Re: [PATCH rdma-next] RDMA/core: Silence oversized kvmalloc() warning

[Leon Romanovsky <leon@kernel.org>]

On Tue, Apr 01, 2025 at 01:35:04PM -0300, Jason Gunthorpe wrote:
> On Mon, Mar 31, 2025 at 10:04:05PM +0300, Leon Romanovsky wrote:
> > 
> > 
> > On Mon, Mar 31, 2025, at 20:45, Jason Gunthorpe wrote:
> > > On Wed, Mar 26, 2025 at 12:58:54PM +0200, Leon Romanovsky wrote:
> > >> On Wed, Mar 19, 2025 at 02:23:49PM -0300, Jason Gunthorpe wrote:
> > >> > On Wed, Mar 19, 2025 at 02:42:21PM +0200, Leon Romanovsky wrote:
> > >> > > From: Shay Drory <shayd@nvidia.com>
> > >> > > 
> > >> > > syzkaller triggered an oversized kvmalloc() warning.
> > >> > > Silence it by adding __GFP_NOWARN.
> > >> > 
> > >> > I don't think GFP_NOWARN is the right thing..
> > >> > 
> > >> > We've hit this before and I think we ended up adding a size limit
> > >> > check prior to the kvmalloc to prevent the overflow triggered warning.
> > >> 
> > >> The size check was needed before this commit was merged:
> > >> 0708a0afe291 ("mm: Consider __GFP_NOWARN flag for oversized kvmalloc() calls")
> > >> 
> > >> From that point, the correct solution is simply provide __GFP_NOWARN flag.
> > >
> > > I'm not sure, NOWARN removes all warnings, even normal OOM warnings
> > > from regually sized allocations which we don't want to remove.
> > 
> > I disagree, this allocation is called from user space. We can safely
> > skip OOM messages and error here will be enough.
> 
> That is not the standard, we OOM splat on all userspace allocations
> too.
> 
> GFP_NOWARN is supposed to be used in cases where the OOM has a
> recovery and nothing will fail.

NULL returned back to the user is best way to recover.

> 
> I think the right thing here is to limit the size, though I'm not
> really sure what the limit should be.

And I still think that NOWARN is the best solution here.

Thanks

> 
> Jason