Re: [syzbot] [mm?] KMSAN: kernel-infoleak in bpf_probe_write_user

[Alexander Potapenko]

On Tue, Apr 16, 2024 at 5:16 PM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Tue, Apr 16, 2024 at 1:52 AM Alexander Potapenko <glider@xxxxxxxxxx> wrote:
> >
> > On Mon, Apr 15, 2024 at 11:06 PM Alexei Starovoitov
> > <alexei.starovoitov@xxxxxxxxx> wrote:
> > >
> > > Hi,
> > >
> > > syzbot folks, please disable such "bug" reporting.
> > > The whole point of bpf is to pass such info to userspace.
> > > probe_write_user, various ring buffers, bpf_*_printk-s, bpf maps
> > > all serve this purpose of "infoleak".
> > >
> >
> > Hi Alexei,
> >
> > From KMSAN's perspective it is fine to pass information to the
> > userspace, unless it is marked as uninitialized.
> > It could be that we are missing some initialization in kernel/bpf/core.c though.
> > Do you know which part of the code is supposed to initialize the stack
> > in PROG_NAME?
>
> cap_bpf + cap_perfmon bpf program are allowed to read uninitialized stack.

Out of curiosity, is this feature supposed to be used in production kernels?

> And recently we added
> commit e8742081db7d ("bpf: Mark bpf prog stack with
> kmsan_unposion_memory in interpreter mode")
> to shut up syzbot.

I checked that the report in question is not reproducible with this
patch anymore. Let's just wait until it reaches the mainline.