[xdp-hints] Re: [PATCH bpf-next v3 2/3] net: stmmac: add Launch Time support to XDP ZC

[Stanislav Fomichev]

On 12/06, Magnus Karlsson wrote:
> On Tue, 5 Dec 2023 at 20:39, Stanislav Fomichev <sdf at google.com> wrote:
> >
> > On 12/05, Willem de Bruijn wrote:
> > > Stanislav Fomichev wrote:
> > > > On Tue, Dec 5, 2023 at 7:34 AM Florian Bezdeka
> > > > <florian.bezdeka at siemens.com> wrote:
> > > > >
> > > > > On Tue, 2023-12-05 at 15:25 +0000, Song, Yoong Siang wrote:
> > > > > > On Monday, December 4, 2023 10:55 PM, Willem de Bruijn wrote:
> > > > > > > Jesper Dangaard Brouer wrote:
> > > > > > > >
> > > > > > > >
> > > > > > > > On 12/3/23 17:51, Song Yoong Siang wrote:
> > > > > > > > > This patch enables Launch Time (Time-Based Scheduling) support to XDP zero
> > > > > > > > > copy via XDP Tx metadata framework.
> > > > > > > > >
> > > > > > > > > Signed-off-by: Song Yoong Siang<yoong.siang.song at intel.com>
> > > > > > > > > ---
> > > > > > > > >   drivers/net/ethernet/stmicro/stmmac/stmmac.h      |  2 ++
> > > > > > > >
> > > > > > > > As requested before, I think we need to see another driver implementing
> > > > > > > > this.
> > > > > > > >
> > > > > > > > I propose driver igc and chip i225.
> > > > > >
> > > > > > Sure. I will include igc patches in next version.
> > > > > >
> > > > > > > >
> > > > > > > > The interesting thing for me is to see how the LaunchTime max 1 second
> > > > > > > > into the future[1] is handled code wise. One suggestion is to add a
> > > > > > > > section to Documentation/networking/xsk-tx-metadata.rst per driver that
> > > > > > > > mentions/documents these different hardware limitations.  It is natural
> > > > > > > > that different types of hardware have limitations.  This is a close-to
> > > > > > > > hardware-level abstraction/API, and IMHO as long as we document the
> > > > > > > > limitations we can expose this API without too many limitations for more
> > > > > > > > capable hardware.
> > > > > >
> > > > > > Sure. I will try to add hardware limitations in documentation.
> > > > > >
> > > > > > >
> > > > > > > I would assume that the kfunc will fail when a value is passed that
> > > > > > > cannot be programmed.
> > > > > > >
> > > > > >
> > > > > > In current design, the xsk_tx_metadata_request() dint got return value.
> > > > > > So user won't know if their request is fail.
> > > > > > It is complex to inform user which request is failing.
> > > > > > Therefore, IMHO, it is good that we let driver handle the error silently.
> > > > > >
> > > > >
> > > > > If the programmed value is invalid, the packet will be "dropped" / will
> > > > > never make it to the wire, right?
> > >
> > > Programmable behavior is to either drop or cap to some boundary
> > > value, such as the farthest programmable time in the future: the
> > > horizon. In fq:
> > >
> > >                 /* Check if packet timestamp is too far in the future. */
> > >                 if (fq_packet_beyond_horizon(skb, q, now)) {
> > >                         if (q->horizon_drop) {
> > >                                         q->stat_horizon_drops++;
> > >                                         return qdisc_drop(skb, sch, to_free);
> > >                         }
> > >                         q->stat_horizon_caps++;
> > >                         skb->tstamp = now + q->horizon;
> > >                 }
> > >                 fq_skb_cb(skb)->time_to_send = skb->tstamp;
> > >
> > > Drop is the more obviously correct mode.
> > >
> > > Programming with a clock source that the driver does not support will
> > > then be a persistent failure.
> > >
> > > Preferably, this driver capability can be queried beforehand (rather
> > > than only through reading error counters afterwards).
> > >
> > > Perhaps it should not be a driver task to convert from possibly
> > > multiple clock sources to the device native clock. Right now, we do
> > > use per-device timecounters for this, implemented in the driver.
> > >
> > > As for which clocks are relevant. For PTP, I suppose the device PHC,
> > > converted to nsec. For pacing offload, TCP uses CLOCK_MONOTONIC.
> >
> > Do we need to expose some generic netdev netlink apis to query/adjust
> > nic clock sources (or maybe there is something existing already)?
> > Then the userspace can be responsible for syncing/converting the
> > timestamps to the internal nic clocks. +1 to trying to avoid doing
> > this in the drivers.
> >
> > > > > That is clearly a situation that the user should be informed about. For
> > > > > RT systems this normally means that something is really wrong regarding
> > > > > timing / cycle overflow. Such systems have to react on that situation.
> > > >
> > > > In general, af_xdp is a bit lacking in this 'notify the user that they
> > > > somehow messed up' area :-(
> > > > For example, pushing a tx descriptor with a wrong addr/len in zc mode
> > > > will not give any visible signal back (besides driver potentially
> > > > spilling something into dmesg as it was in the mlx case).
> > > > We can probably start with having some counters for these events?
> > >
> > > This is because the AF_XDP completion queue descriptor format is only
> > > a u64 address?
> >
> > Yeah. XDP_COPY mode has the descriptor validation which is exported via
> > recvmsg errno, but zerocopy path seems to be too deep in the stack
> > to report something back. And there is no place, as you mention,
> > in the completion ring to report the status.
> >
> > > Could error conditions be reported on tx completion in the metadata,
> > > using xsk_tx_metadata_complete?
> >
> > That would be one way to do it, yes. But then the error reporting depends
> > on the metadata opt-in. Having a separate ring to export the errors,
> > or having a v2 tx-completions layout with extra 'status' field would also
> > work.
> 
> There are error counters for the non-metadata and offloading cases
> above that can be retrieved with the XDP_STATISTICS getsockopt(). From
> if_xdp.h:
> 
> struct xdp_statistics {
>         __u64 rx_dropped; /* Dropped for other reasons */
>         __u64 rx_invalid_descs; /* Dropped due to invalid descriptor */
>         __u64 tx_invalid_descs; /* Dropped due to invalid descriptor */
>         __u64 rx_ring_full; /* Dropped due to rx ring being full */
>         __u64 rx_fill_ring_empty_descs; /* Failed to retrieve item
> from fill ring */
>         __u64 tx_ring_empty_descs; /* Failed to retrieve item from tx ring */
> };
> 
> Albeit, these are aggregate statistics and do not say anything about
> which packet that caused it. Works well for things that are
> programming bugs that should not occur (such as rx_invalid_descs and
> tx_invalid_descs) and requires the programmer to debug and fix his or
> her program, but it does not work for requests that might fail even
> though the program is correct and need to be handled on a packet by
> packet basis. So something needs to be added for that as you both say.
> 
> Would prefer if we could avoid a v2 completion descriptor format or
> another ring that needs to be checked all the time, so if we could
> live with providing the error status in the metadata field of the
> packet at completion time, that would be good. Though having the error
> status in the completion ring would be faster as that cache line is
> hot, while the metadata section of the packet is likely not at
> completion time. So that speaks for a v2 completion ring format. Just
> thinking out loud here.

In this case, maybe adding tx_over_horizon_dropped to XDP_STATISTICS
is all we need here? We can have some new api to query this horizon
per netdev.