Date:   Fri, 15 Sep 2023 14:44:41 +0200
From:   Martin Tůma <tumic@...see.org>
To:     Hans Verkuil <hverkuil-cisco@...all.nl>,
        Mauro Carvalho Chehab <mchehab@...nel.org>
Cc:     linux-media@...r.kernel.org, linux-kernel@...r.kernel.org,
        Martin Tůma <martin.tuma@...iteqautomotive.com>
Subject: Re: [RESEND PATCH v9 1/2] Added Digiteq Automotive MGB4 driver

Hi Hans,

On 15. 09. 23 13:50, Hans Verkuil wrote:
> Hi Martin,
> 
> On 12/09/2023 14:07, tumic@...see.org wrote:
>> From: Martin Tůma <martin.tuma@...iteqautomotive.com>
>>
>> Digiteq Automotive MGB4 is a modular frame grabber PCIe card for automotive
>> video interfaces. As for now, two modules - FPD-Link and GMSL - are
>> available and supported by the driver. The card has two inputs and two
>> outputs (FPD-Link only).
>>
>> In addition to the video interfaces it also provides a trigger signal
>> interface and a MTD interface for FPGA firmware upload.
>>
>> Signed-off-by: Martin Tůma <martin.tuma@...iteqautomotive.com>
>> ---
> 
> <snip>
> 
>> +static ssize_t video_source_show(struct device *dev,
>> +				 struct device_attribute *attr, char *buf)
>> +{
>> +	struct video_device *vdev = to_video_device(dev);
>> +	struct mgb4_vout_dev *voutdev = video_get_drvdata(vdev);
>> +	u32 config = mgb4_read_reg(&voutdev->mgbdev->video,
>> +	  voutdev->config->regs.config);
>> +
>> +	return sprintf(buf, "%u\n", (config & 0xc) >> 2);
>> +}
>> +
>> +static ssize_t video_source_store(struct device *dev,
>> +				  struct device_attribute *attr,
>> +				  const char *buf, size_t count)
>> +{
>> +	struct video_device *vdev = to_video_device(dev);
>> +	struct mgb4_vout_dev *voutdev = video_get_drvdata(vdev);
>> +	struct mgb4_dev *mgbdev = voutdev->mgbdev;
>> +	struct mgb4_vin_dev *loopin_new = 0, *loopin_old = 0;
>> +	unsigned long val;
>> +	unsigned long flags_in[MGB4_VIN_DEVICES], flags_out[MGB4_VOUT_DEVICES];
>> +	ssize_t ret;
>> +	u32 config;
>> +	int i;
>> +
>> +	memset(flags_in, 0, sizeof(flags_in));
>> +	memset(flags_out, 0, sizeof(flags_out));
>> +
>> +	ret = kstrtoul(buf, 10, &val);
>> +	if (ret)
>> +		return ret;
>> +	if (val > 3)
>> +		return -EINVAL;
>> +
>> +	for (i = 0; i < MGB4_VIN_DEVICES; i++)
>> +		if (mgbdev->vin[i])
>> +			spin_lock_irqsave(&mgbdev->vin[i]->vdev.fh_lock,
>> +					  flags_in[i]);
>> +	for (i = 0; i < MGB4_VOUT_DEVICES; i++)
>> +		if (mgbdev->vout[i])
>> +			spin_lock_irqsave(&mgbdev->vout[i]->vdev.fh_lock,
>> +					  flags_out[i]);
> 
> I missed this in my review: never use fh_lock, that's for internal use only.
> 
> Instead, use the vdev->lock mutex.
> 
>> +
>> +	ret = -EBUSY;
>> +	for (i = 0; i < MGB4_VIN_DEVICES; i++)
>> +		if (mgbdev->vin[i] && !list_empty(&mgbdev->vin[i]->vdev.fh_list))
>> +			goto error;
> 
> This is also wrong, the key moment at which such changes are no longer allowed
> is when you allocate buffers for the first time. Use vb2_is_busy() to detect that.
> 
>  From what I can see in the code most store functions operate on a single video device,
> so for those using mutex_lock_interruptible(vdev->lock) and vb2_is_busy() will do the
> trick.
> 

Ok, I will change the check to vb2_is_busy() under the vdev->lock for 
all the "standard" cases (which are all, except the video_source_store() 
case).

> For this store function that needs to check all devices it is more complicated.
> 
> I think the best approach would be to create a top-level field 'reconfig_in_progress'
> (or something along those lines), protected by a spinlock.
> 
> At the start this function will take the lock, check if reconfig is already in progress
> and just return EBUSY here. Otherwise it sets reconfig_in_progress to true and unlocks
> the lock.
> 
> After that for each video device you take the vdev->lock, check if vb2_is_busy() and
> unlock vdev->lock. If any is busy, then set reconfig_in_progress back to false and
> return EBUSY here.
> 
> To prevent another process from creating buffers, you have to check in the queue_setup
> functions if a reconfig is in progress and return EBUSY there.
> 
> This scheme avoids having to lock ALL video devices (bad practice), and it should
> also avoid this smatch warning.
> 

Ok, for this "ugly" one configuration I will add a PCIe device-based 
field and check it in all the v4l2 device's queue_setup() functions. 
Sounds better than locking all the devices at the same time.

Thanks for your help, I will send you a v10 with this two changes 
(together with the sparse/smatch fixes) soon.

M.

> Note: instead of using a spinlock I think you can also use test_and_set_bit: that's
> atomic and you can avoid adding a spinlock.
> 
> Regards,
> 
> 	Hans
> 
>> +	for (i = 0; i < MGB4_VOUT_DEVICES; i++)
>> +		if (mgbdev->vout[i] && !list_empty(&mgbdev->vout[i]->vdev.fh_list))
>> +			goto error;
>> +
>> +	config = mgb4_read_reg(&mgbdev->video, voutdev->config->regs.config);
>> +
>> +	if (((config & 0xc) >> 2) < MGB4_VIN_DEVICES)
>> +		loopin_old = mgbdev->vin[(config & 0xc) >> 2];
>> +	if (val < MGB4_VIN_DEVICES)
>> +		loopin_new = mgbdev->vin[val];
>> +	if (loopin_old && loopin_cnt(loopin_old) == 1)
>> +		mgb4_mask_reg(&mgbdev->video, loopin_old->config->regs.config,
>> +			      0x2, 0x0);
>> +	if (loopin_new)
>> +		mgb4_mask_reg(&mgbdev->video, loopin_new->config->regs.config,
>> +			      0x2, 0x2);
>> +
>> +	if (val == voutdev->config->id + MGB4_VIN_DEVICES)
>> +		mgb4_write_reg(&mgbdev->video, voutdev->config->regs.config,
>> +			       config & ~(1 << 1));
>> +	else
>> +		mgb4_write_reg(&mgbdev->video, voutdev->config->regs.config,
>> +			       config | (1U << 1));
>> +
>> +	mgb4_mask_reg(&mgbdev->video, voutdev->config->regs.config, 0xc,
>> +		      val << 2);
>> +
>> +	ret = count;
>> +
>> +error:
>> +	for (i = MGB4_VOUT_DEVICES - 1; i >= 0; i--)
>> +		if (mgbdev->vout[i])
>> +			spin_unlock_irqrestore(&mgbdev->vout[i]->vdev.fh_lock,
>> +					       flags_out[i]);
>> +	for (i = MGB4_VIN_DEVICES - 1; i >= 0; i--)
>> +		if (mgbdev->vin[i])
>> +			spin_unlock_irqrestore(&mgbdev->vin[i]->vdev.fh_lock,
>> +					       flags_in[i]);
>> +
>> +	return ret;
>> +}
> 

Powered by blists - more mailing lists