[Intel-wired-lan] [PATCH] ixgbe: let the xdpdrv work with more than 64 cpus

[Jason Xing]

On Tue, Aug 24, 2021 at 9:32 PM Jesper Dangaard Brouer
<jbrouer at redhat.com> wrote:
>
>
>
> On 24/08/2021 12.49, kerneljasonxing at gmail.com wrote:
> > From: Jason Xing <xingwanli at kuaishou.com>
> >
> > Originally, ixgbe driver doesn't allow the mounting of xdpdrv if the
> > server is equipped with more than 64 cpus online. So it turns out that
> > the loading of xdpdrv causes the "NOMEM" failure.
> >
> > Actually, we can adjust the algorithm and then make it work, which has
> > no harm at all, only if we set the maxmium number of xdp queues.
>
> This is not true, it can cause harm, because XDP transmission queues are
> used without locking. See drivers ndo_xdp_xmit function ixgbe_xdp_xmit().
> As driver assumption is that each CPU have its own XDP TX-queue.
>

Point taken. I indeed miss that part which would cause bad behavior if it
happens.

At this point, I think I should find all the allocation and use of XDP
related, say,
queues and rings, then adjust them all?

Let's say if the server is shipped with 128 cpus, we could map 128 cpus to 64
rings in the function ixgbe_xdp_xmit(). However, it sounds a little bit odd.

Do you think that it makes any sense?

Thanks,
Jason

> This patch is not a proper fix.
>
> I do think we need a proper fix for this issue on ixgbe.
>
>
> > Fixes: 33fdc82f08 ("ixgbe: add support for XDP_TX action")
> > Co-developed-by: Shujin Li <lishujin at kuaishou.com>
> > Signed-off-by: Shujin Li <lishujin at kuaishou.com>
> > Signed-off-by: Jason Xing <xingwanli at kuaishou.com>
> > ---
> >   drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c  | 2 +-
> >   drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 3 ---
> >   2 files changed, 1 insertion(+), 4 deletions(-)
> >
> > diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c
> > index 0218f6c..5953996 100644
> > --- a/drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c
> > +++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c
> > @@ -299,7 +299,7 @@ static void ixgbe_cache_ring_register(struct ixgbe_adapter *adapter)
> >
> >   static int ixgbe_xdp_queues(struct ixgbe_adapter *adapter)
> >   {
> > -     return adapter->xdp_prog ? nr_cpu_ids : 0;
> > +     return adapter->xdp_prog ? min_t(int, MAX_XDP_QUEUES, nr_cpu_ids) : 0;
> >   }
> >
> >   #define IXGBE_RSS_64Q_MASK  0x3F
> > diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> > index 14aea40..b36d16b 100644
> > --- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> > +++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> > @@ -10130,9 +10130,6 @@ static int ixgbe_xdp_setup(struct net_device *dev, struct bpf_prog *prog)
> >                       return -EINVAL;
> >       }
> >
> > -     if (nr_cpu_ids > MAX_XDP_QUEUES)
> > -             return -ENOMEM;
> > -
> >       old_prog = xchg(&adapter->xdp_prog, prog);
> >       need_reset = (!!prog != !!old_prog);
> >
> >
>