Quoting Sebastian Andrzej Siewior (2019-05-29 08:25:40)
> From: Hugh Dickins <hughd@xxxxxxxxxx>
>
> Since commit
>
> d9c9ce34ed5c8 ("x86/fpu: Fault-in user stack if copy_fpstate_to_sigframe() fails")
>
> we use get_user_pages_unlocked() to pre-faulting user's memory if a
> write generates a pagefault while the handler is disabled.
> This works in general and uncovered a bug as reported by Mike Rapoport.
> It has been pointed out that this function may be fragile and a
> simple pre-fault as in fault_in_pages_writeable() would be a better
> solution. Better as in taste and simplicity: That write (as performed by
> the alternative function) performs exactly the same faulting of memory
> that we had before. This was suggested by Hugh Dickins and Andrew
> Morton.
>
> Use fault_in_pages_writeable() for pre-faulting of user's stack.
>
> Fixes: d9c9ce34ed5c8 ("x86/fpu: Fault-in user stack if copy_fpstate_to_sigframe() fails")
> Suggested-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
> [bigeasy: patch description]
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
I am able to reliably hit the bug here by putting the system under
mempressure, and afterwards processes would die as the exit. This patch
also greatly reduces cycletest latencies while under that mempressure,
~320ms -> ~16ms (on a bxt while also spinning on i915.ko).
Tested-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
-Chris
