On Thu, Dec 3, 2015 at 12:16 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
On Wed 02-12-15 22:43:31, Chris Snook wrote:
> On Tue, Dec 1, 2015 at 12:35 PM David Miller <davem@xxxxxxxxxxxxx> wrote:
>
> > From: Michal Hocko <mhocko@xxxxxxxxxx>
> > Date: Mon, 30 Nov 2015 14:21:29 +0100
> >
> > > On Sat 28-11-15 15:51:13, Pavel Machek wrote:
> > >>
> > >> atl1c driver is doing order-4 allocation with GFP_ATOMIC
> > >> priority. That often breaks networking after resume. Switch to
> > >> GFP_KERNEL. Still not ideal, but should be significantly better.
> > >
> > > It is not clear why GFP_KERNEL can replace GFP_ATOMIC safely neither
> > > from the changelog nor from the patch context.
> >
> > Earlier in the function we do a GFP_KERNEL kmalloc so:
> >
> > ¯\_(ツ)_/¯
> >
> > It should be fine.
> >
>
> AFAICT, the people who benefit from GFP_ATOMIC are the people running all
> their storage over NFS/iSCSI who are suspending their machines while
> they're so busy they don't have any clean order 4 pagecache to drop, and
> want the machine to panic rather than hang.
Why would GFP_KERNEL order-4 allocation hang? It will fail if there are
not >=4 order pages available even after reclaim and/or compaction.
GFP_ATOMIC allocations should be used only when an access to memory
reserves is really required. If the allocation just doesn't want to
invoke direct reclaim then GFP_NOWAIT is a more suitable alternative.
The *machine* may hang if you can't bring back the interface that's required to access the storage. It's a ridiculous use case, as Pavel noted. I only pointed it out to note that there exists a rationale for GFP_ATOMIC. It just isn't nearly as good as the rationale for using GFP_KERNEL.
-- Chris