xref: /netbsd-src/libexec/lfs_cleanerd/lfs_cleanerd.c (revision c2f76ff004a2cb67efe5b12d97bd3ef7fe89e18d)
1 /* $NetBSD: lfs_cleanerd.c,v 1.27 2010/12/23 18:08:41 mlelstv Exp $	 */
2 
3 /*-
4  * Copyright (c) 2005 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 /*
33  * The cleaner daemon for the NetBSD Log-structured File System.
34  * Only tested for use with version 2 LFSs.
35  */
36 
37 #include <sys/syslog.h>
38 #include <sys/param.h>
39 #include <sys/mount.h>
40 #include <sys/stat.h>
41 #include <ufs/ufs/inode.h>
42 #include <ufs/lfs/lfs.h>
43 
44 #include <assert.h>
45 #include <err.h>
46 #include <errno.h>
47 #include <fcntl.h>
48 #include <semaphore.h>
49 #include <stdio.h>
50 #include <stdlib.h>
51 #include <string.h>
52 #include <unistd.h>
53 #include <time.h>
54 #include <util.h>
55 
56 #include "bufcache.h"
57 #include "vnode.h"
58 #include "lfs_user.h"
59 #include "fdfs.h"
60 #include "cleaner.h"
61 #include "kernelops.h"
62 #include "mount_lfs.h"
63 
64 /*
65  * Global variables.
66  */
67 /* XXX these top few should really be fs-specific */
68 int use_fs_idle;	/* Use fs idle rather than cpu idle time */
69 int use_bytes;		/* Use bytes written rather than segments cleaned */
70 int load_threshold;	/* How idle is idle (CPU idle) */
71 int atatime;		/* How many segments (bytes) to clean at a time */
72 
73 int nfss;		/* Number of filesystems monitored by this cleanerd */
74 struct clfs **fsp;	/* Array of extended filesystem structures */
75 int segwait_timeout;	/* Time to wait in lfs_segwait() */
76 int do_quit;		/* Quit after one cleaning loop */
77 int do_coalesce;	/* Coalesce filesystem */
78 int do_small;		/* Use small writes through markv */
79 char *copylog_filename; /* File to use for fs debugging analysis */
80 int inval_segment;	/* Segment to invalidate */
81 int stat_report;	/* Report statistics for this period of cycles */
82 int debug;		/* Turn on debugging */
83 struct cleaner_stats {
84 	double	util_tot;
85 	double	util_sos;
86 	off_t	bytes_read;
87 	off_t	bytes_written;
88 	off_t	segs_cleaned;
89 	off_t	segs_empty;
90 	off_t	segs_error;
91 } cleaner_stats;
92 
93 extern u_int32_t cksum(void *, size_t);
94 extern u_int32_t lfs_sb_cksum(struct dlfs *);
95 extern u_int32_t lfs_cksum_part(void *, size_t, u_int32_t);
96 extern int ufs_getlbns(struct lfs *, struct uvnode *, daddr_t, struct indir *, int *);
97 
98 /* Compat */
99 void pwarn(const char *unused, ...) { /* Does nothing */ };
100 
101 /*
102  * Log a message if debugging is turned on.
103  */
104 void
105 dlog(const char *fmt, ...)
106 {
107 	va_list ap;
108 
109 	if (debug == 0)
110 		return;
111 
112 	va_start(ap, fmt);
113 	vsyslog(LOG_DEBUG, fmt, ap);
114 	va_end(ap);
115 }
116 
117 /*
118  * Remove the specified filesystem from the list, due to its having
119  * become unmounted or other error condition.
120  */
121 void
122 handle_error(struct clfs **cfsp, int n)
123 {
124 	syslog(LOG_NOTICE, "%s: detaching cleaner", cfsp[n]->lfs_fsmnt);
125 	free(cfsp[n]);
126 	if (n != nfss - 1)
127 		cfsp[n] = cfsp[nfss - 1];
128 	--nfss;
129 }
130 
131 /*
132  * Reinitialize a filesystem if, e.g., its size changed.
133  */
134 int
135 reinit_fs(struct clfs *fs)
136 {
137 	char fsname[MNAMELEN];
138 
139 	strncpy(fsname, (char *)fs->lfs_fsmnt, MNAMELEN);
140 	kops.ko_close(fs->clfs_ifilefd);
141 	kops.ko_close(fs->clfs_devfd);
142 	fd_reclaim(fs->clfs_devvp);
143 	fd_reclaim(fs->lfs_ivnode);
144 	free(fs->clfs_dev);
145 	free(fs->clfs_segtab);
146 	free(fs->clfs_segtabp);
147 
148 	return init_fs(fs, fsname);
149 }
150 
151 #ifdef REPAIR_ZERO_FINFO
152 /*
153  * Use fsck's lfs routines to load the Ifile from an unmounted fs.
154  * We interpret "fsname" as the name of the raw disk device.
155  */
156 int
157 init_unmounted_fs(struct clfs *fs, char *fsname)
158 {
159 	struct lfs *disc_fs;
160 	int i;
161 
162 	fs->clfs_dev = fsname;
163 	if ((fs->clfs_devfd = kops.ko_open(fs->clfs_dev, O_RDWR)) < 0) {
164 		syslog(LOG_ERR, "couldn't open device %s read/write",
165 		       fs->clfs_dev);
166 		return -1;
167 	}
168 
169 	disc_fs = lfs_init(fs->clfs_devfd, 0, 0, 0, 0);
170 
171 	fs->lfs_dlfs = disc_fs->lfs_dlfs; /* Structure copy */
172 	strncpy(fs->lfs_fsmnt, fsname, MNAMELEN);
173 	fs->lfs_ivnode = (struct uvnode *)disc_fs->lfs_ivnode;
174 	fs->clfs_devvp = fd_vget(fs->clfs_devfd, fs->lfs_fsize, fs->lfs_ssize,
175 				 atatime);
176 
177 	/* Allocate and clear segtab */
178 	fs->clfs_segtab = (struct clfs_seguse *)malloc(fs->lfs_nseg *
179 						sizeof(*fs->clfs_segtab));
180 	fs->clfs_segtabp = (struct clfs_seguse **)malloc(fs->lfs_nseg *
181 						sizeof(*fs->clfs_segtabp));
182 	for (i = 0; i < fs->lfs_nseg; i++) {
183 		fs->clfs_segtabp[i] = &(fs->clfs_segtab[i]);
184 		fs->clfs_segtab[i].flags = 0x0;
185 	}
186 	syslog(LOG_NOTICE, "%s: unmounted cleaner starting", fsname);
187 
188 	return 0;
189 }
190 #endif
191 
192 /*
193  * Set up the file descriptors, including the Ifile descriptor.
194  * If we can't get the Ifile, this is not an LFS (or the kernel is
195  * too old to support the fcntl).
196  * XXX Merge this and init_unmounted_fs, switching on whether
197  * XXX "fsname" is a dir or a char special device.  Should
198  * XXX also be able to read unmounted devices out of fstab, the way
199  * XXX fsck does.
200  */
201 int
202 init_fs(struct clfs *fs, char *fsname)
203 {
204 	struct statvfs sf;
205 	int rootfd;
206 	int i;
207 	void *sbuf;
208 	char *bn;
209 
210 	/*
211 	 * Get the raw device from the block device.
212 	 * XXX this is ugly.  Is there a way to discover the raw device
213 	 * XXX for a given mount point?
214 	 */
215 	if (kops.ko_statvfs(fsname, &sf, ST_WAIT) < 0)
216 		return -1;
217 	fs->clfs_dev = malloc(strlen(sf.f_mntfromname) + 2);
218 	if (fs->clfs_dev == NULL) {
219 		syslog(LOG_ERR, "couldn't malloc device name string: %m");
220 		return -1;
221 	}
222 	bn = strrchr(sf.f_mntfromname, '/');
223 	bn = bn ? bn+1 : sf.f_mntfromname;
224 	strlcpy(fs->clfs_dev, sf.f_mntfromname, bn - sf.f_mntfromname + 1);
225 	strcat(fs->clfs_dev, "r");
226 	strcat(fs->clfs_dev, bn);
227 	if ((fs->clfs_devfd = kops.ko_open(fs->clfs_dev, O_RDONLY, 0)) < 0) {
228 		syslog(LOG_ERR, "couldn't open device %s for reading",
229 			fs->clfs_dev);
230 		return -1;
231 	}
232 
233 	/* Find the Ifile and open it */
234 	if ((rootfd = kops.ko_open(fsname, O_RDONLY, 0)) < 0)
235 		return -2;
236 	if (kops.ko_fcntl(rootfd, LFCNIFILEFH, &fs->clfs_ifilefh) < 0)
237 		return -3;
238 	if ((fs->clfs_ifilefd = kops.ko_fhopen(&fs->clfs_ifilefh,
239 	    sizeof(fs->clfs_ifilefh), O_RDONLY)) < 0)
240 		return -4;
241 	kops.ko_close(rootfd);
242 
243 	sbuf = malloc(LFS_SBPAD);
244 	if (sbuf == NULL) {
245 		syslog(LOG_ERR, "couldn't malloc superblock buffer");
246 		return -1;
247 	}
248 
249 	/* Load in the superblock */
250 	if (kops.ko_pread(fs->clfs_devfd, sbuf, LFS_SBPAD, LFS_LABELPAD) < 0) {
251 		free(sbuf);
252 		return -1;
253 	}
254 
255 	memcpy(&(fs->lfs_dlfs), sbuf, sizeof(struct dlfs));
256 	free(sbuf);
257 
258 	/* If this is not a version 2 filesystem, complain and exit */
259 	if (fs->lfs_version != 2) {
260 		syslog(LOG_ERR, "%s: not a version 2 LFS", fsname);
261 		return -1;
262 	}
263 
264 	/* Assume fsname is the mounted name */
265 	strncpy((char *)fs->lfs_fsmnt, fsname, MNAMELEN);
266 
267 	/* Set up vnodes for Ifile and raw device */
268 	fs->lfs_ivnode = fd_vget(fs->clfs_ifilefd, fs->lfs_bsize, 0, 0);
269 	fs->clfs_devvp = fd_vget(fs->clfs_devfd, fs->lfs_fsize, fs->lfs_ssize,
270 				 atatime);
271 
272 	/* Allocate and clear segtab */
273 	fs->clfs_segtab = (struct clfs_seguse *)malloc(fs->lfs_nseg *
274 						sizeof(*fs->clfs_segtab));
275 	fs->clfs_segtabp = (struct clfs_seguse **)malloc(fs->lfs_nseg *
276 						sizeof(*fs->clfs_segtabp));
277 	if (fs->clfs_segtab == NULL || fs->clfs_segtabp == NULL) {
278 		syslog(LOG_ERR, "%s: couldn't malloc segment table: %m",
279 			fs->clfs_dev);
280 		return -1;
281 	}
282 
283 	for (i = 0; i < fs->lfs_nseg; i++) {
284 		fs->clfs_segtabp[i] = &(fs->clfs_segtab[i]);
285 		fs->clfs_segtab[i].flags = 0x0;
286 	}
287 
288 	syslog(LOG_NOTICE, "%s: attaching cleaner", fsname);
289 	return 0;
290 }
291 
292 /*
293  * Invalidate all the currently held Ifile blocks so they will be
294  * reread when we clean.  Check the size while we're at it, and
295  * resize the buffer cache if necessary.
296  */
297 void
298 reload_ifile(struct clfs *fs)
299 {
300 	struct ubuf *bp;
301 	struct stat st;
302 	int ohashmax;
303 	extern int hashmax;
304 
305 	while ((bp = LIST_FIRST(&fs->lfs_ivnode->v_dirtyblkhd)) != NULL) {
306 		bremfree(bp);
307 		buf_destroy(bp);
308 	}
309 	while ((bp = LIST_FIRST(&fs->lfs_ivnode->v_cleanblkhd)) != NULL) {
310 		bremfree(bp);
311 		buf_destroy(bp);
312 	}
313 
314 	/* If Ifile is larger than buffer cache, rehash */
315 	fstat(fs->clfs_ifilefd, &st);
316 	if (st.st_size / fs->lfs_bsize > hashmax) {
317 		ohashmax = hashmax;
318 		bufrehash(st.st_size / fs->lfs_bsize);
319 		dlog("%s: resized buffer hash from %d to %d",
320 		     fs->lfs_fsmnt, ohashmax, hashmax);
321 	}
322 }
323 
324 /*
325  * Get IFILE entry for the given inode, store in ifpp.	The buffer
326  * which contains that data is returned in bpp, and must be brelse()d
327  * by the caller.
328  */
329 void
330 lfs_ientry(IFILE **ifpp, struct clfs *fs, ino_t ino, struct ubuf **bpp)
331 {
332 	int error;
333 
334 	error = bread(fs->lfs_ivnode, ino / fs->lfs_ifpb + fs->lfs_cleansz +
335 		      fs->lfs_segtabsz, fs->lfs_bsize, NOCRED, 0, bpp);
336 	if (error)
337 		syslog(LOG_ERR, "%s: ientry failed for ino %d",
338 			fs->lfs_fsmnt, (int)ino);
339 	*ifpp = (IFILE *)(*bpp)->b_data + ino % fs->lfs_ifpb;
340 	return;
341 }
342 
343 #ifdef TEST_PATTERN
344 /*
345  * Check ROOTINO for file data.	 The assumption is that we are running
346  * the "twofiles" test with the rest of the filesystem empty.  Files
347  * created by "twofiles" match the test pattern, but ROOTINO and the
348  * executable itself (assumed to be inode 3) should not match.
349  */
350 static void
351 check_test_pattern(BLOCK_INFO *bip)
352 {
353 	int j;
354 	unsigned char *cp = bip->bi_bp;
355 
356 	/* Check inode sanity */
357 	if (bip->bi_lbn == LFS_UNUSED_LBN) {
358 		assert(((struct ufs1_dinode *)bip->bi_bp)->di_inumber ==
359 			bip->bi_inode);
360 	}
361 
362 	/* These can have the test pattern and it's all good */
363 	if (bip->bi_inode > 3)
364 		return;
365 
366 	for (j = 0; j < bip->bi_size; j++) {
367 		if (cp[j] != (j & 0xff))
368 			break;
369 	}
370 	assert(j < bip->bi_size);
371 }
372 #endif /* TEST_PATTERN */
373 
374 /*
375  * Parse the partial segment at daddr, adding its information to
376  * bip.	 Return the address of the next partial segment to read.
377  */
378 int32_t
379 parse_pseg(struct clfs *fs, daddr_t daddr, BLOCK_INFO **bipp, int *bic)
380 {
381 	SEGSUM *ssp;
382 	IFILE *ifp;
383 	BLOCK_INFO *bip, *nbip;
384 	int32_t *iaddrp, idaddr, odaddr;
385 	FINFO *fip;
386 	struct ubuf *ifbp;
387 	struct ufs1_dinode *dip;
388 	u_int32_t ck, vers;
389 	int fic, inoc, obic;
390 	int i;
391 	char *cp;
392 
393 	odaddr = daddr;
394 	obic = *bic;
395 	bip = *bipp;
396 
397 	/*
398 	 * Retrieve the segment header, set up the SEGSUM pointer
399 	 * as well as the first FINFO and inode address pointer.
400 	 */
401 	cp = fd_ptrget(fs->clfs_devvp, daddr);
402 	ssp = (SEGSUM *)cp;
403 	iaddrp = ((int32_t *)(cp + fs->lfs_ibsize)) - 1;
404 	fip = (FINFO *)(cp + sizeof(SEGSUM));
405 
406 	/*
407 	 * Check segment header magic and checksum
408 	 */
409 	if (ssp->ss_magic != SS_MAGIC) {
410 		syslog(LOG_WARNING, "%s: sumsum magic number bad at 0x%x:"
411 		       " read 0x%x, expected 0x%x", fs->lfs_fsmnt,
412 		       (int32_t)daddr, ssp->ss_magic, SS_MAGIC);
413 		return 0x0;
414 	}
415 	ck = cksum(&ssp->ss_datasum, fs->lfs_sumsize - sizeof(ssp->ss_sumsum));
416 	if (ck != ssp->ss_sumsum) {
417 		syslog(LOG_WARNING, "%s: sumsum checksum mismatch at 0x%x:"
418 		       " read 0x%x, computed 0x%x", fs->lfs_fsmnt,
419 		       (int32_t)daddr, ssp->ss_sumsum, ck);
420 		return 0x0;
421 	}
422 
423 	/* Initialize data sum */
424 	ck = 0;
425 
426 	/* Point daddr at next block after segment summary */
427 	++daddr;
428 
429 	/*
430 	 * Loop over file info and inode pointers.  We always move daddr
431 	 * forward here because we are also computing the data checksum
432 	 * as we go.
433 	 */
434 	fic = inoc = 0;
435 	while (fic < ssp->ss_nfinfo || inoc < ssp->ss_ninos) {
436 		/*
437 		 * We must have either a file block or an inode block.
438 		 * If we don't have either one, it's an error.
439 		 */
440 		if (fic >= ssp->ss_nfinfo && *iaddrp != daddr) {
441 			syslog(LOG_WARNING, "%s: bad pseg at %x (seg %d)",
442 			       fs->lfs_fsmnt, odaddr, dtosn(fs, odaddr));
443 			*bipp = bip;
444 			return 0x0;
445 		}
446 
447 		/*
448 		 * Note each inode from the inode blocks
449 		 */
450 		if (inoc < ssp->ss_ninos && *iaddrp == daddr) {
451 			cp = fd_ptrget(fs->clfs_devvp, daddr);
452 			ck = lfs_cksum_part(cp, sizeof(u_int32_t), ck);
453 			dip = (struct ufs1_dinode *)cp;
454 			for (i = 0; i < fs->lfs_inopb; i++) {
455 				if (dip[i].di_inumber == 0)
456 					break;
457 
458 				/*
459 				 * Check currency before adding it
460 				 */
461 #ifndef REPAIR_ZERO_FINFO
462 				lfs_ientry(&ifp, fs, dip[i].di_inumber, &ifbp);
463 				idaddr = ifp->if_daddr;
464 				brelse(ifbp, 0);
465 				if (idaddr != daddr)
466 #endif
467 					continue;
468 
469 				/*
470 				 * A current inode.  Add it.
471 				 */
472 				++*bic;
473 				nbip = (BLOCK_INFO *)realloc(bip, *bic *
474 							     sizeof(*bip));
475 				if (nbip)
476 					bip = nbip;
477 				else {
478 					--*bic;
479 					*bipp = bip;
480 					return 0x0;
481 				}
482 				bip[*bic - 1].bi_inode = dip[i].di_inumber;
483 				bip[*bic - 1].bi_lbn = LFS_UNUSED_LBN;
484 				bip[*bic - 1].bi_daddr = daddr;
485 				bip[*bic - 1].bi_segcreate = ssp->ss_create;
486 				bip[*bic - 1].bi_version = dip[i].di_gen;
487 				bip[*bic - 1].bi_bp = &(dip[i]);
488 				bip[*bic - 1].bi_size = DINODE1_SIZE;
489 			}
490 			inoc += i;
491 			daddr += btofsb(fs, fs->lfs_ibsize);
492 			--iaddrp;
493 			continue;
494 		}
495 
496 		/*
497 		 * Note each file block from the finfo blocks
498 		 */
499 		if (fic >= ssp->ss_nfinfo)
500 			continue;
501 
502 		/* Count this finfo, whether or not we use it */
503 		++fic;
504 
505 		/*
506 		 * If this finfo has nblocks==0, it was written wrong.
507 		 * Kernels with this problem always wrote this zero-sized
508 		 * finfo last, so just ignore it.
509 		 */
510 		if (fip->fi_nblocks == 0) {
511 #ifdef REPAIR_ZERO_FINFO
512 			struct ubuf *nbp;
513 			SEGSUM *nssp;
514 
515 			syslog(LOG_WARNING, "fixing short FINFO at %x (seg %d)",
516 			       odaddr, dtosn(fs, odaddr));
517 			bread(fs->clfs_devvp, odaddr, fs->lfs_fsize,
518 			    NOCRED, 0, &nbp);
519 			nssp = (SEGSUM *)nbp->b_data;
520 			--nssp->ss_nfinfo;
521 			nssp->ss_sumsum = cksum(&nssp->ss_datasum,
522 				fs->lfs_sumsize - sizeof(nssp->ss_sumsum));
523 			bwrite(nbp);
524 #endif
525 			syslog(LOG_WARNING, "zero-length FINFO at %x (seg %d)",
526 			       odaddr, dtosn(fs, odaddr));
527 			continue;
528 		}
529 
530 		/*
531 		 * Check currency before adding blocks
532 		 */
533 #ifdef REPAIR_ZERO_FINFO
534 		vers = -1;
535 #else
536 		lfs_ientry(&ifp, fs, fip->fi_ino, &ifbp);
537 		vers = ifp->if_version;
538 		brelse(ifbp, 0);
539 #endif
540 		if (vers != fip->fi_version) {
541 			size_t size;
542 
543 			/* Read all the blocks from the data summary */
544 			for (i = 0; i < fip->fi_nblocks; i++) {
545 				size = (i == fip->fi_nblocks - 1) ?
546 					fip->fi_lastlength : fs->lfs_bsize;
547 				cp = fd_ptrget(fs->clfs_devvp, daddr);
548 				ck = lfs_cksum_part(cp, sizeof(u_int32_t), ck);
549 				daddr += btofsb(fs, size);
550 			}
551 			fip = (FINFO *)(fip->fi_blocks + fip->fi_nblocks);
552 			continue;
553 		}
554 
555 		/* Add all the blocks from the finfos (current or not) */
556 		nbip = (BLOCK_INFO *)realloc(bip, (*bic + fip->fi_nblocks) *
557 					     sizeof(*bip));
558 		if (nbip)
559 			bip = nbip;
560 		else {
561 			*bipp = bip;
562 			return 0x0;
563 		}
564 
565 		for (i = 0; i < fip->fi_nblocks; i++) {
566 			bip[*bic + i].bi_inode = fip->fi_ino;
567 			bip[*bic + i].bi_lbn = fip->fi_blocks[i];
568 			bip[*bic + i].bi_daddr = daddr;
569 			bip[*bic + i].bi_segcreate = ssp->ss_create;
570 			bip[*bic + i].bi_version = fip->fi_version;
571 			bip[*bic + i].bi_size = (i == fip->fi_nblocks - 1) ?
572 				fip->fi_lastlength : fs->lfs_bsize;
573 			cp = fd_ptrget(fs->clfs_devvp, daddr);
574 			ck = lfs_cksum_part(cp, sizeof(u_int32_t), ck);
575 			bip[*bic + i].bi_bp = cp;
576 			daddr += btofsb(fs, bip[*bic + i].bi_size);
577 
578 #ifdef TEST_PATTERN
579 			check_test_pattern(bip + *bic + i); /* XXXDEBUG */
580 #endif
581 		}
582 		*bic += fip->fi_nblocks;
583 		fip = (FINFO *)(fip->fi_blocks + fip->fi_nblocks);
584 	}
585 
586 #ifndef REPAIR_ZERO_FINFO
587 	if (ssp->ss_datasum != ck) {
588 		syslog(LOG_WARNING, "%s: data checksum bad at 0x%x:"
589 		       " read 0x%x, computed 0x%x", fs->lfs_fsmnt, odaddr,
590 		       ssp->ss_datasum, ck);
591 		*bic = obic;
592 		return 0x0;
593 	}
594 #endif
595 
596 	*bipp = bip;
597 	return daddr;
598 }
599 
600 static void
601 log_segment_read(struct clfs *fs, int sn)
602 {
603         FILE *fp;
604 	char *cp;
605 
606         /*
607          * Write the segment read, and its contents, into a log file in
608          * the current directory.  We don't need to log the location of
609          * the segment, since that can be inferred from the segments up
610 	 * to this point (ss_nextseg field of the previously written segment).
611 	 *
612 	 * We can use this info later to reconstruct the filesystem at any
613 	 * given point in time for analysis, by replaying the log forward
614 	 * indexed by the segment serial numbers; but it is not suitable
615 	 * for everyday use since the copylog will be simply enormous.
616          */
617 	cp = fd_ptrget(fs->clfs_devvp, sntod(fs, sn));
618 
619         fp = fopen(copylog_filename, "ab");
620         if (fp != NULL) {
621                 if (fwrite(cp, (size_t)fs->lfs_ssize, 1, fp) != 1) {
622                         perror("writing segment to copy log");
623                 }
624         }
625         fclose(fp);
626 }
627 
628 /*
629  * Read a segment to populate the BLOCK_INFO structures.
630  * Return the number of partial segments read and parsed.
631  */
632 int
633 load_segment(struct clfs *fs, int sn, BLOCK_INFO **bipp, int *bic)
634 {
635 	int32_t daddr;
636 	int i, npseg;
637 
638 	daddr = sntod(fs, sn);
639 	if (daddr < btofsb(fs, LFS_LABELPAD))
640 		daddr = btofsb(fs, LFS_LABELPAD);
641 	for (i = 0; i < LFS_MAXNUMSB; i++) {
642 		if (fs->lfs_sboffs[i] == daddr) {
643 			daddr += btofsb(fs, LFS_SBPAD);
644 			break;
645 		}
646 	}
647 
648 	/* Preload the segment buffer */
649 	if (fd_preload(fs->clfs_devvp, sntod(fs, sn)) < 0)
650 		return -1;
651 
652 	if (copylog_filename)
653 		log_segment_read(fs, sn);
654 
655 	/* Note bytes read for stats */
656 	cleaner_stats.segs_cleaned++;
657 	cleaner_stats.bytes_read += fs->lfs_ssize;
658 	++fs->clfs_nactive;
659 
660 	npseg = 0;
661 	while(dtosn(fs, daddr) == sn &&
662 	      dtosn(fs, daddr + btofsb(fs, fs->lfs_bsize)) == sn) {
663 		daddr = parse_pseg(fs, daddr, bipp, bic);
664 		if (daddr == 0x0) {
665 			++cleaner_stats.segs_error;
666 			break;
667 		}
668 		++npseg;
669 	}
670 
671 	return npseg;
672 }
673 
674 void
675 calc_cb(struct clfs *fs, int sn, struct clfs_seguse *t)
676 {
677 	time_t now;
678 	int64_t age, benefit, cost;
679 
680 	time(&now);
681 	age = (now < t->lastmod ? 0 : now - t->lastmod);
682 
683 	/* Under no circumstances clean active or already-clean segments */
684 	if ((t->flags & SEGUSE_ACTIVE) || !(t->flags & SEGUSE_DIRTY)) {
685 		t->priority = 0;
686 		return;
687 	}
688 
689 	/*
690 	 * If the segment is empty, there is no reason to clean it.
691 	 * Clear its error condition, if any, since we are never going to
692 	 * try to parse this one.
693 	 */
694 	if (t->nbytes == 0) {
695 		t->flags &= ~SEGUSE_ERROR; /* Strip error once empty */
696 		t->priority = 0;
697 		return;
698 	}
699 
700 	if (t->flags & SEGUSE_ERROR) {	/* No good if not already empty */
701 		/* No benefit */
702 		t->priority = 0;
703 		return;
704 	}
705 
706 	if (t->nbytes > fs->lfs_ssize) {
707 		/* Another type of error */
708 		syslog(LOG_WARNING, "segment %d: bad seguse count %d",
709 		       sn, t->nbytes);
710 		t->flags |= SEGUSE_ERROR;
711 		t->priority = 0;
712 		return;
713 	}
714 
715 	/*
716 	 * The non-degenerate case.  Use Rosenblum's cost-benefit algorithm.
717 	 * Calculate the benefit from cleaning this segment (one segment,
718 	 * minus fragmentation, dirty blocks and a segment summary block)
719 	 * and weigh that against the cost (bytes read plus bytes written).
720 	 * We count the summary headers as "dirty" to avoid cleaning very
721 	 * old and very full segments.
722 	 */
723 	benefit = (int64_t)fs->lfs_ssize - t->nbytes -
724 		  (t->nsums + 1) * fs->lfs_fsize;
725 	if (fs->lfs_bsize > fs->lfs_fsize) /* fragmentation */
726 		benefit -= (fs->lfs_bsize / 2);
727 	if (benefit <= 0) {
728 		t->priority = 0;
729 		return;
730 	}
731 
732 	cost = fs->lfs_ssize + t->nbytes;
733 	t->priority = (256 * benefit * age) / cost;
734 
735 	return;
736 }
737 
738 /*
739  * Comparator for BLOCK_INFO structures.  Anything not in one of the segments
740  * we're looking at sorts higher; after that we sort first by inode number
741  * and then by block number (unsigned, i.e., negative sorts higher) *but*
742  * sort inodes before data blocks.
743  */
744 static int
745 bi_comparator(const void *va, const void *vb)
746 {
747 	const BLOCK_INFO *a, *b;
748 
749 	a = (const BLOCK_INFO *)va;
750 	b = (const BLOCK_INFO *)vb;
751 
752 	/* Check for out-of-place block */
753 	if (a->bi_segcreate == a->bi_daddr &&
754 	    b->bi_segcreate != b->bi_daddr)
755 		return -1;
756 	if (a->bi_segcreate != a->bi_daddr &&
757 	    b->bi_segcreate == b->bi_daddr)
758 		return 1;
759 	if (a->bi_size <= 0 && b->bi_size > 0)
760 		return 1;
761 	if (b->bi_size <= 0 && a->bi_size > 0)
762 		return -1;
763 
764 	/* Check inode number */
765 	if (a->bi_inode != b->bi_inode)
766 		return a->bi_inode - b->bi_inode;
767 
768 	/* Check lbn */
769 	if (a->bi_lbn == LFS_UNUSED_LBN) /* Inodes sort lower than blocks */
770 		return -1;
771 	if (b->bi_lbn == LFS_UNUSED_LBN)
772 		return 1;
773 	if ((u_int32_t)a->bi_lbn > (u_int32_t)b->bi_lbn)
774 		return 1;
775 	else
776 		return -1;
777 
778 	return 0;
779 }
780 
781 /*
782  * Comparator for sort_segments: cost-benefit equation.
783  */
784 static int
785 cb_comparator(const void *va, const void *vb)
786 {
787 	const struct clfs_seguse *a, *b;
788 
789 	a = *(const struct clfs_seguse * const *)va;
790 	b = *(const struct clfs_seguse * const *)vb;
791 	return a->priority > b->priority ? -1 : 1;
792 }
793 
794 void
795 toss_old_blocks(struct clfs *fs, BLOCK_INFO **bipp, int *bic, int *sizep)
796 {
797 	int i, r;
798 	BLOCK_INFO *bip = *bipp;
799 	struct lfs_fcntl_markv /* {
800 		BLOCK_INFO *blkiov;
801 		int blkcnt;
802 	} */ lim;
803 
804 	if (bic == 0 || bip == NULL)
805 		return;
806 
807 	/*
808 	 * Kludge: Store the disk address in segcreate so we know which
809 	 * ones to toss.
810 	 */
811 	for (i = 0; i < *bic; i++)
812 		bip[i].bi_segcreate = bip[i].bi_daddr;
813 
814 	/* Sort the blocks */
815 	heapsort(bip, *bic, sizeof(BLOCK_INFO), bi_comparator);
816 
817 	/* Use bmapv to locate the blocks */
818 	lim.blkiov = bip;
819 	lim.blkcnt = *bic;
820 	if ((r = kops.ko_fcntl(fs->clfs_ifilefd, LFCNBMAPV, &lim)) < 0) {
821 		syslog(LOG_WARNING, "%s: bmapv returned %d (%m)",
822 		       fs->lfs_fsmnt, r);
823 		return;
824 	}
825 
826 	/* Toss blocks not in this segment */
827 	heapsort(bip, *bic, sizeof(BLOCK_INFO), bi_comparator);
828 
829 	/* Get rid of stale blocks */
830 	if (sizep)
831 		*sizep = 0;
832 	for (i = 0; i < *bic; i++) {
833 		if (bip[i].bi_segcreate != bip[i].bi_daddr)
834 			break;
835 		if (sizep)
836 			*sizep += bip[i].bi_size;
837 	}
838 	*bic = i; /* XXX realloc bip? */
839 	*bipp = bip;
840 
841 	return;
842 }
843 
844 /*
845  * Clean a segment and mark it invalid.
846  */
847 int
848 invalidate_segment(struct clfs *fs, int sn)
849 {
850 	BLOCK_INFO *bip;
851 	int i, r, bic;
852 	off_t nb;
853 	double util;
854 	struct lfs_fcntl_markv /* {
855 		BLOCK_INFO *blkiov;
856 		int blkcnt;
857 	} */ lim;
858 
859 	dlog("%s: inval seg %d", fs->lfs_fsmnt, sn);
860 
861 	bip = NULL;
862 	bic = 0;
863 	fs->clfs_nactive = 0;
864 	if (load_segment(fs, sn, &bip, &bic) <= 0)
865 		return -1;
866 	toss_old_blocks(fs, &bip, &bic, NULL);
867 
868 	/* Record statistics */
869 	for (i = nb = 0; i < bic; i++)
870 		nb += bip[i].bi_size;
871 	util = ((double)nb) / (fs->clfs_nactive * fs->lfs_ssize);
872 	cleaner_stats.util_tot += util;
873 	cleaner_stats.util_sos += util * util;
874 	cleaner_stats.bytes_written += nb;
875 
876 	/*
877 	 * Use markv to move the blocks.
878 	 */
879 	lim.blkiov = bip;
880 	lim.blkcnt = bic;
881 	if ((r = kops.ko_fcntl(fs->clfs_ifilefd, LFCNMARKV, &lim)) < 0) {
882 		syslog(LOG_WARNING, "%s: markv returned %d (%m) "
883 		       "for seg %d", fs->lfs_fsmnt, r, sn);
884 		return r;
885 	}
886 
887 	/*
888 	 * Finally call invalidate to invalidate the segment.
889 	 */
890 	if ((r = kops.ko_fcntl(fs->clfs_ifilefd, LFCNINVAL, &sn)) < 0) {
891 		syslog(LOG_WARNING, "%s: inval returned %d (%m) "
892 		       "for seg %d", fs->lfs_fsmnt, r, sn);
893 		return r;
894 	}
895 
896 	return 0;
897 }
898 
899 /*
900  * Check to see if the given ino/lbn pair is represented in the BLOCK_INFO
901  * array we are sending to the kernel, or if the kernel will have to add it.
902  * The kernel will only add each such pair once, though, so keep track of
903  * previous requests in a separate "extra" BLOCK_INFO array.  Returns 1
904  * if the block needs to be added, 0 if it is already represented.
905  */
906 static int
907 check_or_add(ino_t ino, int32_t lbn, BLOCK_INFO *bip, int bic, BLOCK_INFO **ebipp, int *ebicp)
908 {
909 	BLOCK_INFO *t, *ebip = *ebipp;
910 	int ebic = *ebicp;
911 	int k;
912 
913 	for (k = 0; k < bic; k++) {
914 		if (bip[k].bi_inode != ino)
915 			break;
916 		if (bip[k].bi_lbn == lbn) {
917 			return 0;
918 		}
919 	}
920 
921 	/* Look on the list of extra blocks, too */
922 	for (k = 0; k < ebic; k++) {
923 		if (ebip[k].bi_inode == ino && ebip[k].bi_lbn == lbn) {
924 			return 0;
925 		}
926 	}
927 
928 	++ebic;
929 	t = realloc(ebip, ebic * sizeof(BLOCK_INFO));
930 	if (t == NULL)
931 		return 1; /* Note *ebipc is not updated */
932 
933 	ebip = t;
934 	ebip[ebic - 1].bi_inode = ino;
935 	ebip[ebic - 1].bi_lbn = lbn;
936 
937 	*ebipp = ebip;
938 	*ebicp = ebic;
939 	return 1;
940 }
941 
942 /*
943  * Look for indirect blocks we will have to write which are not
944  * contained in this collection of blocks.  This constitutes
945  * a hidden cleaning cost, since we are unaware of it until we
946  * have already read the segments.  Return the total cost, and fill
947  * in *ifc with the part of that cost due to rewriting the Ifile.
948  */
949 static off_t
950 check_hidden_cost(struct clfs *fs, BLOCK_INFO *bip, int bic, off_t *ifc)
951 {
952 	int start;
953 	struct indir in[NIADDR + 1];
954 	int num;
955 	int i, j, ebic;
956 	BLOCK_INFO *ebip;
957 	int32_t lbn;
958 
959 	start = 0;
960 	ebip = NULL;
961 	ebic = 0;
962 	for (i = 0; i < bic; i++) {
963 		if (i == 0 || bip[i].bi_inode != bip[start].bi_inode) {
964 			start = i;
965 			/*
966 			 * Look for IFILE blocks, unless this is the Ifile.
967 			 */
968 			if (bip[i].bi_inode != fs->lfs_ifile) {
969 				lbn = fs->lfs_cleansz + bip[i].bi_inode /
970 							fs->lfs_ifpb;
971 				*ifc += check_or_add(fs->lfs_ifile, lbn,
972 						     bip, bic, &ebip, &ebic);
973 			}
974 		}
975 		if (bip[i].bi_lbn == LFS_UNUSED_LBN)
976 			continue;
977 		if (bip[i].bi_lbn < NDADDR)
978 			continue;
979 
980 		ufs_getlbns((struct lfs *)fs, NULL, (daddr_t)bip[i].bi_lbn, in, &num);
981 		for (j = 0; j < num; j++) {
982 			check_or_add(bip[i].bi_inode, in[j].in_lbn,
983 				     bip + start, bic - start, &ebip, &ebic);
984 		}
985 	}
986 	return ebic;
987 }
988 
989 /*
990  * Select segments to clean, add blocks from these segments to a cleaning
991  * list, and send this list through lfs_markv() to move them to new
992  * locations on disk.
993  */
994 int
995 clean_fs(struct clfs *fs, CLEANERINFO *cip)
996 {
997 	int i, j, ngood, sn, bic, r, npos;
998 	int bytes, totbytes;
999 	struct ubuf *bp;
1000 	SEGUSE *sup;
1001 	static BLOCK_INFO *bip;
1002 	struct lfs_fcntl_markv /* {
1003 		BLOCK_INFO *blkiov;
1004 		int blkcnt;
1005 	} */ lim;
1006 	int mc;
1007 	BLOCK_INFO *mbip;
1008 	int inc;
1009 	off_t nb;
1010 	off_t goal;
1011 	off_t extra, if_extra;
1012 	double util;
1013 
1014 	/* Read the segment table into our private structure */
1015 	npos = 0;
1016 	for (i = 0; i < fs->lfs_nseg; i+= fs->lfs_sepb) {
1017 		bread(fs->lfs_ivnode, fs->lfs_cleansz + i / fs->lfs_sepb,
1018 		      fs->lfs_bsize, NOCRED, 0, &bp);
1019 		for (j = 0; j < fs->lfs_sepb && i + j < fs->lfs_nseg; j++) {
1020 			sup = ((SEGUSE *)bp->b_data) + j;
1021 			fs->clfs_segtab[i + j].nbytes  = sup->su_nbytes;
1022 			fs->clfs_segtab[i + j].nsums = sup->su_nsums;
1023 			fs->clfs_segtab[i + j].lastmod = sup->su_lastmod;
1024 			/* Keep error status but renew other flags */
1025 			fs->clfs_segtab[i + j].flags  &= SEGUSE_ERROR;
1026 			fs->clfs_segtab[i + j].flags  |= sup->su_flags;
1027 
1028 			/* Compute cost-benefit coefficient */
1029 			calc_cb(fs, i + j, fs->clfs_segtab + i + j);
1030 			if (fs->clfs_segtab[i + j].priority > 0)
1031 				++npos;
1032 		}
1033 		brelse(bp, 0);
1034 	}
1035 
1036 	/* Sort segments based on cleanliness, fulness, and condition */
1037 	heapsort(fs->clfs_segtabp, fs->lfs_nseg, sizeof(struct clfs_seguse *),
1038 		 cb_comparator);
1039 
1040 	/* If no segment is cleanable, just return */
1041 	if (fs->clfs_segtabp[0]->priority == 0) {
1042 		dlog("%s: no segment cleanable", fs->lfs_fsmnt);
1043 		return 0;
1044 	}
1045 
1046 	/* Load some segments' blocks into bip */
1047 	bic = 0;
1048 	fs->clfs_nactive = 0;
1049 	ngood = 0;
1050 	if (use_bytes) {
1051 		/* Set attainable goal */
1052 		goal = fs->lfs_ssize * atatime;
1053 		if (goal > (cip->clean - 1) * fs->lfs_ssize / 2)
1054 			goal = MAX((cip->clean - 1) * fs->lfs_ssize,
1055 				   fs->lfs_ssize) / 2;
1056 
1057 		dlog("%s: cleaning with goal %" PRId64
1058 		     " bytes (%d segs clean, %d cleanable)",
1059 		     fs->lfs_fsmnt, goal, cip->clean, npos);
1060 		syslog(LOG_INFO, "%s: cleaning with goal %" PRId64
1061 		       " bytes (%d segs clean, %d cleanable)",
1062 		       fs->lfs_fsmnt, goal, cip->clean, npos);
1063 		totbytes = 0;
1064 		for (i = 0; i < fs->lfs_nseg && totbytes < goal; i++) {
1065 			if (fs->clfs_segtabp[i]->priority == 0)
1066 				break;
1067 			/* Upper bound on number of segments at once */
1068 			if (ngood * fs->lfs_ssize > 4 * goal)
1069 				break;
1070 			sn = (fs->clfs_segtabp[i] - fs->clfs_segtab);
1071 			dlog("%s: add seg %d prio %" PRIu64
1072 			     " containing %ld bytes",
1073 			     fs->lfs_fsmnt, sn, fs->clfs_segtabp[i]->priority,
1074 			     fs->clfs_segtabp[i]->nbytes);
1075 			if ((r = load_segment(fs, sn, &bip, &bic)) > 0) {
1076 				++ngood;
1077 				toss_old_blocks(fs, &bip, &bic, &bytes);
1078 				totbytes += bytes;
1079 			} else if (r == 0)
1080 				fd_release(fs->clfs_devvp);
1081 			else
1082 				break;
1083 		}
1084 	} else {
1085 		/* Set attainable goal */
1086 		goal = atatime;
1087 		if (goal > cip->clean - 1)
1088 			goal = MAX(cip->clean - 1, 1);
1089 
1090 		dlog("%s: cleaning with goal %d segments (%d clean, %d cleanable)",
1091 		       fs->lfs_fsmnt, (int)goal, cip->clean, npos);
1092 		for (i = 0; i < fs->lfs_nseg && ngood < goal; i++) {
1093 			if (fs->clfs_segtabp[i]->priority == 0)
1094 				break;
1095 			sn = (fs->clfs_segtabp[i] - fs->clfs_segtab);
1096 			dlog("%s: add seg %d prio %" PRIu64,
1097 			     fs->lfs_fsmnt, sn, fs->clfs_segtabp[i]->priority);
1098 			if ((r = load_segment(fs, sn, &bip, &bic)) > 0)
1099 				++ngood;
1100 			else if (r == 0)
1101 				fd_release(fs->clfs_devvp);
1102 			else
1103 				break;
1104 		}
1105 		toss_old_blocks(fs, &bip, &bic, NULL);
1106 	}
1107 
1108 	/* If there is nothing to do, try again later. */
1109 	if (bic == 0) {
1110 		dlog("%s: no blocks to clean in %d cleanable segments",
1111 		       fs->lfs_fsmnt, (int)ngood);
1112 		fd_release_all(fs->clfs_devvp);
1113 		return 0;
1114 	}
1115 
1116 	/* Record statistics */
1117 	for (i = nb = 0; i < bic; i++)
1118 		nb += bip[i].bi_size;
1119 	util = ((double)nb) / (fs->clfs_nactive * fs->lfs_ssize);
1120 	cleaner_stats.util_tot += util;
1121 	cleaner_stats.util_sos += util * util;
1122 	cleaner_stats.bytes_written += nb;
1123 
1124 	/*
1125 	 * Check out our blocks to see if there are hidden cleaning costs.
1126 	 * If there are, we might be cleaning ourselves deeper into a hole
1127 	 * rather than doing anything useful.
1128 	 * XXX do something about this.
1129 	 */
1130 	if_extra = 0;
1131 	extra = fs->lfs_bsize * (off_t)check_hidden_cost(fs, bip, bic, &if_extra);
1132 	if_extra *= fs->lfs_bsize;
1133 
1134 	/*
1135 	 * Use markv to move the blocks.
1136 	 */
1137 	if (do_small)
1138 		inc = MAXPHYS / fs->lfs_bsize - 1;
1139 	else
1140 		inc = LFS_MARKV_MAXBLKCNT / 2;
1141 	for (mc = 0, mbip = bip; mc < bic; mc += inc, mbip += inc) {
1142 		lim.blkiov = mbip;
1143 		lim.blkcnt = (bic - mc > inc ? inc : bic - mc);
1144 #ifdef TEST_PATTERN
1145 		dlog("checking blocks %d-%d", mc, mc + lim.blkcnt - 1);
1146 		for (i = 0; i < lim.blkcnt; i++) {
1147 			check_test_pattern(mbip + i);
1148 		}
1149 #endif /* TEST_PATTERN */
1150 		dlog("sending blocks %d-%d", mc, mc + lim.blkcnt - 1);
1151 		if ((r = kops.ko_fcntl(fs->clfs_ifilefd, LFCNMARKV, &lim))<0) {
1152 			syslog(LOG_WARNING, "%s: markv returned %d (%m)",
1153 			       fs->lfs_fsmnt, r);
1154 			if (errno != EAGAIN && errno != ESHUTDOWN) {
1155 				fd_release_all(fs->clfs_devvp);
1156 				return r;
1157 			}
1158 		}
1159 	}
1160 
1161 	/*
1162 	 * Report progress (or lack thereof)
1163 	 */
1164 	syslog(LOG_INFO, "%s: wrote %" PRId64 " dirty + %"
1165 	       PRId64 " supporting indirect + %"
1166 	       PRId64 " supporting Ifile = %"
1167 	       PRId64 " bytes to clean %d segs (%" PRId64 "%% recovery)",
1168 	       fs->lfs_fsmnt, (int64_t)nb, (int64_t)(extra - if_extra),
1169 	       (int64_t)if_extra, (int64_t)(nb + extra), ngood,
1170 	       (ngood ? (int64_t)(100 - (100 * (nb + extra)) /
1171 					 (ngood * fs->lfs_ssize)) :
1172 		(int64_t)0));
1173 	if (nb + extra >= ngood * fs->lfs_ssize)
1174 		syslog(LOG_WARNING, "%s: cleaner not making forward progress",
1175 		       fs->lfs_fsmnt);
1176 
1177 	/*
1178 	 * Finally call reclaim to prompt cleaning of the segments.
1179 	 */
1180 	kops.ko_fcntl(fs->clfs_ifilefd, LFCNRECLAIM, NULL);
1181 
1182 	fd_release_all(fs->clfs_devvp);
1183 	return 0;
1184 }
1185 
1186 /*
1187  * Read the cleanerinfo block and apply cleaning policy to determine whether
1188  * the given filesystem needs to be cleaned.  Returns 1 if it does, 0 if it
1189  * does not, or -1 on error.
1190  */
1191 int
1192 needs_cleaning(struct clfs *fs, CLEANERINFO *cip)
1193 {
1194 	struct ubuf *bp;
1195 	struct stat st;
1196 	daddr_t fsb_per_seg, max_free_segs;
1197 	time_t now;
1198 	double loadavg;
1199 
1200 	/* If this fs is "on hold", don't clean it. */
1201 	if (fs->clfs_onhold)
1202 		return 0;
1203 
1204 	/*
1205 	 * Read the cleanerinfo block from the Ifile.  We don't want
1206 	 * the cached information, so invalidate the buffer before
1207 	 * handing it back.
1208 	 */
1209 	if (bread(fs->lfs_ivnode, 0, fs->lfs_bsize, NOCRED, 0, &bp)) {
1210 		syslog(LOG_ERR, "%s: can't read inode", fs->lfs_fsmnt);
1211 		return -1;
1212 	}
1213 	*cip = *(CLEANERINFO *)bp->b_data; /* Structure copy */
1214 	brelse(bp, B_INVAL);
1215 	cleaner_stats.bytes_read += fs->lfs_bsize;
1216 
1217 	/*
1218 	 * If the number of segments changed under us, reinit.
1219 	 * We don't have to start over from scratch, however,
1220 	 * since we don't hold any buffers.
1221 	 */
1222 	if (fs->lfs_nseg != cip->clean + cip->dirty) {
1223 		if (reinit_fs(fs) < 0) {
1224 			/* The normal case for unmount */
1225 			syslog(LOG_NOTICE, "%s: filesystem unmounted", fs->lfs_fsmnt);
1226 			return -1;
1227 		}
1228 		syslog(LOG_NOTICE, "%s: nsegs changed", fs->lfs_fsmnt);
1229 	}
1230 
1231 	/* Compute theoretical "free segments" maximum based on usage */
1232 	fsb_per_seg = segtod(fs, 1);
1233 	max_free_segs = MAX(cip->bfree, 0) / fsb_per_seg + fs->lfs_minfreeseg;
1234 
1235 	dlog("%s: bfree = %d, avail = %d, clean = %d/%d",
1236 	     fs->lfs_fsmnt, cip->bfree, cip->avail, cip->clean, fs->lfs_nseg);
1237 
1238 	/* If the writer is waiting on us, clean it */
1239 	if (cip->clean <= fs->lfs_minfreeseg ||
1240 	    (cip->flags & LFS_CLEANER_MUST_CLEAN))
1241 		return 1;
1242 
1243 	/* If there are enough segments, don't clean it */
1244 	if (cip->bfree - cip->avail <= fsb_per_seg &&
1245 	    cip->avail > fsb_per_seg)
1246 		return 0;
1247 
1248 	/* If we are in dire straits, clean it */
1249 	if (cip->bfree - cip->avail > fsb_per_seg &&
1250 	    cip->avail <= fsb_per_seg)
1251 		return 1;
1252 
1253 	/* If under busy threshold, clean regardless of load */
1254 	if (cip->clean < max_free_segs * BUSY_LIM)
1255 		return 1;
1256 
1257 	/* Check busy status; clean if idle and under idle limit */
1258 	if (use_fs_idle) {
1259 		/* Filesystem idle */
1260 		time(&now);
1261 		if (fstat(fs->clfs_ifilefd, &st) < 0) {
1262 			syslog(LOG_ERR, "%s: failed to stat ifile",
1263 			       fs->lfs_fsmnt);
1264 			return -1;
1265 		}
1266 		if (now - st.st_mtime > segwait_timeout &&
1267 		    cip->clean < max_free_segs * IDLE_LIM)
1268 			return 1;
1269 	} else {
1270 		/* CPU idle - use one-minute load avg */
1271 		if (getloadavg(&loadavg, 1) == -1) {
1272 			syslog(LOG_ERR, "%s: failed to get load avg",
1273 			       fs->lfs_fsmnt);
1274 			return -1;
1275 		}
1276 		if (loadavg < load_threshold &&
1277 		    cip->clean < max_free_segs * IDLE_LIM)
1278 			return 1;
1279 	}
1280 
1281 	return 0;
1282 }
1283 
1284 /*
1285  * Report statistics.  If the signal was SIGUSR2, clear the statistics too.
1286  * If the signal was SIGINT, exit.
1287  */
1288 static void
1289 sig_report(int sig)
1290 {
1291 	double avg = 0.0, stddev;
1292 
1293 	avg = cleaner_stats.util_tot / MAX(cleaner_stats.segs_cleaned, 1.0);
1294 	stddev = cleaner_stats.util_sos / MAX(cleaner_stats.segs_cleaned -
1295 					      avg * avg, 1.0);
1296 	syslog(LOG_INFO, "bytes read:	     %" PRId64, cleaner_stats.bytes_read);
1297 	syslog(LOG_INFO, "bytes written:     %" PRId64, cleaner_stats.bytes_written);
1298 	syslog(LOG_INFO, "segments cleaned:  %" PRId64, cleaner_stats.segs_cleaned);
1299 #if 0
1300 	/* "Empty segments" is meaningless, since the kernel handles those */
1301 	syslog(LOG_INFO, "empty segments:    %" PRId64, cleaner_stats.segs_empty);
1302 #endif
1303 	syslog(LOG_INFO, "error segments:    %" PRId64, cleaner_stats.segs_error);
1304 	syslog(LOG_INFO, "utilization total: %g", cleaner_stats.util_tot);
1305 	syslog(LOG_INFO, "utilization sos:   %g", cleaner_stats.util_sos);
1306 	syslog(LOG_INFO, "utilization avg:   %4.2f", avg);
1307 	syslog(LOG_INFO, "utilization sdev:  %9.6f", stddev);
1308 
1309 	if (debug)
1310 		bufstats();
1311 
1312 	if (sig == SIGUSR2)
1313 		memset(&cleaner_stats, 0, sizeof(cleaner_stats));
1314 	if (sig == SIGINT)
1315 		exit(0);
1316 }
1317 
1318 static void
1319 sig_exit(int sig)
1320 {
1321 	exit(0);
1322 }
1323 
1324 static void
1325 usage(void)
1326 {
1327 	errx(1, "usage: lfs_cleanerd [-bcdfmqs] [-i segnum] [-l load] "
1328 	     "[-n nsegs] [-r report_freq] [-t timeout] fs_name ...");
1329 }
1330 
1331 #ifndef LFS_CLEANER_AS_LIB
1332 /*
1333  * Main.
1334  */
1335 int
1336 main(int argc, char **argv)
1337 {
1338 
1339 	return lfs_cleaner_main(argc, argv);
1340 }
1341 #endif
1342 
1343 int
1344 lfs_cleaner_main(int argc, char **argv)
1345 {
1346 	int i, opt, error, r, loopcount, nodetach;
1347 	struct timeval tv;
1348 	sem_t *semaddr = NULL;
1349 	CLEANERINFO ci;
1350 #ifndef USE_CLIENT_SERVER
1351 	char *cp, *pidname;
1352 #endif
1353 
1354 	/*
1355 	 * Set up defaults
1356 	 */
1357 	atatime	 = 1;
1358 	segwait_timeout = 300; /* Five minutes */
1359 	load_threshold	= 0.2;
1360 	stat_report	= 0;
1361 	inval_segment	= -1;
1362 	copylog_filename = NULL;
1363 	nodetach        = 0;
1364 
1365 	/*
1366 	 * Parse command-line arguments
1367 	 */
1368 	while ((opt = getopt(argc, argv, "bC:cdDfi:l:mn:qr:sS:t:")) != -1) {
1369 		switch (opt) {
1370 		    case 'b':	/* Use bytes written, not segments read */
1371 			    use_bytes = 1;
1372 			    break;
1373 		    case 'C':	/* copy log */
1374 			    copylog_filename = optarg;
1375 			    break;
1376 		    case 'c':	/* Coalesce files */
1377 			    do_coalesce++;
1378 			    break;
1379 		    case 'd':	/* Debug mode. */
1380 			    nodetach++;
1381 			    debug++;
1382 			    break;
1383 		    case 'D':	/* stay-on-foreground */
1384 			    nodetach++;
1385 			    break;
1386 		    case 'f':	/* Use fs idle time rather than cpu idle */
1387 			    use_fs_idle = 1;
1388 			    break;
1389 		    case 'i':	/* Invalidate this segment */
1390 			    inval_segment = atoi(optarg);
1391 			    break;
1392 		    case 'l':	/* Load below which to clean */
1393 			    load_threshold = atof(optarg);
1394 			    break;
1395 		    case 'm':	/* [compat only] */
1396 			    break;
1397 		    case 'n':	/* How many segs to clean at once */
1398 			    atatime = atoi(optarg);
1399 			    break;
1400 		    case 'q':	/* Quit after one run */
1401 			    do_quit = 1;
1402 			    break;
1403 		    case 'r':	/* Report every stat_report segments */
1404 			    stat_report = atoi(optarg);
1405 			    break;
1406 		    case 's':	/* Small writes */
1407 			    do_small = 1;
1408 			    break;
1409 		    case 'S':	/* semaphore */
1410 #ifndef LFS_CLEANER_AS_LIB
1411 			    usage();
1412 			    /*NOTREACHED*/
1413 #endif
1414 			    semaddr = (void*)(uintptr_t)strtoull(optarg,NULL,0);
1415 			    break;
1416 		    case 't':	/* timeout */
1417 			    segwait_timeout = atoi(optarg);
1418 			    break;
1419 		    default:
1420 			    usage();
1421 			    /* NOTREACHED */
1422 		}
1423 	}
1424 	argc -= optind;
1425 	argv += optind;
1426 
1427 	if (argc < 1)
1428 		usage();
1429 	if (inval_segment >= 0 && argc != 1) {
1430 		errx(1, "lfs_cleanerd: may only specify one filesystem when "
1431 		     "using -i flag");
1432 	}
1433 
1434 	if (do_coalesce) {
1435 		errx(1, "lfs_cleanerd: -c disabled due to reports of file "
1436 		     "corruption; you may re-enable it by rebuilding the "
1437 		     "cleaner");
1438 	}
1439 
1440 	/*
1441 	 * Set up daemon mode or foreground mode
1442 	 */
1443 	if (nodetach) {
1444 		openlog("lfs_cleanerd", LOG_NDELAY | LOG_PID | LOG_PERROR,
1445 			LOG_DAEMON);
1446 		signal(SIGINT, sig_report);
1447 	} else {
1448 		if (daemon(0, 0) == -1)
1449 			err(1, "lfs_cleanerd: couldn't become a daemon!");
1450 		openlog("lfs_cleanerd", LOG_NDELAY | LOG_PID, LOG_DAEMON);
1451 		signal(SIGINT, sig_exit);
1452 	}
1453 
1454 	/*
1455 	 * Look for an already-running master daemon.  If there is one,
1456 	 * send it our filesystems to add to its list and exit.
1457 	 * If there is none, become the master.
1458 	 */
1459 #ifdef USE_CLIENT_SERVER
1460 	try_to_become_master(argc, argv);
1461 #else
1462 	/* XXX think about this */
1463 	asprintf(&pidname, "lfs_cleanerd:m:%s", argv[0]);
1464 	if (pidname == NULL) {
1465 		syslog(LOG_ERR, "malloc failed: %m");
1466 		exit(1);
1467 	}
1468 	for (cp = pidname; cp != NULL; cp = strchr(cp, '/'))
1469 		*cp = '|';
1470 	pidfile(pidname);
1471 #endif
1472 
1473 	/*
1474 	 * Signals mean daemon should report its statistics
1475 	 */
1476 	memset(&cleaner_stats, 0, sizeof(cleaner_stats));
1477 	signal(SIGUSR1, sig_report);
1478 	signal(SIGUSR2, sig_report);
1479 
1480 	/*
1481 	 * Start up buffer cache.  We only use this for the Ifile,
1482 	 * and we will resize it if necessary, so it can start small.
1483 	 */
1484 	bufinit(4);
1485 
1486 #ifdef REPAIR_ZERO_FINFO
1487 	{
1488 		BLOCK_INFO *bip = NULL;
1489 		int bic = 0;
1490 
1491 		nfss = 1;
1492 		fsp = (struct clfs **)malloc(sizeof(*fsp));
1493 		fsp[0] = (struct clfs *)calloc(1, sizeof(**fsp));
1494 
1495 		if (init_unmounted_fs(fsp[0], argv[0]) < 0) {
1496 			err(1, "init_unmounted_fs");
1497 		}
1498 		dlog("Filesystem has %d segments", fsp[0]->lfs_nseg);
1499 		for (i = 0; i < fsp[0]->lfs_nseg; i++) {
1500 			load_segment(fsp[0], i, &bip, &bic);
1501 			bic = 0;
1502 		}
1503 		exit(0);
1504 	}
1505 #endif
1506 
1507 	/*
1508 	 * Initialize cleaning structures, open devices, etc.
1509 	 */
1510 	nfss = argc;
1511 	fsp = (struct clfs **)malloc(nfss * sizeof(*fsp));
1512 	if (fsp == NULL) {
1513 		syslog(LOG_ERR, "couldn't allocate fs table: %m");
1514 		exit(1);
1515 	}
1516 	for (i = 0; i < nfss; i++) {
1517 		fsp[i] = (struct clfs *)calloc(1, sizeof(**fsp));
1518 		if ((r = init_fs(fsp[i], argv[i])) < 0) {
1519 			syslog(LOG_ERR, "%s: couldn't init: error code %d",
1520 			       argv[i], r);
1521 			handle_error(fsp, i);
1522 			--i; /* Do the new #i over again */
1523 		}
1524 	}
1525 
1526 	/*
1527 	 * If asked to coalesce, do so and exit.
1528 	 */
1529 	if (do_coalesce) {
1530 		for (i = 0; i < nfss; i++)
1531 			clean_all_inodes(fsp[i]);
1532 		exit(0);
1533 	}
1534 
1535 	/*
1536 	 * If asked to invalidate a segment, do that and exit.
1537 	 */
1538 	if (inval_segment >= 0) {
1539 		invalidate_segment(fsp[0], inval_segment);
1540 		exit(0);
1541 	}
1542 
1543 	/*
1544 	 * Main cleaning loop.
1545 	 */
1546 	loopcount = 0;
1547 #ifdef LFS_CLEANER_AS_LIB
1548 	if (semaddr)
1549 		sem_post(semaddr);
1550 #endif
1551 	error = 0;
1552 	while (nfss > 0) {
1553 		int cleaned_one;
1554 		do {
1555 #ifdef USE_CLIENT_SERVER
1556 			check_control_socket();
1557 #endif
1558 			cleaned_one = 0;
1559 			for (i = 0; i < nfss; i++) {
1560 				if ((error = needs_cleaning(fsp[i], &ci)) < 0) {
1561 					handle_error(fsp, i);
1562 					continue;
1563 				}
1564 				if (error == 0) /* No need to clean */
1565 					continue;
1566 
1567 				reload_ifile(fsp[i]);
1568 				if (clean_fs(fsp[i], &ci) < 0) {
1569 					handle_error(fsp, i);
1570 					continue;
1571 				}
1572 				++cleaned_one;
1573 			}
1574 			++loopcount;
1575 			if (stat_report && loopcount % stat_report == 0)
1576 				sig_report(0);
1577 			if (do_quit)
1578 				exit(0);
1579 		} while(cleaned_one);
1580 		tv.tv_sec = segwait_timeout;
1581 		tv.tv_usec = 0;
1582 		/* XXX: why couldn't others work if fsp socket is shutdown? */
1583 		error = kops.ko_fcntl(fsp[0]->clfs_ifilefd,LFCNSEGWAITALL,&tv);
1584 		if (error) {
1585 			if (errno == ESHUTDOWN) {
1586 				for (i = 0; i < nfss; i++) {
1587 					handle_error(fsp, i);
1588 					assert(nfss == 0);
1589 				}
1590 			} else {
1591 #ifdef LFS_CLEANER_AS_LIB
1592 				error = ESHUTDOWN;
1593 				break;
1594 #else
1595 				err(1, "LFCNSEGWAITALL");
1596 #endif
1597 			}
1598 		}
1599 	}
1600 
1601 	/* NOTREACHED */
1602 	return error;
1603 }
1604