Re: [PATCH v10 0/6] Re-introduce TX FIFO resize for larger EP bursting

From: Pavel Hofman
Date: Tue Aug 24 2021 - 01:39:51 EST




Dne 24. 08. 21 v 0:50 Thinh Nguyen napsal(a):
Pavel Hofman wrote:


Dne 23. 08. 21 v 17:21 Andy Shevchenko napsal(a):
On Mon, Aug 23, 2021 at 5:59 PM Pavel Hofman
<pavel.hofman@xxxxxxxxxxx> wrote:
Dne 22. 08. 21 v 21:43 Ferry Toth napsal(a):
Op 19-08-2021 om 23:04 schreef Pavel Hofman:
Dne 19. 08. 21 v 22:10 Ferry Toth napsal(a):
Op 19-08-2021 om 09:51 schreef Pavel Hofman:
Dne 18. 08. 21 v 21:07 Ferry Toth napsal(a):
Op 18-08-2021 om 00:00 schreef Ferry Toth:

...

So, where do we go from here?

I know the patches have been tested on dwc2 (by me and others).  I
do not know if Ruslan or Jerome tested them on dwc3 but probably
not. Ruslan has talked about RPi (my case too) and BeagleboneBlack,
both with dwc2. Perhaps the dwc2 behaves a bit differently than
dwc3?

The patches add a new EP-IN for async feedback. I am sorry I have
not followed your long thread (it started as unrelated to uac). Does
the problem appear with f_uac1 or f_uac2? Please how have you
reached the above problem?

I'm sorry too. I first believed the issue was related to the patch
mentioned in the subject line.

The problem appaers with f_uac2. I bost Edison_Arduino board in host
mode (there is a switch allowing to select host/device mode). When
flipping the switch to device mode udev run a script:
But as I am using configfs (excerpt follows) and just disabling the
last 2 line resolves the issue, I'm guessing uac2 is the issue. Or
exceeding the available resources.

# Create directory structure
mkdir "${GADGET_BASE_DIR}"
cd "${GADGET_BASE_DIR}"
mkdir -p configs/c.1/strings/0x409
mkdir -p strings/0x409

# Serial device
mkdir functions/gser.usb0
ln -s functions/gser.usb0 configs/c.1/
###

# Ethernet device
mkdir functions/eem.usb0
echo "${DEV_ETH_ADDR}" > functions/eem.usb0/dev_addr
echo "${HOST_ETH_ADDR}" > functions/eem.usb0/host_addr
ln -s functions/eem.usb0 configs/c.1/

# Mass Storage device
mkdir functions/mass_storage.usb0
echo 1 > functions/mass_storage.usb0/stall
echo 0 > functions/mass_storage.usb0/lun.0/cdrom
echo 0 > functions/mass_storage.usb0/lun.0/ro
echo 0 > functions/mass_storage.usb0/lun.0/nofua
echo "${USBDISK}" > functions/mass_storage.usb0/lun.0/file
ln -s functions/mass_storage.usb0 configs/c.1/

# UAC2 device
mkdir functions/uac2.usb0
ln -s functions/uac2.usb0 configs/c.1
....

As you say, could perhaps the reason be that the extra EP-IN added in
those patches (previously 1, now 2 with the default config you use)
exceeds your EP-IN max count or available fifos somehow?  You have a
number of functions initialized. If you change the load order of the
functions, do you get the error later with a different function? Just
guessing...

You should be able to switch the default async EP-OUT (which
configures the new feedback EP-IN ) to adaptive EP-OUT (which requires
no feedback EP) with c_sync=8 parameter of f_uac2.

https://urldefense.com/v3/__https://elixir.bootlin.com/linux/v5.14-rc6/source/drivers/usb/gadget/function/f_uac2.c*L47__;Iw!!A4F2R9G_pg!LBySrM_zgMGV0x-zZ7nSrs54yJw1GlnpUVUVxdQE8PeszSMZ6OkFX8lSoigwRbWQzLcU$

https://urldefense.com/v3/__https://elixir.bootlin.com/linux/v5.14-rc6/source/drivers/usb/gadget/function/f_uac2.c*L830__;Iw!!A4F2R9G_pg!LBySrM_zgMGV0x-zZ7nSrs54yJw1GlnpUVUVxdQE8PeszSMZ6OkFX8lSoigwRfP5TdZC$

https://urldefense.com/v3/__https://elixir.bootlin.com/linux/v5.14-rc6/source/include/uapi/linux/usb/ch9.h*L453__;Iw!!A4F2R9G_pg!LBySrM_zgMGV0x-zZ7nSrs54yJw1GlnpUVUVxdQE8PeszSMZ6OkFX8lSoigwRejzMbWO$

Does that fix the problem?

Not sure how to do that. Do you mean the module should have a parameter
called c_sync? `modinfo` list no parameters for usb_f_uac2.

Those are configfs params, not available in modinfo.

I checked and the value is "adaptive"
https://urldefense.com/v3/__https://elixir.bootlin.com/linux/v5.14-rc7/source/drivers/usb/gadget/function/f_uac2.c*L1312__;Iw!!A4F2R9G_pg!LBySrM_zgMGV0x-zZ7nSrs54yJw1GlnpUVUVxdQE8PeszSMZ6OkFX8lSoigwRTETcbsN$


In your configfs script:

Kernel shouldn't crash with any available set of configuration
parameters, right? So, even if the parameter would fix the crash the
series is buggy and has to be reverted in my opinion.

Sure, no problem with reverting. I am just trying to start up some
troubleshooting. A resource exhaustion was mentioned here, that's why I
suggested a way to test the patch with the original number of endpoints
allocated. I do not get any crashes on my setup which uses fewer gadget
functions. That's why I asked what happens if the functions load order
is changed. I am afraid this thread is so complex that the actual
problem has been burried in the history.


As I pointed out previously, the crash is probably because of f_uac2
prematurely freeing feedback request before its completion.
usb_ep_dequeue() is asynchronous. dwc2() may treat it as a synchronous
call so you didn't get a crash.

Thanks for your hint, it greatly helps.


I'm not sure how easy it is for you to obtain/test a device with dwc3,
but it would be great to also have it as part of your testing suite. :)

Can you recommend a reasonably priced device with viable kernel updates for the testing? Optionally with SS which you mentioned last time? Thanks.

Best regards,

Pavel.