Hi Alexei,
On Tue, Aug 22, 2023 at 11:57 AM Alexei Starovoitov
<[email protected]> wrote:
>
> On Tue, Aug 22, 2023 at 10:48 AM Chris Li <[email protected]> wrote:
> >
> > Hi Mel,
> >
> > Adding Alexei to the discussion.
> >
> > On Mon, Aug 21, 2023 at 3:32 AM Mel Gorman <[email protected]> wrote:
> > >
> > > On Thu, Aug 17, 2023 at 11:05:22PM -0700, Chris Li wrote:
> > > > In this patch series I want to safeguard
> > > > the free_pcppage_bulk against change in the
> > > > pcp->count outside of this function. e.g.
> > > > by BPF program inject on the function tracepoint.
> > > >
> > > > I break up the patches into two seperate patches
> > > > for the safeguard and clean up.
> > > >
> > > > Hopefully that is easier to review.
> > > >
> > > > Signed-off-by: Chris Li <[email protected]>
> > >
> > > This sounds like a maintenance nightmare if internal state can be arbitrary
> > > modified by a BPF program and still expected to work properly in all cases.
> > > Every review would have to take into account "what if a BPF script modifies
> > > state behind our back?"
>
> Where did this concern come from?
> Since when BPF can modify arbitrary state?
>
> But I wasn't cc-ed on the original patch, so not sure what it attempts to do.
> Maybe concern is valid.
Sorry I did not CC you on the original patch submission. I should.
Here is the link for the 1/2 patch, which has the step to reproduce.
https://lore.kernel.org/linux-mm/[email protected]/
It is using an older version of the BPF program. That spinlock
allocation was fixed
in commit c66a36af7ba3a628.
Chris
>
> > Thanks for the feedback.
> >
> > I agree that it is hard to support if we allow BPF to change any internal
> > stage as a rule. That is why it is a RFC. Would you consider it case
> > by case basis?
> > The kernel panic is bad, the first patch is actually very small. I can
> > also change it
> > to generate warnings if we detect the inconsistent state.
>
> panic and warns because of bpf prog?!
> bpf infra takes all the precaution to make sure that bpf progs
> can never cause such damage.
>
> >
> > How about the second (clean up) patch or Keming's clean up version? I can modify
> > it to take out the pcp->count if the verdict is just not supporting
> > BPF changing internal
> > state at all. I do wish to get rid of the pindex_min and pindex_max.
> >
> > Thanks
> >
> > Chris
>