[PATCH 1/2] ash: Set the CLOEXEC bit in fcntl_F_DUPFD

[Mark Marshall]

Hi.

I have a small reproducer.  Place the following script on the target device:

z1.sh
=====
sleep 20 </dev/null >/dev/null 2>/dev/null &
PID=$!
echo "self"
ls -lv /proc/self/fd
echo "child"
ls -lv /proc/${PID}/fd

Then, run this script via ssh/sshd from another machine, with 1.24.1
the output is as follows:

marmar01 at marmar13:~/develop/oeltron$ ssh root at 192.168.0.252 "/tmp/z1.sh"
self
lr-x------    1 root     root            64 Mar 30 10:14 0 -> pipe:[7874]
l-wx------    1 root     root            64 Mar 30 10:14 1 -> pipe:[7875]
l-wx------    1 root     root            64 Mar 30 10:14 2 -> pipe:[7876]
lr-x------    1 root     root            64 Mar 30 10:14 3
ls: /proc/self/fd/3: cannot read link: No such file or directory
child
lr-x------    1 root     root            64 Mar 30 10:14 0 -> /dev/null
l-wx------    1 root     root            64 Mar 30 10:14 1 -> /dev/null
l-wx------    1 root     root            64 Mar 30 10:14 2 -> /dev/null


With 1.28.1 the output is this:

marmar01 at marmar13:~/develop/oeltron$ ssh root at 192.168.0.252 "/tmp/z1.sh"
self
lr-x------    1 root     root            64 Mar 30 10:16 0 -> pipe:[7736]
l-wx------    1 root     root            64 Mar 30 10:16 1 -> pipe:[7737]
l-wx------    1 root     root            64 Mar 30 10:16 2 -> pipe:[7738]
lr-x------    1 root     root            64 Mar 30 10:16 3
ls: /proc/self/fd/3: cannot read link: No such file or directory
child
lr-x------    1 root     root            64 Mar 30 10:16 0 -> /dev/null
l-wx------    1 root     root            64 Mar 30 10:16 1 -> /dev/null
l-wx------    1 root     root            64 Mar 30 10:16 2 -> /dev/null
lr-x------    1 root     root            64 Mar 30 10:16 10 -> /dev/null
l-wx------    1 root     root            64 Mar 30 10:16 11 -> pipe:[7737]
l-wx------    1 root     root            64 Mar 30 10:16 12 -> pipe:[7738]

Additionally (and more importantly) with the broken 1.28.1 case the
ssh session waits until the sleep
finishes (20 seconds).  In the real case that I care about the program
being run is not a sleep, but is
a daemon that never exits.

My patch, on top of 1.28.1, makes things work like 1.24.1 again.

Regards,
Mark


On 28 March 2018 at 17:06, Denys Vlasenko <vda.linux at googlemail.com> wrote:
> On Mon, Mar 19, 2018 at 10:31 AM,  <mark.marshall at omicronenergy.com> wrote:
>> From: Mark Marshall <mark.marshall at omicronenergy.com>
>>
>> The function fcntl_F_DUPFD was introduced as some re-factoring at commit
>> 035486c7500c09616a6c1040d8e70923532a5c2d
>> "ash: significant overhaul of redirect saving logic".
>> The old code used to set the CLOEXEC bit, but this got lost in the
>> re-factoring, so this change makes the new code do the same.
>>
>> This was found because if a script was run via ssh, and this script ran
>> a second script, redirecting its stdin and stdout, and putting it into
>> the background, the ssh would hang waiting for the backgrounded script
>> to terminate.  From /proc/999/fd it can be seen that the background script
>> still has the stdin pipe from sshd open (at file handle 10+).
>
> Please provide a reproducer. I tried this:
>
> z1
> ==
> ./busybox ash z2 </dev/null >/dev/null 2>/dev/null &
> echo Ok
>
> z2
> ==
> ls -l /proc/$$/fd   >z2.out1 2>&1
> ls -l /proc/self/fd >z2.out2 2>&1
>
> I do see some extra fds.
>
> The thing is, when I try it, there is no difference before
> and after commit 035486c7500c09616a6c1040d8e70923532a5c2d.
> _______________________________________________
> busybox mailing list
> busybox at busybox.net
> http://lists.busybox.net/mailman/listinfo/busybox