Re: [PATCH RFC v2 1/8] xdp: Infrastructure to generalize XDP

On 14/02/17 21:07, Tom Herbert wrote:

Off the top of my head... I'd say may we might be able to have a
minimally invasive interface with something like:

XDP_RUN(hook, xdp, drv_xdp_handle_action)

This replaces xdp_run and return codes are processed in the called
functions. Its a macro so that xdp_handle_action can be inlined.

I don't see why callbacks are needed, since XDP programs (I assume)
 aren't supposed to block.  This XDP_RUN ends up looking a lot like
 NF_HOOK, for no good reason that I can see (unlike NF hooks, we never
 do things like NF_QUEUE).

Batching could then be done in the backend XDP so that it would be
transparent to the driver.

I also don't see how you can transparently batch and still allow the
 handler to be inlined - you'd have to stash a function pointer that
 you could call later when you decide to dispatch a batch of packets.

To me, the sensible interface (which makes the batching explicit to
 the driver, which I think is necessary) is to have an int (or maybe
 unsigned int, which is the return type of xdp_hookfn, I'm not sure
 which is intended) member in struct xdp_buff.
Then the driver can call something like
	XDP_RUN_ARRAY(napi, xdp_array, array_len);
which is semantically equivalent to
	unsigned int i;
	for (i = 0; i < array_len; i++)
		xdp_array[i].ret = xdp_hook_run(napi, xdp_array + i);
except that it may run the hooks in 'row-major order'.
No callbacks needed, the driver can just loop over xdp_array reading
 the .ret and applying the relevant action to each packet.

This also has the advantage that the driver knows how many packets it
 might have to process in a single batch (i.e. NAPI_POLL_WEIGHT) and
 can allocate the array statically, whereas an XDP hook that tried to
 transparently be 'helpful' would have to guess and/or use kmalloc.

-Ed