[PATCH v3 0/8] gve: precursor control plane changes before bare-metal support

[Stephen Hemminger]

On Fri,  9 Oct 2026 17:59:32 -0700
Joshua Washington <joshwash at google.com> wrote:

> This patch series includes a number of precursor changes to the GVE
> ethdev driver layer in preparation for introducing support for a new
> GVE Mailbox control plane (an alternative to the existing AdminQ control
> plane).
> 
> The most major change in this series is the introduction of a new
> control_ops struct that will be used by both AQ and Mailbox for various
> operations which may need to communicate with the device. There are also
> some changes to how RSS and timestamping are handled in ethdev due to
> slight feature differences between the control planes.

Applied to next-net. I fixed the commit messages based on AI review
for patch 6 and 8. The last one is not really needed.

Subject: Re: [PATCH v3 0/8] net/gve control ops rework

Applies cleanly to main; every commit builds with -Dwerror=true.

v3 addresses the v2 errors and warnings. The remaining items are
commit message wording; fix them if there is a v4, otherwise they can
be adjusted on apply.

Info

[PATCH v3 6/8] net/gve: update RSS config just after successful
programming

1. Body starts with "Moving this" with no antecedent. Say what moves:
   "Move gve_update_priv_rss_config() into gve_adminq_configure_rss()
   so the cache is only updated after the device accepts the config."

[PATCH v3 7/8] net/gve: add RSS cache boolean flag

2. The commit message says the cache is invalidated when a configure
   message is sent if queries are supported. No code does that; the
   only place rss_cache_dirty is set true is gve_init_priv(). Describe
   it as a requirement on the query backend. When that backend lands,
   note that gve_dev_configure() returns 0 without regenerating the
   RETA if the query fails with indir set.

[PATCH v3 8/8] net/gve: refactor timestamp support to clock read type

3. On GQI, rte_eth_read_clock() now returns -EOPNOTSUPP; before this
   patch it worked whenever the device offered the timestamp option.
   Mention it alongside the memzone and thread note.

Reviewed-by: Stephen Hemminger <stephen at networkplumber.org>