Magnus Karlsson <magnus.karlsson@xxxxxxxxx> writes: > On Mon, 18 Mar 2024 at 19:41, Christian Deacon > <christian.m.deacon@xxxxxxxxx> wrote: >> >> Resending the following email to the XDP Newbies mailing list since it >> was rejected due to HTML contents (I've switched email clients and >> forgot to disable HTML, I apologize). >> >> Hey everyone, >> >> I was wondering if there was an update to this. I'm currently running >> into the same issue with a similar setup. >> >> When running the XDP program on a bonding device via native mode, >> packets redirected to the AF_XDP sockets with `bpf_redirect_map()` >> inside the XDP program do not make it to the AF_XDP sockets. Switching >> between zero copy and copy mode does not make a difference along with >> setting the need wakeup flag. >> >> I've tried the latest mainline kernel `6.8.1-060801`, but that did not >> make a difference. If the XDP program is attached with SKB mode, >> packets do show up on the AF_XDP sockets as mentioned in this thread >> already. >> >> While I haven't confirmed it on my side, I'm assuming the >> `xsk_rcv_check()` function is the issue here. I'm unsure if skipping >> this check for the time being would work for my needs, but I'm hoping >> a better solution will be implemented to the mainline kernel. >> >> I also saw there was another similar issue on this mailing list with >> the title "Switching packets between queues in XDP program". However, >> judging from the last reply in that thread, the fix implemented >> wouldn't help with the bonding driver. >> >> Any help is appreciated and thank you for your time! > > You are correct in that the fix above does not address the bonding > case and that the problem is indeed that XDP reports the device as the > real NIC and that the AF_XDP socket is bound to the bonding device. > Therefore xdp->dev != xsk->dev (in principle, not the actual code) and > all packets will be discarded. I got as far as sketching on a solution > but I do not have the bandwidth at the moment to implement it. > Unfortunately it is not a one-liner or even just one hundred lines of > code. Let me know what you think, or if someone can come up with an > easier solution. > > *** Suggestion on how to implement AF_XDP for the bond device > > Two steps: XDP_DRV mode then zero-copy mode > > * XDP_DRV: > > For XDP_DRV mode, the problem to overcome is this piece of code > in xsk_rcv_check(): > > struct net_device *dev = xdp->rxq->dev; > u32 qid = xdp->rxq->queue_index; > > if (!dev->_rx[qid].pool || xs->umem != dev->_rx[qid].pool->umem) > return -EINVAL; > > xs is the socket that was bound to the bonding device e.g., bond0. So > xs->dev points to bond0. xdp->rxq->dev, on the other hand, comes from > XDP and the real driver e.g. eth0, thus xs->dev != xdp->rxq->dev. The > problem here is that only _rx[] of bond0 is populated with the pool > pointer at bind time, so dev->_rx[qid].pool is NULL as it refers to > the _rx of eth0 that was never set. The solution here is then to make > sure that the _rx[] of bond0 is propagated to eth0 (and any other device > bonded to bond0). > > Two new features are needed to support this: > > 1) A helper that copies _rx[].pool from one struct to another > 2) A new xsk_bind netdev event that a driver can subscribe to. Will be called > whenever a xsk socket is bound to a device. > > In the case the socket is bound to bond0 before eth0 is bonded to > bond0, only 1) needs to be used in the bonding driver. > > In the case the socket is bound to bond0 after bonding of eth0 to > bond0, the bonding driver need to subscribe to 2) and in the event > handle call 1). > > * ZERO-COPY > > 1) Relay through the XDP_SETUP_XSK_POOL command in NDO_BPF to the > bonded devices. > > 2) Relay through the ndo_xsk_wakeup to the bonded devices. > > Standby mode seems straight-forward to support. > > How to deal with round-robin mode in the bonding driver? Not possible > to have multiple bonded devices access the same ring. Would require > multiple rings and copying to them. Also not clear how to propagate > the need_wakeup flags of the individual network devices to the one of > the bond device. I think this kind of functionality is much better > performed in user-space with a lib. Simpler and faster. I think this goes for all the things you mentioned above. There is no way we can make this consistent with the in-kernel bond behaviour, so it's going to be a pretty leaky abstraction anyway. So I don't think we should add all this complexity, it's better to handle this in userspace (and just attach to the component interfaces). In fact, I think supporting XDP at all on the bond interface was a mistake; let's not exacerbate it :/ -Toke
|