xref: /netbsd-src/sys/ufs/lfs/lfs_bio.c (revision de4fa6c51a9708fc05f88b618fa6fad87c9508ec)
1 /*	$NetBSD: lfs_bio.c,v 1.114 2008/05/06 18:43:45 ad Exp $	*/
2 
3 /*-
4  * Copyright (c) 1999, 2000, 2001, 2002, 2003, 2008 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Konrad E. Schroder <perseant@hhhh.org>.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 /*
32  * Copyright (c) 1991, 1993
33  *	The Regents of the University of California.  All rights reserved.
34  *
35  * Redistribution and use in source and binary forms, with or without
36  * modification, are permitted provided that the following conditions
37  * are met:
38  * 1. Redistributions of source code must retain the above copyright
39  *    notice, this list of conditions and the following disclaimer.
40  * 2. Redistributions in binary form must reproduce the above copyright
41  *    notice, this list of conditions and the following disclaimer in the
42  *    documentation and/or other materials provided with the distribution.
43  * 3. Neither the name of the University nor the names of its contributors
44  *    may be used to endorse or promote products derived from this software
45  *    without specific prior written permission.
46  *
47  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
48  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
49  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
50  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
51  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
52  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
53  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
54  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
55  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
56  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
57  * SUCH DAMAGE.
58  *
59  *	@(#)lfs_bio.c	8.10 (Berkeley) 6/10/95
60  */
61 
62 #include <sys/cdefs.h>
63 __KERNEL_RCSID(0, "$NetBSD: lfs_bio.c,v 1.114 2008/05/06 18:43:45 ad Exp $");
64 
65 #include <sys/param.h>
66 #include <sys/systm.h>
67 #include <sys/proc.h>
68 #include <sys/buf.h>
69 #include <sys/vnode.h>
70 #include <sys/resourcevar.h>
71 #include <sys/mount.h>
72 #include <sys/kernel.h>
73 #include <sys/kauth.h>
74 
75 #include <ufs/ufs/inode.h>
76 #include <ufs/ufs/ufsmount.h>
77 #include <ufs/ufs/ufs_extern.h>
78 
79 #include <ufs/lfs/lfs.h>
80 #include <ufs/lfs/lfs_extern.h>
81 
82 #include <uvm/uvm.h>
83 
84 /*
85  * LFS block write function.
86  *
87  * XXX
88  * No write cost accounting is done.
89  * This is almost certainly wrong for synchronous operations and NFS.
90  *
91  * protected by lfs_lock.
92  */
93 int	locked_queue_count   = 0;	/* Count of locked-down buffers. */
94 long	locked_queue_bytes   = 0L;	/* Total size of locked buffers. */
95 int	lfs_subsys_pages     = 0L;	/* Total number LFS-written pages */
96 int	lfs_fs_pagetrip	     = 0;	/* # of pages to trip per-fs write */
97 int	lfs_writing	     = 0;	/* Set if already kicked off a writer
98 					   because of buffer space */
99 
100 /* Lock and condition variables for above. */
101 kcondvar_t	locked_queue_cv;
102 kcondvar_t	lfs_writing_cv;
103 kmutex_t	lfs_lock;
104 
105 extern int lfs_dostats;
106 
107 /*
108  * reserved number/bytes of locked buffers
109  */
110 int locked_queue_rcount = 0;
111 long locked_queue_rbytes = 0L;
112 
113 int lfs_fits_buf(struct lfs *, int, int);
114 int lfs_reservebuf(struct lfs *, struct vnode *vp, struct vnode *vp2,
115     int, int);
116 int lfs_reserveavail(struct lfs *, struct vnode *vp, struct vnode *vp2, int);
117 
118 int
119 lfs_fits_buf(struct lfs *fs, int n, int bytes)
120 {
121 	int count_fit, bytes_fit;
122 
123 	ASSERT_NO_SEGLOCK(fs);
124 	KASSERT(mutex_owned(&lfs_lock));
125 
126 	count_fit =
127 	    (locked_queue_count + locked_queue_rcount + n < LFS_WAIT_BUFS);
128 	bytes_fit =
129 	    (locked_queue_bytes + locked_queue_rbytes + bytes < LFS_WAIT_BYTES);
130 
131 #ifdef DEBUG
132 	if (!count_fit) {
133 		DLOG((DLOG_AVAIL, "lfs_fits_buf: no fit count: %d + %d + %d >= %d\n",
134 		      locked_queue_count, locked_queue_rcount,
135 		      n, LFS_WAIT_BUFS));
136 	}
137 	if (!bytes_fit) {
138 		DLOG((DLOG_AVAIL, "lfs_fits_buf: no fit bytes: %ld + %ld + %d >= %ld\n",
139 		      locked_queue_bytes, locked_queue_rbytes,
140 		      bytes, LFS_WAIT_BYTES));
141 	}
142 #endif /* DEBUG */
143 
144 	return (count_fit && bytes_fit);
145 }
146 
147 /* ARGSUSED */
148 int
149 lfs_reservebuf(struct lfs *fs, struct vnode *vp,
150     struct vnode *vp2, int n, int bytes)
151 {
152 	ASSERT_MAYBE_SEGLOCK(fs);
153 	KASSERT(locked_queue_rcount >= 0);
154 	KASSERT(locked_queue_rbytes >= 0);
155 
156 	mutex_enter(&lfs_lock);
157 	while (n > 0 && !lfs_fits_buf(fs, n, bytes)) {
158 		int error;
159 
160 		lfs_flush(fs, 0, 0);
161 
162 		error = cv_timedwait_sig(&locked_queue_cv, &lfs_lock,
163 		    hz * LFS_BUFWAIT);
164 		if (error && error != EWOULDBLOCK) {
165 			mutex_exit(&lfs_lock);
166 			return error;
167 		}
168 	}
169 
170 	locked_queue_rcount += n;
171 	locked_queue_rbytes += bytes;
172 
173 	mutex_exit(&lfs_lock);
174 
175 	KASSERT(locked_queue_rcount >= 0);
176 	KASSERT(locked_queue_rbytes >= 0);
177 
178 	return 0;
179 }
180 
181 /*
182  * Try to reserve some blocks, prior to performing a sensitive operation that
183  * requires the vnode lock to be honored.  If there is not enough space, give
184  * up the vnode lock temporarily and wait for the space to become available.
185  *
186  * Called with vp locked.  (Note nowever that if fsb < 0, vp is ignored.)
187  *
188  * XXX YAMT - it isn't safe to unlock vp here
189  * because the node might be modified while we sleep.
190  * (eg. cached states like i_offset might be stale,
191  *  the vnode might be truncated, etc..)
192  * maybe we should have a way to restart the vnodeop (EVOPRESTART?)
193  * or rearrange vnodeop interface to leave vnode locking to file system
194  * specific code so that each file systems can have their own vnode locking and
195  * vnode re-using strategies.
196  */
197 int
198 lfs_reserveavail(struct lfs *fs, struct vnode *vp,
199     struct vnode *vp2, int fsb)
200 {
201 	CLEANERINFO *cip;
202 	struct buf *bp;
203 	int error, slept;
204 
205 	ASSERT_MAYBE_SEGLOCK(fs);
206 	slept = 0;
207 	mutex_enter(&lfs_lock);
208 	while (fsb > 0 && !lfs_fits(fs, fsb + fs->lfs_ravail + fs->lfs_favail)) {
209 		mutex_exit(&lfs_lock);
210 #if 0
211 		/*
212 		 * XXX ideally, we should unlock vnodes here
213 		 * because we might sleep very long time.
214 		 */
215 		VOP_UNLOCK(vp, 0);
216 		if (vp2 != NULL) {
217 			VOP_UNLOCK(vp2, 0);
218 		}
219 #else
220 		/*
221 		 * XXX since we'll sleep for cleaner with vnode lock holding,
222 		 * deadlock will occur if cleaner tries to lock the vnode.
223 		 * (eg. lfs_markv -> lfs_fastvget -> getnewvnode -> vclean)
224 		 */
225 #endif
226 
227 		if (!slept) {
228 			DLOG((DLOG_AVAIL, "lfs_reserve: waiting for %ld (bfree = %d,"
229 			      " est_bfree = %d)\n",
230 			      fsb + fs->lfs_ravail + fs->lfs_favail,
231 			      fs->lfs_bfree, LFS_EST_BFREE(fs)));
232 		}
233 		++slept;
234 
235 		/* Wake up the cleaner */
236 		LFS_CLEANERINFO(cip, fs, bp);
237 		LFS_SYNC_CLEANERINFO(cip, fs, bp, 0);
238 		lfs_wakeup_cleaner(fs);
239 
240 		mutex_enter(&lfs_lock);
241 		/* Cleaner might have run while we were reading, check again */
242 		if (lfs_fits(fs, fsb + fs->lfs_ravail + fs->lfs_favail))
243 			break;
244 
245 		error = mtsleep(&fs->lfs_avail, PCATCH | PUSER, "lfs_reserve",
246 				0, &lfs_lock);
247 #if 0
248 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); /* XXX use lockstatus */
249 		vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY); /* XXX use lockstatus */
250 #endif
251 		if (error) {
252 			mutex_exit(&lfs_lock);
253 			return error;
254 		}
255 	}
256 #ifdef DEBUG
257 	if (slept) {
258 		DLOG((DLOG_AVAIL, "lfs_reserve: woke up\n"));
259 	}
260 #endif
261 	fs->lfs_ravail += fsb;
262 	mutex_exit(&lfs_lock);
263 
264 	return 0;
265 }
266 
267 #ifdef DIAGNOSTIC
268 int lfs_rescount;
269 int lfs_rescountdirop;
270 #endif
271 
272 int
273 lfs_reserve(struct lfs *fs, struct vnode *vp, struct vnode *vp2, int fsb)
274 {
275 	int error;
276 	int cantwait;
277 
278 	ASSERT_MAYBE_SEGLOCK(fs);
279 	if (vp2) {
280 		/* Make sure we're not in the process of reclaiming vp2 */
281 		mutex_enter(&lfs_lock);
282 		while(fs->lfs_flags & LFS_UNDIROP) {
283 			mtsleep(&fs->lfs_flags, PRIBIO + 1, "lfsrundirop", 0,
284 			    &lfs_lock);
285 		}
286 		mutex_exit(&lfs_lock);
287 	}
288 
289 	KASSERT(fsb < 0 || VOP_ISLOCKED(vp));
290 	KASSERT(vp2 == NULL || fsb < 0 || VOP_ISLOCKED(vp2));
291 	KASSERT(vp2 == NULL || !(VTOI(vp2)->i_flag & IN_ADIROP));
292 	KASSERT(vp2 == NULL || vp2 != fs->lfs_unlockvp);
293 
294 	cantwait = (VTOI(vp)->i_flag & IN_ADIROP) || fs->lfs_unlockvp == vp;
295 #ifdef DIAGNOSTIC
296 	if (cantwait) {
297 		if (fsb > 0)
298 			lfs_rescountdirop++;
299 		else if (fsb < 0)
300 			lfs_rescountdirop--;
301 		if (lfs_rescountdirop < 0)
302 			panic("lfs_rescountdirop");
303 	}
304 	else {
305 		if (fsb > 0)
306 			lfs_rescount++;
307 		else if (fsb < 0)
308 			lfs_rescount--;
309 		if (lfs_rescount < 0)
310 			panic("lfs_rescount");
311 	}
312 #endif
313 	if (cantwait)
314 		return 0;
315 
316 	/*
317 	 * XXX
318 	 * vref vnodes here so that cleaner doesn't try to reuse them.
319 	 * (see XXX comment in lfs_reserveavail)
320 	 */
321 	mutex_enter(&vp->v_interlock);
322 	lfs_vref(vp);
323 	if (vp2 != NULL) {
324 		mutex_enter(&vp2->v_interlock);
325 		lfs_vref(vp2);
326 	}
327 
328 	error = lfs_reserveavail(fs, vp, vp2, fsb);
329 	if (error)
330 		goto done;
331 
332 	/*
333 	 * XXX just a guess. should be more precise.
334 	 */
335 	error = lfs_reservebuf(fs, vp, vp2,
336 	    fragstoblks(fs, fsb), fsbtob(fs, fsb));
337 	if (error)
338 		lfs_reserveavail(fs, vp, vp2, -fsb);
339 
340 done:
341 	lfs_vunref(vp);
342 	if (vp2 != NULL) {
343 		lfs_vunref(vp2);
344 	}
345 
346 	return error;
347 }
348 
349 int
350 lfs_bwrite(void *v)
351 {
352 	struct vop_bwrite_args /* {
353 		struct buf *a_bp;
354 	} */ *ap = v;
355 	struct buf *bp = ap->a_bp;
356 
357 #ifdef DIAGNOSTIC
358 	if (VTOI(bp->b_vp)->i_lfs->lfs_ronly == 0 && (bp->b_flags & B_ASYNC)) {
359 		panic("bawrite LFS buffer");
360 	}
361 #endif /* DIAGNOSTIC */
362 	return lfs_bwrite_ext(bp, 0);
363 }
364 
365 /*
366  * Determine if there is enough room currently available to write fsb
367  * blocks.  We need enough blocks for the new blocks, the current
368  * inode blocks (including potentially the ifile inode), a summary block,
369  * and the segment usage table, plus an ifile block.
370  */
371 int
372 lfs_fits(struct lfs *fs, int fsb)
373 {
374 	int needed;
375 
376 	ASSERT_NO_SEGLOCK(fs);
377 	needed = fsb + btofsb(fs, fs->lfs_sumsize) +
378 		 ((howmany(fs->lfs_uinodes + 1, INOPB(fs)) + fs->lfs_segtabsz +
379 		   1) << (fs->lfs_blktodb - fs->lfs_fsbtodb));
380 
381 	if (needed >= fs->lfs_avail) {
382 #ifdef DEBUG
383 		DLOG((DLOG_AVAIL, "lfs_fits: no fit: fsb = %ld, uinodes = %ld, "
384 		      "needed = %ld, avail = %ld\n",
385 		      (long)fsb, (long)fs->lfs_uinodes, (long)needed,
386 		      (long)fs->lfs_avail));
387 #endif
388 		return 0;
389 	}
390 	return 1;
391 }
392 
393 int
394 lfs_availwait(struct lfs *fs, int fsb)
395 {
396 	int error;
397 	CLEANERINFO *cip;
398 	struct buf *cbp;
399 
400 	ASSERT_NO_SEGLOCK(fs);
401 	/* Push cleaner blocks through regardless */
402 	mutex_enter(&lfs_lock);
403 	if (LFS_SEGLOCK_HELD(fs) &&
404 	    fs->lfs_sp->seg_flags & (SEGM_CLEAN | SEGM_FORCE_CKP)) {
405 		mutex_exit(&lfs_lock);
406 		return 0;
407 	}
408 	mutex_exit(&lfs_lock);
409 
410 	while (!lfs_fits(fs, fsb)) {
411 		/*
412 		 * Out of space, need cleaner to run.
413 		 * Update the cleaner info, then wake it up.
414 		 * Note the cleanerinfo block is on the ifile
415 		 * so it CANT_WAIT.
416 		 */
417 		LFS_CLEANERINFO(cip, fs, cbp);
418 		LFS_SYNC_CLEANERINFO(cip, fs, cbp, 0);
419 
420 #ifdef DEBUG
421 		DLOG((DLOG_AVAIL, "lfs_availwait: out of available space, "
422 		      "waiting on cleaner\n"));
423 #endif
424 
425 		lfs_wakeup_cleaner(fs);
426 #ifdef DIAGNOSTIC
427 		if (LFS_SEGLOCK_HELD(fs))
428 			panic("lfs_availwait: deadlock");
429 #endif
430 		error = tsleep(&fs->lfs_avail, PCATCH | PUSER, "cleaner", 0);
431 		if (error)
432 			return (error);
433 	}
434 	return 0;
435 }
436 
437 int
438 lfs_bwrite_ext(struct buf *bp, int flags)
439 {
440 	struct lfs *fs;
441 	struct inode *ip;
442 	struct vnode *vp;
443 	int fsb;
444 
445 	vp = bp->b_vp;
446 	fs = VFSTOUFS(vp->v_mount)->um_lfs;
447 
448 	ASSERT_MAYBE_SEGLOCK(fs);
449 	KASSERT(bp->b_cflags & BC_BUSY);
450 	KASSERT(flags & BW_CLEAN || !LFS_IS_MALLOC_BUF(bp));
451 	KASSERT(((bp->b_oflags | bp->b_flags) & (BO_DELWRI|B_LOCKED))
452 	    != BO_DELWRI);
453 
454 	/*
455 	 * Don't write *any* blocks if we're mounted read-only, or
456 	 * if we are "already unmounted".
457 	 *
458 	 * In particular the cleaner can't write blocks either.
459 	 */
460 	if (fs->lfs_ronly || (fs->lfs_pflags & LFS_PF_CLEAN)) {
461 		bp->b_oflags &= ~BO_DELWRI;
462 		bp->b_flags |= B_READ;
463 		bp->b_error = 0;
464 		mutex_enter(&bufcache_lock);
465 		LFS_UNLOCK_BUF(bp);
466 		if (LFS_IS_MALLOC_BUF(bp))
467 			bp->b_cflags &= ~BC_BUSY;
468 		else
469 			brelsel(bp, 0);
470 		mutex_exit(&bufcache_lock);
471 		return (fs->lfs_ronly ? EROFS : 0);
472 	}
473 
474 	/*
475 	 * Set the delayed write flag and use reassignbuf to move the buffer
476 	 * from the clean list to the dirty one.
477 	 *
478 	 * Set the B_LOCKED flag and unlock the buffer, causing brelse to move
479 	 * the buffer onto the LOCKED free list.  This is necessary, otherwise
480 	 * getnewbuf() would try to reclaim the buffers using bawrite, which
481 	 * isn't going to work.
482 	 *
483 	 * XXX we don't let meta-data writes run out of space because they can
484 	 * come from the segment writer.  We need to make sure that there is
485 	 * enough space reserved so that there's room to write meta-data
486 	 * blocks.
487 	 */
488 	if ((bp->b_flags & B_LOCKED) == 0) {
489 		fsb = fragstofsb(fs, numfrags(fs, bp->b_bcount));
490 
491 		ip = VTOI(vp);
492 		mutex_enter(&lfs_lock);
493 		if (flags & BW_CLEAN) {
494 			LFS_SET_UINO(ip, IN_CLEANING);
495 		} else {
496 			LFS_SET_UINO(ip, IN_MODIFIED);
497 		}
498 		mutex_exit(&lfs_lock);
499 		fs->lfs_avail -= fsb;
500 
501 		mutex_enter(&bufcache_lock);
502 		mutex_enter(&vp->v_interlock);
503 		bp->b_oflags = (bp->b_oflags | BO_DELWRI) & ~BO_DONE;
504 		LFS_LOCK_BUF(bp);
505 		bp->b_flags &= ~B_READ;
506 		bp->b_error = 0;
507 		reassignbuf(bp, bp->b_vp);
508 		mutex_exit(&vp->v_interlock);
509 	} else {
510 		mutex_enter(&bufcache_lock);
511 	}
512 
513 	if (bp->b_iodone != NULL)
514 		bp->b_cflags &= ~BC_BUSY;
515 	else
516 		brelsel(bp, 0);
517 	mutex_exit(&bufcache_lock);
518 
519 	return (0);
520 }
521 
522 /*
523  * Called and return with the lfs_lock held.
524  */
525 void
526 lfs_flush_fs(struct lfs *fs, int flags)
527 {
528 	ASSERT_NO_SEGLOCK(fs);
529 	KASSERT(mutex_owned(&lfs_lock));
530 	if (fs->lfs_ronly)
531 		return;
532 
533 	if (lfs_dostats)
534 		++lfs_stats.flush_invoked;
535 
536 	mutex_exit(&lfs_lock);
537 	lfs_writer_enter(fs, "fldirop");
538 	lfs_segwrite(fs->lfs_ivnode->v_mount, flags);
539 	lfs_writer_leave(fs);
540 	mutex_enter(&lfs_lock);
541 	fs->lfs_favail = 0; /* XXX */
542 }
543 
544 /*
545  * This routine initiates segment writes when LFS is consuming too many
546  * resources.  Ideally the pageout daemon would be able to direct LFS
547  * more subtly.
548  * XXX We have one static count of locked buffers;
549  * XXX need to think more about the multiple filesystem case.
550  *
551  * Called and return with lfs_lock held.
552  * If fs != NULL, we hold the segment lock for fs.
553  */
554 void
555 lfs_flush(struct lfs *fs, int flags, int only_onefs)
556 {
557 	extern u_int64_t locked_fakequeue_count;
558 	struct mount *mp, *nmp;
559 	struct lfs *tfs;
560 
561 	KASSERT(mutex_owned(&lfs_lock));
562 	KDASSERT(fs == NULL || !LFS_SEGLOCK_HELD(fs));
563 
564 	if (lfs_dostats)
565 		++lfs_stats.write_exceeded;
566 	/* XXX should we include SEGM_CKP here? */
567 	if (lfs_writing && !(flags & SEGM_SYNC)) {
568 		DLOG((DLOG_FLUSH, "lfs_flush: not flushing because another flush is active\n"));
569 		return;
570 	}
571 	while (lfs_writing)
572 		cv_wait(&lfs_writing_cv, &lfs_lock);
573 	lfs_writing = 1;
574 
575 	mutex_exit(&lfs_lock);
576 
577 	if (only_onefs) {
578 		KASSERT(fs != NULL);
579 		if (vfs_busy(fs->lfs_ivnode->v_mount, NULL))
580 			goto errout;
581 		mutex_enter(&lfs_lock);
582 		lfs_flush_fs(fs, flags);
583 		mutex_exit(&lfs_lock);
584 		vfs_unbusy(fs->lfs_ivnode->v_mount, false, NULL);
585 	} else {
586 		locked_fakequeue_count = 0;
587 		mutex_enter(&mountlist_lock);
588 		for (mp = CIRCLEQ_FIRST(&mountlist); mp != (void *)&mountlist;
589 		     mp = nmp) {
590 			if (vfs_busy(mp, &nmp)) {
591 				DLOG((DLOG_FLUSH, "lfs_flush: fs vfs_busy\n"));
592 				continue;
593 			}
594 			if (strncmp(&mp->mnt_stat.f_fstypename[0], MOUNT_LFS,
595 			    sizeof(mp->mnt_stat.f_fstypename)) == 0) {
596 				tfs = VFSTOUFS(mp)->um_lfs;
597 				mutex_enter(&lfs_lock);
598 				lfs_flush_fs(tfs, flags);
599 				mutex_exit(&lfs_lock);
600 			}
601 			vfs_unbusy(mp, false, &nmp);
602 		}
603 		mutex_exit(&mountlist_lock);
604 	}
605 	LFS_DEBUG_COUNTLOCKED("flush");
606 	wakeup(&lfs_subsys_pages);
607 
608     errout:
609 	mutex_enter(&lfs_lock);
610 	KASSERT(lfs_writing);
611 	lfs_writing = 0;
612 	wakeup(&lfs_writing);
613 }
614 
615 #define INOCOUNT(fs) howmany((fs)->lfs_uinodes, INOPB(fs))
616 #define INOBYTES(fs) ((fs)->lfs_uinodes * sizeof (struct ufs1_dinode))
617 
618 /*
619  * make sure that we don't have too many locked buffers.
620  * flush buffers if needed.
621  */
622 int
623 lfs_check(struct vnode *vp, daddr_t blkno, int flags)
624 {
625 	int error;
626 	struct lfs *fs;
627 	struct inode *ip;
628 	extern pid_t lfs_writer_daemon;
629 
630 	error = 0;
631 	ip = VTOI(vp);
632 
633 	/* If out of buffers, wait on writer */
634 	/* XXX KS - if it's the Ifile, we're probably the cleaner! */
635 	if (ip->i_number == LFS_IFILE_INUM)
636 		return 0;
637 	/* If we're being called from inside a dirop, don't sleep */
638 	if (ip->i_flag & IN_ADIROP)
639 		return 0;
640 
641 	fs = ip->i_lfs;
642 
643 	ASSERT_NO_SEGLOCK(fs);
644 
645 	/*
646 	 * If we would flush below, but dirops are active, sleep.
647 	 * Note that a dirop cannot ever reach this code!
648 	 */
649 	mutex_enter(&lfs_lock);
650 	while (fs->lfs_dirops > 0 &&
651 	       (locked_queue_count + INOCOUNT(fs) > LFS_MAX_BUFS ||
652 		locked_queue_bytes + INOBYTES(fs) > LFS_MAX_BYTES ||
653 		lfs_subsys_pages > LFS_MAX_PAGES ||
654 		fs->lfs_dirvcount > LFS_MAX_FSDIROP(fs) ||
655 		lfs_dirvcount > LFS_MAX_DIROP || fs->lfs_diropwait > 0))
656 	{
657 		++fs->lfs_diropwait;
658 		mtsleep(&fs->lfs_writer, PRIBIO+1, "bufdirop", 0,
659 			&lfs_lock);
660 		--fs->lfs_diropwait;
661 	}
662 
663 #ifdef DEBUG
664 	if (locked_queue_count + INOCOUNT(fs) > LFS_MAX_BUFS)
665 		DLOG((DLOG_FLUSH, "lfs_check: lqc = %d, max %d\n",
666 		      locked_queue_count + INOCOUNT(fs), LFS_MAX_BUFS));
667 	if (locked_queue_bytes + INOBYTES(fs) > LFS_MAX_BYTES)
668 		DLOG((DLOG_FLUSH, "lfs_check: lqb = %ld, max %ld\n",
669 		      locked_queue_bytes + INOBYTES(fs), LFS_MAX_BYTES));
670 	if (lfs_subsys_pages > LFS_MAX_PAGES)
671 		DLOG((DLOG_FLUSH, "lfs_check: lssp = %d, max %d\n",
672 		      lfs_subsys_pages, LFS_MAX_PAGES));
673 	if (lfs_fs_pagetrip && fs->lfs_pages > lfs_fs_pagetrip)
674 		DLOG((DLOG_FLUSH, "lfs_check: fssp = %d, trip at %d\n",
675 		      fs->lfs_pages, lfs_fs_pagetrip));
676 	if (lfs_dirvcount > LFS_MAX_DIROP)
677 		DLOG((DLOG_FLUSH, "lfs_check: ldvc = %d, max %d\n",
678 		      lfs_dirvcount, LFS_MAX_DIROP));
679 	if (fs->lfs_dirvcount > LFS_MAX_FSDIROP(fs))
680 		DLOG((DLOG_FLUSH, "lfs_check: lfdvc = %d, max %d\n",
681 		      fs->lfs_dirvcount, LFS_MAX_FSDIROP(fs)));
682 	if (fs->lfs_diropwait > 0)
683 		DLOG((DLOG_FLUSH, "lfs_check: ldvw = %d\n",
684 		      fs->lfs_diropwait));
685 #endif
686 
687 	/* If there are too many pending dirops, we have to flush them. */
688 	if (fs->lfs_dirvcount > LFS_MAX_FSDIROP(fs) ||
689 	    lfs_dirvcount > LFS_MAX_DIROP || fs->lfs_diropwait > 0) {
690 		flags |= SEGM_CKP;
691 	}
692 
693 	if (locked_queue_count + INOCOUNT(fs) > LFS_MAX_BUFS ||
694 	    locked_queue_bytes + INOBYTES(fs) > LFS_MAX_BYTES ||
695 	    lfs_subsys_pages > LFS_MAX_PAGES ||
696 	    fs->lfs_dirvcount > LFS_MAX_FSDIROP(fs) ||
697 	    lfs_dirvcount > LFS_MAX_DIROP || fs->lfs_diropwait > 0) {
698 		lfs_flush(fs, flags, 0);
699 	} else if (lfs_fs_pagetrip && fs->lfs_pages > lfs_fs_pagetrip) {
700 		/*
701 		 * If we didn't flush the whole thing, some filesystems
702 		 * still might want to be flushed.
703 		 */
704 		++fs->lfs_pdflush;
705 		wakeup(&lfs_writer_daemon);
706 	}
707 
708 	while (locked_queue_count + INOCOUNT(fs) > LFS_WAIT_BUFS ||
709 		locked_queue_bytes + INOBYTES(fs) > LFS_WAIT_BYTES ||
710 		lfs_subsys_pages > LFS_WAIT_PAGES ||
711 		fs->lfs_dirvcount > LFS_MAX_FSDIROP(fs) ||
712 		lfs_dirvcount > LFS_MAX_DIROP) {
713 
714 		if (lfs_dostats)
715 			++lfs_stats.wait_exceeded;
716 		DLOG((DLOG_AVAIL, "lfs_check: waiting: count=%d, bytes=%ld\n",
717 		      locked_queue_count, locked_queue_bytes));
718 		error = cv_timedwait_sig(&locked_queue_cv, &lfs_lock,
719 		    hz * LFS_BUFWAIT);
720 		if (error != EWOULDBLOCK)
721 			break;
722 
723 		/*
724 		 * lfs_flush might not flush all the buffers, if some of the
725 		 * inodes were locked or if most of them were Ifile blocks
726 		 * and we weren't asked to checkpoint.	Try flushing again
727 		 * to keep us from blocking indefinitely.
728 		 */
729 		if (locked_queue_count + INOCOUNT(fs) > LFS_MAX_BUFS ||
730 		    locked_queue_bytes + INOBYTES(fs) > LFS_MAX_BYTES) {
731 			lfs_flush(fs, flags | SEGM_CKP, 0);
732 		}
733 	}
734 	mutex_exit(&lfs_lock);
735 	return (error);
736 }
737 
738 /*
739  * Allocate a new buffer header.
740  */
741 struct buf *
742 lfs_newbuf(struct lfs *fs, struct vnode *vp, daddr_t daddr, size_t size, int type)
743 {
744 	struct buf *bp;
745 	size_t nbytes;
746 
747 	ASSERT_MAYBE_SEGLOCK(fs);
748 	nbytes = roundup(size, fsbtob(fs, 1));
749 
750 	bp = getiobuf(NULL, true);
751 	if (nbytes) {
752 		bp->b_data = lfs_malloc(fs, nbytes, type);
753 		/* memset(bp->b_data, 0, nbytes); */
754 	}
755 #ifdef DIAGNOSTIC
756 	if (vp == NULL)
757 		panic("vp is NULL in lfs_newbuf");
758 	if (bp == NULL)
759 		panic("bp is NULL after malloc in lfs_newbuf");
760 #endif
761 
762 	bp->b_bufsize = size;
763 	bp->b_bcount = size;
764 	bp->b_lblkno = daddr;
765 	bp->b_blkno = daddr;
766 	bp->b_error = 0;
767 	bp->b_resid = 0;
768 	bp->b_iodone = lfs_callback;
769 	bp->b_cflags = BC_BUSY | BC_NOCACHE;
770 	bp->b_private = fs;
771 
772 	mutex_enter(&bufcache_lock);
773 	mutex_enter(&vp->v_interlock);
774 	bgetvp(vp, bp);
775 	mutex_exit(&vp->v_interlock);
776 	mutex_exit(&bufcache_lock);
777 
778 	return (bp);
779 }
780 
781 void
782 lfs_freebuf(struct lfs *fs, struct buf *bp)
783 {
784 	struct vnode *vp;
785 
786 	if ((vp = bp->b_vp) != NULL) {
787 		mutex_enter(&bufcache_lock);
788 		mutex_enter(&vp->v_interlock);
789 		brelvp(bp);
790 		mutex_exit(&vp->v_interlock);
791 		mutex_exit(&bufcache_lock);
792 	}
793 	if (!(bp->b_cflags & BC_INVAL)) { /* BC_INVAL indicates a "fake" buffer */
794 		lfs_free(fs, bp->b_data, LFS_NB_UNKNOWN);
795 		bp->b_data = NULL;
796 	}
797 	putiobuf(bp);
798 }
799 
800 /*
801  * Count buffers on the "locked" queue, and compare it to a pro-forma count.
802  * Don't count malloced buffers, since they don't detract from the total.
803  */
804 void
805 lfs_countlocked(int *count, long *bytes, const char *msg)
806 {
807 	struct buf *bp;
808 	int n = 0;
809 	long int size = 0L;
810 
811 	mutex_enter(&bufcache_lock);
812 	TAILQ_FOREACH(bp, &bufqueues[BQ_LOCKED].bq_queue, b_freelist) {
813 		KASSERT(bp->b_iodone == NULL);
814 		n++;
815 		size += bp->b_bufsize;
816 #ifdef DIAGNOSTIC
817 		if (n > nbuf)
818 			panic("lfs_countlocked: this can't happen: more"
819 			      " buffers locked than exist");
820 #endif
821 	}
822 	/*
823 	 * Theoretically this function never really does anything.
824 	 * Give a warning if we have to fix the accounting.
825 	 */
826 	if (n != *count) {
827 		DLOG((DLOG_LLIST, "lfs_countlocked: %s: adjusted buf count"
828 		      " from %d to %d\n", msg, *count, n));
829 	}
830 	if (size != *bytes) {
831 		DLOG((DLOG_LLIST, "lfs_countlocked: %s: adjusted byte count"
832 		      " from %ld to %ld\n", msg, *bytes, size));
833 	}
834 	*count = n;
835 	*bytes = size;
836 	mutex_exit(&bufcache_lock);
837 	return;
838 }
839 
840 int
841 lfs_wait_pages(void)
842 {
843 	int active, inactive;
844 
845 	uvm_estimatepageable(&active, &inactive);
846 	return LFS_WAIT_RESOURCE(active + inactive + uvmexp.free, 1);
847 }
848 
849 int
850 lfs_max_pages(void)
851 {
852 	int active, inactive;
853 
854 	uvm_estimatepageable(&active, &inactive);
855 	return LFS_MAX_RESOURCE(active + inactive + uvmexp.free, 1);
856 }
857