Re: [PATCH 1/2] media: iris: Fix iova allocation from restrict region

[Bryan O'Donoghue <bod@kernel.org>]

On 12/08/2026 11:55, Vishnu Reddy wrote:
> The VPU issues DMA through several SMMU streams, and the hardware does
> not give every stream the same addressable range. The non-pixel stream
> cannot address the low 600MB of IOVA space, while the pixel stream can
> address the full range:
>      +-----------------------------------------------------------+
>      | non-pixel stream addressable range (600 MB - 3.5 GB)      |
>      | 0x25800000 - 0xe0000000                                   |
>      +-----------------------------------------------------------+
>      | pixel stream addressable range (0 - 3.5 GB)               |
>      | 0x00000000 - 0xe0000000                                   |
>      +-----------------------------------------------------------+
> A single "iommus" property on the video-codec node puts every stream
> in one IOMMU domain sharing one IOVA allocator, so nothing restricts a
> non-pixel buffer to avoid 0 to 600MB. Once an allocation lands below that
> boundary the hardware faults, which shows up as unhandled SMMU page
> faults and spontaneous reboots.
> https://gitlab.freedesktop.org/drm/msm/-/work_items/100
> 
> A series to reserve the 0-600MB IOVA range via "iommu-addresses" was
> already posted here:
> https://lore.kernel.org/all/20260807-iris_iova_600mb_fix-v1-0-3996f67e33f9@oss.qualcomm.com

Something for a cover letter not a commit log - ongoing debates about 
how to change the behaviour will be irrelevant in 15 years after this 
stuff has landed.

The commit log should

- State the problem
- State the fix

Additional narrative about other series is for the series overview not 
the commit log.

> Those changes involve DT binding and DT node changes, and discussion is
> still ongoing on how to handle those for stable and for the upcoming
> sub-node design, with no conclusion reached yet. Thereby a critical reset
> issue is still open.

Drop.

> This is an alternate solution to fix the unhandled SMMU page fault

Kill this "alternative solution" stuff - if this lands in mainline then 
this _is_ the solution unless/until it is superseded.

> by restricting the IOVA range in the video driver, which also makes it
> easier and faster to land on mainline and stable kernels. At the same
> time the patch only reserves in the IOVA space without allocating
> any physical memory.

You should specify how you are making that restriction in the commit log.

Reading code...

OK you reserve below the boundary. Please add that to the commit log, 
its the salient piece of information.
> 
> Currently sub-nodes are not yet present, and only a single device is
> available, so the restriction is applied to both non-pixel and pixel
> stream IDs. This makes the solution unoptimal while fixing the issue
> considering all scenarios.
> Once sub-nodes for non-pixel, pixel, and secure streams become available,
> the restriction can be made stream specific.

Yeah all good information for the overview but if someone lands from 
Trantor in 10,000 years to pick through the detritus of our civilisation 
finding an old computer with kernel version 55.20 running on it and no 
sign of venus sub-nodes they might march off to the next planet to try 
to find them.

Drop the sub-nodes discussion from your commit log. Clearly state the 
problem and its remediation in your log, no need to reference ongoing 
bikeshedding elsewhere.

> 
> Fixes: d7378f84e94e ("media: iris: introduce iris core state management with shared queues")
> Cc: stable@vger.kernel.org
> Signed-off-by: Vishnu Reddy <busanna.reddy@oss.qualcomm.com>
> ---
>   drivers/media/platform/qcom/iris/iris_core.h  |  6 +++
>   drivers/media/platform/qcom/iris/iris_probe.c | 69 ++++++++++++++++++++++++++-
>   2 files changed, 74 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/media/platform/qcom/iris/iris_core.h b/drivers/media/platform/qcom/iris/iris_core.h
> index 24da60448cf2..79bc342a25a2 100644
> --- a/drivers/media/platform/qcom/iris/iris_core.h
> +++ b/drivers/media/platform/qcom/iris/iris_core.h
> @@ -7,6 +7,7 @@
>   #define __IRIS_CORE_H__
> 
>   #include <linux/types.h>
> +#include <linux/dma-mapping.h>
>   #include <linux/pm_domain.h>
>   #include <media/v4l2-device.h>
> 
> @@ -25,6 +26,9 @@ struct icc_info {
>   #define IRIS_FW_VERSION_LENGTH		128
>   #define IFACEQ_CORE_PKT_SIZE		(1024 * 4)
> 
> +#define IRIS_NP_RESERVE_IOVA_START	0x0
> +#define IRIS_NP_RESERVE_IOVA_SIZE	0x25800000
> +
>   enum domain_type {
>   	ENCODER	= BIT(0),
>   	DECODER	= BIT(1),
> @@ -77,6 +81,7 @@ struct qcom_ubwc_cfg_data;
>    * @instances: a list_head of all instances
>    * @inst_fw_caps_dec: an array of supported instance capabilities by decoder
>    * @inst_fw_caps_enc: an array of supported instance capabilities by encoder
> + * @iova_state: an array of dma_iova_state entries reserved for the restricted IOVA region
>    */
> 
>   struct iris_core {
> @@ -123,6 +128,7 @@ struct iris_core {
>   	/* encoder and decoder have overlapping caps, so two different arrays are required */
>   	struct platform_inst_fw_cap		inst_fw_caps_dec[INST_FW_CAP_MAX];
>   	struct platform_inst_fw_cap		inst_fw_caps_enc[INST_FW_CAP_MAX];
> +	struct dma_iova_state			*iova_state;
>   };
> 
>   int iris_core_init(struct iris_core *core);
> diff --git a/drivers/media/platform/qcom/iris/iris_probe.c b/drivers/media/platform/qcom/iris/iris_probe.c
> index e4acf4a74f94..f44305ee81b6 100644
> --- a/drivers/media/platform/qcom/iris/iris_probe.c
> +++ b/drivers/media/platform/qcom/iris/iris_probe.c
> @@ -150,6 +150,64 @@ static int iris_init_resources(struct iris_core *core)
>   	return iris_init_resets(core);
>   }
> 
> +static int iris_reserve_iova_region(struct device *dev, struct dma_iova_state **iova_state,
> +				    unsigned long start, unsigned long size)
> +{
> +	unsigned long mask = dma_get_mask(dev);
> +	unsigned long end, rem, chunk;
> +	struct dma_iova_state *state;
> +	unsigned int count = 0;
> +	int ret;
> +
> +	state = kcalloc(BITS_PER_TYPE(dma_addr_t) + 1, sizeof(*state), GFP_KERNEL);
> +	if (!state)
> +		return -ENOMEM;

devm_kcalloc() is less work.

> +
> +	end = start + size;
> +	rem = end - max(start, PAGE_SIZE);
> +
> +	ret = dma_set_mask_and_coherent(dev, end - 1);
> +	if (ret)
> +		goto err_free_mem;


I believe you should set dev->bus_dma_limit instead so drop this mask 
operation.

drivers/ata/ahci.c:	* bogus, platform code should use dev->bus_dma_limit 
instead..

=>

dev->bus_dma_limit = IRIS_NP_RESERVE_IOVA_SIZE -1;

> +
> +	while (rem) {
> +		chunk = min(end & -end, (u64)1 << (fls64(rem) - 1));

> +		if (!dma_iova_try_alloc(dev, &state[count], 0, chunk)) {
> +			ret = -ENOMEM;
> +			goto err_free_iova;
> +		}
> +

check that state[count].addr == end - chunk
error out if it does not.

> +		rem -= chunk;
> +		end -= chunk;
> +		count++;
> +	}
> +
> +	*iova_state = state;
> +	dma_set_mask_and_coherent(dev, mask);
> +
> +	return 0;
> +
> +err_free_iova:
> +	while (count--)
> +		dma_iova_free(dev, &state[count]);
> +	dma_set_mask_and_coherent(dev, mask);
\n> +err_free_mem:
> +	kfree(state);
> +
> +	return ret;


> +}
> +
> +static void iris_unreserve_iova_region(struct device *dev, struct dma_iova_state *iova_state)
> +{
> +	unsigned int i;
> +
> +	for (i = 0; dma_iova_size(&iova_state[i]); i++)
> +		dma_iova_free(dev, &iova_state[i]);
> +
> +	kfree(iova_state);
> +}
> +
>   static int iris_register_video_device(struct iris_core *core, enum domain_type type)
>   {
>   	struct video_device *vdev;
> @@ -207,6 +265,8 @@ static void iris_remove(struct platform_device *pdev)
> 
>   	v4l2_device_unregister(&core->v4l2_dev);
> 
> +	iris_unreserve_iova_region(core->dev, core->iova_state);
> +
>   	mutex_destroy(&core->lock);
>   }
> 
> @@ -292,14 +352,21 @@ static int iris_probe(struct platform_device *pdev)
>   	dma_set_max_seg_size(&pdev->dev, DMA_BIT_MASK(32));
>   	dma_set_seg_boundary(&pdev->dev, DMA_BIT_MASK(32));
> 
> +	ret = iris_reserve_iova_region(dev, &core->iova_state, IRIS_NP_RESERVE_IOVA_START,
> +				       IRIS_NP_RESERVE_IOVA_SIZE);
> +	if (ret)
> +		goto err_vdev_unreg_enc;
> +

This reservation should come before video_register_device() since its 
possible for user-space to race this otherwise.

Should come pretty much up the top

>   	pm_runtime_set_autosuspend_delay(core->dev, AUTOSUSPEND_DELAY_VALUE);
>   	pm_runtime_use_autosuspend(core->dev);
>   	ret = devm_pm_runtime_enable(core->dev);
>   	if (ret)
> -		goto err_vdev_unreg_enc;
> +		goto err_unresv_iova_region;
> 
>   	return 0;
> 
> +err_unresv_iova_region:
> +	iris_unreserve_iova_region(dev, core->iova_state);
>   err_vdev_unreg_enc:
>   	video_unregister_device(core->vdev_enc);
>   err_vdev_unreg_dec:
> 
> --
> 2.34.1
> 

---
bod