Re: [PATCH 2/3] Linux: Use rseq in sched_getcpu if available (v9)

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

----- On Jul 6, 2020, at 2:11 PM, Florian Weimer fweimer@redhat.com wrote:

> * Mathieu Desnoyers:
> 
>> ----- On Jul 6, 2020, at 1:50 PM, Florian Weimer fweimer@redhat.com wrote:
>>
>>> * Mathieu Desnoyers:
>>> 
>>>> Now we need to discuss how we introduce that fix in a way that will
>>>> allow user-space to trust the __rseq_abi.cpu_id field's content.
>>> 
>>> I don't think that's necessary.  We can mention it in the glibc
>>> distribution notes on the wiki.
>>> 
>>>> The usual approach to kernel bug fixing is typically to push the fix,
>>>> mark it for stable kernels, and expect everyone to pick up the
>>>> fixes. I wonder how comfortable glibc would be to replace its
>>>> sched_getcpu implementation with a broken-until-fixed kernel rseq
>>>> implementation without any mechanism in place to know whether it can
>>>> trust the value of the cpu_id field. I am extremely reluctant to do
>>>> so.
>>> 
>>> We have already had similar regressions in sched_getcpu, and we didn't
>>> put anything into glibc to deal with those.
>>
>> Was that acceptable because having a wrong cpu number would never trigger
>> corruption, only slowdowns ?
> 
> First of all, it's a kernel bug.  It's rare that we put workarounds for
> kernel bugs into glibc.
> 
> And yes, in pretty much all cases it's just a performance issue for
> sched_getcpu.  When you know the CPU ID of a thread due to pinning to a
> single CPU, why would you call sched_getcpu?  (That's the case where you
> could get corruption in theory.)
> 
>> In the case of rseq, having the wrong cpu_id value is a real issue
>> which will lead to corruption and crashes. So I maintain my reluctance
>> to introduce the fix without any way for userspace to know whether the
>> cpu_id field value is reliable.
> 
> Yes, for rseq itself, the scenario is somewhat different.  Still, it's
> just another kernel bug.  There will be others. 8-/
> 
> From a schedule point of view, it looks tough to get the magic flag into
> the mainline kernel in time for the upcoming glibc 2.32 release.  If you
> insist on registering rseq only if the bug is not present, we'll
> probably have to back out some or all of the rseq changes.

I've just submitted the fix and a the new rseq flag as RFC to lkml:

https://lore.kernel.org/lkml/20200706204913.20347-1-mathieu.desnoyers@efficios.com/

Let's see how quickly we can come to an agreement on this on the kernel
side.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com