On Tue, Jan 30, 2018 at 3:49 PM, Greentime Hu <green.hu@xxxxxxxxx> wrote:
> Hi, Arnd:
>
> 2018-01-30 21:33 GMT+08:00 Arnd Bergmann <arnd@xxxxxxxx>:
>> On Tue, Jan 30, 2018 at 11:01 AM, Vincent Chen <deanbo422@xxxxxxxxx> wrote:
>>> 2018-01-24 19:10 GMT+08:00 Arnd Bergmann <arnd@xxxxxxxx>:
>>>> On Wed, Jan 24, 2018 at 12:09 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
>>>>> On Wed, Jan 24, 2018 at 11:53 AM, Vincent Chen <deanbo422@xxxxxxxxx> wrote:
>>>>>> 2018-01-18 18:14 GMT+08:00 Arnd Bergmann <arnd@xxxxxxxx>:
>>>>
>>>>> Ok. I still wonder about the kernel part of this though: is it a good idea
>>>>> for user space to configure whether the kernel does unaligned
>>>>> accesses? I would think that the kernel should just be fixed in such
>>>>> a case.
>>>>
>>>> To clarify: I'm asking only about unaligned accesses from kernel code itself,
>>>> which is generally considered a bug when
>>>> CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS is disabled.
>>>>
>>>> Arnd
>>>
>>> Thanks for your comments.
>>>
>>> For performance, we decide always disable
>>> CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS even if hardware supports
>>> unaligned accessing. Therefore, I will remove kernel unaligned accessing from
>>> nds32/mm/alignment.c. In other words, alignment.c only addresses unaligned
>>> accessing for user space.
>>
>> I'm not really following that logic, let's go through that again so I understand
>> the situation better.
>>
>> CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS should be set if and
>> only if you have a CPU that does not need to trap on unaligned accesses.
>>
>> What are the hardware capabilities on nds32? Do you have all three
>> categories:
>>
>> a) some CPUs that always trap on unaligned access
>> b) some CPUs that never trap on unaligned access
>> c) some CPUs that can be configured to either trap or not trap by
>> the kernel?
>>
> We have type a and c.
> We use CONFIG_ALIGNMENT_TRAP for a and
> CONFIG_HW_SUPPORT_UNALIGNMENT_ACCESS for c.
Ok, got it.
> Since unaligned access in kernel code itself should be considered as a
> bug, we will remove the emulation code to handle the kernel code
> unaligned accessed case.
> We think CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS and
> CONFIG_HW_SUPPORT_UNALIGNMENT_ACCESS have different purposes because
> it will still be more efficient to access by byte even if hardware
> support unaligned access.
> CONFIG_HW_SUPPORT_UNALIGNMENT_ACCESS is used to prevent generating
> unaligned access exception.
Hmm, this is a bit tricky. Most architectures actually assume that those
two are the same, and nothing else has a
HW_SUPPORT_UNALIGNMENT_ACCESS option.
We do actually have a related problem on 32-bit ARM, where the current
generation of processors (ARMv6 and higher) support unaligned
accesses for almost all instructions with the exception of those
instructions that operate on multiple memory locations (ldm/stm
and ldrd/strd). We can control the use of those instructions in inline
assembler, and gcc never uses them when it knows that a pointer
is unaligned, but when CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
is set, the kernel sometimes intentionally contains code sequences
that lead the compiler to believe that a variable is aligned when it
is not, so we end up needing a trap handler here.
We might at some point want to clean this up by going through
all uses of CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
and changing them in a way that leads to better results on both
arm32 and nds32.
Arnd
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
