On Thu, 2019-02-07 at 08:48 +0100, Jesper Dangaard Brouer wrote: > On Wed, 6 Feb 2019 00:06:33 +0000 Saeed Mahameed <sae...@mellanox.com > > wrote: > > > On Mon, 2019-02-04 at 19:13 -0800, David Ahern wrote: > [...] > > > mlx5 needs some work. As I recall it still has the bug/panic > > > removing xdp programs - at least I don't recall seeing a patch > > > for > > > it. > > > > Only when xdp_redirect to mlx5, and removing the program while > > redirect is happening, this is actually due to a lack of > > synchronization means between different drivers, we have some ideas > > to overcome this using a standard XDP API, or just use a hack in > > mlx5 > > driver which i don't like: > > > > https://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux.git/commit/?h=topic/xdp-redirect-fix&id=a3652d03cc35fd3ad62744986c8ccaca74c9f20c > > > > I will be working on this towards the end of this week. > > Toke and I have been discussing how to solve this. > > The main idea for fixing this is to tie resource allocation to > interface > insertion into interface maps (kernel/bpf/devmap.c). As the =devmap= > already have the needed synchronisation mechanisms and steps for > safely > adding and removing =net_devices= (e.g. stopping RX side, flushing > remaining frames, waiting RCU period before freeing objects, etc.) > > As described here: > > https://github.com/xdp-project/xdp-project/blob/master/xdp-project.org#better-ndo_xdp_xmit-resource-management > > --Jesper
Yes you already suggested this approach @LPC: So 1) on dev_map_update_elem() we will call dev->dev->ndo_bpf() to notify the device on the intention to start/stop redirect, and wait for it to create/destroy the HW resources before/after actually updating the map But: 2) this won't totally solve our problem, since sometimes the driver can decide to recreate (change of configuration) hw resources on the fly while redirect/devmap is already happening, so we need some kind of a dev_map_notification or a flag with rcu synch, for when the driver want to make the xdp redirect resources unavailable. Thanks, Saeed.
