nilfs2 fails to mount after io/error

nilfs2 fails to mount after io/error

[Adrian Ulrich]

Hi,

I just had an usb-harddrive (with nilfs2 and ext3 mounted) throw some IO-Errors (Most likely due to a crappy usb controller)
and now nilfs2 refuses to mount the partition:

NILFS warning: Sequence number mismatch
NILFS: error searching super root.

(ext3 still works fine)

Any idea?

Regards,
 Adrian





-- 
 RFC 1925:
   (11) Every old idea will be proposed again with a different name and
        a different presentation, regardless of whether it works.

Re: nilfs2 fails to mount after io/error

[Reinoud Zandijk]

Hi,

On Wed, Nov 26, 2008 at 07:22:12PM +0100, Adrian Ulrich wrote:
> I just had an usb-harddrive (with nilfs2 and ext3 mounted) throw some IO-Errors (Most likely due to a crappy usb controller)
> and now nilfs2 refuses to mount the partition:
> 
> NILFS warning: Sequence number mismatch
> NILFS: error searching super root.
> 
> (ext3 still works fine)

Well you could try to mount on an older snapshot or create a snapshot 
first?? i dont have a running nilfs2 around here so i can't try out the 
advise :)

With regards,
Reinoud

Re: nilfs2 fails to mount after io/error

[Adrian Ulrich]

> Well you could try to mount on an older snapshot or create a snapshot 
> first?? i dont have a running nilfs2 around here so i can't try out the 
> advise :)

I do not have any Snapshots (only CPs) and chcp requires a mounted filesystem :-/

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

Hi,

On Wed, 26 Nov 2008 19:22:12 +0100, Adrian Ulrich wrote:
> Hi,
> 
> I just had an usb-harddrive (with nilfs2 and ext3 mounted) throw some IO-Errors (Most likely due to a crappy usb controller)
> and now nilfs2 refuses to mount the partition:
> 
> NILFS warning: Sequence number mismatch
> NILFS: error searching super root.
> 
> (ext3 still works fine)
> 
> Any idea?

Hmm, this error seems to be caused by some kind of bug in NILFS2
write routines.  I'll verify their error handlings.

In the meantime, please try the contained patch as follows:

$ tar jxf nilfs-2.0.5.tar.bz2
$ cd nilfs-2.0.5
$ patch -p1 < this-mail.patch
$ make
$ sudo make install

It disables the check of sequence numbers and would allow you to
mount the partition.

With regards,
Ryusuke Konishi

--
diff --git a/fs/recovery.c b/fs/recovery.c
index bd8e175..1144609 100644
--- a/fs/recovery.c
+++ b/fs/recovery.c
@@ -240,10 +240,12 @@ load_segment_summary(struct nilfs_sb_info *sbi, sector_t pseg_start,
 		goto failed;
 	}
 	store_segsum_info(ssi, sum, sbi->s_super->s_blocksize);
+#if 0
 	if (seg_seq != ssi->seg_seq) {
 		ret = NILFS_SEG_FAIL_SEQ;
 		goto failed;
 	}
+#endif
 	if (full_check) {
 		offset = sizeof(sum->ss_datasum);
 		check_bytes =

Re: nilfs2 fails to mount after io/error

[Adrian Ulrich]

Hi,

> In the meantime, please try the contained patch as follows:

mount.nilfs2 now 'hangs' while mounting the filesystem:

 - CPU-Usage ~ 10%
 - 100% io-wait


Do you think that mount.nilfs2 will ever finish? (It's running since ~15 minutes)


Regards,
 Adrian

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

Hi,

On Fri, 28 Nov 2008 15:37:32 +0100, Adrian Ulrich wrote:
> Hi,
> 
> > In the meantime, please try the contained patch as follows:
> 
> mount.nilfs2 now 'hangs' while mounting the filesystem:
> 
>  - CPU-Usage ~ 10%
>  - 100% io-wait
> 
> 
> Do you think that mount.nilfs2 will ever finish? (It's running since ~15 minutes)

Oh, that seems unusual.  Please stop applying the patch.

The check of sequence number might be really correct.

In that case, recovery logic is suspicious.  Or there was actually a
critical violation of write order on the device.

For nilfs2, if a superblock was written ahead of the segment pointed
by the superblock, this can happen.


Anyhow, I'll review the source code some more.

With regards,
Ryusuke Konishi

Re: nilfs2 fails to mount after io/error

[Adrian Ulrich]

Hi,

> 
> For nilfs2, if a superblock was written ahead of the segment pointed
> by the superblock, this can happen.
> 
> Anyhow, I'll review the source code some more.

Ok. Let me know if you need something from me.
(metadata dump, etc..)


Regards,
 Adrian

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

On Mon, 1 Dec 2008 18:14:35 +0100, Adrian Ulrich wrote:
> Hi,
> 
> > 
> > For nilfs2, if a superblock was written ahead of the segment pointed
> > by the superblock, this can happen.
> > 
> > Anyhow, I'll review the source code some more.
> 
> Ok. Let me know if you need something from me.
> (metadata dump, etc..)

Thanks, 

Could you get debug log with a DEBUG option?

By commenting out the following line in nilfs2/fs/Makefile,
you can make the nilfs2 module for debugging purpose.

 # CONFIG_NILFS_DEBUG=y

Compilation and installation are as usual.

Then write the following option to a procfs file 'nilfs2/debug_option'
before mounting the partition.

 # modprobe nilfs2
 # echo "-vvv recovery" > /proc/fs/nilfs2/debug_option
 # mount -t nilfs2 -r /dev/xxx /test
 ...

This will log trace messages of recovery function.


Thanks in advance,
Ryusuke Konishi

Re: nilfs2 fails to mount after io/error

[Adrian Ulrich]

Debug output:

NILFS nilfs_fill_super: start(silent=0)
NILFS: INFO: recovery required for readonly filesystem.
NILFS: write access will be enabled during recovery.
NILFS(recovery) nilfs_search_super_root: looking segment (seg_start=174272512, seg_end=174274559, segnum=85094, seg_seq=181656)
NILFS(recovery) load_segment_summary: checking segment (pseg_start=174272512, full_check=1)
NILFS(recovery) load_segment_summary: done (ret=4)
NILFS(recovery) nilfs_search_super_root: strayed: scan_newer=0, ret=4
NILFS warning: Sequence number mismatch
NILFS: error searching super root.
NILFS nilfs_fill_super: aborted
NILFS put_nilfs: the_nilfs on bdev sdb5 was freed



Do you also need the output with your first patch applied?

Regards,
 Adrian

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

Hi,

On Tue, 2 Dec 2008 07:44:09 +0100, Adrian Ulrich wrote:
> Debug output:
> 
> NILFS nilfs_fill_super: start(silent=0)
> NILFS: INFO: recovery required for readonly filesystem.
> NILFS: write access will be enabled during recovery.
> NILFS(recovery) nilfs_search_super_root: looking segment (seg_start=174272512, seg_end=174274559, segnum=85094, seg_seq=181656)
> NILFS(recovery) load_segment_summary: checking segment (pseg_start=174272512, full_check=1)
> NILFS(recovery) load_segment_summary: done (ret=4)
> NILFS(recovery) nilfs_search_super_root: strayed: scan_newer=0, ret=4
> NILFS warning: Sequence number mismatch
> NILFS: error searching super root.
> NILFS nilfs_fill_super: aborted
> NILFS put_nilfs: the_nilfs on bdev sdb5 was freed

According to this log, there is actually a sequence number mismatch
between superblock and the segment pointed by the superblock.  The
recovery function seems to work normally.

> 
> Do you also need the output with your first patch applied?

Instead of the previous patch, could you try the following patch?

It forces to mount a filesystem from the segment pointed by the
superblock if there is no other failure.

If it succeeds, copy filesystem contents (just to be safe) and
unmount it.  Then reinstall the original module. 

If successful, it may be able to re-mount as if nothing had happened.

Regards,
Ryusuke Konishi

--
diff --git a/fs/recovery.c b/fs/recovery.c
index bd8e175..874f855 100644
--- a/fs/recovery.c
+++ b/fs/recovery.c
@@ -240,10 +240,12 @@ load_segment_summary(struct nilfs_sb_info *sbi, sector_t pseg_start,
 		goto failed;
 	}
 	store_segsum_info(ssi, sum, sbi->s_super->s_blocksize);
+#if 0
 	if (seg_seq != ssi->seg_seq) {
 		ret = NILFS_SEG_FAIL_SEQ;
 		goto failed;
 	}
+#endif
 	if (full_check) {
 		offset = sizeof(sum->ss_datasum);
 		check_bytes =
@@ -960,6 +962,8 @@ int nilfs_search_super_root(struct the_nilfs *nilfs, struct nilfs_sb_info *sbi,
 		nilfs->ns_cno = cno;  /* nilfs->ns_cno = ri->ri_cno + 1 */
 		nilfs->ns_ctime = ssi.ctime;
 		nilfs->ns_nextnum = nextnum;
+		ri->ri_need_recovery = NILFS_RECOVERY_SR_UPDATED;
+		goto super_root_found;
 
 		if (scan_newer)
 			ri->ri_need_recovery = NILFS_RECOVERY_SR_UPDATED;

Re: nilfs2 fails to mount after io/error

[Adrian Ulrich]

Still fails to mount :

# insmod fs/nilfs2.ko 
# echo "-vvv recovery" > /proc/fs/nilfs2/debug_option
# mount -t nilfs2 -r /dev/wdigital5 /scratch/

Dec  3 06:47:50 echelon kernel: NILFS nilfs_fill_super: start(silent=0)
Dec  3 06:47:50 echelon kernel: NILFS(recovery) nilfs_search_super_root: looking segment (seg_start=174272512, seg_end=174274559, segnum=85094, seg_seq=181656)
Dec  3 06:47:50 echelon kernel: NILFS(recovery) load_segment_summary: checking segment (pseg_start=174272512, full_check=1)
Dec  3 06:47:51 echelon kernel: NILFS(recovery) load_segment_summary: done (ret=0)
Dec  3 06:47:51 echelon kernel: NILFS(recovery) nilfs_search_super_root: found super root: segnum=85094, seq=181656, pseg_start=174272512, pseg_offset=295
Dec  3 06:47:51 echelon kernel: NILFS nilfs_attach_checkpoint: failed to attach ifile (checkpoint number=93986, err=-22)
Dec  3 06:47:51 echelon kernel: NILFS nilfs_fill_super: aborted
Dec  3 06:47:51 echelon kernel: NILFS put_nilfs: the_nilfs on bdev sdb5 was freed

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

On Wed, 3 Dec 2008 07:49:56 +0100, Adrian Ulrich wrote:
> Still fails to mount :
> 
> # insmod fs/nilfs2.ko 
> # echo "-vvv recovery" > /proc/fs/nilfs2/debug_option
> # mount -t nilfs2 -r /dev/wdigital5 /scratch/
> 
> Dec  3 06:47:50 echelon kernel: NILFS nilfs_fill_super: start(silent=0)
> Dec  3 06:47:50 echelon kernel: NILFS(recovery) nilfs_search_super_root: looking segment (seg_start=174272512, seg_end=174274559, segnum=85094, seg_seq=181656)
> Dec  3 06:47:50 echelon kernel: NILFS(recovery) load_segment_summary: checking segment (pseg_start=174272512, full_check=1)
> Dec  3 06:47:51 echelon kernel: NILFS(recovery) load_segment_summary: done (ret=0)
> Dec  3 06:47:51 echelon kernel: NILFS(recovery) nilfs_search_super_root: found super root: segnum=85094, seq=181656, pseg_start=174272512, pseg_offset=295
> Dec  3 06:47:51 echelon kernel: NILFS nilfs_attach_checkpoint: failed to attach ifile (checkpoint number=93986, err=-22)
> Dec  3 06:47:51 echelon kernel: NILFS nilfs_fill_super: aborted
> Dec  3 06:47:51 echelon kernel: NILFS put_nilfs: the_nilfs on bdev sdb5 was freed

The segment pointed by the superblock was likely not the latest one.


Could you get a summary of segments as follows, and send it to me
off-line?

 bash# for (( i = 0; i <= 85096; i++ )); do dumpseg /dev/wdigital5 $i | grep -e creation -e segment -e "ino = 3,"; done > segdump.txt
 bash# bzip2 segdump.txt


This will generate the summary as follows, and it helps
to identify the the last segment actually written.

segment: segnum = 0
  partial segment
    creation time = 2008-10-10 15:30:36
      ino = 3, cno = 1, nblocks = 3, ndatblk = 3
  partial segment
    creation time = 2008-10-10 15:30:48
      ino = 3, cno = 0, nblocks = 39, ndatblk = 38
  partial segment
    creation time = 2008-10-10 15:30:51
segment: segnum = 1
  partial segment
    creation time = 2008-10-10 15:30:52
segment: segnum = 2
  partial segment
    creation time = 2008-10-10 15:30:52
  partial segment
    creation time = 2008-10-10 15:30:54
...


If you can find the latest segment from 'creation time' fields,
then get a full dump, too.

 # dumpseg /dev/wdigital5 [the number of the newest segment] > latest-seg.txt


Regards,
Ryusuke Konishi

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

Hi Adrian,

Thanks for dump logs.

 segment: segnum = 85092
   partial segment
     creation time = 2008-11-26 17:41:05
   partial segment
     creation time = 2008-11-26 17:41:06
 segment: segnum = 85093
   partial segment           -> pseg 174270464
     creation time = 2008-11-26 17:41:06
       ino = 3, cno = 0, nblocks = 86, ndatblk = 80
   partial segment           -> pseg 174271740
     creation time = 2008-11-26 17:41:11
       ino = 3, cno = 0, nblocks = 8, ndatblk = 3
   partial segment           -> pseg 174271758
     creation time = 2008-11-26 17:42:47
       ino = 3, cno = 0, nblocks = 735, ndatblk = 722
 segment: segnum = 85094
   partial segment           -> pseg 174272512    <-- pointed by superblock
     creation time = 2008-10-28 20:27:55
       ino = 3, cno = 0, nblocks = 61, ndatblk = 51
   partial segment
     creation time = 2008-10-28 20:29:14
 segment: segnum = 85095
   partial segment
     creation time = 2008-10-28 20:29:15


According to your logs, the segment pointed by
the superblock was not written as suspected.

This never occurs if the block device is
supporting 'write barrier' feature properly.

Could you tell me the kernel version that you are using?

I feel that an additional disk flush operation may be
needed for older kernels or such devices. (Of course,
first priority should be on making fsck.)


Regards,
Ryusuke Konishi

Re: nilfs2 fails to mount after io/error

[Adrian Ulrich]

Hi,

> This never occurs if the block device is
> supporting 'write barrier' feature properly.
> Could you tell me the kernel version that you are using?

adrian@echelon:~$ uname -r
2.6.27.4

root@echelon:~# udevinfo -a -p /sys/block/sdb/|grep model
    ATTRS{model}=="10EAVS External "
root@echelon:~# lsusb |grep Wes
Bus 001 Device 003: ID 1058:1001 Western Digital Technologies, Inc. External Hard Disk


> (Of course, first priority should be on making fsck.)

Well: Is there any way to recover my data 'soon' or should
i re-format the partition? 

Regards,
 Adrian

Re: nilfs2 fails to mount after io/error

[Chris Samuel]

[-- Attachment #1.1: Type: text/plain, Size: 729 bytes --]

On Thu, 4 Dec 2008 7:06:54 pm Ryusuke Konishi wrote:

> This never occurs if the block device is
> supporting 'write barrier' feature properly.
>
> Could you tell me the kernel version that you are using?
>
> I feel that an additional disk flush operation may be
> needed for older kernels or such devices. (Of course,
> first priority should be on making fsck.)

It's worth remembering that (unless something has changed recently) filesystems 
on a logical volume using LVM won't have barriers either.

cheers,
Chris
-- 
 Chris Samuel  :  http://www.csamuel.org/  :  Melbourne, VIC

This email may come with a PGP signature as a file. Do not panic.
For more info see: http://en.wikipedia.org/wiki/OpenPGP


[-- Attachment #1.2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 481 bytes --]

[-- Attachment #2: Type: text/plain, Size: 158 bytes --]

_______________________________________________
users mailing list
users-JrjvKiOkagjYtjvyW6yDsg@public.gmane.org
https://www.nilfs.org/mailman/listinfo/users

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

Hi,

On Thu, 4 Dec 2008 19:30:59 +1100, Chris Samuel wrote:
> On Thu, 4 Dec 2008 7:06:54 pm Ryusuke Konishi wrote:
> 
> > This never occurs if the block device is
> > supporting 'write barrier' feature properly.
> >
> > Could you tell me the kernel version that you are using?
> >
> > I feel that an additional disk flush operation may be
> > needed for older kernels or such devices. (Of course,
> > first priority should be on making fsck.)
> 
> It's worth remembering that (unless something has changed recently) filesystems 
> on a logical volume using LVM won't have barriers either.
> 
> cheers,
> Chris

That's right!  It should be counted.

Thanks,
Ryusuke Konishi

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

On Thu, 4 Dec 2008 09:29:47 +0100, Adrian Ulrich wrote:
> Hi,
> 
> > This never occurs if the block device is
> > supporting 'write barrier' feature properly.
> > Could you tell me the kernel version that you are using?
> 
> adrian@echelon:~$ uname -r
> 2.6.27.4

Uum, this kernel should support barriers also for USB devices.

> root@echelon:~# udevinfo -a -p /sys/block/sdb/|grep model
>     ATTRS{model}=="10EAVS External "
> root@echelon:~# lsusb |grep Wes
> Bus 001 Device 003: ID 1058:1001 Western Digital Technologies, Inc. External Hard Disk

Does the device cache operate in writeback or writethough? ( FUA or not ? )
These parameters decide how barriers are implemented in scsi layer.
You can show these by

 # cat /sys/class/scsi_disk/x:x:x:x/{FUA,cache_type}

 where x:x:x:x is the address displayed by lsscsi command.

> > (Of course, first priority should be on making fsck.)
> 
> Well: Is there any way to recover my data 'soon' or should
> i re-format the partition? 

Wait for a moment.  I'm now writing an impromptu tool
to correct this kind of problem.

Regards,
Ryusuke Konishi

Re: nilfs2 fails to mount after io/error

[Adrian Ulrich]

Hi,

> [...]  where x:x:x:x is the address displayed by lsscsi command.

bash-3.1# lsscsi | grep 10EAVS
[4:0:0:0]    disk    WD       10EAVS External  1.05  /dev/wdigital    
bash-3.1# cat /sys/class/scsi_disk/4\:0\:0\:0/cache_type 
write through
bash-3.1# cat /sys/class/scsi_disk/4\:0\:0\:0/FUA        
0


> Wait for a moment.  I'm now writing an impromptu tool
> to correct this kind of problem.

Oh, great. I'll be glad to test it ;-)

Regards,
 Adrian

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

Hi,

On Thu, 4 Dec 2008 14:04:31 +0100, Adrian Ulrich wrote:
> > Wait for a moment.  I'm now writing an impromptu tool
> > to correct this kind of problem.
> 
> Oh, great. I'll be glad to test it ;-)

Sorry to keep you waiting. ;-)

I will attach a patch against nilfs2-utils-2.0.6, which adds the
tool 'fsck0.nilfs2'.

 # fsck0.nilfs2 <device name>

will, hopefully, correct your partition.
(I'm not sure because it's not yet tested enough)

Regards,
Ryusuke Konishi
---
From ce7bac99115bdf1fdadc002ba2b14bcb1ae4c19c Mon Sep 17 00:00:00 2001
From: Ryusuke Konishi <konishi.ryusuke-Zyj7fXuS5i5L9jVzuh4AOg@public.gmane.org>
Date: Fri, 5 Dec 2008 00:53:21 +0900
Subject: [PATCH] nilfs2-utils: add test tool to correct log pointer in super block

Signed-off-by: Ryusuke Konishi <konishi.ryusuke-Zyj7fXuS5i5L9jVzuh4AOg@public.gmane.org>
---
 sbin/mkfs/Makefile.am    |    6 +-
 sbin/mkfs/fsck0.nilfs2.c |  958 ++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 963 insertions(+), 1 deletions(-)
 create mode 100644 sbin/mkfs/fsck0.nilfs2.c

diff --git a/sbin/mkfs/Makefile.am b/sbin/mkfs/Makefile.am
index 965094e..5cbe02a 100644
--- a/sbin/mkfs/Makefile.am
+++ b/sbin/mkfs/Makefile.am
@@ -4,13 +4,17 @@
 ##
 ## Written by Koji Sato <koji-sG5X7nlA6pw@public.gmane.org>.
 
-sbin_PROGRAMS = mkfs.nilfs2
+sbin_PROGRAMS = mkfs.nilfs2 fsck0.nilfs2
 
 mkfs_nilfs2_SOURCES = mkfs.c bitops.c ../../lib/crc32.c mkfs.h
 mkfs_nilfs2_CFLAGS = -Wall
 mkfs_nilfs2_CPPFLAGS = -I$(top_srcdir)/include
 mkfs_nilfs2_LDADD = -luuid
 
+fsck0_nilfs2_SOURCES = fsck0.nilfs2.c ../../lib/crc32.c mkfs.h
+fsck0_nilfs2_CFLAGS = -Wall
+fsck0_nilfs2_CPPFLAGS = -I$(top_srcdir)/include
+
 install-exec-hook:
 	list='$(sbin_PROGRAMS)'; \
 	for p in $$list; do \
diff --git a/sbin/mkfs/fsck0.nilfs2.c b/sbin/mkfs/fsck0.nilfs2.c
new file mode 100644
index 0000000..d3d335f
--- /dev/null
+++ b/sbin/mkfs/fsck0.nilfs2.c
@@ -0,0 +1,958 @@
+/*
+ * fsck0.nilfs2.c - correct inconsistencies of nilfs2 volume
+ *
+ * Licensed under GPLv2: the complete text of the GNU General Public License
+ * can be found in COPYING file of the nilfs2-utils package.
+ *
+ * Copyright (C) 2008 Nippon Telegraph and Telephone Corporation.
+ * Written by Ryusuke Konishi <ryusuke-sG5X7nlA6pw@public.gmane.org>
+ */
+#define _LARGEFILE64_SOURCE
+#define _XOPEN_SOURCE 600
+
+#include <sys/types.h>
+#include <linux/types.h>
+#include <endian.h>
+#include <byteswap.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <strings.h>
+#include <stdarg.h>
+#include <stddef.h>
+#include <time.h>
+#include <string.h>
+#include <assert.h>
+#include "mkfs.h"
+
+#define MOUNTS			"/etc/mtab"
+#define LINE_BUFFER_SIZE	256  /* Line buffer size for reading mtab */
+#define MAX_SCAN_SEGMENT	50   /* Maximum number of segments which are
+					tested for the latest segment search */
+#define SCAN_INDICATOR_SPEED	3    /* Indicator speed (smaller value for
+					higher speed) */
+#define SCAN_SEGMENT_MASK	((1U << SCAN_INDICATOR_SPEED) - 1)
+
+# define DIV_ROUND_UP(n, d) (((n) + (d) - 1) / (d))
+
+char *progname = NULL;
+
+struct nilfs_pseg_ref {
+	__u64 blocknr;   /* start blocknumber */
+	__u64 seqnum;    /* sequence number */
+	__u64 cno;       /* checkpoint number */
+	__u64 ctime;     /* creation time */
+};
+
+static int devfd = -1;
+static int blocksize;
+static __u32 crc_seed;
+static __u32 blocks_per_segment;
+static __u64 first_data_block;
+static __u64 nsegments;
+static __u16 checkpoint_size;
+
+static int first_checkpoint_offset;
+static int ncheckpoints_per_block;
+
+/*
+ * Generic routines
+ */
+void die(const char *fmt, ...)
+{
+	va_list args;
+
+	va_start(args, fmt);
+	fprintf(stderr, "%s: ", progname);
+	vfprintf(stderr, fmt, args);
+	fprintf(stderr, "\n");
+	va_end(args);
+
+	if (devfd >= 0)
+		close(devfd);
+	exit(1);
+}
+
+static void (*nilfs_shrink)(void) = NULL;
+
+void *nilfs_malloc(size_t size)
+{
+	void *p = malloc(size);
+	if (!p) {
+		if (nilfs_shrink)
+			nilfs_shrink();
+		p = malloc(size);
+		if (!p)
+			die("memory allocation failure");
+	}
+	return p;
+}
+
+static inline void *nilfs_zalloc(size_t size)
+{
+	void *p = nilfs_malloc(size);
+	memset(p, 0, size);
+	return p;
+}
+
+/*
+ * Block buffer
+ */
+static void *block_buffer = NULL;
+
+static void destroy_block_buffer(void)
+{
+	if (block_buffer) {
+		free(block_buffer);
+		block_buffer = NULL;
+	}
+}
+
+static void init_block_buffer(void)
+{
+	block_buffer = nilfs_malloc(blocksize);
+	atexit(destroy_block_buffer);
+}
+
+static void read_block(int fd, __u64 blocknr, void *buf,
+		       unsigned long size)
+{
+	if (lseek64(fd, blocknr * blocksize, SEEK_SET) < 0 ||
+	    read(fd, buf, size) < size)
+		die("cannot read block (blocknr = %llu)",
+		    (unsigned long long)blocknr);
+}
+
+static inline __u64 segment_start_blocknr(unsigned long segnum)
+{
+	return segnum > 0 ? blocks_per_segment * segnum : first_data_block;
+}
+
+static int pseg_valid(int fd, __u64 pseg_start,
+		      struct nilfs_segment_summary *ss)
+{
+	__u32 crc, sum;
+	int offset = sizeof(ss->ss_datasum);
+	int nblocks = le32_to_cpu(ss->ss_nblocks);
+	__u64 blocknr = pseg_start;
+
+	if (le32_to_cpu(ss->ss_magic) != NILFS_SEGSUM_MAGIC)
+		return 0;
+
+	if (nblocks == 0 || nblocks > blocks_per_segment)
+		return 0;
+
+	sum = le32_to_cpu(ss->ss_datasum);
+
+	posix_fadvise(fd, pseg_start * blocksize, nblocks * blocksize,
+		      POSIX_FADV_SEQUENTIAL);
+
+	read_block(fd, blocknr++, block_buffer, blocksize);
+	crc = nilfs_crc32(crc_seed, block_buffer + offset, blocksize - offset);
+	while (--nblocks > 0) {
+		read_block(fd, blocknr++, block_buffer, blocksize);
+		crc = nilfs_crc32(crc, block_buffer, blocksize);
+	}
+	return crc == sum;
+}
+
+/*
+ * Routines to handle partial segment list
+ */
+struct nilfs_list {  /* use struct list_head in kernel land */
+	struct nilfs_list *prev;
+	struct nilfs_list *next;
+};
+
+static inline void nilfs_list_init(struct nilfs_list *list)
+{
+	list->prev = list->next = list;
+}
+
+static inline int nilfs_list_empty(struct nilfs_list *list)
+{
+	return list->next == list;
+}
+
+static inline void nilfs_list_del(struct nilfs_list *list)
+{
+	struct nilfs_list *p = list->prev, *n = list->next;
+
+	p->next = n;
+	n->prev = p;
+	list->prev = list->next = list;
+}
+
+static inline void nilfs_list_add(struct nilfs_list *list,
+				  struct nilfs_list *item)
+{
+	struct nilfs_list *p = list->prev;
+
+	item->prev = p;
+	item->next = list;
+	p->next = list->prev = item;
+}
+
+/* partial segment information */
+struct nilfs_pseg_info {
+	struct nilfs_list list;
+	__u64 pseg_start;  /* start blocknr */
+	__u32 nblocks;
+	struct nilfs_segment_summary segsum;  /* on-disk log header */
+	__u16 flags;
+};
+
+static inline struct nilfs_pseg_info *
+nilfs_pseg_list_entry(struct nilfs_list *p)
+{
+	return (void *)p - offsetof(struct nilfs_pseg_info, list);
+}
+
+struct nilfs_pseg_info *new_pseg_info(__u64 blocknr)
+{
+	struct nilfs_pseg_info *pseginfo = nilfs_zalloc(sizeof(*pseginfo));
+
+	pseginfo->pseg_start = blocknr;
+	nilfs_list_init(&pseginfo->list);
+	return pseginfo;
+}
+
+static void dispose_pseg_list(struct nilfs_list *list)
+{
+	struct nilfs_list *p, *n;
+
+	for (p = list->next; n = p->next, p != list; p = n) {
+		nilfs_list_del(p);
+		free(nilfs_pseg_list_entry(p));
+	}    
+}
+
+/*
+ * Segment information
+ */
+struct nilfs_segment_info {
+	struct nilfs_list list;
+	struct nilfs_list pseg_list;	/* partial segment list */
+	__u64 seg_start;		/* start blocknr of the segment */
+	__u64 next;			/* pointer to the next full segment */
+	__u64 segseq;			/* sequence number of the segment */
+	unsigned long segnum;		/* the number of the segment */
+	int npsegs;			/* number of partial segments */
+	int refcnt;
+};
+
+static struct nilfs_list segment_cache;
+
+static inline struct nilfs_segment_info *
+nilfs_segment_list_entry(struct nilfs_list *p)
+{
+	return (void *)p - offsetof(struct nilfs_segment_info, list);
+}
+
+struct nilfs_segment_info *new_segment_info(unsigned long segnum)
+{
+	struct nilfs_segment_info *seginfo;
+
+	seginfo = nilfs_zalloc(sizeof(*seginfo));
+	seginfo->segnum = segnum;
+	seginfo->seg_start = segment_start_blocknr(segnum);
+	seginfo->refcnt = 1;
+
+	nilfs_list_init(&seginfo->pseg_list);
+	nilfs_list_add(&segment_cache, &seginfo->list);
+	return seginfo;
+}
+
+void destroy_segment_info(struct nilfs_segment_info *seginfo)
+{
+	nilfs_list_del(&seginfo->list);
+	dispose_pseg_list(&seginfo->pseg_list);
+	free(seginfo);
+}
+
+static inline struct nilfs_segment_info *
+get_segment_info(struct nilfs_segment_info *seginfo)
+{
+	seginfo->refcnt++;
+	return seginfo;
+}
+
+static inline void put_segment_info(struct nilfs_segment_info *seginfo)
+{
+	assert(seginfo->refcnt > 0);
+	seginfo->refcnt--;
+}
+
+/*
+ * Segment cache
+ */
+void destroy_segment_cache(void)
+{
+	struct nilfs_list *p, *n;
+
+	for (p = segment_cache.next; n = p->next, p != &segment_cache; p = n) {
+		destroy_segment_info(nilfs_segment_list_entry(p));
+	}    
+}
+
+void shrink_segment_cache(void)
+{
+	struct nilfs_list *p, *n;
+	struct nilfs_segment_info *seginfo;
+
+	for (p = segment_cache.next; n = p->next, p != &segment_cache; p = n) {
+		seginfo = nilfs_segment_list_entry(p);
+		if (seginfo->refcnt == 0)
+			destroy_segment_info(seginfo);
+	}    
+}
+
+void init_segment_cache(void)
+{
+	nilfs_list_init(&segment_cache);
+	nilfs_shrink = shrink_segment_cache;
+	atexit(destroy_segment_cache);
+}
+
+struct nilfs_segment_info *lookup_segment(unsigned long segnum)
+{
+	struct nilfs_segment_info *seginfo;
+	struct nilfs_list *p;
+
+	for (p = segment_cache.next; p != &segment_cache; p = p->next) {
+		seginfo = nilfs_segment_list_entry(p);
+		if (seginfo->segnum == segnum) {
+			get_segment_info(seginfo);
+			return seginfo;
+		}
+	}
+	return NULL;
+}
+
+struct nilfs_segment_info *load_segment(int fd, unsigned long segnum)
+{
+	struct nilfs_segment_info *seginfo;
+	struct nilfs_pseg_info *pseginfo;
+	struct nilfs_segment_summary *ss;
+	__u64 blocknr, end;
+
+	seginfo = lookup_segment(segnum);
+	if (seginfo)
+		return seginfo;
+
+	seginfo = new_segment_info(segnum);
+	blocknr = seginfo->seg_start;
+
+	pseginfo = new_pseg_info(blocknr);
+	nilfs_list_add(&seginfo->pseg_list, &pseginfo->list);
+
+	ss = &pseginfo->segsum;
+	read_block(fd, blocknr, ss, sizeof(*ss));
+
+	if (!pseg_valid(fd, blocknr, ss)) {
+		put_segment_info(seginfo);
+		fprintf(stderr, "empty or bad segment: "
+			"segnum = %lu, blocknr = %llu\n", segnum,
+			(unsigned long long)segment_start_blocknr(segnum));
+		return NULL; /* no valid partial segment found */
+	}
+
+	seginfo->segseq = le64_to_cpu(ss->ss_seq);
+	seginfo->next = le64_to_cpu(ss->ss_next);
+
+	end = blocknr + blocks_per_segment;
+	do {
+		seginfo->npsegs++;
+
+		pseginfo->nblocks = le32_to_cpu(ss->ss_nblocks);
+		pseginfo->flags = le16_to_cpu(ss->ss_flags);
+
+		blocknr += pseginfo->nblocks;
+		if (blocknr >= end)
+			return seginfo;
+
+		pseginfo = new_pseg_info(blocknr);
+		nilfs_list_add(&seginfo->pseg_list, &pseginfo->list);
+
+		ss = &pseginfo->segsum;
+		read_block(fd, blocknr, ss, sizeof(*ss));
+
+	} while (pseg_valid(fd, blocknr, ss) &&
+		 le64_to_cpu(ss->ss_seq) == seginfo->segseq);
+
+	nilfs_list_del(&pseginfo->list);
+	free(pseginfo);
+
+	return seginfo;
+}
+
+/*
+ * Operations on segment_info structure
+ */
+struct nilfs_pseg_info *lookup_pseg(struct nilfs_segment_info *seginfo,
+				    __u64 blocknr)
+{
+	struct nilfs_pseg_info *pseginfo;
+	struct nilfs_list *p;
+
+	for (p = seginfo->pseg_list.next; p != &seginfo->pseg_list;
+	     p = p->next) {
+		pseginfo = nilfs_pseg_list_entry(p);
+		if (pseginfo->pseg_start == blocknr)
+			return pseginfo;
+	}
+	return NULL;
+}
+
+struct nilfs_pseg_info *first_pseg(struct nilfs_segment_info *seginfo)
+{
+	return nilfs_list_empty(&seginfo->pseg_list) ? NULL :
+		nilfs_pseg_list_entry(seginfo->pseg_list.next);
+}
+
+struct nilfs_pseg_info *last_pseg(struct nilfs_segment_info *seginfo)
+{
+	return nilfs_list_empty(&seginfo->pseg_list) ? NULL :
+		nilfs_pseg_list_entry(seginfo->pseg_list.prev);
+}
+
+struct nilfs_pseg_info *next_pseg(struct nilfs_segment_info *seginfo,
+				  struct nilfs_pseg_info *pseginfo)
+{
+	return pseginfo->list.next == &seginfo->pseg_list ? NULL :
+		nilfs_pseg_list_entry(pseginfo->list.next);
+}
+
+struct nilfs_pseg_info *prev_pseg(struct nilfs_segment_info *seginfo,
+				  struct nilfs_pseg_info *pseginfo)
+{
+	return pseginfo->list.prev == &seginfo->pseg_list ? NULL :
+		nilfs_pseg_list_entry(pseginfo->list.prev);
+}
+
+struct nilfs_pseg_info *
+lookup_last_super_root(struct nilfs_segment_info *seginfo)
+{
+	struct nilfs_pseg_info *pseginfo;
+
+	for (pseginfo = last_pseg(seginfo); pseginfo != NULL;
+	     pseginfo = prev_pseg(seginfo, pseginfo)) {
+		if (pseginfo->flags & NILFS_SS_SR)
+			return pseginfo;
+	}
+	return NULL;
+}
+
+unsigned long log_length(struct nilfs_segment_info *seginfo)
+{
+	return nilfs_list_empty(&seginfo->pseg_list) ? 0 :
+		nilfs_pseg_list_entry(seginfo->pseg_list.prev)->pseg_start -
+		seginfo->seg_start +
+		nilfs_pseg_list_entry(seginfo->pseg_list.prev)->nblocks;
+}
+
+/*
+ * Routines to get latest checkpoint number
+ */
+static __u64 find_latest_checkpoint(int fd, __u64 cpblocknr, __u64 blkoff)
+{
+	struct nilfs_checkpoint *cp;
+	int i, ncp;
+	__u64 cno = 0;
+
+	read_block(fd, cpblocknr, block_buffer, blocksize);
+	if (blkoff == 0) {
+		cp = block_buffer + first_checkpoint_offset * checkpoint_size;
+		ncp = ncheckpoints_per_block - first_checkpoint_offset;
+	} else {
+		cp = block_buffer;
+		ncp = ncheckpoints_per_block;
+	}
+	
+	for (i = 0; i < ncp; i++, cp = (void *)cp + checkpoint_size) {
+		if (!nilfs_checkpoint_invalid(cp) &&
+		    le64_to_cpu(cp->cp_cno) > cno)
+			cno = le64_to_cpu(cp->cp_cno);
+	}
+	return cno;
+}
+
+static void *next_ss_entry(int fd, __u64 *blocknrp,
+			   unsigned *offsetp, unsigned entry_size)
+{
+	void *p;
+
+	if (*offsetp + entry_size > blocksize) {
+		(*blocknrp)++;
+		read_block(fd, *blocknrp, block_buffer, blocksize);
+		*offsetp = 0;
+	}
+	p = block_buffer + *offsetp;
+	(*offsetp) += entry_size;
+	return p;
+}
+
+static __u64 get_latest_cno(int fd, __u64 pseg_start)
+{
+	struct nilfs_segment_summary *ss;
+	struct nilfs_finfo *finfo;
+	__u32 nfinfo;
+	__u32 nblocks, ndatablk, nnodeblk;
+	__u64 ino;
+	__u64 latest_cno = 0, cno;
+	__u64 blocknr = pseg_start, fblocknr;
+	unsigned offset;
+	int i, j;
+
+	read_block(fd, blocknr, block_buffer, blocksize);
+	ss = block_buffer;
+	nfinfo = le32_to_cpu(ss->ss_nfinfo);
+	offset = le16_to_cpu(ss->ss_bytes);
+	fblocknr = blocknr + DIV_ROUND_UP(le32_to_cpu(ss->ss_sumbytes),
+					  blocksize);
+
+	for (i = 0; i < nfinfo; i++) {
+		finfo = next_ss_entry(fd, &blocknr, &offset, sizeof(*finfo));
+	
+		nblocks = le32_to_cpu(finfo->fi_nblocks);
+		ndatablk = le32_to_cpu(finfo->fi_ndatablk);
+		nnodeblk = nblocks - ndatablk;
+		ino = le64_to_cpu(finfo->fi_ino);
+
+		if (ino == NILFS_DAT_INO) {
+			__le64 *blkoff;
+			struct nilfs_binfo_dat *binfo_dat;
+
+			for (j = 0; j < ndatablk; j++, fblocknr++) {
+				blkoff = next_ss_entry(fd, &blocknr,
+						       &offset,
+						       sizeof(*blkoff));
+			}
+			for (j = 0; j < nnodeblk; j++, fblocknr++) {
+				binfo_dat = next_ss_entry(fd, &blocknr,
+							  &offset,
+							  sizeof(*binfo_dat));
+			}
+		} else {
+			struct nilfs_binfo_v *binfo_v;
+			__le64 *vblocknr;
+
+			for (j = 0; j < ndatablk; j++, fblocknr++) {
+				binfo_v = next_ss_entry(fd, &blocknr,
+							&offset,
+							sizeof(*binfo_v));
+			}
+			if (ino == NILFS_CPFILE_INO && ndatablk > 0) {
+				cno = find_latest_checkpoint(
+					fd, fblocknr - 1,
+					le64_to_cpu(binfo_v->bi_blkoff));
+				if (cno > latest_cno)
+					latest_cno = cno;
+			}
+			for (j = 0; j < nnodeblk; j++, fblocknr++) {
+				vblocknr = next_ss_entry(fd, &blocknr,
+							 &offset,
+							 sizeof(*vblocknr));
+			}
+		}
+	}
+
+	return latest_cno;
+}
+
+__u64 find_latest_cno_in_logical_segment(int fd,
+					 struct nilfs_segment_info *seginfo,
+					 struct nilfs_pseg_info *start)
+{
+	struct nilfs_pseg_info *pseginfo = start ? : last_pseg(seginfo);
+	__u64 cno, latest_cno = 0;
+	__u64 seq;
+	int i = 0;
+
+	if (pseginfo == NULL)
+		return 0;
+
+	get_segment_info(seginfo);
+	do {
+		cno = get_latest_cno(fd, pseginfo->pseg_start);
+		if (cno > latest_cno)
+			latest_cno = cno;
+		
+		if (pseginfo->flags & NILFS_SS_LOGBGN)
+			break;
+
+		pseginfo = prev_pseg(seginfo, pseginfo);
+		if (pseginfo == NULL) {
+			unsigned long segnum = seginfo->segnum;
+
+			if (++i > MAX_SCAN_SEGMENT)
+				break;
+			segnum = (segnum == 0) ? nsegments - 1 : segnum - 1;
+			seq = seginfo->segseq;
+
+			put_segment_info(seginfo);
+			seginfo = load_segment(fd, segnum);
+
+			if (!seginfo || seginfo->segseq != seq - 1)
+				break;
+			pseginfo = last_pseg(seginfo);
+		}
+	} while (pseginfo != NULL && !(pseginfo->flags & NILFS_SS_LOGEND));
+
+	if (seginfo)
+		put_segment_info(seginfo);
+	return latest_cno;
+}
+
+void print_pseg_message(struct nilfs_pseg_ref *pseg_ref, const char *fmt, ...)
+{
+	const char *cp;
+	va_list args;
+
+	va_start(args, fmt);
+	vfprintf(stderr, fmt, args);
+	fprintf(stderr, ": blocknr = %llu\n",
+		(unsigned long long)pseg_ref->blocknr);
+
+	for (cp = fmt; *cp == ' '; cp++)
+		fputc(' ', stderr);
+	fprintf(stderr, "    segnum = %lu, seq = %llu, cno=%llu\n",
+		(unsigned long)pseg_ref->blocknr / blocks_per_segment,
+		(unsigned long long)pseg_ref->seqnum,
+		(unsigned long long)pseg_ref->cno);
+	if (pseg_ref->ctime) {
+		char tmbuf[LINE_BUFFER_SIZE];
+		struct tm tm;
+		time_t t = (time_t)le64_to_cpu(pseg_ref->ctime);
+
+		localtime_r(&t, &tm);
+		strftime(tmbuf, LINE_BUFFER_SIZE, "%F %T", &tm);
+		for (cp = fmt; *cp == ' '; cp++)
+			fputc(' ', stderr);
+		fprintf(stderr, "    creation time = %s\n", tmbuf);
+	}
+	va_end(args);
+}
+
+struct nilfs_pseg_info *
+find_latest_super_root(int fd, unsigned long segnum, __u64 blocknr,
+		       struct nilfs_segment_info **seginfop)
+{
+	struct nilfs_segment_info *seginfo;
+	struct nilfs_segment_info *seginfo_sr = NULL;
+		/* seginfo which has the last super root */
+	struct nilfs_pseg_info *pseg_sr = NULL;
+	int cont = 0, invert = 0;
+	int i;
+
+	seginfo = load_segment(fd, segnum);
+	if (seginfo) {
+		pseg_sr = lookup_last_super_root(seginfo);
+		if (pseg_sr)
+			seginfo_sr = get_segment_info(seginfo);
+
+		if (blocknr < seginfo->seg_start + log_length(seginfo))
+			cont = 1;
+	}
+
+	for (i = 0; i < MAX_SCAN_SEGMENT; i++) {
+		struct nilfs_segment_info *seginfo2;
+
+		/*
+		 * Look into the previous segment.
+		 *
+		 * This code depends on the current GC policy; discontinuously
+		 * allocated segments are not supported.
+		 */
+		if (!(i & SCAN_SEGMENT_MASK))
+			fputc('.', stderr);
+		segnum = (segnum == 0) ? nsegments - 1 : segnum - 1;
+
+		seginfo2 = load_segment(fd, segnum);
+		if (!seginfo2) {
+			if (pseg_sr && cont) {
+				pseg_sr = NULL;
+				put_segment_info(seginfo_sr);
+				seginfo_sr = NULL;
+			}
+			cont = 0;
+			if (seginfo) {
+				put_segment_info(seginfo);
+				seginfo = NULL;
+			}
+			continue;
+		}
+
+		if (!seginfo) {
+			seginfo = seginfo2;
+			seginfo2 = NULL;
+
+			if (pseg_sr)
+				put_segment_info(seginfo_sr);
+			pseg_sr = lookup_last_super_root(seginfo);
+			if (pseg_sr)
+				seginfo_sr = get_segment_info(seginfo);
+			continue;
+		}
+
+		if (seginfo2->segseq + 1 != seginfo->segseq)
+			cont = 0;
+
+		if (seginfo2->segseq > seginfo->segseq) {
+			invert++;
+			if (pseg_sr) {
+				pseg_sr = NULL;
+				put_segment_info(seginfo_sr);
+				seginfo_sr = NULL;
+			}
+		}
+		if (invert && !pseg_sr) {
+			pseg_sr = lookup_last_super_root(seginfo2);
+			if (pseg_sr) {
+				put_segment_info(seginfo);
+				*seginfop = seginfo2;
+				fputc('\n', stderr);
+				return pseg_sr; /* latest segment was found */
+			}
+		}
+
+		if (!cont && !pseg_sr) {
+			pseg_sr = lookup_last_super_root(seginfo2);
+			if (pseg_sr)
+				seginfo_sr = get_segment_info(seginfo2);
+		}
+
+		put_segment_info(seginfo);
+		seginfo = seginfo2;
+		seginfo2 = NULL;
+	}
+	fputc('\n', stderr);
+	if (seginfo)
+		put_segment_info(seginfo);
+
+	if (pseg_sr && !cont) {
+		*seginfop = seginfo_sr;
+		return pseg_sr; /* regard second-ranking candidate
+				   as the latest segment */
+	}
+	if (seginfo_sr)
+		put_segment_info(seginfo_sr);
+	return NULL;
+}
+
+static void check_mount(int fd, const char *device)
+{
+	FILE *fp;
+	char line[LINE_BUFFER_SIZE];
+
+	fp = fopen(MOUNTS, "r");
+	if (fp == NULL)
+		die("cannot open %s!", MOUNTS);
+
+	while (fgets(line, LINE_BUFFER_SIZE, fp) != NULL) {
+		if (strncmp(strtok(line, " "), device, strlen(device)) == 0) {
+			fclose(fp);
+			die("%s is currently mounted.", device);
+		}
+	}
+	fclose(fp);
+}
+
+static void check_super_block(struct nilfs_super_block *sb)
+{
+	char tmbuf[LINE_BUFFER_SIZE];
+	struct tm tm;
+	time_t t;
+	__u32 sbsum, crc;
+	int sb_bytes = le16_to_cpu(sb->s_bytes);
+
+	if (le16_to_cpu(sb->s_magic) != NILFS_SUPER_MAGIC)
+		die("Not nilfs2 partition");
+
+	fprintf(stderr, "Super-block:\n");
+
+	crc_seed = le32_to_cpu(sb->s_crc_seed);
+	sbsum = le32_to_cpu(sb->s_sum);
+
+	sb->s_sum = 0;
+	crc = nilfs_crc32(crc_seed, (unsigned char *)sb, sb_bytes);
+	sb->s_sum = cpu_to_le32(sb->s_sum);
+
+	fprintf(stderr, "    revision = %d.%d, checksum = %s\n",
+		le32_to_cpu(sb->s_rev_level),
+		le16_to_cpu(sb->s_minor_rev_level),
+		crc == sbsum ? "OK" : "error");
+
+	blocksize = 1 << (le32_to_cpu(sb->s_log_block_size) + 10);
+	blocks_per_segment = le32_to_cpu(sb->s_blocks_per_segment);
+	first_data_block = le64_to_cpu(sb->s_first_data_block);
+	nsegments = le64_to_cpu(sb->s_nsegments);
+	checkpoint_size = le16_to_cpu(sb->s_checkpoint_size);
+
+	first_checkpoint_offset =
+		DIV_ROUND_UP(sizeof(struct nilfs_cpfile_header),
+			     checkpoint_size);
+	ncheckpoints_per_block = blocksize / checkpoint_size;
+
+	t = (time_t)le64_to_cpu(sb->s_wtime);
+	localtime_r(&t, &tm);
+	strftime(tmbuf, LINE_BUFFER_SIZE, "%F %T", &tm);
+
+	fprintf(stderr, "    blocksize = %d\n", blocksize);
+	fprintf(stderr, "    write time = %s\n", tmbuf);
+}
+
+static void commit_super_block(struct nilfs_super_block *sb,
+			       struct nilfs_pseg_ref *pseg_ref)
+{
+	__u32 sbsum;
+	int sb_bytes = le16_to_cpu(sb->s_bytes);
+
+	sb->s_last_pseg = cpu_to_le64(pseg_ref->blocknr);
+	sb->s_last_seq = cpu_to_le64(pseg_ref->seqnum);
+	sb->s_last_cno = cpu_to_le64(pseg_ref->cno);
+
+	sb->s_state = cpu_to_le16(le16_to_cpu(sb->s_state) & ~NILFS_VALID_FS);
+
+	/* fill in crc */
+	sb->s_sum = 0;
+	sbsum = nilfs_crc32(crc_seed, (unsigned char *)sb, sb_bytes);
+	sb->s_sum = cpu_to_le32(sbsum);
+}
+
+static int nilfs_do_recovery(int fd, struct nilfs_pseg_ref *pseg_ref)
+{
+	struct nilfs_pseg_info *pseginfo;
+	struct nilfs_segment_info *seginfo;
+	unsigned long segnum;
+	int ret = 0;
+
+	init_block_buffer();
+	init_segment_cache();
+
+	/*
+	 * check the partial segment pointed by superblock.
+	 */
+	segnum = pseg_ref->blocknr / blocks_per_segment;
+	seginfo = load_segment(fd, segnum);
+	if (seginfo) {
+		pseginfo = lookup_pseg(seginfo, pseg_ref->blocknr);
+		if (pseginfo &&
+		    seginfo->segseq == pseg_ref->seqnum &&
+		    pseginfo->flags & NILFS_SS_SR) {
+			pseg_ref->ctime = le64_to_cpu(pseginfo->segsum.ss_create);
+
+			print_pseg_message(pseg_ref,
+					   "Valid segment is pointed by "
+					   "superblock (No change needed)");
+			goto out;
+		}
+		put_segment_info(seginfo);
+	}
+
+	/*
+	 * check logs in the current and prior full segments.
+	 */
+	fprintf(stderr, "The latest segment is lost. "
+		"Trying rollback recovery..\n");
+
+	pseginfo = find_latest_super_root(fd, segnum, pseg_ref->blocknr,
+					  &seginfo);
+	if (!pseginfo)
+		die("Cannot find super root");
+
+	pseg_ref->blocknr = pseginfo->pseg_start;
+	pseg_ref->seqnum = seginfo->segseq;
+	pseg_ref->ctime = le64_to_cpu(pseginfo->segsum.ss_create);
+	
+	fprintf(stderr, "Searching the latest checkpoint.\n");
+	pseg_ref->cno = find_latest_cno_in_logical_segment(
+		fd, seginfo, pseginfo);
+	if (pseg_ref->cno == 0)
+		die("Cannot identify the latest checkpoint");
+
+	print_pseg_message(pseg_ref, "Selected segment");
+	ret = 1;
+ out:
+	destroy_segment_cache();
+	destroy_block_buffer();
+	return ret;
+}
+
+static void nilfs_fsck(const char *device)
+{
+	struct nilfs_super_block sb;
+	struct nilfs_pseg_ref pseg_ref;
+	ssize_t size;
+	int c;
+
+	if ((devfd = open(device, O_RDONLY | O_LARGEFILE)) < 0)
+		die("cannot open device %s", device);
+
+	size = pread(devfd, &sb, sizeof(sb), NILFS_SB_OFFSET_BYTES);
+	if (size < sizeof(sb))
+		die("cannot read super block (device=%s)", device);
+
+	check_mount(devfd, device);
+	check_super_block(&sb);
+
+	pseg_ref.blocknr = le64_to_cpu(sb.s_last_pseg);
+	pseg_ref.seqnum = le64_to_cpu(sb.s_last_seq);
+	pseg_ref.cno = le64_to_cpu(sb.s_last_cno);
+	pseg_ref.ctime = 0;
+	print_pseg_message(&pseg_ref, "    indicated partial segment");
+	fputc('\n', stderr);
+
+	if (le16_to_cpu(sb.s_state) & NILFS_VALID_FS) {
+		fprintf(stderr, "Clean FS.\n");
+		goto out_clean;
+	}
+
+	if (nilfs_do_recovery(devfd, &pseg_ref) == 0)
+		goto out;
+
+	/*
+	 * Reopen device to update superblock
+	 */
+	close(devfd);
+	devfd = -1;
+	if ((devfd = open(device, O_RDWR | O_LARGEFILE)) < 0)
+		die("cannot open device %s in read/write mode",
+		    device);
+
+	fprintf(stderr, "Do you wish to overwrite super block (y/N)? ");
+	if ((c = getchar()) == 'y') {
+		commit_super_block(&sb, &pseg_ref);
+		size = pwrite(devfd, &sb, sizeof(sb), NILFS_SB_OFFSET_BYTES);
+		if (size < sizeof(sb))
+			die("cannot write out super block (device=%s)", device);
+		fsync(devfd);
+	}
+ out:
+	fprintf(stderr, "Recovery will complete on mount.\n");
+
+ out_clean:
+	close(devfd);
+}
+
+int main(int argc, char *argv[])
+{
+	char *device;
+
+	if ((progname = strrchr(argv[0], '/')) != NULL)
+		progname++;
+	else
+		progname = argv[0];
+
+	if (argc < 2) {
+		fprintf(stderr, "Usage: %s <device>\n", progname);
+		exit(0);
+	}
+	
+	device = argv[1];
+	nilfs_fsck(device);
+
+	return 0;
+}
-- 
1.5.6.5

Re: nilfs2 fails to mount after io/error

[Reinoud Zandijk]

Hi folks, hi Ryusuke,

On Sun, Dec 07, 2008 at 04:05:09AM +0900, Ryusuke Konishi wrote:
> On Thu, 4 Dec 2008 14:04:31 +0100, Adrian Ulrich wrote:
> Sorry to keep you waiting. ;-)
> 
> I will attach a patch against nilfs2-utils-2.0.6, which adds the
> tool 'fsck0.nilfs2'.
> 
>  # fsck0.nilfs2 <device name>
> 
> will, hopefully, correct your partition.
> (I'm not sure because it's not yet tested enough)

looks nice Ryusuke. Could the problem be averted though by *only* writing the
superblock on dismount or when the origional block would be overwritten.
Seperated by a disc flush of course :)

With regards,
Reinoud

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

On Sat, 6 Dec 2008 20:53:08 +0100, Reinoud Zandijk <reinoud-S783fYmB3Ccdnm+yROfE0A@public.gmane.org> wrote:
> Hi folks, hi Ryusuke,
> 
> On Sun, Dec 07, 2008 at 04:05:09AM +0900, Ryusuke Konishi wrote:
> > On Thu, 4 Dec 2008 14:04:31 +0100, Adrian Ulrich wrote:
> > Sorry to keep you waiting. ;-)
> > 
> > I will attach a patch against nilfs2-utils-2.0.6, which adds the
> > tool 'fsck0.nilfs2'.
> > 
> >  # fsck0.nilfs2 <device name>
> > 
> > will, hopefully, correct your partition.
> > (I'm not sure because it's not yet tested enough)
> 
> looks nice Ryusuke. Could the problem be averted though by *only* writing the
> superblock on dismount or when the origional block would be overwritten.
> Seperated by a disc flush of course :)

Well, unless NILFS itself had destroyed data by bugs, the filesystem
should have a very good chance to be salvaged by only modifying the
superblock like this tool does.

The GC may have reused data blocks needed by the re-selected old
checkpoint, but it is likely very few since there is time lag from
when blocks are marked unused to when they are actually overwritten
under the current GC policy (i.e. FIFO order poclicy).

Is this answering your question?

Regards,
Ryusuke Konishi

Re: nilfs2 fails to mount after io/error

[Adrian Ulrich]

Hi Ryusuke,

> will, hopefully, correct your partition.
> (I'm not sure because it's not yet tested enough)

Thanks a lot for the patch: Running fsck0 fixed the problem and i was
able to mount the filesystem.

So far everything looks fine (data is stll there and i can write to the filesystem),
but: I got this in my syslog:

 segctord starting. Construction interval = 5 seconds, CP frequency < 30 seconds
 NILFS warning: mounting fs with errors


Should i worry about this?


Regards,
 Adrian

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

Hi Adrian,
On Sun, 7 Dec 2008 13:23:35 +0100, Adrian Ulrich wrote:
> Hi Ryusuke,
> 
> > will, hopefully, correct your partition.
> > (I'm not sure because it's not yet tested enough)
> 
> Thanks a lot for the patch: Running fsck0 fixed the problem and i was
> able to mount the filesystem.
> 
> So far everything looks fine (data is stll there and i can write to the filesystem),

That's the good news!

> but: I got this in my syslog:
> 
>  segctord starting. Construction interval = 5 seconds, CP frequency < 30 seconds
>  NILFS warning: mounting fs with errors
> 
> 
> Should i worry about this?

This message means that the filesystem has detected some sort of
inconsistency, but you don't have to worry in your case; the error
flag can be marked imprecisely when a filesystem has detected I/O
errors even if no inconsistency occurs.

This flag is not cleared by fsck0 because I haven't yet implemented
the feature which fully verifies coherency of a filesystem. ;)

Well, if you can't ignore the message, I can add a dummy force check
option that does nothing but drop the flag.

Regards,
Ryusuke Konishi

Re: nilfs2 fails to mount after io/error

[Adrian Ulrich]

Hi,

> Well, if you can't ignore the message, I can add a dummy force check
> option that does nothing but drop the flag.

I can life with it.. Just wanted to ask if it's 'normal' :-)

Regards,
 Adrian

Re: nilfs2 fails to mount after io/error

[Ryusuke Konishi]

On Mon, 8 Dec 2008 17:57:32 +0100, Adrian Ulrich wrote:
> Hi,
> 
> > Well, if you can't ignore the message, I can add a dummy force check
> > option that does nothing but drop the flag.
> 
> I can life with it.. Just wanted to ask if it's 'normal' :-)
> 
> Regards,
>  Adrian

All right, I will support the option if someone wants it or
after I will have implemented meaningful checks.

Thanks,
Ryusuke