Skip to contentSkip to conversation

Add global CMA reserve area

PATCH Applied Devarsh Thakkar 5 participants 15 messages Jun 13, 2024 – Jul 5, 2024

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover Add global CMA reserve area

Add global CMA reserve area for AM62x, AM62A and AM62P SoCs. These SoCs do not have MMU and hence require contiguous memory pool to support various multimedia use-cases.

Brandon Brnich (1):
arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

Devarsh Thakkar (2):
arm64: dts: ti: k3-am62x-sk-common: Reserve 128MiB of global CMA
arm64: dts: ti: k3-am62a7-sk: Reserve 576MiB of global CMA

 arch/arm64/boot/dts/ti/k3-am62a7-sk.dts        | 9 +++++++++
 arch/arm64/boot/dts/ti/k3-am62p5-sk.dts        | 7 +++++++
 arch/arm64/boot/dts/ti/k3-am62x-sk-common.dtsi | 8 ++++++++
 3 files changed, 24 insertions(+)
Signature · 1 line
2.39.1
#1 lore raw

PATCH 1/3 arm64: dts: ti: k3-am62x-sk-common: Reserve 128MiB of global CMA

Reserve 128MiB of global CMA which is also marked as re-usable so that OS can also use the same if peripheral drivers are not using the same.

AM62x supports multimedia components such as GPU, dual Display and Camera. Assuming the worst-case scenario where all 3 are run in parallel below is the calculation :

1) OV5640 camera sensor supports 1920x1080 resolution -> 1920 width x 1080 height x 2 bytesperpixel x 8 buffers (default in yavta) : 32MiB

2) 1920x1200 Microtips LVDS panel supported
-> 1920 width x 1080 height x 4 bytesperpixel x 2 buffers :
16 MiB

3) 1920x1080 HDMI display supported
-> 1920 width x 1080 height x 4 bytesperpixel x 2 buffers :
15.82 MiB which is ~16 MiB

4) IMG GPU shares with display allocated buffers while rendering but in case some dedicated operation viz color conversion, keeping same window of ~16 MiB for GPU too.

Total is 80 MiB and adding 32 MiB for other peripherals and extra 16 MiB to keep as buffer for fragmentation thus rounding total to 128 MiB.

---
 arch/arm64/boot/dts/ti/k3-am62x-sk-common.dtsi | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/arm64/boot/dts/ti/k3-am62x-sk-common.dtsi +8 -0
index f4948b937627..52231bfe60fe 100644
@@ -48,6 +48,14 @@ ramoops@9ca00000 {
 			pmsg-size = <0x8000>;
 		};
 
+		/* global cma region */
+		linux,cma {
+			compatible = "shared-dma-pool";
+			reusable;
+			size = <0x00 0x8000000>;
+			linux,cma-default;
+		};
+
 		secure_tfa_ddr: tfa@9e780000 {
 			reg = <0x00 0x9e780000 0x00 0x80000>;
 			alignment = <0x1000>;
Signature · 1 line
2.39.1
#2 lore raw

PATCH 2/3 arm64: dts: ti: k3-am62a7-sk: Reserve 576MiB of global CMA

In reply to Devarsh Thakkar

Reserve 576MiB of CMA as global CMA pool starting after initial 1GiB of DDR.

AM62ax has different multimedia components such as Camera, Display, H.264 VPU and JPEG Encoder which use CMA for buffer allocations.

The 12x 720x480 realtime VPU decode use-case requires 544MiB of CMA, additional 32MiB is kept as buffer in case some other peripheral also require it while VPU is running.

The reason to choose latter 1GiB is to not overlap with existing memory map which is utilizing initial 1GiB for remoteproc firmwares as shared here [1].

Also some drivers such as JPEG require 32bit addressing so not allocating from higher DDR address.

---
 arch/arm64/boot/dts/ti/k3-am62a7-sk.dts | 9 +++++++++
 1 file changed, 9 insertions(+)
arch/arm64/boot/dts/ti/k3-am62a7-sk.dts +9 -0
index e026f65738b3..67faf46d7a35 100644
@@ -40,6 +40,15 @@ reserved-memory {
 		#size-cells = <2>;
 		ranges;
 
+		/* global cma region */
+		linux,cma {
+			compatible = "shared-dma-pool";
+			reusable;
+			size = <0x00 0x24000000>;
+			alloc-ranges = <0x00 0xc0000000 0x00 0x24000000>;
+			linux,cma-default;
+		};
+
 		secure_tfa_ddr: tfa@9e780000 {
 			reg = <0x00 0x9e780000 0x00 0x80000>;
 			alignment = <0x1000>;
Signature · 1 line
2.39.1
Tested-by on v1 2/3
#3 lore raw

Re: [PATCH 2/3] arm64: dts: ti: k3-am62a7-sk: Reserve 576MiB of global CMA

Hi Devarsh,

Quoted 14 lines · Devarsh Thakkar

Reserve 576MiB of CMA as global CMA pool starting after initial 1GiB of DDR.

AM62ax has different multimedia components such as Camera, Display, H.264 VPU and JPEG Encoder which use CMA for buffer allocations.

The 12x 720x480 realtime VPU decode use-case requires 544MiB of CMA, additional 32MiB is kept as buffer in case some other peripheral also require it while VPU is running.

The reason to choose latter 1GiB is to not overlap with existing memory map which is utilizing initial 1GiB for remoteproc firmwares as shared here [1].

Quoted 5 lines

Also some drivers such as JPEG require 32bit addressing so not allocating from higher DDR address.

I have validated that this patch works with VPU.

Best,
Brandon

Quoted 27 lines
---
 arch/arm64/boot/dts/ti/k3-am62a7-sk.dts | 9 +++++++++
 1 file changed, 9 insertions(+)
arch/arm64/boot/dts/ti/k3-am62a7-sk.dts +9 -0
index e026f65738b3..67faf46d7a35 100644
@@ -40,6 +40,15 @@ reserved-memory {
 		#size-cells = <2>;
 		ranges;
 
+		/* global cma region */
+		linux,cma {
+			compatible = "shared-dma-pool";
+			reusable;
+			size = <0x00 0x24000000>;
+			alloc-ranges = <0x00 0xc0000000 0x00 0x24000000>;
+			linux,cma-default;
+		};
+
 		secure_tfa_ddr: tfa@9e780000 {
 			reg = <0x00 0x9e780000 0x00 0x80000>;
 			alignment = <0x1000>;
Signature · 1 line
2.39.1
#4 lore raw

PATCH 3/3 arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

In reply to Devarsh Thakkar

From: Brandon Brnich <[email protected]>

AM62p has different multimedia components such as Camera, Display, H264 Video Codec which uses CMA for buffer allocations. We require 576MiB for 12 channel decode-to-encode 720x480@30FPS use case.

---
 arch/arm64/boot/dts/ti/k3-am62p5-sk.dts | 7 +++++++
 1 file changed, 7 insertions(+)
arch/arm64/boot/dts/ti/k3-am62p5-sk.dts +7 -0
index fb980d46e304..5ef74d9f8eea 100644
@@ -48,6 +48,13 @@ reserved-memory {
 		#size-cells = <2>;
 		ranges;
 
+		linux,cma {
+			compatible = "shared-dma-pool";
+			reusable;
+			size = <0x00 0x24000000>;
+			linux,cma-default;
+		};
+
 		secure_tfa_ddr: tfa@9e780000 {
 			reg = <0x00 0x9e780000 0x00 0x80000>;
 			no-map;
Signature · 1 line
2.39.1
#5 lore raw

Re: [PATCH 3/3] arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

Hi Devarsh,

Quoted 17 lines · Devarsh Thakkar

From: Brandon Brnich <[email protected]>

AM62p has different multimedia components such as Camera, Display, H264 Video Codec which uses CMA for buffer allocations. We require 576MiB for 12 channel decode-to-encode 720x480@30FPS use case.

---
 arch/arm64/boot/dts/ti/k3-am62p5-sk.dts | 7 +++++++
 1 file changed, 7 insertions(+)
arch/arm64/boot/dts/ti/k3-am62p5-sk.dts +0 -0
index fb980d46e304..5ef74d9f8eea 100644

arch/arm64/boot/dts/ti/k3-am62p5-sk.dts reserved-memory { · line 56

Quoted 9 lines
arch/arm64/boot/dts/ti/k3-am62p5-sk.dts +6 -0
@@ -48,6 +48,13 @@ reserved-memory {
 		#size-cells = <2>;
 		ranges;
 
+		linux,cma {
+			compatible = "shared-dma-pool";
+			reusable;
+			size = <0x00 0x24000000>;
+			linux,cma-default;
+		};

Since AM62p has 8gb memory, this allocation can come from upper portion. Doing so breaks Wave5 encoding/decoding as the driver can not yet handle 48 bit addressing. 48bit support is scheduled to be upstreamed, but unsure of when this will actually make it in.

Could we force this into lower 32bits using same
alloc-ranges as done in your AM62a patch[0]?

[0]: https://patchwork.kernel.org/project/linux-arm-kernel/patch/[email protected]/

Best,
Brandon

Quoted 7 lines
arch/arm64/boot/dts/ti/k3-am62p5-sk.dts +1 -0
+
 		secure_tfa_ddr: tfa@9e780000 {
 			reg = <0x00 0x9e780000 0x00 0x80000>;
 			no-map;
Signature · 1 line
2.39.1
#6 lore raw

Re: [PATCH 3/3] arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

Hi Vignesh,

On 14/06/24 22:34, Brandon Brnich wrote:
[..]

Quoted 20 lines
Quoted 5 lines
arch/arm64/boot/dts/ti/k3-am62p5-sk.dts +0 -0
index fb980d46e304..5ef74d9f8eea 100644

arch/arm64/boot/dts/ti/k3-am62p5-sk.dts reserved-memory { · line 56

Quoted 9 lines
arch/arm64/boot/dts/ti/k3-am62p5-sk.dts +6 -0
@@ -48,6 +48,13 @@ reserved-memory {
 		#size-cells = <2>;
 		ranges;
 
+		linux,cma {
+			compatible = "shared-dma-pool";
+			reusable;
+			size = <0x00 0x24000000>;
+			linux,cma-default;
+		};

Since AM62p has 8gb memory, this allocation can come from upper portion. Doing so breaks Wave5 encoding/decoding as the driver can not yet handle 48 bit addressing. 48bit support is scheduled to be upstreamed, but unsure of when this will actually make it in.

Quoted 3 lines

Could we force this into lower 32bits using same
alloc-ranges as done in your AM62a patch[0]?

I would like to take a look at this separately, is it possible for you to drop this patch [3/3] from series due to above comments and take the rest two from the series ?

Regards
Devarsh

Reviewed-by on the v1 series
#7 lore raw
In reply to Devarsh Thakkar
Quoted 7 lines · Devarsh Thakkar

Add global CMA reserve area for AM62x, AM62A and AM62P SoCs. These SoCs do not have MMU and hence require contiguous memory pool to support various multimedia use-cases.

Brandon Brnich (1):
arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

Quoted 8 lines

Devarsh Thakkar (2):
arm64: dts: ti: k3-am62x-sk-common: Reserve 128MiB of global CMA
arm64: dts: ti: k3-am62a7-sk: Reserve 576MiB of global CMA

 arch/arm64/boot/dts/ti/k3-am62a7-sk.dts        | 9 +++++++++
 arch/arm64/boot/dts/ti/k3-am62p5-sk.dts        | 7 +++++++
 arch/arm64/boot/dts/ti/k3-am62x-sk-common.dtsi | 8 ++++++++
 3 files changed, 24 insertions(+)

I'm still a little torn about putting this allocation into the device tree directly as the actual required allocation size depends on the task.

If it's allowed though, this series is fine for introducing those changes. This uses the long-tested values we've been using on our tree for a bit now. The only thing that's a little worrying is the missing range definitions for devices with more than 32bits of addressable memory as Brandon has pointed out. Once that's addressed:

Specifying these regions using the kernel cmdline parameter via u-boot was brought up as a potential workaround. This is fine until you get into distro boot methods which will almost certainly attempt to override those. I don't know. Still a little odd. Curious how the community feels about it.

Technically the user or distro can still override it with the cmdline parameter if necessary, so this may be the best way to have a useful default.

#8 lore raw
Quoted 12 lines · Randolph Sapp
Quoted 11 lines · Devarsh Thakkar

Add global CMA reserve area for AM62x, AM62A and AM62P SoCs. These SoCs do not have MMU and hence require contiguous memory pool to support various multimedia use-cases.

Brandon Brnich (1):
arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

Devarsh Thakkar (2):
arm64: dts: ti: k3-am62x-sk-common: Reserve 128MiB of global CMA
arm64: dts: ti: k3-am62a7-sk: Reserve 576MiB of global CMA

Quoted 8 lines
Quoted 4 lines
  arch/arm64/boot/dts/ti/k3-am62a7-sk.dts        | 9 +++++++++
  arch/arm64/boot/dts/ti/k3-am62p5-sk.dts        | 7 +++++++
  arch/arm64/boot/dts/ti/k3-am62x-sk-common.dtsi | 8 ++++++++
  3 files changed, 24 insertions(+)

I'm still a little torn about putting this allocation into the device tree directly as the actual required allocation size depends on the task.

That is the exact reason this does not belong in DT. For everyone *not* using the most extreme case (12x decodes at the same time), this is all wasted space. If one is running out of CMA, they can add more on the kernel cmdline.

Quoted 8 lines

If it's allowed though, this series is fine for introducing those changes. This uses the long-tested values we've been using on our tree for a bit now. The only thing that's a little worrying is the missing range definitions for devices with more than 32bits of addressable memory as Brandon has pointed out. Once that's addressed:

Quoted 8 lines

Specifying these regions using the kernel cmdline parameter via u-boot was brought up as a potential workaround. This is fine until you get into distro boot methods which will almost certainly attempt to override those. I don't know. Still a little odd. Curious how the community feels about it.

Technically the user or distro can still override it with the cmdline parameter if necessary, so this may be the best way to have a useful default.

The most useful default is one that doesn't eat 576 MiB of memory "just in case". Needing that much CMA is the exception case and should be the one that requires adding something to the kernel cmdline.

If the kernel cmdline option does not work in some cases then we should fix that instead of hard-coding a workaround here in DT. We are robbing ourselves of a better solution here.

Andrew

#9 lore raw

Hi Andrew, Vignesh,

Quoted 18 lines · Andrew Davis
Quoted 17 lines · Randolph Sapp
Quoted 15 lines · Devarsh Thakkar

Add global CMA reserve area for AM62x, AM62A and AM62P SoCs. These SoCs do not have MMU and hence require contiguous memory pool to support various multimedia use-cases.

Brandon Brnich (1):
   arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

Devarsh Thakkar (2):
   arm64: dts: ti: k3-am62x-sk-common: Reserve 128MiB of global CMA
   arm64: dts: ti: k3-am62a7-sk: Reserve 576MiB of global CMA

  arch/arm64/boot/dts/ti/k3-am62a7-sk.dts        | 9 +++++++++ arch/arm64/boot/dts/ti/k3-am62p5-sk.dts        | 7 +++++++ arch/arm64/boot/dts/ti/k3-am62x-sk-common.dtsi | 8 ++++++++

  3 files changed, 24 insertions(+)
Quoted 9 lines
Quoted 3 lines

I'm still a little torn about putting this allocation into the device tree directly as the actual required allocation size depends on the task.

That is the exact reason this does not belong in DT. For everyone *not* using the most extreme case (12x decodes at the same time), this is all wasted space. If one is running out of CMA, they can add more on the kernel cmdline.

I disagree with this. The 12x decode for 480p is not an extreme use-case this is something VPU is capable to run at optimum frame-rate (12x 1080p it can't) and as the AM62A7 is meant to be AI + multimedia centric device, per the device definition we were given the requirements to support a list of multimedia use-cases which should work out of box and 12x decode for 480p was one of them as device is very much capable of doing that with optimum performance and I don't think it is right to change these requirements on the fly.

The AM62A7 board has 4 GiB of DDR and we have been using this CMA value since more than a year, I have never heard anyone complain about out of memory or CMA starvation and it suffices to requirements of *most use-cases*, but if for some specific use-case it doesn't suffice, user can change it via kernel cmdline.

The kernelcmdline suggestion doesn't suffice out of box experience required, we don't want to ask the user to reboot the board everytime they run out of CMA.

Quoted 10 lines
Quoted 10 lines

If it's allowed though, this series is fine for introducing those changes. This uses the long-tested values we've been using on our tree for a bit now. The only
thing that's a little worrying is the missing range definitions for devices with
more than 32bits of addressable memory as Brandon has pointed out. Once that's addressed:

Quoted 9 lines
Quoted 8 lines

Specifying these regions using the kernel cmdline parameter via u-boot was brought up as a potential workaround. This is fine until you get into distro boot methods which will almost certainly attempt to override those. I don't know. Still a little odd. Curious how the community feels about it.

Technically the user or distro can still override it with the cmdline parameter if necessary, so this may be the best way to have a useful default.

Unlike above, this solution is independent of distro as it should be as we want that all the supported multimedia use-cases should work out of box. This solution is nothing illegal as CMA region carveouts are not a kernel deprecated feature.

Quoted 4 lines

The most useful default is one that doesn't eat 576 MiB of memory "just in case". Needing that much CMA is the exception case and should be the one that requires adding something to the kernel cmdline.

I disagree with this, I would have agreed if this point was made in context of generic device, but we are forgetting here that AM62A7 is a AI+multimedia centric device and customers expect multimedia use-cases to work out of box.

We have 4 GiB RAM and if carving out 576 MiB is helping achieve all major multimedia use-cases then what's the problem ? What exactly is failing for you ? If some specific scenarios are getting hurt then in that case overlays or kernel cmdline option can be used to override the cma.

I feel we are over-complicating things here and going back-and-forth each cycle even though there are no competing alternatives present today. And this blocks out of box experience, as today even the basic HDMI and GPU use-cases don't work out of box.

I had also discussed around this with community on last OSS summit as discussed here [1] were it was suggested too to use this solution as adopted by other vendors.

[1]: https://lore.kernel.org/all/[email protected]/

Regards
Devarsh

#10 lore raw
Quoted 68 lines · Devarsh Thakkar

Hi Andrew, Vignesh,

Quoted 18 lines · Andrew Davis
Quoted 17 lines · Randolph Sapp
Quoted 15 lines · Devarsh Thakkar

Add global CMA reserve area for AM62x, AM62A and AM62P SoCs. These SoCs do not have MMU and hence require contiguous memory pool to support various multimedia use-cases.

Brandon Brnich (1):
   arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

Devarsh Thakkar (2):
   arm64: dts: ti: k3-am62x-sk-common: Reserve 128MiB of global CMA
   arm64: dts: ti: k3-am62a7-sk: Reserve 576MiB of global CMA

  arch/arm64/boot/dts/ti/k3-am62a7-sk.dts        | 9 +++++++++ arch/arm64/boot/dts/ti/k3-am62p5-sk.dts        | 7 +++++++ arch/arm64/boot/dts/ti/k3-am62x-sk-common.dtsi | 8 ++++++++

  3 files changed, 24 insertions(+)
Quoted 9 lines
Quoted 3 lines

I'm still a little torn about putting this allocation into the device tree directly as the actual required allocation size depends on the task.

That is the exact reason this does not belong in DT. For everyone *not* using the most extreme case (12x decodes at the same time), this is all wasted space. If one is running out of CMA, they can add more on the kernel cmdline.

I disagree with this. The 12x decode for 480p is not an extreme use-case this is something VPU is capable to run at optimum frame-rate (12x 1080p it can't) and as the AM62A7 is meant to be AI + multimedia centric device, per the device definition we were given the requirements to support a list of multimedia use-cases which should work out of box and 12x decode for 480p was one of them as device is very much capable of doing that with optimum performance and I don't think it is right to change these requirements on the fly.

The AM62A7 board has 4 GiB of DDR and we have been using this CMA value since more than a year, I have never heard anyone complain about out of memory or CMA starvation and it suffices to requirements of *most use-cases*, but if for some specific use-case it doesn't suffice, user can change it via kernel cmdline.

The kernelcmdline suggestion doesn't suffice out of box experience required, we don't want to ask the user to reboot the board everytime they run out of CMA.

Quoted 10 lines
Quoted 10 lines

If it's allowed though, this series is fine for introducing those changes. This uses the long-tested values we've been using on our tree for a bit now. The only
thing that's a little worrying is the missing range definitions for devices with
more than 32bits of addressable memory as Brandon has pointed out. Once that's addressed:

Quoted 9 lines
Quoted 8 lines

Specifying these regions using the kernel cmdline parameter via u-boot was brought up as a potential workaround. This is fine until you get into distro boot methods which will almost certainly attempt to override those. I don't know. Still a little odd. Curious how the community feels about it.

Technically the user or distro can still override it with the cmdline parameter if necessary, so this may be the best way to have a useful default.

Quoted 4 lines

Unlike above, this solution is independent of distro as it should be as we want that all the supported multimedia use-cases should work out of box. This solution is nothing illegal as CMA region carveouts are not a kernel deprecated feature.

Right. I support this change for at least introducing a usable default. 32M of CMA is barely enough to run glmark2 under Weston once everything's up and running.

As I said before, the user or distro can still override the dt CMA block with the cma kernel parameter if they aren't happy with the default block. Unfortunately this is about the only way to have a usable default value to fall back on.

Quoted 19 lines
Quoted 4 lines

The most useful default is one that doesn't eat 576 MiB of memory "just in case". Needing that much CMA is the exception case and should be the one that requires adding something to the kernel cmdline.

I disagree with this, I would have agreed if this point was made in context of generic device, but we are forgetting here that AM62A7 is a AI+multimedia centric device and customers expect multimedia use-cases to work out of box.

We have 4 GiB RAM and if carving out 576 MiB is helping achieve all major multimedia use-cases then what's the problem ? What exactly is failing for you ? If some specific scenarios are getting hurt then in that case overlays or kernel cmdline option can be used to override the cma.

I feel we are over-complicating things here and going back-and-forth each cycle even though there are no competing alternatives present today. And this blocks out of box experience, as today even the basic HDMI and GPU use-cases don't work out of box.

Quoted 8 lines

I had also discussed around this with community on last OSS summit as discussed here [1] were it was suggested too to use this solution as adopted by other vendors.

[1]: https://lore.kernel.org/all/[email protected]/

Regards
Devarsh

If the community accepts it, it's fine by me.

- Randolph

#11 lore raw
Quoted 74 lines · Randolph Sapp
Quoted 68 lines · Devarsh Thakkar

Hi Andrew, Vignesh,

Quoted 18 lines · Andrew Davis
Quoted 17 lines · Randolph Sapp
Quoted 15 lines · Devarsh Thakkar

Add global CMA reserve area for AM62x, AM62A and AM62P SoCs. These SoCs do not have MMU and hence require contiguous memory pool to support various multimedia use-cases.

Brandon Brnich (1):
   arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

Devarsh Thakkar (2):
   arm64: dts: ti: k3-am62x-sk-common: Reserve 128MiB of global CMA
   arm64: dts: ti: k3-am62a7-sk: Reserve 576MiB of global CMA

  arch/arm64/boot/dts/ti/k3-am62a7-sk.dts        | 9 +++++++++ arch/arm64/boot/dts/ti/k3-am62p5-sk.dts        | 7 +++++++ arch/arm64/boot/dts/ti/k3-am62x-sk-common.dtsi | 8 ++++++++

  3 files changed, 24 insertions(+)
Quoted 9 lines
Quoted 3 lines

I'm still a little torn about putting this allocation into the device tree directly as the actual required allocation size depends on the task.

That is the exact reason this does not belong in DT. For everyone *not* using the most extreme case (12x decodes at the same time), this is all wasted space. If one is running out of CMA, they can add more on the kernel cmdline.

I disagree with this. The 12x decode for 480p is not an extreme use-case this is something VPU is capable to run at optimum frame-rate (12x 1080p it can't) and as the AM62A7 is meant to be AI + multimedia centric device, per the device definition we were given the requirements to support a list of multimedia use-cases which should work out of box and 12x decode for 480p was one of them as device is very much capable of doing that with optimum performance and I don't think it is right to change these requirements on the fly.

The AM62A7 board has 4 GiB of DDR and we have been using this CMA value since more than a year, I have never heard anyone complain about out of memory or CMA starvation and it suffices to requirements of *most use-cases*, but if for some specific use-case it doesn't suffice, user can change it via kernel cmdline.

The kernelcmdline suggestion doesn't suffice out of box experience required, we don't want to ask the user to reboot the board everytime they run out of CMA.

Quoted 10 lines
Quoted 10 lines

If it's allowed though, this series is fine for introducing those changes. This uses the long-tested values we've been using on our tree for a bit now. The only
thing that's a little worrying is the missing range definitions for devices with
more than 32bits of addressable memory as Brandon has pointed out. Once that's addressed:

Quoted 9 lines
Quoted 8 lines

Specifying these regions using the kernel cmdline parameter via u-boot was brought up as a potential workaround. This is fine until you get into distro boot methods which will almost certainly attempt to override those. I don't know. Still a little odd. Curious how the community feels about it.

Technically the user or distro can still override it with the cmdline parameter if necessary, so this may be the best way to have a useful default.

Quoted 4 lines

Unlike above, this solution is independent of distro as it should be as we want that all the supported multimedia use-cases should work out of box. This solution is nothing illegal as CMA region carveouts are not a kernel deprecated feature.

Quoted 9 lines

Right. I support this change for at least introducing a usable default. 32M of CMA is barely enough to run glmark2 under Weston once everything's up and running.

As I said before, the user or distro can still override the dt CMA block with the cma kernel parameter if they aren't happy with the default block. Unfortunately this is about the only way to have a usable default value to fall back on.

Given the number of SoMs and non TI EVMs that are about to come out with AM62A/P and AM67s, we need to provide a consistent way of being able to support multimedia IPs out of the box. Modifying cmdline may not always be feasible given distro defaults don't always provide a way to do so.

So I am inclined to queue first 2 patches unless there is another way t achieve this.

[...]

Signature · 2 lines
Regards
Vignesh
#12 lore raw
Quoted 88 lines · Vignesh Raghavendra
Quoted 74 lines · Randolph Sapp
Quoted 68 lines · Devarsh Thakkar

Hi Andrew, Vignesh,

Quoted 18 lines · Andrew Davis
Quoted 17 lines · Randolph Sapp
Quoted 15 lines · Devarsh Thakkar

Add global CMA reserve area for AM62x, AM62A and AM62P SoCs. These SoCs do not have MMU and hence require contiguous memory pool to support various multimedia use-cases.

Brandon Brnich (1):
   arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

Devarsh Thakkar (2):
   arm64: dts: ti: k3-am62x-sk-common: Reserve 128MiB of global CMA
   arm64: dts: ti: k3-am62a7-sk: Reserve 576MiB of global CMA

   arch/arm64/boot/dts/ti/k3-am62a7-sk.dts        | 9 +++++++++
   arch/arm64/boot/dts/ti/k3-am62p5-sk.dts        | 7 +++++++
   arch/arm64/boot/dts/ti/k3-am62x-sk-common.dtsi | 8 ++++++++
   3 files changed, 24 insertions(+)
Quoted 9 lines
Quoted 3 lines

I'm still a little torn about putting this allocation into the device tree directly as the actual required allocation size depends on the task.

That is the exact reason this does not belong in DT. For everyone *not* using the most extreme case (12x decodes at the same time), this is all wasted space. If one is running out of CMA, they can add more on the kernel cmdline.

I disagree with this. The 12x decode for 480p is not an extreme use-case this is something VPU is capable to run at optimum frame-rate (12x 1080p it can't) and as the AM62A7 is meant to be AI + multimedia centric device, per the device definition we were given the requirements to support a list of multimedia use-cases which should work out of box and 12x decode for 480p was one of them as device is very much capable of doing that with optimum performance and I don't think it is right to change these requirements on the fly.

The AM62A7 board has 4 GiB of DDR and we have been using this CMA value since more than a year, I have never heard anyone complain about out of memory or CMA starvation and it suffices to requirements of *most use-cases*, but if for some specific use-case it doesn't suffice, user can change it via kernel cmdline.

The kernelcmdline suggestion doesn't suffice out of box experience required, we don't want to ask the user to reboot the board everytime they run out of CMA.

Quoted 10 lines
Quoted 10 lines

If it's allowed though, this series is fine for introducing those changes. This uses the long-tested values we've been using on our tree for a bit now. The only
thing that's a little worrying is the missing range definitions for devices with
more than 32bits of addressable memory as Brandon has pointed out. Once that's addressed:

Quoted 9 lines
Quoted 8 lines

Specifying these regions using the kernel cmdline parameter via u-boot was brought up as a potential workaround. This is fine until you get into distro boot methods which will almost certainly attempt to override those. I don't know. Still a little odd. Curious how the community feels about it.

Technically the user or distro can still override it with the cmdline parameter if necessary, so this may be the best way to have a useful default.

Quoted 4 lines

Unlike above, this solution is independent of distro as it should be as we want that all the supported multimedia use-cases should work out of box. This solution is nothing illegal as CMA region carveouts are not a kernel deprecated feature.

Quoted 9 lines

Right. I support this change for at least introducing a usable default. 32M of CMA is barely enough to run glmark2 under Weston once everything's up and running.

As I said before, the user or distro can still override the dt CMA block with the cma kernel parameter if they aren't happy with the default block. Unfortunately this is about the only way to have a usable default value to fall back on.

Quoted 5 lines

Given the number of SoMs and non TI EVMs that are about to come out with AM62A/P and AM67s, we need to provide a consistent way of being able to support multimedia IPs out of the box. Modifying cmdline may not always be feasible given distro defaults don't always provide a way to do so.

We need to keep thinking then. I empathize with desire to put configuration in Device Tree. DT feels like a great spot for it, it is ubiquitous on these boards and has a good bit of tooling around it. We are already describing the hardware, why not configure it here too? But the reason we do not want to go down that road is simple:
DT takes away use-case flexibility. A lack of flexibility is fine for hardware which is unchanging, but not for configuration.

Device policy and configuration must be left to userspace.

It is not for us to decide how folks should use our hardware, and that is what we are doing when we configure it in DT.

For configuration that must happen in early boot before userspace is available (such as kernel stdout and memory carveouts) we have the kernel cmdline. If we find something that cannot be done today though the cmdline, then we should add that support to the cmdline, not give up and just hide the configuration in DT.

What this series does is already available on the kernel cmdline. Our bootloader can provide sane defaults on the cmdline today. If the worry is that distros will override this default then go fix those distros.

Quoted 3 lines

So I am inclined to queue first 2 patches unless there is another way t achieve this.

Our lack of creativity in finding better solutions to this issue is not an excuse to add more junk to DT..

Andrew

Quoted 2 lines

[...]

#13 lore raw
Quoted 95 lines · Andrew Davis
Quoted 88 lines · Vignesh Raghavendra
Quoted 74 lines · Randolph Sapp
Quoted 68 lines · Devarsh Thakkar

Hi Andrew, Vignesh,

Quoted 18 lines · Andrew Davis
Quoted 17 lines · Randolph Sapp
Quoted 15 lines · Devarsh Thakkar

Add global CMA reserve area for AM62x, AM62A and AM62P SoCs. These SoCs do not have MMU and hence require contiguous memory pool to support various multimedia use-cases.

Brandon Brnich (1):
   arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

Devarsh Thakkar (2):
   arm64: dts: ti: k3-am62x-sk-common: Reserve 128MiB of global CMA
   arm64: dts: ti: k3-am62a7-sk: Reserve 576MiB of global CMA

   arch/arm64/boot/dts/ti/k3-am62a7-sk.dts        | 9 +++++++++
   arch/arm64/boot/dts/ti/k3-am62p5-sk.dts        | 7 +++++++
   arch/arm64/boot/dts/ti/k3-am62x-sk-common.dtsi | 8 ++++++++
   3 files changed, 24 insertions(+)
Quoted 9 lines
Quoted 3 lines

I'm still a little torn about putting this allocation into the device tree directly as the actual required allocation size depends on the task.

That is the exact reason this does not belong in DT. For everyone *not* using the most extreme case (12x decodes at the same time), this is all wasted space. If one is running out of CMA, they can add more on the kernel cmdline.

I disagree with this. The 12x decode for 480p is not an extreme use-case this is something VPU is capable to run at optimum frame-rate (12x 1080p it can't) and as the AM62A7 is meant to be AI + multimedia centric device, per the device definition we were given the requirements to support a list of multimedia use-cases which should work out of box and 12x decode for 480p was one of them as device is very much capable of doing that with optimum performance and I don't think it is right to change these requirements on the fly.

The AM62A7 board has 4 GiB of DDR and we have been using this CMA value since more than a year, I have never heard anyone complain about out of memory or CMA starvation and it suffices to requirements of *most use-cases*, but if for some specific use-case it doesn't suffice, user can change it via kernel cmdline.

The kernelcmdline suggestion doesn't suffice out of box experience required, we don't want to ask the user to reboot the board everytime they run out of CMA.

Quoted 10 lines
Quoted 10 lines

If it's allowed though, this series is fine for introducing those changes. This uses the long-tested values we've been using on our tree for a bit now. The only
thing that's a little worrying is the missing range definitions for devices with
more than 32bits of addressable memory as Brandon has pointed out. Once that's addressed:

Quoted 9 lines
Quoted 8 lines

Specifying these regions using the kernel cmdline parameter via u-boot was brought up as a potential workaround. This is fine until you get into distro boot methods which will almost certainly attempt to override those. I don't know. Still a little odd. Curious how the community feels about it.

Technically the user or distro can still override it with the cmdline parameter if necessary, so this may be the best way to have a useful default.

Quoted 4 lines

Unlike above, this solution is independent of distro as it should be as we want that all the supported multimedia use-cases should work out of box. This solution is nothing illegal as CMA region carveouts are not a kernel deprecated feature.

Quoted 9 lines

Right. I support this change for at least introducing a usable default. 32M of CMA is barely enough to run glmark2 under Weston once everything's up and running.

As I said before, the user or distro can still override the dt CMA block with the cma kernel parameter if they aren't happy with the default block. Unfortunately this is about the only way to have a usable default value to fall back on.

Quoted 5 lines

Given the number of SoMs and non TI EVMs that are about to come out with AM62A/P and AM67s, we need to provide a consistent way of being able to support multimedia IPs out of the box. Modifying cmdline may not always be feasible given distro defaults don't always provide a way to do so.

Quoted 7 lines

We need to keep thinking then. I empathize with desire to put configuration in Device Tree. DT feels like a great spot for it, it is ubiquitous on these boards and has a good bit of tooling around it. We are already describing the hardware, why not configure it here too? But the reason we do not want to go down that road is simple:
DT takes away use-case flexibility. A lack of flexibility is fine for hardware which is unchanging, but not for configuration.

I agree with the sentiment here, but this explicit case is an exception. This is no more than a default value. Userspace can still change this allocation with the kernel cma parameter. That still takes precedence.

Quoted 11 lines

Device policy and configuration must be left to userspace.

It is not for us to decide how folks should use our hardware, and that is what we are doing when we configure it in DT.

For configuration that must happen in early boot before userspace is available (such as kernel stdout and memory carveouts) we have the kernel cmdline. If we find something that cannot be done today though the cmdline, then we should add that support to the cmdline, not give up and just hide the configuration in DT.

Quoted 4 lines

What this series does is already available on the kernel cmdline. Our bootloader can provide sane defaults on the cmdline today. If the worry is that distros will override this default then go fix those distros.

Our bootloaders doing anything to kernel cmdline parameters is inherently non-standard. You just brought up userspace control. They clobber cmdline parameters all the time. Unsetting this bootloader value will be the first thing they do. This will be done passively, since they normally inject a new intermediary boot stage that they can actually control between u-boot and their kernel.

- Randolph

Quoted 11 lines
Quoted 3 lines

So I am inclined to queue first 2 patches unless there is another way t achieve this.

Our lack of creativity in finding better solutions to this issue is not an excuse to add more junk to DT..

Andrew

Quoted 2 lines

[...]

maintainer Applied
#14 lore raw

Re: (subset)[PATCH 0/3] Add global CMA reserve area

In reply to Devarsh Thakkar

Hi Devarsh Thakkar,

Quoted 8 lines · Devarsh Thakkar

Add global CMA reserve area for AM62x, AM62A and AM62P SoCs. These SoCs do not have MMU and hence require contiguous memory pool to support various multimedia use-cases.

Brandon Brnich (1):
arm64: dts: ti: k3-am62p5-sk: Reserve 576 MiB of global CMA

[...]

I have applied the following to branch ti-k3-dts-next on [1]. Thank you!

[1/3] arm64: dts: ti: k3-am62x-sk-common: Reserve 128MiB of global CMA

      commit: 9e8560556f9c41da28118af3385b4e9dc832ae2b
[2/3] arm64: dts: ti: k3-am62a7-sk: Reserve 576MiB of global CMA
      commit: 6406c5d5512c0814b8c155df7f833a98d9069a72

All being well this means that it will be integrated into the linux-next tree (usually sometime in the next 24 hours) and sent up the chain during the next merge window (or sooner if it is a relevant bug fix), however if problems are discovered then the patch may be dropped or reverted.

You may get further e-mails resulting from automated or manual testing and review of the tree, please engage with people reporting problems and send followup patches addressing any issues that are reported if needed.

If any updates are required or you are submitting further changes they should be sent as incremental updates against current git, existing patches will not be replaced.

Please add any relevant lists and maintainers to the CCs when replying to this mail.

[1] https://git.kernel.org/pub/scm/linux/kernel/git/ti/linux.git
--
Vignesh