xref: /netbsd-src/sys/ufs/lfs/lfs_bio.c (revision 9fc453562f6ebe8eabdfd51e21ae0a0058906d4f)
1 /*	$NetBSD: lfs_bio.c,v 1.149 2020/09/05 16:30:13 riastradh 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 1.149 2020/09/05 16:30:13 riastradh 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/lfs/ulfs_inode.h>
76 #include <ufs/lfs/ulfsmount.h>
77 #include <ufs/lfs/ulfs_extern.h>
78 
79 #include <ufs/lfs/lfs.h>
80 #include <ufs/lfs/lfs_accessors.h>
81 #include <ufs/lfs/lfs_extern.h>
82 #include <ufs/lfs/lfs_kernel.h>
83 
84 #include <uvm/uvm_extern.h>
85 
86 /*
87  * LFS block write function.
88  *
89  * XXX
90  * No write cost accounting is done.
91  * This is almost certainly wrong for synchronous operations and NFS.
92  *
93  * protected by lfs_lock.
94  */
95 int	locked_queue_count   = 0;	/* Count of locked-down buffers. */
96 long	locked_queue_bytes   = 0L;	/* Total size of locked buffers. */
97 int	lfs_subsys_pages     = 0L;	/* Total number LFS-written pages */
98 int	lfs_fs_pagetrip	     = 0;	/* # of pages to trip per-fs write */
99 int	lfs_writing	     = 0;	/* Set if already kicked off a writer
100 					   because of buffer space */
101 int	locked_queue_waiters = 0;	/* Number of processes waiting on lq */
102 
103 /* Lock and condition variables for above. */
104 kcondvar_t	locked_queue_cv;
105 kcondvar_t	lfs_writing_cv;
106 kmutex_t	lfs_lock;
107 
108 extern int lfs_dostats;
109 
110 /*
111  * reserved number/bytes of locked buffers
112  */
113 int locked_queue_rcount = 0;
114 long locked_queue_rbytes = 0L;
115 
116 static int lfs_fits_buf(struct lfs *, int, int);
117 static int lfs_reservebuf(struct lfs *, struct vnode *vp, struct vnode *vp2,
118     int, int);
119 static int lfs_reserveavail(struct lfs *, struct vnode *vp, struct vnode *vp2,
120     int);
121 
122 static int
lfs_fits_buf(struct lfs * fs,int n,int bytes)123 lfs_fits_buf(struct lfs *fs, int n, int bytes)
124 {
125 	int count_fit, bytes_fit;
126 
127 	ASSERT_NO_SEGLOCK(fs);
128 	KASSERT(mutex_owned(&lfs_lock));
129 
130 	count_fit =
131 	    (locked_queue_count + locked_queue_rcount + n <= LFS_WAIT_BUFS);
132 	bytes_fit =
133 	    (locked_queue_bytes + locked_queue_rbytes + bytes <= LFS_WAIT_BYTES);
134 
135 #ifdef DEBUG
136 	if (!count_fit) {
137 		DLOG((DLOG_AVAIL, "lfs_fits_buf: no fit count: %d + %d + %d >= %d\n",
138 		      locked_queue_count, locked_queue_rcount,
139 		      n, LFS_WAIT_BUFS));
140 	}
141 	if (!bytes_fit) {
142 		DLOG((DLOG_AVAIL, "lfs_fits_buf: no fit bytes: %ld + %ld + %d >= %ld\n",
143 		      locked_queue_bytes, locked_queue_rbytes,
144 		      bytes, LFS_WAIT_BYTES));
145 	}
146 #endif /* DEBUG */
147 
148 	return (count_fit && bytes_fit);
149 }
150 
151 /* ARGSUSED */
152 static int
lfs_reservebuf(struct lfs * fs,struct vnode * vp,struct vnode * vp2,int n,int bytes)153 lfs_reservebuf(struct lfs *fs, struct vnode *vp,
154     struct vnode *vp2, int n, int bytes)
155 {
156 	int cantwait;
157 
158 	ASSERT_MAYBE_SEGLOCK(fs);
159 	KASSERT(locked_queue_rcount >= 0);
160 	KASSERT(locked_queue_rbytes >= 0);
161 
162 	cantwait = (VTOI(vp)->i_state & IN_ADIROP) || fs->lfs_unlockvp == vp;
163 	mutex_enter(&lfs_lock);
164 	while (!cantwait && n > 0 && !lfs_fits_buf(fs, n, bytes)) {
165 		int error;
166 
167 		lfs_flush(fs, 0, 0);
168 
169 		DLOG((DLOG_AVAIL, "lfs_reservebuf: waiting: count=%d, bytes=%ld\n",
170 		      locked_queue_count, locked_queue_bytes));
171 		++locked_queue_waiters;
172 		error = cv_timedwait_sig(&locked_queue_cv, &lfs_lock,
173 		    hz * LFS_BUFWAIT);
174 		--locked_queue_waiters;
175 		if (error && error != EWOULDBLOCK) {
176 			mutex_exit(&lfs_lock);
177 			return error;
178 		}
179 	}
180 
181 	locked_queue_rcount += n;
182 	locked_queue_rbytes += bytes;
183 
184 	if (n < 0 && locked_queue_waiters > 0) {
185 		DLOG((DLOG_AVAIL, "lfs_reservebuf: broadcast: count=%d, bytes=%ld\n",
186 		      locked_queue_count, locked_queue_bytes));
187 		cv_broadcast(&locked_queue_cv);
188 	}
189 
190 	mutex_exit(&lfs_lock);
191 
192 	KASSERT(locked_queue_rcount >= 0);
193 	KASSERT(locked_queue_rbytes >= 0);
194 
195 	return 0;
196 }
197 
198 /*
199  * Try to reserve some blocks, prior to performing a sensitive operation that
200  * requires the vnode lock to be honored.  If there is not enough space, wait
201  * for the space to become available.
202  *
203  * Called with vp locked.  (Note nowever that if fsb < 0, vp is ignored.)
204  */
205 static int
lfs_reserveavail(struct lfs * fs,struct vnode * vp,struct vnode * vp2,int fsb)206 lfs_reserveavail(struct lfs *fs, struct vnode *vp,
207     struct vnode *vp2, int fsb)
208 {
209 	CLEANERINFO *cip;
210 	struct buf *bp;
211 	int error, slept;
212 	int cantwait;
213 
214 	ASSERT_MAYBE_SEGLOCK(fs);
215 	slept = 0;
216 	mutex_enter(&lfs_lock);
217 	cantwait = (VTOI(vp)->i_state & IN_ADIROP) || fs->lfs_unlockvp == vp;
218 	while (!cantwait && fsb > 0 &&
219 	       !lfs_fits(fs, fsb + fs->lfs_ravail + fs->lfs_favail)) {
220 		mutex_exit(&lfs_lock);
221 
222 		if (!slept) {
223 			DLOG((DLOG_AVAIL, "lfs_reserve: waiting for %ld (bfree = %jd,"
224 			      " est_bfree = %jd)\n",
225 			      fsb + fs->lfs_ravail + fs->lfs_favail,
226 			      (intmax_t)lfs_sb_getbfree(fs),
227 			      (intmax_t)LFS_EST_BFREE(fs)));
228 		}
229 		++slept;
230 
231 		/* Wake up the cleaner */
232 		LFS_CLEANERINFO(cip, fs, bp);
233 		LFS_SYNC_CLEANERINFO(cip, fs, bp, 0);
234 		lfs_wakeup_cleaner(fs);
235 
236 		mutex_enter(&lfs_lock);
237 		/* Cleaner might have run while we were reading, check again */
238 		if (lfs_fits(fs, fsb + fs->lfs_ravail + fs->lfs_favail))
239 			break;
240 
241 		error = mtsleep(&fs->lfs_availsleep, PCATCH | PUSER,
242 				"lfs_reserve", 0, &lfs_lock);
243 		if (error) {
244 			mutex_exit(&lfs_lock);
245 			return error;
246 		}
247 	}
248 #ifdef DEBUG
249 	if (slept) {
250 		DLOG((DLOG_AVAIL, "lfs_reserve: woke up\n"));
251 	}
252 #endif
253 	fs->lfs_ravail += fsb;
254 	mutex_exit(&lfs_lock);
255 
256 	return 0;
257 }
258 
259 #ifdef DIAGNOSTIC
260 int lfs_rescount;
261 int lfs_rescountdirop;
262 #endif
263 
264 int
lfs_reserve(struct lfs * fs,struct vnode * vp,struct vnode * vp2,int fsb)265 lfs_reserve(struct lfs *fs, struct vnode *vp, struct vnode *vp2, int fsb)
266 {
267 	int error;
268 
269 	ASSERT_MAYBE_SEGLOCK(fs);
270 	if (vp2) {
271 		/* Make sure we're not in the process of reclaiming vp2 */
272 		mutex_enter(&lfs_lock);
273 		while(fs->lfs_flags & LFS_UNDIROP) {
274 			mtsleep(&fs->lfs_flags, PRIBIO + 1, "lfsrundirop", 0,
275 			    &lfs_lock);
276 		}
277 		mutex_exit(&lfs_lock);
278 	}
279 
280 	KASSERT(fsb < 0 || VOP_ISLOCKED(vp));
281 	KASSERT(vp2 == NULL || fsb < 0 || VOP_ISLOCKED(vp2));
282 	KASSERT(vp2 == NULL || vp2 != fs->lfs_unlockvp);
283 
284 #ifdef DIAGNOSTIC
285 	mutex_enter(&lfs_lock);
286 	if (fsb > 0)
287 		lfs_rescount++;
288 	else if (fsb < 0)
289 		lfs_rescount--;
290 	if (lfs_rescount < 0)
291 		panic("lfs_rescount");
292 	mutex_exit(&lfs_lock);
293 #endif
294 
295 	error = lfs_reserveavail(fs, vp, vp2, fsb);
296 	if (error)
297 		return error;
298 
299 	/*
300 	 * XXX just a guess. should be more precise.
301 	 */
302 	error = lfs_reservebuf(fs, vp, vp2, fsb, lfs_fsbtob(fs, fsb));
303 	if (error)
304 		lfs_reserveavail(fs, vp, vp2, -fsb);
305 
306 	return error;
307 }
308 
309 int
lfs_max_bufs(void)310 lfs_max_bufs(void)
311 {
312 
313 	return LFS_MAX_RESOURCE(buf_nbuf(), 1);
314 }
315 
316 int
lfs_wait_bufs(void)317 lfs_wait_bufs(void)
318 {
319 
320 	return LFS_WAIT_RESOURCE(buf_nbuf(), 1);
321 }
322 
323 int
lfs_bwrite(void * v)324 lfs_bwrite(void *v)
325 {
326 	struct vop_bwrite_args /* {
327 		struct vnode *a_vp;
328 		struct buf *a_bp;
329 	} */ *ap = v;
330 	struct buf *bp = ap->a_bp;
331 
332 	KASSERTMSG((VTOI(bp->b_vp)->i_lfs->lfs_ronly ||
333 		!(bp->b_flags & B_ASYNC)),
334 	    "bawrite LFS buffer");
335 	return lfs_bwrite_ext(bp, 0);
336 }
337 
338 /*
339  * Determine if there is enough room currently available to write fsb
340  * blocks.  We need enough blocks for the new blocks, the current
341  * inode blocks (including potentially the ifile inode), a summary block,
342  * and the segment usage table, plus an ifile block.
343  */
344 int
lfs_fits(struct lfs * fs,int fsb)345 lfs_fits(struct lfs *fs, int fsb)
346 {
347 	int64_t needed;
348 
349 	ASSERT_NO_SEGLOCK(fs);
350 	needed = fsb + lfs_btofsb(fs, lfs_sb_getsumsize(fs)) +
351 		 ((howmany(lfs_sb_getuinodes(fs) + 1, LFS_INOPB(fs)) +
352 		   lfs_sb_getsegtabsz(fs) +
353 		   1) << (lfs_sb_getbshift(fs) - lfs_sb_getffshift(fs)));
354 
355 	if (needed >= lfs_sb_getavail(fs)) {
356 #ifdef DEBUG
357 		DLOG((DLOG_AVAIL, "lfs_fits: no fit: fsb = %ld, uinodes = %ld, "
358 		      "needed = %jd, avail = %jd\n",
359 		      (long)fsb, (long)lfs_sb_getuinodes(fs), (intmax_t)needed,
360 		      (intmax_t)lfs_sb_getavail(fs)));
361 #endif
362 		return 0;
363 	}
364 	return 1;
365 }
366 
367 int
lfs_availwait(struct lfs * fs,int fsb)368 lfs_availwait(struct lfs *fs, int fsb)
369 {
370 	int error;
371 	CLEANERINFO *cip;
372 	struct buf *cbp;
373 
374 	ASSERT_NO_SEGLOCK(fs);
375 	/* Push cleaner blocks through regardless */
376 	mutex_enter(&lfs_lock);
377 	if (LFS_SEGLOCK_HELD(fs) &&
378 	    fs->lfs_sp->seg_flags & (SEGM_CLEAN | SEGM_FORCE_CKP)) {
379 		mutex_exit(&lfs_lock);
380 		return 0;
381 	}
382 	mutex_exit(&lfs_lock);
383 
384 	while (!lfs_fits(fs, fsb)) {
385 		/*
386 		 * Out of space, need cleaner to run.
387 		 * Update the cleaner info, then wake it up.
388 		 * Note the cleanerinfo block is on the ifile
389 		 * so it CANT_WAIT.
390 		 */
391 		LFS_CLEANERINFO(cip, fs, cbp);
392 		LFS_SYNC_CLEANERINFO(cip, fs, cbp, 0);
393 
394 #ifdef DEBUG
395 		DLOG((DLOG_AVAIL, "lfs_availwait: out of available space, "
396 		      "waiting on cleaner\n"));
397 #endif
398 
399 		lfs_wakeup_cleaner(fs);
400 		KASSERTMSG(!LFS_SEGLOCK_HELD(fs), "lfs_availwait: deadlock");
401 		error = tsleep(&fs->lfs_availsleep, PCATCH | PUSER,
402 			       "cleaner", 0);
403 		if (error)
404 			return (error);
405 	}
406 	return 0;
407 }
408 
409 int
lfs_bwrite_ext(struct buf * bp,int flags)410 lfs_bwrite_ext(struct buf *bp, int flags)
411 {
412 	struct lfs *fs;
413 	struct inode *ip;
414 	struct vnode *vp;
415 	int fsb;
416 
417 	vp = bp->b_vp;
418 	fs = VFSTOULFS(vp->v_mount)->um_lfs;
419 
420 	ASSERT_MAYBE_SEGLOCK(fs);
421 	KASSERT(bp->b_cflags & BC_BUSY);
422 	KASSERT(flags & BW_CLEAN || !LFS_IS_MALLOC_BUF(bp));
423 	KASSERT((bp->b_flags & B_LOCKED) || !(bp->b_oflags & BO_DELWRI));
424 
425 	/*
426 	 * Don't write *any* blocks if we're mounted read-only, or
427 	 * if we are "already unmounted".
428 	 *
429 	 * In particular the cleaner can't write blocks either.
430 	 */
431 	if (fs->lfs_ronly || (lfs_sb_getpflags(fs) & LFS_PF_CLEAN)) {
432 		bp->b_oflags &= ~BO_DELWRI;
433 		bp->b_flags |= B_READ; /* XXX is this right? --ks */
434 		bp->b_error = 0;
435 		mutex_enter(&bufcache_lock);
436 		LFS_UNLOCK_BUF(bp);
437 		if (LFS_IS_MALLOC_BUF(bp))
438 			bp->b_cflags &= ~BC_BUSY;
439 		else
440 			brelsel(bp, 0);
441 		mutex_exit(&bufcache_lock);
442 		return (fs->lfs_ronly ? EROFS : 0);
443 	}
444 
445 	/*
446 	 * Set the delayed write flag and use reassignbuf to move the buffer
447 	 * from the clean list to the dirty one.
448 	 *
449 	 * Set the B_LOCKED flag and unlock the buffer, causing brelse to move
450 	 * the buffer onto the LOCKED free list.  This is necessary, otherwise
451 	 * getnewbuf() would try to reclaim the buffers using bawrite, which
452 	 * isn't going to work.
453 	 *
454 	 * XXX we don't let meta-data writes run out of space because they can
455 	 * come from the segment writer.  We need to make sure that there is
456 	 * enough space reserved so that there's room to write meta-data
457 	 * blocks.
458 	 */
459 	if ((bp->b_flags & B_LOCKED) == 0) {
460 		fsb = lfs_numfrags(fs, bp->b_bcount);
461 
462 		ip = VTOI(vp);
463 		mutex_enter(&lfs_lock);
464 		if (flags & BW_CLEAN) {
465 			LFS_SET_UINO(ip, IN_CLEANING);
466 		} else {
467 			LFS_SET_UINO(ip, IN_MODIFIED);
468 		}
469 		mutex_exit(&lfs_lock);
470 		lfs_sb_subavail(fs, fsb);
471 
472 		mutex_enter(&bufcache_lock);
473 		mutex_enter(vp->v_interlock);
474 		bp->b_oflags = (bp->b_oflags | BO_DELWRI) & ~BO_DONE;
475 		LFS_LOCK_BUF(bp);
476 		bp->b_flags &= ~B_READ;
477 		bp->b_error = 0;
478 		reassignbuf(bp, bp->b_vp);
479 		mutex_exit(vp->v_interlock);
480 	} else {
481 		mutex_enter(&bufcache_lock);
482 	}
483 
484 	if (bp->b_iodone != NULL)
485 		bp->b_cflags &= ~BC_BUSY;
486 	else
487 		brelsel(bp, 0);
488 	mutex_exit(&bufcache_lock);
489 
490 	return (0);
491 }
492 
493 /*
494  * Called and return with the lfs_lock held.
495  */
496 void
lfs_flush_fs(struct lfs * fs,int flags)497 lfs_flush_fs(struct lfs *fs, int flags)
498 {
499 	ASSERT_NO_SEGLOCK(fs);
500 	KASSERT(mutex_owned(&lfs_lock));
501 	if (fs->lfs_ronly)
502 		return;
503 
504 	if (lfs_dostats)
505 		++lfs_stats.flush_invoked;
506 
507 	fs->lfs_pdflush = 0;
508 	mutex_exit(&lfs_lock);
509 	lfs_writer_enter(fs, "fldirop");
510 	lfs_segwrite(fs->lfs_ivnode->v_mount, flags);
511 	lfs_writer_leave(fs);
512 	mutex_enter(&lfs_lock);
513 	fs->lfs_favail = 0; /* XXX */
514 }
515 
516 /*
517  * This routine initiates segment writes when LFS is consuming too many
518  * resources.  Ideally the pageout daemon would be able to direct LFS
519  * more subtly.
520  * XXX We have one static count of locked buffers;
521  * XXX need to think more about the multiple filesystem case.
522  *
523  * Called and return with lfs_lock held.
524  * If fs != NULL, we hold the segment lock for fs.
525  */
526 void
lfs_flush(struct lfs * fs,int flags,int only_onefs)527 lfs_flush(struct lfs *fs, int flags, int only_onefs)
528 {
529 	extern u_int64_t locked_fakequeue_count;
530 	mount_iterator_t *iter;
531 	struct mount *mp;
532 	struct lfs *tfs;
533 
534 	KASSERT(mutex_owned(&lfs_lock));
535 	KDASSERT(fs == NULL || !LFS_SEGLOCK_HELD(fs));
536 
537 	if (lfs_dostats)
538 		++lfs_stats.write_exceeded;
539 	/* XXX should we include SEGM_CKP here? */
540 	if (lfs_writing && !(flags & SEGM_SYNC)) {
541 		DLOG((DLOG_FLUSH, "lfs_flush: not flushing because another flush is active\n"));
542 		return;
543 	}
544 	while (lfs_writing)
545 		cv_wait(&lfs_writing_cv, &lfs_lock);
546 	lfs_writing = 1;
547 
548 	mutex_exit(&lfs_lock);
549 
550 	if (only_onefs) {
551 		KASSERT(fs != NULL);
552 		if (vfs_busy(fs->lfs_ivnode->v_mount))
553 			goto errout;
554 		mutex_enter(&lfs_lock);
555 		lfs_flush_fs(fs, flags);
556 		mutex_exit(&lfs_lock);
557 		vfs_unbusy(fs->lfs_ivnode->v_mount);
558 	} else {
559 		locked_fakequeue_count = 0;
560 		mountlist_iterator_init(&iter);
561 		while ((mp = mountlist_iterator_next(iter)) != NULL) {
562 			if (strncmp(&mp->mnt_stat.f_fstypename[0], MOUNT_LFS,
563 			    sizeof(mp->mnt_stat.f_fstypename)) == 0) {
564 				tfs = VFSTOULFS(mp)->um_lfs;
565 				mutex_enter(&lfs_lock);
566 				lfs_flush_fs(tfs, flags);
567 				mutex_exit(&lfs_lock);
568 			}
569 		}
570 		mountlist_iterator_destroy(iter);
571 	}
572 	wakeup(&lfs_subsys_pages);
573 
574     errout:
575 	mutex_enter(&lfs_lock);
576 	KASSERT(lfs_writing);
577 	lfs_writing = 0;
578 	wakeup(&lfs_writing);
579 }
580 
581 #define INOCOUNT(fs) howmany(lfs_sb_getuinodes(fs), LFS_INOPB(fs))
582 #define INOBYTES(fs) (lfs_sb_getuinodes(fs) * DINOSIZE(fs))
583 
584 /*
585  * make sure that we don't have too many locked buffers.
586  * flush buffers if needed.
587  */
588 int
lfs_check(struct vnode * vp,daddr_t blkno,int flags)589 lfs_check(struct vnode *vp, daddr_t blkno, int flags)
590 {
591 	int error;
592 	struct lfs *fs;
593 	struct inode *ip;
594 	extern kcondvar_t lfs_writerd_cv;
595 
596 	error = 0;
597 	ip = VTOI(vp);
598 
599 	/* If out of buffers, wait on writer */
600 	/* XXX KS - if it's the Ifile, we're probably the cleaner! */
601 	if (ip->i_number == LFS_IFILE_INUM)
602 		return 0;
603 	/* If we're being called from inside a dirop, don't sleep */
604 	if (ip->i_state & IN_ADIROP)
605 		return 0;
606 
607 	fs = ip->i_lfs;
608 
609 	ASSERT_NO_SEGLOCK(fs);
610 
611 	/*
612 	 * If we would flush below, but dirops are active, sleep.
613 	 * Note that a dirop cannot ever reach this code!
614 	 */
615 	mutex_enter(&lfs_lock);
616 	while (fs->lfs_dirops > 0 &&
617 	       (locked_queue_count + INOCOUNT(fs) > LFS_MAX_BUFS ||
618 		locked_queue_bytes + INOBYTES(fs) > LFS_MAX_BYTES ||
619 		lfs_subsys_pages > LFS_MAX_PAGES ||
620 		fs->lfs_dirvcount > LFS_MAX_FSDIROP(fs) ||
621 		lfs_dirvcount > LFS_MAX_DIROP || fs->lfs_diropwait > 0))
622 	{
623 		++fs->lfs_diropwait;
624 		mtsleep(&fs->lfs_writer, PRIBIO+1, "bufdirop", 0,
625 			&lfs_lock);
626 		--fs->lfs_diropwait;
627 	}
628 
629 #ifdef DEBUG
630 	if (locked_queue_count + INOCOUNT(fs) > LFS_MAX_BUFS)
631 		DLOG((DLOG_FLUSH, "lfs_check: lqc = %d, max %d\n",
632 		      locked_queue_count + INOCOUNT(fs), LFS_MAX_BUFS));
633 	if (locked_queue_bytes + INOBYTES(fs) > LFS_MAX_BYTES)
634 		DLOG((DLOG_FLUSH, "lfs_check: lqb = %ld, max %ld\n",
635 		      locked_queue_bytes + INOBYTES(fs), LFS_MAX_BYTES));
636 	if (lfs_subsys_pages > LFS_MAX_PAGES)
637 		DLOG((DLOG_FLUSH, "lfs_check: lssp = %d, max %d\n",
638 		      lfs_subsys_pages, LFS_MAX_PAGES));
639 	if (lfs_fs_pagetrip && fs->lfs_pages > lfs_fs_pagetrip)
640 		DLOG((DLOG_FLUSH, "lfs_check: fssp = %d, trip at %d\n",
641 		      fs->lfs_pages, lfs_fs_pagetrip));
642 	if (lfs_dirvcount > LFS_MAX_DIROP)
643 		DLOG((DLOG_FLUSH, "lfs_check: ldvc = %d, max %d\n",
644 		      lfs_dirvcount, LFS_MAX_DIROP));
645 	if (fs->lfs_dirvcount > LFS_MAX_FSDIROP(fs))
646 		DLOG((DLOG_FLUSH, "lfs_check: lfdvc = %d, max %d\n",
647 		      fs->lfs_dirvcount, LFS_MAX_FSDIROP(fs)));
648 	if (fs->lfs_diropwait > 0)
649 		DLOG((DLOG_FLUSH, "lfs_check: ldvw = %d\n",
650 		      fs->lfs_diropwait));
651 #endif
652 
653 	/* If there are too many pending dirops, we have to flush them. */
654 	if (fs->lfs_dirvcount > LFS_MAX_FSDIROP(fs) ||
655 	    lfs_dirvcount > LFS_MAX_DIROP || fs->lfs_diropwait > 0) {
656 		KASSERT(fs->lfs_dirops == 0);
657 		fs->lfs_writer++;
658 		mutex_exit(&lfs_lock);
659 		lfs_flush_dirops(fs);
660 		mutex_enter(&lfs_lock);
661 		if (--fs->lfs_writer == 0)
662 			cv_broadcast(&fs->lfs_diropscv);
663 		KASSERT(fs->lfs_dirops == 0);
664 	} else if (locked_queue_count + INOCOUNT(fs) > LFS_MAX_BUFS ||
665 	    locked_queue_bytes + INOBYTES(fs) > LFS_MAX_BYTES ||
666 	    lfs_subsys_pages > LFS_MAX_PAGES ||
667 	    fs->lfs_dirvcount > LFS_MAX_FSDIROP(fs) ||
668 	    lfs_dirvcount > LFS_MAX_DIROP || fs->lfs_diropwait > 0) {
669 		lfs_flush(fs, flags, 0);
670 	} else if (lfs_fs_pagetrip && fs->lfs_pages > lfs_fs_pagetrip) {
671 		/*
672 		 * If we didn't flush the whole thing, some filesystems
673 		 * still might want to be flushed.
674 		 */
675 		++fs->lfs_pdflush;
676 		cv_broadcast(&lfs_writerd_cv);
677 	}
678 
679 	while (locked_queue_count + INOCOUNT(fs) >= LFS_WAIT_BUFS ||
680 		locked_queue_bytes + INOBYTES(fs) >= LFS_WAIT_BYTES ||
681 		lfs_subsys_pages > LFS_WAIT_PAGES ||
682 		fs->lfs_dirvcount > LFS_MAX_FSDIROP(fs) ||
683 		lfs_dirvcount > LFS_MAX_DIROP) {
684 
685 		if (lfs_dostats)
686 			++lfs_stats.wait_exceeded;
687 		DLOG((DLOG_AVAIL, "lfs_check: waiting: count=%d, bytes=%ld\n",
688 		      locked_queue_count, locked_queue_bytes));
689 		++locked_queue_waiters;
690 		error = cv_timedwait_sig(&locked_queue_cv, &lfs_lock,
691 		    hz * LFS_BUFWAIT);
692 		--locked_queue_waiters;
693 		if (error != EWOULDBLOCK)
694 			break;
695 
696 		/*
697 		 * lfs_flush might not flush all the buffers, if some of the
698 		 * inodes were locked or if most of them were Ifile blocks
699 		 * and we weren't asked to checkpoint.	Try flushing again
700 		 * to keep us from blocking indefinitely.
701 		 */
702 		if (locked_queue_count + INOCOUNT(fs) >= LFS_MAX_BUFS ||
703 		    locked_queue_bytes + INOBYTES(fs) >= LFS_MAX_BYTES) {
704 			lfs_flush(fs, flags | SEGM_CKP, 0);
705 		}
706 	}
707 	mutex_exit(&lfs_lock);
708 	return (error);
709 }
710 
711 /*
712  * Allocate a new buffer header.
713  */
714 struct buf *
lfs_newbuf(struct lfs * fs,struct vnode * vp,daddr_t daddr,size_t size,int type)715 lfs_newbuf(struct lfs *fs, struct vnode *vp, daddr_t daddr, size_t size, int type)
716 {
717 	struct buf *bp;
718 	size_t nbytes;
719 
720 	ASSERT_MAYBE_SEGLOCK(fs);
721 	nbytes = roundup(size, lfs_fsbtob(fs, 1));
722 
723 	bp = getiobuf(NULL, true);
724 	if (nbytes) {
725 		bp->b_data = lfs_malloc(fs, nbytes, type);
726 		/* memset(bp->b_data, 0, nbytes); */
727 	}
728 	KASSERT(vp != NULL);
729 	KASSERT(bp != NULL);
730 
731 	bp->b_bufsize = size;
732 	bp->b_bcount = size;
733 	bp->b_lblkno = daddr;
734 	bp->b_blkno = daddr;
735 	bp->b_error = 0;
736 	bp->b_resid = 0;
737 	bp->b_iodone = lfs_free_aiodone;
738 	bp->b_cflags |= BC_BUSY | BC_NOCACHE;
739 	bp->b_private = fs;
740 
741 	mutex_enter(&bufcache_lock);
742 	mutex_enter(vp->v_interlock);
743 	bgetvp(vp, bp);
744 	mutex_exit(vp->v_interlock);
745 	mutex_exit(&bufcache_lock);
746 
747 	return (bp);
748 }
749 
750 void
lfs_freebuf(struct lfs * fs,struct buf * bp)751 lfs_freebuf(struct lfs *fs, struct buf *bp)
752 {
753 	struct vnode *vp;
754 
755 	if ((vp = bp->b_vp) != NULL) {
756 		mutex_enter(&bufcache_lock);
757 		mutex_enter(vp->v_interlock);
758 		brelvp(bp);
759 		mutex_exit(vp->v_interlock);
760 		mutex_exit(&bufcache_lock);
761 	}
762 	if (!(bp->b_cflags & BC_INVAL)) { /* BC_INVAL indicates a "fake" buffer */
763 		lfs_free(fs, bp->b_data, LFS_NB_UNKNOWN);
764 		bp->b_data = NULL;
765 	}
766 	putiobuf(bp);
767 }
768 
769 int
lfs_wait_pages(void)770 lfs_wait_pages(void)
771 {
772 	int active, inactive;
773 
774 	uvm_estimatepageable(&active, &inactive);
775 	return LFS_WAIT_RESOURCE(active + inactive + uvm_availmem(false), 1);
776 }
777 
778 int
lfs_max_pages(void)779 lfs_max_pages(void)
780 {
781 	int active, inactive;
782 
783 	uvm_estimatepageable(&active, &inactive);
784 	return LFS_MAX_RESOURCE(active + inactive + uvm_availmem(false), 1);
785 }
786