On Mon, Aug 23, 2021 at 4:27 PM Frank Rowand [off-list ref] wrote:

On 8/23/21 1:46 PM, Rob Herring wrote:

quoted

On Sun, Aug 22, 2021 at 11:01:15PM -0500, Frank Rowand wrote:

quoted

Hi Marc,

On 8/20/21 9:47 AM, Marc Zyngier wrote:

quoted

Trying to boot without SYSFS, but with OF_DYNAMIC quickly
results in a crash:

[    0.088460] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000070
[...]
[    0.103927] CPU: 1 PID: 1 Comm: swapper/0 Not tainted 5.14.0-rc3 #4179
[    0.105810] Hardware name: linux,dummy-virt (DT)
[  0.107147] pstate: 80000005 (Nzcv daif -PAN -UAO -TCO BTYPE=--)
[    0.108876] pc : kernfs_find_and_get_ns+0x3c/0x7c
[    0.110244] lr : kernfs_find_and_get_ns+0x3c/0x7c
[...]
[    0.134087] Call trace:
[    0.134800]  kernfs_find_and_get_ns+0x3c/0x7c
[    0.136054]  safe_name+0x4c/0xd0
[    0.136994]  __of_attach_node_sysfs+0xf8/0x124
[    0.138287]  of_core_init+0x90/0xfc
[    0.139296]  driver_init+0x30/0x4c
[    0.140283]  kernel_init_freeable+0x160/0x1b8
[    0.141543]  kernel_init+0x30/0x140
[    0.142561]  ret_from_fork+0x10/0x18

While not having sysfs isn't a very common option these days,
it is still expected that such configuration would work.

Paper over it by bailing out from __of_attach_node_sysfs() if
CONFIG_SYSFS isn't enabled.

CONFIG_SYSFS should be automatically selected when CONFIG_OF_DYNAMIC
is enabled, and it should not be possible to disable CONFIG_SYSFS
in this case.

That used to be true, but isn't now.

OK.  I agree with you, but when I investigated the original patch
email I came to a different conclusion because of the way that I
used make menuconfig to debug the situation.

It is true when I start with a .config created from 'make qcom_defconfig',
then select OF_UNITTEST, which is the only way I can see OF_DYNAMIC.  It
is the "if OF_UNITTEST" that means SYSFS can not be disabled.

Not really. Disabling SYSFS has nothing to do with the DT code. It's
not super easy though. It required setting EXPERT and disabling
CONFIGFS_FS and things selecting it (PCIE endpoint and USB gadget).

If I start with the .config that Marc supplied, then make menuconfig
still does not show the OC_DYNAMIC option, but leaves it set since
it was already set.  In this case SYSFS remains disabled because
OF_UNITTEST is also disabled.

I don't see the relationship between SYSFS and OF_UNITTEST.

Using '/OF_DYNAMIC' from within make menuconfig, to get more info
about OF_DYNAMIC tells me that the prompt for OF_DYNAMIC is visible
if OF && OF_UNITTEST.  This is due to the "if OF_UNITTEST" in line 58
of drivers/of/Kconfig in the OF_DYNAMIC specification.

Thus I can't figure out how to use make menuconfig to set OF_DYNAMIC
without setting OF_UNITTEST.  I tried setting OF_UNITTEST, then setting
OF_DYNAMIC, saving the changes, then another make menuconfig to
disable OF_UNITTEST, which then has the side effect of unsetting
OF_DYNAMIC.

Selecting OF_OVERLAY also.

Would you accept a patch that removes the "if OF_UNITTEST" from
the Kconfig entry for OF_DYNAMIC?

I guess. The purpose of making it visible was for compile testing, so
maybe make it 'if COMPILE_TEST' instead?


Looking at occurrences of CONFIG_OF_DYNAMIC, there's more than I'd
like. This for example is a common pattern:

drivers/spi/spi.c:      if (IS_ENABLED(CONFIG_OF_DYNAMIC))
drivers/spi/spi.c-
WARN_ON(of_reconfig_notifier_register(&spi_of_notifier));

Really, of_reconfig_notifier_register() should just return 0 if
!IS_ENABLED(CONFIG_OF_DYNAMIC).

Rob