Re: [PATCHv4 06/14] remoteproc/omap: Initialize and assign reserved memory node

From: Tero Kristo
Date: Tue Jan 07 2020 - 09:25:19 EST

On 07/01/2020 15:37, Andrew F. Davis wrote:
On 1/2/20 8:18 AM, Tero Kristo wrote:
From: Suman Anna <s-anna@xxxxxx>

The reserved memory nodes are not assigned to platform devices by
default in the driver core to avoid the lookup for every platform
device and incur a penalty as the real users are expected to be
only a few devices.

OMAP remoteproc devices fall into the above category and the OMAP
remoteproc driver _requires_ specific CMA pools to be assigned
for each device at the moment to align on the location of the
vrings and vring buffers in the RTOS-side firmware images. So,

Requires only at the moment due to firmware..

This sounds like some firmware images hard-coded their vring addresses
instead of getting them dynamically as they should and we are hacking
around that on the kernel side by giving them the addresses they use as

The firmwares are built on specific device addresses, this includes data + code.

Should we rather make use of the IOMMU attached to the DSP to map any
kernel address to the DSP where the firmware expects it? Or better yet
fixup the firmwares.

Well, we do use IOMMU to map the corresponding memory area to specific device address. What this patch does, is to allocate a contiguous memory area for the remoteproc shared memories. Using completely random memory location would potentially fragment the remoteproc memory around page boundaries, resulting in a complex (read ineffective) IOMMU mapping. Also, we are going to need the reserved memory mechanism for the remoteproc anyways later, as we are going to introduce the support for early-boot / late-attach. Bootloader would pass the used memory regions to the kernel via the reserved memory nodes in that case (unless we assume to use some hardcoded region, which would be worse than passing it via DT.)


DRAM carveouts should be a last resort used only for when hardware
really requires it.


use the of_reserved_mem_device_init/release() API appropriately
to assign the corresponding reserved memory region to the OMAP
remoteproc device. Note that only one region per device is
allowed by the framework.

Signed-off-by: Suman Anna <s-anna@xxxxxx>
Signed-off-by: Tero Kristo <t-kristo@xxxxxx>
Reviewed-by: Bjorn Andersson <bjorn.andersson@xxxxxxxxxx>
drivers/remoteproc/omap_remoteproc.c | 12 +++++++++++-
1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/drivers/remoteproc/omap_remoteproc.c b/drivers/remoteproc/omap_remoteproc.c
index 9ca337f18ac2..8a6dd742a8b1 100644
--- a/drivers/remoteproc/omap_remoteproc.c
+++ b/drivers/remoteproc/omap_remoteproc.c
@@ -17,6 +17,7 @@
#include <linux/module.h>
#include <linux/err.h>
#include <linux/of_device.h>
+#include <linux/of_reserved_mem.h>
#include <linux/platform_device.h>
#include <linux/dma-mapping.h>
#include <linux/remoteproc.h>
@@ -480,14 +481,22 @@ static int omap_rproc_probe(struct platform_device *pdev)
if (ret)
goto free_rproc;
+ ret = of_reserved_mem_device_init(&pdev->dev);
+ if (ret) {
+ dev_err(&pdev->dev, "device does not have specific CMA pool\n");
+ goto free_rproc;
+ }
platform_set_drvdata(pdev, rproc);
ret = rproc_add(rproc);
if (ret)
- goto free_rproc;
+ goto release_mem;
return 0;
+ of_reserved_mem_device_release(&pdev->dev);
return ret;
@@ -499,6 +508,7 @@ static int omap_rproc_remove(struct platform_device *pdev)
+ of_reserved_mem_device_release(&pdev->dev);
return 0;

Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki. Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki