Re: [PATCH 4/4] media: venus: add PIL support

[Vikash Garodia <vgarodia@codeaurora.org>]

Hi Stan,

On 2018-05-22 18:32, Stanimir Varbanov wrote:
> Hi Vikash,
> 
> On 05/17/2018 02:32 PM, Vikash Garodia wrote:
>> This adds support to load the video firmware
>> and bring ARM9 out of reset. This is useful
>> for platforms which does not have trustzone
>> to reset the ARM9.
>> 
>> Signed-off-by: Vikash Garodia <vgarodia@codeaurora.org>
>> ---
>>  .../devicetree/bindings/media/qcom,venus.txt       |   8 +-
>>  drivers/media/platform/qcom/venus/core.c           |  67 +++++++--
>>  drivers/media/platform/qcom/venus/core.h           |   6 +
>>  drivers/media/platform/qcom/venus/firmware.c       | 163 
>> +++++++++++++++++----
>>  drivers/media/platform/qcom/venus/firmware.h       |  10 +-
>>  5 files changed, 217 insertions(+), 37 deletions(-)
>> 
>> diff --git a/Documentation/devicetree/bindings/media/qcom,venus.txt 
>> b/Documentation/devicetree/bindings/media/qcom,venus.txt
>> index 00d0d1b..0ff0f2d 100644
>> --- a/Documentation/devicetree/bindings/media/qcom,venus.txt
>> +++ b/Documentation/devicetree/bindings/media/qcom,venus.txt
> 
> for this change in DT binding you have to cc devicetree ML. And 
> probably
> it could be separate patch.

Will keep it as a separate patch and add the DT reviewers.

>> @@ -53,7 +53,7 @@
>> 
>>  * Subnodes
>>  The Venus video-codec node must contain two subnodes representing
>> -video-decoder and video-encoder.
>> +video-decoder and video-encoder, one optional firmware subnode.
>> 
>>  Every of video-encoder or video-decoder subnode should have:
>> 
>> @@ -79,6 +79,8 @@ Every of video-encoder or video-decoder subnode 
>> should have:
>>  		    power domain which is responsible for collapsing
>>  		    and restoring power to the subcore.
>> 
>> +The firmware sub node must contain the iommus specifiers for ARM9.
>> +
>>  * An Example
>>  	video-codec@1d00000 {
>>  		compatible = "qcom,msm8916-venus";
>> @@ -105,4 +107,8 @@ Every of video-encoder or video-decoder subnode 
>> should have:
>>  			clock-names = "core";
>>  			power-domains = <&mmcc VENUS_CORE1_GDSC>;
>>  		};
>> +		firmware {
> 
> venus-firmware
Ok.

>> +			compatible = "qcom,venus-pil-no-tz";
> 
> this should be following the other subnodes compatible names:
> 
> compatible = "venus-firmware";

Probably "venus-firmware-no-tz". Want to keep "-no-tz" explicitly as 
this node
is not needed for TZ based PIL.

>> +			iommus = <&apps_smmu 0x10b2 0x0>;
>> +		}
>>  	};
>> diff --git a/drivers/media/platform/qcom/venus/core.c 
>> b/drivers/media/platform/qcom/venus/core.c
>> index 1308488..16910558 100644
>> --- a/drivers/media/platform/qcom/venus/core.c
>> +++ b/drivers/media/platform/qcom/venus/core.c
>> @@ -22,6 +22,7 @@
>>  #include <linux/slab.h>
>>  #include <linux/types.h>
>>  #include <linux/pm_runtime.h>
>> +#include <linux/iommu.h>
>>  #include <media/videobuf2-v4l2.h>
>>  #include <media/v4l2-mem2mem.h>
>>  #include <media/v4l2-ioctl.h>
>> @@ -30,6 +31,7 @@
>>  #include "vdec.h"
>>  #include "venc.h"
>>  #include "firmware.h"
>> +#include "hfi_venus.h"
>> 
>>  static void venus_event_notify(struct venus_core *core, u32 event)
>>  {
>> @@ -76,7 +78,7 @@ static void venus_sys_error_handler(struct 
>> work_struct *work)
>>  	hfi_core_deinit(core, true);
>>  	hfi_destroy(core);
>>  	mutex_lock(&core->lock);
>> -	venus_shutdown(core->dev);
>> +	venus_shutdown(core);
>> 
>>  	pm_runtime_put_sync(core->dev);
>> 
>> @@ -84,7 +86,7 @@ static void venus_sys_error_handler(struct 
>> work_struct *work)
>> 
>>  	pm_runtime_get_sync(core->dev);
>> 
>> -	ret |= venus_boot(core->dev, core->res->fwname);
>> +	ret |= venus_boot(core);
>> 
>>  	ret |= hfi_core_resume(core, true);
>> 
>> @@ -179,6 +181,20 @@ static u32 to_v4l2_codec_type(u32 codec)
>>  	}
>>  }
>> 
>> +static int store_firmware_dev(struct device *dev, void *data)
>> +{
>> +	struct venus_core *core;
>> +
>> +	core = (struct venus_core *)data;
>> +	if (!core)
>> +		return -EINVAL;
>> +
>> +	if (of_device_is_compatible(dev->of_node, "qcom,venus-pil-no-tz"))
>> +		core->fw.dev = dev;
>> +
>> +	return 0;
>> +}
>> +
>>  static int venus_enumerate_codecs(struct venus_core *core, u32 type)
>>  {
>>  	const struct hfi_inst_ops dummy_ops = {};
>> @@ -229,6 +245,7 @@ static int venus_probe(struct platform_device 
>> *pdev)
>>  	struct device *dev = &pdev->dev;
>>  	struct venus_core *core;
>>  	struct resource *r;
>> +	struct iommu_domain *iommu_domain;
>>  	int ret;
>> 
>>  	core = devm_kzalloc(dev, sizeof(*core), GFP_KERNEL);
>> @@ -279,7 +296,14 @@ static int venus_probe(struct platform_device 
>> *pdev)
>>  	if (ret < 0)
>>  		goto err_runtime_disable;
>> 
>> -	ret = venus_boot(dev, core->res->fwname);
>> +	ret = of_platform_populate(dev->of_node, NULL, NULL, dev);
>> +	if (ret)
>> +		goto err_runtime_disable;
>> +
>> +	/* Attempt to register child devices */
> 
> This comment is wrong, the child devices are created by
> of_platform_populate above.

Good catch. Intend was to mention something like "Attempt to store 
firmware
device". Will correct it.

>> +	ret = device_for_each_child(dev, core, store_firmware_dev);
> 
> Why we need these complex gymnastics to get struct device pointer when
> that could be done in venus_firmware .probe method?
> 
> I think the answer is because you want to avoid having 
> venus-firmware.ko
> (because you have to have separate struct device for iommu SID). In 
> that
> case it would be better to make venus-firmware.ko.

I can have the venus firmware .probe method without venus-firmware.ko. I 
had
the probe method initially, but since it was just storing the device 
pointer,
i am doing it while iterating over the child nodes.

>> +
>> +	ret = venus_boot(core);
>>  	if (ret)
>>  		goto err_runtime_disable;
>> 
>> @@ -303,14 +327,17 @@ static int venus_probe(struct platform_device 
>> *pdev)
>>  	if (ret)
>>  		goto err_core_deinit;
>> 
>> -	ret = of_platform_populate(dev->of_node, NULL, NULL, dev);
>> +	ret = pm_runtime_put_sync(dev);
>>  	if (ret)
>>  		goto err_dev_unregister;
>> 
>> -	ret = pm_runtime_put_sync(dev);
>> -	if (ret)
>> +	iommu_domain = iommu_get_domain_for_dev(dev);
>> +	if (!iommu_domain)
>>  		goto err_dev_unregister;
>> 
>> +	iommu_domain->geometry.aperture_start = VENUS_FW_MEM_SIZE;
>> +	iommu_domain->geometry.aperture_end = VENUS_MAX_MEM_REGION;
> 
> I don't think that is needed for this struct device (Venus DT node
> struct device). And also why aperture_start is on 6th MB? I think that
> this iommu domain is for venus_non_secure iommu context_bank.

ARM9 cannot accept iova as 0x0 for data buffers. The range is from 
firmware
end address(VENUS_FW_MEM_SIZE) till 3.5 GB. When the driver programs the
register for firmware start and end address, and provides a buffer 
having
iova in the firmware range, it would end up generating a different SID 
and
would lead to issues.

> Those geometry parameters are checked/used only from dma-iommu.c. They
> are checked before entering on venus_probe and only when
> geometry.force_aperture is true. So updating those params here doesn't
> make any sense to iommu?

I am not very sure on this part. We can stay with 
dma_set_mask_and_coherent
to keep the upper limit check. For lower limit, we can keep it as a TODO
for future.

>> +
>>  	return 0;
>> 
>>  err_dev_unregister:
>> @@ -318,7 +345,7 @@ static int venus_probe(struct platform_device 
>> *pdev)
>>  err_core_deinit:
>>  	hfi_core_deinit(core, false);
>>  err_venus_shutdown:
>> -	venus_shutdown(dev);
>> +	venus_shutdown(core);
>>  err_runtime_disable:
>>  	pm_runtime_set_suspended(dev);
>>  	pm_runtime_disable(dev);
>> @@ -339,7 +366,7 @@ static int venus_remove(struct platform_device 
>> *pdev)
>>  	WARN_ON(ret);
>> 
>>  	hfi_destroy(core);
>> -	venus_shutdown(dev);
>> +	venus_shutdown(core);
>>  	of_platform_depopulate(dev);
>> 
>>  	pm_runtime_put_sync(dev);
>> @@ -483,7 +510,29 @@ static __maybe_unused int 
>> venus_runtime_resume(struct device *dev)
>>  		.pm = &venus_pm_ops,
>>  	},
>>  };
>> -module_platform_driver(qcom_venus_driver);
>> +
>> +static int __init venus_init(void)
>> +{
>> +	int ret;
>> +
>> +	ret = platform_driver_register(&qcom_video_firmware_driver);
>> +	if (ret)
>> +		return ret;
> 
> I think that this shouldn't be here, it is clear that firmware loader
> code should be on separate device/driver (even outside of venus DT 
> node).
> 
This is needed to register the driver with platform bus. Otherwise iommu 
group
for firmware device will not be created and iommu_domain_alloc would 
fail.

>> +
>> +	ret = platform_driver_register(&qcom_venus_driver);
>> +	if (ret)
>> +		platform_driver_unregister(&qcom_video_firmware_driver);
>> +
>> +	return ret;
>> +}
>> +module_init(venus_init);
>> +
>> +static void __exit venus_exit(void)
>> +{
>> +	platform_driver_unregister(&qcom_venus_driver);
>> +	platform_driver_unregister(&qcom_video_firmware_driver);
>> +}
>> +module_exit(venus_exit);
>> 
>>  MODULE_ALIAS("platform:qcom-venus");
>>  MODULE_DESCRIPTION("Qualcomm Venus video encoder and decoder 
>> driver");
>> diff --git a/drivers/media/platform/qcom/venus/core.h 
>> b/drivers/media/platform/qcom/venus/core.h
>> index 85e66e2..68fc8af 100644
>> --- a/drivers/media/platform/qcom/venus/core.h
>> +++ b/drivers/media/platform/qcom/venus/core.h
>> @@ -80,6 +80,11 @@ struct venus_caps {
>>  	bool valid;
>>  };
>> 
>> +struct video_firmware {
>> +	struct device *dev;
>> +	dma_addr_t iova;
>> +	struct iommu_domain *iommu_domain;
>> +};
>>  /**
>>   * struct venus_core - holds core parameters valid for all instances
>>   *
>> @@ -124,6 +129,7 @@ struct venus_core {
>>  	struct device *dev;
>>  	struct device *dev_dec;
>>  	struct device *dev_enc;
>> +	struct video_firmware fw;
>>  	struct mutex lock;
>>  	struct list_head instances;
>>  	atomic_t insts_count;
>> diff --git a/drivers/media/platform/qcom/venus/firmware.c 
>> b/drivers/media/platform/qcom/venus/firmware.c
>> index 8f25375..614c805 100644
>> --- a/drivers/media/platform/qcom/venus/firmware.c
>> +++ b/drivers/media/platform/qcom/venus/firmware.c
>> @@ -12,8 +12,12 @@
>>   *
>>   */
>> 
>> +#include <linux/module.h>
>> +#include <linux/of_device.h>
>> +#include <linux/platform_device.h>
>>  #include <linux/device.h>
>>  #include <linux/firmware.h>
>> +#include <linux/iommu.h>
>>  #include <linux/delay.h>
>>  #include <linux/kernel.h>
>>  #include <linux/io.h>
>> @@ -27,8 +31,10 @@
>>  #include "firmware.h"
>>  #include "hfi_venus_io.h"
>> 
>> -#define VENUS_PAS_ID			9
>> -#define VENUS_FW_MEM_SIZE		(6 * SZ_1M)
>> +static const struct of_device_id firmware_dt_match[] = {
>> +	{ .compatible = "qcom,venus-pil-no-tz" },
>> +	{ }
>> +};
>> 
>>  void venus_reset_hw(struct venus_core *core)
>>  {
>> @@ -53,40 +59,37 @@ void venus_reset_hw(struct venus_core *core)
>>  	/* Bring Arm9 out of reset */
>>  	writel_relaxed(0, reg_base + WRAPPER_A9SS_SW_RESET);
>>  }
>> -int venus_boot(struct device *dev, const char *fwname)
>> +EXPORT_SYMBOL_GPL(venus_reset_hw);
>> +
>> +int venus_load_fw(struct device *dev, const char *fwname,
>> +		phys_addr_t *mem_phys, size_t *mem_size)
>>  {
>>  	const struct firmware *mdt;
>>  	struct device_node *node;
>> -	phys_addr_t mem_phys;
>>  	struct resource r;
>>  	ssize_t fw_size;
>> -	size_t mem_size;
>>  	void *mem_va;
>>  	int ret;
>> 
>> -	if (!IS_ENABLED(CONFIG_QCOM_MDT_LOADER) || !qcom_scm_is_available())
>> -		return -EPROBE_DEFER;
>> -
>>  	node = of_parse_phandle(dev->of_node, "memory-region", 0);
>>  	if (!node) {
>>  		dev_err(dev, "no memory-region specified\n");
>>  		return -EINVAL;
>>  	}
>> -
>>  	ret = of_address_to_resource(node, 0, &r);
>>  	if (ret)
>>  		return ret;
>> 
>> -	mem_phys = r.start;
>> -	mem_size = resource_size(&r);
>> +	*mem_phys = r.start;
>> +	*mem_size = resource_size(&r);
>> 
>> -	if (mem_size < VENUS_FW_MEM_SIZE)
>> +	if (*mem_size < VENUS_FW_MEM_SIZE)
>>  		return -EINVAL;
>> 
>> -	mem_va = memremap(r.start, mem_size, MEMREMAP_WC);
>> +	mem_va = memremap(r.start, *mem_size, MEMREMAP_WC);
>>  	if (!mem_va) {
>>  		dev_err(dev, "unable to map memory region: %pa+%zx\n",
>> -			&r.start, mem_size);
>> +			&r.start, *mem_size);
>>  		return -ENOMEM;
>>  	}
>> 
>> @@ -101,24 +104,134 @@ int venus_boot(struct device *dev, const char 
>> *fwname)
>>  		goto err_unmap;
>>  	}
>> 
>> -	ret = qcom_mdt_load(dev, mdt, fwname, VENUS_PAS_ID, mem_va, 
>> mem_phys,
>> -			    mem_size, NULL);
>> +	ret = qcom_mdt_load(dev, mdt, fwname, VENUS_PAS_ID, mem_va, 
>> *mem_phys,
>> +			    *mem_size, NULL);
>> 
>>  	release_firmware(mdt);
>> 
>> -	if (ret)
>> -		goto err_unmap;
>> -
>> -	ret = qcom_scm_pas_auth_and_reset(VENUS_PAS_ID);
>> -	if (ret)
>> -		goto err_unmap;
>> -
>>  err_unmap:
>>  	memunmap(mem_va);
>>  	return ret;
>>  }
>> 
>> -int venus_shutdown(struct device *dev)
>> +int venus_boot_noTZ(struct venus_core *core, phys_addr_t mem_phys,
>> +							size_t mem_size)
>>  {
>> -	return qcom_scm_pas_shutdown(VENUS_PAS_ID);
>> +	struct iommu_domain *iommu;
>> +	struct device *dev;
>> +	int ret;
>> +
>> +	if (!core->fw.dev)
>> +		return -EPROBE_DEFER;
>> +
>> +	dev = core->fw.dev;
>> +
>> +	iommu = iommu_domain_alloc(&platform_bus_type);
>> +	if (!iommu) {
>> +		dev_err(dev, "Failed to allocate iommu domain\n");
>> +		return -ENOMEM;
>> +	}
>> +
>> +	iommu->geometry.aperture_start = 0x0;
>> +	iommu->geometry.aperture_end = VENUS_FW_MEM_SIZE;
> 
> The same comment for geometry params as for venus_probe is valid here.
As this is used only for firmware context bank, i can remove these
explicit iommu configuration.

>> +
>> +	ret = iommu_attach_device(iommu, dev);
>> +	if (ret) {
>> +		dev_err(dev, "could not attach device\n");
>> +		goto err_attach;
>> +	}
>> +
>> +	ret = iommu_map(iommu, core->fw.iova, mem_phys, mem_size,
>> +			IOMMU_READ|IOMMU_WRITE|IOMMU_PRIV);
> 
> iova is not initialized and is zero, maybe we don't need that variable
> in the venus_firmware structure?

As the iova will be always zero here, i can hard-code as well.

>> +	if (ret) {
>> +		dev_err(dev, "could not map video firmware region\n");
>> +		goto err_map;
>> +	}
>> +	core->fw.iommu_domain = iommu;
>> +	venus_reset_hw(core);
>> +
>> +	return 0;
>> +
>> +err_map:
>> +	iommu_detach_device(iommu, dev);
>> +err_attach:
>> +	iommu_domain_free(iommu);
>> +	return ret;
>>  }
>> +
>> +int venus_shutdown_noTZ(struct venus_core *core)
>> +{
>> +	struct iommu_domain *iommu;
>> +	u32 reg;
>> +	struct device *dev = core->fw.dev;
>> +	void __iomem *reg_base = core->base;
>> +
>> +	/* Assert the reset to ARM9 */
>> +	reg = readl_relaxed(reg_base + WRAPPER_A9SS_SW_RESET);
>> +	reg |= BIT(4);
>> +	writel_relaxed(reg, reg_base + WRAPPER_A9SS_SW_RESET);
>> +
>> +	/* Make sure reset is asserted before the mapping is removed */
>> +	mb();
>> +
>> +	iommu = core->fw.iommu_domain;
>> +
>> +	iommu_unmap(iommu, core->fw.iova, VENUS_FW_MEM_SIZE);
>> +	iommu_detach_device(iommu, dev);
>> +	iommu_domain_free(iommu);
> 
> check iommu APIs for errors.
Ok.