Re: [RFC PATCH 2/2] bpf: let bpf_warn_invalid_xdp_action() report more info

Hello,

On Wed, 2021-11-17 at 16:48 -0800, Alexei Starovoitov wrote:

On Mon, Nov 15, 2021 at 06:09:28PM +0100, Toke Høiland-Jørgensen wrote:

quoted

Paolo Abeni [off-list ref] writes:

quoted

quoted

quoted

-	pr_warn_once("%s XDP return value %u, expect packet loss!\n",
+	pr_warn_once("%s XDP return value %u on prog %d dev %s attach type %d, expect packet loss!\n",
 		     act > act_max ? "Illegal" : "Driver unsupported",
-		     act);
+		     act, prog->aux->id, dev->name, prog->expected_attach_type);

This will only ever trigger once per reboot even if the message differs,
right? Which makes it less useful as a debugging aid; so I'm not sure if
it's really worth it with this intrusive change unless we also do
something to add a proper debugging aid (like a tracepoint)...

Yes, the idea would be to add a tracepoint there, if there is general
agreement about this change.

I think this patch is needed because the WARN_ONCE splat gives
implicitly information about the related driver and attach type.
Replacing with a simple printk we lose them.

Ah, right, good point. Pointing that out in the commit message might be
a good idea; otherwise people may miss that ;)

Though it's quite a churn across the drivers I think extra verbosity here is justified.
I'd only suggest to print stable things. Like prog->aux->id probably has
little value for the person looking at the logs. That prog id is likely gone.
If it was prog->aux->name it would be more helpful.
Same with expected_attach_type. Why print it at all?
tracepoint is probably good idea too.

Thanks for the feedback.

I tried to select the additional arguments to allow the user/admin
tracking down which program is causing the issue and why. I'm a
complete newbe wrt XDP, so likely my choice were naive.

I thought the id identifies the program in an unambiguous manner. I
understand the program could be unloaded meanwhile, but if that is not
the case the id should be quite useful. Perhaps we could dump both the
id and the name?

I included the attach type as different types support/allow different
actions: the same program could cause the warning or not depending on
it. If that is not useful I can drop the attach type from the next
iteration.

Thanks!

Paolo