About LPAE supporting on EXYNOS5440

About LPAE supporting on EXYNOS5440

[Kukjin Kim]

Hi all,

The exynos5440 can support LPAE and referece boards, SSDK5440 and SD5v1 have
over 8GiB memory.

Current exynos_defconfig selects every EXYNOS SoCs including exynos5440 it
means supporting LPAE will be selected for other EXYNOS SoCs.

I'm wondering if seleting LPAE causes some problem for multiplatform. Can we
support LPAE enabled and disabled platform together with single zImage?

Thanks,
Kukjin

About LPAE supporting on EXYNOS5440

[Arnd Bergmann]

On Tuesday 11 June 2013 17:33:33 Kukjin Kim wrote:
> Hi all,
> 
> The exynos5440 can support LPAE and referece boards, SSDK5440 and SD5v1 have
> over 8GiB memory.
> 
> Current exynos_defconfig selects every EXYNOS SoCs including exynos5440 it
> means supporting LPAE will be selected for other EXYNOS SoCs.
> 
> I'm wondering if seleting LPAE causes some problem for multiplatform. Can we
> support LPAE enabled and disabled platform together with single zImage?

No, that isn't possible. I think we should enable all exynos support in the
regular multi_v7_defconfig and leave LPAE turned off for that, but create
another multi_lpae_defconfig that has all LPAE capable machines turned on.
This is also what I expect the distros to do.

	Arnd

About LPAE supporting on EXYNOS5440

[Kukjin Kim]

Arnd Bergmann wrote:
> 
> On Tuesday 11 June 2013 17:33:33 Kukjin Kim wrote:
> > Hi all,
> >
> > The exynos5440 can support LPAE and referece boards, SSDK5440 and SD5v1
> have
> > over 8GiB memory.
> >
> > Current exynos_defconfig selects every EXYNOS SoCs including exynos5440
> it
> > means supporting LPAE will be selected for other EXYNOS SoCs.
> >
> > I'm wondering if seleting LPAE causes some problem for multiplatform.
> Can we
> > support LPAE enabled and disabled platform together with single zImage?
> 
> No, that isn't possible.

Yeah, I know that.

> I think we should enable all exynos support in
> the
> regular multi_v7_defconfig and leave LPAE turned off for that,

But if so, problem will happen in the memory parsing without LPAE
enabling...so I'm wondering I should deselect exynos5440 in the defconfig
for other exynos SoCs...

> but create
> another multi_lpae_defconfig that has all LPAE capable machines turned on.
> This is also what I expect the distros to do.
> 
Yeah, it could be a solution for multiplatform at this moment. And if so,
probably I should deselect exynos5440 in the exynos_defconfig and need to
add another defconfig for exynos lpae stuff.

BTW, do you know what SoCs can support LPAE now?

Thanks,
Kukjin

About LPAE supporting on EXYNOS5440

[Arnd Bergmann]

On Tuesday 11 June 2013 19:05:22 Kukjin Kim wrote:
> Arnd Bergmann wrote:
> > 
> > On Tuesday 11 June 2013 17:33:33 Kukjin Kim wrote:
> > > Hi all,
> > >
> > > The exynos5440 can support LPAE and referece boards, SSDK5440 and SD5v1
> > have
> > > over 8GiB memory.
> > >
> > > Current exynos_defconfig selects every EXYNOS SoCs including exynos5440
> > it
> > > means supporting LPAE will be selected for other EXYNOS SoCs.
> > >
> > > I'm wondering if seleting LPAE causes some problem for multiplatform.
> > Can we
> > > support LPAE enabled and disabled platform together with single zImage?
> > 
> > No, that isn't possible.
> 
> Yeah, I know that.
> 
> > I think we should enable all exynos support in
> > the
> > regular multi_v7_defconfig and leave LPAE turned off for that,
> 
> But if so, problem will happen in the memory parsing without LPAE
> enabling...so I'm wondering I should deselect exynos5440 in the defconfig
> for other exynos SoCs...

I think there is a bug in the memory parsing code and already two people
have submitted independent fixes for that. It really should just work
even if you have more memory that ends up getting ignored.

> > but create
> > another multi_lpae_defconfig that has all LPAE capable machines turned on.
> > This is also what I expect the distros to do.
> > 
> Yeah, it could be a solution for multiplatform at this moment. And if so,
> probably I should deselect exynos5440 in the exynos_defconfig and need to
> add another defconfig for exynos lpae stuff.

I would definitely leave on exynos5440 support in defconfig. It's not a
lot of extra code, and if you have a system with less than 4GB memory,
you really don't want to enable LPAE because of the overhead.
 
> BTW, do you know what SoCs can support LPAE now?

Upstream I think we support exynos, vexpress (using A15, A7 and A12
core tiles), omap5 (partially), mvebu (Armada XP), highbank (Midway)
and some of the shmobile parts we added in 3.10.

In mach-sunxi, there will soon be support for Allwinner-A20 and -A31,
which are both Cortex-A7 based. MSM also has support for Cortex-A7
but not upstream. A couple of new platforms are under discussion
already that will also have LPAE, mostly server parts.

	Arnd

About LPAE supporting on EXYNOS5440

[Olof Johansson]

On Tue, Jun 11, 2013 at 11:53:36AM +0200, Arnd Bergmann wrote:
> On Tuesday 11 June 2013 17:33:33 Kukjin Kim wrote:
> > Hi all,
> > 
> > The exynos5440 can support LPAE and referece boards, SSDK5440 and SD5v1 have
> > over 8GiB memory.
> > 
> > Current exynos_defconfig selects every EXYNOS SoCs including exynos5440 it
> > means supporting LPAE will be selected for other EXYNOS SoCs.
> > 
> > I'm wondering if seleting LPAE causes some problem for multiplatform. Can we
> > support LPAE enabled and disabled platform together with single zImage?
> 
> No, that isn't possible. I think we should enable all exynos support in the
> regular multi_v7_defconfig and leave LPAE turned off for that, but create
> another multi_lpae_defconfig that has all LPAE capable machines turned on.
> This is also what I expect the distros to do.

I think it would be unfortunate to split the defconfig just for LPAE,
since all the other parts should be identical (drivers, etc).

I wonder if it makes more sense to just include a config fragment to
turn on LPAE together with the existing multi_v7_defconfig.


-Olof

About LPAE supporting on EXYNOS5440

[Subash Patel]

On 06/11/2013 06:13 PM, Arnd Bergmann wrote:
> On Tuesday 11 June 2013 19:05:22 Kukjin Kim wrote:
>> Arnd Bergmann wrote:
>>>
>>> On Tuesday 11 June 2013 17:33:33 Kukjin Kim wrote:
>>>> Hi all,
>>>>
>>>> The exynos5440 can support LPAE and referece boards, SSDK5440 and SD5v1
>>> have
>>>> over 8GiB memory.
>>>>
>>>> Current exynos_defconfig selects every EXYNOS SoCs including exynos5440
>>> it
>>>> means supporting LPAE will be selected for other EXYNOS SoCs.
>>>>
>>>> I'm wondering if seleting LPAE causes some problem for multiplatform.
>>> Can we
>>>> support LPAE enabled and disabled platform together with single zImage?
>>>
>>> No, that isn't possible.
>>
>> Yeah, I know that.
>>
>>> I think we should enable all exynos support in
>>> the
>>> regular multi_v7_defconfig and leave LPAE turned off for that,
>>
>> But if so, problem will happen in the memory parsing without LPAE
>> enabling...so I'm wondering I should deselect exynos5440 in the defconfig
>> for other exynos SoCs...
>
> I think there is a bug in the memory parsing code and already two people
> have submitted independent fixes for that. It really should just work
> even if you have more memory that ends up getting ignored.
>
>>> but create
>>> another multi_lpae_defconfig that has all LPAE capable machines turned on.
>>> This is also what I expect the distros to do.
>>>
>> Yeah, it could be a solution for multiplatform at this moment. And if so,
>> probably I should deselect exynos5440 in the exynos_defconfig and need to
>> add another defconfig for exynos lpae stuff.
>
> I would definitely leave on exynos5440 support in defconfig. It's not a
> lot of extra code, and if you have a system with less than 4GB memory,
> you really don't want to enable LPAE because of the overhead.

Even if we have <= 4GB memory, and the system designers have placed the 
memory banks in addresses > 2^32 bytes, then we need LPAE support. I 
think we may have such systems in future. So memory capacity is not the 
only parameter to judge if we need to enable LPAE or not.

Regards,
Subash

>
>> BTW, do you know what SoCs can support LPAE now?
>
> Upstream I think we support exynos, vexpress (using A15, A7 and A12
> core tiles), omap5 (partially), mvebu (Armada XP), highbank (Midway)
> and some of the shmobile parts we added in 3.10.
>
> In mach-sunxi, there will soon be support for Allwinner-A20 and -A31,
> which are both Cortex-A7 based. MSM also has support for Cortex-A7
> but not upstream. A couple of new platforms are under discussion
> already that will also have LPAE, mostly server parts.
>
> 	Arnd
>
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>

About LPAE supporting on EXYNOS5440

[Arnd Bergmann]

On Wednesday 12 June 2013, Subash Patel wrote:
> > I would definitely leave on exynos5440 support in defconfig. It's not a
> > lot of extra code, and if you have a system with less than 4GB memory,
> > you really don't want to enable LPAE because of the overhead.
> 
> Even if we have <= 4GB memory, and the system designers have placed the 
> memory banks in addresses > 2^32 bytes, then we need LPAE support. I 
> think we may have such systems in future. So memory capacity is not the 
> only parameter to judge if we need to enable LPAE or not.

Yes, but that wasn't the point. You can always build systems with
exynos5440 that don't need LPAE, so we should not assume that it is
not a reasonable configuration.

I would certainly advise system designers to put all the RAM and I/O
into normally addressable locations if possible to avoid the need
for LPAE, but of course that isn't always possible.

We should probably have someone measure the performance impact
of LPAE as well. If it's less than a few percent, we might not
care all that much.

	Arnd

About LPAE supporting on EXYNOS5440

[Kukjin Kim]

Arnd Bergmann wrote:
> 
> On Wednesday 12 June 2013, Subash Patel wrote:
> > > I would definitely leave on exynos5440 support in defconfig. It's not
> a
> > > lot of extra code, and if you have a system with less than 4GB memory,
> > > you really don't want to enable LPAE because of the overhead.
> >
> > Even if we have <= 4GB memory, and the system designers have placed the
> > memory banks in addresses > 2^32 bytes, then we need LPAE support. I
> > think we may have such systems in future. So memory capacity is not the
> > only parameter to judge if we need to enable LPAE or not.
> 
> Yes, but that wasn't the point. You can always build systems with
> exynos5440 that don't need LPAE, so we should not assume that it is
> not a reasonable configuration.
> 
Agreed.

> I would certainly advise system designers to put all the RAM and I/O
> into normally addressable locations if possible to avoid the need
> for LPAE, but of course that isn't always possible.
> 
Yes, same here. Actually, some IPs sometimes can request to alloc/use
something like dma buffer over 32 bit area after LPAE enabling, but will not
be happened without LPAE.

> We should probably have someone measure the performance impact
> of LPAE as well. If it's less than a few percent, we might not
> care all that much.
> 
Yes, after testing on exynos5440 board, LPAE doesn't cause any performance
degradation with lmbench tool.

- Kukjin