Re: [PATCH 4/8] media: venus: migrate firmware loading to qcom_mdt_pas_load() and qcom_pas_prepare_and_auth_reset()

[Gourav Kumar <gourav.kumar@oss.qualcomm.com>]

On 9/25/2026 9:27 PM, Mukesh Ojha wrote:
> venus_load_fw() uses qcom_mdt_load() for the TZ path and a bare
> memremap()/qcom_mdt_load_no_init()/memunmap() sequence for the non-TZ
> path.  venus_boot() then calls qcom_pas_auth_and_reset() which skips the
> shmbridge prepare step required before TrustZone authentication.
> 
> For the TZ path, add a pas_ctx pointer to struct video_firmware so the
> context can be threaded from venus_load_fw() where it is allocated to
> venus_boot() where authentication happens.  Replace qcom_mdt_load() with
> devm_qcom_pas_context_alloc() and qcom_mdt_pas_load(), storing the
> context in core->fw.pas_ctx, then pass it to
> qcom_pas_prepare_and_auth_reset() in venus_boot().
> 
> For the non-TZ path, pass NULL to qcom_mdt_load_no_init() directly,
> removing the explicit memremap()/memunmap() since qcom_mdt_load_no_init()
> handles the mapping internally when mem_region is NULL.

That is only true after patch 7/8, which adds the NULL mem_region
handling to qcom_mdt_load_no_init(). Until then a NULL mem_region is
rejected, so the !use_tz path fails between patches 4 and 7, which
breaks bisection. I would suggest rearranging/splitting the patches so
the qcom_mdt_load_no_init() change comes before this one.

Gourav Kumar
> 
> Signed-off-by: Mukesh Ojha <mukesh.ojha@oss.qualcomm.com>
> ---
>  drivers/media/platform/qcom/venus/core.h     |  2 ++
>  drivers/media/platform/qcom/venus/firmware.c | 29 +++++++++-----------
>  2 files changed, 15 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/media/platform/qcom/venus/core.h b/drivers/media/platform/qcom/venus/core.h
> index 46705a666776..adcf67043bf6 100644
> --- a/drivers/media/platform/qcom/venus/core.h
> +++ b/drivers/media/platform/qcom/venus/core.h
> @@ -9,6 +9,7 @@
>  
>  #include <linux/bitops.h>
>  #include <linux/list.h>
> +#include <linux/firmware/qcom/qcom_pas.h>
>  #include <media/videobuf2-v4l2.h>
>  #include <media/v4l2-ctrls.h>
>  #include <media/v4l2-device.h>
> @@ -223,6 +224,7 @@ struct venus_core {
>  		size_t mapped_mem_size;
>  		phys_addr_t mem_phys;
>  		size_t mem_size;
> +		struct qcom_pas_context *pas_ctx;
>  	} fw;
>  	struct mutex lock;
>  	struct list_head instances;
> diff --git a/drivers/media/platform/qcom/venus/firmware.c b/drivers/media/platform/qcom/venus/firmware.c
> index 3a38ff985822..4ea4c7769ac3 100644
> --- a/drivers/media/platform/qcom/venus/firmware.c
> +++ b/drivers/media/platform/qcom/venus/firmware.c
> @@ -86,7 +86,6 @@ static int venus_load_fw(struct venus_core *core, const char *fwname,
>  	struct resource res;
>  	struct device *dev;
>  	ssize_t fw_size;
> -	void *mem_va;
>  	int ret;
>  
>  	*mem_phys = 0;
> @@ -117,21 +116,19 @@ static int venus_load_fw(struct venus_core *core, const char *fwname,
>  		goto err_release_fw;
>  	}
>  
> -	mem_va = memremap(*mem_phys, *mem_size, MEMREMAP_WC);
> -	if (!mem_va) {
> -		dev_err(dev, "unable to map memory region %pa size %#zx\n", mem_phys, *mem_size);
> -		ret = -ENOMEM;
> -		goto err_release_fw;
> -	}
> -
> -	if (core->use_tz)
> -		ret = qcom_mdt_load(dev, mdt, fwname, VENUS_PAS_ID,
> -				    mem_va, *mem_phys, *mem_size, NULL);
> -	else
> -		ret = qcom_mdt_load_no_init(dev, mdt, fwname, mem_va,
> +	if (core->use_tz) {
> +		core->fw.pas_ctx = devm_qcom_pas_context_alloc(dev, VENUS_PAS_ID,
> +							       *mem_phys, *mem_size);
> +		if (IS_ERR(core->fw.pas_ctx)) {
> +			ret = PTR_ERR(core->fw.pas_ctx);
> +			core->fw.pas_ctx = NULL;
> +			goto err_release_fw;
> +		}
> +		ret = qcom_mdt_pas_load(core->fw.pas_ctx, mdt, fwname, NULL);

qcom_mdt_pas_load() passes the ctx to qcom_pas_init_image(), which keeps
the metadata buffer in ctx->ptr, and qcom_pas_prepare_and_auth_reset()
does not free it. Who is expected to call qcom_pas_metadata_release()
for venus?

Gourav Kumar

> +	} else {
> +		ret = qcom_mdt_load_no_init(dev, mdt, fwname, NULL,
>  					    *mem_phys, *mem_size, NULL);
> -
> -	memunmap(mem_va);
> +	}
>  err_release_fw:
>  	release_firmware(mdt);
>  	return ret;
> @@ -237,7 +234,7 @@ int venus_boot(struct venus_core *core)
>  	core->fw.mem_phys = mem_phys;
>  
>  	if (core->use_tz)
> -		ret = qcom_pas_auth_and_reset(VENUS_PAS_ID);
> +		ret = qcom_pas_prepare_and_auth_reset(core->fw.pas_ctx);

ctx->use_tzmem is not set for venus, so the shmbridge creation in
qcom_scm_pas_prepare_and_auth_reset() is skipped and the behaviour is
the same as the old qcom_pas_auth_and_reset(). Please reword the commit
message accordingly, since it currently implies that the prepare step is
now performed by the new API as a fix.

Gourav Kumar

>  	else
>  		ret = venus_boot_no_tz(core, mem_phys, mem_size);
>