Re: [RFC PATCH glibc 3/8] nptl: Start new threads with all signals blocked [BZ #25098]

[Carlos O'Donell via Libc-alpha <libc-alpha@sourceware.org>]

On 3/23/20 1:05 PM, Christian Brauner wrote:
> On Mon, Mar 23, 2020 at 01:02:12PM -0400, Mathieu Desnoyers via Libc-alpha wrote:
>> ----- On Mar 23, 2020, at 11:31 AM, Christian Brauner christian.brauner@ubuntu.com wrote:
>>
>>> On Mon, Mar 23, 2020 at 09:16:02AM -0400, Mathieu Desnoyers via Libc-alpha
>>> wrote:
>>>> From: Florian Weimer <fweimer@redhat.com>
>>>>
>>>> New threads inherit the signal mask from the current thread.  This
>>>> means that signal handlers can run on the newly created thread
>>>> immediately after the kernel has created the userspace thread, even
>>>> before glibc has initialized the TCB.  Consequently, new threads can
>>>> observe uninitialized ctype data, among other things.
>>>>
>>>> To address this, block all signals before starting the thread, and
>>>> pass the original signal mask to the start routine wrapper.  On the
>>>> new thread, first perform all thread initialization, and then unblock
>>>> signals.
>>>>
>>>> The cost of doing this is two rt_sigprocmask system calls on the old
>>>> thread, and one rt_sigprocmask system call on the new thread.  (If
>>>> there was a way to clone a new thread with a signals disabled, this
>>>
>>> This could be a new clone3() flag. If someone wants to send a patch I'd
>>> take it.
>>
>> I agree that it would be a nice improvement to alleviate the overhead of
>> tweaking the signal masks on thread creation. I suspect the code proposed in
>> this patch is still needed, because glibc would have to support the currently
>> existing kernels. The improvement you envision involves adding this new flag
>> into the Linux kernel clone3 system call and then wiring up glibc support.
>>
>> I don't expect this should delay rseq integration into glibc ?
> 
> Oh no, I wasn't trying to say that rseq() in glibc should be blocked on
> this, not at all. I was just saying we should probably do this since
> it's a valuable improvement.

That's up to you to decide, and it depends on the workload. I'd be hesitant
to spend a flag bit on this. To cleanup the glibc code would require you to
(a) implement this in the kernel
(b) have glibc move the minimum kernel line above the new kernel that has this
We're talking a minor cleanup that would take 10-15 years to arrive... if ever.
Moving the minimum kernel baseline has container workload impacts that are
going to become real problems soon.

Mathieu, Florian, and I talked about this very issue at GNU Tools
Cauldron 2019 (Montreal). We agreed that even if we could fix it immediately
we would still need the above implementation.

IMO you are doing something wrong if this shows up in your profiling though.
You should not be starting up and killing threads that fast. You should be
using thread pools etc. The kernel reaping of tasks is quite slow and even
in glibc we often hit EAGAIN limits on heavily loaded test boxes.

The above code is required for correctness so certain state is not observable,
but it need not be that fast. Someone correct me if I'm wrong though and I've
missed a use case.

-- 
Cheers,
Carlos.