On Mon, Nov 30, 2015 at 02:02:00PM +0100, Michal Hocko wrote:
[Let's CC Will - see the question at the end of the email please]
> > There is no reference to OOM possibility in the email that
I can see. Can
> > you give examples of the OOM messages that shows the problem sites? It was
> > suspected that there may be some callers that were accidentally depending
> > on access to emergency reserves. If so, either they need to be fixed (if
> > the case is extremely rare) or a small reserve will have to be created
> > for callers that are not high priority but still cannot reclaim.
virtqueue_add_sgs(vq, sgs, num_out, num_in, vbr, GFP_ATOMIC)
gfp &= ~(__GFP_HIGHMEM | __GFP_HIGH)
So this is true __GFP_ATOMIC, we just drop __GFP_HIGH so it doesn't get
access to more reserves. It still does ALLOC_HARDER. So I think the real
issue is somewhere else when something should have triggered kswapd and
it doesn't do that anymore. I have tried to find that offender the last
time but didn't manage to find any.
Btw. I completely miss why b92b1b89a33c ("virtio: force vring
descriptors to be allocated from lowmem") had to clear __GFP_HIGH. Will
do you remember why you have dropped that flag as well?
Right, that looks unnecessary, but it could be that we were masking a
bug somewhere else.
Also I do not seem to find any user of alloc_indirect which would do
__GFP_HIGHMEM. All of them are either GFP_KERNEL or GFP_ATOMIC. So
either I am missing something or this is not really needed. Maybe the
situation was different back in 2012.
I tried to revisit the thread leading to that patch, but it doesn't make
a whole lot of sense:
I certainly remember debugging the failure (i.e. it wasn't theoretical),
and we were ending up with highmem addresses being passed in the virtio
ring (due to the zero-copy stuff in 9p) and also for the descriptors
themselves. The discussion at the time makes it sound like GFP_ATOMIC
was giving us those...