break;
942 				}
943 
944 				/*
945 				 * Ok, it's passed all the tests,
946 				 * so remove it from the free list
947 				 * and mark it busy. We will use it.
948 				 */
949 				bremfree(tbp);
950 				tbp->b_flags &= ~B_DONE;
951 			} /* end of code for non-first buffers only */
952 			/*
953 			 * If the IO is via the VM then we do some
954 			 * special VM hackery (yuck).  Since the buffer's
955 			 * block size may not be page-aligned it is possible
956 			 * for a page to be shared between two buffers.  We
957 			 * have to get rid of the duplication when building
958 			 * the cluster.
959 			 */
960 			if (tbp->b_flags & B_VMIO) {
961 				vm_page_t m;
962 
963 				VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
964 				if (i == 0) {
965 					vfs_drain_busy_pages(tbp);
966 				} else { /* if not first buffer */
967 					for (j = 0; j < tbp->b_npages; j += 1) {
968 						m = tbp->b_pages[j];
969 						if (vm_page_xbusied(m)) {
970 							VM_OBJECT_WUNLOCK(
971 							    tbp->b_object);
972 							bqrelse(tbp);
973 							goto finishcluster;
974 						}
975 					}
976 				}
977 				for (j = 0; j < tbp->b_npages; j += 1) {
978 					m = tbp->b_pages[j];
979 					vm_page_sbusy(m);
980 					vm_object_pip_add(m->object, 1);
981 					if ((bp->b_npages == 0) ||
982 					  (bp->b_pages[bp->b_npages - 1] != m)) {
983 						bp->b_pages[bp->b_npages] = m;
984 						bp->b_npages++;
985 					}
986 				}
987 				VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
988 			}
989 			bp->b_bcount += size;
990 			bp->b_bufsize += size;
991 			/*
992 			 * If any of the clustered buffers have their
993 			 * B_BARRIER flag set, transfer that request to
994 			 * the cluster.
995 			 */
996 			bp->b_flags |= (tbp->b_flags & B_BARRIER);
997 			tbp->b_flags &= ~(B_DONE | B_BARRIER);
998 			tbp->b_flags |= B_ASYNC;
999 			tbp->b_ioflags &= ~BIO_ERROR;
1000 			tbp->b_iocmd = BIO_WRITE;
1001 			bundirty(tbp);
1002 			reassignbuf(tbp);		/* put on clean list */
1003 			bufobj_wref(tbp->b_bufobj);
1004 			BUF_KERNPROC(tbp);
1005 			TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
1006 				tbp, b_cluster.cluster_entry);
1007 		}
1008 	finishcluster:
1009 		if ((bp->b_flags & B_UNMAPPED) == 0) {
1010 			pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1011 			    (vm_page_t *)bp->b_pages, bp->b_npages);
1012 		}
1013 		if (bp->b_bufsize > bp->b_kvasize)
1014 			panic(
1015 			    "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
1016 			    bp->b_bufsize, bp->b_kvasize);
1017 		bp->b_kvasize = bp->b_bufsize;
1018 		totalwritten += bp->b_bufsize;
1019 		bp->b_dirtyoff = 0;
1020 		bp->b_dirtyend = bp->b_bufsize;
1021 		bawrite(bp);
1022 
1023 		len -= i;
1024 	}
1025 	return totalwritten;
1026 }
1027 
1028 /*
1029  * Collect together all the buffers in a cluster.
1030  * Plus add one additional buffer.
1031  */
1032 static struct cluster_save *
1033 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags)
1034 {
1035 	struct cluster_save *buflist;
1036 	struct buf *bp;
1037 	daddr_t lbn;
1038 	int i, len;
1039 
1040 	len = vp->v_lastw - vp->v_cstart + 1;
1041 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1042 	    M_SEGMENT, M_WAITOK);
1043 	buflist->bs_nchildren = 0;
1044 	buflist->bs_children = (struct buf **) (buflist + 1);
1045 	for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1046 		(void)bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1047 		    gbflags, &bp);
1048 		buflist->bs_children[i] = bp;
1049 		if (bp->b_blkno == bp->b_lblkno)
1050 			VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1051 				NULL, NULL);
1052 	}
1053 	buflist->bs_children[i] = bp = last_bp;
1054 	if (bp->b_blkno == bp->b_lblkno)
1055 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1056 	buflist->bs_nchildren = i + 1;
1057 	return (buflist);
1058 }
1059