[Barebox] Polling infrastructure causes reentrancy problem

[Christian Eggers]

Hi Sascha,

On Monday, 8 November 2021, 11:34:26 CET, Sascha Hauer wrote:
> Hi Christian,
> 
> On Fri, Nov 05, 2021 at 01:40:52PM +0100, Christian Eggers wrote:
> > Disclaimer: I stumbled over this on barebox 2020.01.
> > I haven't tried to reproduce this on the current version.
> 
> Please update to a recent barebox. The fastboot commands now run in a
> workqueue which itself only runs when we are outside any critical pathes
> in barebox.
thanks for the hint. I just upgraded to 2021.10.

> Particularly fastboot commands will only execute when you called
> command_slice_release() beforehand. You'll have to call
> command_slice_acquire() before entering any critical pathes like probing
> for USB devices.
I guess that more or less all commands can be critical because
is_timeout() is called from many places. As parallelism for fastboot
is not mandatory, I implemented it the following way: 

	struct usbgadget_funcs const fastboot_funcs= {
		.flags = USBGADGET_FASTBOOT,
		.fastboot_opts = "/dev/null(null)",
	};

	/*
	 *  Register fastboot gadget (will run in background)
	 */
	ret = usbgadget_register(&fastboot_funcs);
	if (ret)
		goto error_usb_gadget;

	/* fastboot runs within a work queue. Allow fastboot execution within
	 * the next 5 seconds...
	 */
	command_slice_release();
	mdelay(5000);
	command_slice_acquire();


BTW: I use a modified/stripped down version of the Android fastboot for booting
Barebox FIT images. Is there any interest integrating this into barebox/scripts/?
- more or less stand alone (about 15 files at all)
- currently USB-only
- Apache-2.0 / BSD-2-clause
- C++
- CMake (only 60 lines, so it can be ported to Kbuild)

regards
Christian