[ovs-dev] [PATCHv2] netdev-afxdp: Add interrupt mode netdev class.

William Tu u9012063 at gmail.com
Thu Feb 27 19:52:31 UTC 2020


On Thu, Feb 27, 2020 at 11:10 AM William Tu <u9012063 at gmail.com> wrote:
>
> On Thu, Feb 27, 2020 at 9:54 AM Ilya Maximets <i.maximets at ovn.org> wrote:
> >
> > On 2/27/20 6:51 PM, William Tu wrote:
> > > On Thu, Feb 27, 2020 at 9:42 AM Ilya Maximets <i.maximets at ovn.org> wrote:
> > >>
> > >> On 2/27/20 6:30 PM, William Tu wrote:
> > >>> The patch adds a new netdev class 'afxdp-nonpmd' to enable afxdp
> > >>> interrupt mode. This is similar to 'type=afxdp', except that the
> > >>> is_pmd field is set to false. As a result, the packet processing
> > >>> is handled by main thread, not pmd thread. This avoids burning
> > >>> the CPU to always 100% when there is no traffic.
> > >>>
> > >>> Tested-at: https://travis-ci.org/williamtu/ovs-travis/builds/655885506
> > >>> Signed-off-by: William Tu <u9012063 at gmail.com>
> > >>> ---
> > >>>  NEWS                  |  3 +++
> > >>>  lib/netdev-afxdp.c    | 20 +++++++++++++++++++-
> > >>>  lib/netdev-linux.c    | 20 ++++++++++++++++++++
> > >>>  lib/netdev-provider.h |  1 +
> > >>>  lib/netdev.c          |  1 +
> > >>>  tests/system-afxdp.at | 23 +++++++++++++++++++++++
> > >>>  6 files changed, 67 insertions(+), 1 deletion(-)
> > >>>
> > >>> diff --git a/NEWS b/NEWS
> > >>> index f62ef1f47ea8..594c55dc11d6 100644
> > >>> --- a/NEWS
> > >>> +++ b/NEWS
> > >>> @@ -4,6 +4,9 @@ Post-v2.13.0
> > >>>     - OpenFlow:
> > >>>       * The OpenFlow ofp_desc/serial_num may now be configured by setting the
> > >>>         value of other-config:dp-sn in the Bridge table.
> > >>> +   - AF_XDP:
> > >>> +     * New netdev class 'afxdp-nonpmd' for netdev-afxdp to save CPU cycles
> > >>> +       by enabling interrupt mode.
> > >>>
> > >>>
> > >>>  v2.13.0 - 14 Feb 2020
> > >>> diff --git a/lib/netdev-afxdp.c b/lib/netdev-afxdp.c
> > >>> index 482400d8d135..cd2c7c381139 100644
> > >>> --- a/lib/netdev-afxdp.c
> > >>> +++ b/lib/netdev-afxdp.c
> > >>> @@ -169,6 +169,12 @@ struct netdev_afxdp_tx_lock {
> > >>>      );
> > >>>  };
> > >>>
> > >>> +static int nonpmd_cnt;  /* Number of afxdp netdevs in non-pmd mode. */
> > >>> +static bool
> > >>> +netdev_is_afxdp_nonpmd(struct netdev *netdev) {
> > >>> +    return netdev_get_class(netdev) == &netdev_afxdp_nonpmd_class;
> > >>> +}
> > >>> +
> > >>>  #ifdef HAVE_XDP_NEED_WAKEUP
> > >>>  static inline void
> > >>>  xsk_rx_wakeup_if_needed(struct xsk_umem_info *umem,
> > >>> @@ -1115,7 +1121,10 @@ netdev_afxdp_batch_send(struct netdev *netdev, int qid,
> > >>>      struct netdev_linux *dev;
> > >>>      int ret;
> > >>>
> > >>> -    if (concurrent_txq) {
> > >>> +    /* Lock is required when mixing afxdp pmd and nonpmd mode.
> > >>> +     * ex: one device is created 'afxdp', the other is 'afxdp-nonpmd'.
> > >>> +     */
> > >>> +    if (concurrent_txq || (nonpmd_cnt != 0)) {
> > >>
> > >> I'm confused about this change.  Are you expecting same device being
> > >> opened twice with different netdev type?  Database should not allow this.
> > >
> > > Not the same device.
> > >
> > > For example, dev1 opened as 'afxdp', and dev2 opened as 'afxdp-nonpmd'.
> > > When dev2 receives a packet and send to dev1, the main thread used by dev2
> > > is calling the send(), while it is possible that dev2's pmd thread is
> > > also calling
> > > send(). So need a lock here.
> >
> > But they will send to different tx queues.
>
> OK,  I think you are talking about the XPS feature (dynamic txqs).
> I thought XPS can only work when all between pmd threads.
> So adding a lock here.
> The patch currently doesn't work without the lock. Let me investigate more.
>
More details.
On my test using veth (only have 1 rxq 1 txq)
Somehow the main thread is assigned to use queue id  = 2, which causes segfault.
I will send v3 patch.
Thanks
William


More information about the dev mailing list