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

[Vladyslav Tymoshchuk <tvlad744@gmail.com>]

On Thu, Nov 05, 2026 at 12:02:46AM +0000, Thinh Nguyen wrote:
> Hi Vladyslav,
>
> Could you test this change instead to see if it resolves the issue?

It does. OnePlus Pad 3 (SM8750, UCSI over pmic_glink), v7.3-rc5 plus
board patches, with my dwc3-qcom v2 reverted and nothing else holding
back pmic_glink notifications during suspend. I added debug prints to
dwc3_usb_role_switch_set(), to __dwc3_set_mode() before the glue call,
and to dwc3_qcom_set_role_notifier(), with the device's PM state.

The same kernel without your change resets the SoC when the cable is
unplugged in s2idle. The last lines in the ramoops console:

  PM: suspend-to-idle
  role_switch_set ->0 (pm suspended 1 late 1 noirq 1)
  qcom notifier: role 2->0, vbus override 0 (pm suspended 1 noirq 1)

With your change there was no reset. Each suspend run was 150 s of
s2idle ended by an RTC alarm, with the cable changed by hand during it:

  unplug                  2 runs
  plug in                 1 run
  unplug, plug back in    1 run
  cable left plugged in   1 run

plus unplug and plug in while awake. In the suspend runs every role
switch call came with dwc3 in the noirq phase and only stored the role.
The glue was called once, with the last role, when the freezable
workqueue was thawed:

  PM: suspend-to-idle
  role_switch_set ->0 (pm suspended 1 late 1 noirq 1)
  PM: resume of devices complete after 108.388 msecs
  Restarting tasks: Starting
  drd_work passes role 0 to the glue
  qcom notifier: role 2->0, vbus override 0 (pm suspended 0 noirq 0)
  Restarting tasks: Done

Unplug and plug in during one suspend came down to a single DEVICE
call, which dwc3-qcom skips since its role did not change. After every
resume with the cable in, the NCM gadget was configured again within
5 s. While awake, the glue still gets NONE and DEVICE although the dr
mode stays peripheral, so the call has to stay ahead of the early
returns in __dwc3_set_mode(), where you put it.

Host mode, with a USB stick on the port instead of the cable: plugging
and unplugging the stick while awake works as before, and so did two
suspends with the stick left in. With the stick unplugged in s2idle
(3 runs) the NONE call again came in noirq and was only stored, and
drd_work passed it on after thaw:

  PM: suspend-to-idle
  role_switch_set ->0 (pm suspended 1 late 1 noirq 1)
  PM: resume of devices complete after 2554.135 msecs
  Restarting tasks: Starting
  drd_work passes role 0 to the glue
  qcom notifier: role 1->0, vbus override 1 (pm suspended 0 noirq 0)
  xhci-hcd xhci-hcd.3.auto: remove, state 1
  Restarting tasks: Done

The host-mode suspends were on a kernel that also keeps the USB GDSCs
of gcc-sm8750 in retention. Without that the xHCI came back from
s2idle with an SMMU fault and "HC died" in all three runs I tried,
with the stick unplugged or left in. The dwc3 code was the same in
both kernels, so that is a separate problem. I have no result for
plugging the stick in during s2idle: the port did not detect the stick
every time, also while awake, and I have not found out why yet.

Two comments go stale with this change: dwc3_qcom_set_role_notifier()
says role changes are protected by the usb_role_switch mutex, while now
they are serialized by dwc->mutex in drd_work, and the @pre_set_role
kerneldoc could say that it is called from drd_work, so not during
system sleep.

Will you post it, or would you like me to send it as v3 in place of my
dwc3-qcom patch?

Tested-by: Vladyslav Tymoshchuk <tvlad744@gmail.com> # SM8750