Hi Maxime,

On 01-10-18 17:57, Maxime Ripard wrote:

Hi Hans,

It's been a while :)

Yes it has.

On Sun, Sep 30, 2018 at 05:09:27PM +0200, Hans de Goede wrote:

quoted

While doing some brcmfmac driver work I needed to test this also on some
devicetree based boards. So I fired up the good old Cubietruck and when
that would not work a Banana Pro.

With an unmodified 4.17 kernel both boards intermittently would come up
with non working wifi with the following errors:

  brcmfmac: brcmf_sdio_bus_rxctl: resumed on timeout
  brcmfmac: brcmf_bus_started: failed: -110
  brcmfmac: brcmf_attach: dongle is not responding: err=-110
  brcmfmac: brcmf_sdio_firmware_callback: brcmf_attach failed

They would come up this way more often then with actual working wifi,
once this problem happens it seems to require a power-cycle to fix.
Once things work one can safely reboot without hitting the issue.

I've found that disabling OOB interrupts fixes this. This really is more
of a workaround then a proper fix, but it makes the wifi reliable again
and it does not have much of a downside.

Using an OOB IRQ instead of the sdio-IRQ mechanism is mostly important to
allow the MMC controller to go into runtime-suspend which is not really an
issue on these boards since they are (usually) not battery powered.

I've looked at recent brcmfmac and mmc-core changes which may explain this
and I've not found anything. So the most likely culprit is the A20 external
interrupt handling e.g. perhaps it is set to edge instead of level? Either
way I do not have time to further investigate this.

BugLink: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=908438
Signed-off-by: Hans de Goede <redacted>

Unfortunately, I'd really prefer if we were fixing this properly.

I understand, but I really already have spend more time on this then
I wanted to spend on it. If someone has an idea how to fix this I can
*maybe* run some quick tests, but that is it.

You
were saying that the regression has been introduced between 4.17 and
4.18, have you been able to bisect which commit was actually creating
this regression?

Erm, no what I was trying to say is that I can reproduce the issue
with 4.17, which is also the version mentioned in the Debian bug about
the same problem. I've not tried older kernels then 4.17 so I do not
know when this problem got introduced.

As you suggested, one reason could be the runtime_pm
introduction. This can be pretty easily tested by adding a
pm_runtime_get_sync call in the probe.

I assume you mean the runtime pm support in the sunxi mmc controller
driver, right ? When was that first introduced?

I myself actually suspect the external irq handling code, but I
guess that the runtime pm support code also is a likely cause
of this.

Regards,

Hans