Re: [PATCH net-next v4] net/mlx5e: advertise tcp-data-split support

From
Nimrod Oren <hidden>
Date
2026-11-17 17:49:35
Also in
linux-rdma
Reply to
Dimitri Daskalakis, 2026-11-15
On 15/11/2026 22:59, Dimitri Daskalakis wrote:
On 11/15/26 2:39 AM, Nimrod Oren wrote:
quoted
On 11/11/2026 19:05, Dimitri Daskalakis wrote:
quoted
Changes in v4:
- Nimrod and sashiko pointed out that not reporting DISABLED for
  tcp-data-split makes the feature look unsupported. This conflicted
  with my attempt to block users from setting DISABLED, locally sashiko
  found several scenarios where ring changes would get blocked.
Can you please share those scenarios?

I think blocking 'disabled' was the right direction. Allowing it doesn't
give users anything new, only a way to pin HW-GRO off, at the cost of
another dependency between the two knobs in mlx5e_fix_features().

We can still report 'disabled' when HW-GRO is supported but off.
I should have saved the review, I reapplied v3 and added an else to 
disable tcp-data-split if the merge type is not SHAMPO, here's what sashiko
reports.

 [Medium]
 Does this reject a value that get_ringparam reports?

 When HW-GRO is off, mlx5e_ethtool_get_ringparam() reports DISABLED. That
 is accurate, because SHAMPO is not in use. Before this patch,
 "ethtool -G eth0 tcp-data-split off" was accepted in that state. Now it
 fails with -EINVAL whatever the HW-GRO state is.

 This affects any netlink client that reads the rings config and writes
 the same attributes back. The whole request fails, including any ring
 size change sent with it. The changelog does not mention that existing
 users of "off" will break.

 A narrower fix may be possible. mlx5e_fix_features() could clear
 NETIF_F_GRO_HW when hds_config is DISABLED, like it already handles
 ENABLED. That would also fix the mis-reporting and keep DISABLED usable.

Netlink users that call get then set with the received parameters will
get a failure. This generally applies to users who save the current ring state,
and try to revert back (via netlink or ethtool). A subsequent sashiko run
highlighted the manual use case via ethtool.

I recall sashiko had more severe feedback, but those cases may have been
from the incorrect feature updates I fixed here [1]. The feedback is technically
correct, though I'm not sure how much it matters in practice. v4 implements the
narrower fix sashiko suggests, so it works for users who write back stored ring state.

[1] https://patchwork.kernel.org/project/netdevbpf/patch/20261017005401.1582755-1-dimitri.daskalakis1@gmail.com/
Thanks,

I'm not sure the write-back scenario is worth worrying about. The
drivers/net selftests for example always restore to 'auto' first rather
than a saved value, which seems like the safest way to reset this knob.

Still, there may be some merit to the point that blocking 'disabled'
would break setting it while HW-GRO is off, which works today, even if
it's not a very useful flow.

IMO both approaches are acceptable: v4 as is, or blocking 'disabled' as
in v3 plus reporting it when HW-GRO is supported but off. You can keep
my tag if you go with the latter.

Reviewed-by: Nimrod Oren <redacted>
Thread (8 messages) 8 messages, 5 authors, last reply 22h ago
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help