Re: [PATCH] drm: gem: add an option for supporting the dma-coherent hardware.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi,

On 2023/6/8 15:39, Maxime Ripard wrote:
On Thu, Jun 08, 2023 at 01:18:38AM +0800, Sui Jingfeng wrote:
Hi,

On 2023/6/8 00:12, Paul Cercueil wrote:
Hi Sui,

Le mercredi 07 juin 2023 à 22:38 +0800, Sui Jingfeng a écrit :
Hi,  welcome to discussion.


I have limited skills in manipulating English.

It may not express what I'm really means in the short time.

Part of word in the sentence may not as accurate as your.

Well, please don't misunderstand, I'm not doing the rude to you.
No problem.

I will explain it with more details.

See below:


On 2023/6/7 20:09, Paul Cercueil wrote:
Hi Sui,

Le mercredi 07 juin 2023 à 18:30 +0800, Sui Jingfeng a écrit :
Hi,


On 2023/6/7 17:36, Paul Cercueil wrote:
Hi Sui,

Le mercredi 07 juin 2023 à 13:30 +0800, Sui Jingfeng a écrit :
The single map_noncoherent member of struct
drm_gem_dma_object
may
not
sufficient for describing the backing memory of the GEM
buffer
object.

Especially on dma-coherent systems, the backing memory is
both
cached
coherent for multi-core CPUs and dma-coherent for peripheral
device.
Say architectures like X86-64, LoongArch64, Loongson Mips64,
etc.

Whether a peripheral device is dma-coherent or not can be
implementation-dependent. The single map_noncoherent option
is
not
enough
to reflect real hardware anymore. For example, the Loongson
LS3A4000
CPU
and LS2K2000/LS2K1000 SoC, peripheral device of such hardware
platform
allways snoop CPU's cache. Doing the allocation with
dma_alloc_coherent
function is preferred. The return buffer is cached, it should
not
using
the default write-combine mapping. While with the current
implement,
there
no way to tell the drm core to reflect this.

This patch adds cached and coherent members to struct
drm_gem_dma_object.
which allow driver implements to inform the core. Introducing
new
mappings
while keeping the original default behavior unchanged.
Did you try to simply set the "dma-coherent" property to the
device's
node?
But this approach can only be applied for the device driver with
DT
support.

X86-64, Loongson ls3a4000 mips64, Loongson ls3a5000 CPU typically
do
not
have DT support.

They using ACPI to pass parameter from the firmware to Linux
kernel.

You approach will lost the effectiveness on such a case.
Well, I don't really know how ACPI handles it - but it should just
be a
matter of setting dev->dma_coherent. That's basically what the DT
code
does.

Some MIPS boards set it in their setup code for instance.

This is a *strategy*, not a *mechanism*.

In this case, DT is just used to describing the hardware.

(It is actually a hardware feature describing language, the
granularity
is large)

It does not changing the state of the hardware.

It's your platform firmware or kernel setting up code who actually do
such a things.


It's just that it works on *one* platform, it does not guarantee it
will
works on others.
If you add the "dma-coherent" property in a device node in DT, you
effectively specify that the device is DMA-coherent; so you describe
the hardware, which is what DT is for, and you are not changing the
state of the hardware.

Note that some MIPS platforms (arch/mips/alchemy/common/setup.c)
default to DMA-coherent mapping; I believe you could do something
similar with your Loongson LS3A4000 CPU and LS2K2000/LS2K1000 SoC.

The preblem is that device driver can have various demand.

It probably want to create different kind of buffers for different thing
simultaneously.

Say, one allocated with dma_alloc_coherent for command buffer or dma
descriptor

another one allocated with  dma_alloc_wc for uploading shader etc.

also has the third one allocated with dma_alloc_noncoherent() for doing some
else.
And it will work just fine.

struct device dma_coherent, or DT's dma-coherent property define that
the device doesn't need any kind of cache maintenance, ever. If it's
missing, we need to perform cache maintenance to keep coherency.

dma_alloc_* functions provide guarantees to the driver. With
dma_alloc_wc and dma_alloc_coherent, the buffer is coherent, and thus
you don't need to perform cache maintenance operations by hand in the
driver.

With dma_alloc_noncoherent, the buffer is non-coherent and the driver
needs to perform them when relevant.

How those buffers are created is platform specific, but the guarantees
provided *to the driver* are always there.

A buffer allocated with dma_alloc_coherent might be provided by
different means (at the hardware level with a coherency unit, by mapping
it non-cacheable), but as far as the driver is concerned it's always
going to be coherent.

Similarly, a driver using dma_alloc_noncoherent will always require
cache maintenance operations to use the API properly, even if the
hardware provides coherency (in which case, those operations will be
nop).

So, yeah, like I was saying in the other mail, it looks like you're
confusing a bunch of things. dma_alloc_* functions are about the driver
expectations and guarantees. DT's dma-coherent property is about how we
can implement them on a given platform.

They don't have to match, and that's precisely how we can have drivers
that run on any combination of platforms: the driver only cares about
the buffer guarantees, the platform description takes care of how they
are implemented.

You are right in overall.

Yeah, you have better understanding than me.


But let me give you an example which may made people confusing:


The the drm/ingenic and drm/etnaviv (KMS-RO) as an example:



when drm/etnaviv's etnaviv_gem_prime_import_sg_table() function get called,

drm/etnaviv is importing buffer from drm/ingenic.

drm/etnaviv is the importer, and drm/ingenic is the exporter.

drm/ingenic choose non-coherent mapping by default for JZ4770(this is gc800 in it).

It's cached, for fast CPU software rendering.


While drm/etnaviv import the buffer, get the SG, using the WC mapping by default.

Dose this cause *cache aliasing* because of different driver using different cache

mapping for the same backing memory ?


Because the imported buffer originally belong to the KMS driver side.

For drm/ingenic(jz4770), the BO will be cached, but their hardware can't guarantee coherency.

when etnaviv finished the rendering, they will do the resolve.

By using the WC mapping, the GPU will write directly to the system RAM.


1)

If CPU flush the cache back to the system RAM(framebuffer when running glmark-es2-drm).

then the image(resolved by GPU) in framebuffer will be overwrite by the stall data in the cache.


2)

Think about occasions when we need the CPU to do the read access to the rendered image.

(snap shoot, or using with X server fake EXA)

The CPU still think the share buffer as cached, it will read from the cache if hit.

while GPU write to RAM directly by using WC mapping.


Even it call dma_sync_single_for_device(), it only get SYNC-ed for the device.

there is no SYNC for the CPU's cached.

I think, In the end, it will lost of coherency.


3)

If the user want to use X server graphic environment,

then the case will be more complex for 3D acceleration support.


Even it hacks somewhere to call the sync for CPU,

they still may need invalid the cache frequently.

In this case, it will not get a good performance.


At any case,  such a KMS-RO combination((cached no-coherent + WC)) will be a misery./
/

While drm/ingenic could give up the hardware acceleration and the 3D acceleration in X environment.

it's OK, as its for low-ended graphic application.


But, at the other hand, it is say also why arm soc adhere to the write-combine.

because they have no choice.

While ingenic is the first exception, thanks Paul's patch which help us to understand a lot thing .


4)

While Loongson LS2K1000 SoC is DMA-coherent,

We are also prefer cached framebuffer for fast CPU software rendering.

I also get the hardware accelerated 3D works successfully,

even only for the GL client (such as glxgears and glmark2).


Therefore, on our hardware platform.

I force both the KMS driver and RO drivers to use cached mapping.

with the hardware maintained cache coherence blessing us.

It turn out that is works.


We don't need ( and don't want ) to call the dma_sync_*() series function,

not because we don't know it is no-op for DMA-coherent.

It is because that we realized that it is not needed with 100% confident.


We want to keep the code clean.

We are ZERO tolerate to the not needed function calls.


Maxime

--
Jingfeng




[Index of Archives]     [Linux DRI Users]     [Linux Intel Graphics]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]
  Powered by Linux