xref: /freebsd/sys/kern/vfs_cluster.c (revision 22cf89c938886d14f5796fc49f9f020c23ea8eaf)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * Modifications/enhancements:
7  * 	Copyright (c) 1995 John S. Dyson.  All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	@(#)vfs_cluster.c	8.7 (Berkeley) 2/13/94
34  */
35 
36 #include <sys/cdefs.h>
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/proc.h>
41 #include <sys/bio.h>
42 #include <sys/buf.h>
43 #include <sys/vnode.h>
44 #include <sys/malloc.h>
45 #include <sys/mount.h>
46 #include <sys/racct.h>
47 #include <sys/resourcevar.h>
48 #include <sys/rwlock.h>
49 #include <sys/vmmeter.h>
50 #include <vm/vm.h>
51 #include <vm/vm_object.h>
52 #include <vm/vm_page.h>
53 #include <sys/sysctl.h>
54 
55 static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer");
56 static uma_zone_t cluster_pbuf_zone;
57 
58 static void cluster_init(void *);
59 static struct cluster_save *cluster_collectbufs(struct vnode *vp,
60 	    struct vn_clusterw *vnc, struct buf *last_bp, int gbflags);
61 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize,
62 	    daddr_t lbn, daddr_t blkno, long size, int run, int gbflags,
63 	    struct buf *fbp);
64 static void cluster_callback(struct buf *);
65 
66 static int write_behind = 1;
67 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
68     "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
69 
70 static int read_max = 64;
71 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
72     "Cluster read-ahead max block count");
73 
74 static int read_min = 1;
75 SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0,
76     "Cluster read min block count");
77 
78 SYSINIT(cluster, SI_SUB_CPU, SI_ORDER_ANY, cluster_init, NULL);
79 
80 static void
81 cluster_init(void *dummy)
82 {
83 
84 	cluster_pbuf_zone = pbuf_zsecond_create("clpbuf", nswbuf / 2);
85 }
86 
87 /*
88  * Read data to a buf, including read-ahead if we find this to be beneficial.
89  * cluster_read replaces bread.
90  */
91 int
92 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
93     struct ucred *cred, long totread, int seqcount, int gbflags,
94     struct buf **bpp)
95 {
96 	struct buf *bp, *rbp, *reqbp;
97 	struct bufobj *bo;
98 	struct thread *td;
99 	daddr_t blkno, origblkno;
100 	int maxra, racluster;
101 	int error, ncontig;
102 	int i;
103 
104 	error = 0;
105 	td = curthread;
106 	bo = &vp->v_bufobj;
107 	if (!unmapped_buf_allowed)
108 		gbflags &= ~GB_UNMAPPED;
109 
110 	/*
111 	 * Try to limit the amount of read-ahead by a few
112 	 * ad-hoc parameters.  This needs work!!!
113 	 */
114 	racluster = vp->v_mount->mnt_iosize_max / size;
115 	maxra = seqcount;
116 	maxra = min(read_max, maxra);
117 	maxra = min(nbuf/8, maxra);
118 	if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
119 		maxra = (filesize / size) - lblkno;
120 
121 	/*
122 	 * get the requested block
123 	 */
124 	error = getblkx(vp, lblkno, lblkno, size, 0, 0, gbflags, &bp);
125 	if (error != 0) {
126 		*bpp = NULL;
127 		return (error);
128 	}
129 	gbflags &= ~GB_NOSPARSE;
130 	origblkno = lblkno;
131 	*bpp = reqbp = bp;
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 		if (!seqcount) {
140 			return 0;
141 		} else if ((bp->b_flags & B_RAM) == 0) {
142 			return 0;
143 		} else {
144 			bp->b_flags &= ~B_RAM;
145 			BO_RLOCK(bo);
146 			for (i = 1; i < maxra; i++) {
147 				/*
148 				 * Stop if the buffer does not exist or it
149 				 * is invalid (about to go away?)
150 				 */
151 				rbp = gbincore(&vp->v_bufobj, lblkno+i);
152 				if (rbp == NULL || (rbp->b_flags & B_INVAL))
153 					break;
154 
155 				/*
156 				 * Set another read-ahead mark so we know
157 				 * to check again. (If we can lock the
158 				 * buffer without waiting)
159 				 */
160 				if ((((i % racluster) == (racluster - 1)) ||
161 				    (i == (maxra - 1)))
162 				    && (0 == BUF_LOCK(rbp,
163 					LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
164 					rbp->b_flags |= B_RAM;
165 					BUF_UNLOCK(rbp);
166 				}
167 			}
168 			BO_RUNLOCK(bo);
169 			if (i >= maxra) {
170 				return 0;
171 			}
172 			lblkno += i;
173 		}
174 		reqbp = bp = NULL;
175 	/*
176 	 * If it isn't in the cache, then get a chunk from
177 	 * disk if sequential, otherwise just get the block.
178 	 */
179 	} else {
180 		off_t firstread = bp->b_offset;
181 		int nblks;
182 		long minread;
183 
184 		KASSERT(bp->b_offset != NOOFFSET,
185 		    ("cluster_read: no buffer offset"));
186 
187 		ncontig = 0;
188 
189 		/*
190 		 * Adjust totread if needed
191 		 */
192 		minread = read_min * size;
193 		if (minread > totread)
194 			totread = minread;
195 
196 		/*
197 		 * Compute the total number of blocks that we should read
198 		 * synchronously.
199 		 */
200 		if (firstread + totread > filesize)
201 			totread = filesize - firstread;
202 		nblks = howmany(totread, size);
203 		if (nblks > racluster)
204 			nblks = racluster;
205 
206 		/*
207 		 * Now compute the number of contiguous blocks.
208 		 */
209 		if (nblks > 1) {
210 	    		error = VOP_BMAP(vp, lblkno, NULL,
211 				&blkno, &ncontig, NULL);
212 			/*
213 			 * If this failed to map just do the original block.
214 			 */
215 			if (error || blkno == -1)
216 				ncontig = 0;
217 		}
218 
219 		/*
220 		 * If we have contiguous data available do a cluster
221 		 * otherwise just read the requested block.
222 		 */
223 		if (ncontig) {
224 			/* Account for our first block. */
225 			ncontig = min(ncontig + 1, nblks);
226 			if (ncontig < nblks)
227 				nblks = ncontig;
228 			bp = cluster_rbuild(vp, filesize, lblkno,
229 			    blkno, size, nblks, gbflags, bp);
230 			lblkno += (bp->b_bufsize / size);
231 		} else {
232 			bp->b_flags |= B_RAM;
233 			bp->b_iocmd = BIO_READ;
234 			lblkno += 1;
235 		}
236 	}
237 
238 	/*
239 	 * handle the synchronous read so that it is available ASAP.
240 	 */
241 	if (bp) {
242 		if ((bp->b_flags & B_CLUSTER) == 0) {
243 			vfs_busy_pages(bp, 0);
244 		}
245 		bp->b_flags &= ~B_INVAL;
246 		bp->b_ioflags &= ~BIO_ERROR;
247 		if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
248 			BUF_KERNPROC(bp);
249 		bp->b_iooffset = dbtob(bp->b_blkno);
250 		bstrategy(bp);
251 #ifdef RACCT
252 		if (racct_enable) {
253 			PROC_LOCK(td->td_proc);
254 			racct_add_buf(td->td_proc, bp, 0);
255 			PROC_UNLOCK(td->td_proc);
256 		}
257 #endif /* RACCT */
258 		td->td_ru.ru_inblock++;
259 	}
260 
261 	/*
262 	 * If we have been doing sequential I/O, then do some read-ahead.
263 	 */
264 	while (lblkno < (origblkno + maxra)) {
265 		error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
266 		if (error)
267 			break;
268 
269 		if (blkno == -1)
270 			break;
271 
272 		/*
273 		 * We could throttle ncontig here by maxra but we might as
274 		 * well read the data if it is contiguous.  We're throttled
275 		 * by racluster anyway.
276 		 */
277 		if (ncontig) {
278 			ncontig = min(ncontig + 1, racluster);
279 			rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
280 			    size, ncontig, gbflags, NULL);
281 			lblkno += (rbp->b_bufsize / size);
282 			if (rbp->b_flags & B_DELWRI) {
283 				bqrelse(rbp);
284 				continue;
285 			}
286 		} else {
287 			rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
288 			lblkno += 1;
289 			if (rbp->b_flags & B_DELWRI) {
290 				bqrelse(rbp);
291 				continue;
292 			}
293 			rbp->b_flags |= B_ASYNC | B_RAM;
294 			rbp->b_iocmd = BIO_READ;
295 			rbp->b_blkno = blkno;
296 		}
297 		if (rbp->b_flags & B_CACHE) {
298 			rbp->b_flags &= ~B_ASYNC;
299 			bqrelse(rbp);
300 			continue;
301 		}
302 		if ((rbp->b_flags & B_CLUSTER) == 0) {
303 			vfs_busy_pages(rbp, 0);
304 		}
305 		rbp->b_flags &= ~B_INVAL;
306 		rbp->b_ioflags &= ~BIO_ERROR;
307 		if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
308 			BUF_KERNPROC(rbp);
309 		rbp->b_iooffset = dbtob(rbp->b_blkno);
310 		bstrategy(rbp);
311 #ifdef RACCT
312 		if (racct_enable) {
313 			PROC_LOCK(td->td_proc);
314 			racct_add_buf(td->td_proc, rbp, 0);
315 			PROC_UNLOCK(td->td_proc);
316 		}
317 #endif /* RACCT */
318 		td->td_ru.ru_inblock++;
319 	}
320 
321 	if (reqbp) {
322 		/*
323 		 * Like bread, always brelse() the buffer when
324 		 * returning an error.
325 		 */
326 		error = bufwait(reqbp);
327 		if (error != 0) {
328 			brelse(reqbp);
329 			*bpp = NULL;
330 		}
331 	}
332 	return (error);
333 }
334 
335 /*
336  * If blocks are contiguous on disk, use this to provide clustered
337  * read ahead.  We will read as many blocks as possible sequentially
338  * and then parcel them up into logical blocks in the buffer hash table.
339  */
340 static struct buf *
341 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
342     daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
343 {
344 	struct buf *bp, *tbp;
345 	daddr_t bn;
346 	off_t off;
347 	long tinc, tsize;
348 	int i, inc, j, k, toff;
349 
350 	KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
351 	    ("cluster_rbuild: size %ld != f_iosize %jd\n",
352 	    size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
353 
354 	/*
355 	 * avoid a division
356 	 */
357 	while ((u_quad_t) size * (lbn + run) > filesize) {
358 		--run;
359 	}
360 
361 	if (fbp) {
362 		tbp = fbp;
363 		tbp->b_iocmd = BIO_READ;
364 	} else {
365 		tbp = getblk(vp, lbn, size, 0, 0, gbflags);
366 		if (tbp->b_flags & B_CACHE)
367 			return tbp;
368 		tbp->b_flags |= B_ASYNC | B_RAM;
369 		tbp->b_iocmd = BIO_READ;
370 	}
371 	tbp->b_blkno = blkno;
372 	if ( (tbp->b_flags & B_MALLOC) ||
373 		((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
374 		return tbp;
375 
376 	bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT);
377 	if (bp == NULL)
378 		return tbp;
379 	MPASS((bp->b_flags & B_MAXPHYS) != 0);
380 
381 	/*
382 	 * We are synthesizing a buffer out of vm_page_t's, but
383 	 * if the block size is not page aligned then the starting
384 	 * address may not be either.  Inherit the b_data offset
385 	 * from the original buffer.
386 	 */
387 	bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
388 	if ((gbflags & GB_UNMAPPED) != 0) {
389 		bp->b_data = unmapped_buf;
390 	} else {
391 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
392 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
393 	}
394 	bp->b_iocmd = BIO_READ;
395 	bp->b_iodone = cluster_callback;
396 	bp->b_blkno = blkno;
397 	bp->b_lblkno = lbn;
398 	bp->b_offset = tbp->b_offset;
399 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
400 	pbgetvp(vp, bp);
401 
402 	TAILQ_INIT(&bp->b_cluster.cluster_head);
403 
404 	bp->b_bcount = 0;
405 	bp->b_bufsize = 0;
406 	bp->b_npages = 0;
407 
408 	inc = btodb(size);
409 	for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
410 		if (i == 0) {
411 			vm_object_pip_add(tbp->b_bufobj->bo_object,
412 			    tbp->b_npages);
413 			vfs_busy_pages_acquire(tbp);
414 		} else {
415 			if ((bp->b_npages * PAGE_SIZE) +
416 			    round_page(size) > vp->v_mount->mnt_iosize_max) {
417 				break;
418 			}
419 
420 			tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
421 			    (gbflags & GB_UNMAPPED));
422 
423 			/* Don't wait around for locked bufs. */
424 			if (tbp == NULL)
425 				break;
426 
427 			/*
428 			 * Stop scanning if the buffer is fully valid
429 			 * (marked B_CACHE), or locked (may be doing a
430 			 * background write), or if the buffer is not
431 			 * VMIO backed.  The clustering code can only deal
432 			 * with VMIO-backed buffers.  The bo lock is not
433 			 * required for the BKGRDINPROG check since it
434 			 * can not be set without the buf lock.
435 			 */
436 			if ((tbp->b_vflags & BV_BKGRDINPROG) ||
437 			    (tbp->b_flags & B_CACHE) ||
438 			    (tbp->b_flags & B_VMIO) == 0) {
439 				bqrelse(tbp);
440 				break;
441 			}
442 
443 			/*
444 			 * The buffer must be completely invalid in order to
445 			 * take part in the cluster.  If it is partially valid
446 			 * then we stop.
447 			 */
448 			off = tbp->b_offset;
449 			tsize = size;
450 			for (j = 0; tsize > 0; j++) {
451 				toff = off & PAGE_MASK;
452 				tinc = tsize;
453 				if (toff + tinc > PAGE_SIZE)
454 					tinc = PAGE_SIZE - toff;
455 				if (vm_page_trysbusy(tbp->b_pages[j]) == 0)
456 					break;
457 				if ((tbp->b_pages[j]->valid &
458 				    vm_page_bits(toff, tinc)) != 0) {
459 					vm_page_sunbusy(tbp->b_pages[j]);
460 					break;
461 				}
462 				vm_object_pip_add(tbp->b_bufobj->bo_object, 1);
463 				off += tinc;
464 				tsize -= tinc;
465 			}
466 			if (tsize > 0) {
467 clean_sbusy:
468 				vm_object_pip_wakeupn(tbp->b_bufobj->bo_object,
469 				    j);
470 				for (k = 0; k < j; k++)
471 					vm_page_sunbusy(tbp->b_pages[k]);
472 				bqrelse(tbp);
473 				break;
474 			}
475 
476 			/*
477 			 * Set a read-ahead mark as appropriate
478 			 */
479 			if ((fbp && (i == 1)) || (i == (run - 1)))
480 				tbp->b_flags |= B_RAM;
481 
482 			/*
483 			 * Set the buffer up for an async read (XXX should
484 			 * we do this only if we do not wind up brelse()ing?).
485 			 * Set the block number if it isn't set, otherwise
486 			 * if it is make sure it matches the block number we
487 			 * expect.
488 			 */
489 			tbp->b_flags |= B_ASYNC;
490 			tbp->b_iocmd = BIO_READ;
491 			if (tbp->b_blkno == tbp->b_lblkno) {
492 				tbp->b_blkno = bn;
493 			} else if (tbp->b_blkno != bn) {
494 				goto clean_sbusy;
495 			}
496 		}
497 		/*
498 		 * XXX fbp from caller may not be B_ASYNC, but we are going
499 		 * to biodone() it in cluster_callback() anyway
500 		 */
501 		BUF_KERNPROC(tbp);
502 		TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
503 			tbp, b_cluster.cluster_entry);
504 		for (j = 0; j < tbp->b_npages; j += 1) {
505 			vm_page_t m;
506 
507 			m = tbp->b_pages[j];
508 			if ((bp->b_npages == 0) ||
509 			    (bp->b_pages[bp->b_npages-1] != m)) {
510 				bp->b_pages[bp->b_npages] = m;
511 				bp->b_npages++;
512 			}
513 			if (vm_page_all_valid(m))
514 				tbp->b_pages[j] = bogus_page;
515 		}
516 
517 		/*
518 		 * Don't inherit tbp->b_bufsize as it may be larger due to
519 		 * a non-page-aligned size.  Instead just aggregate using
520 		 * 'size'.
521 		 */
522 		if (tbp->b_bcount != size)
523 			printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
524 		if (tbp->b_bufsize != size)
525 			printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
526 		bp->b_bcount += size;
527 		bp->b_bufsize += size;
528 	}
529 
530 	/*
531 	 * Fully valid pages in the cluster are already good and do not need
532 	 * to be re-read from disk.  Replace the page with bogus_page
533 	 */
534 	for (j = 0; j < bp->b_npages; j++) {
535 		if (vm_page_all_valid(bp->b_pages[j]))
536 			bp->b_pages[j] = bogus_page;
537 	}
538 	if (bp->b_bufsize > bp->b_kvasize)
539 		panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
540 		    bp->b_bufsize, bp->b_kvasize);
541 
542 	if (buf_mapped(bp)) {
543 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
544 		    (vm_page_t *)bp->b_pages, bp->b_npages);
545 	}
546 	return (bp);
547 }
548 
549 /*
550  * Cleanup after a clustered read or write.
551  * This is complicated by the fact that any of the buffers might have
552  * extra memory (if there were no empty buffer headers at allocbuf time)
553  * that we will need to shift around.
554  */
555 static void
556 cluster_callback(struct buf *bp)
557 {
558 	struct buf *nbp, *tbp;
559 	int error = 0;
560 
561 	/*
562 	 * Must propagate errors to all the components.
563 	 */
564 	if (bp->b_ioflags & BIO_ERROR)
565 		error = bp->b_error;
566 
567 	if (buf_mapped(bp)) {
568 		pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
569 		    bp->b_npages);
570 	}
571 	/*
572 	 * Move memory from the large cluster buffer into the component
573 	 * buffers and mark IO as done on these.
574 	 */
575 	for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
576 		tbp; tbp = nbp) {
577 		nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
578 		if (error) {
579 			tbp->b_ioflags |= BIO_ERROR;
580 			tbp->b_error = error;
581 		} else {
582 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
583 			tbp->b_flags &= ~B_INVAL;
584 			tbp->b_ioflags &= ~BIO_ERROR;
585 			/*
586 			 * XXX the bdwrite()/bqrelse() issued during
587 			 * cluster building clears B_RELBUF (see bqrelse()
588 			 * comment).  If direct I/O was specified, we have
589 			 * to restore it here to allow the buffer and VM
590 			 * to be freed.
591 			 */
592 			if (tbp->b_flags & B_DIRECT)
593 				tbp->b_flags |= B_RELBUF;
594 		}
595 		bufdone(tbp);
596 	}
597 	pbrelvp(bp);
598 	uma_zfree(cluster_pbuf_zone, bp);
599 }
600 
601 /*
602  *	cluster_wbuild_wb:
603  *
604  *	Implement modified write build for cluster.
605  *
606  *		write_behind = 0	write behind disabled
607  *		write_behind = 1	write behind normal (default)
608  *		write_behind = 2	write behind backed-off
609  */
610 
611 static __inline int
612 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
613     int gbflags)
614 {
615 	int r = 0;
616 
617 	switch (write_behind) {
618 	case 2:
619 		if (start_lbn < len)
620 			break;
621 		start_lbn -= len;
622 		/* FALLTHROUGH */
623 	case 1:
624 		r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
625 		/* FALLTHROUGH */
626 	default:
627 		/* FALLTHROUGH */
628 		break;
629 	}
630 	return(r);
631 }
632 
633 /*
634  * Do clustered write for FFS.
635  *
636  * Three cases:
637  *	1. Write is not sequential (write asynchronously)
638  *	Write is sequential:
639  *	2.	beginning of cluster - begin cluster
640  *	3.	middle of a cluster - add to cluster
641  *	4.	end of a cluster - asynchronously write cluster
642  */
643 void
644 cluster_write(struct vnode *vp, struct vn_clusterw *vnc, struct buf *bp,
645     u_quad_t filesize, int seqcount, int gbflags)
646 {
647 	daddr_t lbn, pbn;
648 	int maxclen, cursize;
649 	int lblocksize;
650 	int async;
651 
652 	if (!unmapped_buf_allowed)
653 		gbflags &= ~GB_UNMAPPED;
654 
655 	if (vp->v_type == VREG) {
656 		async = DOINGASYNC(vp);
657 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
658 	} else {
659 		async = 0;
660 		lblocksize = bp->b_bufsize;
661 	}
662 	lbn = bp->b_lblkno;
663 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
664 
665 	/* Initialize vnode to beginning of file. */
666 	if (lbn == 0)
667 		vnc->v_lasta = vnc->v_clen = vnc->v_cstart = vnc->v_lastw = 0;
668 
669 	if (vnc->v_clen == 0 || lbn != vnc->v_lastw + 1 ||
670 	    (bp->b_blkno != vnc->v_lasta + btodb(lblocksize))) {
671 		maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
672 		if (vnc->v_clen != 0) {
673 			/*
674 			 * Next block is not sequential.
675 			 *
676 			 * If we are not writing at end of file, the process
677 			 * seeked to another point in the file since its last
678 			 * write, or we have reached our maximum cluster size,
679 			 * then push the previous cluster. Otherwise try
680 			 * reallocating to make it sequential.
681 			 *
682 			 * Change to algorithm: only push previous cluster if
683 			 * it was sequential from the point of view of the
684 			 * seqcount heuristic, otherwise leave the buffer
685 			 * intact so we can potentially optimize the I/O
686 			 * later on in the buf_daemon or update daemon
687 			 * flush.
688 			 */
689 			cursize = vnc->v_lastw - vnc->v_cstart + 1;
690 			if ((u_quad_t)bp->b_offset + lblocksize != filesize ||
691 			    lbn != vnc->v_lastw + 1 || vnc->v_clen <= cursize) {
692 				if (!async && seqcount > 0) {
693 					cluster_wbuild_wb(vp, lblocksize,
694 					    vnc->v_cstart, cursize, gbflags);
695 				}
696 			} else {
697 				struct buf **bpp, **endbp;
698 				struct cluster_save *buflist;
699 
700 				buflist = cluster_collectbufs(vp, vnc, bp,
701 				    gbflags);
702 				if (buflist == NULL) {
703 					/*
704 					 * Cluster build failed so just write
705 					 * it now.
706 					 */
707 					bawrite(bp);
708 					return;
709 				}
710 				endbp = &buflist->bs_children
711 				    [buflist->bs_nchildren - 1];
712 				if (VOP_REALLOCBLKS(vp, buflist)) {
713 					/*
714 					 * Failed, push the previous cluster
715 					 * if *really* writing sequentially
716 					 * in the logical file (seqcount > 1),
717 					 * otherwise delay it in the hopes that
718 					 * the low level disk driver can
719 					 * optimize the write ordering.
720 					 */
721 					for (bpp = buflist->bs_children;
722 					     bpp < endbp; bpp++)
723 						brelse(*bpp);
724 					free(buflist, M_SEGMENT);
725 					if (seqcount > 1) {
726 						cluster_wbuild_wb(vp,
727 						    lblocksize, vnc->v_cstart,
728 						    cursize, gbflags);
729 					}
730 				} else {
731 					/*
732 					 * Succeeded, keep building cluster.
733 					 */
734 					for (bpp = buflist->bs_children;
735 					     bpp <= endbp; bpp++)
736 						bdwrite(*bpp);
737 					free(buflist, M_SEGMENT);
738 					vnc->v_lastw = lbn;
739 					vnc->v_lasta = bp->b_blkno;
740 					return;
741 				}
742 			}
743 		}
744 		/*
745 		 * Consider beginning a cluster. If at end of file, make
746 		 * cluster as large as possible, otherwise find size of
747 		 * existing cluster.
748 		 */
749 		if (vp->v_type == VREG &&
750 		    (u_quad_t) bp->b_offset + lblocksize != filesize &&
751 		    bp->b_blkno == bp->b_lblkno &&
752 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen,
753 		    NULL) != 0 || bp->b_blkno == -1)) {
754 			pbn = bp->b_blkno;
755 			bawrite(bp);
756 			vnc->v_clen = 0;
757 			vnc->v_lasta = pbn;
758 			vnc->v_cstart = lbn + 1;
759 			vnc->v_lastw = lbn;
760 			return;
761 		}
762 		vnc->v_clen = maxclen;
763 		pbn = bp->b_blkno;
764 		if (!async && maxclen == 0) {	/* I/O not contiguous */
765 			vnc->v_cstart = lbn + 1;
766 			bawrite(bp);
767 		} else {	/* Wait for rest of cluster */
768 			vnc->v_cstart = lbn;
769 			bdwrite(bp);
770 		}
771 	} else if (lbn == vnc->v_cstart + vnc->v_clen) {
772 		/*
773 		 * At end of cluster, write it out if seqcount tells us we
774 		 * are operating sequentially, otherwise let the buf or
775 		 * update daemon handle it.
776 		 */
777 		pbn = bp->b_blkno;
778 		bdwrite(bp);
779 		if (seqcount > 1) {
780 			cluster_wbuild_wb(vp, lblocksize, vnc->v_cstart,
781 			    vnc->v_clen + 1, gbflags);
782 		}
783 		vnc->v_clen = 0;
784 		vnc->v_cstart = lbn + 1;
785 	} else if (vm_page_count_severe()) {
786 		/*
787 		 * We are low on memory, get it going NOW
788 		 */
789 		pbn = bp->b_blkno;
790 		bawrite(bp);
791 	} else {
792 		/*
793 		 * In the middle of a cluster, so just delay the I/O for now.
794 		 */
795 		pbn = bp->b_blkno;
796 		bdwrite(bp);
797 	}
798 	vnc->v_lastw = lbn;
799 	vnc->v_lasta = pbn;
800 }
801 
802 /*
803  * This is an awful lot like cluster_rbuild...wish they could be combined.
804  * The last lbn argument is the current block on which I/O is being
805  * performed.  Check to see that it doesn't fall in the middle of
806  * the current block (if last_bp == NULL).
807  */
808 int
809 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
810     int gbflags)
811 {
812 	struct buf *bp, *tbp;
813 	struct bufobj *bo;
814 	int i, j;
815 	int totalwritten = 0;
816 	int dbsize = btodb(size);
817 
818 	if (!unmapped_buf_allowed)
819 		gbflags &= ~GB_UNMAPPED;
820 
821 	bo = &vp->v_bufobj;
822 	while (len > 0) {
823 		/*
824 		 * If the buffer is not delayed-write (i.e. dirty), or it
825 		 * is delayed-write but either locked or inval, it cannot
826 		 * partake in the clustered write.
827 		 */
828 		BO_LOCK(bo);
829 		if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
830 		    (tbp->b_vflags & BV_BKGRDINPROG)) {
831 			BO_UNLOCK(bo);
832 			++start_lbn;
833 			--len;
834 			continue;
835 		}
836 		if (BUF_LOCK(tbp,
837 		    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) {
838 			++start_lbn;
839 			--len;
840 			continue;
841 		}
842 		if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
843 			BUF_UNLOCK(tbp);
844 			++start_lbn;
845 			--len;
846 			continue;
847 		}
848 		bremfree(tbp);
849 		tbp->b_flags &= ~B_DONE;
850 
851 		/*
852 		 * Extra memory in the buffer, punt on this buffer.
853 		 * XXX we could handle this in most cases, but we would
854 		 * have to push the extra memory down to after our max
855 		 * possible cluster size and then potentially pull it back
856 		 * up if the cluster was terminated prematurely--too much
857 		 * hassle.
858 		 */
859 		if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
860 		     (B_CLUSTEROK | B_VMIO)) ||
861 		  (tbp->b_bcount != tbp->b_bufsize) ||
862 		  (tbp->b_bcount != size) ||
863 		  (len == 1) ||
864 		  ((bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT)) == NULL)) {
865 			totalwritten += tbp->b_bufsize;
866 			bawrite(tbp);
867 			++start_lbn;
868 			--len;
869 			continue;
870 		}
871 		MPASS((bp->b_flags & B_MAXPHYS) != 0);
872 
873 		/*
874 		 * We got a pbuf to make the cluster in.
875 		 * so initialise it.
876 		 */
877 		TAILQ_INIT(&bp->b_cluster.cluster_head);
878 		bp->b_bcount = 0;
879 		bp->b_bufsize = 0;
880 		bp->b_npages = 0;
881 		if (tbp->b_wcred != NOCRED)
882 			bp->b_wcred = crhold(tbp->b_wcred);
883 
884 		bp->b_blkno = tbp->b_blkno;
885 		bp->b_lblkno = tbp->b_lblkno;
886 		bp->b_offset = tbp->b_offset;
887 
888 		/*
889 		 * We are synthesizing a buffer out of vm_page_t's, but
890 		 * if the block size is not page aligned then the starting
891 		 * address may not be either.  Inherit the b_data offset
892 		 * from the original buffer.
893 		 */
894 		if ((gbflags & GB_UNMAPPED) == 0 ||
895 		    (tbp->b_flags & B_VMIO) == 0) {
896 			bp->b_data = (char *)((vm_offset_t)bp->b_data |
897 			    ((vm_offset_t)tbp->b_data & PAGE_MASK));
898 		} else {
899 			bp->b_data = unmapped_buf;
900 		}
901 		bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
902 		    B_NEEDCOMMIT));
903 		bp->b_iodone = cluster_callback;
904 		pbgetvp(vp, bp);
905 		/*
906 		 * From this location in the file, scan forward to see
907 		 * if there are buffers with adjacent data that need to
908 		 * be written as well.
909 		 */
910 		for (i = 0; i < len; ++i, ++start_lbn) {
911 			if (i != 0) { /* If not the first buffer */
912 				/*
913 				 * If the adjacent data is not even in core it
914 				 * can't need to be written.
915 				 */
916 				BO_LOCK(bo);
917 				if ((tbp = gbincore(bo, start_lbn)) == NULL ||
918 				    (tbp->b_vflags & BV_BKGRDINPROG)) {
919 					BO_UNLOCK(bo);
920 					break;
921 				}
922 
923 				/*
924 				 * If it IS in core, but has different
925 				 * characteristics, or is locked (which
926 				 * means it could be undergoing a background
927 				 * I/O or be in a weird state), then don't
928 				 * cluster with it.
929 				 */
930 				if (BUF_LOCK(tbp,
931 				    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
932 				    BO_LOCKPTR(bo)))
933 					break;
934 
935 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
936 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
937 				    != (B_DELWRI | B_CLUSTEROK |
938 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
939 				    tbp->b_wcred != bp->b_wcred) {
940 					BUF_UNLOCK(tbp);
941 					break;
942 				}
943 
944 				/*
945 				 * Check that the combined cluster
946 				 * would make sense with regard to pages
947 				 * and would not be too large
948 				 */
949 				if ((tbp->b_bcount != size) ||
950 				  ((bp->b_blkno + (dbsize * i)) !=
951 				    tbp->b_blkno) ||
952 				  ((tbp->b_npages + bp->b_npages) >
953 				    (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
954 					BUF_UNLOCK(tbp);
955 					break;
956 				}
957 
958 				/*
959 				 * Ok, it's passed all the tests,
960 				 * so remove it from the free list
961 				 * and mark it busy. We will use it.
962 				 */
963 				bremfree(tbp);
964 				tbp->b_flags &= ~B_DONE;
965 			} /* end of code for non-first buffers only */
966 			/*
967 			 * If the IO is via the VM then we do some
968 			 * special VM hackery (yuck).  Since the buffer's
969 			 * block size may not be page-aligned it is possible
970 			 * for a page to be shared between two buffers.  We
971 			 * have to get rid of the duplication when building
972 			 * the cluster.
973 			 */
974 			if (tbp->b_flags & B_VMIO) {
975 				vm_page_t m;
976 
977 				if (i == 0) {
978 					vfs_busy_pages_acquire(tbp);
979 				} else { /* if not first buffer */
980 					for (j = 0; j < tbp->b_npages; j += 1) {
981 						m = tbp->b_pages[j];
982 						if (vm_page_trysbusy(m) == 0) {
983 							for (j--; j >= 0; j--)
984 								vm_page_sunbusy(
985 								    tbp->b_pages[j]);
986 							bqrelse(tbp);
987 							goto finishcluster;
988 						}
989 					}
990 				}
991 				vm_object_pip_add(tbp->b_bufobj->bo_object,
992 				    tbp->b_npages);
993 				for (j = 0; j < tbp->b_npages; j += 1) {
994 					m = tbp->b_pages[j];
995 					if ((bp->b_npages == 0) ||
996 					  (bp->b_pages[bp->b_npages - 1] != m)) {
997 						bp->b_pages[bp->b_npages] = m;
998 						bp->b_npages++;
999 					}
1000 				}
1001 			}
1002 			bp->b_bcount += size;
1003 			bp->b_bufsize += size;
1004 			/*
1005 			 * If any of the clustered buffers have their
1006 			 * B_BARRIER flag set, transfer that request to
1007 			 * the cluster.
1008 			 */
1009 			bp->b_flags |= (tbp->b_flags & B_BARRIER);
1010 			tbp->b_flags &= ~(B_DONE | B_BARRIER);
1011 			tbp->b_flags |= B_ASYNC;
1012 			tbp->b_ioflags &= ~BIO_ERROR;
1013 			tbp->b_iocmd = BIO_WRITE;
1014 			bundirty(tbp);
1015 			reassignbuf(tbp);		/* put on clean list */
1016 			bufobj_wref(tbp->b_bufobj);
1017 			BUF_KERNPROC(tbp);
1018 			buf_track(tbp, __func__);
1019 			TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
1020 				tbp, b_cluster.cluster_entry);
1021 		}
1022 	finishcluster:
1023 		if (buf_mapped(bp)) {
1024 			pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1025 			    (vm_page_t *)bp->b_pages, bp->b_npages);
1026 		}
1027 		if (bp->b_bufsize > bp->b_kvasize)
1028 			panic(
1029 			    "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
1030 			    bp->b_bufsize, bp->b_kvasize);
1031 		totalwritten += bp->b_bufsize;
1032 		bp->b_dirtyoff = 0;
1033 		bp->b_dirtyend = bp->b_bufsize;
1034 		bawrite(bp);
1035 
1036 		len -= i;
1037 	}
1038 	return totalwritten;
1039 }
1040 
1041 /*
1042  * Collect together all the buffers in a cluster.
1043  * Plus add one additional buffer.
1044  */
1045 static struct cluster_save *
1046 cluster_collectbufs(struct vnode *vp, struct vn_clusterw *vnc,
1047     struct buf *last_bp, int gbflags)
1048 {
1049 	struct cluster_save *buflist;
1050 	struct buf *bp;
1051 	daddr_t lbn;
1052 	int i, j, len, error;
1053 
1054 	len = vnc->v_lastw - vnc->v_cstart + 1;
1055 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1056 	    M_SEGMENT, M_WAITOK);
1057 	buflist->bs_nchildren = 0;
1058 	buflist->bs_children = (struct buf **) (buflist + 1);
1059 	for (lbn = vnc->v_cstart, i = 0; i < len; lbn++, i++) {
1060 		error = bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1061 		    gbflags, &bp);
1062 		if (error != 0) {
1063 			/*
1064 			 * If read fails, release collected buffers
1065 			 * and return failure.
1066 			 */
1067 			for (j = 0; j < i; j++)
1068 				brelse(buflist->bs_children[j]);
1069 			free(buflist, M_SEGMENT);
1070 			return (NULL);
1071 		}
1072 		buflist->bs_children[i] = bp;
1073 		if (bp->b_blkno == bp->b_lblkno)
1074 			VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1075 				NULL, NULL);
1076 	}
1077 	buflist->bs_children[i] = bp = last_bp;
1078 	if (bp->b_blkno == bp->b_lblkno)
1079 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1080 	buflist->bs_nchildren = i + 1;
1081 	return (buflist);
1082 }
1083 
1084 void
1085 cluster_init_vn(struct vn_clusterw *vnc)
1086 {
1087 	vnc->v_lasta = 0;
1088 	vnc->v_clen = 0;
1089 	vnc->v_cstart = 0;
1090 	vnc->v_lastw = 0;
1091 }
1092