Re: [RFC][PATCH] exec: Freeze the other threads during a multi-threaded exec

[Eric W. Biederman]

ebiederm@xxxxxxxxxxxx (Eric W. Biederman) writes:

> Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> writes:
>
>> It also makes for a possible _huge_ latency regression for execve(),
>> since freezing really has never been a very low-latency operation.
>>
>> Other threads doing IO can now basically block execve() for a long
>> long long time.
>
> Hmm.  Potentially.  The synchronization with the other threads must
> happen in a multi-threaded exec in de_thread.
>
> So I need to look at the differences between where de_thread thread
> can kill a thread and the freezer can not freeze a thread.  I am hoping
> that the freezer has already instrumented most of those sleeps but I
> admit I have not looked yet.

Alright I have looked at the freezer a bit more and I now see that the
point of marking things freezable is for kernel threads rather that user
space threads.  I think there are 5 maybe 6 places the code sleeps
reachable by userspace threads that are marked as freezable and most
of those are callable from get_signal.

For exec all I care about are user space threads.  So it appears the
freezer infrastructure adds very little.

Now to see if I can find another way to divert a task into a slow path
as it wakes up, so I don't need to manually wrap all of the sleeping
calls.  Something that plays nice with the scheduler.

Eric