xref: /dflybsd-src/sys/kern/vfs_cluster.c (revision 1c4f2fa48567f20b32ef2963ca51c6d456c4b58b)
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  * 3. All advertising materials mentioning features or use of this software
16  *    must display the following acknowledgement:
17  *	This product includes software developed by the University of
18  *	California, Berkeley and its contributors.
19  * 4. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	@(#)vfs_cluster.c	8.7 (Berkeley) 2/13/94
36  * $FreeBSD: src/sys/kern/vfs_cluster.c,v 1.92.2.9 2001/11/18 07:10:59 dillon Exp $
37  * $DragonFly: src/sys/kern/vfs_cluster.c,v 1.40 2008/07/14 03:09:00 dillon Exp $
38  */
39 
40 #include "opt_debug_cluster.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/proc.h>
46 #include <sys/buf.h>
47 #include <sys/vnode.h>
48 #include <sys/malloc.h>
49 #include <sys/mount.h>
50 #include <sys/resourcevar.h>
51 #include <sys/vmmeter.h>
52 #include <vm/vm.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_page.h>
55 #include <sys/sysctl.h>
56 #include <sys/buf2.h>
57 #include <vm/vm_page2.h>
58 
59 #include <machine/limits.h>
60 
61 #if defined(CLUSTERDEBUG)
62 #include <sys/sysctl.h>
63 static int	rcluster= 0;
64 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
65 #endif
66 
67 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");
68 
69 static struct cluster_save *
70 	cluster_collectbufs (struct vnode *vp, struct buf *last_bp,
71 			    int blksize);
72 static struct buf *
73 	cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
74 			    off_t doffset, int blksize, int run,
75 			    struct buf *fbp);
76 static void cluster_callback (struct bio *);
77 static void cluster_setram (struct buf *);
78 
79 static int write_behind = 1;
80 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
81 
82 extern vm_page_t	bogus_page;
83 
84 extern int cluster_pbuf_freecnt;
85 
86 /*
87  * Maximum number of blocks for read-ahead.
88  */
89 #define MAXRA 32
90 
91 /*
92  * This replaces bread.
93  */
94 int
95 cluster_read(struct vnode *vp, off_t filesize, off_t loffset,
96 	     int blksize, size_t resid, int seqcount, struct buf **bpp)
97 {
98 	struct buf *bp, *rbp, *reqbp;
99 	off_t origoffset;
100 	off_t doffset;
101 	int error;
102 	int i;
103 	int maxra, racluster;
104 	int totread;
105 
106 	error = 0;
107 	totread = (resid > INT_MAX) ? INT_MAX : (int)resid;
108 
109 	/*
110 	 * racluster - calculate maximum cluster IO size (limited by
111 	 *	       backing block device).
112 	 *
113 	 * Try to limit the amount of read-ahead by a few ad-hoc parameters.
114 	 * This needs work!!!
115 	 *
116 	 * NOTE!  The BMAP operations may involve synchronous I/O so we
117 	 *	  really want several cluster IOs in progress to absorb
118 	 *	  the time lag.
119 	 */
120 	racluster = vmaxiosize(vp) / blksize;
121 	maxra = 2 * racluster + (totread / blksize);
122 	if (maxra > MAXRA)
123 		maxra = MAXRA;
124 	if (maxra > nbuf / 8)
125 		maxra = nbuf / 8;
126 
127 	/*
128 	 * Get the requested block.
129 	 */
130 	*bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0);
131 	origoffset = loffset;
132 
133 	/*
134 	 * if it is in the cache, then check to see if the reads have been
135 	 * sequential.  If they have, then try some read-ahead, otherwise
136 	 * back-off on prospective read-aheads.
137 	 */
138 	if (bp->b_flags & B_CACHE) {
139 		/*
140 		 * Not sequential, do not do any read-ahead
141 		 */
142 		if (seqcount == 0 || maxra == 0)
143 			return 0;
144 
145 		/*
146 		 * No read-ahead mark, do not do any read-ahead
147 		 * yet.
148 		 */
149 		if ((bp->b_flags & B_RAM) == 0)
150 			return 0;
151 
152 		/*
153 		 * We hit a read-ahead-mark, figure out how much read-ahead
154 		 * to do (maxra) and where to start (loffset).
155 		 *
156 		 * Shortcut the scan.  Typically the way this works is that
157 		 * we've built up all the blocks inbetween except for the
158 		 * last in previous iterations, so if the second-to-last
159 		 * block is present we just skip ahead to it.
160 		 *
161 		 * This algorithm has O(1) cpu in the steady state no
162 		 * matter how large maxra is.
163 		 */
164 		bp->b_flags &= ~B_RAM;
165 
166 		if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST))
167 			i = maxra - 1;
168 		else
169 			i = 1;
170 		while (i < maxra) {
171 			if (findblk(vp, loffset + i * blksize,
172 				    FINDBLK_TEST) == NULL) {
173 				break;
174 			}
175 			++i;
176 		}
177 		if (i >= maxra)
178 			return 0;
179 		maxra -= i;
180 		loffset += i * blksize;
181 		reqbp = bp = NULL;
182 	} else {
183 		off_t firstread = bp->b_loffset;
184 		int nblks;
185 
186 		/*
187 		 * Set-up synchronous read for bp.
188 		 */
189 		bp->b_cmd = BUF_CMD_READ;
190 		bp->b_bio1.bio_done = biodone_sync;
191 		bp->b_bio1.bio_flags |= BIO_SYNC;
192 
193 		KASSERT(firstread != NOOFFSET,
194 			("cluster_read: no buffer offset"));
195 		if (firstread + totread > filesize)
196 			totread = (int)(filesize - firstread);
197 		nblks = totread / blksize;
198 		if (nblks) {
199 			int burstbytes;
200 
201 			if (nblks > racluster)
202 				nblks = racluster;
203 
204 	    		error = VOP_BMAP(vp, loffset, &doffset,
205 					 &burstbytes, NULL, BUF_CMD_READ);
206 			if (error)
207 				goto single_block_read;
208 			if (doffset == NOOFFSET)
209 				goto single_block_read;
210 			if (burstbytes < blksize * 2)
211 				goto single_block_read;
212 			if (nblks > burstbytes / blksize)
213 				nblks = burstbytes / blksize;
214 
215 			bp = cluster_rbuild(vp, filesize, loffset,
216 					    doffset, blksize, nblks, bp);
217 			loffset += bp->b_bufsize;
218 			maxra -= (bp->b_bufsize - blksize) / blksize;
219 		} else {
220 single_block_read:
221 			/*
222 			 * if it isn't in the cache, then get a chunk from
223 			 * disk if sequential, otherwise just get the block.
224 			 */
225 			cluster_setram(bp);
226 			loffset += blksize;
227 		}
228 	}
229 
230 	/*
231 	 * If B_CACHE was not set issue bp.  bp will either be an
232 	 * asynchronous cluster buf or a synchronous single-buf.
233 	 * If it is a single buf it will be the same as reqbp.
234 	 *
235 	 * NOTE: Once an async cluster buf is issued bp becomes invalid.
236 	 */
237 	if (bp) {
238 #if defined(CLUSTERDEBUG)
239 		if (rcluster)
240 			kprintf("S(%lld,%d,%d) ",
241 			    bp->b_loffset, bp->b_bcount, seqcount);
242 #endif
243 		if ((bp->b_flags & B_CLUSTER) == 0)
244 			vfs_busy_pages(vp, bp);
245 		bp->b_flags &= ~(B_ERROR|B_INVAL);
246 		vn_strategy(vp, &bp->b_bio1);
247 		error = 0;
248 		/* bp invalid now */
249 	}
250 
251 	/*
252 	 * If we have been doing sequential I/O, then do some read-ahead.
253 	 * The code above us should have positioned us at the next likely
254 	 * offset.
255 	 *
256 	 * Only mess with buffers which we can immediately lock.  HAMMER
257 	 * will do device-readahead irrespective of what the blocks
258 	 * represent.
259 	 */
260 	while (!error && seqcount && maxra > 0 &&
261 	       loffset + blksize <= filesize) {
262 		int nblksread;
263 		int ntoread;
264 		int burstbytes;
265 		int tmp_error;
266 
267 		rbp = getblk(vp, loffset, blksize,
268 			     GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
269 		if (rbp == NULL)
270 			goto no_read_ahead;
271 		if ((rbp->b_flags & B_CACHE)) {
272 			bqrelse(rbp);
273 			goto no_read_ahead;
274 		}
275 
276 		/*
277 		 * An error from the read-ahead bmap has nothing to do
278 		 * with the caller's original request.
279 		 */
280 		tmp_error = VOP_BMAP(vp, loffset, &doffset,
281 				     &burstbytes, NULL, BUF_CMD_READ);
282 		if (tmp_error || doffset == NOOFFSET) {
283 			rbp->b_flags |= B_INVAL;
284 			brelse(rbp);
285 			rbp = NULL;
286 			goto no_read_ahead;
287 		}
288 		ntoread = burstbytes / blksize;
289 		nblksread = (totread + blksize - 1) / blksize;
290 		if (seqcount < nblksread)
291 			seqcount = nblksread;
292 		if (ntoread > seqcount)
293 			ntoread = seqcount;
294 
295 		/*
296 		 * rbp: async read
297 		 */
298 		rbp->b_cmd = BUF_CMD_READ;
299 		/*rbp->b_flags |= B_AGE*/;
300 		cluster_setram(rbp);
301 
302 		if (burstbytes) {
303 			rbp = cluster_rbuild(vp, filesize, loffset,
304 					     doffset, blksize,
305 					     ntoread, rbp);
306 		} else {
307 			rbp->b_bio2.bio_offset = doffset;
308 		}
309 #if defined(CLUSTERDEBUG)
310 		if (rcluster) {
311 			if (bp)
312 				kprintf("A+(%lld,%d,%lld,%d) ",
313 				    rbp->b_loffset, rbp->b_bcount,
314 				    rbp->b_loffset - origoffset,
315 				    seqcount);
316 			else
317 				kprintf("A(%lld,%d,%lld,%d) ",
318 				    rbp->b_loffset, rbp->b_bcount,
319 				    rbp->b_loffset - origoffset,
320 				    seqcount);
321 		}
322 #endif
323 		rbp->b_flags &= ~(B_ERROR|B_INVAL);
324 
325 		if ((rbp->b_flags & B_CLUSTER) == 0)
326 			vfs_busy_pages(vp, rbp);
327 		BUF_KERNPROC(rbp);
328 		loffset += rbp->b_bufsize;
329 		maxra -= rbp->b_bufsize / blksize;
330 		vn_strategy(vp, &rbp->b_bio1);
331 		/* rbp invalid now */
332 	}
333 
334 	/*
335 	 * Wait for our original buffer to complete its I/O.  reqbp will
336 	 * be NULL if the original buffer was B_CACHE.  We are returning
337 	 * (*bpp) which is the same as reqbp when reqbp != NULL.
338 	 */
339 no_read_ahead:
340 	if (reqbp) {
341 		KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
342 		error = biowait(&reqbp->b_bio1, "clurd");
343 	}
344 	return (error);
345 }
346 
347 /*
348  * If blocks are contiguous on disk, use this to provide clustered
349  * read ahead.  We will read as many blocks as possible sequentially
350  * and then parcel them up into logical blocks in the buffer hash table.
351  *
352  * This function either returns a cluster buf or it returns fbp.  fbp is
353  * already expected to be set up as a synchronous or asynchronous request.
354  *
355  * If a cluster buf is returned it will always be async.
356  */
357 static struct buf *
358 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
359 	       int blksize, int run, struct buf *fbp)
360 {
361 	struct buf *bp, *tbp;
362 	off_t boffset;
363 	int i, j;
364 	int maxiosize = vmaxiosize(vp);
365 
366 	/*
367 	 * avoid a division
368 	 */
369 	while (loffset + run * blksize > filesize) {
370 		--run;
371 	}
372 
373 	tbp = fbp;
374 	tbp->b_bio2.bio_offset = doffset;
375 	if((tbp->b_flags & B_MALLOC) ||
376 	    ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
377 		return tbp;
378 	}
379 
380 	bp = trypbuf(&cluster_pbuf_freecnt);
381 	if (bp == NULL) {
382 		return tbp;
383 	}
384 
385 	/*
386 	 * We are synthesizing a buffer out of vm_page_t's, but
387 	 * if the block size is not page aligned then the starting
388 	 * address may not be either.  Inherit the b_data offset
389 	 * from the original buffer.
390 	 */
391 	bp->b_data = (char *)((vm_offset_t)bp->b_data |
392 	    ((vm_offset_t)tbp->b_data & PAGE_MASK));
393 	bp->b_flags |= B_CLUSTER | B_VMIO;
394 	bp->b_cmd = BUF_CMD_READ;
395 	bp->b_bio1.bio_done = cluster_callback;		/* default to async */
396 	bp->b_bio1.bio_caller_info1.cluster_head = NULL;
397 	bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
398 	bp->b_loffset = loffset;
399 	bp->b_bio2.bio_offset = doffset;
400 	KASSERT(bp->b_loffset != NOOFFSET,
401 		("cluster_rbuild: no buffer offset"));
402 
403 	bp->b_bcount = 0;
404 	bp->b_bufsize = 0;
405 	bp->b_xio.xio_npages = 0;
406 
407 	for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
408 		if (i) {
409 			if ((bp->b_xio.xio_npages * PAGE_SIZE) +
410 			    round_page(blksize) > maxiosize) {
411 				break;
412 			}
413 
414 			/*
415 			 * Shortcut some checks and try to avoid buffers that
416 			 * would block in the lock.  The same checks have to
417 			 * be made again after we officially get the buffer.
418 			 */
419 			tbp = getblk(vp, loffset + i * blksize, blksize,
420 				     GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
421 			if (tbp == NULL)
422 				break;
423 			for (j = 0; j < tbp->b_xio.xio_npages; j++) {
424 				if (tbp->b_xio.xio_pages[j]->valid)
425 					break;
426 			}
427 			if (j != tbp->b_xio.xio_npages) {
428 				bqrelse(tbp);
429 				break;
430 			}
431 
432 			/*
433 			 * Stop scanning if the buffer is fuly valid
434 			 * (marked B_CACHE), or locked (may be doing a
435 			 * background write), or if the buffer is not
436 			 * VMIO backed.  The clustering code can only deal
437 			 * with VMIO-backed buffers.
438 			 */
439 			if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
440 			    (tbp->b_flags & B_VMIO) == 0 ||
441 			    (LIST_FIRST(&tbp->b_dep) != NULL &&
442 			     buf_checkread(tbp))
443 			) {
444 				bqrelse(tbp);
445 				break;
446 			}
447 
448 			/*
449 			 * The buffer must be completely invalid in order to
450 			 * take part in the cluster.  If it is partially valid
451 			 * then we stop.
452 			 */
453 			for (j = 0;j < tbp->b_xio.xio_npages; j++) {
454 				if (tbp->b_xio.xio_pages[j]->valid)
455 					break;
456 			}
457 			if (j != tbp->b_xio.xio_npages) {
458 				bqrelse(tbp);
459 				break;
460 			}
461 
462 			/*
463 			 * Set a read-ahead mark as appropriate
464 			 */
465 			if (i == 1 || i == (run - 1))
466 				cluster_setram(tbp);
467 
468 			/*
469 			 * Depress the priority of buffers not explicitly
470 			 * requested.
471 			 */
472 			/* tbp->b_flags |= B_AGE; */
473 
474 			/*
475 			 * Set the block number if it isn't set, otherwise
476 			 * if it is make sure it matches the block number we
477 			 * expect.
478 			 */
479 			if (tbp->b_bio2.bio_offset == NOOFFSET) {
480 				tbp->b_bio2.bio_offset = boffset;
481 			} else if (tbp->b_bio2.bio_offset != boffset) {
482 				brelse(tbp);
483 				break;
484 			}
485 		}
486 
487 		/*
488 		 * The passed-in tbp (i == 0) will already be set up for
489 		 * async or sync operation.  All other tbp's acquire in
490 		 * our loop are set up for async operation.
491 		 */
492 		tbp->b_cmd = BUF_CMD_READ;
493 		BUF_KERNPROC(tbp);
494 		cluster_append(&bp->b_bio1, tbp);
495 		for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
496 			vm_page_t m;
497 			m = tbp->b_xio.xio_pages[j];
498 			vm_page_io_start(m);
499 			vm_object_pip_add(m->object, 1);
500 			if ((bp->b_xio.xio_npages == 0) ||
501 				(bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
502 				bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
503 				bp->b_xio.xio_npages++;
504 			}
505 			if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
506 				tbp->b_xio.xio_pages[j] = bogus_page;
507 		}
508 		/*
509 		 * XXX shouldn't this be += size for both, like in
510 		 * cluster_wbuild()?
511 		 *
512 		 * Don't inherit tbp->b_bufsize as it may be larger due to
513 		 * a non-page-aligned size.  Instead just aggregate using
514 		 * 'size'.
515 		 */
516 		if (tbp->b_bcount != blksize)
517 		    kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
518 		if (tbp->b_bufsize != blksize)
519 		    kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
520 		bp->b_bcount += blksize;
521 		bp->b_bufsize += blksize;
522 	}
523 
524 	/*
525 	 * Fully valid pages in the cluster are already good and do not need
526 	 * to be re-read from disk.  Replace the page with bogus_page
527 	 */
528 	for (j = 0; j < bp->b_xio.xio_npages; j++) {
529 		if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
530 		    VM_PAGE_BITS_ALL) {
531 			bp->b_xio.xio_pages[j] = bogus_page;
532 		}
533 	}
534 	if (bp->b_bufsize > bp->b_kvasize) {
535 		panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
536 		    bp->b_bufsize, bp->b_kvasize);
537 	}
538 	pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
539 		(vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
540 	BUF_KERNPROC(bp);
541 	return (bp);
542 }
543 
544 /*
545  * Cleanup after a clustered read or write.
546  * This is complicated by the fact that any of the buffers might have
547  * extra memory (if there were no empty buffer headers at allocbuf time)
548  * that we will need to shift around.
549  *
550  * The returned bio is &bp->b_bio1
551  */
552 void
553 cluster_callback(struct bio *bio)
554 {
555 	struct buf *bp = bio->bio_buf;
556 	struct buf *tbp;
557 	int error = 0;
558 
559 	/*
560 	 * Must propogate errors to all the components.  A short read (EOF)
561 	 * is a critical error.
562 	 */
563 	if (bp->b_flags & B_ERROR) {
564 		error = bp->b_error;
565 	} else if (bp->b_bcount != bp->b_bufsize) {
566 		panic("cluster_callback: unexpected EOF on cluster %p!", bio);
567 	}
568 
569 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
570 	/*
571 	 * Move memory from the large cluster buffer into the component
572 	 * buffers and mark IO as done on these.  Since the memory map
573 	 * is the same, no actual copying is required.
574 	 */
575 	while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
576 		bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
577 		if (error) {
578 			tbp->b_flags |= B_ERROR;
579 			tbp->b_error = error;
580 		} else {
581 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
582 			tbp->b_flags &= ~(B_ERROR|B_INVAL);
583 			/*
584 			 * XXX the bdwrite()/bqrelse() issued during
585 			 * cluster building clears B_RELBUF (see bqrelse()
586 			 * comment).  If direct I/O was specified, we have
587 			 * to restore it here to allow the buffer and VM
588 			 * to be freed.
589 			 */
590 			if (tbp->b_flags & B_DIRECT)
591 				tbp->b_flags |= B_RELBUF;
592 		}
593 		biodone(&tbp->b_bio1);
594 	}
595 	relpbuf(bp, &cluster_pbuf_freecnt);
596 }
597 
598 /*
599  *	cluster_wbuild_wb:
600  *
601  *	Implement modified write build for cluster.
602  *
603  *		write_behind = 0	write behind disabled
604  *		write_behind = 1	write behind normal (default)
605  *		write_behind = 2	write behind backed-off
606  */
607 
608 static __inline int
609 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
610 {
611 	int r = 0;
612 
613 	switch(write_behind) {
614 	case 2:
615 		if (start_loffset < len)
616 			break;
617 		start_loffset -= len;
618 		/* fall through */
619 	case 1:
620 		r = cluster_wbuild(vp, blksize, start_loffset, len);
621 		/* fall through */
622 	default:
623 		/* fall through */
624 		break;
625 	}
626 	return(r);
627 }
628 
629 /*
630  * Do clustered write for FFS.
631  *
632  * Three cases:
633  *	1. Write is not sequential (write asynchronously)
634  *	Write is sequential:
635  *	2.	beginning of cluster - begin cluster
636  *	3.	middle of a cluster - add to cluster
637  *	4.	end of a cluster - asynchronously write cluster
638  */
639 void
640 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
641 {
642 	struct vnode *vp;
643 	off_t loffset;
644 	int maxclen, cursize;
645 	int async;
646 
647 	vp = bp->b_vp;
648 	if (vp->v_type == VREG)
649 		async = vp->v_mount->mnt_flag & MNT_ASYNC;
650 	else
651 		async = 0;
652 	loffset = bp->b_loffset;
653 	KASSERT(bp->b_loffset != NOOFFSET,
654 		("cluster_write: no buffer offset"));
655 
656 	/* Initialize vnode to beginning of file. */
657 	if (loffset == 0)
658 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
659 
660 	if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
661 	    bp->b_bio2.bio_offset == NOOFFSET ||
662 	    (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
663 		maxclen = vmaxiosize(vp);
664 		if (vp->v_clen != 0) {
665 			/*
666 			 * Next block is not sequential.
667 			 *
668 			 * If we are not writing at end of file, the process
669 			 * seeked to another point in the file since its last
670 			 * write, or we have reached our maximum cluster size,
671 			 * then push the previous cluster. Otherwise try
672 			 * reallocating to make it sequential.
673 			 *
674 			 * Change to algorithm: only push previous cluster if
675 			 * it was sequential from the point of view of the
676 			 * seqcount heuristic, otherwise leave the buffer
677 			 * intact so we can potentially optimize the I/O
678 			 * later on in the buf_daemon or update daemon
679 			 * flush.
680 			 */
681 			cursize = vp->v_lastw - vp->v_cstart + blksize;
682 			if (bp->b_loffset + blksize != filesize ||
683 			    loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
684 				if (!async && seqcount > 0) {
685 					cluster_wbuild_wb(vp, blksize,
686 						vp->v_cstart, cursize);
687 				}
688 			} else {
689 				struct buf **bpp, **endbp;
690 				struct cluster_save *buflist;
691 
692 				buflist = cluster_collectbufs(vp, bp, blksize);
693 				endbp = &buflist->bs_children
694 				    [buflist->bs_nchildren - 1];
695 				if (VOP_REALLOCBLKS(vp, buflist)) {
696 					/*
697 					 * Failed, push the previous cluster
698 					 * if *really* writing sequentially
699 					 * in the logical file (seqcount > 1),
700 					 * otherwise delay it in the hopes that
701 					 * the low level disk driver can
702 					 * optimize the write ordering.
703 					 */
704 					for (bpp = buflist->bs_children;
705 					     bpp < endbp; bpp++)
706 						brelse(*bpp);
707 					kfree(buflist, M_SEGMENT);
708 					if (seqcount > 1) {
709 						cluster_wbuild_wb(vp,
710 						    blksize, vp->v_cstart,
711 						    cursize);
712 					}
713 				} else {
714 					/*
715 					 * Succeeded, keep building cluster.
716 					 */
717 					for (bpp = buflist->bs_children;
718 					     bpp <= endbp; bpp++)
719 						bdwrite(*bpp);
720 					kfree(buflist, M_SEGMENT);
721 					vp->v_lastw = loffset;
722 					vp->v_lasta = bp->b_bio2.bio_offset;
723 					return;
724 				}
725 			}
726 		}
727 		/*
728 		 * Consider beginning a cluster. If at end of file, make
729 		 * cluster as large as possible, otherwise find size of
730 		 * existing cluster.
731 		 */
732 		if ((vp->v_type == VREG) &&
733 		    bp->b_loffset + blksize != filesize &&
734 		    (bp->b_bio2.bio_offset == NOOFFSET) &&
735 		    (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
736 		     bp->b_bio2.bio_offset == NOOFFSET)) {
737 			bawrite(bp);
738 			vp->v_clen = 0;
739 			vp->v_lasta = bp->b_bio2.bio_offset;
740 			vp->v_cstart = loffset + blksize;
741 			vp->v_lastw = loffset;
742 			return;
743 		}
744 		if (maxclen > blksize)
745 			vp->v_clen = maxclen - blksize;
746 		else
747 			vp->v_clen = 0;
748 		if (!async && vp->v_clen == 0) { /* I/O not contiguous */
749 			vp->v_cstart = loffset + blksize;
750 			bawrite(bp);
751 		} else {	/* Wait for rest of cluster */
752 			vp->v_cstart = loffset;
753 			bdwrite(bp);
754 		}
755 	} else if (loffset == vp->v_cstart + vp->v_clen) {
756 		/*
757 		 * At end of cluster, write it out if seqcount tells us we
758 		 * are operating sequentially, otherwise let the buf or
759 		 * update daemon handle it.
760 		 */
761 		bdwrite(bp);
762 		if (seqcount > 1)
763 			cluster_wbuild_wb(vp, blksize, vp->v_cstart,
764 					  vp->v_clen + blksize);
765 		vp->v_clen = 0;
766 		vp->v_cstart = loffset + blksize;
767 	} else if (vm_page_count_severe()) {
768 		/*
769 		 * We are low on memory, get it going NOW
770 		 */
771 		bawrite(bp);
772 	} else {
773 		/*
774 		 * In the middle of a cluster, so just delay the I/O for now.
775 		 */
776 		bdwrite(bp);
777 	}
778 	vp->v_lastw = loffset;
779 	vp->v_lasta = bp->b_bio2.bio_offset;
780 }
781 
782 
783 /*
784  * This is an awful lot like cluster_rbuild...wish they could be combined.
785  * The last lbn argument is the current block on which I/O is being
786  * performed.  Check to see that it doesn't fall in the middle of
787  * the current block (if last_bp == NULL).
788  */
789 int
790 cluster_wbuild(struct vnode *vp, int blksize, off_t start_loffset, int bytes)
791 {
792 	struct buf *bp, *tbp;
793 	int i, j;
794 	int totalwritten = 0;
795 	int maxiosize = vmaxiosize(vp);
796 
797 	while (bytes > 0) {
798 		/*
799 		 * If the buffer is not delayed-write (i.e. dirty), or it
800 		 * is delayed-write but either locked or inval, it cannot
801 		 * partake in the clustered write.
802 		 */
803 		tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK);
804 		if (tbp == NULL ||
805 		    (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI ||
806 		    (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
807 			if (tbp)
808 				BUF_UNLOCK(tbp);
809 			start_loffset += blksize;
810 			bytes -= blksize;
811 			continue;
812 		}
813 		bremfree(tbp);
814 		KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
815 
816 		/*
817 		 * Extra memory in the buffer, punt on this buffer.
818 		 * XXX we could handle this in most cases, but we would
819 		 * have to push the extra memory down to after our max
820 		 * possible cluster size and then potentially pull it back
821 		 * up if the cluster was terminated prematurely--too much
822 		 * hassle.
823 		 */
824 		if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
825 		    (tbp->b_bcount != tbp->b_bufsize) ||
826 		    (tbp->b_bcount != blksize) ||
827 		    (bytes == blksize) ||
828 		    ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
829 			totalwritten += tbp->b_bufsize;
830 			bawrite(tbp);
831 			start_loffset += blksize;
832 			bytes -= blksize;
833 			continue;
834 		}
835 
836 		/*
837 		 * Set up the pbuf.  Track our append point with b_bcount
838 		 * and b_bufsize.  b_bufsize is not used by the device but
839 		 * our caller uses it to loop clusters and we use it to
840 		 * detect a premature EOF on the block device.
841 		 */
842 		bp->b_bcount = 0;
843 		bp->b_bufsize = 0;
844 		bp->b_xio.xio_npages = 0;
845 		bp->b_loffset = tbp->b_loffset;
846 		bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
847 
848 		/*
849 		 * We are synthesizing a buffer out of vm_page_t's, but
850 		 * if the block size is not page aligned then the starting
851 		 * address may not be either.  Inherit the b_data offset
852 		 * from the original buffer.
853 		 */
854 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
855 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
856 		bp->b_flags &= ~B_ERROR;
857 		bp->b_flags |= B_CLUSTER | B_BNOCLIP |
858 			(tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
859 		bp->b_bio1.bio_caller_info1.cluster_head = NULL;
860 		bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
861 
862 		/*
863 		 * From this location in the file, scan forward to see
864 		 * if there are buffers with adjacent data that need to
865 		 * be written as well.
866 		 */
867 		for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
868 			if (i != 0) { /* If not the first buffer */
869 				tbp = findblk(vp, start_loffset,
870 					      FINDBLK_NBLOCK);
871 				/*
872 				 * Buffer not found or could not be locked
873 				 * non-blocking.
874 				 */
875 				if (tbp == NULL)
876 					break;
877 
878 				/*
879 				 * If it IS in core, but has different
880 				 * characteristics, then don't cluster
881 				 * with it.
882 				 */
883 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
884 				     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
885 				    != (B_DELWRI | B_CLUSTEROK |
886 				     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
887 				    (tbp->b_flags & B_LOCKED) ||
888 				    (LIST_FIRST(&tbp->b_dep) &&
889 				     buf_checkwrite(tbp))
890 				) {
891 					BUF_UNLOCK(tbp);
892 					break;
893 				}
894 
895 				/*
896 				 * Check that the combined cluster
897 				 * would make sense with regard to pages
898 				 * and would not be too large
899 				 */
900 				if ((tbp->b_bcount != blksize) ||
901 				  ((bp->b_bio2.bio_offset + i) !=
902 				    tbp->b_bio2.bio_offset) ||
903 				  ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
904 				    (maxiosize / PAGE_SIZE))) {
905 					BUF_UNLOCK(tbp);
906 					break;
907 				}
908 				/*
909 				 * Ok, it's passed all the tests,
910 				 * so remove it from the free list
911 				 * and mark it busy. We will use it.
912 				 */
913 				bremfree(tbp);
914 				KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
915 			} /* end of code for non-first buffers only */
916 
917 			/*
918 			 * If the IO is via the VM then we do some
919 			 * special VM hackery (yuck).  Since the buffer's
920 			 * block size may not be page-aligned it is possible
921 			 * for a page to be shared between two buffers.  We
922 			 * have to get rid of the duplication when building
923 			 * the cluster.
924 			 */
925 			if (tbp->b_flags & B_VMIO) {
926 				vm_page_t m;
927 
928 				if (i != 0) { /* if not first buffer */
929 					for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
930 						m = tbp->b_xio.xio_pages[j];
931 						if (m->flags & PG_BUSY) {
932 							bqrelse(tbp);
933 							goto finishcluster;
934 						}
935 					}
936 				}
937 
938 				for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
939 					m = tbp->b_xio.xio_pages[j];
940 					vm_page_io_start(m);
941 					vm_object_pip_add(m->object, 1);
942 					if ((bp->b_xio.xio_npages == 0) ||
943 					  (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
944 						bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
945 						bp->b_xio.xio_npages++;
946 					}
947 				}
948 			}
949 			bp->b_bcount += blksize;
950 			bp->b_bufsize += blksize;
951 
952 			bundirty(tbp);
953 			tbp->b_flags &= ~B_ERROR;
954 			tbp->b_cmd = BUF_CMD_WRITE;
955 			BUF_KERNPROC(tbp);
956 			cluster_append(&bp->b_bio1, tbp);
957 
958 			/*
959 			 * check for latent dependencies to be handled
960 			 */
961 			if (LIST_FIRST(&tbp->b_dep) != NULL)
962 				buf_start(tbp);
963 		}
964 	finishcluster:
965 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
966 			(vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
967 		if (bp->b_bufsize > bp->b_kvasize) {
968 			panic(
969 			    "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
970 			    bp->b_bufsize, bp->b_kvasize);
971 		}
972 		totalwritten += bp->b_bufsize;
973 		bp->b_dirtyoff = 0;
974 		bp->b_dirtyend = bp->b_bufsize;
975 		bp->b_bio1.bio_done = cluster_callback;
976 		bp->b_cmd = BUF_CMD_WRITE;
977 
978 		vfs_busy_pages(vp, bp);
979 		bp->b_runningbufspace = bp->b_bufsize;
980 		if (bp->b_runningbufspace) {
981 			runningbufspace += bp->b_runningbufspace;
982 			++runningbufcount;
983 		}
984 		BUF_KERNPROC(bp);
985 		vn_strategy(vp, &bp->b_bio1);
986 
987 		bytes -= i;
988 	}
989 	return totalwritten;
990 }
991 
992 /*
993  * Collect together all the buffers in a cluster.
994  * Plus add one additional buffer.
995  */
996 static struct cluster_save *
997 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
998 {
999 	struct cluster_save *buflist;
1000 	struct buf *bp;
1001 	off_t loffset;
1002 	int i, len;
1003 
1004 	len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
1005 	buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1006 			 M_SEGMENT, M_WAITOK);
1007 	buflist->bs_nchildren = 0;
1008 	buflist->bs_children = (struct buf **) (buflist + 1);
1009 	for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
1010 		(void) bread(vp, loffset, last_bp->b_bcount, &bp);
1011 		buflist->bs_children[i] = bp;
1012 		if (bp->b_bio2.bio_offset == NOOFFSET) {
1013 			VOP_BMAP(bp->b_vp, bp->b_loffset,
1014 				 &bp->b_bio2.bio_offset,
1015 				 NULL, NULL, BUF_CMD_WRITE);
1016 		}
1017 	}
1018 	buflist->bs_children[i] = bp = last_bp;
1019 	if (bp->b_bio2.bio_offset == NOOFFSET) {
1020 		VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1021 			 NULL, NULL, BUF_CMD_WRITE);
1022 	}
1023 	buflist->bs_nchildren = i + 1;
1024 	return (buflist);
1025 }
1026 
1027 void
1028 cluster_append(struct bio *bio, struct buf *tbp)
1029 {
1030 	tbp->b_cluster_next = NULL;
1031 	if (bio->bio_caller_info1.cluster_head == NULL) {
1032 		bio->bio_caller_info1.cluster_head = tbp;
1033 		bio->bio_caller_info2.cluster_tail = tbp;
1034 	} else {
1035 		bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1036 		bio->bio_caller_info2.cluster_tail = tbp;
1037 	}
1038 }
1039 
1040 static
1041 void
1042 cluster_setram (struct buf *bp)
1043 {
1044 	bp->b_flags |= B_RAM;
1045 	if (bp->b_xio.xio_npages)
1046 		vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM);
1047 }
1048