[PATCH net] net: stmmac: hold runtime PM reference in setup_tc

[Lorenzo Bianconi]

> The qdisc offload callbacks invoked by stmmac_setup_tc() program
> MTL/MAC registers, but they can be reached while the interface is down,
> when stmmac_release() has dropped the runtime PM usage counter and the
> device may be suspended with its clocks gated. Accessing the registers
> in that state can trigger a bus error.
> 
> Hold a runtime PM reference while configuring the register-touching
> qdisc offloads (mqprio, cbs and taprio) so the device is active, and its
> clocks enabled, whenever the MTL/MAC registers are programmed.
> 
> The TC block callback stmmac_setup_tc_block_cb() programs the MTL/MAC
> registers as well, but it runs asynchronously from stmmac_setup_tc(),
> outside the runtime PM reference held there. Hold a runtime PM reference
> for the whole stmmac_setup_tc_block_cb() call as well, covering the
> cls_u32/cls_flower setup and the queue enable/disable accesses.
> 
> No reference is held for the TC_SETUP_BLOCK bookkeeping itself, the
> TC_QUERY_CAPS query or the tc-etf path, since none of them touch the
> registers synchronously. In particular the block bind/unbind must reach
> flow_block_cb_setup_simple() even when the device is suspended, so the
> driver never leaves a stale flow_block_cb on its block list.
> 
> Fixes: 1f705bc61aee ("net: stmmac: Add support for CBS QDISC")
> Fixes: 4dbbe8dde848 ("net: stmmac: Add support for U32 TC filter using Flexible RX Parser")
> Signed-off-by: Lorenzo Bianconi <lorenzo.bianconi at oss.qualcomm.com>
> ---
>  drivers/net/ethernet/stmicro/stmmac/stmmac_main.c | 34 ++++++++++++++++++++---
>  1 file changed, 30 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> index b2b7d0242dd3..4baf40fb01dc 100644
> --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> @@ -6392,9 +6392,13 @@ static int stmmac_setup_tc_block_cb(enum tc_setup_type type, void *type_data,
>  				    void *cb_priv)
>  {
>  	struct stmmac_priv *priv = cb_priv;
> -	int ret = -EOPNOTSUPP;
> +	int ret;
>  
>  	if (!tc_cls_can_offload_and_chain0(priv->dev, type_data))
> +		return -EOPNOTSUPP;
> +
> +	ret = pm_runtime_resume_and_get(priv->device);
> +	if (ret < 0)
>  		return ret;
>  
>  	__stmmac_disable_all_queues(priv);
> @@ -6411,6 +6415,8 @@ static int stmmac_setup_tc_block_cb(enum tc_setup_type type, void *type_data,
>  	}
>  
>  	stmmac_enable_all_queues(priv);
> +	pm_runtime_put(priv->device);

commenting on sashiko's report:
https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260824-stmmac-setup-tc-enable-pm-v1-1-45172d241a4b%40oss.qualcomm.com

- Dropping the -EOPNOTSUPP initializer of ret changes what this callback
  returns for tc_setup_type values the driver does not handle. ret is now
  first assigned by pm_runtime_resume_and_get(), which returns exactly 0 on
  success
  - I will fix it in v2

> +
>  	return ret;
>  }
>  
> @@ -6420,26 +6426,46 @@ static int stmmac_setup_tc(struct net_device *ndev, enum tc_setup_type type,
>  			   void *type_data)
>  {
>  	struct stmmac_priv *priv = netdev_priv(ndev);
> +	int ret;
>  
>  	switch (type) {
>  	case TC_QUERY_CAPS:
>  		return stmmac_tc_query_caps(priv, priv, type_data);
>  	case TC_SETUP_QDISC_MQPRIO:
> -		return stmmac_tc_setup_mqprio(priv, priv, type_data);
> +		ret = pm_runtime_resume_and_get(priv->device);
> +		if (ret < 0)
> +			return ret;
> +
> +		ret = stmmac_tc_setup_mqprio(priv, priv, type_data);
> +		break;
>  	case TC_SETUP_BLOCK:
>  		return flow_block_cb_setup_simple(type_data,
>  						  &stmmac_block_cb_list,
>  						  stmmac_setup_tc_block_cb,
>  						  priv, priv, true);
>  	case TC_SETUP_QDISC_CBS:
> -		return stmmac_tc_setup_cbs(priv, priv, type_data);
> +		ret = pm_runtime_resume_and_get(priv->device);
> +		if (ret < 0)
> +			return ret;
> +
> +		ret = stmmac_tc_setup_cbs(priv, priv, type_data);
> +		break;
>  	case TC_SETUP_QDISC_TAPRIO:
> -		return stmmac_tc_setup_taprio(priv, priv, type_data);
> +		ret = pm_runtime_resume_and_get(priv->device);
> +		if (ret < 0)
> +			return ret;
> +
> +		ret = stmmac_tc_setup_taprio(priv, priv, type_data);

- This is a pre-existing issue, but the patch now explicitly sanctions
  running taprio (and cls_u32/cls_flower in the block callback) with only
  the bus/CSR clocks resumed, without the rest of the hardware state those
  sequences depend on
  - This is fixed in the following patch:
    https://lore.kernel.org/netdev/20260825-stmmac-est-reapply-after-open-v1-1-dfa80735e0a1@oss.qualcomm.com/

Regards,
Lorenzo

> +		break;
>  	case TC_SETUP_QDISC_ETF:
>  		return stmmac_tc_setup_etf(priv, priv, type_data);
>  	default:
>  		return -EOPNOTSUPP;
>  	}
> +
> +	pm_runtime_put(priv->device);
> +
> +	return ret;
>  }
>  
>  static u16 stmmac_select_queue(struct net_device *dev, struct sk_buff *skb,
> 
> ---
> base-commit: 7cbfb180945ce529608e4d4e24a6d483699fab1e
> change-id: 20260824-stmmac-setup-tc-enable-pm-149aa563d797
> 
> Best regards,
> -- 
> Lorenzo Bianconi <lorenzo.bianconi at oss.qualcomm.com>
> 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 228 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20260827/605a1d4f/attachment.sig>