Re: mm allocation failure and hang when running xfstests generic/269 on xfs

[Brian Foster <bfoster@redhat.com>]

On Thu, Mar 02, 2017 at 02:27:55PM +0100, Michal Hocko wrote:
> On Thu 02-03-17 08:00:09, Brian Foster wrote:
> > On Thu, Mar 02, 2017 at 01:49:09PM +0100, Michal Hocko wrote:
> > > On Thu 02-03-17 07:24:27, Brian Foster wrote:
> > > > On Thu, Mar 02, 2017 at 11:35:20AM +0100, Michal Hocko wrote:
> > > > > On Thu 02-03-17 19:04:48, Tetsuo Handa wrote:
> > > > > [...]
> > > > > > So, commit 5d17a73a2ebeb8d1("vmalloc: back off when the current task is
> > > > > > killed") implemented __GFP_KILLABLE flag and automatically applied that
> > > > > > flag. As a result, those who are not ready to fail upon SIGKILL are
> > > > > > confused. ;-)
> > > > > 
> > > > > You are right! The function is documented it might fail but the code
> > > > > doesn't really allow that. This seems like a bug to me. What do you
> > > > > think about the following?
> > > > > ---
> > > > > From d02cb0285d8ce3344fd64dc7e2912e9a04bef80d Mon Sep 17 00:00:00 2001
> > > > > From: Michal Hocko <mhocko@suse.com>
> > > > > Date: Thu, 2 Mar 2017 11:31:11 +0100
> > > > > Subject: [PATCH] xfs: allow kmem_zalloc_greedy to fail
> > > > > 
> > > > > Even though kmem_zalloc_greedy is documented it might fail the current
> > > > > code doesn't really implement this properly and loops on the smallest
> > > > > allowed size for ever. This is a problem because vzalloc might fail
> > > > > permanently. Since 5d17a73a2ebe ("vmalloc: back off when the current
> > > > > task is killed") such a failure is much more probable than it used to
> > > > > be. Fix this by bailing out if the minimum size request failed.
> > > > > 
> > > > > This has been noticed by a hung generic/269 xfstest by Xiong Zhou.
> > > > > 
> > > > > Reported-by: Xiong Zhou <xzhou@redhat.com>
> > > > > Analyzed-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
> > > > > Signed-off-by: Michal Hocko <mhocko@suse.com>
> > > > > ---
> > > > >  fs/xfs/kmem.c | 2 ++
> > > > >  1 file changed, 2 insertions(+)
> > > > > 
> > > > > diff --git a/fs/xfs/kmem.c b/fs/xfs/kmem.c
> > > > > index 339c696bbc01..ee95f5c6db45 100644
> > > > > --- a/fs/xfs/kmem.c
> > > > > +++ b/fs/xfs/kmem.c
> > > > > @@ -34,6 +34,8 @@ kmem_zalloc_greedy(size_t *size, size_t minsize, size_t maxsize)
> > > > >  	size_t		kmsize = maxsize;
> > > > >  
> > > > >  	while (!(ptr = vzalloc(kmsize))) {
> > > > > +		if (kmsize == minsize)
> > > > > +			break;
> > > > >  		if ((kmsize >>= 1) <= minsize)
> > > > >  			kmsize = minsize;
> > > > >  	}
> > > > 
> > > > More consistent with the rest of the kmem code might be to accept a
> > > > flags argument and do something like this based on KM_MAYFAIL.
> > > 
> > > Well, vmalloc doesn't really support GFP_NOFAIL semantic right now for
> > > the same reason it doesn't support GFP_NOFS. So I am not sure this is a
> > > good idea.
> > > 
> > 
> > Not sure I follow..? I'm just suggesting to control the loop behavior
> > based on the KM_ flag, not to do or change anything wrt to GFP_ flags.
> 
> As Tetsuo already pointed out, vmalloc cannot really support never-fail
> semantic with the current implementation so the semantic would have
> to be implemented in kmem_zalloc_greedy and the only way to do that
> would be to loop there and this is rather nasty as you can see from the
> reported issue because the vmalloc failure might be permanent so there
> won't be any way to make a forward progress. Breaking out of the loop
> on fatal_signal_pending pending would break the non-failing sementic.
> 

Sure..

> Besides that, there doesn't really seem to be any demand for this
> semantic in the first place so why to make this more complicated than
> necessary?
> 

That may very well be the case. I'm not necessarily against this...

> I see your argument about being in sync with other kmem helpers but
> those are bit different because regular page/slab allocators allow never
> fail semantic (even though this is mostly ignored by those helpers which
> implement their own retries but that is a different topic).
> 

... but what I'm trying to understand here is whether this failure
scenario is specific to vmalloc() or whether the other kmem_*()
functions are susceptible to the same problem. For example, suppose we
replaced this kmem_zalloc_greedy() call with a kmem_zalloc(PAGE_SIZE,
KM_SLEEP) call. Could we hit the same problem if the process is killed?

Brian

> -- 
> Michal Hocko
> SUSE Labs
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org.  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>