Re: [PATCH] seqlock: serialize against writers

[Gregory Haskins]

Andi Kleen wrote:
> On Fri, Aug 29, 2008 at 12:29:42PM -0400, Gregory Haskins wrote:
>   
>> Andi Kleen wrote:
>>     
>>>> Im running it on a x86_64 box as we speak.  How can I tell if there is a
>>>> certain mode that is permitting this?
>>>>     
>>>>         
>>> If the boot up says you're running with PMtimer then it uses the fallback
>>> (usually happens on pre Fam10h AMD boxes). A typical Intel box
>>> would use the faster ring 3 only TSC path and then explode with your
>>> change I bet. 
>>>   
>>>       
>> Thinking about this some more, perhaps the issue is I am not hitting the
>> contended path in vsyscall?
>>     
>
> Yes it will be only contended when gettimeofday() races with the timer 
> interrupt.  You could try to run gettimeofday() in a loop and see how
> long it holds up.
>
> But anyways from the theory you should crash when it happens. 
> Writes to kernel data are not allowed in vsyscalls and your read_lock clearly 
> does a write.
>   

Oh I don't deny that it does.  The compiler neatly reminded me that it
could no longer be "const".  I just was ignorant of the userspace
requirement ;)

But we *do* have a serious problem here.  A non-preemptible seqlock_t
will do bad things if it preempts the writer, so we need some kind of
solution here, one way or the other.  So suggestions welcome :)  I
realize this is only an issue currently in PREEMPT_RT so perhaps most
will not care... but I do need to solve this at least for this branch.

I currently do not see a way to solve this problem that doesn't involve
some heavier involvement in the read path (read: const seqlock_t need
not apply).  Given that, we either need to address the const requirement
for userspace, alter userspace's usage, or find another way to address
the deadlock.  Any ideas?

-Greg

Attachment: signature.asc
Description: OpenPGP digital signature