Re: [PATCH v8 3/5] usb: xhci: Add support for Renesas controller with memory

[Christian Lamparter]

Hello,

On Wednesday, 1 April 2020 14:57:48 CEST Vinod Koul wrote:
> On 26-03-20, 17:21, Vinod Koul wrote:
> > On 26-03-20, 13:29, Mathias Nyman wrote:
> > > On 23.3.2020 19.05, Vinod Koul wrote:
> > > > Some rensas controller like uPD720201 and uPD720202 need firmware to be
> > > > loaded. Add these devices in table and invoke renesas firmware loader
> > > > functions to check and load the firmware into device memory when
> > > > required.
> > > > 
> > > > Signed-off-by: Vinod Koul <vkoul@xxxxxxxxxx>
> > > > ---
> > > >  drivers/usb/host/xhci-pci-renesas.c |  1 +
> > > >  drivers/usb/host/xhci-pci.c         | 29 ++++++++++++++++++++++++++++-
> > > >  drivers/usb/host/xhci-pci.h         |  3 +++
> > > >  3 files changed, 32 insertions(+), 1 deletion(-)
> > > > 
> > > 
> > > It's unfortunate if firmware loading couldn't be initiated in a PCI fixup hook
> > > for this Renesas controller. What was the reason it failed?
> > > 
> > > Nicolas Saenz Julienne just submitted a solution like that for Raspberry Pi 4
> > > where firmware loading is initiated in pci-quirks.c quirk_usb_early_handoff()
> > > 
> > > https://lore.kernel.org/lkml/20200324182812.20420-1-nsaenzjulienne@xxxxxxx
> > > 
> > > Is he doing something different than what was done for the Renesas controller?
> > 
> > I tried and everytime ended up not getting firmware. Though I did not
> > investigate a lot. Christian seemed to have tested sometime back as
> > well.
> > 
> > Another problem is that we dont get driver_data in the quirk and there
> > didnt seem a way to find the firmware name.
> > 
> > > > diff --git a/drivers/usb/host/xhci-pci-renesas.c b/drivers/usb/host/xhci-pci-renesas.c
> > > > index c588277ac9b8..d413d53df94b 100644
> > > > --- a/drivers/usb/host/xhci-pci-renesas.c
> > > > +++ b/drivers/usb/host/xhci-pci-renesas.c
> > > > @@ -336,6 +336,7 @@ static void renesas_fw_callback(const struct firmware *fw,
> > > >  		goto cleanup;
> > > >  	}
> > > >  
> > > > +	xhci_pci_probe(pdev, ctx->id);
> > > >  	return;
> > > 
> > > I haven't looked into this but instead of calling xhci_pci_probe() here in the async fw
> > > loading callback could we just return -EPROBE_DEFER until firmware is loaded when
> > > xhci_pci_probe() is originally called?
> > 
> > Hmm, initially my thinking was how to tell device core to probe again,
> > and then digging up I saw wait_for_device_probe() which can be used, let
> > me try that
> 
> Sorry to report back that it doesn't work as planned :(
> 
> I modified the code to invoke the request_firmware_nowait() which will load
> the firmware and provide the firmware in callback. Meanwhile return -EPROBE_DEFER.
> 
> After a bit, the core invokes the driver probe again and we hit the
> roadblock. The request_firmware uses devres and allocates resources for
> loading the firmware. The problem is that device core checks for this:
> 
> bus: 'pci': really_probe: probing driver xhci_hcd_pci with device 0000:01:00.0
> pci 0000:01:00.0: Resources present before probing
> 
> And here the probe fails. In some cases the firmware_callback finishes
> before this and we can probe again, but that is not very reliable.
> 
> I tested another way to use request_firmware() (sync version) and then
> load the firmware in probe and load. The request is done only for
> renesas devices if they dont have firmware already running.
> So rest of the devices wont have any impact.
> 
> Now should we continue this way in the patchset or move to sync version.
> Am okay either way.

Just a word of caution.

The problem with the usage of "sync" request_firmware in drivers is that if the
code is built into the kernel the request_firmware() could be called before the
(root) filesystem on which the firmware resides is ready.... So this will get
weird during boot because what is the sync request_firmware() going to do? From what
I know, this is why the funny _async firmware request APIs are even a thing...

(I took a quick peek into the RPI 4 code, but unlike this code it seems to fetch
from nvmem/eeprom, is this right? I had a tons-of-fun dealing with caldata and
firmware on UBIFS in UBI Volumes. So I'm prepared to test this cases. :D )

(Another possibility would be to use request_firmware_direct() here.
Though, I don't know if it would be considered API Abuse to -EPROBE_DEFER
on errors of request_firmware_direct() in order to wait for FSes to appear )

Regards,
Christian