loopback-mounted ext4 sees hole-filling (on rawhide, but not F15)

loopback-mounted ext4 sees hole-filling (on rawhide, but not F15)

[Jim Meyering]

When I run the following script on rawhide (2.6.39-0.rc3.git2.0.fc16.x86_64)
it shows that a file with holes sometimes ends up with some non-NUL bytes
where the holes should be.

For best results, run it in an empty directory, preferably on a tmpfs FS.
I've run it on ext4, too.

It creates a 4MB ext4 file system, mounts it, and writes two files,
one with holes, the other identical, but with NULs instead,
and then runs sync, compares the two files, and unmounts
and removes the temporary file and the mount point directory.

The key is that it unlinks the backing file after mounting
and before writing the files.

=====================================
#!/bin/sh
set -e
dd if=/dev/zero of=blob bs=4k count=1000 >/dev/null 2>&1
mkdir mnt
mkfs -q -t ext4 -F blob
mount -oloop blob mnt

# Removing the backing file is key.
rm -f blob
cwd=$PWD
cd mnt

# Create a reference file.  Just like the following one,
# but with explicit NULs in place of holes.
perl -e '$n=1024; for (1..71) { print "\0"x$n, chr($_)x$n };' \
  -e 'close *STDOUT or die "$!"' > ref

# Seek 1KB, write 1KB of data, seek 1KB, write 1KB of data, etc....
perl -e '$n = 1 * 1024; *F = *STDOUT;' \
  -e 'for (1..71) { sysseek (*F, $n, 1)' \
  -e '&& syswrite (*F, chr($_)x$n) or die "$!"}' > j1

# filefrag -vs j1

sync
cmp -s ref j1 && fail=0 || { cmp -l ref j1|head -20; fail=1; }

cd /; umount "$cwd/mnt"
rm -rf "$cwd/blob" "$cwd/mnt"

exit $fail
=====================================

Interestingly, if you remove the "sync", it exits 0 every time.

Beware that sometimes you (accidentally?) get
all NUL blocks, so the script does occasionally exit 0.
To demonstrate, I turned off the cmp -l ...|head bit and ran it
for a while, printing only exit status:

    $ while :; do /t/loop-bug; printf $?; done
    110101111111111111111111111111101111110111111111111111101111111110101111101^C
    [Exit 130 (INT)]

Jim

P.S., this all started because I accidentally removed a backing
file, and that made some in-progress FIEMAP-related tests fail.

PPS, here's sample output, when it exits nonzero:

  # bash /t/bug-demo                                                    :
       9   0  41
      17   0 240
      18   0 216
      19   0 125
      20   0   1
      29   0   3
      30   0   3
      31   0   3
      32   0   3
      33   0 142
      34   0  40
      35   0 155
      36   0 156
      37   0 164
      39   0 142
      41   0  41
      49   0 146
      50   0 141
      51   0 151
      52   0 154

Re: loopback-mounted ext4 sees hole-filling (on rawhide, but not F15)

[Eric Sandeen]

On 4/19/11 11:16 AM, Jim Meyering wrote:
> When I run the following script on rawhide (2.6.39-0.rc3.git2.0.fc16.x86_64)
> it shows that a file with holes sometimes ends up with some non-NUL bytes
> where the holes should be.
> 
> For best results, run it in an empty directory, preferably on a tmpfs FS.
> I've run it on ext4, too.
> 
> It creates a 4MB ext4 file system, mounts it, and writes two files,
> one with holes, the other identical, but with NULs instead,
> and then runs sync, compares the two files, and unmounts
> and removes the temporary file and the mount point directory.
> 
> The key is that it unlinks the backing file after mounting
> and before writing the files.

I've found that disabling mblk_io_submit by adding ",nomblk_io_submit"
to the mount options in the testcase fixes this corruption case.

(6fd7a46781999c32f423025767e43b349b967d57 ext4: enable mblk_io_submit by default
turned it back on by default after the last corruption was found & fixed)

-Eric

> =====================================
> #!/bin/sh
> set -e
> dd if=/dev/zero of=blob bs=4k count=1000 >/dev/null 2>&1
> mkdir mnt
> mkfs -q -t ext4 -F blob
> mount -oloop blob mnt
> 
> # Removing the backing file is key.
> rm -f blob
> cwd=$PWD
> cd mnt
> 
> # Create a reference file.  Just like the following one,
> # but with explicit NULs in place of holes.
> perl -e '$n=1024; for (1..71) { print "\0"x$n, chr($_)x$n };' \
>   -e 'close *STDOUT or die "$!"' > ref
> 
> # Seek 1KB, write 1KB of data, seek 1KB, write 1KB of data, etc....
> perl -e '$n = 1 * 1024; *F = *STDOUT;' \
>   -e 'for (1..71) { sysseek (*F, $n, 1)' \
>   -e '&& syswrite (*F, chr($_)x$n) or die "$!"}' > j1
> 
> # filefrag -vs j1
> 
> sync
> cmp -s ref j1 && fail=0 || { cmp -l ref j1|head -20; fail=1; }
> 
> cd /; umount "$cwd/mnt"
> rm -rf "$cwd/blob" "$cwd/mnt"
> 
> exit $fail
> =====================================
> 
> Interestingly, if you remove the "sync", it exits 0 every time.
> 
> Beware that sometimes you (accidentally?) get
> all NUL blocks, so the script does occasionally exit 0.
> To demonstrate, I turned off the cmp -l ...|head bit and ran it
> for a while, printing only exit status:
> 
>     $ while :; do /t/loop-bug; printf $?; done
>     110101111111111111111111111111101111110111111111111111101111111110101111101^C
>     [Exit 130 (INT)]
> 
> Jim
> 
> P.S., this all started because I accidentally removed a backing
> file, and that made some in-progress FIEMAP-related tests fail.
> 
> PPS, here's sample output, when it exits nonzero:
> 
>   # bash /t/bug-demo                                                    :
>        9   0  41
>       17   0 240
>       18   0 216
>       19   0 125
>       20   0   1
>       29   0   3
>       30   0   3
>       31   0   3
>       32   0   3
>       33   0 142
>       34   0  40
>       35   0 155
>       36   0 156
>       37   0 164
>       39   0 142
>       41   0  41
>       49   0 146
>       50   0 141
>       51   0 151
>       52   0 154
> --
> To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html