xref: /dflybsd-src/sys/kern/vfs_cluster.c (revision 9bbdde3598baabd3206445e589eb185bfed745d2)
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 
57 #include <sys/buf2.h>
58 #include <vm/vm_page2.h>
59 
60 #include <machine/limits.h>
61 
62 #if defined(CLUSTERDEBUG)
63 #include <sys/sysctl.h>
64 static int	rcluster= 0;
65 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
66 #endif
67 
68 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");
69 
70 static struct cluster_save *
71 	cluster_collectbufs (struct vnode *vp, struct buf *last_bp,
72 			    int blksize);
73 static struct buf *
74 	cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
75 			    off_t doffset, int blksize, int run,
76 			    struct buf *fbp);
77 static void cluster_callback (struct bio *);
78 static void cluster_setram (struct buf *);
79 static int cluster_wbuild(struct vnode *vp, struct buf **bpp, int blksize,
80 			    off_t start_loffset, int bytes);
81 
82 static int write_behind = 1;
83 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
84     "Cluster write-behind setting");
85 static int max_readahead = 2 * 1024 * 1024;
86 SYSCTL_INT(_vfs, OID_AUTO, max_readahead, CTLFLAG_RW, &max_readahead, 0,
87     "Limit in bytes for desired cluster read-ahead");
88 
89 extern vm_page_t	bogus_page;
90 
91 extern int cluster_pbuf_freecnt;
92 
93 /*
94  * This replaces bread.
95  *
96  * filesize	- read-ahead @ blksize will not cross this boundary
97  * loffset	- loffset for returned *bpp
98  * blksize	- blocksize for returned *bpp and read-ahead bps
99  * minreq	- minimum (not a hard minimum) in bytes, typically reflects
100  *		  a higher level uio resid.
101  * maxreq	- maximum (sequential heuristic) in bytes (highet typ ~2MB)
102  * bpp		- return buffer (*bpp) for (loffset,blksize)
103  */
104 int
105 cluster_readx(struct vnode *vp, off_t filesize, off_t loffset,
106 	     int blksize, size_t minreq, size_t maxreq, struct buf **bpp)
107 {
108 	struct buf *bp, *rbp, *reqbp;
109 	off_t origoffset;
110 	off_t doffset;
111 	int error;
112 	int i;
113 	int maxra;
114 	int maxrbuild;
115 
116 	error = 0;
117 
118 	/*
119 	 * Calculate the desired read-ahead in blksize'd blocks (maxra).
120 	 * To do this we calculate maxreq.
121 	 *
122 	 * maxreq typically starts out as a sequential heuristic.  If the
123 	 * high level uio/resid is bigger (minreq), we pop maxreq up to
124 	 * minreq.  This represents the case where random I/O is being
125 	 * performed by the userland is issuing big read()'s.
126 	 *
127 	 * Then we limit maxreq to max_readahead to ensure it is a reasonable
128 	 * value.
129 	 *
130 	 * Finally we must ensure that (loffset + maxreq) does not cross the
131 	 * boundary (filesize) for the current blocksize.  If we allowed it
132 	 * to cross we could end up with buffers past the boundary with the
133 	 * wrong block size (HAMMER large-data areas use mixed block sizes).
134 	 * minreq is also absolutely limited to filesize.
135 	 */
136 	if (maxreq < minreq)
137 		maxreq = minreq;
138 	/* minreq not used beyond this point */
139 
140 	if (maxreq > max_readahead) {
141 		maxreq = max_readahead;
142 		if (maxreq > 16 * 1024 * 1024)
143 			maxreq = 16 * 1024 * 1024;
144 	}
145 	if (maxreq < blksize)
146 		maxreq = blksize;
147 	if (loffset + maxreq > filesize) {
148 		if (loffset > filesize)
149 			maxreq = 0;
150 		else
151 			maxreq = filesize - loffset;
152 	}
153 
154 	maxra = (int)(maxreq / blksize);
155 
156 	/*
157 	 * Get the requested block.
158 	 */
159 	if (*bpp)
160 		reqbp = bp = *bpp;
161 	else
162 		*bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0);
163 	origoffset = loffset;
164 
165 	/*
166 	 * Calculate the maximum cluster size for a single I/O, used
167 	 * by cluster_rbuild().
168 	 */
169 	maxrbuild = vmaxiosize(vp) / blksize;
170 
171 	/*
172 	 * if it is in the cache, then check to see if the reads have been
173 	 * sequential.  If they have, then try some read-ahead, otherwise
174 	 * back-off on prospective read-aheads.
175 	 */
176 	if (bp->b_flags & B_CACHE) {
177 		/*
178 		 * Not sequential, do not do any read-ahead
179 		 */
180 		if (maxra <= 1)
181 			return 0;
182 
183 		/*
184 		 * No read-ahead mark, do not do any read-ahead
185 		 * yet.
186 		 */
187 		if ((bp->b_flags & B_RAM) == 0)
188 			return 0;
189 
190 		/*
191 		 * We hit a read-ahead-mark, figure out how much read-ahead
192 		 * to do (maxra) and where to start (loffset).
193 		 *
194 		 * Shortcut the scan.  Typically the way this works is that
195 		 * we've built up all the blocks inbetween except for the
196 		 * last in previous iterations, so if the second-to-last
197 		 * block is present we just skip ahead to it.
198 		 *
199 		 * This algorithm has O(1) cpu in the steady state no
200 		 * matter how large maxra is.
201 		 */
202 		bp->b_flags &= ~B_RAM;
203 
204 		if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST))
205 			i = maxra - 1;
206 		else
207 			i = 1;
208 		while (i < maxra) {
209 			if (findblk(vp, loffset + i * blksize,
210 				    FINDBLK_TEST) == NULL) {
211 				break;
212 			}
213 			++i;
214 		}
215 
216 		/*
217 		 * We got everything or everything is in the cache, no
218 		 * point continuing.
219 		 */
220 		if (i >= maxra)
221 			return 0;
222 
223 		/*
224 		 * Calculate where to start the read-ahead and how much
225 		 * to do.  Generally speaking we want to read-ahead by
226 		 * (maxra) when we've found a read-ahead mark.  We do
227 		 * not want to reduce maxra here as it will cause
228 		 * successive read-ahead I/O's to be smaller and smaller.
229 		 *
230 		 * However, we have to make sure we don't break the
231 		 * filesize limitation for the clustered operation.
232 		 */
233 		loffset += i * blksize;
234 		reqbp = bp = NULL;
235 
236 		if (loffset >= filesize)
237 			return 0;
238 		if (loffset + maxra * blksize > filesize) {
239 			maxreq = filesize - loffset;
240 			maxra = (int)(maxreq / blksize);
241 		}
242 	} else {
243 		__debugvar off_t firstread = bp->b_loffset;
244 		int nblks;
245 
246 		/*
247 		 * Set-up synchronous read for bp.
248 		 */
249 		bp->b_cmd = BUF_CMD_READ;
250 		bp->b_bio1.bio_done = biodone_sync;
251 		bp->b_bio1.bio_flags |= BIO_SYNC;
252 
253 		KASSERT(firstread != NOOFFSET,
254 			("cluster_read: no buffer offset"));
255 
256 		/*
257 		 * nblks is our cluster_rbuild request size, limited
258 		 * primarily by the device.
259 		 */
260 		if ((nblks = maxra) > maxrbuild)
261 			nblks = maxrbuild;
262 
263 		if (nblks > 1) {
264 			int burstbytes;
265 
266 	    		error = VOP_BMAP(vp, loffset, &doffset,
267 					 &burstbytes, NULL, BUF_CMD_READ);
268 			if (error)
269 				goto single_block_read;
270 			if (nblks > burstbytes / blksize)
271 				nblks = burstbytes / blksize;
272 			if (doffset == NOOFFSET)
273 				goto single_block_read;
274 			if (nblks <= 1)
275 				goto single_block_read;
276 
277 			bp = cluster_rbuild(vp, filesize, loffset,
278 					    doffset, blksize, nblks, bp);
279 			loffset += bp->b_bufsize;
280 			maxra -= bp->b_bufsize / blksize;
281 		} else {
282 single_block_read:
283 			/*
284 			 * If it isn't in the cache, then get a chunk from
285 			 * disk if sequential, otherwise just get the block.
286 			 */
287 			cluster_setram(bp);
288 			loffset += blksize;
289 			--maxra;
290 		}
291 	}
292 
293 	/*
294 	 * If B_CACHE was not set issue bp.  bp will either be an
295 	 * asynchronous cluster buf or a synchronous single-buf.
296 	 * If it is a single buf it will be the same as reqbp.
297 	 *
298 	 * NOTE: Once an async cluster buf is issued bp becomes invalid.
299 	 */
300 	if (bp) {
301 #if defined(CLUSTERDEBUG)
302 		if (rcluster)
303 			kprintf("S(%012jx,%d,%d)\n",
304 			    (intmax_t)bp->b_loffset, bp->b_bcount, maxra);
305 #endif
306 		if ((bp->b_flags & B_CLUSTER) == 0)
307 			vfs_busy_pages(vp, bp);
308 		bp->b_flags &= ~(B_ERROR|B_INVAL);
309 		vn_strategy(vp, &bp->b_bio1);
310 		error = 0;
311 		/* bp invalid now */
312 	}
313 
314 	/*
315 	 * If we have been doing sequential I/O, then do some read-ahead.
316 	 * The code above us should have positioned us at the next likely
317 	 * offset.
318 	 *
319 	 * Only mess with buffers which we can immediately lock.  HAMMER
320 	 * will do device-readahead irrespective of what the blocks
321 	 * represent.
322 	 */
323 	while (error == 0 && maxra > 0) {
324 		int burstbytes;
325 		int tmp_error;
326 		int nblks;
327 
328 		rbp = getblk(vp, loffset, blksize,
329 			     GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
330 		if (rbp == NULL)
331 			goto no_read_ahead;
332 		if ((rbp->b_flags & B_CACHE)) {
333 			bqrelse(rbp);
334 			goto no_read_ahead;
335 		}
336 
337 		/*
338 		 * An error from the read-ahead bmap has nothing to do
339 		 * with the caller's original request.
340 		 */
341 		tmp_error = VOP_BMAP(vp, loffset, &doffset,
342 				     &burstbytes, NULL, BUF_CMD_READ);
343 		if (tmp_error || doffset == NOOFFSET) {
344 			rbp->b_flags |= B_INVAL;
345 			brelse(rbp);
346 			rbp = NULL;
347 			goto no_read_ahead;
348 		}
349 		if ((nblks = maxra) > maxrbuild)
350 			nblks = maxrbuild;
351 		if (nblks > burstbytes / blksize)
352 			nblks = burstbytes / blksize;
353 
354 		/*
355 		 * rbp: async read
356 		 */
357 		rbp->b_cmd = BUF_CMD_READ;
358 		/*rbp->b_flags |= B_AGE*/;
359 		cluster_setram(rbp);
360 
361 		if (nblks > 1) {
362 			rbp = cluster_rbuild(vp, filesize, loffset,
363 					     doffset, blksize,
364 					     nblks, rbp);
365 		} else {
366 			rbp->b_bio2.bio_offset = doffset;
367 		}
368 
369 #if defined(CLUSTERDEBUG)
370 		if (rcluster) {
371 			if (bp) {
372 				kprintf("A+(%012jx,%d,%jd) "
373 					"doff=%012jx minr=%zd ra=%d\n",
374 				    (intmax_t)loffset, rbp->b_bcount,
375 				    (intmax_t)(loffset - origoffset),
376 				    (intmax_t)doffset, minreq, maxra);
377 			} else {
378 				kprintf("A-(%012jx,%d,%jd) "
379 					"doff=%012jx minr=%zd ra=%d\n",
380 				    (intmax_t)rbp->b_loffset, rbp->b_bcount,
381 				    (intmax_t)(loffset - origoffset),
382 				    (intmax_t)doffset, minreq, maxra);
383 			}
384 		}
385 #endif
386 		rbp->b_flags &= ~(B_ERROR|B_INVAL);
387 
388 		if ((rbp->b_flags & B_CLUSTER) == 0)
389 			vfs_busy_pages(vp, rbp);
390 		BUF_KERNPROC(rbp);
391 		loffset += rbp->b_bufsize;
392 		maxra -= rbp->b_bufsize / blksize;
393 		vn_strategy(vp, &rbp->b_bio1);
394 		/* rbp invalid now */
395 	}
396 
397 	/*
398 	 * Wait for our original buffer to complete its I/O.  reqbp will
399 	 * be NULL if the original buffer was B_CACHE.  We are returning
400 	 * (*bpp) which is the same as reqbp when reqbp != NULL.
401 	 */
402 no_read_ahead:
403 	if (reqbp) {
404 		KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
405 		error = biowait(&reqbp->b_bio1, "clurd");
406 	}
407 	return (error);
408 }
409 
410 /*
411  * If blocks are contiguous on disk, use this to provide clustered
412  * read ahead.  We will read as many blocks as possible sequentially
413  * and then parcel them up into logical blocks in the buffer hash table.
414  *
415  * This function either returns a cluster buf or it returns fbp.  fbp is
416  * already expected to be set up as a synchronous or asynchronous request.
417  *
418  * If a cluster buf is returned it will always be async.
419  */
420 static struct buf *
421 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
422 	       int blksize, int run, struct buf *fbp)
423 {
424 	struct buf *bp, *tbp;
425 	off_t boffset;
426 	int i, j;
427 	int maxiosize = vmaxiosize(vp);
428 
429 	/*
430 	 * avoid a division
431 	 */
432 	while (loffset + run * blksize > filesize) {
433 		--run;
434 	}
435 
436 	tbp = fbp;
437 	tbp->b_bio2.bio_offset = doffset;
438 	if((tbp->b_flags & B_MALLOC) ||
439 	    ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
440 		return tbp;
441 	}
442 
443 	bp = trypbuf_kva(&cluster_pbuf_freecnt);
444 	if (bp == NULL) {
445 		return tbp;
446 	}
447 
448 	/*
449 	 * We are synthesizing a buffer out of vm_page_t's, but
450 	 * if the block size is not page aligned then the starting
451 	 * address may not be either.  Inherit the b_data offset
452 	 * from the original buffer.
453 	 */
454 	bp->b_data = (char *)((vm_offset_t)bp->b_data |
455 	    ((vm_offset_t)tbp->b_data & PAGE_MASK));
456 	bp->b_flags |= B_CLUSTER | B_VMIO;
457 	bp->b_cmd = BUF_CMD_READ;
458 	bp->b_bio1.bio_done = cluster_callback;		/* default to async */
459 	bp->b_bio1.bio_caller_info1.cluster_head = NULL;
460 	bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
461 	bp->b_loffset = loffset;
462 	bp->b_bio2.bio_offset = doffset;
463 	KASSERT(bp->b_loffset != NOOFFSET,
464 		("cluster_rbuild: no buffer offset"));
465 
466 	bp->b_bcount = 0;
467 	bp->b_bufsize = 0;
468 	bp->b_xio.xio_npages = 0;
469 
470 	for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
471 		if (i) {
472 			if ((bp->b_xio.xio_npages * PAGE_SIZE) +
473 			    round_page(blksize) > maxiosize) {
474 				break;
475 			}
476 
477 			/*
478 			 * Shortcut some checks and try to avoid buffers that
479 			 * would block in the lock.  The same checks have to
480 			 * be made again after we officially get the buffer.
481 			 */
482 			tbp = getblk(vp, loffset + i * blksize, blksize,
483 				     GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
484 			if (tbp == NULL)
485 				break;
486 			for (j = 0; j < tbp->b_xio.xio_npages; j++) {
487 				if (tbp->b_xio.xio_pages[j]->valid)
488 					break;
489 			}
490 			if (j != tbp->b_xio.xio_npages) {
491 				bqrelse(tbp);
492 				break;
493 			}
494 
495 			/*
496 			 * Stop scanning if the buffer is fuly valid
497 			 * (marked B_CACHE), or locked (may be doing a
498 			 * background write), or if the buffer is not
499 			 * VMIO backed.  The clustering code can only deal
500 			 * with VMIO-backed buffers.
501 			 */
502 			if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
503 			    (tbp->b_flags & B_VMIO) == 0 ||
504 			    (LIST_FIRST(&tbp->b_dep) != NULL &&
505 			     buf_checkread(tbp))
506 			) {
507 				bqrelse(tbp);
508 				break;
509 			}
510 
511 			/*
512 			 * The buffer must be completely invalid in order to
513 			 * take part in the cluster.  If it is partially valid
514 			 * then we stop.
515 			 */
516 			for (j = 0;j < tbp->b_xio.xio_npages; j++) {
517 				if (tbp->b_xio.xio_pages[j]->valid)
518 					break;
519 			}
520 			if (j != tbp->b_xio.xio_npages) {
521 				bqrelse(tbp);
522 				break;
523 			}
524 
525 			/*
526 			 * Set a read-ahead mark as appropriate.  Always
527 			 * set the read-ahead mark at (run - 1).  It is
528 			 * unclear why we were also setting it at i == 1.
529 			 */
530 			if (/*i == 1 ||*/ i == (run - 1))
531 				cluster_setram(tbp);
532 
533 			/*
534 			 * Depress the priority of buffers not explicitly
535 			 * requested.
536 			 */
537 			/* tbp->b_flags |= B_AGE; */
538 
539 			/*
540 			 * Set the block number if it isn't set, otherwise
541 			 * if it is make sure it matches the block number we
542 			 * expect.
543 			 */
544 			if (tbp->b_bio2.bio_offset == NOOFFSET) {
545 				tbp->b_bio2.bio_offset = boffset;
546 			} else if (tbp->b_bio2.bio_offset != boffset) {
547 				brelse(tbp);
548 				break;
549 			}
550 		}
551 
552 		/*
553 		 * The passed-in tbp (i == 0) will already be set up for
554 		 * async or sync operation.  All other tbp's acquire in
555 		 * our loop are set up for async operation.
556 		 */
557 		tbp->b_cmd = BUF_CMD_READ;
558 		BUF_KERNPROC(tbp);
559 		cluster_append(&bp->b_bio1, tbp);
560 		for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
561 			vm_page_t m;
562 
563 			m = tbp->b_xio.xio_pages[j];
564 			vm_page_busy_wait(m, FALSE, "clurpg");
565 			vm_page_io_start(m);
566 			vm_page_wakeup(m);
567 			vm_object_pip_add(m->object, 1);
568 			if ((bp->b_xio.xio_npages == 0) ||
569 				(bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
570 				bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
571 				bp->b_xio.xio_npages++;
572 			}
573 			if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
574 				tbp->b_xio.xio_pages[j] = bogus_page;
575 		}
576 		/*
577 		 * XXX shouldn't this be += size for both, like in
578 		 * cluster_wbuild()?
579 		 *
580 		 * Don't inherit tbp->b_bufsize as it may be larger due to
581 		 * a non-page-aligned size.  Instead just aggregate using
582 		 * 'size'.
583 		 */
584 		if (tbp->b_bcount != blksize)
585 		    kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
586 		if (tbp->b_bufsize != blksize)
587 		    kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
588 		bp->b_bcount += blksize;
589 		bp->b_bufsize += blksize;
590 	}
591 
592 	/*
593 	 * Fully valid pages in the cluster are already good and do not need
594 	 * to be re-read from disk.  Replace the page with bogus_page
595 	 */
596 	for (j = 0; j < bp->b_xio.xio_npages; j++) {
597 		if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
598 		    VM_PAGE_BITS_ALL) {
599 			bp->b_xio.xio_pages[j] = bogus_page;
600 		}
601 	}
602 	if (bp->b_bufsize > bp->b_kvasize) {
603 		panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
604 		    bp->b_bufsize, bp->b_kvasize);
605 	}
606 	pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
607 		(vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
608 	BUF_KERNPROC(bp);
609 	return (bp);
610 }
611 
612 /*
613  * Cleanup after a clustered read or write.
614  * This is complicated by the fact that any of the buffers might have
615  * extra memory (if there were no empty buffer headers at allocbuf time)
616  * that we will need to shift around.
617  *
618  * The returned bio is &bp->b_bio1
619  */
620 void
621 cluster_callback(struct bio *bio)
622 {
623 	struct buf *bp = bio->bio_buf;
624 	struct buf *tbp;
625 	int error = 0;
626 
627 	/*
628 	 * Must propogate errors to all the components.  A short read (EOF)
629 	 * is a critical error.
630 	 */
631 	if (bp->b_flags & B_ERROR) {
632 		error = bp->b_error;
633 	} else if (bp->b_bcount != bp->b_bufsize) {
634 		panic("cluster_callback: unexpected EOF on cluster %p!", bio);
635 	}
636 
637 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
638 	/*
639 	 * Move memory from the large cluster buffer into the component
640 	 * buffers and mark IO as done on these.  Since the memory map
641 	 * is the same, no actual copying is required.
642 	 */
643 	while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
644 		bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
645 		if (error) {
646 			tbp->b_flags |= B_ERROR | B_IODEBUG;
647 			tbp->b_error = error;
648 		} else {
649 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
650 			tbp->b_flags &= ~(B_ERROR|B_INVAL);
651 			tbp->b_flags |= B_IODEBUG;
652 			/*
653 			 * XXX the bdwrite()/bqrelse() issued during
654 			 * cluster building clears B_RELBUF (see bqrelse()
655 			 * comment).  If direct I/O was specified, we have
656 			 * to restore it here to allow the buffer and VM
657 			 * to be freed.
658 			 */
659 			if (tbp->b_flags & B_DIRECT)
660 				tbp->b_flags |= B_RELBUF;
661 		}
662 		biodone(&tbp->b_bio1);
663 	}
664 	relpbuf(bp, &cluster_pbuf_freecnt);
665 }
666 
667 /*
668  *	cluster_wbuild_wb:
669  *
670  *	Implement modified write build for cluster.
671  *
672  *		write_behind = 0	write behind disabled
673  *		write_behind = 1	write behind normal (default)
674  *		write_behind = 2	write behind backed-off
675  */
676 
677 static __inline int
678 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
679 {
680 	int r = 0;
681 
682 	switch(write_behind) {
683 	case 2:
684 		if (start_loffset < len)
685 			break;
686 		start_loffset -= len;
687 		/* fall through */
688 	case 1:
689 		r = cluster_wbuild(vp, NULL, blksize, start_loffset, len);
690 		/* fall through */
691 	default:
692 		/* fall through */
693 		break;
694 	}
695 	return(r);
696 }
697 
698 /*
699  * Do clustered write for FFS.
700  *
701  * Three cases:
702  *	1. Write is not sequential (write asynchronously)
703  *	Write is sequential:
704  *	2.	beginning of cluster - begin cluster
705  *	3.	middle of a cluster - add to cluster
706  *	4.	end of a cluster - asynchronously write cluster
707  */
708 void
709 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
710 {
711 	struct vnode *vp;
712 	off_t loffset;
713 	int maxclen, cursize;
714 	int async;
715 
716 	vp = bp->b_vp;
717 	if (vp->v_type == VREG)
718 		async = vp->v_mount->mnt_flag & MNT_ASYNC;
719 	else
720 		async = 0;
721 	loffset = bp->b_loffset;
722 	KASSERT(bp->b_loffset != NOOFFSET,
723 		("cluster_write: no buffer offset"));
724 
725 	/* Initialize vnode to beginning of file. */
726 	if (loffset == 0)
727 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
728 
729 	if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
730 	    bp->b_bio2.bio_offset == NOOFFSET ||
731 	    (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
732 		maxclen = vmaxiosize(vp);
733 		if (vp->v_clen != 0) {
734 			/*
735 			 * Next block is not sequential.
736 			 *
737 			 * If we are not writing at end of file, the process
738 			 * seeked to another point in the file since its last
739 			 * write, or we have reached our maximum cluster size,
740 			 * then push the previous cluster. Otherwise try
741 			 * reallocating to make it sequential.
742 			 *
743 			 * Change to algorithm: only push previous cluster if
744 			 * it was sequential from the point of view of the
745 			 * seqcount heuristic, otherwise leave the buffer
746 			 * intact so we can potentially optimize the I/O
747 			 * later on in the buf_daemon or update daemon
748 			 * flush.
749 			 */
750 			cursize = vp->v_lastw - vp->v_cstart + blksize;
751 			if (bp->b_loffset + blksize < filesize ||
752 			    loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
753 				if (!async && seqcount > 0) {
754 					cluster_wbuild_wb(vp, blksize,
755 						vp->v_cstart, cursize);
756 				}
757 			} else {
758 				struct buf **bpp, **endbp;
759 				struct cluster_save *buflist;
760 
761 				buflist = cluster_collectbufs(vp, bp, blksize);
762 				endbp = &buflist->bs_children
763 				    [buflist->bs_nchildren - 1];
764 				if (VOP_REALLOCBLKS(vp, buflist)) {
765 					/*
766 					 * Failed, push the previous cluster
767 					 * if *really* writing sequentially
768 					 * in the logical file (seqcount > 1),
769 					 * otherwise delay it in the hopes that
770 					 * the low level disk driver can
771 					 * optimize the write ordering.
772 					 */
773 					for (bpp = buflist->bs_children;
774 					     bpp < endbp; bpp++)
775 						brelse(*bpp);
776 					kfree(buflist, M_SEGMENT);
777 					if (seqcount > 1) {
778 						cluster_wbuild_wb(vp,
779 						    blksize, vp->v_cstart,
780 						    cursize);
781 					}
782 				} else {
783 					/*
784 					 * Succeeded, keep building cluster.
785 					 */
786 					for (bpp = buflist->bs_children;
787 					     bpp <= endbp; bpp++)
788 						bdwrite(*bpp);
789 					kfree(buflist, M_SEGMENT);
790 					vp->v_lastw = loffset;
791 					vp->v_lasta = bp->b_bio2.bio_offset;
792 					return;
793 				}
794 			}
795 		}
796 		/*
797 		 * Consider beginning a cluster. If at end of file, make
798 		 * cluster as large as possible, otherwise find size of
799 		 * existing cluster.
800 		 */
801 		if ((vp->v_type == VREG) &&
802 		    bp->b_loffset + blksize < filesize &&
803 		    (bp->b_bio2.bio_offset == NOOFFSET) &&
804 		    (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
805 		     bp->b_bio2.bio_offset == NOOFFSET)) {
806 			bawrite(bp);
807 			vp->v_clen = 0;
808 			vp->v_lasta = bp->b_bio2.bio_offset;
809 			vp->v_cstart = loffset + blksize;
810 			vp->v_lastw = loffset;
811 			return;
812 		}
813 		if (maxclen > blksize)
814 			vp->v_clen = maxclen - blksize;
815 		else
816 			vp->v_clen = 0;
817 		if (!async && vp->v_clen == 0) { /* I/O not contiguous */
818 			vp->v_cstart = loffset + blksize;
819 			bawrite(bp);
820 		} else {	/* Wait for rest of cluster */
821 			vp->v_cstart = loffset;
822 			bdwrite(bp);
823 		}
824 	} else if (loffset == vp->v_cstart + vp->v_clen) {
825 		/*
826 		 * At end of cluster, write it out if seqcount tells us we
827 		 * are operating sequentially, otherwise let the buf or
828 		 * update daemon handle it.
829 		 */
830 		bdwrite(bp);
831 		if (seqcount > 1)
832 			cluster_wbuild_wb(vp, blksize, vp->v_cstart,
833 					  vp->v_clen + blksize);
834 		vp->v_clen = 0;
835 		vp->v_cstart = loffset + blksize;
836 	} else if (vm_page_count_severe()) {
837 		/*
838 		 * We are low on memory, get it going NOW
839 		 */
840 		bawrite(bp);
841 	} else {
842 		/*
843 		 * In the middle of a cluster, so just delay the I/O for now.
844 		 */
845 		bdwrite(bp);
846 	}
847 	vp->v_lastw = loffset;
848 	vp->v_lasta = bp->b_bio2.bio_offset;
849 }
850 
851 /*
852  * This is the clustered version of bawrite().  It works similarly to
853  * cluster_write() except I/O on the buffer is guaranteed to occur.
854  */
855 int
856 cluster_awrite(struct buf *bp)
857 {
858 	int total;
859 
860 	/*
861 	 * Don't bother if it isn't clusterable.
862 	 */
863 	if ((bp->b_flags & B_CLUSTEROK) == 0 ||
864 	    bp->b_vp == NULL ||
865 	    (bp->b_vp->v_flag & VOBJBUF) == 0) {
866 		total = bp->b_bufsize;
867 		bawrite(bp);
868 		return (total);
869 	}
870 
871 	total = cluster_wbuild(bp->b_vp, &bp, bp->b_bufsize,
872 			       bp->b_loffset, vmaxiosize(bp->b_vp));
873 	if (bp)
874 		bawrite(bp);
875 
876 	return total;
877 }
878 
879 /*
880  * This is an awful lot like cluster_rbuild...wish they could be combined.
881  * The last lbn argument is the current block on which I/O is being
882  * performed.  Check to see that it doesn't fall in the middle of
883  * the current block (if last_bp == NULL).
884  *
885  * cluster_wbuild() normally does not guarantee anything.  If bpp is
886  * non-NULL and cluster_wbuild() is able to incorporate it into the
887  * I/O it will set *bpp to NULL, otherwise it will leave it alone and
888  * the caller must dispose of *bpp.
889  */
890 static int
891 cluster_wbuild(struct vnode *vp, struct buf **bpp,
892 	       int blksize, off_t start_loffset, int bytes)
893 {
894 	struct buf *bp, *tbp;
895 	int i, j;
896 	int totalwritten = 0;
897 	int must_initiate;
898 	int maxiosize = vmaxiosize(vp);
899 
900 	while (bytes > 0) {
901 		/*
902 		 * If the buffer matches the passed locked & removed buffer
903 		 * we used the passed buffer (which might not be B_DELWRI).
904 		 *
905 		 * Otherwise locate the buffer and determine if it is
906 		 * compatible.
907 		 */
908 		if (bpp && (*bpp)->b_loffset == start_loffset) {
909 			tbp = *bpp;
910 			*bpp = NULL;
911 			bpp = NULL;
912 		} else {
913 			tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK);
914 			if (tbp == NULL ||
915 			    (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) !=
916 			     B_DELWRI ||
917 			    (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
918 				if (tbp)
919 					BUF_UNLOCK(tbp);
920 				start_loffset += blksize;
921 				bytes -= blksize;
922 				continue;
923 			}
924 			bremfree(tbp);
925 		}
926 		KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
927 
928 		/*
929 		 * Extra memory in the buffer, punt on this buffer.
930 		 * XXX we could handle this in most cases, but we would
931 		 * have to push the extra memory down to after our max
932 		 * possible cluster size and then potentially pull it back
933 		 * up if the cluster was terminated prematurely--too much
934 		 * hassle.
935 		 */
936 		if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
937 		    (tbp->b_bcount != tbp->b_bufsize) ||
938 		    (tbp->b_bcount != blksize) ||
939 		    (bytes == blksize) ||
940 		    ((bp = getpbuf_kva(&cluster_pbuf_freecnt)) == NULL)) {
941 			totalwritten += tbp->b_bufsize;
942 			bawrite(tbp);
943 			start_loffset += blksize;
944 			bytes -= blksize;
945 			continue;
946 		}
947 
948 		/*
949 		 * Set up the pbuf.  Track our append point with b_bcount
950 		 * and b_bufsize.  b_bufsize is not used by the device but
951 		 * our caller uses it to loop clusters and we use it to
952 		 * detect a premature EOF on the block device.
953 		 */
954 		bp->b_bcount = 0;
955 		bp->b_bufsize = 0;
956 		bp->b_xio.xio_npages = 0;
957 		bp->b_loffset = tbp->b_loffset;
958 		bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
959 
960 		/*
961 		 * We are synthesizing a buffer out of vm_page_t's, but
962 		 * if the block size is not page aligned then the starting
963 		 * address may not be either.  Inherit the b_data offset
964 		 * from the original buffer.
965 		 */
966 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
967 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
968 		bp->b_flags &= ~B_ERROR;
969 		bp->b_flags |= B_CLUSTER | B_BNOCLIP |
970 			(tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
971 		bp->b_bio1.bio_caller_info1.cluster_head = NULL;
972 		bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
973 
974 		/*
975 		 * From this location in the file, scan forward to see
976 		 * if there are buffers with adjacent data that need to
977 		 * be written as well.
978 		 *
979 		 * IO *must* be initiated on index 0 at this point
980 		 * (particularly when called from cluster_awrite()).
981 		 */
982 		for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
983 			if (i == 0) {
984 				must_initiate = 1;
985 			} else {
986 				/*
987 				 * Not first buffer.
988 				 */
989 				must_initiate = 0;
990 				tbp = findblk(vp, start_loffset,
991 					      FINDBLK_NBLOCK);
992 				/*
993 				 * Buffer not found or could not be locked
994 				 * non-blocking.
995 				 */
996 				if (tbp == NULL)
997 					break;
998 
999 				/*
1000 				 * If it IS in core, but has different
1001 				 * characteristics, then don't cluster
1002 				 * with it.
1003 				 */
1004 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
1005 				     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
1006 				    != (B_DELWRI | B_CLUSTEROK |
1007 				     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
1008 				    (tbp->b_flags & B_LOCKED)
1009 				) {
1010 					BUF_UNLOCK(tbp);
1011 					break;
1012 				}
1013 
1014 				/*
1015 				 * Check that the combined cluster
1016 				 * would make sense with regard to pages
1017 				 * and would not be too large
1018 				 *
1019 				 * WARNING! buf_checkwrite() must be the last
1020 				 *	    check made.  If it returns 0 then
1021 				 *	    we must initiate the I/O.
1022 				 */
1023 				if ((tbp->b_bcount != blksize) ||
1024 				  ((bp->b_bio2.bio_offset + i) !=
1025 				    tbp->b_bio2.bio_offset) ||
1026 				  ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
1027 				    (maxiosize / PAGE_SIZE)) ||
1028 				  (LIST_FIRST(&tbp->b_dep) &&
1029 				   buf_checkwrite(tbp))
1030 				) {
1031 					BUF_UNLOCK(tbp);
1032 					break;
1033 				}
1034 				if (LIST_FIRST(&tbp->b_dep))
1035 					must_initiate = 1;
1036 				/*
1037 				 * Ok, it's passed all the tests,
1038 				 * so remove it from the free list
1039 				 * and mark it busy. We will use it.
1040 				 */
1041 				bremfree(tbp);
1042 				KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
1043 			}
1044 
1045 			/*
1046 			 * If the IO is via the VM then we do some
1047 			 * special VM hackery (yuck).  Since the buffer's
1048 			 * block size may not be page-aligned it is possible
1049 			 * for a page to be shared between two buffers.  We
1050 			 * have to get rid of the duplication when building
1051 			 * the cluster.
1052 			 */
1053 			if (tbp->b_flags & B_VMIO) {
1054 				vm_page_t m;
1055 
1056 				/*
1057 				 * Try to avoid deadlocks with the VM system.
1058 				 * However, we cannot abort the I/O if
1059 				 * must_initiate is non-zero.
1060 				 */
1061 				if (must_initiate == 0) {
1062 					for (j = 0;
1063 					     j < tbp->b_xio.xio_npages;
1064 					     ++j) {
1065 						m = tbp->b_xio.xio_pages[j];
1066 						if (m->flags & PG_BUSY) {
1067 							bqrelse(tbp);
1068 							goto finishcluster;
1069 						}
1070 					}
1071 				}
1072 
1073 				for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
1074 					m = tbp->b_xio.xio_pages[j];
1075 					vm_page_busy_wait(m, FALSE, "clurpg");
1076 					vm_page_io_start(m);
1077 					vm_page_wakeup(m);
1078 					vm_object_pip_add(m->object, 1);
1079 					if ((bp->b_xio.xio_npages == 0) ||
1080 					  (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
1081 						bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
1082 						bp->b_xio.xio_npages++;
1083 					}
1084 				}
1085 			}
1086 			bp->b_bcount += blksize;
1087 			bp->b_bufsize += blksize;
1088 
1089 			bundirty(tbp);
1090 			tbp->b_flags &= ~B_ERROR;
1091 			tbp->b_cmd = BUF_CMD_WRITE;
1092 			BUF_KERNPROC(tbp);
1093 			cluster_append(&bp->b_bio1, tbp);
1094 
1095 			/*
1096 			 * check for latent dependencies to be handled
1097 			 */
1098 			if (LIST_FIRST(&tbp->b_dep) != NULL)
1099 				buf_start(tbp);
1100 		}
1101 	finishcluster:
1102 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1103 			    (vm_page_t *)bp->b_xio.xio_pages,
1104 			    bp->b_xio.xio_npages);
1105 		if (bp->b_bufsize > bp->b_kvasize) {
1106 			panic("cluster_wbuild: b_bufsize(%d) "
1107 			      "> b_kvasize(%d)\n",
1108 			      bp->b_bufsize, bp->b_kvasize);
1109 		}
1110 		totalwritten += bp->b_bufsize;
1111 		bp->b_dirtyoff = 0;
1112 		bp->b_dirtyend = bp->b_bufsize;
1113 		bp->b_bio1.bio_done = cluster_callback;
1114 		bp->b_cmd = BUF_CMD_WRITE;
1115 
1116 		vfs_busy_pages(vp, bp);
1117 		bsetrunningbufspace(bp, bp->b_bufsize);
1118 		BUF_KERNPROC(bp);
1119 		vn_strategy(vp, &bp->b_bio1);
1120 
1121 		bytes -= i;
1122 	}
1123 	return totalwritten;
1124 }
1125 
1126 /*
1127  * Collect together all the buffers in a cluster.
1128  * Plus add one additional buffer.
1129  */
1130 static struct cluster_save *
1131 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
1132 {
1133 	struct cluster_save *buflist;
1134 	struct buf *bp;
1135 	off_t loffset;
1136 	int i, len;
1137 
1138 	len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
1139 	buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1140 			 M_SEGMENT, M_WAITOK);
1141 	buflist->bs_nchildren = 0;
1142 	buflist->bs_children = (struct buf **) (buflist + 1);
1143 	for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
1144 		(void) bread(vp, loffset, last_bp->b_bcount, &bp);
1145 		buflist->bs_children[i] = bp;
1146 		if (bp->b_bio2.bio_offset == NOOFFSET) {
1147 			VOP_BMAP(bp->b_vp, bp->b_loffset,
1148 				 &bp->b_bio2.bio_offset,
1149 				 NULL, NULL, BUF_CMD_WRITE);
1150 		}
1151 	}
1152 	buflist->bs_children[i] = bp = last_bp;
1153 	if (bp->b_bio2.bio_offset == NOOFFSET) {
1154 		VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1155 			 NULL, NULL, BUF_CMD_WRITE);
1156 	}
1157 	buflist->bs_nchildren = i + 1;
1158 	return (buflist);
1159 }
1160 
1161 void
1162 cluster_append(struct bio *bio, struct buf *tbp)
1163 {
1164 	tbp->b_cluster_next = NULL;
1165 	if (bio->bio_caller_info1.cluster_head == NULL) {
1166 		bio->bio_caller_info1.cluster_head = tbp;
1167 		bio->bio_caller_info2.cluster_tail = tbp;
1168 	} else {
1169 		bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1170 		bio->bio_caller_info2.cluster_tail = tbp;
1171 	}
1172 }
1173 
1174 static
1175 void
1176 cluster_setram (struct buf *bp)
1177 {
1178 	bp->b_flags |= B_RAM;
1179 	if (bp->b_xio.xio_npages)
1180 		vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM);
1181 }
1182