xref: /freebsd/sys/kern/vfs_cluster.c (revision d0ba1baed3f6e4936a0c1b89c25f6c59168ef6de)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * Modifications/enhancements:
7  * 	Copyright (c) 1995 John S. Dyson.  All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	@(#)vfs_cluster.c	8.7 (Berkeley) 2/13/94
34  */
35 
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38 
39 #include "opt_debug_cluster.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/proc.h>
45 #include <sys/bio.h>
46 #include <sys/buf.h>
47 #include <sys/vnode.h>
48 #include <sys/malloc.h>
49 #include <sys/mount.h>
50 #include <sys/racct.h>
51 #include <sys/resourcevar.h>
52 #include <sys/rwlock.h>
53 #include <sys/vmmeter.h>
54 #include <vm/vm.h>
55 #include <vm/vm_object.h>
56 #include <vm/vm_page.h>
57 #include <sys/sysctl.h>
58 
59 #if defined(CLUSTERDEBUG)
60 static int	rcluster= 0;
61 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0,
62     "Debug VFS clustering code");
63 #endif
64 
65 static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer");
66 
67 static struct cluster_save *cluster_collectbufs(struct vnode *vp,
68 	    struct buf *last_bp, int gbflags);
69 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize,
70 	    daddr_t lbn, daddr_t blkno, long size, int run, int gbflags,
71 	    struct buf *fbp);
72 static void cluster_callback(struct buf *);
73 
74 static int write_behind = 1;
75 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
76     "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
77 
78 static int read_max = 64;
79 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
80     "Cluster read-ahead max block count");
81 
82 static int read_min = 1;
83 SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0,
84     "Cluster read min block count");
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 	struct thread *td;
98 	daddr_t blkno, origblkno;
99 	int maxra, racluster;
100 	int error, ncontig;
101 	int i;
102 
103 	error = 0;
104 	td = curthread;
105 	bo = &vp->v_bufobj;
106 	if (!unmapped_buf_allowed)
107 		gbflags &= ~GB_UNMAPPED;
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 	error = getblkx(vp, lblkno, size, 0, 0, gbflags, &bp);
124 	if (error != 0) {
125 		*bpp = NULL;
126 		return (error);
127 	}
128 	gbflags &= ~GB_NOSPARSE;
129 	origblkno = lblkno;
130 	*bpp = reqbp = bp;
131 
132 	/*
133 	 * if it is in the cache, then check to see if the reads have been
134 	 * sequential.  If they have, then try some read-ahead, otherwise
135 	 * back-off on prospective read-aheads.
136 	 */
137 	if (bp->b_flags & B_CACHE) {
138 		if (!seqcount) {
139 			return 0;
140 		} else if ((bp->b_flags & B_RAM) == 0) {
141 			return 0;
142 		} else {
143 			bp->b_flags &= ~B_RAM;
144 			BO_RLOCK(bo);
145 			for (i = 1; i < maxra; i++) {
146 				/*
147 				 * Stop if the buffer does not exist or it
148 				 * is invalid (about to go away?)
149 				 */
150 				rbp = gbincore(&vp->v_bufobj, lblkno+i);
151 				if (rbp == NULL || (rbp->b_flags & B_INVAL))
152 					break;
153 
154 				/*
155 				 * Set another read-ahead mark so we know
156 				 * to check again. (If we can lock the
157 				 * buffer without waiting)
158 				 */
159 				if ((((i % racluster) == (racluster - 1)) ||
160 				    (i == (maxra - 1)))
161 				    && (0 == BUF_LOCK(rbp,
162 					LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
163 					rbp->b_flags |= B_RAM;
164 					BUF_UNLOCK(rbp);
165 				}
166 			}
167 			BO_RUNLOCK(bo);
168 			if (i >= maxra) {
169 				return 0;
170 			}
171 			lblkno += i;
172 		}
173 		reqbp = bp = NULL;
174 	/*
175 	 * If it isn't in the cache, then get a chunk from
176 	 * disk if sequential, otherwise just get the block.
177 	 */
178 	} else {
179 		off_t firstread = bp->b_offset;
180 		int nblks;
181 		long minread;
182 
183 		KASSERT(bp->b_offset != NOOFFSET,
184 		    ("cluster_read: no buffer offset"));
185 
186 		ncontig = 0;
187 
188 		/*
189 		 * Adjust totread if needed
190 		 */
191 		minread = read_min * size;
192 		if (minread > totread)
193 			totread = minread;
194 
195 		/*
196 		 * Compute the total number of blocks that we should read
197 		 * synchronously.
198 		 */
199 		if (firstread + totread > filesize)
200 			totread = filesize - firstread;
201 		nblks = howmany(totread, size);
202 		if (nblks > racluster)
203 			nblks = racluster;
204 
205 		/*
206 		 * Now compute the number of contiguous blocks.
207 		 */
208 		if (nblks > 1) {
209 	    		error = VOP_BMAP(vp, lblkno, NULL,
210 				&blkno, &ncontig, NULL);
211 			/*
212 			 * If this failed to map just do the original block.
213 			 */
214 			if (error || blkno == -1)
215 				ncontig = 0;
216 		}
217 
218 		/*
219 		 * If we have contiguous data available do a cluster
220 		 * otherwise just read the requested block.
221 		 */
222 		if (ncontig) {
223 			/* Account for our first block. */
224 			ncontig = min(ncontig + 1, nblks);
225 			if (ncontig < nblks)
226 				nblks = ncontig;
227 			bp = cluster_rbuild(vp, filesize, lblkno,
228 			    blkno, size, nblks, gbflags, bp);
229 			lblkno += (bp->b_bufsize / size);
230 		} else {
231 			bp->b_flags |= B_RAM;
232 			bp->b_iocmd = BIO_READ;
233 			lblkno += 1;
234 		}
235 	}
236 
237 	/*
238 	 * handle the synchronous read so that it is available ASAP.
239 	 */
240 	if (bp) {
241 		if ((bp->b_flags & B_CLUSTER) == 0) {
242 			vfs_busy_pages(bp, 0);
243 		}
244 		bp->b_flags &= ~B_INVAL;
245 		bp->b_ioflags &= ~BIO_ERROR;
246 		if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
247 			BUF_KERNPROC(bp);
248 		bp->b_iooffset = dbtob(bp->b_blkno);
249 		bstrategy(bp);
250 #ifdef RACCT
251 		if (racct_enable) {
252 			PROC_LOCK(td->td_proc);
253 			racct_add_buf(td->td_proc, bp, 0);
254 			PROC_UNLOCK(td->td_proc);
255 		}
256 #endif /* RACCT */
257 		td->td_ru.ru_inblock++;
258 	}
259 
260 	/*
261 	 * If we have been doing sequential I/O, then do some read-ahead.
262 	 */
263 	while (lblkno < (origblkno + maxra)) {
264 		error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
265 		if (error)
266 			break;
267 
268 		if (blkno == -1)
269 			break;
270 
271 		/*
272 		 * We could throttle ncontig here by maxra but we might as
273 		 * well read the data if it is contiguous.  We're throttled
274 		 * by racluster anyway.
275 		 */
276 		if (ncontig) {
277 			ncontig = min(ncontig + 1, racluster);
278 			rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
279 			    size, ncontig, gbflags, NULL);
280 			lblkno += (rbp->b_bufsize / size);
281 			if (rbp->b_flags & B_DELWRI) {
282 				bqrelse(rbp);
283 				continue;
284 			}
285 		} else {
286 			rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
287 			lblkno += 1;
288 			if (rbp->b_flags & B_DELWRI) {
289 				bqrelse(rbp);
290 				continue;
291 			}
292 			rbp->b_flags |= B_ASYNC | B_RAM;
293 			rbp->b_iocmd = BIO_READ;
294 			rbp->b_blkno = blkno;
295 		}
296 		if (rbp->b_flags & B_CACHE) {
297 			rbp->b_flags &= ~B_ASYNC;
298 			bqrelse(rbp);
299 			continue;
300 		}
301 		if ((rbp->b_flags & B_CLUSTER) == 0) {
302 			vfs_busy_pages(rbp, 0);
303 		}
304 		rbp->b_flags &= ~B_INVAL;
305 		rbp->b_ioflags &= ~BIO_ERROR;
306 		if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
307 			BUF_KERNPROC(rbp);
308 		rbp->b_iooffset = dbtob(rbp->b_blkno);
309 		bstrategy(rbp);
310 #ifdef RACCT
311 		if (racct_enable) {
312 			PROC_LOCK(td->td_proc);
313 			racct_add_buf(td->td_proc, rbp, 0);
314 			PROC_UNLOCK(td->td_proc);
315 		}
316 #endif /* RACCT */
317 		td->td_ru.ru_inblock++;
318 	}
319 
320 	if (reqbp) {
321 		/*
322 		 * Like bread, always brelse() the buffer when
323 		 * returning an error.
324 		 */
325 		error = bufwait(reqbp);
326 		if (error != 0) {
327 			brelse(reqbp);
328 			*bpp = NULL;
329 		}
330 	}
331 	return (error);
332 }
333 
334 /*
335  * If blocks are contiguous on disk, use this to provide clustered
336  * read ahead.  We will read as many blocks as possible sequentially
337  * and then parcel them up into logical blocks in the buffer hash table.
338  */
339 static struct buf *
340 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
341     daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
342 {
343 	struct buf *bp, *tbp;
344 	daddr_t bn;
345 	off_t off;
346 	long tinc, tsize;
347 	int i, inc, j, k, toff;
348 
349 	KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
350 	    ("cluster_rbuild: size %ld != f_iosize %jd\n",
351 	    size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
352 
353 	/*
354 	 * avoid a division
355 	 */
356 	while ((u_quad_t) size * (lbn + run) > filesize) {
357 		--run;
358 	}
359 
360 	if (fbp) {
361 		tbp = fbp;
362 		tbp->b_iocmd = BIO_READ;
363 	} else {
364 		tbp = getblk(vp, lbn, size, 0, 0, gbflags);
365 		if (tbp->b_flags & B_CACHE)
366 			return tbp;
367 		tbp->b_flags |= B_ASYNC | B_RAM;
368 		tbp->b_iocmd = BIO_READ;
369 	}
370 	tbp->b_blkno = blkno;
371 	if( (tbp->b_flags & B_MALLOC) ||
372 		((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
373 		return tbp;
374 
375 	bp = trypbuf(&cluster_pbuf_freecnt);
376 	if (bp == NULL)
377 		return tbp;
378 
379 	/*
380 	 * We are synthesizing a buffer out of vm_page_t's, but
381 	 * if the block size is not page aligned then the starting
382 	 * address may not be either.  Inherit the b_data offset
383 	 * from the original buffer.
384 	 */
385 	bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
386 	if ((gbflags & GB_UNMAPPED) != 0) {
387 		bp->b_data = unmapped_buf;
388 	} else {
389 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
390 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
391 	}
392 	bp->b_iocmd = BIO_READ;
393 	bp->b_iodone = cluster_callback;
394 	bp->b_blkno = blkno;
395 	bp->b_lblkno = lbn;
396 	bp->b_offset = tbp->b_offset;
397 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
398 	pbgetvp(vp, bp);
399 
400 	TAILQ_INIT(&bp->b_cluster.cluster_head);
401 
402 	bp->b_bcount = 0;
403 	bp->b_bufsize = 0;
404 	bp->b_npages = 0;
405 
406 	inc = btodb(size);
407 	for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
408 		if (i == 0) {
409 			VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
410 			vfs_drain_busy_pages(tbp);
411 			vm_object_pip_add(tbp->b_bufobj->bo_object,
412 			    tbp->b_npages);
413 			for (k = 0; k < tbp->b_npages; k++)
414 				vm_page_sbusy(tbp->b_pages[k]);
415 			VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
416 		} else {
417 			if ((bp->b_npages * PAGE_SIZE) +
418 			    round_page(size) > vp->v_mount->mnt_iosize_max) {
419 				break;
420 			}
421 
422 			tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
423 			    (gbflags & GB_UNMAPPED));
424 
425 			/* Don't wait around for locked bufs. */
426 			if (tbp == NULL)
427 				break;
428 
429 			/*
430 			 * Stop scanning if the buffer is fully valid
431 			 * (marked B_CACHE), or locked (may be doing a
432 			 * background write), or if the buffer is not
433 			 * VMIO backed.  The clustering code can only deal
434 			 * with VMIO-backed buffers.  The bo lock is not
435 			 * required for the BKGRDINPROG check since it
436 			 * can not be set without the buf lock.
437 			 */
438 			if ((tbp->b_vflags & BV_BKGRDINPROG) ||
439 			    (tbp->b_flags & B_CACHE) ||
440 			    (tbp->b_flags & B_VMIO) == 0) {
441 				bqrelse(tbp);
442 				break;
443 			}
444 
445 			/*
446 			 * The buffer must be completely invalid in order to
447 			 * take part in the cluster.  If it is partially valid
448 			 * then we stop.
449 			 */
450 			off = tbp->b_offset;
451 			tsize = size;
452 			VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
453 			for (j = 0; tsize > 0; j++) {
454 				toff = off & PAGE_MASK;
455 				tinc = tsize;
456 				if (toff + tinc > PAGE_SIZE)
457 					tinc = PAGE_SIZE - toff;
458 				VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object);
459 				if ((tbp->b_pages[j]->valid &
460 				    vm_page_bits(toff, tinc)) != 0)
461 					break;
462 				if (vm_page_xbusied(tbp->b_pages[j]))
463 					break;
464 				vm_object_pip_add(tbp->b_bufobj->bo_object, 1);
465 				vm_page_sbusy(tbp->b_pages[j]);
466 				off += tinc;
467 				tsize -= tinc;
468 			}
469 			if (tsize > 0) {
470 clean_sbusy:
471 				vm_object_pip_add(tbp->b_bufobj->bo_object, -j);
472 				for (k = 0; k < j; k++)
473 					vm_page_sunbusy(tbp->b_pages[k]);
474 				VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
475 				bqrelse(tbp);
476 				break;
477 			}
478 			VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
479 
480 			/*
481 			 * Set a read-ahead mark as appropriate
482 			 */
483 			if ((fbp && (i == 1)) || (i == (run - 1)))
484 				tbp->b_flags |= B_RAM;
485 
486 			/*
487 			 * Set the buffer up for an async read (XXX should
488 			 * we do this only if we do not wind up brelse()ing?).
489 			 * Set the block number if it isn't set, otherwise
490 			 * if it is make sure it matches the block number we
491 			 * expect.
492 			 */
493 			tbp->b_flags |= B_ASYNC;
494 			tbp->b_iocmd = BIO_READ;
495 			if (tbp->b_blkno == tbp->b_lblkno) {
496 				tbp->b_blkno = bn;
497 			} else if (tbp->b_blkno != bn) {
498 				VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
499 				goto clean_sbusy;
500 			}
501 		}
502 		/*
503 		 * XXX fbp from caller may not be B_ASYNC, but we are going
504 		 * to biodone() it in cluster_callback() anyway
505 		 */
506 		BUF_KERNPROC(tbp);
507 		TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
508 			tbp, b_cluster.cluster_entry);
509 		VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
510 		for (j = 0; j < tbp->b_npages; j += 1) {
511 			vm_page_t m;
512 			m = tbp->b_pages[j];
513 			if ((bp->b_npages == 0) ||
514 			    (bp->b_pages[bp->b_npages-1] != m)) {
515 				bp->b_pages[bp->b_npages] = m;
516 				bp->b_npages++;
517 			}
518 			if (m->valid == VM_PAGE_BITS_ALL)
519 				tbp->b_pages[j] = bogus_page;
520 		}
521 		VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
522 		/*
523 		 * Don't inherit tbp->b_bufsize as it may be larger due to
524 		 * a non-page-aligned size.  Instead just aggregate using
525 		 * 'size'.
526 		 */
527 		if (tbp->b_bcount != size)
528 			printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
529 		if (tbp->b_bufsize != size)
530 			printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
531 		bp->b_bcount += size;
532 		bp->b_bufsize += size;
533 	}
534 
535 	/*
536 	 * Fully valid pages in the cluster are already good and do not need
537 	 * to be re-read from disk.  Replace the page with bogus_page
538 	 */
539 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
540 	for (j = 0; j < bp->b_npages; j++) {
541 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object);
542 		if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
543 			bp->b_pages[j] = bogus_page;
544 	}
545 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
546 	if (bp->b_bufsize > bp->b_kvasize)
547 		panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
548 		    bp->b_bufsize, bp->b_kvasize);
549 
550 	if (buf_mapped(bp)) {
551 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
552 		    (vm_page_t *)bp->b_pages, bp->b_npages);
553 	}
554 	return (bp);
555 }
556 
557 /*
558  * Cleanup after a clustered read or write.
559  * This is complicated by the fact that any of the buffers might have
560  * extra memory (if there were no empty buffer headers at allocbuf time)
561  * that we will need to shift around.
562  */
563 static void
564 cluster_callback(bp)
565 	struct buf *bp;
566 {
567 	struct buf *nbp, *tbp;
568 	int error = 0;
569 
570 	/*
571 	 * Must propagate errors to all the components.
572 	 */
573 	if (bp->b_ioflags & BIO_ERROR)
574 		error = bp->b_error;
575 
576 	if (buf_mapped(bp)) {
577 		pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
578 		    bp->b_npages);
579 	}
580 	/*
581 	 * Move memory from the large cluster buffer into the component
582 	 * buffers and mark IO as done on these.
583 	 */
584 	for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
585 		tbp; tbp = nbp) {
586 		nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
587 		if (error) {
588 			tbp->b_ioflags |= BIO_ERROR;
589 			tbp->b_error = error;
590 		} else {
591 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
592 			tbp->b_flags &= ~B_INVAL;
593 			tbp->b_ioflags &= ~BIO_ERROR;
594 			/*
595 			 * XXX the bdwrite()/bqrelse() issued during
596 			 * cluster building clears B_RELBUF (see bqrelse()
597 			 * comment).  If direct I/O was specified, we have
598 			 * to restore it here to allow the buffer and VM
599 			 * to be freed.
600 			 */
601 			if (tbp->b_flags & B_DIRECT)
602 				tbp->b_flags |= B_RELBUF;
603 		}
604 		bufdone(tbp);
605 	}
606 	pbrelvp(bp);
607 	relpbuf(bp, &cluster_pbuf_freecnt);
608 }
609 
610 /*
611  *	cluster_wbuild_wb:
612  *
613  *	Implement modified write build for cluster.
614  *
615  *		write_behind = 0	write behind disabled
616  *		write_behind = 1	write behind normal (default)
617  *		write_behind = 2	write behind backed-off
618  */
619 
620 static __inline int
621 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
622     int gbflags)
623 {
624 	int r = 0;
625 
626 	switch (write_behind) {
627 	case 2:
628 		if (start_lbn < len)
629 			break;
630 		start_lbn -= len;
631 		/* FALLTHROUGH */
632 	case 1:
633 		r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
634 		/* FALLTHROUGH */
635 	default:
636 		/* FALLTHROUGH */
637 		break;
638 	}
639 	return(r);
640 }
641 
642 /*
643  * Do clustered write for FFS.
644  *
645  * Three cases:
646  *	1. Write is not sequential (write asynchronously)
647  *	Write is sequential:
648  *	2.	beginning of cluster - begin cluster
649  *	3.	middle of a cluster - add to cluster
650  *	4.	end of a cluster - asynchronously write cluster
651  */
652 void
653 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount,
654     int gbflags)
655 {
656 	daddr_t lbn;
657 	int maxclen, cursize;
658 	int lblocksize;
659 	int async;
660 
661 	if (!unmapped_buf_allowed)
662 		gbflags &= ~GB_UNMAPPED;
663 
664 	if (vp->v_type == VREG) {
665 		async = DOINGASYNC(vp);
666 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
667 	} else {
668 		async = 0;
669 		lblocksize = bp->b_bufsize;
670 	}
671 	lbn = bp->b_lblkno;
672 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
673 
674 	/* Initialize vnode to beginning of file. */
675 	if (lbn == 0)
676 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
677 
678 	if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
679 	    (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
680 		maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
681 		if (vp->v_clen != 0) {
682 			/*
683 			 * Next block is not sequential.
684 			 *
685 			 * If we are not writing at end of file, the process
686 			 * seeked to another point in the file since its last
687 			 * write, or we have reached our maximum cluster size,
688 			 * then push the previous cluster. Otherwise try
689 			 * reallocating to make it sequential.
690 			 *
691 			 * Change to algorithm: only push previous cluster if
692 			 * it was sequential from the point of view of the
693 			 * seqcount heuristic, otherwise leave the buffer
694 			 * intact so we can potentially optimize the I/O
695 			 * later on in the buf_daemon or update daemon
696 			 * flush.
697 			 */
698 			cursize = vp->v_lastw - vp->v_cstart + 1;
699 			if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
700 			    lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
701 				if (!async && seqcount > 0) {
702 					cluster_wbuild_wb(vp, lblocksize,
703 					    vp->v_cstart, cursize, gbflags);
704 				}
705 			} else {
706 				struct buf **bpp, **endbp;
707 				struct cluster_save *buflist;
708 
709 				buflist = cluster_collectbufs(vp, bp, gbflags);
710 				endbp = &buflist->bs_children
711 				    [buflist->bs_nchildren - 1];
712 				if (VOP_REALLOCBLKS(vp, buflist)) {
713 					/*
714 					 * Failed, push the previous cluster
715 					 * if *really* writing sequentially
716 					 * in the logical file (seqcount > 1),
717 					 * otherwise delay it in the hopes that
718 					 * the low level disk driver can
719 					 * optimize the write ordering.
720 					 */
721 					for (bpp = buflist->bs_children;
722 					     bpp < endbp; bpp++)
723 						brelse(*bpp);
724 					free(buflist, M_SEGMENT);
725 					if (seqcount > 1) {
726 						cluster_wbuild_wb(vp,
727 						    lblocksize, vp->v_cstart,
728 						    cursize, gbflags);
729 					}
730 				} else {
731 					/*
732 					 * Succeeded, keep building cluster.
733 					 */
734 					for (bpp = buflist->bs_children;
735 					     bpp <= endbp; bpp++)
736 						bdwrite(*bpp);
737 					free(buflist, M_SEGMENT);
738 					vp->v_lastw = lbn;
739 					vp->v_lasta = bp->b_blkno;
740 					return;
741 				}
742 			}
743 		}
744 		/*
745 		 * Consider beginning a cluster. If at end of file, make
746 		 * cluster as large as possible, otherwise find size of
747 		 * existing cluster.
748 		 */
749 		if ((vp->v_type == VREG) &&
750 			((u_quad_t) bp->b_offset + lblocksize) != filesize &&
751 		    (bp->b_blkno == bp->b_lblkno) &&
752 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
753 		     bp->b_blkno == -1)) {
754 			bawrite(bp);
755 			vp->v_clen = 0;
756 			vp->v_lasta = bp->b_blkno;
757 			vp->v_cstart = lbn + 1;
758 			vp->v_lastw = lbn;
759 			return;
760 		}
761 		vp->v_clen = maxclen;
762 		if (!async && maxclen == 0) {	/* I/O not contiguous */
763 			vp->v_cstart = lbn + 1;
764 			bawrite(bp);
765 		} else {	/* Wait for rest of cluster */
766 			vp->v_cstart = lbn;
767 			bdwrite(bp);
768 		}
769 	} else if (lbn == vp->v_cstart + vp->v_clen) {
770 		/*
771 		 * At end of cluster, write it out if seqcount tells us we
772 		 * are operating sequentially, otherwise let the buf or
773 		 * update daemon handle it.
774 		 */
775 		bdwrite(bp);
776 		if (seqcount > 1) {
777 			cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
778 			    vp->v_clen + 1, gbflags);
779 		}
780 		vp->v_clen = 0;
781 		vp->v_cstart = lbn + 1;
782 	} else if (vm_page_count_severe()) {
783 		/*
784 		 * We are low on memory, get it going NOW
785 		 */
786 		bawrite(bp);
787 	} else {
788 		/*
789 		 * In the middle of a cluster, so just delay the I/O for now.
790 		 */
791 		bdwrite(bp);
792 	}
793 	vp->v_lastw = lbn;
794 	vp->v_lasta = bp->b_blkno;
795 }
796 
797 
798 /*
799  * This is an awful lot like cluster_rbuild...wish they could be combined.
800  * The last lbn argument is the current block on which I/O is being
801  * performed.  Check to see that it doesn't fall in the middle of
802  * the current block (if last_bp == NULL).
803  */
804 int
805 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
806     int gbflags)
807 {
808 	struct buf *bp, *tbp;
809 	struct bufobj *bo;
810 	int i, j;
811 	int totalwritten = 0;
812 	int dbsize = btodb(size);
813 
814 	if (!unmapped_buf_allowed)
815 		gbflags &= ~GB_UNMAPPED;
816 
817 	bo = &vp->v_bufobj;
818 	while (len > 0) {
819 		/*
820 		 * If the buffer is not delayed-write (i.e. dirty), or it
821 		 * is delayed-write but either locked or inval, it cannot
822 		 * partake in the clustered write.
823 		 */
824 		BO_LOCK(bo);
825 		if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
826 		    (tbp->b_vflags & BV_BKGRDINPROG)) {
827 			BO_UNLOCK(bo);
828 			++start_lbn;
829 			--len;
830 			continue;
831 		}
832 		if (BUF_LOCK(tbp,
833 		    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) {
834 			++start_lbn;
835 			--len;
836 			continue;
837 		}
838 		if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
839 			BUF_UNLOCK(tbp);
840 			++start_lbn;
841 			--len;
842 			continue;
843 		}
844 		bremfree(tbp);
845 		tbp->b_flags &= ~B_DONE;
846 
847 		/*
848 		 * Extra memory in the buffer, punt on this buffer.
849 		 * XXX we could handle this in most cases, but we would
850 		 * have to push the extra memory down to after our max
851 		 * possible cluster size and then potentially pull it back
852 		 * up if the cluster was terminated prematurely--too much
853 		 * hassle.
854 		 */
855 		if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
856 		     (B_CLUSTEROK | B_VMIO)) ||
857 		  (tbp->b_bcount != tbp->b_bufsize) ||
858 		  (tbp->b_bcount != size) ||
859 		  (len == 1) ||
860 		  ((bp = (vp->v_vflag & VV_MD) != 0 ?
861 		  trypbuf(&cluster_pbuf_freecnt) :
862 		  getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
863 			totalwritten += tbp->b_bufsize;
864 			bawrite(tbp);
865 			++start_lbn;
866 			--len;
867 			continue;
868 		}
869 
870 		/*
871 		 * We got a pbuf to make the cluster in.
872 		 * so initialise it.
873 		 */
874 		TAILQ_INIT(&bp->b_cluster.cluster_head);
875 		bp->b_bcount = 0;
876 		bp->b_bufsize = 0;
877 		bp->b_npages = 0;
878 		if (tbp->b_wcred != NOCRED)
879 			bp->b_wcred = crhold(tbp->b_wcred);
880 
881 		bp->b_blkno = tbp->b_blkno;
882 		bp->b_lblkno = tbp->b_lblkno;
883 		bp->b_offset = tbp->b_offset;
884 
885 		/*
886 		 * We are synthesizing a buffer out of vm_page_t's, but
887 		 * if the block size is not page aligned then the starting
888 		 * address may not be either.  Inherit the b_data offset
889 		 * from the original buffer.
890 		 */
891 		if ((gbflags & GB_UNMAPPED) == 0 ||
892 		    (tbp->b_flags & B_VMIO) == 0) {
893 			bp->b_data = (char *)((vm_offset_t)bp->b_data |
894 			    ((vm_offset_t)tbp->b_data & PAGE_MASK));
895 		} else {
896 			bp->b_data = unmapped_buf;
897 		}
898 		bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
899 		    B_NEEDCOMMIT));
900 		bp->b_iodone = cluster_callback;
901 		pbgetvp(vp, bp);
902 		/*
903 		 * From this location in the file, scan forward to see
904 		 * if there are buffers with adjacent data that need to
905 		 * be written as well.
906 		 */
907 		for (i = 0; i < len; ++i, ++start_lbn) {
908 			if (i != 0) { /* If not the first buffer */
909 				/*
910 				 * If the adjacent data is not even in core it
911 				 * can't need to be written.
912 				 */
913 				BO_LOCK(bo);
914 				if ((tbp = gbincore(bo, start_lbn)) == NULL ||
915 				    (tbp->b_vflags & BV_BKGRDINPROG)) {
916 					BO_UNLOCK(bo);
917 					break;
918 				}
919 
920 				/*
921 				 * If it IS in core, but has different
922 				 * characteristics, or is locked (which
923 				 * means it could be undergoing a background
924 				 * I/O or be in a weird state), then don't
925 				 * cluster with it.
926 				 */
927 				if (BUF_LOCK(tbp,
928 				    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
929 				    BO_LOCKPTR(bo)))
930 					break;
931 
932 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
933 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
934 				    != (B_DELWRI | B_CLUSTEROK |
935 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
936 				    tbp->b_wcred != bp->b_wcred) {
937 					BUF_UNLOCK(tbp);
938 					break;
939 				}
940 
941 				/*
942 				 * Check that the combined cluster
943 				 * would make sense with regard to pages
944 				 * and would not be too large
945 				 */
946 				if ((tbp->b_bcount != size) ||
947 				  ((bp->b_blkno + (dbsize * i)) !=
948 				    tbp->b_blkno) ||
949 				  ((tbp->b_npages + bp->b_npages) >
950 				    (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
951 					BUF_UNLOCK(tbp);
952 					break;
953 				}
954 
955 				/*
956 				 * Ok, it's passed all the tests,
957 				 * so remove it from the free list
958 				 * and mark it busy. We will use it.
959 				 */
960 				bremfree(tbp);
961 				tbp->b_flags &= ~B_DONE;
962 			} /* end of code for non-first buffers only */
963 			/*
964 			 * If the IO is via the VM then we do some
965 			 * special VM hackery (yuck).  Since the buffer's
966 			 * block size may not be page-aligned it is possible
967 			 * for a page to be shared between two buffers.  We
968 			 * have to get rid of the duplication when building
969 			 * the cluster.
970 			 */
971 			if (tbp->b_flags & B_VMIO) {
972 				vm_page_t m;
973 
974 				VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
975 				if (i == 0) {
976 					vfs_drain_busy_pages(tbp);
977 				} else { /* if not first buffer */
978 					for (j = 0; j < tbp->b_npages; j += 1) {
979 						m = tbp->b_pages[j];
980 						if (vm_page_xbusied(m)) {
981 							VM_OBJECT_WUNLOCK(
982 							    tbp->b_object);
983 							bqrelse(tbp);
984 							goto finishcluster;
985 						}
986 					}
987 				}
988 				for (j = 0; j < tbp->b_npages; j += 1) {
989 					m = tbp->b_pages[j];
990 					vm_page_sbusy(m);
991 					vm_object_pip_add(m->object, 1);
992 					if ((bp->b_npages == 0) ||
993 					  (bp->b_pages[bp->b_npages - 1] != m)) {
994 						bp->b_pages[bp->b_npages] = m;
995 						bp->b_npages++;
996 					}
997 				}
998 				VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
999 			}
1000 			bp->b_bcount += size;
1001 			bp->b_bufsize += size;
1002 			/*
1003 			 * If any of the clustered buffers have their
1004 			 * B_BARRIER flag set, transfer that request to
1005 			 * the cluster.
1006 			 */
1007 			bp->b_flags |= (tbp->b_flags & B_BARRIER);
1008 			tbp->b_flags &= ~(B_DONE | B_BARRIER);
1009 			tbp->b_flags |= B_ASYNC;
1010 			tbp->b_ioflags &= ~BIO_ERROR;
1011 			tbp->b_iocmd = BIO_WRITE;
1012 			bundirty(tbp);
1013 			reassignbuf(tbp);		/* put on clean list */
1014 			bufobj_wref(tbp->b_bufobj);
1015 			BUF_KERNPROC(tbp);
1016 			buf_track(tbp, __func__);
1017 			TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
1018 				tbp, b_cluster.cluster_entry);
1019 		}
1020 	finishcluster:
1021 		if (buf_mapped(bp)) {
1022 			pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1023 			    (vm_page_t *)bp->b_pages, bp->b_npages);
1024 		}
1025 		if (bp->b_bufsize > bp->b_kvasize)
1026 			panic(
1027 			    "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
1028 			    bp->b_bufsize, bp->b_kvasize);
1029 		totalwritten += bp->b_bufsize;
1030 		bp->b_dirtyoff = 0;
1031 		bp->b_dirtyend = bp->b_bufsize;
1032 		bawrite(bp);
1033 
1034 		len -= i;
1035 	}
1036 	return totalwritten;
1037 }
1038 
1039 /*
1040  * Collect together all the buffers in a cluster.
1041  * Plus add one additional buffer.
1042  */
1043 static struct cluster_save *
1044 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags)
1045 {
1046 	struct cluster_save *buflist;
1047 	struct buf *bp;
1048 	daddr_t lbn;
1049 	int i, len;
1050 
1051 	len = vp->v_lastw - vp->v_cstart + 1;
1052 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1053 	    M_SEGMENT, M_WAITOK);
1054 	buflist->bs_nchildren = 0;
1055 	buflist->bs_children = (struct buf **) (buflist + 1);
1056 	for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1057 		(void)bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1058 		    gbflags, &bp);
1059 		buflist->bs_children[i] = bp;
1060 		if (bp->b_blkno == bp->b_lblkno)
1061 			VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1062 				NULL, NULL);
1063 	}
1064 	buflist->bs_children[i] = bp = last_bp;
1065 	if (bp->b_blkno == bp->b_lblkno)
1066 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1067 	buflist->bs_nchildren = i + 1;
1068 	return (buflist);
1069 }
1070