Re: [PATCH] xhci: Fix HIK camera control timeout on Etron hosts

[Kuangyi Chiang]

Hi,

Thanks for the review.

Michal Pecio <michal.pecio@xxxxxxxxx> 於 2026年9月14日週一 下午7:09寫道：
>
> On Mon, 14 Sep 2026 09:30:33 +0800, Kuangyi Chiang wrote:
> > Control transfers to the HIK 1080P Camera (2bdf:0280) can time out when
> > the device is connected through an Etron xHCI host:
> >
> > [   48.363671] uvcvideo: Found UVC 1.00 device HIK 1080P Camera (2bdf:0280)
> > [   53.465662] uvcvideo: Failed to set UVC probe control : -110 (exp. 26).
> > [   58.585653] uvcvideo: Failed to query (129) UVC probe control : -110 (exp. 26).
> > [   58.592967] uvcvideo: Failed to initialize the device (-5).
> >
> > Testing shows that at least 125 us between control transfers is required
> > to avoid the timeout. Limit the workaround to the affected Etron host and
> > USB device combination to avoid changing control-transfer timing for other
> > devices.
>
> Any idea why it happens or why the delay helps?

We don't know the root cause yet, but we did some more testing.

We tested delays of 50us, 125us, and 200us. The issue was still
reproducible with 50us, while 125us and 200us passed our test.

Our pass criterion is 30 successful plug/unplug cycles on each port.
So 125us is the minimum delay we tested that consistently works.

We also tried USB_QUIRK_DELAY_CTRL_MSG, but its 200ms delay is too
long and causes UVC initialization to fail.

> How many other HCs did you test this camera with?

The camera works with the EHCI host we tested without any delay.
We can also reproduce a similar issue with an ASMedia ASM1142 xHCI
host (1b21:1242). This patch, however, is limited to the Etron/HIK
combination.

> Does it help to reconnect and try again?

No. The issue is still reproducible after reconnecting the camera.

>
> FWIW, I tried a few USB 2.0 cameras with my Etron and they mostly work,
> except one. It passes probe just fine, but fails on streaming attempt.
> Adding 125us delay doesn't help, but 1000 works sometimes and 4000 even
> more often.
>
> Probably a HW bug. Honestly, I never use Etron for anything serious.
>
> > Perform the delay before taking xhci->lock so the existing state
> > validation and TD queueing remain in the same critical section. Use
> > udelay() because urb_enqueue() may be called from atomic context and
> > cannot use a sleeping delay.
>
> Actually, you would need to do that under lock if you want to ensure
> that multiple callers can't submit URBs closely together. But that's
> dodgy and we don't even know what exactly triggers it or whether this
> solution really is the best one.

Agreed. For the UVC control transfers involved here, the next transfer
is submitted only after the previous one completes, so they are
serialized.

The delay before xhci->lock does not prevent other callers from
submitting control URBs close together. However, this workaround is
limited to control transfers for the affected Etron/HIK combination.

Regards,
Kuangyi

>
> Regards,
> Michal