Re: Linux 5.12-rc5

[Frank Rowand <frowand.list@gmail.com>]

On 3/29/21 1:41 PM, Rob Herring wrote:
> n Mon, Mar 29, 2021 at 1:05 PM Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
>>
>> On Sun, Mar 28, 2021 at 7:07 PM Guenter Roeck <linux@roeck-us.net> wrote:
>>>
>>> This is not really a new problem. I enabled devicetree unit tests
>>> in the openrisc kernel and was rewarded with a crash.
>>> https://lore.kernel.org/lkml/20210327224116.69309-1-linux@roeck-us.net/
>>> has all the glorious details.
>>
>> Hmm.
>>
>> I'm not sure I love that patch.
>>
>> I don't think the patch is _wrong_ per se, but if that "require 8 byte
>> alignment" is a problem, then this seems to be papering over the issue
>> rather than fixing it.
>>
>> So your patch protects from a NULL pointer dereference, but the
>> underlying issue seems to be a regression, and the fix sounds like the
>> kernel shouldn't be so strict about alignment requirements.
> 
> In the interest of the DT unittests not panicking and halting boot, I
> think we should handle NULL pointer.

Agreed.

> 
>> I guess we could make ARCH_SLAB_MINALIGN be at least 8 (perhaps only
>> if the allocations is >= 8) but honestly, I don't think libfdt merits
>> making such a big change. Small allocations are actually not uncommon
>> in the kernel, and on 32-bit architectures I think 4-byte allocations
>> are normal.
>>
>> So I'd be inclined to just remove the new
>>
>>         /* The device tree must be at an 8-byte aligned address */
>>         if ((uintptr_t)fdt & 7)
>>                 return -FDT_ERR_ALIGNMENT;
>>
>> check in scripts/dtc/libfdt/fdt.c which I assume is the source of the
>> problem. Rob?
> 
> That is the source, but I'd rather not remove it as we try to avoid
> any modifications from upstream. And we've found a couple of cases of
> not following documented alignment requirements.

Agreed to not remove.  We can be properly aligned without changing
kmemdup().

> 
>> Your patch to then avoid the NULL pointer dereference seems to be then
>> an additional safety, but not fixing the actual regression.
> 
> I think the right fix is not using kmemdup which copies the unittest dtb.A

This is not the only place a kmemdup() is used by overlays.

I'll create a patch this week to fix all of the kmemdup() locations and add
the null pointer check.

-Frank

> 
> Rob
>