On 08/11/2019 15:34, Daniel Borkmann wrote: > On 11/8/19 3:08 PM, Mickaël Salaün wrote: >> On 06/11/2019 22:45, KP Singh wrote: >>> On 06-Nov 17:55, Mickaël Salaün wrote: >>>> On 06/11/2019 11:06, KP Singh wrote: >>>>> On 05-Nov 11:34, Alexei Starovoitov wrote: >>>>>> On Tue, Nov 05, 2019 at 07:01:41PM +0100, Mickaël Salaün wrote: >>>>>>> On 05/11/2019 18:18, Alexei Starovoitov wrote: > [...] >>> * Use a single BPF program type; this is necessary for a key requirement >>> of KRSI, i.e. runtime instrumentation. The upcoming prototype should >>> illustrate how this works for KRSI - note that it’s possible to vary >>> the context types exposed by different hooks. >> >> Why a single BPF program type? Do you mean *attach* types? Landlock only >> use one program type, but will use multiple attach types. >> >> Why do you think it is necessary for KRSI or for runtime instrumentation? >> >> If it is justified, it could be a dedicated program attach type (e.g. >> BPF_LANDLOCK_INTROSPECTION). >> >> What is the advantage to have the possibility to vary the context types >> over dedicated *typed* contexts? I don't see any advantages, but at >> least one main drawback: to require runtime checks (when helpers use >> this generic context) instead of load time checks (thanks to static type >> checking of the context). > > Lets take security_sock_rcv_skb(struct sock *sk, struct sk_buff *skb) > as one specific example here: the running kernel has its own internal > btf_vmlinux and therefore a complete description of itself. From verifier > side we can retrieve & introspect the security_sock_rcv_skb signatue OK, this is indeed the signature defined by the kernel API. What happen if this API change (e.g. if struct sock is replaced with a struct sock_meta)? > and > thus know that the given BPF attachment point has struct sock and struct > sk_buff as input arguments How does the verifier know a given BPF attachment point for a program without relying on its type or attach type? How and where is registered this mapping? To say it another way, if there is no way to differentiate two program targeting different hook, I don't understand how the verifier could check if a given program can legitimately call a helper which could read the tracer and tracee fields (legitimate for a ptrace hook), whereas this program may be attached to a sock_rcv_skb hook (and there is no way to know that). > which can then be accessed generically by the > prog in order to allow sk_filter_trim_cap() to pass or to drop the skb. > The same generic approach can be done for many of the other lsm hooks, so > single program type would be enough there and context is derived > automatically, > no dedicated extra context per attach type would be needed and no runtime > checks as you mentioned above since its still all asserted at verification > time. I mentioned runtime check because I though a helper should handle the case when it doesn't make sense for a program attached to a specific point/hook (e.g. ptrace) to use an input argument (e.g. sk) defined for another point/hook (e.g. sock_rcv_skb). > > Thanks, > Daniel > Thanks for this explanation Daniel.
