On Wed, Feb 08, 2023 at 02:00:14PM +0000, Parav Pandit wrote: > > From: Michael S. Tsirkin <m...@redhat.com> > > Sent: Wednesday, February 8, 2023 8:52 AM > > > > On Wed, Feb 08, 2023 at 01:38:36PM +0000, Parav Pandit wrote: > > > > > > > From: Michael S. Tsirkin <m...@redhat.com> > > > > Sent: Wednesday, February 8, 2023 8:32 AM > > > > > > > > On Wed, Feb 08, 2023 at 05:18:32AM +0000, Parav Pandit wrote: > > > > > > From: Heng Qi <hen...@linux.alibaba.com> > > > > > > Sent: Tuesday, February 7, 2023 10:25 PM > > > > > > > > > > [..] > > > > > > >> > > > > > > >> Do you think we need both hash_types and hash_tunnel_types? > > > > > > > In struct virtio_net_config we need two fields. > > > > > > > a. supported_hash_types (already exists) b. > > > > > > > supported_hash_tunnel_type > > > > > > > -> bitmap indicating for which outer headers, inner hash > > > > > > > -> calculation is > > > > > > supported. > > > > > > > > > > > > Thanks for the suggestion, we seem to have reached an agreement. > > > > > > > > > > > > > > > > > > > > In struct virtio_net_hdr we need two fields. > > > > > > > a. hash_report (already exists) b. hash_tunnel_type 8 bits -> > > > > > > > absolute value indicating which outer header > > > > > > exists when inner header hash calculated. > > > > > > > You already have it in your patch named as hash_report_tunnel. > > > > > > > May be better to name as hash_report_tunnel_type to make it > > > > > > > clearer that its > > > > > > type. > > > > > > > > > > > > Sure. > > > > > > > > > > > > Thanks for your reply. > > > > > > > > > > I had one last question. Why do we need to inform the > > > > hash_report_tunnel_type of the outer header in the virtio_net_hdr? > > > > > Is this for debug? Or is there a use case that will process this > > > > > value? > > > > > > > > Well we have hash_report which is kind of similar (and also kind of > > > > pointless but I think it's there because WHQL wants it). > > > Hash_report is useful. It tells hash_value is in which namespace > > > (ipv4-tcp/ipv4 > > udp etc). > > > OS can use this value to find tcp connection in a given namespace. > > > > > > > Maybe we can steal some bits > > > > from there instead of a new field? > > > > > > > I do not have problem adding extra bits. I just don't find that just > > > telling that > > its vxlan or nvgre to the OS is useful. > > > If OS needs to know about outer header details, it needs to know the VNI > > information than just telling vxlan. > > > > This does make sense. > > > > > > > > > > > > I have a follow up question though: are we only hashing the inner > > > > header or both inner and outer header? Somewhat confused on this. > > > > > > > I understood as inner header. But worth to describe it. May be there. > > > Need to > > read v8 patch. > > > > Hmm. I just realized that there's a security problem with hashing just the > > inner > > header: it allow users inside the tunnel control queueing outside. > > By observing packet loss some information leaks between tunnels. > > > I likely didn't understand. Can you please explain? > > Queuing is always done on the inner header with/without encapsulation. > Hash is always reported for inner header. > It is only adding the ability to hash even when outer header exists.
If hashing just on outer header (currently the only option) then a given tunnel all lands in a given queue. Just keep that queue separate and users of this tunnel can not learn whether other queues are overflowing, and can not overflow other queues. If you hash inner header then user can flood device with packets of a given connection and the same connection in a different tunnel hashes to the same queue. Now one tunnel can - cause DoS for another tunnel - cause packet loss or latency triggering possible security bugs within guest - detect that another tunnel is using the connection by detecting its own packet loss or increased latency > If queuing to be decided based on outer header (hash), then that is different. > Hashing both inner and outer in a flat q structure unlikely works, right? > Because both hashes can result in different q selection. That's the point. Is there any precedent in OSes for configuring things like this that we can look at? > > > > Ideas for solving this they all involve hashing both inner and outer > > header: > > 1- report two sets of hashes. overkill? > > 2- hash both headers together > > 2- add salt. can come from driver or device itself > > > > More ideas? > > > > -- > > MST --------------------------------------------------------------------- To unsubscribe, e-mail: virtio-dev-unsubscr...@lists.oasis-open.org For additional commands, e-mail: virtio-dev-h...@lists.oasis-open.org
