Jim,

Your assumption that the same file descriptor is being re-opened is
wrong!
The file descriptor retrieved via /proc is a new one. It is not the
same as the
initial read-only.

Do a strace on your test and you will see that the 'file descriptor'
in /proc
will be accessed as an ordinairy file. After checking the directory
permissions
of that particular file an open will be performed on '/proc/self/fd/0'
and a new
O_WRONLY file descriptor is being created.

As Martin Rex already explained yesterday, /proc is all virtual.
The item referred as fd in /proc is not a real file descriptor and as
of that, that 'not-tfor-real file descriptor' is also not re-opend and
so does
not become read-write.

The entire discussion about the file descriptor behavior the past days,
including your statement below is all based on false assumptions...
I'll show you a snip out of my strace of the original scenario, being
performed by
Pavel. But the same mechanism is being performed by you, Jim, in the
following step:
# su nobody -c 'echo "hacked" >/proc/self/fd/0' < /dir/file.txt

All you do is just open the FILE via the path of /proc, not via the
assumed
path via /tmp (or /dir in your example) nor access the assumed read-
only fd being
presented via /proc. Therefor it's totally of no influence what you do
with the original
directory permission. File access has nothing to do with directory
permissions...!

Imagen:
- a house surrounded with a fence with all doors unlocked (file with
perm 0666)
- a drive-way leads to the gate in the fence and the gate is unlocked
(dir with perms 777)
- next we put a lock on the gate and don't give guest the key (dir
with perms 700)
- guest cannot access the house because he can't pass the gate
- now we take an airplane and parachute guest straight into the
perimeter of the fence (/proc access)
- guest can access the house (write the file), because the house has
all doors unlocked

There is no such mechanism that gives guest the ability to break the
lock / steel or duplicate
the key of the lock on the gate (circumvent the original directory
permissions or
upgrade the read-only fd to be read-write), because guest never takes
that path via the drive-way and gate
if he accesses the file via /proc

Please do your own strace of the mentioned scenario's, so you'll see
it with your own eyes.
People tent to debate mechanisms, based on the assumption that a
certain mechanisme is
being involved here. But the assumed mechanisms are not in play in
this particular case.

Below you'll find the parts of the strace output that proves it all;

7649 stat64("/proc", {st_mode=S_IFDIR|0555, st_size=0, ...}) = 0
7649 stat64("/proc/self", {st_mode=S_IFDIR|0555, st_size=0, ...}) = 0
7649 stat64("/proc/self/fd", {st_mode=S_IFDIR|0500, st_size=0, ...})
= 0
7649 stat64("/proc/self/fd/3", {st_mode=S_IFREG|0666,
st_size=35, ...}) = 0
<snip>
7649 rt_sigaction(SIGINT, {0x8085067, [], 0}, {0x8085067, [], 0}, 8)
= 0
7649 time(NULL) = 1256668021
7649 open("/proc/self/fd/3", O_WRONLY|O_CREAT|O_TRUNC|O_LARGEFILE,
0666) = 4
7649 fcntl64(1, F_GETFD) = 0
7649 fcntl64(1, F_DUPFD, 10) = 10
7649 fcntl64(1, F_GETFD) = 0
7649 fcntl64(10, F_SETFD, FD_CLOEXEC) = 0
7649 close(1) = 0
7649 dup2(4, 1) = 1
7649 close(4) = 0
7649 write(1, "got you\n", 8) = 8
7649 close(1) = 0

Best regards,
Marco

On 29 okt 2009, at 21:10, Jim Paris wrote:

>>> 0700 mode from the origin, you would be right, and procfs wouldn't
>>> allow
>>> opening files in that directory too, but if you let others to
>>> traverse
>>> that directory and open your believed to be secure files from the
>>> origin,
>>> it's your fault.
>>
>> I can do the example with fd passing and 700 directory, but it would
>> be lot of C code. Feel free to play, my example was not nearly the
>> only way to demonstrate it, and no, it was not racy.
>
> Here is an example that shows the behavior where a passed read-only fd
> can become read-write by reopening it through /proc, when file
> permissions allow it (but directory permissions do not):
>
> $ sudo su
> # mkdir -m 0700 /dir
> # echo "safe" > /dir/file.txt
> # chmod 0666 /dir/file.txt
> # ls -al /dir
> total 12
> drwx------ 2 root root 4096 2009-10-29 00:28 .
> drwxr-xr-x 27 root root 4096 2009-10-29 00:28 ..
> -rw-rw-rw- 1 root root 7 2009-10-29 00:43 file.txt
> # cat /dir/file.txt
> safe
>
> Now user "nobody" cannot read or write this file:
>
> # su nobody -c 'cat /dir/file.txt'
> sh: /dir/file.txt: Permission denied
> # su nobody -c 'echo "hacked" > /dir/file.txt'
> sh: /dir/file.txt: Permission denied
> # cat /dir/file.txt
> safe
>
> If we provide an open read-only file descriptor (as stdin, fd 0), they
> can read it:
>
> # su nobody -c 'cat <&0' < /dir/file.txt
> safe
>
> But they still can't write to this descriptor:
>
> # su nobody -c 'echo "hacked" >&0' < /dir/file.txt
> sh: line 0: echo: write error: Bad file descriptor
>
> Unless we re-open the file using the magic link in /proc:
>
> # su nobody -c 'echo "hacked" >/proc/self/fd/0' < /dir/file.txt
> # cat /dir/file.txt
> hacked
>
> Again, debatable whether this is a bug, but it's certainly
> non-obvious. There is no other way (that I'm aware) for the "nobody"
> user to gain write access to /dir/file.txt, even when given a
> read-only fd, without using /proc.
>
> -jim

[ reply ]