> On Tue, Jun 09, 2020 at 09:41:10AM +0200, Lorenzo Bianconi wrote: > > > On Tue, Jun 09, 2020 at 12:02:39AM +0200, Lorenzo Bianconi wrote: > > > > Disable frames injection in mvneta_xdp_xmit routine during hw > > > > re-configuration in order to avoid hardware hangs > > > > > > Hi Lorenzo > > > > > > Why does mvneta_tx() also not need the same protection? > > > > > > Andrew > > > > Hi Andrew, > > > > So far I have not been able to trigger the issue in the legacy tx path. > > Even if you have not hit the issue, do you still think it is possible? > If it is hard to trigger, maybe it is worth protecting against it, > just in case.
The issue occurs putting the device down while it is still transmitting. In particular mvneta_port_down() fails to stop tx (TIMEOUT for TX stopped status=...) and the device is not able to recover. The above pattern can occur with XDP because if we remove the program from a running interface, we will put the interface down for DMA buffers reconfiguration while mvneta_xdp_xmit() is concurrently running on a remote cpu. Looking at the code I do not think it can occurs in the legacy tx path (mvneta_tx()) since __dev_close() (trigger by userspace) will run dev_deactivate_many() before running mvneta_stop(). > > > I hit the problem adding the capability to attach an eBPF program to CPUMAP > > entries [1]. In particular I am redirecting traffic to mvneta and > > concurrently > > attaching/removing a XDP program to/from it. > > I am not sure this can occur running mvneta_tx(). > > Moreover it seems a common pattern for .ndo_xdp_xmit() in other drivers > > (e.g ixgbe, bnxt, mlx5) > > I was wondering if this should be solved at a higher level. And if you > say there are more MAC drivers with this issue, maybe it should. Not > sure how though. It seems like MTU change and rx mode change wound > need to be protected, which at a higher level is harder to do. What > exactly do you need to protect, in a generic way? Yes, we can think about it but I guess we should fix the issue first since it is already there and it will be easy to backport the fix, agree? Regards, Lorenzo > > Andrew
signature.asc
Description: PGP signature
