Linux kernel and device drivers for NXP i.MX platforms
 help / color / mirror / Atom feed
From: "Ming Qian(OSS)" <ming.qian@oss.nxp.com>
To: Nicolas Dufresne <nicolas@ndufresne.ca>, linux-media@vger.kernel.org
Cc: mchehab@kernel.org, hverkuil-cisco@xs4all.nl,
	benjamin.gaignard@collabora.com, p.zabel@pengutronix.de,
	sebastian.fricke@collabora.com, shawnguo@kernel.org,
	ulf.hansson@linaro.org, s.hauer@pengutronix.de,
	kernel@pengutronix.de, festevam@gmail.com, linux-imx@nxp.com,
	l.stach@pengutronix.de, Frank.li@nxp.com, peng.fan@nxp.com,
	eagle.zhou@nxp.com, imx@lists.linux.dev,
	linux-pm@vger.kernel.org, linux-kernel@vger.kernel.org,
	linux-arm-kernel@lists.infradead.org
Subject: Re: [PATCH v2 2/2] media: verisilicon: Avoid G2 bus error while decoding H.264 and HEVC
Date: Mon, 1 Dec 2025 09:52:19 +0800	[thread overview]
Message-ID: <64ad3d66-4edf-4a0b-8992-0db0e90a32b0@oss.nxp.com> (raw)
In-Reply-To: <46e9a5a881946d879d1b2af3421d574b26256ae3.camel@ndufresne.ca>

Hi Nicolas,

On 11/29/2025 5:28 AM, Nicolas Dufresne wrote:
> Hi,
> 
> Le vendredi 28 novembre 2025 à 10:51 +0800, ming.qian@oss.nxp.com a écrit :
>> From: Ming Qian <ming.qian@oss.nxp.com>
>>
>> For the i.MX8MQ platform, there is a hardware limitation: the g1 VPU and
>> g2 VPU cannot decode simultaneously; otherwise, it will cause below bus
>> error and produce corrupted pictures, even led to system hang.
>>
>> [  110.527986] hantro-vpu 38310000.video-codec: frame decode timed out.
>> [  110.583517] hantro-vpu 38310000.video-codec: bus error detected.
>>
>> Therefore, it is necessary to ensure that g1 and g2 operate alternately.
>> Then this allows for successful multi-instance decoding of H.264 and HEVC.
>>
>> To achieve this, we can have g1 and g2 share the same v4l2_m2m_dev, and
>> then the v4l2_m2m_dev can handle the scheduling.
>>
>> Fixes: cb5dd5a0fa518 ("media: hantro: Introduce G2/HEVC decoder")
>> Signed-off-by: Ming Qian <ming.qian@oss.nxp.com>
> 
> I like where this is going.
>> ---
>> v2
>> - Abandon the waiting approach.
>> - Switch to a shared v4l2_m2m_dev solution.
>>
>>   drivers/media/platform/verisilicon/hantro.h   |  2 +
>>   .../media/platform/verisilicon/hantro_drv.c   | 41 +++++++++++++++++--
>>   .../media/platform/verisilicon/imx8m_vpu_hw.c |  2 +
>>   3 files changed, 42 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/media/platform/verisilicon/hantro.h b/drivers/media/platform/verisilicon/hantro.h
>> index e0fdc4535b2d..8a0583f95417 100644
>> --- a/drivers/media/platform/verisilicon/hantro.h
>> +++ b/drivers/media/platform/verisilicon/hantro.h
>> @@ -77,6 +77,7 @@ struct hantro_irq {
>>    * @double_buffer:		core needs double buffering
>>    * @legacy_regs:		core uses legacy register set
>>    * @late_postproc:		postproc must be set up at the end of the job
>> + * @shared_resource:		flag of shared resource
>>    */
>>   struct hantro_variant {
>>   	unsigned int enc_offset;
>> @@ -101,6 +102,7 @@ struct hantro_variant {
>>   	unsigned int double_buffer : 1;
>>   	unsigned int legacy_regs : 1;
>>   	unsigned int late_postproc : 1;
>> +	unsigned int shared_resource : 1;
> 
> Instead of that, I'd make an array of compatible node we are going to share the
> with.
> 
>>   };
>>   
>>   /**
>> diff --git a/drivers/media/platform/verisilicon/hantro_drv.c b/drivers/media/platform/verisilicon/hantro_drv.c
>> index 60b95b5d8565..f42b2df86806 100644
>> --- a/drivers/media/platform/verisilicon/hantro_drv.c
>> +++ b/drivers/media/platform/verisilicon/hantro_drv.c
>> @@ -35,6 +35,10 @@ module_param_named(debug, hantro_debug, int, 0644);
>>   MODULE_PARM_DESC(debug,
>>   		 "Debug level - higher value produces more verbose messages");
>>   
>> +static DEFINE_MUTEX(shared_dev_lock);
>> +static atomic_t shared_dev_ref_cnt = ATOMIC_INIT(0);
>> +static struct v4l2_m2m_dev *shared_m2m_dev;
>> +
>>   void *hantro_get_ctrl(struct hantro_ctx *ctx, u32 id)
>>   {
>>   	struct v4l2_ctrl *ctrl;
>> @@ -1035,6 +1039,37 @@ static int hantro_disable_multicore(struct hantro_dev *vpu)
>>   	return 0;
>>   }
>>   
>> +static struct v4l2_m2m_dev *hantro_get_v4l2_m2m_dev(struct hantro_dev *vpu)
>> +{
>> +	if (!vpu->variant || !vpu->variant->shared_resource)
>> +		return v4l2_m2m_init(&vpu_m2m_ops);
>> +
>> +	scoped_guard(mutex, &shared_dev_lock) {
>> +		if (atomic_inc_return(&shared_dev_ref_cnt) == 1) {
>> +			shared_m2m_dev = v4l2_m2m_init(&vpu_m2m_ops);
>> +			if (IS_ERR(shared_m2m_dev))
>> +				atomic_dec(&shared_dev_ref_cnt);
>> +		}
>> +	}
>> +
>> +	return shared_m2m_dev;
>> +}
>> +
>> +static void hantro_put_v4l2_m2m_dev(struct hantro_dev *vpu)
>> +{
>> +	if (!vpu->variant || !vpu->variant->shared_resource)
>> +		v4l2_m2m_release(vpu->m2m_dev);
>> +
>> +	scoped_guard(mutex, &shared_dev_lock) {
>> +		if (!atomic_dec_return(&shared_dev_ref_cnt)) {
>> +			if (!IS_ERR(shared_m2m_dev)) {
>> +				v4l2_m2m_release(shared_m2m_dev);
>> +				shared_m2m_dev = NULL;
>> +			}
>> +		}
>> +	}
>> +}
> 
> Then instead of this, I would like to add a kref to v4l2_m2m_dec, I checked
> already and this is both safe and backward compatible.
> 
> Then in the get function, you will walk over the compatible that are different
> from the currently probe node. If you find one that is initialized, you will
> call v4l2_m2m_get() to obtained a shared reference. In _remove() you will simply
> do v4l2_m2m_put() instead of v4l2_m2m_release().
> 
> Hope the others are happy with this. For Hantro drivers, this will make it a
> much more powerfull mechanism.
> 
> Nicolas
> 

For v4l2_m2m_get() and v4l2_m2m_put(), do you mean defining these two
functions in v4l2 m2m, instead of just adding them in the hantro driver?

Regards,
Ming

>> +
>>   static int hantro_probe(struct platform_device *pdev)
>>   {
>>   	const struct of_device_id *match;
>> @@ -1186,7 +1221,7 @@ static int hantro_probe(struct platform_device *pdev)
>>   	}
>>   	platform_set_drvdata(pdev, vpu);
>>   
>> -	vpu->m2m_dev = v4l2_m2m_init(&vpu_m2m_ops);
>> +	vpu->m2m_dev = hantro_get_v4l2_m2m_dev(vpu);
>>   	if (IS_ERR(vpu->m2m_dev)) {
>>   		v4l2_err(&vpu->v4l2_dev, "Failed to init mem2mem device\n");
>>   		ret = PTR_ERR(vpu->m2m_dev);
>> @@ -1225,7 +1260,7 @@ static int hantro_probe(struct platform_device *pdev)
>>   	hantro_remove_enc_func(vpu);
>>   err_m2m_rel:
>>   	media_device_cleanup(&vpu->mdev);
>> -	v4l2_m2m_release(vpu->m2m_dev);
>> +	hantro_put_v4l2_m2m_dev(vpu)
> 
>>   err_v4l2_unreg:
>>   	v4l2_device_unregister(&vpu->v4l2_dev);
>>   err_clk_unprepare:
>> @@ -1248,7 +1283,7 @@ static void hantro_remove(struct platform_device *pdev)
>>   	hantro_remove_dec_func(vpu);
>>   	hantro_remove_enc_func(vpu);
>>   	media_device_cleanup(&vpu->mdev);
>> -	v4l2_m2m_release(vpu->m2m_dev);
>> +	hantro_put_v4l2_m2m_dev(vpu);
>>   	v4l2_device_unregister(&vpu->v4l2_dev);
>>   	clk_bulk_unprepare(vpu->variant->num_clocks, vpu->clocks);
>>   	reset_control_assert(vpu->resets);
>> diff --git a/drivers/media/platform/verisilicon/imx8m_vpu_hw.c b/drivers/media/platform/verisilicon/imx8m_vpu_hw.c
>> index 5be0e2e76882..5111ce5c36f3 100644
>> --- a/drivers/media/platform/verisilicon/imx8m_vpu_hw.c
>> +++ b/drivers/media/platform/verisilicon/imx8m_vpu_hw.c
>> @@ -356,6 +356,7 @@ const struct hantro_variant imx8mq_vpu_g1_variant = {
>>   	.num_irqs = ARRAY_SIZE(imx8mq_irqs),
>>   	.clk_names = imx8mq_g1_clk_names,
>>   	.num_clocks = ARRAY_SIZE(imx8mq_g1_clk_names),
>> +	.shared_resource = 1,
>>   };
>>   
>>   const struct hantro_variant imx8mq_vpu_g2_variant = {
>> @@ -371,6 +372,7 @@ const struct hantro_variant imx8mq_vpu_g2_variant = {
>>   	.num_irqs = ARRAY_SIZE(imx8mq_g2_irqs),
>>   	.clk_names = imx8mq_g2_clk_names,
>>   	.num_clocks = ARRAY_SIZE(imx8mq_g2_clk_names),
>> +	.shared_resource = 1,
>>   };
>>   
>>   const struct hantro_variant imx8mm_vpu_g1_variant = {

  reply	other threads:[~2025-12-01  1:52 UTC|newest]

Thread overview: 9+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2025-11-28  2:51 [PATCH v2 1/2] pmdomain: imx8m-blk-ctrl: Remove separate rst and clk mask for 8mq vpu ming.qian
2025-11-28  2:51 ` [PATCH v2 2/2] media: verisilicon: Avoid G2 bus error while decoding H.264 and HEVC ming.qian
2025-11-28 21:28   ` Nicolas Dufresne
2025-12-01  1:52     ` Ming Qian(OSS) [this message]
2025-12-02 20:21       ` Nicolas Dufresne
2025-12-02 20:57         ` Nicolas Dufresne
2025-11-28 10:29 ` [PATCH v2 1/2] pmdomain: imx8m-blk-ctrl: Remove separate rst and clk mask for 8mq vpu Peng Fan
2025-11-28 10:38 ` Lucas Stach
2025-12-01  1:43   ` Ming Qian(OSS)

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=64ad3d66-4edf-4a0b-8992-0db0e90a32b0@oss.nxp.com \
    --to=ming.qian@oss.nxp.com \
    --cc=Frank.li@nxp.com \
    --cc=benjamin.gaignard@collabora.com \
    --cc=eagle.zhou@nxp.com \
    --cc=festevam@gmail.com \
    --cc=hverkuil-cisco@xs4all.nl \
    --cc=imx@lists.linux.dev \
    --cc=kernel@pengutronix.de \
    --cc=l.stach@pengutronix.de \
    --cc=linux-arm-kernel@lists.infradead.org \
    --cc=linux-imx@nxp.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-media@vger.kernel.org \
    --cc=linux-pm@vger.kernel.org \
    --cc=mchehab@kernel.org \
    --cc=nicolas@ndufresne.ca \
    --cc=p.zabel@pengutronix.de \
    --cc=peng.fan@nxp.com \
    --cc=s.hauer@pengutronix.de \
    --cc=sebastian.fricke@collabora.com \
    --cc=shawnguo@kernel.org \
    --cc=ulf.hansson@linaro.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox