Re: [PATCH] mtd: fix oops when writing to phram device on arm64

[Florian Fainelli <f.fainelli@xxxxxxxxx>]

On 9/21/2019 11:03 PM, Huaijie Yi wrote:
>> On Mon, Aug 19, 2019 at 3:39 PM Huaijie Yi <yihuaijie@xxxxxxxxxx> wrote:
>>>
>>> From: yihuaijie <yihuaijie@xxxxxxxxxx>
>>>
>>> memcpy() to memory remaped by ioremap() at an address not aligned to 
>>> 8-bytes will cause oops on arm64 platform, fixing this by using 
>>> ioremap_wc().
>>
>> This might be the right thing on arm64, but I fear not for all other archs and configurations.
>>
>> We had a similar chain of problems already with our sram driver.
>> See:
>> eb43e023130b ("misc: sram: add optional ioremap without write combining")
>> 0ab163ad1ea0 ("misc: sram: switch to ioremap_wc from ioremap")
>>
>> --
>> Thanks,
>> //richard
> 
> Hi, richard
> 
> I verified this on x86_64, and there isn't the problem.
> 
> I will resend a V2 to fix this only on ARM64.

The SRAM driver also takes care of using memcpy_{to,from}io() which the
PHRAM does not make use, that sounds like a possible issue too. There is
the alignment that is important, but also how you access the underlying
memory (e.g.: beats that you generate on the bus connecting you to that
memory area). ARM64 may be forgiving and allow memcpy() to work against
a WC mapping, other architectures, maybe not so much.

It sounds like adopting the solution that the SRAM driver did, whereby
adding an optional boolean Device Tree property to indicate whether
ioremap_wc() or ioremap() is to be used seems like the most flexible
solution.
-- 
Florian

______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/