Re: [PATCH net] gve: clear flow rule cache state when freeing its buffers

From: Harshitha Ramamurthy

Date: Wed Nov 18 2026 - 18:42:28 EST


On Tue, Nov 17, 2026 at 6:35 PM Zhao Li <enderaoelyther@xxxxxxxxx> wrote:
>
> Freeing the flow-rule cache clears its pointers but retains the cached
> counts and synchronization flag. A successful rule-reset command
> invalidates the cache, but a failed reset does not. Recovery can then
> allocate zeroed buffers while retaining a nonzero count and synced flag,

Yes, this is possible when gve_adminq_reset_flow_rules() fails in the
gve_reset() path however....

> causing a later flow-rule query to return zeroed rules as if they were a
> synchronized cache.

The user visible case is narrower and will only show up on an ethtool
-n <dev> rule 0 with a stale value. Because gve_get_flow_rule_entry()
will end up re-querying for every other location.

Also, the more holistic fix is up in gve_adminq_configure_flow_rule():

diff --git a/drivers/net/ethernet/google/gve/gve_adminq.c
b/drivers/net/ethernet/google/gve/gve_adminq.c
index 8b215fca9c1b..ee9d5bc20e7e 100644
--- a/drivers/net/ethernet/google/gve/gve_adminq.c
+++ b/drivers/net/ethernet/google/gve/gve_adminq.c
@@ -1399,8 +1399,7 @@ gve_adminq_configure_flow_rule(struct gve_priv *priv,
sizeof(struct gve_adminq_configure_flow_rule),
flow_rule_cmd);

- if (!err)
- priv->flow_rules_cache.rules_cache_synced = false;
+ priv->flow_rules_cache.rules_cache_synced = false;

return err;
}

When a new flow rule configuration is _attempted_ as opposed to just
successful, we can treat the cache as out of sync because the command
could have failed for any reason and the device state is unknown. This
covers both the reset path and a failed add/del command.

I'd suggest a v2 with this diff instead of the memset describing the
problem a bit more in detail in the commit message. Feel free to use:
Suggested-by: Harshitha Ramamurthy <hramamurthy@xxxxxxxxxx>

pw-bot: cr

>
> Clear the complete cache state after freeing its buffers so reallocation
> always starts unsynchronized with zero cached entries.
>
> Found by code review. Compile-tested with W=1.
>
> AI assistance was used.
>
> Fixes: 57718b60df9b ("gve: Add flow steering adminq commands")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Zhao Li <enderaoelyther@xxxxxxxxx>
> ---
> No new W=1, sparse or smatch warnings.
>
> drivers/net/ethernet/google/gve/gve_main.c | 3 +--
> 1 file changed, 1 insertion(+), 2 deletions(-)
>
> diff --git a/drivers/net/ethernet/google/gve/gve_main.c b/drivers/net/ethernet/google/gve/gve_main.c
> index 35cf0f06a0f2..747f5229d7ef 100644
> --- a/drivers/net/ethernet/google/gve/gve_main.c
> +++ b/drivers/net/ethernet/google/gve/gve_main.c
> @@ -189,9 +189,8 @@ static void gve_free_flow_rule_caches(struct gve_priv *priv)
> struct gve_flow_rules_cache *flow_rules_cache = &priv->flow_rules_cache;
>
> kvfree(flow_rules_cache->rule_ids_cache);
> - flow_rules_cache->rule_ids_cache = NULL;
> kvfree(flow_rules_cache->rules_cache);
> - flow_rules_cache->rules_cache = NULL;
> + memset(flow_rules_cache, 0, sizeof(*flow_rules_cache));
> }
>
> static int gve_alloc_rss_config_cache(struct gve_priv *priv)
>
> base-commit: a2239d5dbbda3a61b6323df8836f605921a29f7f
> --
> 2.56.1
>