Andi Kleen wrote:
quoted text
>> I could just force all of the seqbegins to hit the slowpath by hacking=

quoted text
>> the code and see what happens (aside from slowing down, of course ;)
>>    =20
>
> Only if you don't believe it will really crash? I think it's pretty
> clear even without trying it.
>  =20

Well, I guess it was just to prove to myself that I broke something
because I dont understand how the vsyscall interface works.  But given
your expertise here, I have no problem with just taking your word for it.=



quoted text
>  =20
>> Question: Which seqlock_t does userspace use?  I assume it uses
>> seqlock_t and not raw_seqlock_t.=20
>>    =20
>
>  =20
>> But the only reason that I ask is that
>> I converted raw_seqlock_t to use the new style as well to be consisten=
t,
quoted text
>>    =20
>
> There's no raw_seqlock_t anywhere in mainline?
>  =20
Yeah, understood.  There is both in -rt and I was just saying that we
technically only need the seqlock_t fix in -rt.  So if raw_seqlock_t
could be left pristine and solve this problem, that is an acceptable
compromise to me.


quoted text
> Anyways the variable is declared (in mainline) in asm-x86/vgtod.h=20
>
>  =20
>> even though it is not strictly necessary for the same reasons.  So if
>> perchance userspace uses the raw variant, I could solve this issue by
>> only re-working the seqlock_t variant.  Kind of a long shot, but figur=
ed
quoted text
>> I would mention it :)
>>    =20
>
> I guess you could define a new seqlock_t which is explicitely user spac=
e
quoted text
> safe. That might avoid such issues in the future. But then
> that would likely require some code duplication and be ugly.
>
> On the other hand whatever problem you fixing in the kernel
> (to be honest it's still unclear to me what the problem is)
> needs to be likely fixed for the userland lock too.
>  =20

Yeah, it would possibly be a problem in both cases.

The problem I am addressing only exists in -rt since it has seqlock_t
and raw_seqlock_t (with the former using preemptible spinlock_t's).=20
Since the underlying seqlock_t->spinlock_t is preemptible, you can see
that one thread that does:

{
    write_seqlock();
    /* asl */
    write_sequnlock();
}

while other high-prio threads do

do { read_seqbegin(); /* asl */; } while (read_seqretry());

The readers could preempt the writer mid critical section and enter a
live-locked loop.

raw_seqlock_t (which is equivalent to a mainline seqlock_t) do not have
this problem because the spinlock acquisition inside the write_seqlock
disables preemption.

HTH

-Greg