[xdp-hints] Re: [RFC bpf-next 05/23] ice: Introduce ice_xdp_buff

[Larysa Zaremba <larysa.zaremba@intel.com>]

On Thu, Sep 07, 2023 at 06:43:58PM +0200, Maciej Fijalkowski wrote:
> On Thu, Sep 07, 2023 at 06:42:33PM +0200, Maciej Fijalkowski wrote:
> > On Thu, Sep 07, 2023 at 09:33:14AM -0700, Stanislav Fomichev wrote:
> > > On Thu, Sep 7, 2023 at 7:27 AM Larysa Zaremba <larysa.zaremba@intel.com> wrote:
> > > >
> > > > On Tue, Sep 05, 2023 at 07:53:03PM +0200, Maciej Fijalkowski wrote:
> > > > > On Mon, Sep 04, 2023 at 08:11:09PM +0200, Larysa Zaremba wrote:
> > > > > > On Mon, Sep 04, 2023 at 05:32:14PM +0200, Maciej Fijalkowski wrote:
> > > > > > > On Thu, Aug 24, 2023 at 09:26:44PM +0200, Larysa Zaremba wrote:
> > > > > > > > In order to use XDP hints via kfuncs we need to put
> > > > > > > > RX descriptor and ring pointers just next to xdp_buff.
> > > > > > > > Same as in hints implementations in other drivers, we achieve
> > > > > > > > this through putting xdp_buff into a child structure.
> > > > > > >
> > > > > > > Don't you mean a parent struct? xdp_buff will be 'child' of ice_xdp_buff
> > > > > > > if i'm reading this right.
> > > > > > >
> > > > > >
> > > > > > ice_xdp_buff is a child in terms of inheritance (pointer to ice_xdp_buff could
> > > > > > replace pointer to xdp_buff, but not in reverse).
> > > > > >
> > > > > > > >
> > > > > > > > Currently, xdp_buff is stored in the ring structure,
> > > > > > > > so replace it with union that includes child structure.
> > > > > > > > This way enough memory is available while existing XDP code
> > > > > > > > remains isolated from hints.
> > > > > > > >
> > > > > > > > Minimum size of the new child structure (ice_xdp_buff) is exactly
> > > > > > > > 64 bytes (single cache line). To place it at the start of a cache line,
> > > > > > > > move 'next' field from CL1 to CL3, as it isn't used often. This still
> > > > > > > > leaves 128 bits available in CL3 for packet context extensions.
> > > > > > >
> > > > > > > I believe ice_xdp_buff will be beefed up in later patches, so what is the
> > > > > > > point of moving 'next' ? We won't be able to keep ice_xdp_buff in a single
> > > > > > > CL anyway.
> > > > > > >
> > > > > >
> > > > > > It is to at least keep xdp_buff and descriptor pointer (used for every hint) in
> > > > > > a single CL, other fields are situational.
> > > > >
> > > > > Right, something must be moved...still, would be good to see perf
> > > > > before/after :)
> > > > >
> > > > > >
> > > > > > > >
> > > > > > > > Signed-off-by: Larysa Zaremba <larysa.zaremba@intel.com>
> > > > > > > > ---
> > > > > > > >  drivers/net/ethernet/intel/ice/ice_txrx.c     |  7 +++--
> > > > > > > >  drivers/net/ethernet/intel/ice/ice_txrx.h     | 26 ++++++++++++++++---
> > > > > > > >  drivers/net/ethernet/intel/ice/ice_txrx_lib.h | 10 +++++++
> > > > > > > >  3 files changed, 38 insertions(+), 5 deletions(-)
> > > > > > > >
> > > > > > > > diff --git a/drivers/net/ethernet/intel/ice/ice_txrx.c b/drivers/net/ethernet/intel/ice/ice_txrx.c
> > > > > > > > index 40f2f6dabb81..4e6546d9cf85 100644
> > > > > > > > --- a/drivers/net/ethernet/intel/ice/ice_txrx.c
> > > > > > > > +++ b/drivers/net/ethernet/intel/ice/ice_txrx.c
> > > > > > > > @@ -557,13 +557,14 @@ ice_rx_frame_truesize(struct ice_rx_ring *rx_ring, const unsigned int size)
> > > > > > > >   * @xdp_prog: XDP program to run
> > > > > > > >   * @xdp_ring: ring to be used for XDP_TX action
> > > > > > > >   * @rx_buf: Rx buffer to store the XDP action
> > > > > > > > + * @eop_desc: Last descriptor in packet to read metadata from
> > > > > > > >   *
> > > > > > > >   * Returns any of ICE_XDP_{PASS, CONSUMED, TX, REDIR}
> > > > > > > >   */
> > > > > > > >  static void
> > > > > > > >  ice_run_xdp(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp,
> > > > > > > >             struct bpf_prog *xdp_prog, struct ice_tx_ring *xdp_ring,
> > > > > > > > -           struct ice_rx_buf *rx_buf)
> > > > > > > > +           struct ice_rx_buf *rx_buf, union ice_32b_rx_flex_desc *eop_desc)
> > > > > > > >  {
> > > > > > > >         unsigned int ret = ICE_XDP_PASS;
> > > > > > > >         u32 act;
> > > > > > > > @@ -571,6 +572,8 @@ ice_run_xdp(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp,
> > > > > > > >         if (!xdp_prog)
> > > > > > > >                 goto exit;
> > > > > > > >
> > > > > > > > +       ice_xdp_meta_set_desc(xdp, eop_desc);
> > > > > > >
> > > > > > > I am currently not sure if for multi-buffer case HW repeats all the
> > > > > > > necessary info within each descriptor for every frag? IOW shouldn't you be
> > > > > > > using the ice_rx_ring::first_desc?
> > > > > > >
> > > > > > > Would be good to test hints for mbuf case for sure.
> > > > > > >
> > > > > >
> > > > > > In the skb path, we take metadata from the last descriptor only, so this should
> > > > > > be fine. Really worth testing with mbuf though.
> > > >
> > > > I retract my promise to test this with mbuf, as for now hints and mbuf are not
> > > > supposed to go together [0].
> > > 
> > > Hm, I don't think it's intentional. I don't see why mbuf and hints
> > > can't coexist.
> > 
> > They should coexist, xdp mbuf support is an integral part of driver as we
> > know:)
> > 
> > > Anything pops into your mind? Otherwise, can change that mask to be
> > > ~(BPF_F_XDP_DEV_BOUND_ONLY|BPF_F_XDP_HAS_FRAGS) as part of the series
> > > (or separately, up to you).
> > 
> > +1
> 
> IMHO that should be a standalone patch.
>

Sorry for not answering, I was stuck in testing and debugging, wanted to come 
back with a definitive answer. Fortunately, the problems were not caused by
hints and mbuf clashing on some fundamental level, everything works now, so I 
will send the patch that allows to combine them tomorrow.

> > 
> > > 
> > > > Making sure they can co-exist peacefully can be a topic for another series.
> > > > For now I just can just say with high confidence that in case of multi-buffer
> > > > frames, we do have all the supported metadata in the EoP descriptor.
> > > >
> > > > [0] https://elixir.bootlin.com/linux/v6.5.2/source/kernel/bpf/offload.c#L234
> > > >
> > > > >
> > > > > Ok, thanks!
> > > > >
> > >