Re: [PATCH net-next] net/mlx5: remove the recent devlink params

On Tue, 2021-10-26 at 09:47 -0700, Jakub Kicinski wrote:

On Tue, 26 Oct 2021 15:53:17 +0000 Parav Pandit wrote:

quoted

Hi Jakub,

quoted

From: Jakub Kicinski <kuba@kernel.org>
Sent: Tuesday, October 26, 2021 9:00 PM
To: davem@davemloft.net
Cc: Saeed Mahameed <saeedm@nvidia.com>; netdev@vger.kernel.org;
Leon
Romanovsky [off-list ref]; Jakub Kicinski [off-list ref]
Subject: [PATCH net-next] net/mlx5: remove the recent devlink
params

revert commit 46ae40b94d88 ("net/mlx5: Let user configure
io_eq_size
param") revert commit a6cb08daa3b4 ("net/mlx5: Let user configure
event_eq_size param") revert commit 554604061979 ("net/mlx5: Let
user
configure max_macs param")

The EQE parameters are applicable to more drivers, they should be
configured
via standard API, probably ethtool. Example of another driver
needing
something similar:  

ethool is not a good choice for following reasons.

1. EQs of the mlx5 core devlink instance is used by multiple
auxiliary devices, namely net, rdma, vdpa.
2. And sometime netdevice doesn't even exists to operate via
ethtool (but devlink instance exist to serve other devices).
3. ethtool doesn't have notion set the config and apply (like
devlink reload)
Such reload operation is useful when user wants to configure more
than one parameter and initialize the device only once.
Otherwise dynamically changing parameter results in multiple device
re-init sequence that increases device setup time.

Sure these are good points. OTOH devlink doesn't have any notion of
queues, IRQ vectors etc today so it doesn't really fit there, either.

The point is, mlx5 has different architecture compared to other
vendors, EQs, IRQs are considered shared resources of upper layer
protocols/interfaces.

ethtool is netdev specific. 
EQ concept might not apply to all vendors even with last gen HW.

netdev/ethtool shouldn't be controlling concrete/arch specific
resources. better if we kept ethtool abstract (queues, rings, and
ethernet specific ..)

quoted

Should we define the devlink resources in the devlink layer,
instead of driver?

I'm not sure how the EQE fits into the existing devlink objects.
params are not much of an API, it's a garbage bag.

devlink is good place, it is what we decide it should be, a garbage bag
or a standard place for all generic networking device related knobs.

quoted

So that multiple drivers can make use of them without redefinition?

Yes, please.

let's send a RFC, CC Jiri of course and other vendors who might be
interested according to Jakub's grep (huawei/hinic, ibm/ehea,
microsoft/mana, qlogic/qed) ..