cx23885-input.c does in fact use a workqueue....

cx23885-input.c does in fact use a workqueue....

[Andy Walls]

Tejun,

I just noticed this commit:

commit 8c71778cbf2c8beaefaa2dee5478aa0622d96682
Author: Tejun Heo <tj@kernel.org>
Date:   Fri Dec 24 16:14:20 2010 +0100

    media/video: don't use flush_scheduled_work()
    
    This patch converts the remaining users of flush_scheduled_work() in
    media/video.
    
    * bttv-input.c and cx23885-input.c don't use workqueue at all.  No
      need to flush.
[...]


The cx23885 driver does in fact schedule work for IR input handling:

Here's where it is scheduled for CX23888 chips:

http://git.linuxtv.org/media_tree.git?a=blob;f=drivers/media/video/cx23885/cx23885-ir.c;h=7125247dd25558678c823ee3262675570c9aa630;hb=HEAD#l76

Here's where it is scheduled for CX23885 chips:

http://git.linuxtv.org/media_tree.git?a=blob;f=drivers/media/video/cx23885/cx23885-core.c;h=359882419b7f588b7c698dbcfb6a39ddb1603301;hb=HEAD#l1861


The two different chips are handled slightly differently because

a. the CX23888 IR unit is accessable via a PCI register block.  The IR
IRQ can be acknowledged with direct PCI register accesses in an
interrupt context, and the IR pulse FIFO serviced later in a workqueue
context.

b. the CX23885 IR unit is accessed over an I2C bus.  The CX23885 A/V IRQ
has to be masked in an interrupt context (with PCI registers accesses).
Then the CX23885 A/V unit's IR IRQ is ack'ed over I2C in a workqueue
context and the IR pulse FIFO is also serviced over I2C in a workqueue
context.


So what should be done about the flush_scheduled_work()?  I think it
belongs there.

Regards,
Andy

Re: cx23885-input.c does in fact use a workqueue....

[Dmitry Torokhov]

On Sun, Feb 13, 2011 at 08:35:22PM -0500, Andy Walls wrote:
> Tejun,
> 
> I just noticed this commit:
> 
> commit 8c71778cbf2c8beaefaa2dee5478aa0622d96682
> Author: Tejun Heo <tj@kernel.org>
> Date:   Fri Dec 24 16:14:20 2010 +0100
> 
>     media/video: don't use flush_scheduled_work()
>     
>     This patch converts the remaining users of flush_scheduled_work() in
>     media/video.
>     
>     * bttv-input.c and cx23885-input.c don't use workqueue at all.  No
>       need to flush.
> [...]
> 
> 
> The cx23885 driver does in fact schedule work for IR input handling:
> 
> Here's where it is scheduled for CX23888 chips:
> 
> http://git.linuxtv.org/media_tree.git?a=blob;f=drivers/media/video/cx23885/cx23885-ir.c;h=7125247dd25558678c823ee3262675570c9aa630;hb=HEAD#l76
> 
> Here's where it is scheduled for CX23885 chips:
> 
> http://git.linuxtv.org/media_tree.git?a=blob;f=drivers/media/video/cx23885/cx23885-core.c;h=359882419b7f588b7c698dbcfb6a39ddb1603301;hb=HEAD#l1861
> 
> 
> The two different chips are handled slightly differently because
> 
> a. the CX23888 IR unit is accessable via a PCI register block.  The IR
> IRQ can be acknowledged with direct PCI register accesses in an
> interrupt context, and the IR pulse FIFO serviced later in a workqueue
> context.
> 
> b. the CX23885 IR unit is accessed over an I2C bus.  The CX23885 A/V IRQ
> has to be masked in an interrupt context (with PCI registers accesses).
> Then the CX23885 A/V unit's IR IRQ is ack'ed over I2C in a workqueue
> context and the IR pulse FIFO is also serviced over I2C in a workqueue
> context.
> 
> 
> So what should be done about the flush_scheduled_work()?  I think it
> belongs there.
> 

Convert to using threaded irq?

-- 
Dmitry

Re: cx23885-input.c does in fact use a workqueue....

[Tejun Heo]

Hello,

On Sun, Feb 13, 2011 at 08:33:55PM -0800, Dmitry Torokhov wrote:
> > The cx23885 driver does in fact schedule work for IR input handling:
> > 
> > Here's where it is scheduled for CX23888 chips:
> > 
> > http://git.linuxtv.org/media_tree.git?a=blob;f=drivers/media/video/cx23885/cx23885-ir.c;h=7125247dd25558678c823ee3262675570c9aa630;hb=HEAD#l76
> > 
> > Here's where it is scheduled for CX23885 chips:
> > 
> > http://git.linuxtv.org/media_tree.git?a=blob;f=drivers/media/video/cx23885/cx23885-core.c;h=359882419b7f588b7c698dbcfb6a39ddb1603301;hb=HEAD#l1861

Ah, sorry about missing those.

> > The two different chips are handled slightly differently because
> > 
> > a. the CX23888 IR unit is accessable via a PCI register block.  The IR
> > IRQ can be acknowledged with direct PCI register accesses in an
> > interrupt context, and the IR pulse FIFO serviced later in a workqueue
> > context.
> > 
> > b. the CX23885 IR unit is accessed over an I2C bus.  The CX23885 A/V IRQ
> > has to be masked in an interrupt context (with PCI registers accesses).
> > Then the CX23885 A/V unit's IR IRQ is ack'ed over I2C in a workqueue
> > context and the IR pulse FIFO is also serviced over I2C in a workqueue
> > context.
> > 
> > 
> > So what should be done about the flush_scheduled_work()?  I think it
> > belongs there.
> > 
> 
> Convert to using threaded irq?

Or

1. Just flush the work items explicitly using flush_work_sync().

2. Create a dedicated workqueue to serve as flushing domain.

The first would look like the following.  Does this look correct?

Thanks.

diff --git a/drivers/media/video/cx23885/cx23885-input.c b/drivers/media/video/cx23885/cx23885-input.c
index 199b996..e27cedb 100644
--- a/drivers/media/video/cx23885/cx23885-input.c
+++ b/drivers/media/video/cx23885/cx23885-input.c
@@ -229,6 +229,9 @@ static void cx23885_input_ir_stop(struct cx23885_dev *dev)
 		v4l2_subdev_call(dev->sd_ir, ir, rx_s_parameters, &params);
 		v4l2_subdev_call(dev->sd_ir, ir, rx_g_parameters, &params);
 	}
+	flush_work_sync(&dev->cx25840_work);
+	flush_work_sync(&dev->ir_rx_work);
+	flush_work_sync(&dev->ir_tx_work);
 }
 
 static void cx23885_input_ir_close(struct rc_dev *rc)

-- 
tejun

Re: cx23885-input.c does in fact use a workqueue....

[Andy Walls]

On Mon, 2011-02-14 at 12:03 +0100, Tejun Heo wrote:
> Hello,
> 
> On Sun, Feb 13, 2011 at 08:33:55PM -0800, Dmitry Torokhov wrote:
> > > The cx23885 driver does in fact schedule work for IR input handling:
> > > 
> > > Here's where it is scheduled for CX23888 chips:
> > > 
> > > http://git.linuxtv.org/media_tree.git?a=blob;f=drivers/media/video/cx23885/cx23885-ir.c;h=7125247dd25558678c823ee3262675570c9aa630;hb=HEAD#l76
> > > 
> > > Here's where it is scheduled for CX23885 chips:
> > > 
> > > http://git.linuxtv.org/media_tree.git?a=blob;f=drivers/media/video/cx23885/cx23885-core.c;h=359882419b7f588b7c698dbcfb6a39ddb1603301;hb=HEAD#l1861
> 
> Ah, sorry about missing those.

There's a lot of indirection in drivers under drivers/media/video so it
doesn't surprise me when someone misses something.

(BTW, I only checked cx23885 since I wrote the IR handling and added the
work queue usage for IR that driver.)



> > > So what should be done about the flush_scheduled_work()?  I think it
> > > belongs there.
> > > 
> > 
> > Convert to using threaded irq?

Too big of a regression testing job for me to take on at the moment.


> Or
> 
> 1. Just flush the work items explicitly using flush_work_sync().

That will do for now.

> 2. Create a dedicated workqueue to serve as flushing domain.

I have gotten reports of the IR Rx FIFO overflows for the CX23885 IR Rx
unit (the I2C connected one).  I eventually should either set the Rx
FIFO service interrupt watermark down from 4 measurments to 1
measurment, or use a kthread_worker with some higher priority to respond
to the IR Rx FIFO service interrupt. 



> The first would look like the following.  Does this look correct?

Yes, your patch below looks sane to me.

Reviewed-by: Andy Walls <awalls@md.metrocast.net>

Thanks,
Andy

> Thanks.
> 
> diff --git a/drivers/media/video/cx23885/cx23885-input.c b/drivers/media/video/cx23885/cx23885-input.c
> index 199b996..e27cedb 100644
> --- a/drivers/media/video/cx23885/cx23885-input.c
> +++ b/drivers/media/video/cx23885/cx23885-input.c
> @@ -229,6 +229,9 @@ static void cx23885_input_ir_stop(struct cx23885_dev *dev)
>  		v4l2_subdev_call(dev->sd_ir, ir, rx_s_parameters, &params);
>  		v4l2_subdev_call(dev->sd_ir, ir, rx_g_parameters, &params);
>  	}
> +	flush_work_sync(&dev->cx25840_work);
> +	flush_work_sync(&dev->ir_rx_work);
> +	flush_work_sync(&dev->ir_tx_work);
>  }
>  
>  static void cx23885_input_ir_close(struct rc_dev *rc)
> 

Re: cx23885-input.c does in fact use a workqueue....

[Tejun Heo]

Hello,

On Mon, Feb 14, 2011 at 07:54:36PM -0500, Andy Walls wrote:
> > 1. Just flush the work items explicitly using flush_work_sync().
> 
> That will do for now.
> 
> > 2. Create a dedicated workqueue to serve as flushing domain.
> 
> I have gotten reports of the IR Rx FIFO overflows for the CX23885 IR Rx
> unit (the I2C connected one).  I eventually should either set the Rx
> FIFO service interrupt watermark down from 4 measurments to 1
> measurment, or use a kthread_worker with some higher priority to respond
> to the IR Rx FIFO service interrupt. 

Hmmm... please consider playing with WQ_HIGHPRI before going forward
with dedicated thread.

> > The first would look like the following.  Does this look correct?
> 
> Yes, your patch below looks sane to me.
> 
> Reviewed-by: Andy Walls <awalls@md.metrocast.net>

Thanks.  Will send patch with proper description soon.

-- 
tejun

[PATCH] cx23885: restore flushes of cx23885_dev work items

[Tejun Heo]

Commit 8c71778c (media/video: don't use flush_scheduled_work())
dropped flush_scheduled_work() from cx23885_input_ir_stop()
incorrectly assuming that it didn't use any work item; however,
cx23885_dev makes use of three work items - cx25840_work and
ir_{r|t}x_work.

Make cx23885_input_ir_stop() sync flush all three work items before
returning.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Andy Walls <awalls@md.metrocast.net>
Reviewed-by: Andy Walls <awalls@md.metrocast.net>
Cc: Dmitry Torokhov <dmitry.torokhov@gmail.com>
Cc: Mauro Carvalho Chehab <mchehab@redhat.com>
---
 drivers/media/video/cx23885/cx23885-input.c |    3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/media/video/cx23885/cx23885-input.c b/drivers/media/video/cx23885/cx23885-input.c
index 199b996..e27cedb 100644
--- a/drivers/media/video/cx23885/cx23885-input.c
+++ b/drivers/media/video/cx23885/cx23885-input.c
@@ -229,6 +229,9 @@ static void cx23885_input_ir_stop(struct cx23885_dev *dev)
 		v4l2_subdev_call(dev->sd_ir, ir, rx_s_parameters, &params);
 		v4l2_subdev_call(dev->sd_ir, ir, rx_g_parameters, &params);
 	}
+	flush_work_sync(&dev->cx25840_work);
+	flush_work_sync(&dev->ir_rx_work);
+	flush_work_sync(&dev->ir_tx_work);
 }
 
 static void cx23885_input_ir_close(struct rc_dev *rc)

Re: [PATCH] cx23885: restore flushes of cx23885_dev work items

[Mauro Carvalho Chehab]

Em 15-02-2011 07:20, Tejun Heo escreveu:
> Commit 8c71778c (media/video: don't use flush_scheduled_work())
> dropped flush_scheduled_work() from cx23885_input_ir_stop()
> incorrectly assuming that it didn't use any work item; however,
> cx23885_dev makes use of three work items - cx25840_work and
> ir_{r|t}x_work.
> 
> Make cx23885_input_ir_stop() sync flush all three work items before
> returning.
> 
> Signed-off-by: Tejun Heo <tj@kernel.org>
> Reported-by: Andy Walls <awalls@md.metrocast.net>
> Reviewed-by: Andy Walls <awalls@md.metrocast.net>
> Cc: Dmitry Torokhov <dmitry.torokhov@gmail.com>
> Cc: Mauro Carvalho Chehab <mchehab@redhat.com>

Acked-by: Mauro Carvalho Chehab <mchehab@redhat.com>

> ---
>  drivers/media/video/cx23885/cx23885-input.c |    3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/drivers/media/video/cx23885/cx23885-input.c b/drivers/media/video/cx23885/cx23885-input.c
> index 199b996..e27cedb 100644
> --- a/drivers/media/video/cx23885/cx23885-input.c
> +++ b/drivers/media/video/cx23885/cx23885-input.c
> @@ -229,6 +229,9 @@ static void cx23885_input_ir_stop(struct cx23885_dev *dev)
>  		v4l2_subdev_call(dev->sd_ir, ir, rx_s_parameters, &params);
>  		v4l2_subdev_call(dev->sd_ir, ir, rx_g_parameters, &params);
>  	}
> +	flush_work_sync(&dev->cx25840_work);
> +	flush_work_sync(&dev->ir_rx_work);
> +	flush_work_sync(&dev->ir_tx_work);
>  }
>  
>  static void cx23885_input_ir_close(struct rc_dev *rc)