Re: [PATCH v3 3/3] usb: dwc3: xilinx: unwind ZynqMP platform init on probe failure and remove

[Thinh Nguyen <Thinh.Nguyen@synopsys.com>]

On Mon, Sep 28, 2026, Radhey Shyam Pandey wrote:
> dwc3_xlnx_init_zynqmp() leaves the controller out of reset with the PHY
> powered on, but nothing undoes that if probe fails later or when the
> driver is removed. The resets stay deasserted and the PHY stays on with
> no driver bound.
> 
> Register the teardown with devm_add_action_or_reset() once init has
> completed, so it runs on probe failure without the driver having to
> track how far init progressed. The teardown only undoes the reset cycle
> init performed, so it carries the same usb3-phy guard: with no PHY the
> resets were never asserted, and asserting them could break a USB3
> configuration that is live but missing from the device tree.
> 
> The resets have to be asserted while the clocks are still running, so
> both the probe error path and remove() run the teardown explicitly
> before gating the clocks rather than leaving it to devres, which would
> otherwise run it afterwards. devm_release_action() unlinks the action
> before calling it, so devres cannot run it a second time.
> 
> Running it from remove() also covers .shutdown, which shares
> dwc3_xlnx_remove(): device_shutdown() does not call
> devres_release_all(), so a teardown left to devres would never run
> there.
> 
> Fixes: 84770f028fab ("usb: dwc3: Add driver for Xilinx platforms")
> Cc: stable@vger.kernel.org
> Suggested-by: Philipp Zabel <p.zabel@pengutronix.de>
> Assisted-by: Claude:claude-opus-5
> Signed-off-by: Radhey Shyam Pandey <radhey.shyam.pandey@amd.com>
> ---
> Changes in v3:
> - Register the teardown with devm_add_action_or_reset() instead of a
>   pltfm_exit pointer and the usb_resets_released flag, per Philipp.
> - Guard the teardown's reset asserts on usb3_phy; v2 asserted
>   unconditionally, resetting a no-PHY board that init never asserted.
> - Replace v2's err_pltfm_exit label with dwc3_xlnx_release_teardown(),
>   called from both the probe error path and remove() so the resets are
>   unwound before the clocks are gated in either.
> - Explain in remove() why the teardown is not left to devres:
>   device_shutdown() does not call devres_release_all().
> - Clocks left unmanaged: devm_pm_runtime_enable()'s cleanup runs
>   runtime_suspend() before a clock devres node would release, which
>   double-disables on ZynqMP.
> 
> Changes in v2:
> - Reworked so the fix no longer depends on the platform-data cleanup.
>   v1 registered the teardown as plat->exit in struct dwc3_xlnx_platdata,
>   which is introduced by one of the cleanup patches; that made the fix
>   unbackportable.  It now uses a pltfm_exit pointer alongside the
>   existing pltfm_init in struct dwc3_xlnx, assigned by
>   dwc3_xlnx_init_zynqmp() once init has succeeded.  The follow-up
>   cleanup series folds both pointers into the platform data struct.
> - Made dwc3_xlnx_exit_zynqmp() idempotent by returning early when
>   usb_resets_released is clear, rather than guarding only the reset
>   assertions.  phy_power_off() and phy_exit() decrement their counts
>   unconditionally, so an unbalanced second call would underflow them.
> - Rewrote the commit message to describe the bug rather than the
>   implementation, since the callback it referred to no longer exists at
>   this point in the series.
> - Added Cc: stable.
> ---
>  drivers/usb/dwc3/dwc3-xilinx.c | 44 +++++++++++++++++++++++++++++++++-
>  1 file changed, 43 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/usb/dwc3/dwc3-xilinx.c b/drivers/usb/dwc3/dwc3-xilinx.c
> index d2315573b3d3..475bc19a1746 100644
> --- a/drivers/usb/dwc3/dwc3-xilinx.c
> +++ b/drivers/usb/dwc3/dwc3-xilinx.c
> @@ -112,6 +112,39 @@ static int dwc3_xlnx_init_versal(struct dwc3_xlnx *priv_data)
>  	return 0;
>  }
>  
> +static void dwc3_xlnx_zynqmp_teardown(void *data)
> +{
> +	struct dwc3_xlnx *priv_data = data;
> +
> +	phy_power_off(priv_data->usb3_phy);
> +
> +	/*
> +	 * Undo only the reset cycle init performed.  As on the init error
> +	 * path, the unconditional deasserts are not paired with an assert
> +	 * here: with no usb3-phy, resetting the core could break a USB3
> +	 * configuration that is live but missing from the device tree.
> +	 */
> +	if (priv_data->usb3_phy) {
> +		reset_control_assert(priv_data->usb_hibrst);
> +		reset_control_assert(priv_data->usb_crst);
> +		reset_control_assert(priv_data->usb_apbrst);
> +	}
> +
> +	phy_exit(priv_data->usb3_phy);
> +}
> +
> +/*
> + * Run the platform teardown explicitly, so the resets are asserted while the
> + * clocks are still running.  devm_release_action() unlinks the action before
> + * calling it, so devres will not run it a second time.
> + */
> +static void dwc3_xlnx_release_teardown(struct device *dev,
> +				       struct dwc3_xlnx *priv_data)
> +{
> +	if (devm_is_action_added(dev, dwc3_xlnx_zynqmp_teardown, priv_data))
> +		devm_release_action(dev, dwc3_xlnx_zynqmp_teardown, priv_data);
> +}
> +
>  static int dwc3_xlnx_init_zynqmp(struct dwc3_xlnx *priv_data)
>  {
>  	struct device		*dev = priv_data->dev;
> @@ -228,7 +261,8 @@ static int dwc3_xlnx_init_zynqmp(struct dwc3_xlnx *priv_data)
>  
>  	dwc3_xlnx_set_coherency(priv_data, XLNX_USB_TRAFFIC_ROUTE_CONFIG);
>  
> -	return 0;
> +	return devm_add_action_or_reset(dev, dwc3_xlnx_zynqmp_teardown,
> +					priv_data);
>  
>  err_phy_power_off:
>  	phy_power_off(priv_data->usb3_phy);
> @@ -355,6 +389,7 @@ static int dwc3_xlnx_probe(struct platform_device *pdev)
>  	pm_runtime_set_suspended(dev);
>  
>  err_clk_put:
> +	dwc3_xlnx_release_teardown(dev, priv_data);
>  	clk_bulk_disable_unprepare(priv_data->num_clocks, priv_data->clks);
>  
>  	return ret;
> @@ -367,6 +402,13 @@ static void dwc3_xlnx_remove(struct platform_device *pdev)
>  
>  	of_platform_depopulate(dev);
>  
> +	/*
> +	 * This is also what makes .shutdown behave the same as .remove:
> +	 * device_shutdown() does not call devres_release_all(), so a teardown
> +	 * left to devres would never run on the shutdown path.
> +	 */
> +	dwc3_xlnx_release_teardown(dev, priv_data);
> +
>  	clk_bulk_disable_unprepare(priv_data->num_clocks, priv_data->clks);
>  	priv_data->num_clocks = 0;
>  
> -- 
> 2.44.4
> 

All the inline comments are quite verbose. Unless necessary, I'd prefer
keeping the detailed explanations in the commit message. But it's fine.

Acked-by: Thinh Nguyen <Thinh.Nguyen@synopsys.com>

BR,
Thinh