Re: [PATCH v6 06/11] arm64: dts: ti: k3-am62a7-sk: Enable IPC with remote processors
From: Andrew Davis
Date: Mon Apr 07 2025 - 12:15:24 EST
On 4/7/25 9:13 AM, Judith Mendez wrote:
Hi Devarsh,
On 4/7/25 8:54 AM, Devarsh Thakkar wrote:
Hi Judith,
On 05/04/25 05:45, Judith Mendez wrote:
> From: Devarsh Thakkar <devarsht@xxxxxx>
Thanks for the patch.
For each remote proc, reserve memory for IPC and bind the mailbox
assignments. Two memory regions are reserved for each remote processor.
The first region of 1MB of memory is used for Vring shared buffers
and the second region is used as external memory to the remote processor
for the resource table and for tracebuffer allocations.
Signed-off-by: Devarsh Thakkar <devarsht@xxxxxx>
Signed-off-by: Hari Nagalla <hnagalla@xxxxxx>
Signed-off-by: Judith Mendez <jm@xxxxxx>
---
arch/arm64/boot/dts/ti/k3-am62a7-sk.dts | 96 +++++++++++++++++++++++--
1 file changed, 90 insertions(+), 6 deletions(-)
diff --git a/arch/arm64/boot/dts/ti/k3-am62a7-sk.dts b/arch/arm64/boot/dts/ti/k3-am62a7-sk.dts
index 1c9d95696c839..7d817b447c1d0 100644
--- a/arch/arm64/boot/dts/ti/k3-am62a7-sk.dts
+++ b/arch/arm64/boot/dts/ti/k3-am62a7-sk.dts
@@ -52,6 +52,42 @@ linux,cma {
linux,cma-default;
};
+ c7x_0_dma_memory_region: c7x-dma-memory@99800000 {
+ compatible = "shared-dma-pool";
+ reg = <0x00 0x99800000 0x00 0x100000>;
+ no-map;
+ };
+
+ c7x_0_memory_region: c7x-memory@99900000 {
+ compatible = "shared-dma-pool";
+ reg = <0x00 0x99900000 0x00 0xf00000>;
+ no-map;
+ };
+
+ mcu_r5fss0_core0_dma_memory_region: r5f-dma-memory@9b800000 {
+ compatible = "shared-dma-pool";
+ reg = <0x00 0x9b800000 0x00 0x100000>;
+ no-map;
+ };
+
+ mcu_r5fss0_core0_memory_region: r5f-dma-memory@9b900000 {
+ compatible = "shared-dma-pool";
+ reg = <0x00 0x9b900000 0x00 0xf00000>;
+ no-map;
+ };
+
+ wkup_r5fss0_core0_dma_memory_region: r5f-dma-memory@9c800000 {
+ compatible = "shared-dma-pool";
+ reg = <0x00 0x9c800000 0x00 0x100000>;
+ no-map;
+ };
+
+ wkup_r5fss0_core0_memory_region: r5f-dma-memory@9c900000 {
+ compatible = "shared-dma-pool";
+ reg = <0x00 0x9c900000 0x00 0xf00000>;
+ no-map;
+ };
+
secure_tfa_ddr: tfa@9e780000 {
reg = <0x00 0x9e780000 0x00 0x80000>;
alignment = <0x1000>;
@@ -63,12 +99,6 @@ secure_ddr: optee@9e800000 {
alignment = <0x1000>;
no-map;
};
-
- wkup_r5fss0_core0_memory_region: r5f-dma-memory@9c900000 {
- compatible = "shared-dma-pool";
- reg = <0x00 0x9c900000 0x00 0x01e00000>;
- no-map;
- };
};
This is missing the edgeAI specific remote-core carveouts and RTOS-to-RTOS IPC memory regions [1] being used by edgeAI firmwares which come as pre-packaged in the official SDK release for AM62A.
There is only one official SDK release for AM62A (which is edgeAI based) [2] which packages these edgeAI remoteproc firmwares and in my view it is a fair expectation that remote core careveouts in device-tree should align with firmwares released in SDK.
Why should Linux upstream care about what we do in our current evil vendor SDKs?
We change things around every cycle, and do all kinds of hacky things to just
"make it work" for the current SDK release.
This is because most developers (including me) and vendors download this official SDK release and use it with latest upstream kernel and modules (right now we are applying required patches locally) and this patch won't suffice for this, in-fact it won't work since the remoteproc firmwares are already using regions beyond the reserved-regions from this patch.
Then that firmware team should fix their firmware, Linux should not
do things just because some builds of a firmware did the wrong thing.
Just a random example from the top of my mind here, a while back
someone on the codec firmware team decided to take the standard RPMSG
name service structure and modify it to suit some specific usecase
they had, suddenly all the firmware we made for AM57x devices stopped
working on upstream kernels. Instead of fixing the firmware we just
carried a hack for the same in our vendor kernel trees. Now customers
have no path to use this old firmware on newer kernels as adding the
hack to upstream Linux (rightly) failed [0].
Let's not do that again, if we have firmware that doesn't follow
upstream, then let's fix the firmware, not hack around it upstream.
The edgeAI firmware folks have no issue ignoring existing upstream
IPC carveouts and simply replacing them all with their own custom
ones in their SDK[1], I see no reason they cannot continue doing that
if they don't want to fix their firmware.
I understand your point, currently with this patch remoteproc loading
will not work for some cores. However, the goal here is to standardize
as much as possible the memory carveout sizes, push the "demo firmware"
to request resources the correct way from resource table, and move away
from this dependency and limitations that we have with our firmware. We
should soon be able to generate our own firmware using Zephyr, which
Andrew is pioneering, so with this firmware we should move to the
correct direction upstream. Downstream we are still using the memory
carveout sizes that the firmware folk want so desperately to keep, for
now..
+1
I have this Zephyr based firmware for AM62A working and it uses the
standard IPC regions as specified in this patch. I'll be posting the PR
for it in Zephyr upstream by the end of week.
For this patch as it is:
Acked-by: Andrew Davis <afd@xxxxxx>
Andrew
[0] https://lore.kernel.org/lkml/20241011123922.23135-1-richard@xxxxxx/
[1] https://git.ti.com/cgit/edgeai/meta-edgeai/tree/recipes-kernel/linux/linux-ti-staging/j721e-evm/0001-arm64-dts-ti-Add-DTB-overlays-for-vision-apps-and-ed.patch?h=kirkstone
~ Judith
[1]: https://git.ti.com/cgit/ti-linux-kernel/ti-linux-kernel/tree/arch/arm64/boot/dts/ti/k3-am62a7-sk.dts?h=ti-linux-6.6.y-cicd#n103
[2]: https://www.ti.com/tool/PROCESSOR-SDK-AM62A
Regards
Devarsh
opp-table {
@@ -741,3 +771,57 @@ dpi1_out: endpoint {
};
};
};
+
+&mailbox0_cluster0 {
+ status = "okay";
+
+ mbox_r5_0: mbox-r5-0 {
+ ti,mbox-rx = <0 0 0>;
+ ti,mbox-tx = <1 0 0>;
+ };
+};
+
+&mailbox0_cluster1 {
+ status = "okay";
+
+ mbox_c7x_0: mbox-c7x-0 {
+ ti,mbox-rx = <0 0 0>;
+ ti,mbox-tx = <1 0 0>;
+ };
+};
+
+&mailbox0_cluster2 {
+ status = "okay";
+
+ mbox_mcu_r5_0: mbox-mcu-r5-0 {
+ ti,mbox-rx = <0 0 0>;
+ ti,mbox-tx = <1 0 0>;
+ };
+};
+
+&wkup_r5fss0 {
+ status = "okay";
+};
+
+&wkup_r5fss0_core0 {
+ mboxes = <&mailbox0_cluster0>, <&mbox_r5_0>;
+ memory-region = <&wkup_r5fss0_core0_dma_memory_region>,
+ <&wkup_r5fss0_core0_memory_region>;
+};
+
+&mcu_r5fss0 {
+ status = "okay";
+};
+
+&mcu_r5fss0_core0 {
+ mboxes = <&mailbox0_cluster2>, <&mbox_mcu_r5_0>;
+ memory-region = <&mcu_r5fss0_core0_dma_memory_region>,
+ <&mcu_r5fss0_core0_memory_region>;
+};
+
+&c7x_0 {
+ mboxes = <&mailbox0_cluster1>, <&mbox_c7x_0>;
+ memory-region = <&c7x_0_dma_memory_region>,
+ <&c7x_0_memory_region>;
+ status = "okay";
+};