xref: /freebsd/sys/kern/vfs_cluster.c (revision f126890ac5386406dadf7c4cfa9566cbb56537c5)
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 
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/kernel.h>
37 #include <sys/proc.h>
38 #include <sys/bio.h>
39 #include <sys/buf.h>
40 #include <sys/vnode.h>
41 #include <sys/malloc.h>
42 #include <sys/mount.h>
43 #include <sys/racct.h>
44 #include <sys/resourcevar.h>
45 #include <sys/rwlock.h>
46 #include <sys/vmmeter.h>
47 #include <vm/vm.h>
48 #include <vm/vm_object.h>
49 #include <vm/vm_page.h>
50 #include <sys/sysctl.h>
51 
52 static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer");
53 static uma_zone_t cluster_pbuf_zone;
54 
55 static void cluster_init(void *);
56 static struct cluster_save *cluster_collectbufs(struct vnode *vp,
57 	    struct vn_clusterw *vnc, struct buf *last_bp, int gbflags);
58 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize,
59 	    daddr_t lbn, daddr_t blkno, long size, int run, int gbflags,
60 	    struct buf *fbp);
61 static void cluster_callback(struct buf *);
62 
63 static int write_behind = 1;
64 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
65     "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
66 
67 static int read_max = 64;
68 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
69     "Cluster read-ahead max block count");
70 
71 static int read_min = 1;
72 SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0,
73     "Cluster read min block count");
74 
75 SYSINIT(cluster, SI_SUB_CPU, SI_ORDER_ANY, cluster_init, NULL);
76 
77 static void
78 cluster_init(void *dummy)
79 {
80 
81 	cluster_pbuf_zone = pbuf_zsecond_create("clpbuf", nswbuf / 2);
82 }
83 
84 /*
85  * Read data to a buf, including read-ahead if we find this to be beneficial.
86  * cluster_read replaces bread.
87  */
88 int
89 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
90     struct ucred *cred, long totread, int seqcount, int gbflags,
91     struct buf **bpp)
92 {
93 	struct buf *bp, *rbp, *reqbp;
94 	struct bufobj *bo;
95 	struct thread *td;
96 	daddr_t blkno, origblkno;
97 	int maxra, racluster;
98 	int error, ncontig;
99 	int i;
100 
101 	error = 0;
102 	td = curthread;
103 	bo = &vp->v_bufobj;
104 	if (!unmapped_buf_allowed)
105 		gbflags &= ~GB_UNMAPPED;
106 
107 	/*
108 	 * Try to limit the amount of read-ahead by a few
109 	 * ad-hoc parameters.  This needs work!!!
110 	 */
111 	racluster = vp->v_mount->mnt_iosize_max / size;
112 	maxra = seqcount;
113 	maxra = min(read_max, maxra);
114 	maxra = min(nbuf/8, maxra);
115 	if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
116 		maxra = (filesize / size) - lblkno;
117 
118 	/*
119 	 * get the requested block
120 	 */
121 	error = getblkx(vp, lblkno, lblkno, size, 0, 0, gbflags, &bp);
122 	if (error != 0) {
123 		*bpp = NULL;
124 		return (error);
125 	}
126 	gbflags &= ~GB_NOSPARSE;
127 	origblkno = lblkno;
128 	*bpp = reqbp = bp;
129 
130 	/*
131 	 * if it is in the cache, then check to see if the reads have been
132 	 * sequential.  If they have, then try some read-ahead, otherwise
133 	 * back-off on prospective read-aheads.
134 	 */
135 	if (bp->b_flags & B_CACHE) {
136 		if (!seqcount) {
137 			return 0;
138 		} else if ((bp->b_flags & B_RAM) == 0) {
139 			return 0;
140 		} else {
141 			bp->b_flags &= ~B_RAM;
142 			BO_RLOCK(bo);
143 			for (i = 1; i < maxra; i++) {
144 				/*
145 				 * Stop if the buffer does not exist or it
146 				 * is invalid (about to go away?)
147 				 */
148 				rbp = gbincore(&vp->v_bufobj, lblkno+i);
149 				if (rbp == NULL || (rbp->b_flags & B_INVAL))
150 					break;
151 
152 				/*
153 				 * Set another read-ahead mark so we know
154 				 * to check again. (If we can lock the
155 				 * buffer without waiting)
156 				 */
157 				if ((((i % racluster) == (racluster - 1)) ||
158 				    (i == (maxra - 1)))
159 				    && (0 == BUF_LOCK(rbp,
160 					LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
161 					rbp->b_flags |= B_RAM;
162 					BUF_UNLOCK(rbp);
163 				}
164 			}
165 			BO_RUNLOCK(bo);
166 			if (i >= maxra) {
167 				return 0;
168 			}
169 			lblkno += i;
170 		}
171 		reqbp = bp = NULL;
172 	/*
173 	 * If it isn't in the cache, then get a chunk from
174 	 * disk if sequential, otherwise just get the block.
175 	 */
176 	} else {
177 		off_t firstread = bp->b_offset;
178 		int nblks;
179 		long minread;
180 
181 		KASSERT(bp->b_offset != NOOFFSET,
182 		    ("cluster_read: no buffer offset"));
183 
184 		ncontig = 0;
185 
186 		/*
187 		 * Adjust totread if needed
188 		 */
189 		minread = read_min * size;
190 		if (minread > totread)
191 			totread = minread;
192 
193 		/*
194 		 * Compute the total number of blocks that we should read
195 		 * synchronously.
196 		 */
197 		if (firstread + totread > filesize)
198 			totread = filesize - firstread;
199 		nblks = howmany(totread, size);
200 		if (nblks > racluster)
201 			nblks = racluster;
202 
203 		/*
204 		 * Now compute the number of contiguous blocks.
205 		 */
206 		if (nblks > 1) {
207 	    		error = VOP_BMAP(vp, lblkno, NULL,
208 				&blkno, &ncontig, NULL);
209 			/*
210 			 * If this failed to map just do the original block.
211 			 */
212 			if (error || blkno == -1)
213 				ncontig = 0;
214 		}
215 
216 		/*
217 		 * If we have contiguous data available do a cluster
218 		 * otherwise just read the requested block.
219 		 */
220 		if (ncontig) {
221 			/* Account for our first block. */
222 			ncontig = min(ncontig + 1, nblks);
223 			if (ncontig < nblks)
224 				nblks = ncontig;
225 			bp = cluster_rbuild(vp, filesize, lblkno,
226 			    blkno, size, nblks, gbflags, bp);
227 			lblkno += (bp->b_bufsize / size);
228 		} else {
229 			bp->b_flags |= B_RAM;
230 			bp->b_iocmd = BIO_READ;
231 			lblkno += 1;
232 		}
233 	}
234 
235 	/*
236 	 * handle the synchronous read so that it is available ASAP.
237 	 */
238 	if (bp) {
239 		if ((bp->b_flags & B_CLUSTER) == 0) {
240 			vfs_busy_pages(bp, 0);
241 		}
242 		bp->b_flags &= ~B_INVAL;
243 		bp->b_ioflags &= ~BIO_ERROR;
244 		if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
245 			BUF_KERNPROC(bp);
246 		bp->b_iooffset = dbtob(bp->b_blkno);
247 		bstrategy(bp);
248 #ifdef RACCT
249 		if (racct_enable) {
250 			PROC_LOCK(td->td_proc);
251 			racct_add_buf(td->td_proc, bp, 0);
252 			PROC_UNLOCK(td->td_proc);
253 		}
254 #endif /* RACCT */
255 		td->td_ru.ru_inblock++;
256 	}
257 
258 	/*
259 	 * If we have been doing sequential I/O, then do some read-ahead.
260 	 */
261 	while (lblkno < (origblkno + maxra)) {
262 		error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
263 		if (error)
264 			break;
265 
266 		if (blkno == -1)
267 			break;
268 
269 		/*
270 		 * We could throttle ncontig here by maxra but we might as
271 		 * well read the data if it is contiguous.  We're throttled
272 		 * by racluster anyway.
273 		 */
274 		if (ncontig) {
275 			ncontig = min(ncontig + 1, racluster);
276 			rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
277 			    size, ncontig, gbflags, NULL);
278 			lblkno += (rbp->b_bufsize / size);
279 			if (rbp->b_flags & B_DELWRI) {
280 				bqrelse(rbp);
281 				continue;
282 			}
283 		} else {
284 			rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
285 			lblkno += 1;
286 			if (rbp->b_flags & B_DELWRI) {
287 				bqrelse(rbp);
288 				continue;
289 			}
290 			rbp->b_flags |= B_ASYNC | B_RAM;
291 			rbp->b_iocmd = BIO_READ;
292 			rbp->b_blkno = blkno;
293 		}
294 		if (rbp->b_flags & B_CACHE) {
295 			rbp->b_flags &= ~B_ASYNC;
296 			bqrelse(rbp);
297 			continue;
298 		}
299 		if ((rbp->b_flags & B_CLUSTER) == 0) {
300 			vfs_busy_pages(rbp, 0);
301 		}
302 		rbp->b_flags &= ~B_INVAL;
303 		rbp->b_ioflags &= ~BIO_ERROR;
304 		if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
305 			BUF_KERNPROC(rbp);
306 		rbp->b_iooffset = dbtob(rbp->b_blkno);
307 		bstrategy(rbp);
308 #ifdef RACCT
309 		if (racct_enable) {
310 			PROC_LOCK(td->td_proc);
311 			racct_add_buf(td->td_proc, rbp, 0);
312 			PROC_UNLOCK(td->td_proc);
313 		}
314 #endif /* RACCT */
315 		td->td_ru.ru_inblock++;
316 	}
317 
318 	if (reqbp) {
319 		/*
320 		 * Like bread, always brelse() the buffer when
321 		 * returning an error.
322 		 */
323 		error = bufwait(reqbp);
324 		if (error != 0) {
325 			brelse(reqbp);
326 			*bpp = NULL;
327 		}
328 	}
329 	return (error);
330 }
331 
332 /*
333  * If blocks are contiguous on disk, use this to provide clustered
334  * read ahead.  We will read as many blocks as possible sequentially
335  * and then parcel them up into logical blocks in the buffer hash table.
336  */
337 static struct buf *
338 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
339     daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
340 {
341 	struct buf *bp, *tbp;
342 	daddr_t bn;
343 	off_t off;
344 	long tinc, tsize;
345 	int i, inc, j, k, toff;
346 
347 	KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
348 	    ("cluster_rbuild: size %ld != f_iosize %jd\n",
349 	    size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
350 
351 	/*
352 	 * avoid a division
353 	 */
354 	while ((u_quad_t) size * (lbn + run) > filesize) {
355 		--run;
356 	}
357 
358 	if (fbp) {
359 		tbp = fbp;
360 		tbp->b_iocmd = BIO_READ;
361 	} else {
362 		tbp = getblk(vp, lbn, size, 0, 0, gbflags);
363 		if (tbp->b_flags & B_CACHE)
364 			return tbp;
365 		tbp->b_flags |= B_ASYNC | B_RAM;
366 		tbp->b_iocmd = BIO_READ;
367 	}
368 	tbp->b_blkno = blkno;
369 	if ( (tbp->b_flags & B_MALLOC) ||
370 		((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
371 		return tbp;
372 
373 	bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT);
374 	if (bp == NULL)
375 		return tbp;
376 	MPASS((bp->b_flags & B_MAXPHYS) != 0);
377 
378 	/*
379 	 * We are synthesizing a buffer out of vm_page_t's, but
380 	 * if the block size is not page aligned then the starting
381 	 * address may not be either.  Inherit the b_data offset
382 	 * from the original buffer.
383 	 */
384 	bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
385 	if ((gbflags & GB_UNMAPPED) != 0) {
386 		bp->b_data = unmapped_buf;
387 	} else {
388 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
389 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
390 	}
391 	bp->b_iocmd = BIO_READ;
392 	bp->b_iodone = cluster_callback;
393 	bp->b_blkno = blkno;
394 	bp->b_lblkno = lbn;
395 	bp->b_offset = tbp->b_offset;
396 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
397 	pbgetvp(vp, bp);
398 
399 	TAILQ_INIT(&bp->b_cluster.cluster_head);
400 
401 	bp->b_bcount = 0;
402 	bp->b_bufsize = 0;
403 	bp->b_npages = 0;
404 
405 	inc = btodb(size);
406 	for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
407 		if (i == 0) {
408 			vm_object_pip_add(tbp->b_bufobj->bo_object,
409 			    tbp->b_npages);
410 			vfs_busy_pages_acquire(tbp);
411 		} else {
412 			if ((bp->b_npages * PAGE_SIZE) +
413 			    round_page(size) > vp->v_mount->mnt_iosize_max) {
414 				break;
415 			}
416 
417 			tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
418 			    (gbflags & GB_UNMAPPED));
419 
420 			/* Don't wait around for locked bufs. */
421 			if (tbp == NULL)
422 				break;
423 
424 			/*
425 			 * Stop scanning if the buffer is fully valid
426 			 * (marked B_CACHE), or locked (may be doing a
427 			 * background write), or if the buffer is not
428 			 * VMIO backed.  The clustering code can only deal
429 			 * with VMIO-backed buffers.  The bo lock is not
430 			 * required for the BKGRDINPROG check since it
431 			 * can not be set without the buf lock.
432 			 */
433 			if ((tbp->b_vflags & BV_BKGRDINPROG) ||
434 			    (tbp->b_flags & B_CACHE) ||
435 			    (tbp->b_flags & B_VMIO) == 0) {
436 				bqrelse(tbp);
437 				break;
438 			}
439 
440 			/*
441 			 * The buffer must be completely invalid in order to
442 			 * take part in the cluster.  If it is partially valid
443 			 * then we stop.
444 			 */
445 			off = tbp->b_offset;
446 			tsize = size;
447 			for (j = 0; tsize > 0; j++) {
448 				toff = off & PAGE_MASK;
449 				tinc = tsize;
450 				if (toff + tinc > PAGE_SIZE)
451 					tinc = PAGE_SIZE - toff;
452 				if (vm_page_trysbusy(tbp->b_pages[j]) == 0)
453 					break;
454 				if ((tbp->b_pages[j]->valid &
455 				    vm_page_bits(toff, tinc)) != 0) {
456 					vm_page_sunbusy(tbp->b_pages[j]);
457 					break;
458 				}
459 				vm_object_pip_add(tbp->b_bufobj->bo_object, 1);
460 				off += tinc;
461 				tsize -= tinc;
462 			}
463 			if (tsize > 0) {
464 clean_sbusy:
465 				vm_object_pip_wakeupn(tbp->b_bufobj->bo_object,
466 				    j);
467 				for (k = 0; k < j; k++)
468 					vm_page_sunbusy(tbp->b_pages[k]);
469 				bqrelse(tbp);
470 				break;
471 			}
472 
473 			/*
474 			 * Set a read-ahead mark as appropriate
475 			 */
476 			if ((fbp && (i == 1)) || (i == (run - 1)))
477 				tbp->b_flags |= B_RAM;
478 
479 			/*
480 			 * Set the buffer up for an async read (XXX should
481 			 * we do this only if we do not wind up brelse()ing?).
482 			 * Set the block number if it isn't set, otherwise
483 			 * if it is make sure it matches the block number we
484 			 * expect.
485 			 */
486 			tbp->b_flags |= B_ASYNC;
487 			tbp->b_iocmd = BIO_READ;
488 			if (tbp->b_blkno == tbp->b_lblkno) {
489 				tbp->b_blkno = bn;
490 			} else if (tbp->b_blkno != bn) {
491 				goto clean_sbusy;
492 			}
493 		}
494 		/*
495 		 * XXX fbp from caller may not be B_ASYNC, but we are going
496 		 * to biodone() it in cluster_callback() anyway
497 		 */
498 		BUF_KERNPROC(tbp);
499 		TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
500 			tbp, b_cluster.cluster_entry);
501 		for (j = 0; j < tbp->b_npages; j += 1) {
502 			vm_page_t m;
503 
504 			m = tbp->b_pages[j];
505 			if ((bp->b_npages == 0) ||
506 			    (bp->b_pages[bp->b_npages-1] != m)) {
507 				bp->b_pages[bp->b_npages] = m;
508 				bp->b_npages++;
509 			}
510 			if (vm_page_all_valid(m))
511 				tbp->b_pages[j] = bogus_page;
512 		}
513 
514 		/*
515 		 * Don't inherit tbp->b_bufsize as it may be larger due to
516 		 * a non-page-aligned size.  Instead just aggregate using
517 		 * 'size'.
518 		 */
519 		if (tbp->b_bcount != size)
520 			printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
521 		if (tbp->b_bufsize != size)
522 			printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
523 		bp->b_bcount += size;
524 		bp->b_bufsize += size;
525 	}
526 
527 	/*
528 	 * Fully valid pages in the cluster are already good and do not need
529 	 * to be re-read from disk.  Replace the page with bogus_page
530 	 */
531 	for (j = 0; j < bp->b_npages; j++) {
532 		if (vm_page_all_valid(bp->b_pages[j]))
533 			bp->b_pages[j] = bogus_page;
534 	}
535 	if (bp->b_bufsize > bp->b_kvasize)
536 		panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
537 		    bp->b_bufsize, bp->b_kvasize);
538 
539 	if (buf_mapped(bp)) {
540 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
541 		    (vm_page_t *)bp->b_pages, bp->b_npages);
542 	}
543 	return (bp);
544 }
545 
546 /*
547  * Cleanup after a clustered read or write.
548  * This is complicated by the fact that any of the buffers might have
549  * extra memory (if there were no empty buffer headers at allocbuf time)
550  * that we will need to shift around.
551  */
552 static void
553 cluster_callback(struct buf *bp)
554 {
555 	struct buf *nbp, *tbp;
556 	int error = 0;
557 
558 	/*
559 	 * Must propagate errors to all the components.
560 	 */
561 	if (bp->b_ioflags & BIO_ERROR)
562 		error = bp->b_error;
563 
564 	if (buf_mapped(bp)) {
565 		pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
566 		    bp->b_npages);
567 	}
568 	/*
569 	 * Move memory from the large cluster buffer into the component
570 	 * buffers and mark IO as done on these.
571 	 */
572 	for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
573 		tbp; tbp = nbp) {
574 		nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
575 		if (error) {
576 			tbp->b_ioflags |= BIO_ERROR;
577 			tbp->b_error = error;
578 		} else {
579 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
580 			tbp->b_flags &= ~B_INVAL;
581 			tbp->b_ioflags &= ~BIO_ERROR;
582 			/*
583 			 * XXX the bdwrite()/bqrelse() issued during
584 			 * cluster building clears B_RELBUF (see bqrelse()
585 			 * comment).  If direct I/O was specified, we have
586 			 * to restore it here to allow the buffer and VM
587 			 * to be freed.
588 			 */
589 			if (tbp->b_flags & B_DIRECT)
590 				tbp->b_flags |= B_RELBUF;
591 		}
592 		bufdone(tbp);
593 	}
594 	pbrelvp(bp);
595 	uma_zfree(cluster_pbuf_zone, bp);
596 }
597 
598 /*
599  *	cluster_wbuild_wb:
600  *
601  *	Implement modified write build for cluster.
602  *
603  *		write_behind = 0	write behind disabled
604  *		write_behind = 1	write behind normal (default)
605  *		write_behind = 2	write behind backed-off
606  */
607 
608 static __inline int
609 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
610     int gbflags)
611 {
612 	int r = 0;
613 
614 	switch (write_behind) {
615 	case 2:
616 		if (start_lbn < len)
617 			break;
618 		start_lbn -= len;
619 		/* FALLTHROUGH */
620 	case 1:
621 		r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
622 		/* FALLTHROUGH */
623 	default:
624 		/* FALLTHROUGH */
625 		break;
626 	}
627 	return(r);
628 }
629 
630 /*
631  * Do clustered write for FFS.
632  *
633  * Three cases:
634  *	1. Write is not sequential (write asynchronously)
635  *	Write is sequential:
636  *	2.	beginning of cluster - begin cluster
637  *	3.	middle of a cluster - add to cluster
638  *	4.	end of a cluster - asynchronously write cluster
639  */
640 void
641 cluster_write(struct vnode *vp, struct vn_clusterw *vnc, struct buf *bp,
642     u_quad_t filesize, int seqcount, int gbflags)
643 {
644 	daddr_t lbn, pbn;
645 	int maxclen, cursize;
646 	int lblocksize;
647 	int async;
648 
649 	if (!unmapped_buf_allowed)
650 		gbflags &= ~GB_UNMAPPED;
651 
652 	if (vp->v_type == VREG) {
653 		async = DOINGASYNC(vp);
654 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
655 	} else {
656 		async = 0;
657 		lblocksize = bp->b_bufsize;
658 	}
659 	lbn = bp->b_lblkno;
660 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
661 
662 	/* Initialize vnode to beginning of file. */
663 	if (lbn == 0)
664 		vnc->v_lasta = vnc->v_clen = vnc->v_cstart = vnc->v_lastw = 0;
665 
666 	if (vnc->v_clen == 0 || lbn != vnc->v_lastw + 1 ||
667 	    (bp->b_blkno != vnc->v_lasta + btodb(lblocksize))) {
668 		maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
669 		if (vnc->v_clen != 0) {
670 			/*
671 			 * Next block is not sequential.
672 			 *
673 			 * If we are not writing at end of file, the process
674 			 * seeked to another point in the file since its last
675 			 * write, or we have reached our maximum cluster size,
676 			 * then push the previous cluster. Otherwise try
677 			 * reallocating to make it sequential.
678 			 *
679 			 * Change to algorithm: only push previous cluster if
680 			 * it was sequential from the point of view of the
681 			 * seqcount heuristic, otherwise leave the buffer
682 			 * intact so we can potentially optimize the I/O
683 			 * later on in the buf_daemon or update daemon
684 			 * flush.
685 			 */
686 			cursize = vnc->v_lastw - vnc->v_cstart + 1;
687 			if ((u_quad_t)bp->b_offset + lblocksize != filesize ||
688 			    lbn != vnc->v_lastw + 1 || vnc->v_clen <= cursize) {
689 				if (!async && seqcount > 0) {
690 					cluster_wbuild_wb(vp, lblocksize,
691 					    vnc->v_cstart, cursize, gbflags);
692 				}
693 			} else {
694 				struct buf **bpp, **endbp;
695 				struct cluster_save *buflist;
696 
697 				buflist = cluster_collectbufs(vp, vnc, bp,
698 				    gbflags);
699 				if (buflist == NULL) {
700 					/*
701 					 * Cluster build failed so just write
702 					 * it now.
703 					 */
704 					bawrite(bp);
705 					return;
706 				}
707 				endbp = &buflist->bs_children
708 				    [buflist->bs_nchildren - 1];
709 				if (VOP_REALLOCBLKS(vp, buflist)) {
710 					/*
711 					 * Failed, push the previous cluster
712 					 * if *really* writing sequentially
713 					 * in the logical file (seqcount > 1),
714 					 * otherwise delay it in the hopes that
715 					 * the low level disk driver can
716 					 * optimize the write ordering.
717 					 */
718 					for (bpp = buflist->bs_children;
719 					     bpp < endbp; bpp++)
720 						brelse(*bpp);
721 					free(buflist, M_SEGMENT);
722 					if (seqcount > 1) {
723 						cluster_wbuild_wb(vp,
724 						    lblocksize, vnc->v_cstart,
725 						    cursize, gbflags);
726 					}
727 				} else {
728 					/*
729 					 * Succeeded, keep building cluster.
730 					 */
731 					for (bpp = buflist->bs_children;
732 					     bpp <= endbp; bpp++)
733 						bdwrite(*bpp);
734 					free(buflist, M_SEGMENT);
735 					vnc->v_lastw = lbn;
736 					vnc->v_lasta = bp->b_blkno;
737 					return;
738 				}
739 			}
740 		}
741 		/*
742 		 * Consider beginning a cluster. If at end of file, make
743 		 * cluster as large as possible, otherwise find size of
744 		 * existing cluster.
745 		 */
746 		if (vp->v_type == VREG &&
747 		    (u_quad_t) bp->b_offset + lblocksize != filesize &&
748 		    bp->b_blkno == bp->b_lblkno &&
749 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen,
750 		    NULL) != 0 || bp->b_blkno == -1)) {
751 			pbn = bp->b_blkno;
752 			bawrite(bp);
753 			vnc->v_clen = 0;
754 			vnc->v_lasta = pbn;
755 			vnc->v_cstart = lbn + 1;
756 			vnc->v_lastw = lbn;
757 			return;
758 		}
759 		vnc->v_clen = maxclen;
760 		pbn = bp->b_blkno;
761 		if (!async && maxclen == 0) {	/* I/O not contiguous */
762 			vnc->v_cstart = lbn + 1;
763 			bawrite(bp);
764 		} else {	/* Wait for rest of cluster */
765 			vnc->v_cstart = lbn;
766 			bdwrite(bp);
767 		}
768 	} else if (lbn == vnc->v_cstart + vnc->v_clen) {
769 		/*
770 		 * At end of cluster, write it out if seqcount tells us we
771 		 * are operating sequentially, otherwise let the buf or
772 		 * update daemon handle it.
773 		 */
774 		pbn = bp->b_blkno;
775 		bdwrite(bp);
776 		if (seqcount > 1) {
777 			cluster_wbuild_wb(vp, lblocksize, vnc->v_cstart,
778 			    vnc->v_clen + 1, gbflags);
779 		}
780 		vnc->v_clen = 0;
781 		vnc->v_cstart = lbn + 1;
782 	} else if (vm_page_count_severe()) {
783 		/*
784 		 * We are low on memory, get it going NOW
785 		 */
786 		pbn = bp->b_blkno;
787 		bawrite(bp);
788 	} else {
789 		/*
790 		 * In the middle of a cluster, so just delay the I/O for now.
791 		 */
792 		pbn = bp->b_blkno;
793 		bdwrite(bp);
794 	}
795 	vnc->v_lastw = lbn;
796 	vnc->v_lasta = pbn;
797 }
798 
799 /*
800  * This is an awful lot like cluster_rbuild...wish they could be combined.
801  * The last lbn argument is the current block on which I/O is being
802  * performed.  Check to see that it doesn't fall in the middle of
803  * the current block (if last_bp == NULL).
804  */
805 int
806 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
807     int gbflags)
808 {
809 	struct buf *bp, *tbp;
810 	struct bufobj *bo;
811 	int i, j;
812 	int totalwritten = 0;
813 	int dbsize = btodb(size);
814 
815 	if (!unmapped_buf_allowed)
816 		gbflags &= ~GB_UNMAPPED;
817 
818 	bo = &vp->v_bufobj;
819 	while (len > 0) {
820 		/*
821 		 * If the buffer is not delayed-write (i.e. dirty), or it
822 		 * is delayed-write but either locked or inval, it cannot
823 		 * partake in the clustered write.
824 		 */
825 		BO_LOCK(bo);
826 		if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
827 		    (tbp->b_vflags & BV_BKGRDINPROG)) {
828 			BO_UNLOCK(bo);
829 			++start_lbn;
830 			--len;
831 			continue;
832 		}
833 		if (BUF_LOCK(tbp,
834 		    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) {
835 			++start_lbn;
836 			--len;
837 			continue;
838 		}
839 		if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
840 			BUF_UNLOCK(tbp);
841 			++start_lbn;
842 			--len;
843 			continue;
844 		}
845 		bremfree(tbp);
846 		tbp->b_flags &= ~B_DONE;
847 
848 		/*
849 		 * Extra memory in the buffer, punt on this buffer.
850 		 * XXX we could handle this in most cases, but we would
851 		 * have to push the extra memory down to after our max
852 		 * possible cluster size and then potentially pull it back
853 		 * up if the cluster was terminated prematurely--too much
854 		 * hassle.
855 		 */
856 		if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
857 		     (B_CLUSTEROK | B_VMIO)) ||
858 		  (tbp->b_bcount != tbp->b_bufsize) ||
859 		  (tbp->b_bcount != size) ||
860 		  (len == 1) ||
861 		  ((bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT)) == NULL)) {
862 			totalwritten += tbp->b_bufsize;
863 			bawrite(tbp);
864 			++start_lbn;
865 			--len;
866 			continue;
867 		}
868 		MPASS((bp->b_flags & B_MAXPHYS) != 0);
869 
870 		/*
871 		 * We got a pbuf to make the cluster in.
872 		 * so initialise it.
873 		 */
874 		TAILQ_INIT(&bp->b_cluster.cluster_head);
875 		bp->b_bcount = 0;
876 		bp->b_bufsize = 0;
877 		bp->b_npages = 0;
878 		if (tbp->b_wcred != NOCRED)
879 			bp->b_wcred = crhold(tbp->b_wcred);
880 
881 		bp->b_blkno = tbp->b_blkno;
882 		bp->b_lblkno = tbp->b_lblkno;
883 		bp->b_offset = tbp->b_offset;
884 
885 		/*
886 		 * We are synthesizing a buffer out of vm_page_t's, but
887 		 * if the block size is not page aligned then the starting
888 		 * address may not be either.  Inherit the b_data offset
889 		 * from the original buffer.
890 		 */
891 		if ((gbflags & GB_UNMAPPED) == 0 ||
892 		    (tbp->b_flags & B_VMIO) == 0) {
893 			bp->b_data = (char *)((vm_offset_t)bp->b_data |
894 			    ((vm_offset_t)tbp->b_data & PAGE_MASK));
895 		} else {
896 			bp->b_data = unmapped_buf;
897 		}
898 		bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
899 		    B_NEEDCOMMIT));
900 		bp->b_iodone = cluster_callback;
901 		pbgetvp(vp, bp);
902 		/*
903 		 * From this location in the file, scan forward to see
904 		 * if there are buffers with adjacent data that need to
905 		 * be written as well.
906 		 */
907 		for (i = 0; i < len; ++i, ++start_lbn) {
908 			if (i != 0) { /* If not the first buffer */
909 				/*
910 				 * If the adjacent data is not even in core it
911 				 * can't need to be written.
912 				 */
913 				BO_LOCK(bo);
914 				if ((tbp = gbincore(bo, start_lbn)) == NULL ||
915 				    (tbp->b_vflags & BV_BKGRDINPROG)) {
916 					BO_UNLOCK(bo);
917 					break;
918 				}
919 
920 				/*
921 				 * If it IS in core, but has different
922 				 * characteristics, or is locked (which
923 				 * means it could be undergoing a background
924 				 * I/O or be in a weird state), then don't
925 				 * cluster with it.
926 				 */
927 				if (BUF_LOCK(tbp,
928 				    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
929 				    BO_LOCKPTR(bo)))
930 					break;
931 
932 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
933 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
934 				    != (B_DELWRI | B_CLUSTEROK |
935 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
936 				    tbp->b_wcred != bp->b_wcred) {
937 					BUF_UNLOCK(tbp);
938 					break;
939 				}
940 
941 				/*
942 				 * Check that the combined cluster
943 				 * would make sense with regard to pages
944 				 * and would not be too large
945 				 */
946 				if ((tbp->b_bcount != size) ||
947 				  ((bp->b_blkno + (dbsize * i)) !=
948 				    tbp->b_blkno) ||
949 				  ((tbp->b_npages + bp->b_npages) >
950 				    (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
951 					BUF_UNLOCK(tbp);
952 					break;
953 				}
954 
955 				/*
956 				 * Ok, it's passed all the tests,
957 				 * so remove it from the free list
958 				 * and mark it busy. We will use it.
959 				 */
960 				bremfree(tbp);
961 				tbp->b_flags &= ~B_DONE;
962 			} /* end of code for non-first buffers only */
963 			/*
964 			 * If the IO is via the VM then we do some
965 			 * special VM hackery (yuck).  Since the buffer's
966 			 * block size may not be page-aligned it is possible
967 			 * for a page to be shared between two buffers.  We
968 			 * have to get rid of the duplication when building
969 			 * the cluster.
970 			 */
971 			if (tbp->b_flags & B_VMIO) {
972 				vm_page_t m;
973 
974 				if (i == 0) {
975 					vfs_busy_pages_acquire(tbp);
976 				} else { /* if not first buffer */
977 					for (j = 0; j < tbp->b_npages; j += 1) {
978 						m = tbp->b_pages[j];
979 						if (vm_page_trysbusy(m) == 0) {
980 							for (j--; j >= 0; j--)
981 								vm_page_sunbusy(
982 								    tbp->b_pages[j]);
983 							bqrelse(tbp);
984 							goto finishcluster;
985 						}
986 					}
987 				}
988 				vm_object_pip_add(tbp->b_bufobj->bo_object,
989 				    tbp->b_npages);
990 				for (j = 0; j < tbp->b_npages; j += 1) {
991 					m = tbp->b_pages[j];
992 					if ((bp->b_npages == 0) ||
993 					  (bp->b_pages[bp->b_npages - 1] != m)) {
994 						bp->b_pages[bp->b_npages] = m;
995 						bp->b_npages++;
996 					}
997 				}
998 			}
999 			bp->b_bcount += size;
1000 			bp->b_bufsize += size;
1001 			/*
1002 			 * If any of the clustered buffers have their
1003 			 * B_BARRIER flag set, transfer that request to
1004 			 * the cluster.
1005 			 */
1006 			bp->b_flags |= (tbp->b_flags & B_BARRIER);
1007 			tbp->b_flags &= ~(B_DONE | B_BARRIER);
1008 			tbp->b_flags |= B_ASYNC;
1009 			tbp->b_ioflags &= ~BIO_ERROR;
1010 			tbp->b_iocmd = BIO_WRITE;
1011 			bundirty(tbp);
1012 			reassignbuf(tbp);		/* put on clean list */
1013 			bufobj_wref(tbp->b_bufobj);
1014 			BUF_KERNPROC(tbp);
1015 			buf_track(tbp, __func__);
1016 			TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
1017 				tbp, b_cluster.cluster_entry);
1018 		}
1019 	finishcluster:
1020 		if (buf_mapped(bp)) {
1021 			pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1022 			    (vm_page_t *)bp->b_pages, bp->b_npages);
1023 		}
1024 		if (bp->b_bufsize > bp->b_kvasize)
1025 			panic(
1026 			    "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
1027 			    bp->b_bufsize, bp->b_kvasize);
1028 		totalwritten += bp->b_bufsize;
1029 		bp->b_dirtyoff = 0;
1030 		bp->b_dirtyend = bp->b_bufsize;
1031 		bawrite(bp);
1032 
1033 		len -= i;
1034 	}
1035 	return totalwritten;
1036 }
1037 
1038 /*
1039  * Collect together all the buffers in a cluster.
1040  * Plus add one additional buffer.
1041  */
1042 static struct cluster_save *
1043 cluster_collectbufs(struct vnode *vp, struct vn_clusterw *vnc,
1044     struct buf *last_bp, int gbflags)
1045 {
1046 	struct cluster_save *buflist;
1047 	struct buf *bp;
1048 	daddr_t lbn;
1049 	int i, j, len, error;
1050 
1051 	len = vnc->v_lastw - vnc->v_cstart + 1;
1052 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1053 	    M_SEGMENT, M_WAITOK);
1054 	buflist->bs_nchildren = 0;
1055 	buflist->bs_children = (struct buf **) (buflist + 1);
1056 	for (lbn = vnc->v_cstart, i = 0; i < len; lbn++, i++) {
1057 		error = bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1058 		    gbflags, &bp);
1059 		if (error != 0) {
1060 			/*
1061 			 * If read fails, release collected buffers
1062 			 * and return failure.
1063 			 */
1064 			for (j = 0; j < i; j++)
1065 				brelse(buflist->bs_children[j]);
1066 			free(buflist, M_SEGMENT);
1067 			return (NULL);
1068 		}
1069 		buflist->bs_children[i] = bp;
1070 		if (bp->b_blkno == bp->b_lblkno)
1071 			VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1072 				NULL, NULL);
1073 	}
1074 	buflist->bs_children[i] = bp = last_bp;
1075 	if (bp->b_blkno == bp->b_lblkno)
1076 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1077 	buflist->bs_nchildren = i + 1;
1078 	return (buflist);
1079 }
1080 
1081 void
1082 cluster_init_vn(struct vn_clusterw *vnc)
1083 {
1084 	vnc->v_lasta = 0;
1085 	vnc->v_clen = 0;
1086 	vnc->v_cstart = 0;
1087 	vnc->v_lastw = 0;
1088 }
1089