netdev - Re: [PATCH intel-next 2/2] ice: introduce XDP Tx fallback path

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <87czt5dal0.fsf@toke.dk>
Date:   Tue, 01 Jun 2021 14:38:03 +0200
From:   Toke Høiland-Jørgensen <toke@...hat.com>
To:     Maciej Fijalkowski <maciej.fijalkowski@...el.com>,
        intel-wired-lan@...ts.osuosl.org
Cc:     netdev@...r.kernel.org, bpf@...r.kernel.org, davem@...emloft.net,
        anthony.l.nguyen@...el.com, kuba@...nel.org, bjorn@...nel.org,
        magnus.karlsson@...el.com,
        Maciej Fijalkowski <maciej.fijalkowski@...el.com>
Subject: Re: [PATCH intel-next 2/2] ice: introduce XDP Tx fallback path

Maciej Fijalkowski <maciej.fijalkowski@...el.com> writes:

> Under rare circumstances there might be a situation where a requirement
> of having a XDP Tx queue per core could not be fulfilled and some of the
> Tx resources would have to be shared between cores. This yields a need
> for placing accesses to xdp_rings array onto critical section protected
> by spinlock.
>
> Design of handling such scenario is to at first find out how many queues
> are there that XDP could use. Any number that is not less than the half
> of a count of cores of platform is allowed. XDP queue count < cpu count
> is signalled via new VSI state ICE_VSI_XDP_FALLBACK which carries the
> information further down to Rx rings where new ICE_TX_XDP_LOCKED is set
> based on the mentioned VSI state. This ring flag indicates that locking
> variants for getting/putting xdp_ring need to be used in fast path.
>
> For XDP_REDIRECT the impact on standard case (one XDP ring per CPU) can
> be reduced a bit by providing a separate ndo_xdp_xmit and swap it at
> configuration time. However, due to the fact that net_device_ops struct
> is a const, it is not possible to replace a single ndo, so for the
> locking variant of ndo_xdp_xmit, whole net_device_ops needs to be
> replayed.
>
> It has an impact on performance (1-2 %) of a non-fallback path as
> branches are introduced.

I generally feel this is the right approach, although the performance
impact is a bit unfortunately, obviously. Maybe it could be avoided by
the use of static_branch? I.e., keep a global refcount of how many
netdevs are using the locked path and only activate the check in the
fast path while that refcount is >0?

-Toke