Date:   Fri, 11 Nov 2022 01:31:02 +0000
From:   Thinh Nguyen <Thinh.Nguyen@...opsys.com>
To:     Ferry Toth <fntoth@...il.com>
CC:     Thinh Nguyen <Thinh.Nguyen@...opsys.com>,
        "linux-usb@...r.kernel.org" <linux-usb@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Heikki Krogerus <heikki.krogerus@...ux.intel.com>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        Sean Anderson <sean.anderson@...o.com>,
        Liu Shixin <liushixin2@...wei.com>,
        Andrey Smirnov <andrew.smirnov@...il.com>,
        Andy Shevchenko <andriy.shevchenko@...ux.intel.com>
Subject: Re: [PATCH v1 1/1] usb: ulpi: defer ulpi_register on ulpi_read_id
 timeout

On Thu, Nov 10, 2022, Ferry Toth wrote:
> Hi
> 
> Op 10-11-2022 om 13:45 schreef Ferry Toth:
> > (sorry sent html with previous attempt)
> > 
> > On 10-11-2022 01:06, Thinh Nguyen wrote:
> > > Hi Ferry,
> > > 
> > > On Wed, Nov 09, 2022, Ferry Toth wrote:
> > > > Since commit 0f010171
> > > > Dual Role support on Intel Merrifield platform broke due to rearranging
> > > > the call to dwc3_get_extcon().
> > > > 
> > > > It appears to be caused by ulpi_read_id() on the first test
> > > > write failing
> > > > with -ETIMEDOUT. Currently ulpi_read_id() expects to discover
> > > > the phy via
> > > > DT when the test write fails and returns 0 in that case even if
> > > > DT does not
> > > > provide the phy. Due to the timeout being masked dwc3 probe continues by
> > > > calling dwc3_core_soft_reset() followed by dwc3_get_extcon()
> > > > which happens
> > > > to return -EPROBE_DEFER. On deferred probe ulpi_read_id()
> > > > finally succeeds.
> > > > 
> > > > This patch changes ulpi_read_id() to return -ETIMEDOUT when it
> > > > occurs and
> > > > catches the error in dwc3_core_init(). It handles the error by calling
> > > > dwc3_core_soft_reset() after which it requests -EPROBE_DEFER. On
> > > > deferred
> > > > probe ulpi_read_id() again succeeds.
> > > > 
> > > > Signed-off-by: Ferry Toth<ftoth@...londelft.nl>
> > > > ---
> > > >   drivers/usb/common/ulpi.c | 5 +++--
> > > >   drivers/usb/dwc3/core.c   | 5 ++++-
> > > >   2 files changed, 7 insertions(+), 3 deletions(-)
> > > > 
> > > Can you split the dwc3 change and ulpi change to separate patches?
> > 
> > Thanks for your comments.
> > 
> > I will send v2
> > 
> > > > diff --git a/drivers/usb/common/ulpi.c b/drivers/usb/common/ulpi.c
> > > > index d7c8461976ce..d8f22bc2f9d0 100644
> > > > --- a/drivers/usb/common/ulpi.c
> > > > +++ b/drivers/usb/common/ulpi.c
> > > > @@ -206,8 +206,9 @@ static int ulpi_read_id(struct ulpi *ulpi)
> > > >       /* Test the interface */
> > > >       ret = ulpi_write(ulpi, ULPI_SCRATCH, 0xaa);
> > > > -    if (ret < 0)
> > > > -        goto err;
> > > > +    if (ret < 0) {
> > > > +        return ret;
> > > > +    }
> > > >       ret = ulpi_read(ulpi, ULPI_SCRATCH);
> > > >       if (ret < 0)
> > > > diff --git a/drivers/usb/dwc3/core.c b/drivers/usb/dwc3/core.c
> > > > index 648f1c570021..e293ef70039b 100644
> > > > --- a/drivers/usb/dwc3/core.c
> > > > +++ b/drivers/usb/dwc3/core.c
> > > > @@ -1106,8 +1106,11 @@ static int dwc3_core_init(struct dwc3 *dwc)
> > > >       if (!dwc->ulpi_ready) {
> > > >           ret = dwc3_core_ulpi_init(dwc);
> > > > -        if (ret)
> > > > +        if (ret) {
> > > > +            dwc3_core_soft_reset(dwc);
> > > We shouldn't need to do soft reset here. The controller shouldn't be at
> > > a bad/incorrect state at this point to warrant a soft-reset. There will
> > > be a soft-reset when it goes through the initialization again.
> > 
> > It doesn't go through the initialization again unless we set
> > -EPROBE_DEFER. And when we make ulpi_read_id() return -EPROBE_DEFER it
> > will goto err0 here, so skips dwc3_core_soft_reset.
> > 
> > Do you mean you prefer something like:
> > 
> > if (ret) {
> > 
> >      if (ret == -ETIMEDOUT) ret = -EPROBE_DEFER;
> > 
> >      else goto err0;

Why "else"? But I saw you remove that in the new patch.

> > 
> > }
> 
> I just tested, and calling dwc3_core_soft_reset() proves to be necessary as
> we need to goto err0 directly after. Else ret is overwritten and
> -EPROBE_DEFER lost.

Looks like there's a strange dependency problem here.
 * The setup needs a soft-reset before ulpi registration
 * The ulpi registration needs to go before the phy initialization
 * The soft-reset should be called after the phy initialization

I can't explain the actual issue here, and we can't debug further
because to look into it further would require looking at internal
signals.

This soft-reset and -EPROBE_DEFER seems to be a workaround to this
dependency problem. Instead of using -EPROBE_DEFER, can you do this:

diff --git a/drivers/usb/dwc3/core.c b/drivers/usb/dwc3/core.c
index 2f0a9679686f..5a1aaf3741ec 100644
--- a/drivers/usb/dwc3/core.c
+++ b/drivers/usb/dwc3/core.c
@@ -1097,6 +1097,8 @@ static int dwc3_core_init(struct dwc3 *dwc)
                goto err0;
 
        if (!dwc->ulpi_ready) {
+               /* Add comment */
+               dwc3_core_soft_reset(dwc);
                ret = dwc3_core_ulpi_init(dwc);
                if (ret)
                        goto err0;


> 
> > > > +            ret = -EPROBE_DEFER;
> > > We shouldn't automatically set every error status to correspond to
> > > -EPROBE_DEFER. Check only the approapriate error codes (-ETIMEDOUT +
> > > any other?).
> > Other could be -ENOMEM. I think no need to do any new handling for that.

There's also -EINVAL and some others too. Regardless, if it's -ENOMEM,
then it should not return -EPROBE_DEFER.

> > > >               goto err0;
> > > > +        }
> > > >           dwc->ulpi_ready = true;
> > > >       }
> > > > 

Thanks,
Thinh

Powered by blists - more mailing lists