Re: Memory sharing issue by application on V4L2 based device driver with system mmu.

thank you for your comments.

your second comment has no any problem as I said before, user virtual
addess could be translated in page unit. but the problem, as you said,
is that when cpu access to the memory in user mode, the memory
allocated by malloc, page fault occurs so we can't find pfn to user
virtual address. I missed that. but I think we could resolve this one.

as before, user application allocates memory through malloc function
and then send it to device driver(using userptr feature). if the pfn
is null when device driver translated user virtual address in page
unit then it allocates phsical memory in page unit using some
interface such as alloc_page() and then mapping them. when pfn is
null, to check it and allocate physical memory in page unit could be
processed by videobuf2.

of course, videobuf2 has no any duty considered for system mmu. so
videobuf2 just provides callback for 3rd party and any platform with
system mmu such as Samsung SoC C210 implements the function(allocating
physical memory and mapping it) and registers it to callback of
videobuf2. by doing so, I think your first comment could be cleared.

please, feel free to give me your opinion and pointing out.

thank you.

2011년 1월 7일 오전 8:57, Jonghun Han [off-list ref]님의 말:

Hello,

There are two reasons why malloc isn't suitable for it.

The first is that malloc doesn't allocate memory when malloc is called.
So driver or vb2 cannot find PFN for it in the VIDIOC_QBUF.

The second is that malloc uses 4KB page allocation.
SYS.MMU(IO-MMU) can handle scattered memory. But it has a penalty when TLB
miss is occurred.
So as possible as physically contiguous pages are needed for performance
enhancement.

So new allocator which can clear two main issues is needed.

Best regards,

quoted

-----Original Message-----
From: <redacted> [mailto:linux-media-
owner@vger.kernel.org] On Behalf Of InKi Dae
Sent: Thursday, January 06, 2011 10:25 PM
To: <redacted>
Subject: Memory sharing issue by application on V4L2 based device driver

with

quoted

system mmu.

Hello, all.

I'd like to discuss memory sharing issue by application on v4l2 based

device driver

quoted

with system mmu and get some advices about that.

Now I am working on Samsung SoC C210 platform and this platform has some
multimedia devices with system mmu such as fimc, and mfc also we have
implemented device drivers for them. those drivers are based on V4L2

framework

quoted

with videobuf2. for system mmu of each device, we used VCM(Virtual

Contiguous

quoted

Memory) framework.

99 more quoted lines

--
To unsubscribe from this list: send the line "unsubscribe linux-fbdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Hello,

That's not a translation issue. What I mention is the size of allocation.
The malloc uses 4KB page allocation and SYS.MMU can handle it.
But 64KB or 1MB physically contiguous memory is better than 4KB page in the
point of performance.

Best regards,

-----Original Message-----
From: <redacted> [mailto:linux-media-
owner@vger.kernel.org] On Behalf Of InKi Dae
Sent: Friday, January 07, 2011 11:17 AM
To: Jonghun Han
Cc: <redacted>; <redacted>;

linux-fbdev;

kyungmin.park@samsung.com
Subject: Re: Memory sharing issue by application on V4L2 based device

driver

with system mmu.

thank you for your comments.

your second comment has no any problem as I said before, user virtual

addess

could be translated in page unit. but the problem, as you said, is that

when cpu

access to the memory in user mode, the memory allocated by malloc, page

fault

occurs so we can't find pfn to user virtual address. I missed that. but I

think we

could resolve this one.

as before, user application allocates memory through malloc function and

then

send it to device driver(using userptr feature). if the pfn is null when

device driver

translated user virtual address in page unit then it allocates phsical

memory in

page unit using some interface such as alloc_page() and then mapping

them. when

pfn is null, to check it and allocate physical memory in page unit could

be

processed by videobuf2.

of course, videobuf2 has no any duty considered for system mmu. so
videobuf2 just provides callback for 3rd party and any platform with

system mmu

such as Samsung SoC C210 implements the function(allocating physical

memory

and mapping it) and registers it to callback of videobuf2. by doing so, I

think your

first comment could be cleared.

please, feel free to give me your opinion and pointing out.

thank you.

2011년 1월 7일 오전 8:57, Jonghun Han [off-list ref]님의 말:

quoted

Hello,

There are two reasons why malloc isn't suitable for it.

The first is that malloc doesn't allocate memory when malloc is called.
So driver or vb2 cannot find PFN for it in the VIDIOC_QBUF.

The second is that malloc uses 4KB page allocation.
SYS.MMU(IO-MMU) can handle scattered memory. But it has a penalty when
TLB miss is occurred.
So as possible as physically contiguous pages are needed for
performance enhancement.

So new allocator which can clear two main issues is needed.

Best regards,

quoted

-----Original Message-----
From: <redacted> [mailto:linux-media-
owner@vger.kernel.org] On Behalf Of InKi Dae
Sent: Thursday, January 06, 2011 10:25 PM
To: <redacted>
Subject: Memory sharing issue by application on V4L2 based device
driver

with

quoted

system mmu.

Hello, all.

I'd like to discuss memory sharing issue by application on v4l2 based

device driver

quoted

with system mmu and get some advices about that.

118 more quoted lines

the body

of a message to majordomo@vger.kernel.org More majordomo info at
http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe linux-fbdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Hello, Mr. Jonghun.

please, note that user process's pgd and device's pgd with system mmu
are differnent. first, the memory allocated by user malloc should be
mapped to physical memory in page unit because user app also should be
accessed to own memory region. and when mapping the physical memory to
device virtual address, it creates page table entry for 64k or 1M and
then could map physical address to those entries. therefore what you
mention has no problem.

actual issue is not to need demend paging, some issues we don't
understand may exist as side effects.

thank you.

Jonghun Han 쓴 글:

Hello,

That's not a translation issue. What I mention is the size of allocation.
The malloc uses 4KB page allocation and SYS.MMU can handle it.
But 64KB or 1MB physically contiguous memory is better than 4KB page in the
point of performance.

Best regards,

quoted

-----Original Message-----
From: <redacted> [mailto:linux-media-
owner@vger.kernel.org] On Behalf Of InKi Dae
Sent: Friday, January 07, 2011 11:17 AM
To: Jonghun Han
Cc: <redacted>; <redacted>;

linux-fbdev;

quoted

kyungmin.park@samsung.com
Subject: Re: Memory sharing issue by application on V4L2 based device

driver

quoted

with system mmu.

thank you for your comments.

your second comment has no any problem as I said before, user virtual

addess

quoted

could be translated in page unit. but the problem, as you said, is that

when cpu

quoted

access to the memory in user mode, the memory allocated by malloc, page

fault

quoted

occurs so we can't find pfn to user virtual address. I missed that. but I

think we

quoted

could resolve this one.

as before, user application allocates memory through malloc function and

then

quoted

send it to device driver(using userptr feature). if the pfn is null when

device driver

quoted

translated user virtual address in page unit then it allocates phsical

memory in

quoted

page unit using some interface such as alloc_page() and then mapping

176 more quoted lines

--
To unsubscribe from this list: send the line "unsubscribe linux-fbdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Fre, 2011-01-07 at 11:17 +0900, InKi Dae wrote: 

thank you for your comments.

your second comment has no any problem as I said before, user virtual
addess could be translated in page unit. but the problem, as you said,
is that when cpu access to the memory in user mode, the memory
allocated by malloc, page fault occurs so we can't find pfn to user
virtual address. I missed that. but I think we could resolve this one.

as before, user application allocates memory through malloc function
and then send it to device driver(using userptr feature). if the pfn
is null when device driver translated user virtual address in page
unit then it allocates phsical memory in page unit using some
interface such as alloc_page() and then mapping them. when pfn is
null, to check it and allocate physical memory in page unit could be
processed by videobuf2.

of course, videobuf2 has no any duty considered for system mmu. so
videobuf2 just provides callback for 3rd party and any platform with
system mmu such as Samsung SoC C210 implements the function(allocating
physical memory and mapping it) and registers it to callback of
videobuf2. by doing so, I think your first comment could be cleared.

FWIW, TTM (drivers/gpu/drm/ttm, include/drm/ttm) is designed and used
for managing memory between CPU/GPU and kernel/userspace access. I
haven't looked at your requirements in detail, but if you haven't looked
at TTM yet, it sounds like it might be worth a look.


-- 
Earthling Michel Dänzer           |                http://www.vmware.com
Libre software enthusiast         |          Debian, X and DRI developer

Michel Dänzer 쓴 글:

On Fre, 2011-01-07 at 11:17 +0900, InKi Dae wrote: 

quoted

thank you for your comments.

your second comment has no any problem as I said before, user virtual
addess could be translated in page unit. but the problem, as you said,
is that when cpu access to the memory in user mode, the memory
allocated by malloc, page fault occurs so we can't find pfn to user
virtual address. I missed that. but I think we could resolve this one.

as before, user application allocates memory through malloc function
and then send it to device driver(using userptr feature). if the pfn
is null when device driver translated user virtual address in page
unit then it allocates phsical memory in page unit using some
interface such as alloc_page() and then mapping them. when pfn is
null, to check it and allocate physical memory in page unit could be
processed by videobuf2.

of course, videobuf2 has no any duty considered for system mmu. so
videobuf2 just provides callback for 3rd party and any platform with
system mmu such as Samsung SoC C210 implements the function(allocating
physical memory and mapping it) and registers it to callback of
videobuf2. by doing so, I think your first comment could be cleared.

FWIW, TTM (drivers/gpu/drm/ttm, include/drm/ttm) is designed and used
for managing memory between CPU/GPU and kernel/userspace access. I
haven't looked at your requirements in detail, but if you haven't looked
at TTM yet, it sounds like it might be worth a look.

Thank you for your advice. I would look at TTM driver.

 .. snip..

But 64KB or 1MB physically contiguous memory is better than 4KB page in the
point of performance.

Could you explain that in more details please? I presume you are
talking about a CPU that has a MMU unit, right?

Sorry for late reply.

Samsung SoC C210 has many multimedia IPs.
Each IP has its own MMU named SYSTEM MMU like CPU's MMU.
So it can handle discontiguous memory using device own virtual address.

What Inki Dae wants to discuss is how to allocate the memory and how to
share it with other multimedia IPs. 

Thank you for interesting.

Best regards,

-----Original Message-----
From: <redacted> [mailto:linux-fbdev-
owner@vger.kernel.org] On Behalf Of Konrad Rzeszutek Wilk
Sent: Wednesday, January 12, 2011 2:35 AM
To: Jonghun Han
Cc: 'InKi Dae'; <redacted>;

linux-arm-kernel@lists.infradead.org;

'linux-fbdev'; kyungmin.park@samsung.com
Subject: Re: Memory sharing issue by application on V4L2 based device

driver

with system mmu.


 .. snip..

quoted

But 64KB or 1MB physically contiguous memory is better than 4KB page
in the point of performance.

Could you explain that in more details please? I presume you are talking

about a

CPU that has a MMU unit, right?
--
To unsubscribe from this list: send the line "unsubscribe linux-fbdev" in

the body

of a message to majordomo@vger.kernel.org More majordomo info at
http://vger.kernel.org/majordomo-info.html