Re: [PATCH] mm: do not rely on preempt_count in print_vma_addr

[Michal Hocko]

On Tue 07-11-17 00:16:58, Yang Shi wrote:
> 
> 
> On 11/6/17 5:40 AM, Michal Hocko wrote:
> > On Mon 06-11-17 13:12:22, Michal Hocko wrote:
> > > On Mon 06-11-17 13:00:25, Peter Zijlstra wrote:
> > > > On Mon, Nov 06, 2017 at 11:43:54AM +0100, Michal Hocko wrote:
> > > > > > Yes the comment is very much accurate.
> > > > > 
> > > > > Which suggests that print_vma_addr might be problematic, right?
> > > > > Shouldn't we do trylock on mmap_sem instead?
> > > > 
> > > > Yes that's complete rubbish. trylock will get spurious failures to print
> > > > when the lock is contended.
> > > 
> > > Yes, but I guess that it is acceptable to to not print the state under
> > > that condition.
> > 
> > So what do you think about this? I think this is more robust than
> > playing tricks with the explicit preempt count checks and less tedious
> > than checking to make it conditional on the context. This is on top of
> > Linus tree and if accepted it should replace the patch discussed here.
> > ---
> >  From 0de6d57cbc54ee2686d1f1e4ffcc4ed490ded8aa Mon Sep 17 00:00:00 2001
> > From: Michal Hocko <mhocko@xxxxxxxx>
> > Date: Mon, 6 Nov 2017 14:31:20 +0100
> > Subject: [PATCH] mm: do not rely on preempt_count in print_vma_addr
> > 
> > The preempt count check on print_vma_addr has been added by e8bff74afbdb
> > ("x86: fix "BUG: sleeping function called from invalid context" in
> > print_vma_addr()") and it relied on the elevated preempt count from
> > preempt_conditional_sti because preempt_count check doesn't work on
> > non preemptive kernels by default. The code has evolved though and
> > d99e1bd175f4 ("x86/entry/traps: Refactor preemption and interrupt flag
> > handling") has replaced preempt_conditional_sti by an explicit
> > preempt_disable which is noop on !PREEMPT so the check in print_vma_addr
> > is broken.
> > 
> > Fix the issue by using trylock on mmap_sem rather than chacking the
> 
> s/chacking/checking

ups, fixed

> > preempt count. The allocation we are relying on has to be GFP_NOWAIT
> > as well. There is a chance that we won't dump the vma state if the lock
> > is contended or the memory short but this is acceptable outcome and much
> > less fragile than the not working preemption check or tricks around it.
> > 
> > Fixes: d99e1bd175f4 ("x86/entry/traps: Refactor preemption and interrupt flag handling")
> > Signed-off-by: Michal Hocko <mhocko@xxxxxxxx>
> 
> Acked-by: Yang Shi <yang.s@xxxxxxxxxxxxxxx>

Thanks!
-- 
Michal Hocko
SUSE Labs