[RFC PATCH] remoteproc: imx_dsp_rproc: Fix multiple start/stop commands

[RFC PATCH] remoteproc: imx_dsp_rproc: Fix multiple start/stop commands

[Daniel Baluta]

After commit 67a7bc7f0358 ("remoteproc: Use of reserved_mem_region_*
functions for "memory-region"") following commands with
imx-dsp-rproc started to fail:

$ echo zephyr.elf > /sys/class/remoteproc/remoteproc0/firmware
$ echo start > /sys/class/remoteproc/remoteproc0/state
$ echo stop > /sys/class/remoteproc/remoteproc0/state
$ echo start > /sys/class/remoteproc/remoteproc0/state #! This fails
-sh: echo: write error: Device or resource busy

This happens because aforementioned commit replaced devm_ioremap_wc with
devm_ioremap_resource_wc which will "reserve" the memory region with the
first start and then will fail at the second start if the memory
region is already reserved.

So prefer using devm_ioremap_wc as there is no easy way to undo
devm_iormep_resource_wc manually.

Fixes: 67a7bc7f0358 ("remoteproc: Use of_reserved_mem_region_* functions for "memory-region"")
Signed-off-by: Daniel Baluta <daniel.baluta@nxp.com>
---
 drivers/remoteproc/imx_dsp_rproc.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/remoteproc/imx_dsp_rproc.c b/drivers/remoteproc/imx_dsp_rproc.c
index 5130a35214c9..79d5c15319f3 100644
--- a/drivers/remoteproc/imx_dsp_rproc.c
+++ b/drivers/remoteproc/imx_dsp_rproc.c
@@ -709,10 +709,10 @@ static int imx_dsp_rproc_add_carveout(struct imx_dsp_rproc *priv)
 		if (imx_dsp_rproc_sys_to_da(priv, res.start, resource_size(&res), &da))
 			return -EINVAL;
 
-		cpu_addr = devm_ioremap_resource_wc(dev, &res);
-		if (IS_ERR(cpu_addr)) {
+		cpu_addr = devm_ioremap_wc(dev, res.start, resource_size(&res));
+		if (!cpu_addr) {
 			dev_err(dev, "failed to map memory %pR\n", &res);
-			return PTR_ERR(cpu_addr);
+			return -ENOMEM;
 		}
 
 		/* Register memory region */
-- 
2.45.2

Re: [RFC PATCH] remoteproc: imx_dsp_rproc: Fix multiple start/stop commands

[Rob Herring]

On Tue, Dec 09, 2025 at 04:04:25PM +0200, Daniel Baluta wrote:
> After commit 67a7bc7f0358 ("remoteproc: Use of reserved_mem_region_*
> functions for "memory-region"") following commands with
> imx-dsp-rproc started to fail:
>
> $ echo zephyr.elf > /sys/class/remoteproc/remoteproc0/firmware
> $ echo start > /sys/class/remoteproc/remoteproc0/state
> $ echo stop > /sys/class/remoteproc/remoteproc0/state
> $ echo start > /sys/class/remoteproc/remoteproc0/state #! This fails
> -sh: echo: write error: Device or resource busy

So unlike Marek's case, the first time works. Can you confirm your 
region is fixed address?

> This happens because aforementioned commit replaced devm_ioremap_wc with
> devm_ioremap_resource_wc which will "reserve" the memory region with the
> first start and then will fail at the second start if the memory
> region is already reserved.
> 
> So prefer using devm_ioremap_wc as there is no easy way to undo
> devm_iormep_resource_wc manually.
> 
> Fixes: 67a7bc7f0358 ("remoteproc: Use of_reserved_mem_region_* functions for "memory-region"")
> Signed-off-by: Daniel Baluta <daniel.baluta@nxp.com>
> ---
>  drivers/remoteproc/imx_dsp_rproc.c | 6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/remoteproc/imx_dsp_rproc.c b/drivers/remoteproc/imx_dsp_rproc.c
> index 5130a35214c9..79d5c15319f3 100644
> --- a/drivers/remoteproc/imx_dsp_rproc.c
> +++ b/drivers/remoteproc/imx_dsp_rproc.c
> @@ -709,10 +709,10 @@ static int imx_dsp_rproc_add_carveout(struct imx_dsp_rproc *priv)
>  		if (imx_dsp_rproc_sys_to_da(priv, res.start, resource_size(&res), &da))
>  			return -EINVAL;
>  
> -		cpu_addr = devm_ioremap_resource_wc(dev, &res);
> -		if (IS_ERR(cpu_addr)) {
> +		cpu_addr = devm_ioremap_wc(dev, res.start, resource_size(&res));

Best case this is reusing the same mapping and we just have unnecessary 
ioremap and iounmap calls and devm entries. Worst case, we get a new 
virtual mapping every time. IIRC, arm32 will reuse existing mapping, but 
arm64 does not. But that's 10+ years ago I looked into it.

Seems like devres is not the right lifetime. This should be just 
ioremap_wc() instead.

Rob

Re: [RFC PATCH] remoteproc: imx_dsp_rproc: Fix multiple start/stop commands

[Iuliana Prodan]

On 12/9/2025 4:04 PM, Daniel Baluta wrote:
> After commit 67a7bc7f0358 ("remoteproc: Use of reserved_mem_region_*
> functions for "memory-region"") following commands with
> imx-dsp-rproc started to fail:
> 
> $ echo zephyr.elf > /sys/class/remoteproc/remoteproc0/firmware
> $ echo start > /sys/class/remoteproc/remoteproc0/state
> $ echo stop > /sys/class/remoteproc/remoteproc0/state
> $ echo start > /sys/class/remoteproc/remoteproc0/state #! This fails
> -sh: echo: write error: Device or resource busy
> 
> This happens because aforementioned commit replaced devm_ioremap_wc with
> devm_ioremap_resource_wc which will "reserve" the memory region with the
> first start and then will fail at the second start if the memory
> region is already reserved.
> 
> So prefer using devm_ioremap_wc as there is no easy way to undo
> devm_iormep_resource_wc manually.
> 
> Fixes: 67a7bc7f0358 ("remoteproc: Use of_reserved_mem_region_* functions for "memory-region"")
> Signed-off-by: Daniel Baluta <daniel.baluta@nxp.com>
> ---
>   drivers/remoteproc/imx_dsp_rproc.c | 6 +++---
>   1 file changed, 3 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/remoteproc/imx_dsp_rproc.c b/drivers/remoteproc/imx_dsp_rproc.c
> index 5130a35214c9..79d5c15319f3 100644
> --- a/drivers/remoteproc/imx_dsp_rproc.c
> +++ b/drivers/remoteproc/imx_dsp_rproc.c
> @@ -709,10 +709,10 @@ static int imx_dsp_rproc_add_carveout(struct imx_dsp_rproc *priv)
>   		if (imx_dsp_rproc_sys_to_da(priv, res.start, resource_size(&res), &da))
>   			return -EINVAL;
>   
> -		cpu_addr = devm_ioremap_resource_wc(dev, &res);
> -		if (IS_ERR(cpu_addr)) {
> +		cpu_addr = devm_ioremap_wc(dev, res.start, resource_size(&res));

While this patch technically "works", it’s more of a workaround because:
- it removes important resource protection;
- it doesn’t address the root cause.

The real issue is likely that:
- resources should be released on stop but aren’t;
- or the mapping should happen once during probe, not on every start.

Better alterntives:
1. move the devm_ioremap_resource_wc() to probe() and keep it mapped for 
the device lifetime. This is the typical pattern for remoteproc drivers.
2. use non-devm versions in start/stop callbacks:
/* In start callback */
cpu_addr = ioremap_wc(res.start, resource_size(&res));
/* In stop callback */
iounmap(cpu_addr);

My vote goes for option 1.

> +		if (!cpu_addr) {
>   			dev_err(dev, "failed to map memory %pR\n", &res);
> -			return PTR_ERR(cpu_addr);
> +			return -ENOMEM;
>   		}
>   
>   		/* Register memory region */

Re: [RFC PATCH] remoteproc: imx_dsp_rproc: Fix multiple start/stop commands

[Daniel Baluta]

On Tue, Dec 9, 2025 at 5:27 PM Rob Herring <robh@kernel.org> wrote:
>
> On Tue, Dec 09, 2025 at 04:04:25PM +0200, Daniel Baluta wrote:
> > After commit 67a7bc7f0358 ("remoteproc: Use of reserved_mem_region_*
> > functions for "memory-region"") following commands with
> > imx-dsp-rproc started to fail:
> >
> > $ echo zephyr.elf > /sys/class/remoteproc/remoteproc0/firmware
> > $ echo start > /sys/class/remoteproc/remoteproc0/state
> > $ echo stop > /sys/class/remoteproc/remoteproc0/state
> > $ echo start > /sys/class/remoteproc/remoteproc0/state #! This fails
> > -sh: echo: write error: Device or resource busy
>
> So unlike Marek's case, the first time works. Can you confirm your
> region is fixed address?

Yes, all regions are fixed addresses.

> > +             cpu_addr = devm_ioremap_wc(dev, res.start, resource_size(&res));
>
> Best case this is reusing the same mapping and we just have unnecessary
> ioremap and iounmap calls and devm entries. Worst case, we get a new
> virtual mapping every time. IIRC, arm32 will reuse existing mapping, but
> arm64 does not. But that's 10+ years ago I looked into it.
>
> Seems like devres is not the right lifetime. This should be just
> ioremap_wc() instead.

Yes, looks like for arm64 (imx8mp-evk) will not reuse existing mapping
but creates a new one.

Best course of action would be as Iulia suggested to move the operation
at probe instead of prepare.