Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx> wrote: > On Fri, May 27, 2022 at 03:15:58AM IST, Florian Westphal wrote: > > Lorenzo Bianconi <lorenzo@xxxxxxxxxx> wrote: > > > From: Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx> > > > > > > Since we want to allow user to set some fields in nf_conn after it is > > > allocated but before it is inserted, we can permit BPF_WRITE for normal > > > nf_conn, and then mark return value as read only on insert, preventing > > > further BPF_WRITE. This way, nf_conn can be written to using normal > > > BPF instructions after allocation, but not after insertion. > > > > > > Note that we special nf_conn a bit here, inside the btf_struct_access > > > callback for XDP and TC programs. Since this is the only struct for > > > these programs requiring such adjustments, making this mechanism > > > more generic has been left as an exercise for a future patch adding > > > custom callbacks for more structs. > > > > Are you sure this is safe? > > As far as I can see this allows nf_conn->status = ~0ul. > > I'm fairly sure this isn't a good idea, see nf_ct_delete() for example. > > This only allows writing to an allocated but not yet inserted nf_conn. The idea > was that insert checks whether ct->status only has permitted bits set before > making the entry visible, and then we make nf_conn pointer read only, however > the runtime check seems to be missing right now in patch 12; something to fix in > v5. With that sorted, would it be fine? Its fragile, e.g. what if I set TEMPLATE bit? If refcount goes down to 0, object is released via kfree() instead of kmem_cache_free. What if I clear SNAT_DONE bit? Would it leave the (freed) entry on the bysource hash list (see nf_nat_core.c)? Or is there some magic that prevents this from happening? I have no idea how processing pipeline looks like...
