[PATCH v2] usb: dwc3: qcom: Defer VBUS override update during system suspend

[PATCH v2] usb: dwc3: qcom: Defer VBUS override update during system suspend

[Vladyslav Tymoshchuk]

dwc3_qcom_set_role_notifier() updates the VBUS override in QSCRATCH
synchronously, from the context of the usb_role_switch_set_role()
caller. The glue assumes that role switches only happen from a
freezable workqueue, but that holds for the dwc3 core's mode switch,
not for this notifier: the UCSI connector work, for one, runs on the
non-freezable system workqueue.

On SM8750, when the USB cable is unplugged while the system is in
s2idle, the UCSI connector work for the change runs before the devices
have resumed. It switches the role to USB_ROLE_NONE, and the notifier
accesses QSCRATCH while the glue clocks are still off, before
dwc3_qcom_pm_resume() turns them back on. The access does not complete
and the SoC resets. pm_runtime_resume_and_get() does not prevent this:
runtime PM is disabled at that point and the device is still marked
active, so it returns success without resuming anything.

While the glue is suspended for system sleep, only record the VBUS
override update and apply it from dwc3_qcom_pm_resume() once the
clocks are on. Apply it after the core has resumed, since resuming the
gadget enables the override and the role switch is the more recent
event. Serialize the notifier against the system suspend and resume
callbacks with a mutex, so that a role switch can't run while the
core or the glue is being suspended or resumed.

Fixes: 21188e8d6d75 ("usb: dwc3: qcom: Implement glue callbacks to facilitate runtime suspend")
Cc: stable@vger.kernel.org
Signed-off-by: Vladyslav Tymoshchuk <tvlad744@gmail.com>
---
Changes in v2:
- Hold role_lock across the whole of dwc3_qcom_pm_suspend() and
  dwc3_qcom_pm_resume(). In v1 a role switch could still run during
  dwc3_pm_suspend() or dwc3_pm_resume() and bypass the deferral, as
  the Sashiko review pointed out [4].
- Link to v1: https://lore.kernel.org/all/20261018143126.3978-1-tvlad744@gmail.com/

The review also asked about the direct_complete path, which skips
these callbacks. It is not taken for this driver: every
qcom,snps-dwc3 node in the tree has a power domain, and
genpd_prepare() never returns 1. Where it is taken, the role switch
is dropped as it is without this patch, and the hardware is not
touched.

Tested on SM8750 (OnePlus Pad 3; the board DT is not upstream yet),
v7.3-rc5 plus board patches, with the UCSI connector work still on
system_wq and without the pmic_glink RFC [5]. The test kernel also has
local ucsi_glink and qcom_battmgr changes that make a cable plug or
unplug wake the system from s2idle, similar to Jishnu Prakash's
ucsi_glink patch [3]:
- cable unplugged during s2idle: resumed in 2 of 2 runs; without this
  patch the SoC reset in all 6 earlier runs. The switch to
  USB_ROLE_NONE came 15 ms after the start of the resume, during the
  noirq phase, and was only recorded; dwc3_qcom_pm_resume() cleared
  the override about a second later.
- cable plugged in during s2idle, s2idle with the cable connected,
  and unplug/replug while awake: no change in behaviour.
- with a test-only msleep() under role_lock, so that a role switch
  comes in while the lock is held:
  - 3 s in dwc3_qcom_pm_resume() after dwc3_pm_resume(): the switch
    from a cable plug during s2idle waited for the lock and then set
    the override directly;
  - 20 s in dwc3_qcom_pm_suspend() before dwc3_pm_suspend(): the
    switch from an unplug waited for the lock and was deferred. The
    unplug aborted the suspend, and dwc3_qcom_pm_resume() applied the
    update. No reset.

I have not tested this patch without the wakeup changes.

The UCSI side is discussed in Abel Vesa's RFC to queue the connector
work on system_freezable_wq [1]. Since v1, I have also sent an RFC
that holds back the PMIC GLINK notifications over system sleep [5].
With it, an unplug during s2idle on SM8750 no longer reaches this
notifier. This patch is needed either way: the RFC does not hold back
a role switch that is already in flight when the system suspend
starts, and callers of usb_role_switch_set_role() outside pmic_glink,
TCPM's kthread worker for one, are not frozen.

Faisal Hassan's SA8255P series [2] also changes dwc3_qcom_pm_suspend()
and dwc3_qcom_pm_resume() (the wakeup argument becomes a pm_message_t);
whichever goes in second needs a small rebase.

I used an AI assistant (Claude) for the analysis and to draft this
patch and its description; I reviewed the result and ran the tests on
my device.

[1] https://lore.kernel.org/all/20250205-ucsi-schedule-conn-worker-on-freezable-wq-v1-1-107d1356b77b@linaro.org/
[2] https://lore.kernel.org/all/20261010053246.2991984-3-faisal.hassan@oss.qualcomm.com/
[3] https://lore.kernel.org/all/20260722-ucsi_glink_wakeup-v2-1-b43dee56df24@oss.qualcomm.com/
[4] https://lore.kernel.org/all/sashiko-outbox-176222@kernel.org/
[5] https://lore.kernel.org/all/20261022164151.592-1-tvlad744@gmail.com/

 drivers/usb/dwc3/dwc3-qcom.c | 61 +++++++++++++++++++++++++++++-------
 1 file changed, 50 insertions(+), 11 deletions(-)

diff --git a/drivers/usb/dwc3/dwc3-qcom.c b/drivers/usb/dwc3/dwc3-qcom.c
index ac68b4218b56..b22a72e3c1b9 100644
--- a/drivers/usb/dwc3/dwc3-qcom.c
+++ b/drivers/usb/dwc3/dwc3-qcom.c
@@ -10,6 +10,7 @@
 #include <linux/irq.h>
 #include <linux/of_clk.h>
 #include <linux/module.h>
+#include <linux/mutex.h>
 #include <linux/kernel.h>
 #include <linux/interconnect.h>
 #include <linux/platform_device.h>
@@ -85,6 +86,11 @@ struct dwc3_qcom {
 	struct icc_path		*icc_path_apps;
 
 	enum usb_role		current_role;
+
+	/* Serializes the role notifier against system suspend and resume */
+	struct mutex		role_lock;
+	bool			vbus_override_pending;
+	bool			vbus_override;
 };
 
 #define to_dwc3_qcom(d) container_of((d), struct dwc3_qcom, dwc)
@@ -561,27 +567,40 @@ static int dwc3_qcom_setup_irq(struct dwc3_qcom *qcom, struct platform_device *p
 static void dwc3_qcom_set_role_notifier(struct dwc3 *dwc, enum usb_role next_role)
 {
 	struct dwc3_qcom *qcom = to_dwc3_qcom(dwc);
+	bool enable;
+
+	guard(mutex)(&qcom->role_lock);
 
 	if (qcom->current_role == next_role)
 		return;
 
+	enable = qcom->current_role != USB_ROLE_DEVICE;
+
+	/*
+	 * Not every caller of usb_role_switch_set_role() is frozen during
+	 * system sleep, e.g. the UCSI connector work can run in the noirq
+	 * resume phase. The glue clocks are off then and runtime PM can't
+	 * turn them on, so leave QSCRATCH alone and let
+	 * dwc3_qcom_pm_resume() apply the update.
+	 */
+	if (qcom->pm_suspended) {
+		qcom->vbus_override_pending = true;
+		qcom->vbus_override = enable;
+		qcom->current_role = next_role;
+		return;
+	}
+
 	if (pm_runtime_resume_and_get(qcom->dev)) {
 		dev_dbg(qcom->dev, "Failed to resume device\n");
 		return;
 	}
 
-	if (qcom->current_role == USB_ROLE_DEVICE)
-		dwc3_qcom_vbus_override_enable(qcom, false);
-	else if (qcom->current_role != USB_ROLE_DEVICE)
-		dwc3_qcom_vbus_override_enable(qcom, true);
+	qcom->vbus_override_pending = false;
+	dwc3_qcom_vbus_override_enable(qcom, enable);
 
 	pm_runtime_mark_last_busy(qcom->dev);
 	pm_runtime_put_sync(qcom->dev);
 
-	/*
-	 * Current role changes via usb_role_switch_set_role callback protected
-	 * internally by mutex lock.
-	 */
 	qcom->current_role = next_role;
 }
 
@@ -624,6 +643,10 @@ static int dwc3_qcom_probe(struct platform_device *pdev)
 
 	qcom->dev = &pdev->dev;
 
+	ret = devm_mutex_init(dev, &qcom->role_lock);
+	if (ret)
+		return ret;
+
 	qcom->resets = devm_reset_control_array_get_optional_exclusive(dev);
 	if (IS_ERR(qcom->resets)) {
 		return dev_err_probe(&pdev->dev, PTR_ERR(qcom->resets),
@@ -752,6 +775,12 @@ static int dwc3_qcom_pm_suspend(struct device *dev)
 	bool wakeup = device_may_wakeup(dev);
 	int ret;
 
+	/*
+	 * A role switch that comes in now waits until the glue is suspended
+	 * and is then deferred, see dwc3_qcom_set_role_notifier().
+	 */
+	guard(mutex)(&qcom->role_lock);
+
 	ret = dwc3_pm_suspend(&qcom->dwc);
 	if (ret)
 		return ret;
@@ -772,6 +801,8 @@ static int dwc3_qcom_pm_resume(struct device *dev)
 	bool wakeup = device_may_wakeup(dev);
 	int ret;
 
+	guard(mutex)(&qcom->role_lock);
+
 	ret = dwc3_qcom_resume(qcom, wakeup);
 	if (ret)
 		return ret;
@@ -779,10 +810,18 @@ static int dwc3_qcom_pm_resume(struct device *dev)
 	qcom->pm_suspended = false;
 
 	ret = dwc3_pm_resume(&qcom->dwc);
-	if (ret)
-		return ret;
 
-	return 0;
+	/*
+	 * Apply a VBUS override update from a role switch that came in while
+	 * suspended. Do it after the core has resumed, as resuming the gadget
+	 * enables the override and the role switch is the more recent event.
+	 */
+	if (qcom->vbus_override_pending) {
+		dwc3_qcom_vbus_override_enable(qcom, qcom->vbus_override);
+		qcom->vbus_override_pending = false;
+	}
+
+	return ret;
 }
 
 static void dwc3_qcom_complete(struct device *dev)

base-commit: a90ee4305c4a5df72c11b31dacfdc76e00fcf78a
-- 
2.43.0