On Thu, Nov 02, 2023 at 02:23:02PM +0100, Maciej Fijalkowski wrote:
> On Tue, Oct 31, 2023 at 03:22:31PM +0100, Larysa Zaremba wrote:
> > On Sat, Oct 28, 2023 at 09:55:52PM +0200, Maciej Fijalkowski wrote:
> > > On Mon, Oct 23, 2023 at 11:35:46AM +0200, Larysa Zaremba wrote:
> > > > On Fri, Oct 20, 2023 at 06:32:13PM +0200, Maciej Fijalkowski wrote:
> > > > > On Thu, Oct 12, 2023 at 07:05:17PM +0200, Larysa Zaremba wrote:
> > > > > > Usage of XDP hints requires putting additional information after the
> > > > > > xdp_buff. In basic case, only the descriptor has to be copied on a
> > > > > > per-packet basis, because xdp_buff permanently resides before per-ring
> > > > > > metadata (cached time and VLAN protocol ID).
> > > > > >
> > > > > > However, in ZC mode, xdp_buffs come from a pool, so memory after such
> > > > > > buffer does not contain any reliable information, so everything has to be
> > > > > > copied, damaging the performance.
> > > > > >
> > > > > > Introduce a static key to enable meta sources assignment only when attached
> > > > > > XDP program is device-bound.
> > > > > >
> > > > > > This patch eliminates a 6% performance drop in ZC mode, which was a result
> > > > > > of addition of XDP hints to the driver.
> > > > > >
> > > > > > Signed-off-by: Larysa Zaremba <larysa.zaremba@xxxxxxxxx>
> > > > > > ---
> > > > > > drivers/net/ethernet/intel/ice/ice.h | 1 +
> > > > > > drivers/net/ethernet/intel/ice/ice_main.c | 14 ++++++++++++++
> > > > > > drivers/net/ethernet/intel/ice/ice_txrx.c | 3 ++-
> > > > > > drivers/net/ethernet/intel/ice/ice_xsk.c | 3 +++
> > > > > > 4 files changed, 20 insertions(+), 1 deletion(-)
> > > > > >
> > > > > > diff --git a/drivers/net/ethernet/intel/ice/ice.h b/drivers/net/ethernet/intel/ice/ice.h
> > > > > > index 3d0f15f8b2b8..76d22be878a4 100644
> > > > > > --- a/drivers/net/ethernet/intel/ice/ice.h
> > > > > > +++ b/drivers/net/ethernet/intel/ice/ice.h
> > > > > > @@ -210,6 +210,7 @@ enum ice_feature {
> > > > > > };
> > > > > >
> > > > > > DECLARE_STATIC_KEY_FALSE(ice_xdp_locking_key);
> > > > > > +DECLARE_STATIC_KEY_FALSE(ice_xdp_meta_key);
> > > > > >
> > > > > > struct ice_channel {
> > > > > > struct list_head list;
> > > > > > diff --git a/drivers/net/ethernet/intel/ice/ice_main.c b/drivers/net/ethernet/intel/ice/ice_main.c
> > > > > > index 47e8920e1727..ee0df86d34b7 100644
> > > > > > --- a/drivers/net/ethernet/intel/ice/ice_main.c
> > > > > > +++ b/drivers/net/ethernet/intel/ice/ice_main.c
> > > > > > @@ -48,6 +48,9 @@ MODULE_PARM_DESC(debug, "netif level (0=none,...,16=all)");
> > > > > > DEFINE_STATIC_KEY_FALSE(ice_xdp_locking_key);
> > > > > > EXPORT_SYMBOL(ice_xdp_locking_key);
> > > > > >
> > > > > > +DEFINE_STATIC_KEY_FALSE(ice_xdp_meta_key);
> > > > > > +EXPORT_SYMBOL(ice_xdp_meta_key);
> > > > > > +
> > > > > > /**
> > > > > > * ice_hw_to_dev - Get device pointer from the hardware structure
> > > > > > * @hw: pointer to the device HW structure
> > > > > > @@ -2634,6 +2637,11 @@ static int ice_xdp_alloc_setup_rings(struct ice_vsi *vsi)
> > > > > > return -ENOMEM;
> > > > > > }
> > > > > >
> > > > > > +static bool ice_xdp_prog_has_meta(struct bpf_prog *prog)
> > > > > > +{
> > > > > > + return prog && prog->aux->dev_bound;
> > > > > > +}
> > > > > > +
> > > > > > /**
> > > > > > * ice_vsi_assign_bpf_prog - set or clear bpf prog pointer on VSI
> > > > > > * @vsi: VSI to set the bpf prog on
> > > > > > @@ -2644,10 +2652,16 @@ static void ice_vsi_assign_bpf_prog(struct ice_vsi *vsi, struct bpf_prog *prog)
> > > > > > struct bpf_prog *old_prog;
> > > > > > int i;
> > > > > >
> > > > > > + if (ice_xdp_prog_has_meta(prog))
> > > > > > + static_branch_inc(&ice_xdp_meta_key);
> > > > >
> > > > > i thought boolean key would be enough but inc/dec should serve properly
> > > > > for example prog hotswap cases.
> > > > >
> > > >
> > > > My thought process on using counting instead of boolean was: there can be
> > > > several PFs that use the same driver, so therefore we need to keep track of how
> > > > many od them use hints.
> > >
> > > Very good point. This implies that if PF0 has hints-enabled prog loaded,
> > > PF1 with non-hints prog will "suffer" from it.
> > >
> > > Sorry for such a long delays in responses but I was having a hard time
> > > making up my mind about it. In the end I have come up to some conclusions.
> > > I know the timing for sending this response is not ideal, but I need to
> > > get this off my chest and bring discussion back to life:)
> > >
> > > IMHO having static keys to eliminate ZC overhead does not scale. I assume
> > > every other driver would have to follow that.
> > >
> > > XSK pool allows us to avoid initializing various things per each packet.
> > > Instead, taking xdp_rxq_info as an example, each xdp_buff from pool has
> > > xdp_rxq_info assigned at init time. With this in mind, we should have some
> > > mechanism to set hints-specific things in xdp_buff_xsk::cb, at init time
> > > as well. Such mechanism should not require us to expose driver's private
> > > xdp_buff hints containers (such as ice_pkt_ctx) to XSK pool.
> > >
> > > Right now you moved phctime down to ice_pkt_ctx and to me that's the main
> > > reason we have to copy ice_pkt_ctx to each xdp_buff on ZC. What if we keep
> > > the cached_phctime at original offset in ring but ice_pkt_ctx would get a
> > > pointer to that?
> > >
> > > This would allow us to init the pointer in each xdp_buff from XSK pool at
> > > init time. I have come up with a way to program that via so called XSK
> > > meta descriptors. Each desc would have data to write onto cb, offset
> > > within cb and amount of bytes to write/copy.
> > >
> > > I'll share the diff below but note that I didn't measure how much lower
> > > the performance is degraded. My icelake machine where I used to measure
> > > performance-sensitive code got broke. For now we can't escape initing
> > > eop_desc per each xdp_buff, but I moved it to alloc side, as we mangle
> > > descs there anyway.
> > >
> > > I think mlx5 could benefit from that approach as well with initing the rq
> > > ptr at init time.
> > >
> > > Diff does mostly these things:
> > > - move cached_phctime to old place in ice_rx_ring and add ptr to that in
> > > ice_pkt_ctx
> > > - introduce xsk_pool_set_meta()
> > > - use it from ice side.
> > >
> >
> > Thank you for the code! I will probably send v7 with such changes. Are you OK,
> > if patch with core changes would go with you as an author?
>
> Yes or I can produce a patch and share, up to you.
>
I have already started, your diff does not compile, so I took some creative
liberty. Will send you patches for verification this week.
> >
> > But also, I see a minor problem with that switching VLAN protocol does not
> > trigger buffer allocation, so we have to point to that too, this probably means
> > moving cached time back and finding 16 extra bits in CL3. Single pointer to
> > {cached time, vlan_proto} would be copied to be after xdp_buff.
>
> It's not that it has to trigger buffer allocation, we could stop the
> interface if pool is present and update vlan proto on pool's xdp_buffs
> (from quick glance i don't see that we're stopping iface for setting vlan
> features) but that sounds like more of a hassle to do...
>
> So yeah maybe let's just have a ptr in ice_pkt_ctx as well.
>
> [...]
