[stripping some of the bouncy CCs.]

On Tue, 11 Nov 2025 15:39:03 +0100 Jiri Pirko wrote:

Tue, Nov 11, 2025 at 04:26:34AM +0100, saeed@kernel.org wrote:

quoted

On 10 Nov 15:46, Jakub Kicinski wrote:  

quoted

On Sun, 9 Nov 2025 11:46:37 +0100 Jiri Pirko wrote:  

quoted

quoted

So, I checked a couple of flows internally, and it seems this allows
some flexibility in the FW to decide later on which mode to pick,
based on other parameters, which practically means
"user has no preference on this param". Driver can only find out
after boot, when it reads the runtime capabilities, but still
this is a bug, by the time the driver reads this (in devlink), the
default value should've already been determined by FW, so FW must
return the actual runtime value. Which can only be one of the following  

I don't think it is correct to expose the "default" as a value.

On read, user should see the configured value, either "full_csum" or
"l4_only". Reporting "default" to the user does not make any sense.
On write, user should pass either "full_csum" or "l4_only". Why we would
ever want to pass "default"?  

FWIW I agree that this feels a bit odd. Should the default be a flag
attr? On get flag being present means the value is the FW default (no
override present). On set passing the flag means user wants to reset
to FW default (remove override)?

Y'all did not respond to this part, should we assume that what 
I described is clear and makes sense? I think we should make that
part of the series, unlike the pending indication.

quoted

quoted

quoted

Regardless this patch, since this is param to be reflected on fw reboot
(permanent cmode), I think it would be nice to expose indication if
param value passed to user currently affects the fw, or if it is going
to be applied after fw reboot. Perhaps a simple bool attr would do?  

IIUC we're basically talking about user having no information that
the update is pending? Could this be done by the core? Core can do
a ->get prior to calling ->set and if the ->set succeeds and
cmode != runtime record that the update is pending?
  

Could work if on GET driver reads 'current' value from FW, then it should
be simpler if GET != SET then 'pending', one problem though is if SET was
done by external tool or value wasn't applied after reboot, then we loose
that information, but do we care? I think we shouldn't.
 

quoted

That feels very separate from the series tho, there are 3 permanent
params in mlx5, already. Is there something that makes this one special?  

Agreed. That is why I wrote "regardless this patch". But I think the
pending indication is definitelly nice to have.

Yes, I've been wondering why it's missing since the day devlink params
were added :)

quoted

In mlx5 they all have the same behavior, devlink sets 'next' value, devlink
reads 'next' value. The only special thing about the new param
is that it has a 'device_default' value and when you read that from 'next' it
will always show 'device_default' as the actual value is only
known at run time ,e.g. 'next boot'.

I think the only valid solution for permanent and drv_init params is to
have 'next' and 'current' values reported by driver on read. Or maybe go just
with  'set' != 'get' then 'pending' as discussed above ?  

Hmm, is it possible to rebind the driver without fw going through
next-boot phase? I'm wondering if it wouldn't be safer to have this
pending flag set to be responsibility of the driver...

The downside is that drivers may either have bugs or not implement 
the new feature. So when there's no indication of pending change
the user will have no idea whether its because there's none or the
driver simply does not report both. 

My experience implementing the pending FW version in a couple of
products is that it takes a lot of "discussions" to get FW people 
to implement this sort of a thing right. mlx5 already has the right 
FW APIs so we should allow their full use. But I don't think the 
"what if user change the setting with fwctl", "what if user reloaded
the driver" corner cases should stop us from trying to get the core
to implement 99% of the cases right.

FTR I'm not aware of any Meta-internal products having permanent knobs.
I just don't think we can depend on the random people that submit
drivers these days to get this right. And devlink users will assume
that it's Linux that sucks if it doesn't work right, not vendor X.

Long story short I think we can add the reporting of both values via GET
but I'd definitely still like to see the core trying to do the tracking
automatically.