Re: Preemptible idr_alloc() in QRTR code

From: Matthew Wilcox
Date: Wed Jan 27 2021 - 02:15:00 EST


On Tue, Jan 26, 2021 at 06:36:02PM +0000, Mark Rutland wrote:
> On Tue, Jan 26, 2021 at 11:00:05AM -0600, Bjorn Andersson wrote:
> > On Tue 26 Jan 10:21 CST 2021, Mark Rutland wrote:
> >
> > > On Tue, Jan 26, 2021 at 02:58:33PM +0000, Matthew Wilcox wrote:
> > > > On Tue, Jan 26, 2021 at 10:47:34AM +0000, Mark Rutland wrote:
> > > > > Hi,
> > > > >
> > > > > When fuzzing arm64 with Syzkaller, I'm seeing some splats where
> > > > > this_cpu_ptr() is used in the bowels of idr_alloc(), by way of
> > > > > radix_tree_node_alloc(), in a preemptible context:
> > > >
> > > > I sent a patch to fix this last June. The maintainer seems to be
> > > > under the impression that I care an awful lot more about their
> > > > code than I do.
> > > >
> > > > https://lore.kernel.org/netdev/20200605120037.17427-1-willy@xxxxxxxxxxxxx/
> > >
> > > Ah; I hadn't spotted the (glaringly obvious) GFP_ATOMIC abuse, thanks
> > > for the pointer, and sorry for the noise.
> > >
> >
> > I'm afraid this isn't as obvious to me as it is to you. Are you saying
> > that one must not use GFP_ATOMIC in non-atomic contexts?
> >
> > That said, glancing at the code I'm puzzled to why it would use
> > GFP_ATOMIC.
>
> I'm also not entirely sure about the legitimacy of GFP_ATOMIC outside of
> atomic contexts -- I couldn't spot any documentation saying that wasn't
> legitimate, but Matthew's commit message implies so, and it sticks out
> as odd.

It's actually an assumption in the radix tree code. If you say you
can't be preempted by saying GFP_ATOMIC, it takes you at your word and
does some things which cannot be done in preemptable context.