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

[Michal Hocko]

On Mon 06-11-17 15:19:46, Vlastimil Babka wrote:
> On 11/06/2017 02:40 PM, 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
> > 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.
> 
> If we fail to allocate the page, we could still print the addresses,
> just miss the filename? But that's an improvement, not a fix.

Agreed. Or we could have some preallocated buffer if this is more
widespread pattern

> > Fixes: d99e1bd175f4 ("x86/entry/traps: Refactor preemption and interrupt flag handling")
> > Signed-off-by: Michal Hocko <mhocko@xxxxxxxx>
> 
> Acked-by: Vlastimil Babka <vbabka@xxxxxxx>

Thanks!

> 
> > ---
> >  mm/memory.c | 8 +++-----
> >  1 file changed, 3 insertions(+), 5 deletions(-)
> > 
> > diff --git a/mm/memory.c b/mm/memory.c
> > index a728bed16c20..1e308ac8ca0a 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -4457,17 +4457,15 @@ void print_vma_addr(char *prefix, unsigned long ip)
> >  	struct vm_area_struct *vma;
> >  
> >  	/*
> > -	 * Do not print if we are in atomic
> > -	 * contexts (in exception stacks, etc.):
> > +	 * we might be running from an atomic context so we cannot sleep
> >  	 */
> > -	if (preempt_count())
> > +	if (!down_read_trylock(&mm->mmap_sem))
> >  		return;
> >  
> > -	down_read(&mm->mmap_sem);
> >  	vma = find_vma(mm, ip);
> >  	if (vma && vma->vm_file) {
> >  		struct file *f = vma->vm_file;
> > -		char *buf = (char *)__get_free_page(GFP_KERNEL);
> > +		char *buf = (char *)__get_free_page(GFP_NOWAIT);
> >  		if (buf) {
> >  			char *p;
> >  
> > 

-- 
Michal Hocko
SUSE Labs