[PATCH 1/2] signals: kill(-1) should only signal processes in the same namespace

Eric W. Biederman ebiederm at xmission.com
Thu Jul 17 10:45:13 PDT 2008


"Daniel Hokka Zakrisson" <daniel at hozac.com> writes:

> Pavel Emelyanov wrote:
>> Daniel Hokka Zakrisson wrote:
>>> While moving Linux-VServer to using pid namespaces, I noticed that
>>> kill(-1) from inside a pid namespace is currently signalling every
>>> process in the entire system, including processes that are otherwise
>>> unreachable from the current process.
>>
>> This is not a "news" actually, buy anyway - thanks :)
>
> And yet nobody's fixed it... Kind of a critical thing, if you actually
> want to use them, since most distribution's rc-scripts do a kill(-1,
> SIGTERM), followed by kill(-1, SIGKILL) when halting (which, needless to
> say, would be very bad).
>
>>> This patch fixes it by making sure that only processes which are in
>>> the same pid namespace as current get signalled.
>>
>> This is to be done, indeed, but I do not like the proposed implementation,
>> since you have to walk all the tasks in the system (under tasklist_lock,
>> by the way) to search for a couple of interesting ones. Better look at how
>> zap_pid_ns_processes works (by the way - I saw some patch doing so some
>> time ago).
>
> The way zap_pid_ns_processes does it is worse, since it signals every
> thread in the namespace rather than every thread group. So either we walk
> the global tasklist, or we create a per-namespace one. Is that what we
> want?

Can you please introduce kill_pidns_info and have both
kill_something_info and zap_pid_ns_processes call this common
function?

We want to walk the set of all pids in a pid namespace.  /proc does
this and it is the recommended idiom.  If walking all of the pids in a
pid namespace is not fast enough we can accelerate that.

You are correct signalling every thread in a namespace is worse, in
fact it is semantically incorrect.  zap_pid_ns_processes gets away
with it because it is sending SIGKILL.   Therefore kill_pidns_info 
should skip sending a signal to every task that is not the
thread_group_leader.

We need to hold the tasklist_lock to prevent new processes from
joining the list of all processes.  Otherwise we could run the code
under the rcu_read_lock.

Eric


More information about the Containers mailing list