net: introduce and use generic XDP stats

[alobakin]

This is an almost complete rework of [0].

This series introduces generic XDP statistics infra based on rtnl
xstats, and wires up the drivers which collect appropriate statistics
to this new interface. Finally, it introduces XDP/XSK statistics to
all XDP-capable Intel drivers.

Those counters are:
* packets: number of frames passed to bpf_prog_run_xdp().
* bytes: number of bytes went through bpf_prog_run_xdp().
* errors: number of general XDP errors, if driver has one unified
  counter.
* aborted: number of XDP_ABORTED returns.
* drop: number of XDP_DROP returns.
* invalid: number of returns of unallowed values (i.e. not XDP_*).
* pass: number of XDP_PASS returns.
* redirect: number of successfully performed XDP_REDIRECT requests.
* redirect_errors: number of failed XDP_REDIRECT requests.
* tx: number of successfully performed XDP_TX requests.
* tx_errors: number of failed XDP_TX requests.
* xmit_packets: number of successfully transmitted XDP/XSK frames.
* xmit_bytes: number of successfully transmitted XDP/XSK frames.
* xmit_errors: of XDP/XSK frames failed to transmit.
* xmit_full: number of XDP/XSK queue being full at the moment of
  transmission.

To provide them, developers need to implement .ndo_get_xdp_stats()
and, if they want to expose stats on a per-channel basis,
.ndo_get_xdp_stats_nch(). include/net/xdp.h contains some helper
structs and functions which might be useful for doing this.
It is up to developers to decide whether to implement XDP stats in
their drivers or not, depending on the needs and so on, but if so,
it is implied that they will be implemented using this new infra
rather than custom Ethtool entries. XDPs stat {,type} list can be
expanded if needed as counters are being provided as nested NL attrs.
There's an option to provide XDP and XSK counters separately, and I
used it in mlx5 (has separate RQs and SQs) and Intel (have separate
NAPI poll routines) drivers.

This series doesn't touches existing Ethtool-exposed XDP stats due
to the potential breakage of custom{,er} scripts and other stuff
that might be hardwired on their presence. Developers are free to
drop them on their own if possible. In ideal case we would see
Ethtool stats free from anything related to XDP, but it's unlikely
to happen (:

XDP_PASS kpps on an ice NIC with ~50 Gbps line rate:

Frame size     64    | 128   | 256   | 512   | 1024 | 1532
----------------------------------------------------------
net-next       23557 | 23750 | 20731 | 11723 | 6270 | 4377
This series    23484 | 23812 | 20679 | 11720 | 6270 | 4377

The same situation with XDP_DROP and several more common cases:
nothing past stddev (which is a bit surprising, but not complaining
at all).

A brief series breakdown:
 * 1-2: introduce new infra and APIs, for rtnetlink and drivers
   respectively;
 * 3-19: add needed callback to the existing drivers to export
   their stats using new infra. Some of the patches are cosmetic
   prereqs;
 * 20-25: add XDP/XSK stats to all Intel drivers;
 * 26: mention generic XDP stats in Documentation.

From v1 [0]:
 - use rtnl xstats instead of Ethtool standard -- XDP stats are
   purely software while Ethtool infra was designed for HW stats
   (Jakub);
 - split xmit into xmit_packets and xmit_bytes, add xmit_full;
 - don't touch existing drivers custom XDP stats exposed via
   Ethtool (Jakub);
 - add a bunch of helper structs and functions to reduce boilerplates
   and code duplication in new drivers;
 - merge with the series which adds XDP stats to Intel drivers;
 - add some perf numbers per Jakub's request.

[0] https://lore.kernel.org/all/[email protected]