xref: /dflybsd-src/sys/kern/vfs_cluster.c (revision dae741e33c840b92a8a53bf9f01157ede145e256)
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     "Cluster write-behind setting");
82 static int max_readahead = 2 * 1024 * 1024;
83 SYSCTL_INT(_vfs, OID_AUTO, max_readahead, CTLFLAG_RW, &max_readahead, 0,
84     "Limit in bytes for desired cluster read-ahead");
85 
86 extern vm_page_t	bogus_page;
87 
88 extern int cluster_pbuf_freecnt;
89 
90 /*
91  * This replaces bread.
92  *
93  * filesize	- read-ahead @ blksize will not cross this boundary
94  * loffset	- loffset for returned *bpp
95  * blksize	- blocksize for returned *bpp and read-ahead bps
96  * minreq	- minimum (not a hard minimum) in bytes, typically reflects
97  *		  a higher level uio resid.
98  * maxreq	- maximum (sequential heuristic) in bytes (highet typ ~2MB)
99  * bpp		- return buffer (*bpp) for (loffset,blksize)
100  */
101 int
102 cluster_read(struct vnode *vp, off_t filesize, off_t loffset,
103 	     int blksize, size_t minreq, size_t maxreq, struct buf **bpp)
104 {
105 	struct buf *bp, *rbp, *reqbp;
106 	off_t origoffset;
107 	off_t doffset;
108 	int error;
109 	int i;
110 	int maxra;
111 	int maxrbuild;
112 
113 	error = 0;
114 
115 	/*
116 	 * Calculate the desired read-ahead in blksize'd blocks (maxra).
117 	 * To do this we calculate maxreq.
118 	 *
119 	 * maxreq typically starts out as a sequential heuristic.  If the
120 	 * high level uio/resid is bigger (minreq), we pop maxreq up to
121 	 * minreq.  This represents the case where random I/O is being
122 	 * performed by the userland is issuing big read()'s.
123 	 *
124 	 * Then we limit maxreq to max_readahead to ensure it is a reasonable
125 	 * value.
126 	 *
127 	 * Finally we must ensure that (loffset + maxreq) does not cross the
128 	 * boundary (filesize) for the current blocksize.  If we allowed it
129 	 * to cross we could end up with buffers past the boundary with the
130 	 * wrong block size (HAMMER large-data areas use mixed block sizes).
131 	 * minreq is also absolutely limited to filesize.
132 	 */
133 	if (maxreq < minreq)
134 		maxreq = minreq;
135 	/* minreq not used beyond this point */
136 
137 	if (maxreq > max_readahead) {
138 		maxreq = max_readahead;
139 		if (maxreq > 16 * 1024 * 1024)
140 			maxreq = 16 * 1024 * 1024;
141 	}
142 	if (maxreq < blksize)
143 		maxreq = blksize;
144 	if (loffset + maxreq > filesize) {
145 		if (loffset > filesize)
146 			maxreq = 0;
147 		else
148 			maxreq = filesize - loffset;
149 	}
150 
151 	maxra = (int)(maxreq / blksize);
152 
153 	/*
154 	 * Get the requested block.
155 	 */
156 	*bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0);
157 	origoffset = loffset;
158 
159 	/*
160 	 * Calculate the maximum cluster size for a single I/O, used
161 	 * by cluster_rbuild().
162 	 */
163 	maxrbuild = vmaxiosize(vp) / blksize;
164 
165 	/*
166 	 * if it is in the cache, then check to see if the reads have been
167 	 * sequential.  If they have, then try some read-ahead, otherwise
168 	 * back-off on prospective read-aheads.
169 	 */
170 	if (bp->b_flags & B_CACHE) {
171 		/*
172 		 * Not sequential, do not do any read-ahead
173 		 */
174 		if (maxra <= 1)
175 			return 0;
176 
177 		/*
178 		 * No read-ahead mark, do not do any read-ahead
179 		 * yet.
180 		 */
181 		if ((bp->b_flags & B_RAM) == 0)
182 			return 0;
183 
184 		/*
185 		 * We hit a read-ahead-mark, figure out how much read-ahead
186 		 * to do (maxra) and where to start (loffset).
187 		 *
188 		 * Shortcut the scan.  Typically the way this works is that
189 		 * we've built up all the blocks inbetween except for the
190 		 * last in previous iterations, so if the second-to-last
191 		 * block is present we just skip ahead to it.
192 		 *
193 		 * This algorithm has O(1) cpu in the steady state no
194 		 * matter how large maxra is.
195 		 */
196 		bp->b_flags &= ~B_RAM;
197 
198 		if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST))
199 			i = maxra - 1;
200 		else
201 			i = 1;
202 		while (i < maxra) {
203 			if (findblk(vp, loffset + i * blksize,
204 				    FINDBLK_TEST) == NULL) {
205 				break;
206 			}
207 			++i;
208 		}
209 
210 		/*
211 		 * We got everything or everything is in the cache, no
212 		 * point continuing.
213 		 */
214 		if (i >= maxra)
215 			return 0;
216 		maxra -= i;
217 		loffset += i * blksize;
218 		reqbp = bp = NULL;
219 	} else {
220 		__debugvar off_t firstread = bp->b_loffset;
221 		int nblks;
222 
223 		/*
224 		 * Set-up synchronous read for bp.
225 		 */
226 		bp->b_cmd = BUF_CMD_READ;
227 		bp->b_bio1.bio_done = biodone_sync;
228 		bp->b_bio1.bio_flags |= BIO_SYNC;
229 
230 		KASSERT(firstread != NOOFFSET,
231 			("cluster_read: no buffer offset"));
232 
233 		/*
234 		 * nblks is our cluster_rbuild request size, limited
235 		 * primarily by the device.
236 		 */
237 		if ((nblks = maxra) > maxrbuild)
238 			nblks = maxrbuild;
239 
240 		if (nblks > 1) {
241 			int burstbytes;
242 
243 	    		error = VOP_BMAP(vp, loffset, &doffset,
244 					 &burstbytes, NULL, BUF_CMD_READ);
245 			if (error)
246 				goto single_block_read;
247 			if (nblks > burstbytes / blksize)
248 				nblks = burstbytes / blksize;
249 			if (doffset == NOOFFSET)
250 				goto single_block_read;
251 			if (nblks <= 1)
252 				goto single_block_read;
253 
254 			bp = cluster_rbuild(vp, filesize, loffset,
255 					    doffset, blksize, nblks, bp);
256 			loffset += bp->b_bufsize;
257 			maxra -= bp->b_bufsize / blksize;
258 		} else {
259 single_block_read:
260 			/*
261 			 * If it isn't in the cache, then get a chunk from
262 			 * disk if sequential, otherwise just get the block.
263 			 */
264 			cluster_setram(bp);
265 			loffset += blksize;
266 			--maxra;
267 		}
268 	}
269 
270 	/*
271 	 * If B_CACHE was not set issue bp.  bp will either be an
272 	 * asynchronous cluster buf or a synchronous single-buf.
273 	 * If it is a single buf it will be the same as reqbp.
274 	 *
275 	 * NOTE: Once an async cluster buf is issued bp becomes invalid.
276 	 */
277 	if (bp) {
278 #if defined(CLUSTERDEBUG)
279 		if (rcluster)
280 			kprintf("S(%012jx,%d,%d)\n",
281 			    (intmax_t)bp->b_loffset, bp->b_bcount, maxra);
282 #endif
283 		if ((bp->b_flags & B_CLUSTER) == 0)
284 			vfs_busy_pages(vp, bp);
285 		bp->b_flags &= ~(B_ERROR|B_INVAL);
286 		vn_strategy(vp, &bp->b_bio1);
287 		error = 0;
288 		/* bp invalid now */
289 	}
290 
291 	/*
292 	 * If we have been doing sequential I/O, then do some read-ahead.
293 	 * The code above us should have positioned us at the next likely
294 	 * offset.
295 	 *
296 	 * Only mess with buffers which we can immediately lock.  HAMMER
297 	 * will do device-readahead irrespective of what the blocks
298 	 * represent.
299 	 */
300 	while (error == 0 && maxra > 0) {
301 		int burstbytes;
302 		int tmp_error;
303 		int nblks;
304 
305 		rbp = getblk(vp, loffset, blksize,
306 			     GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
307 		if (rbp == NULL)
308 			goto no_read_ahead;
309 		if ((rbp->b_flags & B_CACHE)) {
310 			bqrelse(rbp);
311 			goto no_read_ahead;
312 		}
313 
314 		/*
315 		 * An error from the read-ahead bmap has nothing to do
316 		 * with the caller's original request.
317 		 */
318 		tmp_error = VOP_BMAP(vp, loffset, &doffset,
319 				     &burstbytes, NULL, BUF_CMD_READ);
320 		if (tmp_error || doffset == NOOFFSET) {
321 			rbp->b_flags |= B_INVAL;
322 			brelse(rbp);
323 			rbp = NULL;
324 			goto no_read_ahead;
325 		}
326 		if ((nblks = maxra) > maxrbuild)
327 			nblks = maxrbuild;
328 		if (nblks > burstbytes / blksize)
329 			nblks = burstbytes / blksize;
330 
331 		/*
332 		 * rbp: async read
333 		 */
334 		rbp->b_cmd = BUF_CMD_READ;
335 		/*rbp->b_flags |= B_AGE*/;
336 		cluster_setram(rbp);
337 
338 		if (nblks > 1) {
339 			rbp = cluster_rbuild(vp, filesize, loffset,
340 					     doffset, blksize,
341 					     nblks, rbp);
342 		} else {
343 			rbp->b_bio2.bio_offset = doffset;
344 		}
345 
346 #if defined(CLUSTERDEBUG)
347 		if (rcluster) {
348 			if (bp) {
349 				kprintf("A+(%012jx,%d,%jd) "
350 					"doff=%012jx minr=%zd ra=%d\n",
351 				    (intmax_t)loffset, rbp->b_bcount,
352 				    (intmax_t)(loffset - origoffset),
353 				    (intmax_t)doffset, minreq, maxra);
354 			} else {
355 				kprintf("A-(%012jx,%d,%jd) "
356 					"doff=%012jx minr=%zd ra=%d\n",
357 				    (intmax_t)rbp->b_loffset, rbp->b_bcount,
358 				    (intmax_t)(loffset - origoffset),
359 				    (intmax_t)doffset, minreq, maxra);
360 			}
361 		}
362 #endif
363 		rbp->b_flags &= ~(B_ERROR|B_INVAL);
364 
365 		if ((rbp->b_flags & B_CLUSTER) == 0)
366 			vfs_busy_pages(vp, rbp);
367 		BUF_KERNPROC(rbp);
368 		loffset += rbp->b_bufsize;
369 		maxra -= rbp->b_bufsize / blksize;
370 		vn_strategy(vp, &rbp->b_bio1);
371 		/* rbp invalid now */
372 	}
373 
374 	/*
375 	 * Wait for our original buffer to complete its I/O.  reqbp will
376 	 * be NULL if the original buffer was B_CACHE.  We are returning
377 	 * (*bpp) which is the same as reqbp when reqbp != NULL.
378 	 */
379 no_read_ahead:
380 	if (reqbp) {
381 		KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
382 		error = biowait(&reqbp->b_bio1, "clurd");
383 	}
384 	return (error);
385 }
386 
387 /*
388  * If blocks are contiguous on disk, use this to provide clustered
389  * read ahead.  We will read as many blocks as possible sequentially
390  * and then parcel them up into logical blocks in the buffer hash table.
391  *
392  * This function either returns a cluster buf or it returns fbp.  fbp is
393  * already expected to be set up as a synchronous or asynchronous request.
394  *
395  * If a cluster buf is returned it will always be async.
396  */
397 static struct buf *
398 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
399 	       int blksize, int run, struct buf *fbp)
400 {
401 	struct buf *bp, *tbp;
402 	off_t boffset;
403 	int i, j;
404 	int maxiosize = vmaxiosize(vp);
405 
406 	/*
407 	 * avoid a division
408 	 */
409 	while (loffset + run * blksize > filesize) {
410 		--run;
411 	}
412 
413 	tbp = fbp;
414 	tbp->b_bio2.bio_offset = doffset;
415 	if((tbp->b_flags & B_MALLOC) ||
416 	    ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
417 		return tbp;
418 	}
419 
420 	bp = trypbuf_kva(&cluster_pbuf_freecnt);
421 	if (bp == NULL) {
422 		return tbp;
423 	}
424 
425 	/*
426 	 * We are synthesizing a buffer out of vm_page_t's, but
427 	 * if the block size is not page aligned then the starting
428 	 * address may not be either.  Inherit the b_data offset
429 	 * from the original buffer.
430 	 */
431 	bp->b_data = (char *)((vm_offset_t)bp->b_data |
432 	    ((vm_offset_t)tbp->b_data & PAGE_MASK));
433 	bp->b_flags |= B_CLUSTER | B_VMIO;
434 	bp->b_cmd = BUF_CMD_READ;
435 	bp->b_bio1.bio_done = cluster_callback;		/* default to async */
436 	bp->b_bio1.bio_caller_info1.cluster_head = NULL;
437 	bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
438 	bp->b_loffset = loffset;
439 	bp->b_bio2.bio_offset = doffset;
440 	KASSERT(bp->b_loffset != NOOFFSET,
441 		("cluster_rbuild: no buffer offset"));
442 
443 	bp->b_bcount = 0;
444 	bp->b_bufsize = 0;
445 	bp->b_xio.xio_npages = 0;
446 
447 	for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
448 		if (i) {
449 			if ((bp->b_xio.xio_npages * PAGE_SIZE) +
450 			    round_page(blksize) > maxiosize) {
451 				break;
452 			}
453 
454 			/*
455 			 * Shortcut some checks and try to avoid buffers that
456 			 * would block in the lock.  The same checks have to
457 			 * be made again after we officially get the buffer.
458 			 */
459 			tbp = getblk(vp, loffset + i * blksize, blksize,
460 				     GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
461 			if (tbp == NULL)
462 				break;
463 			for (j = 0; j < tbp->b_xio.xio_npages; j++) {
464 				if (tbp->b_xio.xio_pages[j]->valid)
465 					break;
466 			}
467 			if (j != tbp->b_xio.xio_npages) {
468 				bqrelse(tbp);
469 				break;
470 			}
471 
472 			/*
473 			 * Stop scanning if the buffer is fuly valid
474 			 * (marked B_CACHE), or locked (may be doing a
475 			 * background write), or if the buffer is not
476 			 * VMIO backed.  The clustering code can only deal
477 			 * with VMIO-backed buffers.
478 			 */
479 			if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
480 			    (tbp->b_flags & B_VMIO) == 0 ||
481 			    (LIST_FIRST(&tbp->b_dep) != NULL &&
482 			     buf_checkread(tbp))
483 			) {
484 				bqrelse(tbp);
485 				break;
486 			}
487 
488 			/*
489 			 * The buffer must be completely invalid in order to
490 			 * take part in the cluster.  If it is partially valid
491 			 * then we stop.
492 			 */
493 			for (j = 0;j < tbp->b_xio.xio_npages; j++) {
494 				if (tbp->b_xio.xio_pages[j]->valid)
495 					break;
496 			}
497 			if (j != tbp->b_xio.xio_npages) {
498 				bqrelse(tbp);
499 				break;
500 			}
501 
502 			/*
503 			 * Set a read-ahead mark as appropriate
504 			 */
505 			if (i == 1 || i == (run - 1))
506 				cluster_setram(tbp);
507 
508 			/*
509 			 * Depress the priority of buffers not explicitly
510 			 * requested.
511 			 */
512 			/* tbp->b_flags |= B_AGE; */
513 
514 			/*
515 			 * Set the block number if it isn't set, otherwise
516 			 * if it is make sure it matches the block number we
517 			 * expect.
518 			 */
519 			if (tbp->b_bio2.bio_offset == NOOFFSET) {
520 				tbp->b_bio2.bio_offset = boffset;
521 			} else if (tbp->b_bio2.bio_offset != boffset) {
522 				brelse(tbp);
523 				break;
524 			}
525 		}
526 
527 		/*
528 		 * The passed-in tbp (i == 0) will already be set up for
529 		 * async or sync operation.  All other tbp's acquire in
530 		 * our loop are set up for async operation.
531 		 */
532 		tbp->b_cmd = BUF_CMD_READ;
533 		BUF_KERNPROC(tbp);
534 		cluster_append(&bp->b_bio1, tbp);
535 		for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
536 			vm_page_t m;
537 
538 			m = tbp->b_xio.xio_pages[j];
539 			vm_page_busy_wait(m, FALSE, "clurpg");
540 			vm_page_io_start(m);
541 			vm_page_wakeup(m);
542 			vm_object_pip_add(m->object, 1);
543 			if ((bp->b_xio.xio_npages == 0) ||
544 				(bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
545 				bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
546 				bp->b_xio.xio_npages++;
547 			}
548 			if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
549 				tbp->b_xio.xio_pages[j] = bogus_page;
550 		}
551 		/*
552 		 * XXX shouldn't this be += size for both, like in
553 		 * cluster_wbuild()?
554 		 *
555 		 * Don't inherit tbp->b_bufsize as it may be larger due to
556 		 * a non-page-aligned size.  Instead just aggregate using
557 		 * 'size'.
558 		 */
559 		if (tbp->b_bcount != blksize)
560 		    kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
561 		if (tbp->b_bufsize != blksize)
562 		    kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
563 		bp->b_bcount += blksize;
564 		bp->b_bufsize += blksize;
565 	}
566 
567 	/*
568 	 * Fully valid pages in the cluster are already good and do not need
569 	 * to be re-read from disk.  Replace the page with bogus_page
570 	 */
571 	for (j = 0; j < bp->b_xio.xio_npages; j++) {
572 		if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
573 		    VM_PAGE_BITS_ALL) {
574 			bp->b_xio.xio_pages[j] = bogus_page;
575 		}
576 	}
577 	if (bp->b_bufsize > bp->b_kvasize) {
578 		panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
579 		    bp->b_bufsize, bp->b_kvasize);
580 	}
581 	pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
582 		(vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
583 	BUF_KERNPROC(bp);
584 	return (bp);
585 }
586 
587 /*
588  * Cleanup after a clustered read or write.
589  * This is complicated by the fact that any of the buffers might have
590  * extra memory (if there were no empty buffer headers at allocbuf time)
591  * that we will need to shift around.
592  *
593  * The returned bio is &bp->b_bio1
594  */
595 void
596 cluster_callback(struct bio *bio)
597 {
598 	struct buf *bp = bio->bio_buf;
599 	struct buf *tbp;
600 	int error = 0;
601 
602 	/*
603 	 * Must propogate errors to all the components.  A short read (EOF)
604 	 * is a critical error.
605 	 */
606 	if (bp->b_flags & B_ERROR) {
607 		error = bp->b_error;
608 	} else if (bp->b_bcount != bp->b_bufsize) {
609 		panic("cluster_callback: unexpected EOF on cluster %p!", bio);
610 	}
611 
612 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
613 	/*
614 	 * Move memory from the large cluster buffer into the component
615 	 * buffers and mark IO as done on these.  Since the memory map
616 	 * is the same, no actual copying is required.
617 	 */
618 	while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
619 		bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
620 		if (error) {
621 			tbp->b_flags |= B_ERROR | B_IODEBUG;
622 			tbp->b_error = error;
623 		} else {
624 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
625 			tbp->b_flags &= ~(B_ERROR|B_INVAL);
626 			tbp->b_flags |= B_IODEBUG;
627 			/*
628 			 * XXX the bdwrite()/bqrelse() issued during
629 			 * cluster building clears B_RELBUF (see bqrelse()
630 			 * comment).  If direct I/O was specified, we have
631 			 * to restore it here to allow the buffer and VM
632 			 * to be freed.
633 			 */
634 			if (tbp->b_flags & B_DIRECT)
635 				tbp->b_flags |= B_RELBUF;
636 		}
637 		biodone(&tbp->b_bio1);
638 	}
639 	relpbuf(bp, &cluster_pbuf_freecnt);
640 }
641 
642 /*
643  *	cluster_wbuild_wb:
644  *
645  *	Implement modified write build for cluster.
646  *
647  *		write_behind = 0	write behind disabled
648  *		write_behind = 1	write behind normal (default)
649  *		write_behind = 2	write behind backed-off
650  */
651 
652 static __inline int
653 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
654 {
655 	int r = 0;
656 
657 	switch(write_behind) {
658 	case 2:
659 		if (start_loffset < len)
660 			break;
661 		start_loffset -= len;
662 		/* fall through */
663 	case 1:
664 		r = cluster_wbuild(vp, blksize, start_loffset, len);
665 		/* fall through */
666 	default:
667 		/* fall through */
668 		break;
669 	}
670 	return(r);
671 }
672 
673 /*
674  * Do clustered write for FFS.
675  *
676  * Three cases:
677  *	1. Write is not sequential (write asynchronously)
678  *	Write is sequential:
679  *	2.	beginning of cluster - begin cluster
680  *	3.	middle of a cluster - add to cluster
681  *	4.	end of a cluster - asynchronously write cluster
682  */
683 void
684 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
685 {
686 	struct vnode *vp;
687 	off_t loffset;
688 	int maxclen, cursize;
689 	int async;
690 
691 	vp = bp->b_vp;
692 	if (vp->v_type == VREG)
693 		async = vp->v_mount->mnt_flag & MNT_ASYNC;
694 	else
695 		async = 0;
696 	loffset = bp->b_loffset;
697 	KASSERT(bp->b_loffset != NOOFFSET,
698 		("cluster_write: no buffer offset"));
699 
700 	/* Initialize vnode to beginning of file. */
701 	if (loffset == 0)
702 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
703 
704 	if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
705 	    bp->b_bio2.bio_offset == NOOFFSET ||
706 	    (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
707 		maxclen = vmaxiosize(vp);
708 		if (vp->v_clen != 0) {
709 			/*
710 			 * Next block is not sequential.
711 			 *
712 			 * If we are not writing at end of file, the process
713 			 * seeked to another point in the file since its last
714 			 * write, or we have reached our maximum cluster size,
715 			 * then push the previous cluster. Otherwise try
716 			 * reallocating to make it sequential.
717 			 *
718 			 * Change to algorithm: only push previous cluster if
719 			 * it was sequential from the point of view of the
720 			 * seqcount heuristic, otherwise leave the buffer
721 			 * intact so we can potentially optimize the I/O
722 			 * later on in the buf_daemon or update daemon
723 			 * flush.
724 			 */
725 			cursize = vp->v_lastw - vp->v_cstart + blksize;
726 			if (bp->b_loffset + blksize != filesize ||
727 			    loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
728 				if (!async && seqcount > 0) {
729 					cluster_wbuild_wb(vp, blksize,
730 						vp->v_cstart, cursize);
731 				}
732 			} else {
733 				struct buf **bpp, **endbp;
734 				struct cluster_save *buflist;
735 
736 				buflist = cluster_collectbufs(vp, bp, blksize);
737 				endbp = &buflist->bs_children
738 				    [buflist->bs_nchildren - 1];
739 				if (VOP_REALLOCBLKS(vp, buflist)) {
740 					/*
741 					 * Failed, push the previous cluster
742 					 * if *really* writing sequentially
743 					 * in the logical file (seqcount > 1),
744 					 * otherwise delay it in the hopes that
745 					 * the low level disk driver can
746 					 * optimize the write ordering.
747 					 */
748 					for (bpp = buflist->bs_children;
749 					     bpp < endbp; bpp++)
750 						brelse(*bpp);
751 					kfree(buflist, M_SEGMENT);
752 					if (seqcount > 1) {
753 						cluster_wbuild_wb(vp,
754 						    blksize, vp->v_cstart,
755 						    cursize);
756 					}
757 				} else {
758 					/*
759 					 * Succeeded, keep building cluster.
760 					 */
761 					for (bpp = buflist->bs_children;
762 					     bpp <= endbp; bpp++)
763 						bdwrite(*bpp);
764 					kfree(buflist, M_SEGMENT);
765 					vp->v_lastw = loffset;
766 					vp->v_lasta = bp->b_bio2.bio_offset;
767 					return;
768 				}
769 			}
770 		}
771 		/*
772 		 * Consider beginning a cluster. If at end of file, make
773 		 * cluster as large as possible, otherwise find size of
774 		 * existing cluster.
775 		 */
776 		if ((vp->v_type == VREG) &&
777 		    bp->b_loffset + blksize != filesize &&
778 		    (bp->b_bio2.bio_offset == NOOFFSET) &&
779 		    (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
780 		     bp->b_bio2.bio_offset == NOOFFSET)) {
781 			bawrite(bp);
782 			vp->v_clen = 0;
783 			vp->v_lasta = bp->b_bio2.bio_offset;
784 			vp->v_cstart = loffset + blksize;
785 			vp->v_lastw = loffset;
786 			return;
787 		}
788 		if (maxclen > blksize)
789 			vp->v_clen = maxclen - blksize;
790 		else
791 			vp->v_clen = 0;
792 		if (!async && vp->v_clen == 0) { /* I/O not contiguous */
793 			vp->v_cstart = loffset + blksize;
794 			bawrite(bp);
795 		} else {	/* Wait for rest of cluster */
796 			vp->v_cstart = loffset;
797 			bdwrite(bp);
798 		}
799 	} else if (loffset == vp->v_cstart + vp->v_clen) {
800 		/*
801 		 * At end of cluster, write it out if seqcount tells us we
802 		 * are operating sequentially, otherwise let the buf or
803 		 * update daemon handle it.
804 		 */
805 		bdwrite(bp);
806 		if (seqcount > 1)
807 			cluster_wbuild_wb(vp, blksize, vp->v_cstart,
808 					  vp->v_clen + blksize);
809 		vp->v_clen = 0;
810 		vp->v_cstart = loffset + blksize;
811 	} else if (vm_page_count_severe()) {
812 		/*
813 		 * We are low on memory, get it going NOW
814 		 */
815 		bawrite(bp);
816 	} else {
817 		/*
818 		 * In the middle of a cluster, so just delay the I/O for now.
819 		 */
820 		bdwrite(bp);
821 	}
822 	vp->v_lastw = loffset;
823 	vp->v_lasta = bp->b_bio2.bio_offset;
824 }
825 
826 
827 /*
828  * This is an awful lot like cluster_rbuild...wish they could be combined.
829  * The last lbn argument is the current block on which I/O is being
830  * performed.  Check to see that it doesn't fall in the middle of
831  * the current block (if last_bp == NULL).
832  */
833 int
834 cluster_wbuild(struct vnode *vp, int blksize, off_t start_loffset, int bytes)
835 {
836 	struct buf *bp, *tbp;
837 	int i, j;
838 	int totalwritten = 0;
839 	int maxiosize = vmaxiosize(vp);
840 
841 	while (bytes > 0) {
842 		/*
843 		 * If the buffer is not delayed-write (i.e. dirty), or it
844 		 * is delayed-write but either locked or inval, it cannot
845 		 * partake in the clustered write.
846 		 */
847 		tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK);
848 		if (tbp == NULL ||
849 		    (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI ||
850 		    (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
851 			if (tbp)
852 				BUF_UNLOCK(tbp);
853 			start_loffset += blksize;
854 			bytes -= blksize;
855 			continue;
856 		}
857 		bremfree(tbp);
858 		KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
859 
860 		/*
861 		 * Extra memory in the buffer, punt on this buffer.
862 		 * XXX we could handle this in most cases, but we would
863 		 * have to push the extra memory down to after our max
864 		 * possible cluster size and then potentially pull it back
865 		 * up if the cluster was terminated prematurely--too much
866 		 * hassle.
867 		 */
868 		if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
869 		    (tbp->b_bcount != tbp->b_bufsize) ||
870 		    (tbp->b_bcount != blksize) ||
871 		    (bytes == blksize) ||
872 		    ((bp = getpbuf_kva(&cluster_pbuf_freecnt)) == NULL)) {
873 			totalwritten += tbp->b_bufsize;
874 			bawrite(tbp);
875 			start_loffset += blksize;
876 			bytes -= blksize;
877 			continue;
878 		}
879 
880 		/*
881 		 * Set up the pbuf.  Track our append point with b_bcount
882 		 * and b_bufsize.  b_bufsize is not used by the device but
883 		 * our caller uses it to loop clusters and we use it to
884 		 * detect a premature EOF on the block device.
885 		 */
886 		bp->b_bcount = 0;
887 		bp->b_bufsize = 0;
888 		bp->b_xio.xio_npages = 0;
889 		bp->b_loffset = tbp->b_loffset;
890 		bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
891 
892 		/*
893 		 * We are synthesizing a buffer out of vm_page_t's, but
894 		 * if the block size is not page aligned then the starting
895 		 * address may not be either.  Inherit the b_data offset
896 		 * from the original buffer.
897 		 */
898 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
899 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
900 		bp->b_flags &= ~B_ERROR;
901 		bp->b_flags |= B_CLUSTER | B_BNOCLIP |
902 			(tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
903 		bp->b_bio1.bio_caller_info1.cluster_head = NULL;
904 		bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
905 
906 		/*
907 		 * From this location in the file, scan forward to see
908 		 * if there are buffers with adjacent data that need to
909 		 * be written as well.
910 		 */
911 		for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
912 			if (i != 0) { /* If not the first buffer */
913 				tbp = findblk(vp, start_loffset,
914 					      FINDBLK_NBLOCK);
915 				/*
916 				 * Buffer not found or could not be locked
917 				 * non-blocking.
918 				 */
919 				if (tbp == NULL)
920 					break;
921 
922 				/*
923 				 * If it IS in core, but has different
924 				 * characteristics, then don't cluster
925 				 * with it.
926 				 */
927 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
928 				     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
929 				    != (B_DELWRI | B_CLUSTEROK |
930 				     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
931 				    (tbp->b_flags & B_LOCKED) ||
932 				    (LIST_FIRST(&tbp->b_dep) &&
933 				     buf_checkwrite(tbp))
934 				) {
935 					BUF_UNLOCK(tbp);
936 					break;
937 				}
938 
939 				/*
940 				 * Check that the combined cluster
941 				 * would make sense with regard to pages
942 				 * and would not be too large
943 				 */
944 				if ((tbp->b_bcount != blksize) ||
945 				  ((bp->b_bio2.bio_offset + i) !=
946 				    tbp->b_bio2.bio_offset) ||
947 				  ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
948 				    (maxiosize / PAGE_SIZE))) {
949 					BUF_UNLOCK(tbp);
950 					break;
951 				}
952 				/*
953 				 * Ok, it's passed all the tests,
954 				 * so remove it from the free list
955 				 * and mark it busy. We will use it.
956 				 */
957 				bremfree(tbp);
958 				KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
959 			} /* end of code for non-first buffers only */
960 
961 			/*
962 			 * If the IO is via the VM then we do some
963 			 * special VM hackery (yuck).  Since the buffer's
964 			 * block size may not be page-aligned it is possible
965 			 * for a page to be shared between two buffers.  We
966 			 * have to get rid of the duplication when building
967 			 * the cluster.
968 			 */
969 			if (tbp->b_flags & B_VMIO) {
970 				vm_page_t m;
971 
972 				if (i != 0) { /* if not first buffer */
973 					for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
974 						m = tbp->b_xio.xio_pages[j];
975 						if (m->flags & PG_BUSY) {
976 							bqrelse(tbp);
977 							goto finishcluster;
978 						}
979 					}
980 				}
981 
982 				for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
983 					m = tbp->b_xio.xio_pages[j];
984 					vm_page_busy_wait(m, FALSE, "clurpg");
985 					vm_page_io_start(m);
986 					vm_page_wakeup(m);
987 					vm_object_pip_add(m->object, 1);
988 					if ((bp->b_xio.xio_npages == 0) ||
989 					  (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
990 						bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
991 						bp->b_xio.xio_npages++;
992 					}
993 				}
994 			}
995 			bp->b_bcount += blksize;
996 			bp->b_bufsize += blksize;
997 
998 			bundirty(tbp);
999 			tbp->b_flags &= ~B_ERROR;
1000 			tbp->b_cmd = BUF_CMD_WRITE;
1001 			BUF_KERNPROC(tbp);
1002 			cluster_append(&bp->b_bio1, tbp);
1003 
1004 			/*
1005 			 * check for latent dependencies to be handled
1006 			 */
1007 			if (LIST_FIRST(&tbp->b_dep) != NULL)
1008 				buf_start(tbp);
1009 		}
1010 	finishcluster:
1011 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1012 			(vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1013 		if (bp->b_bufsize > bp->b_kvasize) {
1014 			panic(
1015 			    "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
1016 			    bp->b_bufsize, bp->b_kvasize);
1017 		}
1018 		totalwritten += bp->b_bufsize;
1019 		bp->b_dirtyoff = 0;
1020 		bp->b_dirtyend = bp->b_bufsize;
1021 		bp->b_bio1.bio_done = cluster_callback;
1022 		bp->b_cmd = BUF_CMD_WRITE;
1023 
1024 		vfs_busy_pages(vp, bp);
1025 		bsetrunningbufspace(bp, bp->b_bufsize);
1026 		BUF_KERNPROC(bp);
1027 		vn_strategy(vp, &bp->b_bio1);
1028 
1029 		bytes -= i;
1030 	}
1031 	return totalwritten;
1032 }
1033 
1034 /*
1035  * Collect together all the buffers in a cluster.
1036  * Plus add one additional buffer.
1037  */
1038 static struct cluster_save *
1039 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
1040 {
1041 	struct cluster_save *buflist;
1042 	struct buf *bp;
1043 	off_t loffset;
1044 	int i, len;
1045 
1046 	len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
1047 	buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1048 			 M_SEGMENT, M_WAITOK);
1049 	buflist->bs_nchildren = 0;
1050 	buflist->bs_children = (struct buf **) (buflist + 1);
1051 	for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
1052 		(void) bread(vp, loffset, last_bp->b_bcount, &bp);
1053 		buflist->bs_children[i] = bp;
1054 		if (bp->b_bio2.bio_offset == NOOFFSET) {
1055 			VOP_BMAP(bp->b_vp, bp->b_loffset,
1056 				 &bp->b_bio2.bio_offset,
1057 				 NULL, NULL, BUF_CMD_WRITE);
1058 		}
1059 	}
1060 	buflist->bs_children[i] = bp = last_bp;
1061 	if (bp->b_bio2.bio_offset == NOOFFSET) {
1062 		VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1063 			 NULL, NULL, BUF_CMD_WRITE);
1064 	}
1065 	buflist->bs_nchildren = i + 1;
1066 	return (buflist);
1067 }
1068 
1069 void
1070 cluster_append(struct bio *bio, struct buf *tbp)
1071 {
1072 	tbp->b_cluster_next = NULL;
1073 	if (bio->bio_caller_info1.cluster_head == NULL) {
1074 		bio->bio_caller_info1.cluster_head = tbp;
1075 		bio->bio_caller_info2.cluster_tail = tbp;
1076 	} else {
1077 		bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1078 		bio->bio_caller_info2.cluster_tail = tbp;
1079 	}
1080 }
1081 
1082 static
1083 void
1084 cluster_setram (struct buf *bp)
1085 {
1086 	bp->b_flags |= B_RAM;
1087 	if (bp->b_xio.xio_npages)
1088 		vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM);
1089 }
1090