At Wed, 15 Oct 2014 18:53:28 -0600, Shuah Khan wrote:

On 10/15/2014 11:05 AM, Takashi Iwai wrote:

...

...

+#if defined(CONFIG_MEDIA_SUPPORT) +extern int media_tknres_create(struct device *dev); +extern int media_tknres_destroy(struct device *dev);

+extern int media_get_tuner_tkn(struct device *dev); +extern int media_put_tuner_tkn(struct device *dev);

+extern int media_get_audio_tkn(struct device *dev); +extern int media_put_audio_tkn(struct device *dev);

The words "tknres" and "tkn" (especially latter) look ugly and not clear to me. IMO, it should be "token".

No problem. I can change that to media_token_create/destroy() and expand token in other functions.

...

...

+struct media_tkn {

• spinlock_t lock;
• unsigned int owner; /* owner task pid */
• unsigned int tgid; /* owner task gid */
• struct task_struct *task;

+};

+struct media_tknres {

• struct media_tkn tuner;
• struct media_tkn audio;

+};

Why do you need to have both tuner and audio tokens? If I understand correctly, no matter whether it's tuner or audio, if it's being used, the open must fail, right?

As it evolved during development, it turns out at the moment I don't have any use-cases that require holding audio and tuner separately. It probably could be collapsed into just a media token. I have to think about this some.

...

...

+static int __media_get_tkn(struct media_tkn *tkn, char *tkn_str) +{

• int rc = 0;
• unsigned tpid;
• unsigned tgid;
• spin_lock(&tkn->lock);

You should use spin_lock_irqsave() here and in all other places. The trigger callback in usb-audio, for example, may be called in irq context.

ok. Good point, will change that.

...

...

• tpid = task_pid_nr(current);
• tgid = task_tgid_nr(current);
• /* allow task in the same group id to release */

IMO, it's not "release" but "steal"... But what happens if the stolen owner calls put? Then it'll be released although the stealing owner still thinks it's being held.

Yeah it could be called a steal. :) Essentially tgid happens to be the real owner. I am overwriting the pid with current pid when tgid is same.

media dvb and v4l apps start two or more threads that all share the tgid and subsequent token gets should be allowed based on the tgid.

Scenario 1:

Please note that there are 3 device files in question and media token resource is the lock for all. Hence the changes to v4l-core, dvb-core, and snd-usb-audio to hold the token for exclusive access when the task or tgid don't match.

program starts:

• first thread opens device file in R/W mode - open gets the token or thread makes ioctls calls that clearly indicates intent to change tuner settings

• creates one thread for audio

• creates another for video or continues video function

• video thread does a close and re-opens the device file

  In this case when thread tries to close, token put fails unless tasks with same tgid are allowed to release. ( I ran into this one of the media applications and it is a valid case to handle, thread can close the file and should be able to open again without running into token busy case )

• or continue to just use the device file

• audio thread does snd_pcm_capture ops - trigger start

program exits:

• video thread closes the device file
• audio thread does snd_pcm_capture ops - trigger stop

This got me thinking about the scenario when an application does a fork() as opposed to create a thread. I have to think about this and see how I can address that.

...

...

• if (tkn->task && ((tkn->task != current) && (tkn->tgid != tgid))) {

Shouldn't the second "&&" be "||" instead? And too many parentheses there.

Right - this is my bad. The comment right above this conditional is a give away that, at some point I did a copy and paste from _put to _get and only changed the first task null check. I am yelling at myself at the moment.

...

...

• 	rc = -EBUSY;
  

Wrong indentation.

Yes. Will fix that.

...

I have a feeling that the whole thing can be a bit more simplified in the end...

Any ideas to simplify are welcome.

There are a few points to make things more consistent and simplified, IMO. First of all, the whole concept can be more generalized. It's not necessarily only for media and audio. Rather we can make it a generic dev_token. Then it'll become more convincing to land into lib/*.

The media-specific handling is only about the pid and gid checks. This can be implemented as a callback that is passed at creating a token, for example. This will reduce the core code in lib/*.

Also, get and put should handle a proper refcount. The point I mentioned in my previous post is the possible unbalance, and you'll see unexpected unlock in the scenario. In addition, with use of kref, the lock can be removed.

So, essentially the lib code would be just a devres_alloc() for an object with kref, and dev_token_get() will take a kref with the exclusion check.

Takashi

On 10/21/2014 01:51 PM, Takashi Iwai wrote:

At Tue, 21 Oct 2014 12:46:03 +0200, Hans Verkuil wrote:

...

Hi Shuah,

As promised, here is my review for this patch series.

On 10/14/2014 04:58 PM, Shuah Khan wrote:

...

Add media token device resource framework to allow sharing resources such as tuner, dma, audio etc. across media drivers and non-media sound drivers that control media hardware. The Media token resource is created at the main struct device that is common to all drivers that claim various pieces of the main media device, which allows them to find the resource using the main struct device. As an example, digital, analog, and snd-usb-audio drivers can use the media token resource API using the main struct device for the interface the media device is attached to.

A shared media tokens resource is created using devres framework for drivers to find and lock/unlock. Creating a shared devres helps avoid creating data structure dependencies between drivers. This media token resource contains media token for tuner, and audio. When tuner token is requested, audio token is issued.

Did you mean: 'tuner token is issued' instead of audio token?

I also have the same question as Takashi: why do we have an audio token in the first place? While you are streaming audio over alsa the underlying tuner must be marked as being in use. It's all about the tuner, since that's the resource that is being shared, not about audio as such.

For the remainder of my review I will ignore the audio-related code and concentrate only on the tuner part.

...

Subsequent token (for tuner and audio) gets from the same task and task in the same tgid succeed. This allows applications that make multiple v4l2 ioctls to work with the first call acquiring the token and applications that create separate threads to handle video and audio functions.

Signed-off-by: Shuah Khan shuahkh@osg.samsung.com

MAINTAINERS | 2 + include/linux/media_tknres.h | 50 +++++++++ lib/Makefile | 2 + lib/media_tknres.c | 237 ++++++++++++++++++++++++++++++++++++++++++

I am still not convinced myself that this should be a generic API. The only reason we need it today is for sharing tuners. Which is almost a purely media thing with USB audio as the single non-media driver that will be affected. Today I see no use case outside of tuners. I would probably want to keep this inside drivers/media.

If this is going to be expanded it can always be moved to lib later.

Well, my argument is that it should be more generic if it were intended to be put in lib. It'd be fine to put into drivers/media, but this code snippet must be a separate module. Otherwise usb-audio would grab the whole media stuff even if not needed at all.

Certainly.

(snip)

...

I also discovered that you are missing MODULE_AUTHOR, MODULE_DESCRIPTION and above all MODULE_LICENSE. Without the MODULE_LICENSE it won't link this module to the GPL devres_* functions. It took me some time to figure that out.

It was a code in lib, so it cannot be a module at all :)

Well, it depends on CONFIG_MEDIA_SUPPORT which is 'm' in my case, so it compiles as a module :-)

Regards,

Hans

On 10/16/2014 06:00 AM, Lars-Peter Clausen wrote:

On 10/14/2014 04:58 PM, Shuah Khan wrote: [...]

...

  switch (cmd) {
  case SNDRV_PCM_TRIGGER_START:

•    err = media_get_audio_tkn(&subs->dev->dev);
  

•    if (err == -EBUSY) {
  

•        dev_info(&subs->dev->dev, "%s device is busy\n",
  

•                __func__);
  

In my opinion this should not dev_info() as this is out of band error signaling and also as the potential to spam the log. The userspace application is already properly notified by the return code.

Yes it has the potential to flood the dmesg especially with alsa, I will remove the dev_info().

thanks, -- Shuah

On 10/16/2014 08:01 AM, Takashi Iwai wrote:

At Thu, 16 Oct 2014 07:10:37 -0600, Shuah Khan wrote:

...

On 10/16/2014 06:00 AM, Lars-Peter Clausen wrote:

...

On 10/14/2014 04:58 PM, Shuah Khan wrote: [...]

...

  switch (cmd) {
  case SNDRV_PCM_TRIGGER_START:

•    err = media_get_audio_tkn(&subs->dev->dev);
  

•    if (err == -EBUSY) {
  

•        dev_info(&subs->dev->dev, "%s device is busy\n",
  

•                __func__);
  

In my opinion this should not dev_info() as this is out of band error signaling and also as the potential to spam the log. The userspace application is already properly notified by the return code.

Yes it has the potential to flood the dmesg especially with alsa, I will remove the dev_info().

Yes. And, I think doing this in the trigger isn't the best. Why not in open & close?

My first cut of this change was in open and close. I saw pulseaudio application go into this loop trying to open the device. To avoid such problems, I went with trigger stat and stop. That made all the pulseaudio continues attempts to open problems go away.

Also, is this token restriction needed only for PCM? No mixer or other controls?

snd_pcm_ops are the only ones media drivers implement and use. So I don't think mixer is needed. If it is needed, it is not to hard to add for that case.

thanks, -- Shuah

On 10/16/2014 08:16 AM, Takashi Iwai wrote:

At Thu, 16 Oct 2014 08:10:52 -0600, Shuah Khan wrote:

...

On 10/16/2014 08:01 AM, Takashi Iwai wrote:

...

At Thu, 16 Oct 2014 07:10:37 -0600, Shuah Khan wrote:

...

On 10/16/2014 06:00 AM, Lars-Peter Clausen wrote:

...

On 10/14/2014 04:58 PM, Shuah Khan wrote: [...]

...

  switch (cmd) {
  case SNDRV_PCM_TRIGGER_START:

•    err = media_get_audio_tkn(&subs->dev->dev);
  

•    if (err == -EBUSY) {
  

•        dev_info(&subs->dev->dev, "%s device is busy\n",
  

•                __func__);
  

In my opinion this should not dev_info() as this is out of band error signaling and also as the potential to spam the log. The userspace application is already properly notified by the return code.

Yes it has the potential to flood the dmesg especially with alsa, I will remove the dev_info().

Yes. And, I think doing this in the trigger isn't the best. Why not in open & close?

My first cut of this change was in open and close. I saw pulseaudio application go into this loop trying to open the device. To avoid such problems, I went with trigger stat and stop. That made all the pulseaudio continues attempts to open problems go away.

But now starting the stream gives the error, and PA would loop it again, no?

...

...

Also, is this token restriction needed only for PCM? No mixer or other controls?

snd_pcm_ops are the only ones media drivers implement and use. So I don't think mixer is needed. If it is needed, it is not to hard to add for that case.

Well, then I wonder what resource does actually conflict with usb-audio and media drivers at all...?

audio for dvb/v4l apps gets disrupted when audio app starts. For example, dvb or v4l app tuned to a channel, and when an audio app starts. audio path needs protected to avoid conflicts between digital and analog applications as well.

-- Shuah

On 10/16/2014 08:48 AM, Takashi Iwai wrote:

At Thu, 16 Oct 2014 08:39:14 -0600, Shuah Khan wrote:

...

On 10/16/2014 08:16 AM, Takashi Iwai wrote:

...

At Thu, 16 Oct 2014 08:10:52 -0600, Shuah Khan wrote:

...

On 10/16/2014 08:01 AM, Takashi Iwai wrote:

...

At Thu, 16 Oct 2014 07:10:37 -0600, Shuah Khan wrote:

...

On 10/16/2014 06:00 AM, Lars-Peter Clausen wrote: > On 10/14/2014 04:58 PM, Shuah Khan wrote: > [...] >> switch (cmd) { >> case SNDRV_PCM_TRIGGER_START: >> + err = media_get_audio_tkn(&subs->dev->dev); >> + if (err == -EBUSY) { >> + dev_info(&subs->dev->dev, "%s device is busy\n", >> + __func__); > > In my opinion this should not dev_info() as this is out of band error > signaling and also as the potential to spam the log. The userspace > application is already properly notified by the return code. >

Yes it has the potential to flood the dmesg especially with alsa, I will remove the dev_info().

Yes. And, I think doing this in the trigger isn't the best. Why not in open & close?

My first cut of this change was in open and close. I saw pulseaudio application go into this loop trying to open the device. To avoid such problems, I went with trigger stat and stop. That made all the pulseaudio continues attempts to open problems go away.

But now starting the stream gives the error, and PA would loop it again, no?

...

...

Also, is this token restriction needed only for PCM? No mixer or other controls?

snd_pcm_ops are the only ones media drivers implement and use. So I don't think mixer is needed. If it is needed, it is not to hard to add for that case.

Well, then I wonder what resource does actually conflict with usb-audio and media drivers at all...?

audio for dvb/v4l apps gets disrupted when audio app starts. For example, dvb or v4l app tuned to a channel, and when an audio app starts. audio path needs protected to avoid conflicts between digital and analog applications as well.

OK, then concentrating on only PCM is fine.

But, I'm still not convinced about doing the token management in the trigger. The reason -EBUSY doesn't work is that it's the very same error code when a PCM device is blocked by other processes. And -EAGAIN is interpreted by PCM core to -EBUSY for historical reasons.

ah. ok your recommendation is to go with open and close. Mauro has some reservations with holding at open when I discussed my observations with pulseaudio when I was holding token in open instead of trigger start. Maybe he can chime with his concerns. I think his concern was breaking applications if token is held in open().

Based on what you are seeing trigger could be worse.

How applications (e.g. PA) should behave if the token get fails? Shouldn't it retry or totally give up?

It would be up to the application I would think. I see that arecord quits right away when it finds the device busy. pluseaudio on the other hand appears to retry. I downloaded pulseaudio sources to understand what it is doing, however I didn't get too far. The way it does audio handling is complex for me to follow without spending a lot of time.

thanks, -- Shuah

At Sat, 18 Oct 2014 20:49:58 +0200, Mauro Carvalho Chehab wrote:

Em Thu, 16 Oct 2014 08:59:29 -0600 Shuah Khan shuahkh@osg.samsung.com escreveu:

...

On 10/16/2014 08:48 AM, Takashi Iwai wrote:

...

At Thu, 16 Oct 2014 08:39:14 -0600, Shuah Khan wrote:

...

On 10/16/2014 08:16 AM, Takashi Iwai wrote:

...

At Thu, 16 Oct 2014 08:10:52 -0600, Shuah Khan wrote:

...

On 10/16/2014 08:01 AM, Takashi Iwai wrote: > At Thu, 16 Oct 2014 07:10:37 -0600, > Shuah Khan wrote: >> >> On 10/16/2014 06:00 AM, Lars-Peter Clausen wrote: >>> On 10/14/2014 04:58 PM, Shuah Khan wrote: >>> [...] >>>> switch (cmd) { >>>> case SNDRV_PCM_TRIGGER_START: >>>> + err = media_get_audio_tkn(&subs->dev->dev); >>>> + if (err == -EBUSY) { >>>> + dev_info(&subs->dev->dev, "%s device is busy\n", >>>> + __func__); >>> >>> In my opinion this should not dev_info() as this is out of band error >>> signaling and also as the potential to spam the log. The userspace >>> application is already properly notified by the return code. >>> >> >> Yes it has the potential to flood the dmesg especially with alsa, >> I will remove the dev_info(). > > Yes. And, I think doing this in the trigger isn't the best. > Why not in open & close?

My first cut of this change was in open and close. I saw pulseaudio application go into this loop trying to open the device. To avoid such problems, I went with trigger stat and stop. That made all the pulseaudio continues attempts to open problems go away.

But now starting the stream gives the error, and PA would loop it again, no?

...

> Also, is this token restriction needed only for PCM? No mixer or > other controls?

snd_pcm_ops are the only ones media drivers implement and use. So I don't think mixer is needed. If it is needed, it is not to hard to add for that case.

Well, then I wonder what resource does actually conflict with usb-audio and media drivers at all...?

audio for dvb/v4l apps gets disrupted when audio app starts. For example, dvb or v4l app tuned to a channel, and when an audio app starts. audio path needs protected to avoid conflicts between digital and analog applications as well.

OK, then concentrating on only PCM is fine.

But, I'm still not convinced about doing the token management in the trigger. The reason -EBUSY doesn't work is that it's the very same error code when a PCM device is blocked by other processes. And -EAGAIN is interpreted by PCM core to -EBUSY for historical reasons.

ah. ok your recommendation is to go with open and close. Mauro has some reservations with holding at open when I discussed my observations with pulseaudio when I was holding token in open instead of trigger start. Maybe he can chime with his concerns. I think his concern was breaking applications if token is held in open().

Yes. My concern is that PA has weird behaviors, and it tries to open and keep opened all audio devices.

PA usually closes the PCM devices when unused. If it doesn't, it must be a bug.

Takashi

At Tue, 21 Oct 2014 17:42:51 +0200, Hans Verkuil wrote:

On 10/16/2014 04:48 PM, Takashi Iwai wrote:

...

At Thu, 16 Oct 2014 08:39:14 -0600, Shuah Khan wrote:

...

On 10/16/2014 08:16 AM, Takashi Iwai wrote:

...

At Thu, 16 Oct 2014 08:10:52 -0600, Shuah Khan wrote:

...

On 10/16/2014 08:01 AM, Takashi Iwai wrote:

...

At Thu, 16 Oct 2014 07:10:37 -0600, Shuah Khan wrote: > > On 10/16/2014 06:00 AM, Lars-Peter Clausen wrote: >> On 10/14/2014 04:58 PM, Shuah Khan wrote: >> [...] >>> switch (cmd) { >>> case SNDRV_PCM_TRIGGER_START: >>> + err = media_get_audio_tkn(&subs->dev->dev); >>> + if (err == -EBUSY) { >>> + dev_info(&subs->dev->dev, "%s device is busy\n", >>> + __func__); >> >> In my opinion this should not dev_info() as this is out of band error >> signaling and also as the potential to spam the log. The userspace >> application is already properly notified by the return code. >> > > Yes it has the potential to flood the dmesg especially with alsa, > I will remove the dev_info().

Yes. And, I think doing this in the trigger isn't the best. Why not in open & close?

My first cut of this change was in open and close. I saw pulseaudio application go into this loop trying to open the device. To avoid such problems, I went with trigger stat and stop. That made all the pulseaudio continues attempts to open problems go away.

But now starting the stream gives the error, and PA would loop it again, no?

...

...

Also, is this token restriction needed only for PCM? No mixer or other controls?

snd_pcm_ops are the only ones media drivers implement and use. So I don't think mixer is needed. If it is needed, it is not to hard to add for that case.

Well, then I wonder what resource does actually conflict with usb-audio and media drivers at all...?

audio for dvb/v4l apps gets disrupted when audio app starts. For example, dvb or v4l app tuned to a channel, and when an audio app starts. audio path needs protected to avoid conflicts between digital and analog applications as well.

OK, then concentrating on only PCM is fine.

But, I'm still not convinced about doing the token management in the trigger. The reason -EBUSY doesn't work is that it's the very same error code when a PCM device is blocked by other processes. And -EAGAIN is interpreted by PCM core to -EBUSY for historical reasons.

How applications (e.g. PA) should behave if the token get fails? Shouldn't it retry or totally give up?

Quite often media apps open the alsa device at the start and then switch between TV, radio or DVB mode. If the alsa device would claim the tuner just by being opened (as opposed to actually using the tuner, which happens when you start streaming),

What about parameter changes? The sound devices have to be configured before using. Don't they influence on others at all, i.e. you can change the PCM sample rate etc during TV, radio or DVB is running?

then that would make it impossible for the application to switch tuner mode. In general you want to avoid that open() will start configuring hardware since that can quite often be slow. Tuner configuration in particular can be slow since several common tuners need to load firmware over i2c. You only want to do that when it is really needed, and not when some application (udev!) opens the device just to examine what sort of device it is.

But most apps close the device soon after that, no? Which programs keep the PCM device (not the control) opened without actually using?

So claiming the tuner in the trigger seems to be the right place. If returning EBUSY is a poor error code for alsa, then we can use something else for that. EACCES perhaps?

Sorry, I'm not convinced by that. If the device has to be controlled exclusively, the right position is the open/close. Otherwise, the program cannot know when it becomes inaccessible out of sudden during its operation.

Takashi

At Tue, 21 Oct 2014 12:08:59 -0400, Devin Heitmueller wrote:

...

Sorry, I'm not convinced by that. If the device has to be controlled exclusively, the right position is the open/close. Otherwise, the program cannot know when it becomes inaccessible out of sudden during its operation.

I can say that I've definitely seen cases where if you configure a device as the "default" capture device in PulseAudio, then pulse will continue to capture from it even if you're not actively capturing the audio from pulse. I only spotted this because I had a USB analyzer on the device and was dumbfounded when the ISOC packets kept arriving even after I had closed VLC.

You might have had an input monitor active in some PA apps? PA shouldn't do it as default. If it does unintentionally, you should report it to PA guys.

Takashi

this seems like a feature, not a bug. PulseAudio starts streaming before clients push any data and likewise keeps sources active even after for some time after clients stop recording. Closing VLC in your example doesn't immediately close the ALSA device. look for module-suspend-on-idle in your default.pa config file.

The ALSA userland emulation in PulseAudio is supposed to faithfully emulate the behavior of the ALSA kernel ABI... except when it doesn't, then it's not a bug but rather a feature. :-)

I also agree that the open/close of the alsa device is the only way to control exclusion.

I was also a proponent that we should have fairly coarse locking done at open/close for the various device nodes (ALSA/V4L/DVB). The challenge here is that we have a large installed based of existing applications that rely on kernel behavior that isn't formally specified in any specification. Hence we're forced to try to come up with a solution that minimizes the risk of ABI breakage.

If we were doing this from scratch then we could lay down some hard/fast rules about things apps aren't supposed to do and how apps are supposed to respond to those exception cases. Unfortunately we don't have that luxury here.

Devin

Em Wed, 22 Oct 2014 14:26:41 -0500 Pierre-Louis Bossart pierre-louis.bossart@linux.intel.com escreveu:

On 10/21/14, 11:08 AM, Devin Heitmueller wrote:

...

...

Sorry, I'm not convinced by that. If the device has to be controlled exclusively, the right position is the open/close. Otherwise, the program cannot know when it becomes inaccessible out of sudden during its operation.

I can say that I've definitely seen cases where if you configure a device as the "default" capture device in PulseAudio, then pulse will continue to capture from it even if you're not actively capturing the audio from pulse. I only spotted this because I had a USB analyzer on the device and was dumbfounded when the ISOC packets kept arriving even after I had closed VLC.

this seems like a feature, not a bug. PulseAudio starts streaming before clients push any data and likewise keeps sources active even after for some time after clients stop recording. Closing VLC in your example doesn't immediately close the ALSA device. look for module-suspend-on-idle in your default.pa config file.

This could be a feature for an audio capture device that is standalone, but for sure it isn't a feature for an audio capture device where the audio is only available if video is also being streamed.

A V4L device with ALSA capture is a different beast than a standalone capture port. In a simplified way, it will basically follow the following state machine:

+---------------+ | start | +---------------+ | | v +--------------------------------+ | idle | <+ +--------------------------------+ | | ^ | | | | | | v | | | +---------------+ | | | | configuration | | | | +---------------+ | | | | | | | | | | | v | | | +---------------+ | | | +> | streaming | -+ | | | +---------------+ | | | | | | | | | | | v v | | +---------------+-----------+----+ | +- | 1 | suspended | 2 | -+ +---------------+-----------+----+

1) start state

This is when the V4L2 device gots probed. It checks if the hardware is present and initializes some vars/registers, turning off everything that can be powered down.

The tuner on put in sleep mode, analog audio/video decoders and the dvb frontend and demux are also turned off.

2) idle state

As the device is powered off, audio won't produce anything.

Depending on the device, reading for audio may return a sequence of zeros, or may even fail, as the DMA engine is not ready yet for streaming.

Also, the audio mixer is muted, but the audio input switch is on a random state.

2) configuration state

When V4L2 node device is opened and configured, the audio mixer will be switched to input audio from the same source of the video stream. The corresponding audio input is also unmuted. Almost all devices have at least two audio/video inputs: TV TUNER and COMPOSITE. Other devices may also have S-VIDEO, COMPOSITE 2, RADIO TUNER, etc.

If the device is set on TUNER mode, on modern devices, a tuner firmware will be loaded. That may require a long time. Typically, most devices take 1/2 seconds to load a firmware, but some devices may take up to 30 seconds. The firmware may depend on the TV standard that will be used, so this can't be loaded at driver warm up state.

Also, the power consumption of the tuners is high (it can be ~100-200 mW or more when powered, and ~16mW when just I2C is powered). We don't want to keep it powered when the device is not used, as this spends battery. Also, the life of the device reduces a lot if we keep it always powered.

During this stage, if an ALSA call is issued, it may interfere at the device settings and/or firmware load, with can cause the audio to fail. On such cases, applications might need to close the V4L2 node and re-open again.

3) streaming state

The change to this staging requires a V4L2 ioctl.

Please notice, however, that some apps will open the audio device before the V4L2 node, while others will open it after that.

In any case, audio will only start to produce data after the V4L2 ioctl at V4L2 that starts the DMA engine there.

After that ioctl: - Audio PCM capture will work; - The mixers will be in a good state: unmuted, and switched to the corresponding input as the video stream.

If the user wants to do something unusual, like mixing the composite audio input with the tuner audio input, it can use the ALSA mixer for doing that. Otherwise, the only part of the ALSA device that will be used is the PCM engine.

4) streaming->stop transition

When the V4L2 device is closed (or an ioctl for stream stop is issued), the driver will release the resources. What will be released vary from driver to driver.

Some drivers will stop both audio and video DMA engine (as they actually have just one DMA engine for both). Other devices will still be streaming audio until the alsa trigger stops it.

In any case, the driver will put the tuner into sleep mode. Some drivers do it immediately. Others will wait for a few seconds before doing that.

Ether way, from the PCM capture audio PoV, the device is now useless: it will either not return anything, or it will just return a stream with all zeros.

5) suspend/resume transitions

When the machine suspends to disk or memory, the DMA engine will stop, and the devices will be powered off.

The logic of returning from suspend will depend on the previous state.

If the device were in stop pode, at resume time, it should return to idle state.

If the device is streaming, it should return to streaming state. In such case, no ALSA ioctls should happen while V4L2 is suspending or resuming, otherwise the device will be on an unknown state, and will only return to work after replugging the device.

I also agree that the open/close of the alsa device is the only way to control exclusion.

This will break apps, as they can open the alsa device before configuring the stream with V4L2. They can also close alsa either before or after stopping the V4L2 stream.

-Pierre

-

As a reference, the state machine above came from this dot config, using graph-easy to produce an ascii output:

digraph media { node [shape=record] suspended[label = "<streaming> 1 | <suspended> suspended | <idle> 2"] start -> idle idle -> configuration configuration -> streaming streaming -> idle idle -> "suspended":idle "suspended":idle -> idle streaming -> "suspended":streaming "suspended":streaming -> streaming }

(re-sending from my third e-mail - somehow, the two emails I have at Samsung didn't seem to be delivering to vger.kernel.org today)

Em Wed, 22 Oct 2014 14:26:41 -0500 Pierre-Louis Bossart pierre-louis.bossart@linux.intel.com escreveu:

On 10/21/14, 11:08 AM, Devin Heitmueller wrote:

...

...

Sorry, I'm not convinced by that. If the device has to be controlled exclusively, the right position is the open/close. Otherwise, the program cannot know when it becomes inaccessible out of sudden during its operation.

I can say that I've definitely seen cases where if you configure a device as the "default" capture device in PulseAudio, then pulse will continue to capture from it even if you're not actively capturing the audio from pulse. I only spotted this because I had a USB analyzer on the device and was dumbfounded when the ISOC packets kept arriving even after I had closed VLC.

this seems like a feature, not a bug. PulseAudio starts streaming before clients push any data and likewise keeps sources active even after for some time after clients stop recording. Closing VLC in your example doesn't immediately close the ALSA device. look for module-suspend-on-idle in your default.pa config file.

This could be a feature for an audio capture device that is standalone, but for sure it isn't a feature for an audio capture device where the audio is only available if video is also being streamed.

A V4L device with ALSA capture is a different beast than a standalone capture port. In a simplified way, it will basically follow the following state machine:

+---------------+ | start | +---------------+ | | v +--------------------------------+ | idle | <+ +--------------------------------+ | | ^ | | | | | | v | | | +---------------+ | | | | configuration | | | | +---------------+ | | | | | | | | | | | v | | | +---------------+ | | | +> | streaming | -+ | | | +---------------+ | | | | | | | | | | | v v | | +---------------+-----------+----+ | +- | 1 | suspended | 2 | -+ +---------------+-----------+----+

1) start state

This is when the V4L2 device gots probed. It checks if the hardware is present and initializes some vars/registers, turning off everything that can be powered down.

The tuner on put in sleep mode, analog audio/video decoders and the dvb frontend and demux are also turned off.

2) idle state

As the device is powered off, audio won't produce anything.

Depending on the device, reading for audio may return a sequence of zeros, or may even fail, as the DMA engine is not ready yet for streaming.

Also, the audio mixer is muted, but the audio input switch is on a random state.

2) configuration state

When V4L2 node device is opened and configured, the audio mixer will be switched to input audio from the same source of the video stream. The corresponding audio input is also unmuted. Almost all devices have at least two audio/video inputs: TV TUNER and COMPOSITE. Other devices may also have S-VIDEO, COMPOSITE 2, RADIO TUNER, etc.

If the device is set on TUNER mode, on modern devices, a tuner firmware will be loaded. That may require a long time. Typically, most devices take 1/2 seconds to load a firmware, but some devices may take up to 30 seconds. The firmware may depend on the TV standard that will be used, so this can't be loaded at driver warm up state.

Also, the power consumption of the tuners is high (it can be ~100-200 mW or more when powered, and ~16mW when just I2C is powered). We don't want to keep it powered when the device is not used, as this spends battery. Also, the life of the device reduces a lot if we keep it always powered.

During this stage, if an ALSA call is issued, it may interfere at the device settings and/or firmware load, with can cause the audio to fail. On such cases, applications might need to close the V4L2 node and re-open again.

3) streaming state

The change to this staging requires a V4L2 ioctl.

Please notice, however, that some apps will open the audio device before the V4L2 node, while others will open it after that.

In any case, audio will only start to produce data after the V4L2 ioctl at V4L2 that starts the DMA engine there.

After that ioctl: - Audio PCM capture will work; - The mixers will be in a good state: unmuted, and switched to the corresponding input as the video stream.

If the user wants to do something unusual, like mixing the composite audio input with the tuner audio input, it can use the ALSA mixer for doing that. Otherwise, the only part of the ALSA device that will be used is the PCM engine.

4) streaming->stop transition

When the V4L2 device is closed (or an ioctl for stream stop is issued), the driver will release the resources. What will be released vary from driver to driver.

Some drivers will stop both audio and video DMA engine (as they actually have just one DMA engine for both). Other devices will still be streaming audio until the alsa trigger stops it.

In any case, the driver will put the tuner into sleep mode. Some drivers do it immediately. Others will wait for a few seconds before doing that.

Ether way, from the PCM capture audio PoV, the device is now useless: it will either not return anything, or it will just return a stream with all zeros.

5) suspend/resume transitions

When the machine suspends to disk or memory, the DMA engine will stop, and the devices will be powered off.

The logic of returning from suspend will depend on the previous state.

If the device were in stop pode, at resume time, it should return to idle state.

If the device is streaming, it should return to streaming state. In such case, no ALSA ioctls should happen while V4L2 is suspending or resuming, otherwise the device will be on an unknown state, and will only return to work after replugging the device.

I also agree that the open/close of the alsa device is the only way to control exclusion.

This will break apps, as they can open the alsa device before configuring the stream with V4L2. They can also close alsa either before or after stopping the V4L2 stream.

-Pierre

-

As a reference, the state machine above came from this dot config, using graph-easy to produce an ascii output:

digraph media { node [shape=record] suspended[label = "<streaming> 1 | <suspended> suspended | <idle> 2"] start -> idle idle -> configuration configuration -> streaming streaming -> idle idle -> "suspended":idle "suspended":idle -> idle streaming -> "suspended":streaming "suspended":streaming -> streaming }

Em Sun, 26 Oct 2014 09:27:40 +0100 Takashi Iwai tiwai@suse.de escreveu:

At Sat, 25 Oct 2014 11:41:15 -0200, Mauro Carvalho Chehab wrote:

...

 +---------------+
 |     start     |
 +---------------+
   |
   |
   v
 +--------------------------------+
 |              idle              | <+
 +--------------------------------+  |
   |                ^           |    |
   |                |           |    |
   v                |           |    |
 +---------------+  |           |    |
 | configuration |  |           |    |
 +---------------+  |           |    |
   |                |           |    |
   |                |           |    |
   v                |           |    |
 +---------------+  |           |    |

+> | streaming | -+ | | | +---------------+ | | | | | | | | | | | v v | | +---------------+-----------+----+ | +- | 1 | suspended | 2 | -+ +---------------+-----------+----+

The above diagram is actually simplified. There's an extra state there that should be mentioned:

idle <-> DVB stream

(actually, DVB stream is a copy of the above, except that ALSA won't have any business to do while the device is on DVB mode)

So, judging from your description, the problem isn't about the exclusive lock, but rather implementation a kind of master/slave devices? Then the proposed patch doesn't look like a correct implementation to me.

You might see it as a master/slave, except that it is not that simple. It is really a hardware lock.

There are actually 3 (sub)drivers that may need to set the hardware: - ALSA driver; - V4L2 driver; - DVB driver.

The goal of the lock is to prevent that more than one driver would try to use a common piece of the hardware that it is already in usage by another driver.

So, for example, the ALSA driver should not reprogram the hardware while the V4L2 driver is doing that.

I agree that, for V4L2/ALSA driver's interaction PoV, it could resemble a sort of master/slave control, in the sense that ALSA capture start only makes sense while V4L2 is at streaming state. Yet, just opening the device or (even start capture on ALSA, for the hardware with multiple DMA engines, like the ones that use snd-usb-audio) won't cause any harm if the V4L2 driver is not reconfiguring the audio registers at the same time.

Also, ALSA open/close should be supported any time, as otherwise it will break existing applications.

However, when the device is streaming on DVB mode, it is not possible to stream on V4L2 mode, as there's just one DMA for both and just one tuner. Also, ALSA capture doesn't make much sense on such case. Still, locking on open/close may eventually break existing applications. Also, it doesn't really make any sense to block the device to move from analog to digital mode just because the ALSA devnodes are opened.

What I (and supposedly Pierre) opposed is the implementation of exclusive lock control in spontaneous callbacks. Especially the trigger callback is a bad place since it's a callback that is supposed to just trigger atomically. In general, the only good place for allowing user-space to *control* the exclusive lock is open/close, unless the finer lock control is exposed.

I see your point. Then perhaps we should expose callbacks from other parts of the ALSA core, perhaps at the logic that calls the trigger callback at read and poll syscalls, plus the corresponding logic that is used when the device is using the mmap syscall.

But, reading through the argument from you guys, the intention of the patch seems like just to raise the conflict from the hardware level to the software level.

Yes.

If so, I doubt whether such an exclusive lock is the best way. For example, audio stream can simply receive an error at any time if something is wrong and reacts accordingly instead of keeping the lock while streaming. Then the master side (video) can set the error flag, let the audio stream stop in the driver (if running), and sync with it.

Hmm... this is actually more complex than that. V4L2 driver doesn't know if ALSA is streaming or not, or even if ALSA device node is opened while he is touching at the hardware configuration or changing the state. I mean: it is not an error to set the hardware. The error only happens if ALSA and V4L2 tries to do it at the same time on an incompatible way.

Also, this won't work for DVB, as on DVB this is really an exclusive lock that would prevent both ALSA and V4L2 drivers to stream while in DVB mode.

Implementing it with a lock seems to be the best approach, at least on my eyes.

That said, we should go back and start discussing the design goal at first.

Surely.

thanks,

Takashi

To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html

Hi Shuah,

I'm understanding that you're collecting comments to write a RFC with the needs by the media token, right?

I'm sending you my contributions to such text. See enclosed.

I suggest to change the subject and submit this on a separate thread, after we finish the review of such document. Anyway, I'm changing the subject of this Thread to reflect that.

Regards, Mauro

Em Tue, 28 Oct 2014 15:15:43 -0600 Shuah Khan shuahkh@osg.samsung.com escreveu:

On 10/27/2014 06:52 AM, Mauro Carvalho Chehab wrote:

...

Em Sun, 26 Oct 2014 09:27:40 +0100 Takashi Iwai tiwai@suse.de escreveu:

...

Hmm... this is actually more complex than that. V4L2 driver doesn't know if ALSA is streaming or not, or even if ALSA device node is opened while he is touching at the hardware configuration or changing the state. I mean: it is not an error to set the hardware. The error only happens if ALSA and V4L2 tries to do it at the same time on an incompatible way.

Also, this won't work for DVB, as on DVB this is really an exclusive lock that would prevent both ALSA and V4L2 drivers to stream while in DVB mode.

Implementing it with a lock seems to be the best approach, at least on my eyes.

...

That said, we should go back and start discussing the design goal at first.

Surely.

This is long, however, hoping it will describe the problem and solution that is being pursued in detail:

Before starting with the description, this is the simplified diagram of a media device (without IR, eeprom and other beasts):

+-----------------------------------------------------------------------------------------+ | | | +----------------+ +------------------+-------+----------------+ | | | demod_video | <-- | analog | tuner | digital | | | +----------------+ +------------------+-------+----------------+ | | | | | | | | | | | v v v v | +--------------+-----+-----------------+ +------------------+ +---------------+ | | dvb | DMA | analog | | demod_audio | | digital_demux | -+ +--------------+-----+-----------------+ +------------------+ +---------------+ | | | | | | v v v +--------------+ +----------------+ +------------------+ | devnode dvr0 | | devnode video0 | | audio DMA | +--------------+ +----------------+ +------------------+ | | v +------------------+ | devnode pcmC1D0c | +------------------+

There are two components that are shared there between analog and digital: the tuner (where the signal is captured) and the DMA engine used to stream analog and Digital TV (dvb).

PS.: the diagram is over-simplified, as the tuner is just one of the possible inputs for the analog part of the device. Other possible inputs are S-Video, composite, HDMI, etc.

Sometimes, the audio DMA is also shared, e. g. just one stream comes from the hardware. It is up to the driver to split audio and video and send them to the V4L2 and ALSA APIs. This is the case of tm6000 driver.

Those shared components can be used either at analog or digital mode, but not at the same time.

Also, programming the V4L2 analog and audio DMA and demods should be done via V4L2 API, as this API allows the selection of the proper audio/video input (almost all devices have multiple analog inputs).

Please notice that, if the tuner is on digital mode, the entire analog path is disabled, including ALSA output.

If the tuner is on analog mode, both ALSA and V4L2 can work at the same time. However, during the period where the tuner firmware is loaded, and during the DMA configuration and input selection time, neither ALSA or V4L2 can stream. Such configuration/firmware load is commanded via V4L2 API, as ALSA knows nothing about tuner or input selection.

At a higher level the problem description is:

There are 3 different device files that get created to control tuner and audio functions on a media device. 3 drivers (dvb, v4l2, alsa), and 3 core apis (dvb-core, v4l-core, audio) that control the tuner and audio hardware and provide user api to these 3 device files.

There's actually a 4th component for some drivers: the mceusb driver, that handles remote controllers. The mceusb handles the Microsoft Media Center Remote Control protocol. It supports standalone remote controller devices, but it also supports a few USB devices that use a separate interface for IR.

There are currently some issues on cx231xx and mceusb, as both drivers can be used at the same time, but, when cx231xx sends certain commands, the mceusb IR polls fail. This is out of the scope of the audio lock, but it also needs to be addressed some day.

User applications, drivers and the core have no knowledge of each other. The only thing that is common across all these drivers is the parent device for the main usb device which is controlled by the main usb driver.

I would add that there are user applications that can handle all 3 APIs like MythTV. But, at least MythTV doesn't know how to associate ALSA, V4L2 and DVB devnodes that belong to the same device.

I mean: if MythTV finds, let's say, 3 V4L2 nodes, 3 ALSA nodes, and 1 DVB node, it doesn't know what device is associated with the DVB node.

Almost all applications that are aware of V4L2 API are also aware of ALSA API and may associate audio and video, as there is a way to associate it using sysfs. However, several apps don't use it.

The premise for the main design idea in this series is creating a common construct at the parent device structure that is visible to all drivers to act as a master access control (lock). Let's call this media token object with two sub-tokens one for tuner and another for audio.

Each of the apis evolved separately, hence have their own backwards compatibility to maintain. Starting with v4l2:

v4l2 case: Multiple v4l2 applications are allowed to open /dev/video0 in read/write mode with no restrictions as long as the tuner is in analog mode. v4l2 core handles conflicting requests between v4l2 applications.

It doesn't have the knowledge that the tuner is in

To be clear: "It" here refers to v4l2 core. The drivers may have this knowledge as, except for one case (bttv driver), they share some data.

use by a dvb and/or audio is in use. As soon as a v4l2 application starts, digital stream glitches and audio glitches.

dvb case: Multiple dvb applications can open the dvb device in read only mode.

There's no issue with ALSA on R/O mode, as the application is not allowed to modify anything at the stream. This is used only to monitor an already opened device in R/W mode.

As soon an application open the device read/write mode a separate kthread is kicked off to handle the request. Only one application can open the device in read/write mode.

Similar to v4l2 case,

s/v4l2/v4l2 core/

dvb-core doesn't have any knowledge that the tuner is in use by v4l2 and/or audio is in use. As soon as a dvb application starts v4l2 video glitches and audio glitches.

audio case: Same scenario is applicable to audio application. When a v4l2 or dvb application starts, audio application gets impacted.

Problems to address:

dvb owns tuner and audio: another dvb, v4l2 app and audio app should detect tuner/audio busy right away and exit.

v4l2 owns tuner and audio: another dvb and audio app should detect tuner/audio busy right away and exit.

Actually, no: audio should not exit. The V4L2 should only hold the token for the required time to initialize the device and/or load the firmware. ALSA applications should wait for V4L2 to finish programming at audio, and should keep working after that.

                       v4l2 app can continue to use it until it
                       tries to change the tuner/audio state.

audio owns audio: dvb and v4l2 apps should detect audio busy and exit.

Actually no. It is, instead:

audio owns audio: dvb apps should detect audio busy and exit. V4L2 apps should work. However, when certain V4L2 ioctls are issued, the audio device driver should not send any command to the hardware. After such commands, the audio mixers may change.

We need two separate tokens because of that: the behavior is different.

This is basically why we need two separate tokens, and because we cannot implement locking at ALSA open/close.

Special cases:

dvb apps. access tuner and audio in exclusive mode. i.e only one dvb app. at a time is allowed to open the device read/write mode.

To be clearer: dvb apps won't use the audio node, but audio should be blocked, as the devices can't use audio while in DVB mode.

As dvb apps. create threads to handle audio and video,

No. DVB apps don't handle audio/video. It receives data as MPEG-TS, using a separate device node. Yet, the same DMA engine that provides video (and, sometimes audio) is used by the DVB devnode.

all threads in that group should be allowed by the higher level construct to access the tuner and audio. dvb application will have to hold tuner and audio tokens so v4l2 and audio apps. know they are in use.

audio apps. access audio in exclusive mode. i.e only one audio app. at a time is allowed to open the device in read/write mode. Audio apps. create threads and thread closes and re-opens the audio device. Threads can do this and hence something that higher level construct has to allow. audio app. has to hold audio token so dvb and v4l2 know that it is in use. (Note: I am not sure if I have the audio scenario right)

v4l2 apps. access tuner and audio in shared v4l2 mode. i.e several v4l2 processes and threads could use tuner and audio at the same time. The higher level construct has to allow multiple v4l2 apps. to access and disallow dvb and audio apps. access when they are in use by v4l2.

Actually, V4L2 core handles concurrency. There's just one file handler with full control to start/stop stream at V4L2 side.

Adding to this, both dvb and v4l2 open audio device and make snd pcm capture callbacks.

Huh? DVB won't need to touch at PCM capture callbacks. It should just avoid audio PCM capture to stream while in DVB mode.

There is no way to tell if dvb or v4l2 or audio app is the one that is making this request.

dvb app would like audio in exclusive mode allowing only one process and its threads to access it.

No. It just wants to disable the part of the hardware that can now be powered off.

v4l2 on the other hand would like audio in shared state accessible to all v4l2 processes.

If dvb-core and v4l2-core get tuner and audio tokens at the same time, the window for having tuner token and not getting audio token go down.

No. It should not be allowed that both dvb-core and v4l2-core to get the tokens at the same time. This is an exclusive lock.

In dvb case when dvb device is opened in read/write mode, and v4l2 case when an app. tries to change the status. Audio callbacks have to detect if audio is busy, if not which mode to request the token in.

Huh?

For dvb and audio app. cases, the audio token should be requested in exclusive mode and in v4l2 case shared mode. The logic for requesting audio token will have to be try to get in exclusive mode, if fails, try to get in shared mode, and if that fails give up.

Huh?

Current status: Combining patch v1 and patch v2 designs by allowing shared mode token hold for v4l2, and deciding on where to hold audio token from alsa driver will solve the above conflict scenarios. That said, the question is "is this the right approach?" or are there other ways to solve the problem. One thing is clear, we need some common higher level construct for all the device drivers and dvb, v4l2, and audio ioctls, callbacks etc, to detect the hardware is in use.

I think that the current status is that we need to finish the spec first. Then check if the patches are doing what's above.

It seems that we agree to not agree at the requirements so far ;)

I do think lock/token approach has the best potential to solve the problems. We are at this point very close to addressing conflicts. At least the ones I am able to test.

thanks, -- Shuah

Regards, Mauro

Em Tue, 28 Oct 2014 22:00:51 -0200 Mauro Carvalho Chehab mchehab@osg.samsung.com escreveu:

Em Tue, 28 Oct 2014 21:42:50 -0200 Mauro Carvalho Chehab mchehab@osg.samsung.com escreveu:

...

Before starting with the description, this is the simplified diagram of a media device (without IR, eeprom and other beasts):

As reference, a more complete diagram would be:

An even more complete (but still simplified) diagram is shown at: http://linuxtv.org/downloads/presentations/typical_hybrid_hardware.png

The dot lines represent the parts of the graph that are switched by the tuner, DMA or input select.

Please notice that the DMA engines, together with the stuff needed to control A/V switches is at one single chip. Changing the registers there can affect the other streams, specially on most sophisticated devices like cx231xx, where it even has a power management IP block that validades if a device to be turned on/off won't exceed the maximum drain current of 500mA. That's basically why we need to do a temporary lock alsa, dvb, v4l and IR drivers when doing certain changes.

Also, please notice that I2C buses that can be as slow as 10kbps are used to control for several devices, like: - the tuner - the Digital TV (DTV) demod - Analog and/or Video demod (sometimes embedded at the main chip) - DTV demux (sometimes embedded at the main chip) - The remote controller (sometimes embedded at the main chip)

For some devices, after powered on, or when certain parameters change, a new firmware (and sometimes a hardware reset) is required. The firmware size can be about 64KB or even bigger.

Also, the A/V switch it is actually two independent switches (or one switch for video and one audio mux for audio) that needs to be changed together when the source changes.

Regards, Mauro

For those curious enough or that are using mutt/pine this is the graph, in text mode, manually adjusted to fit into 80 cols, and with a link added by hand, as graph-easy failed to represent everything on this graph:

+------------------+ | IR | +------------------+ | | v +------------------+ | IR POLL or DMA | +------------------+ | | v +------------------+ | devnode input8 | +------------------+

.................................................. : : : +-------------+-----------+---------+ : | digital | tuner | analog | : +-------------+-----------+---------+ : : : : DTV IF : : off on A/V : v : +------------------+ : | DTV demod | : +------------------+ : | : | MPEG-TS : | off on A/V : v : +------------------+ : | demux | : +------------------+ : | : | MPEG-TS : | off on A/V : v : +---------------+---------+-------------+ : | dvb | DMA | analog | <----+ : +---------------+---------+-------------+ | : : : | : : MPEG-TS Video stream : Video stream | : : off on A/V off on DTV : off on DTV | : v v | : +------------------+ +----------------+ | ATV IF : | | | | | off on DTV : | devnode dvr0 | | devnode video0 | | off on Composite : | | | | | : +------------------+ +----------------+ | : | +----+-------------------------------------------------+ | v +-------------+ | +---------+------------+-----------+ +-----------+ | video demod | ------+ | tuner | A/V switch | composite | <-- | comp. src | +-------------+ +---------+------------+-----------+ +-----------+ ^ : : : : : Audio IF ................................. : off on DTV Video IF v off on DTV +------------------+ | audio demod | +------------------+ | | Audio stream | off on DTV v +------------------+ | audio DMA | +------------------+ | | Audio stream | off on DTV v +------------------+ | devnode pcmC1D0c | +------------------+

And this is the dot file to produce it:

digraph media_hardware { node [shape=record] DMA1[label = "<dvb> dvb | <dma> DMA | <video> analog"] DMA2[label = "audio DMA"] DMA3[label = "IR POLL or DMA"] tuner[label = "<digital> digital | <tuner> tuner | <analog> analog"] input[label = "<tuner> tuner | <switch> A/V switch | <composite> composite"]

"composite src" -> input:composite

DMA1:video -> "devnode video0" [style=dotted label="Video stream\noff on DTV"] DMA1:dvb -> "devnode dvr0" [style=dotted label="MPEG-TS\noff on A/V"]

"demux" -> DMA1:dvb [label="MPEG-TS\noff on A/V"] "video demod" -> DMA1:video [label="Video stream\noff on DTV"]

tuner:digital -> "DTV demod" [style=dotted label="DTV IF\noff on A/V"] "DTV demod" -> "demux" [label="MPEG-TS\noff on A/V"] tuner:analog -> input:tuner [style=dotted label="ATV IF\noff on DTV\noff on Composite"]

input:switch -> "video demod" [style=dotted label="Video IF\noff on DTV"] input:switch -> "audio demod" [style=dotted label="Audio IF\noff on DTV"]

"audio demod" -> DMA2 [label="Audio stream\noff on DTV"]; DMA2 -> "devnode pcmC1D0c" [label="Audio stream\noff on DTV"];

IR -> DMA3 DMA3 -> "devnode input8" }

Em Wed, 29 Oct 2014 10:06:51 -0600 Shuah Khan shuahkh@osg.samsung.com escreveu:

On 10/28/2014 05:42 PM, Mauro Carvalho Chehab wrote:

...

Hi Shuah,

I'm understanding that you're collecting comments to write a RFC with the needs by the media token, right?

I'm sending you my contributions to such text. See enclosed.

I suggest to change the subject and submit this on a separate thread, after we finish the review of such document. Anyway, I'm changing the subject of this Thread to reflect that.

Regards, Mauro

Em Tue, 28 Oct 2014 15:15:43 -0600 Shuah Khan shuahkh@osg.samsung.com escreveu:

...

On 10/27/2014 06:52 AM, Mauro Carvalho Chehab wrote:

...

Em Sun, 26 Oct 2014 09:27:40 +0100 Takashi Iwai tiwai@suse.de escreveu:

...

Hmm... this is actually more complex than that. V4L2 driver doesn't know if ALSA is streaming or not, or even if ALSA device node is opened while he is touching at the hardware configuration or changing the state. I mean: it is not an error to set the hardware. The error only happens if ALSA and V4L2 tries to do it at the same time on an incompatible way.

Also, this won't work for DVB, as on DVB this is really an exclusive lock that would prevent both ALSA and V4L2 drivers to stream while in DVB mode.

Implementing it with a lock seems to be the best approach, at least on my eyes.

...

That said, we should go back and start discussing the design goal at first.

Surely.

This is long, however, hoping it will describe the problem and solution that is being pursued in detail:

Before starting with the description, this is the simplified diagram of a media device (without IR, eeprom and other beasts):

+-----------------------------------------------------------------------------------------+ | | | +----------------+ +------------------+-------+----------------+ | | | demod_video | <-- | analog | tuner | digital | | | +----------------+ +------------------+-------+----------------+ | | | | | | | | | | | v v v v | +--------------+-----+-----------------+ +------------------+ +---------------+ | | dvb | DMA | analog | | demod_audio | | digital_demux | -+ +--------------+-----+-----------------+ +------------------+ +---------------+ | | | | | | v v v +--------------+ +----------------+ +------------------+ | devnode dvr0 | | devnode video0 | | audio DMA | +--------------+ +----------------+ +------------------+ | | v +------------------+ | devnode pcmC1D0c | +------------------+

There are two components that are shared there between analog and digital: the tuner (where the signal is captured) and the DMA engine used to stream analog and Digital TV (dvb).

PS.: the diagram is over-simplified, as the tuner is just one of the possible inputs for the analog part of the device. Other possible inputs are S-Video, composite, HDMI, etc.

Sometimes, the audio DMA is also shared, e. g. just one stream comes from the hardware. It is up to the driver to split audio and video and send them to the V4L2 and ALSA APIs. This is the case of tm6000 driver.

Those shared components can be used either at analog or digital mode, but not at the same time.

Also, programming the V4L2 analog and audio DMA and demods should be done via V4L2 API, as this API allows the selection of the proper audio/video input (almost all devices have multiple analog inputs).

Please notice that, if the tuner is on digital mode, the entire analog path is disabled, including ALSA output.

If the tuner is on analog mode, both ALSA and V4L2 can work at the same time. However, during the period where the tuner firmware is loaded, and during the DMA configuration and input selection time, neither ALSA or V4L2 can stream. Such configuration/firmware load is commanded via V4L2 API, as ALSA knows nothing about tuner or input selection.

...

At a higher level the problem description is:

There are 3 different device files that get created to control tuner and audio functions on a media device. 3 drivers (dvb, v4l2, alsa), and 3 core apis (dvb-core, v4l-core, audio) that control the tuner and audio hardware and provide user api to these 3 device files.

There's actually a 4th component for some drivers: the mceusb driver, that handles remote controllers. The mceusb handles the Microsoft Media Center Remote Control protocol. It supports standalone remote controller devices, but it also supports a few USB devices that use a separate interface for IR.

There are currently some issues on cx231xx and mceusb, as both drivers can be used at the same time, but, when cx231xx sends certain commands, the mceusb IR polls fail. This is out of the scope of the audio lock, but it also needs to be addressed some day.

...

User applications, drivers and the core have no knowledge of each other. The only thing that is common across all these drivers is the parent device for the main usb device which is controlled by the main usb driver.

I would add that there are user applications that can handle all 3 APIs like MythTV. But, at least MythTV doesn't know how to associate ALSA, V4L2 and DVB devnodes that belong to the same device.

I mean: if MythTV finds, let's say, 3 V4L2 nodes, 3 ALSA nodes, and 1 DVB node, it doesn't know what device is associated with the DVB node.

Almost all applications that are aware of V4L2 API are also aware of ALSA API and may associate audio and video, as there is a way to associate it using sysfs. However, several apps don't use it.

...

The premise for the main design idea in this series is creating a common construct at the parent device structure that is visible to all drivers to act as a master access control (lock). Let's call this media token object with two sub-tokens one for tuner and another for audio.

Each of the apis evolved separately, hence have their own backwards compatibility to maintain. Starting with v4l2:

v4l2 case: Multiple v4l2 applications are allowed to open /dev/video0 in read/write mode with no restrictions as long as the tuner is in analog mode. v4l2 core handles conflicting requests between v4l2 applications.

...

It doesn't have the knowledge that the tuner is in

To be clear: "It" here refers to v4l2 core. The drivers may have this knowledge as, except for one case (bttv driver), they share some data.

...

use by a dvb and/or audio is in use. As soon as a v4l2 application starts, digital stream glitches and audio glitches.

dvb case: Multiple dvb applications can open the dvb device in read only mode.

There's no issue with ALSA on R/O mode, as the application is not allowed to modify anything at the stream. This is used only to monitor an already opened device in R/W mode.

...

As soon an application open the device read/write mode a separate kthread is kicked off to handle the request. Only one application can open the device in read/write mode.

...

Similar to v4l2 case,

s/v4l2/v4l2 core/

...

dvb-core doesn't have any knowledge that the tuner is in use by v4l2 and/or audio is in use. As soon as a dvb application starts v4l2 video glitches and audio glitches.

audio case: Same scenario is applicable to audio application. When a v4l2 or dvb application starts, audio application gets impacted.

Problems to address:

dvb owns tuner and audio: another dvb, v4l2 app and audio app should detect tuner/audio busy right away and exit.

v4l2 owns tuner and audio: another dvb and audio app should detect tuner/audio busy right away and exit.

Actually, no: audio should not exit. The V4L2 should only hold the token for the required time to initialize the device and/or load the firmware. ALSA applications should wait for V4L2 to finish programming at audio, and should keep working after that.

...

                       v4l2 app can continue to use it until it
                       tries to change the tuner/audio state.

audio owns audio: dvb and v4l2 apps should detect audio busy and exit.

Actually no. It is, instead:

audio owns audio: dvb apps should detect audio busy and exit. V4L2 apps should work. However, when certain V4L2 ioctls are issued, the audio device driver should not send any command to the hardware. After such commands, the audio mixers may change.

We need two separate tokens because of that: the behavior is different.

This is basically why we need two separate tokens, and because we cannot implement locking at ALSA open/close.

...

Special cases:

dvb apps. access tuner and audio in exclusive mode. i.e only one dvb app. at a time is allowed to open the device read/write mode.

To be clearer: dvb apps won't use the audio node, but audio should be blocked, as the devices can't use audio while in DVB mode.

...

As dvb apps. create threads to handle audio and video,

No. DVB apps don't handle audio/video. It receives data as MPEG-TS, using a separate device node. Yet, the same DMA engine that provides video (and, sometimes audio) is used by the DVB devnode.

...

all threads in that group should be allowed by the higher level construct to access the tuner and audio. dvb application will have to hold tuner and audio tokens so v4l2 and audio apps. know they are in use.

audio apps. access audio in exclusive mode. i.e only one audio app. at a time is allowed to open the device in read/write mode. Audio apps. create threads and thread closes and re-opens the audio device. Threads can do this and hence something that higher level construct has to allow. audio app. has to hold audio token so dvb and v4l2 know that it is in use. (Note: I am not sure if I have the audio scenario right)

v4l2 apps. access tuner and audio in shared v4l2 mode. i.e several v4l2 processes and threads could use tuner and audio at the same time. The higher level construct has to allow multiple v4l2 apps. to access and disallow dvb and audio apps. access when they are in use by v4l2.

Actually, V4L2 core handles concurrency. There's just one file handler with full control to start/stop stream at V4L2 side.

...

Adding to this, both dvb and v4l2 open audio device and make snd pcm capture callbacks.

Huh? DVB won't need to touch at PCM capture callbacks. It should just avoid audio PCM capture to stream while in DVB mode.

...

There is no way to tell if dvb or v4l2 or audio app is the one that is making this request.

...

dvb app would like audio in exclusive mode allowing only one process and its threads to access it.

No. It just wants to disable the part of the hardware that can now be powered off.

Hmm. I am seeing some snd_pcm_lib_ioctls coming from dvb application.

Well, probably it is related to the audio output and not audio capture (or the application you're looking into is an hybrid one).

...

From what you are saying, these could be for poweroff.

...

...

v4l2 on the other hand would like audio in shared state accessible to all v4l2 processes.

...

If dvb-core and v4l2-core get tuner and audio tokens at the same time, the window for having tuner token and not getting audio token go down.

No. It should not be allowed that both dvb-core and v4l2-core to get the tokens at the same time. This is an exclusive lock.

I should have explained this better. What I meant was that dvb-core when it gets the tuner, it should also obtain audio right away. v4l2-core when it gets the tuner, it should get the audio at the same time. When dvb-core has the tuner, v4l2 shouldn't get it and vice versa.

Yep.

...

...

In dvb case when dvb device is opened in read/write mode, and v4l2 case when an app. tries to change the status. Audio callbacks have to detect if audio is busy, if not which mode to request the token in.

Huh?

ok again bad explaining on my part. Let me try again. When dvb-core has the audio locked, audio application should detect the condition and take appropriate action. When v4l2-core has audio locked, audio application should detect the condition and take appropriate action.

...

...

For dvb and audio app. cases, the audio token should be requested in exclusive mode and in v4l2 case shared mode. The logic for requesting audio token will have to be try to get in exclusive mode, if fails, try to get in shared mode, and if that fails give up.

Huh?

Consider this flow:

step 1: dvb-core locks tuner and audio. step 2: audio ioctl is initiated (from the application) step 3: If dvb-core has the audio locked, how does alsa know if it can proceed with the ioctl request or not?

That's simple: it should stop streaming, as part of the hardware can be powered off. It can only return opening the device after DVB releases.

...

...

Current status: Combining patch v1 and patch v2 designs by allowing shared mode token hold for v4l2, and deciding on where to hold audio token from alsa driver will solve the above conflict scenarios. That said, the question is "is this the right approach?" or are there other ways to solve the problem. One thing is clear, we need some common higher level construct for all the device drivers and dvb, v4l2, and audio ioctls, callbacks etc, to detect the hardware is in use.

I think that the current status is that we need to finish the spec first. Then check if the patches are doing what's above.

It seems that we agree to not agree at the requirements so far ;)

I would say it is more of not having the same understanding of the requirements as opposed to wanting to agree to disagree.

True.

Right I agree that developing a clear RFC spec will help us all be on the same page while I continue to find solution for this problem, and continuing work towards patch v3.

I will compile what we discussed so far in RFC and send it out for review in a day or two. I will include your diagrams and the scenarios I put together with your corrections in it.

Ok. After that, we can proceed with the token patch v3, or to see if it is possible to use the media controller here, as Sakari suggested.

Regards, Mauro

Hi Shuah,

This looks good. I have a few remarks, see below...

On 11/05/2014 12:08 AM, Shuah Khan wrote:

Hi Mauro,

Here is the RFC as promised. I also included the Media controller as a an alternative and captured the discussion in the thread on that topic. Please review.

-- Shuah

---

RFC Media Token API Specification

Let's start with a diagram of a media device (without IR, eeprom and others):

http://linuxtv.org/downloads/presentations/typical_hybrid_hardware.png

The dot lines represent the parts of the graph that are switched by the tuner, DMA or input select.

Please notice that the DMA engines, together with the stuff needed to control A/V switches is at one single chip. Changing the registers there can affect the other streams, specially on most sophisticated devices like cx231xx, where it even has a power management IP block that validates if a device to be turned on/off won't exceed the maximum drain current of 500mA. That's basically why we need to do a temporary lock alsa, dvb, v4l and IR drivers when doing certain changes.

Also, please notice that I2C buses that can be as slow as 10kbps are used to control for several devices, like: - the tuner - the Digital TV (DTV) demod - Analog and/or Video demod (sometimes embedded at the main chip) - DTV demux (sometimes embedded at the main chip) - The remote controller (sometimes embedded at the main chip)

For some devices, after powered on, or when certain parameters change, a new firmware (and sometimes a hardware reset) is required. The firmware size can be about 64KB or even bigger.

Also, the A/V switch it is actually two independent switches (or one switch for video and one audio mux for audio) that needs to be changed together when the source changes.

There are two components that are shared there between analog and digital: the tuner (where the signal is captured) and the DMA engine used to stream analog and Digital TV (dvb).

PS.: the diagram is over-simplified, as the tuner is just one of the possible inputs for the analog part of the device. Other possible inputs are S-Video, composite, HDMI, etc.

Sometimes, the audio DMA is also shared, e. g. just one stream comes from the hardware. It is up to the driver to split audio and video and send them to the V4L2 and ALSA APIs. This is the case of tm6000 driver.

Those shared components can be used either at analog or digital mode, but not at the same time.

At least in theory there may be independent DMA engines for analog and digital video, which means that you can have DVB capture and Composite analog capture (or any other analog input as long as it isn't from the tuner). I actually suspect that one or more of the conexant devices can do that. But I doubt we support that in the driver and I'm not sure we want that anyway.

Also, programming the V4L2 analog and audio DMA and demods should be done via V4L2 API, as this API allows the selection of the proper audio/video input (almost all devices have multiple analog inputs).

Please notice that, if the tuner is on digital mode, the entire analog path is disabled, including ALSA output.

If the tuner is on analog mode, both ALSA and V4L2 can work at the same time. However, during the period where the tuner firmware is loaded, and during the DMA configuration and input selection time, neither ALSA or V4L2 can stream. Such configuration/firmware load is commanded via V4L2 API, as ALSA knows nothing about tuner or input selection.

At a higher level the problem description is:

There are 3 different device files that get created to control tuner and audio functions on a media device. 4 drivers (dvb, v4l2, alsa, and the main usb driver for the usb device), and 3 core APIs (dvb-core, v4l-core, audio) that control the tuner and audio hardware and provide user API to these 3 device files.

The above driver model is simplified, there's 4th component for some drivers: the mceusb driver, that handles remote controllers. The mceusb handles the Microsoft Media Center Remote Control protocol. It supports stand alone remote controller devices, but it also supports a few USB devices that use a separate interface for IR.

There are currently some issues on cx231xx and mceusb, as both drivers can be used at the same time, but, when cx231xx sends certain commands, the mceusb IR polls fail. This is out of the scope of the audio lock, but it also needs to be addressed some day.

Most media user applications, drivers and the core have no knowledge of each other. The only thing that is common across all these drivers is the parent device for the main usb device which is controlled by the usb driver.

Some media user applications like MythTV can handle all 3 APIs, however, MythTV doesn't know how to associate ALSA, V4L2 and DVB devnodes that belong to the same device. If MythTV finds, 3 V4L2 nodes, 3 ALSA nodes, and 1 DVB node, it doesn't know which device is associated with the DVB node.

Almost all applications that are aware of V4L2 API are also aware of ALSA API and may associate audio and video, as there is a way to associate it using sysfs. However, several applications don't use it.

The premise for the main design idea in this series is creating a common construct at the parent device structure that is visible to all drivers to act as a master access control (lock). Let's call this media token object with two sub-tokens one for tuner and another for audio.

Each of the APIS evolved separately, hence have their own backwards compatibility to maintain. Starting with v4l2:

V4L2 case: Multiple v4l2 applications are allowed to open /dev/video0 in read/write mode with no restrictions as long as the tuner is in analog mode. V4L2-core handles conflicting requests between v4l2 applications. V4L2-core doesn't have the knowledge that the tuner is in use by a dvb and/or audio is in use. Individual drivers may have this knowledge as, except for one case (bttv driver), they share some data.

As soon as a V4L2 application starts, digital stream glitches and audio glitches.

DVB case: Multiple dvb applications can open the dvb device in read only mode. As soon an application open the device read/write mode a separate kthread is kicked off to handle the request. Only one application can open the device in read/write mode. There's no issue with ALSA in R/O mode, as the application is not allowed to modify anything with the stream. This is used only to monitor an already opened device in R/W mode.

Similar to V4L2-core case, dvb-core doesn't have any knowledge that the tuner is in use by v4l2 and/or audio is in use. As soon as a dvb application starts v4l2 video glitches and audio glitches.

Audio case: Same scenario is applicable to audio application. When a v4l2 or dvb application starts, audio application gets impacted.

Problems to address:

Dvb owns tuner and audio: another dvb, v4l2 application and ALSA application should detect tuner/audio busy right away and exit. Dvb applications don't use audio node, however, devices can't use audio hardware while in DVB mode.

V4l2 owns tuner and audio: dvb should detect tuner/audio busy right away and exit. The V4L2-core should only hold the token for the required time to initialize the device and/or load the firmware. ALSA applications should wait for v4L2-core to finish programming at audio, and should keep working after that.

Audio owns audio: dvb applications should detect audio busy and exit. V4L2 applications should work. However, when certain V4L2 ioctls are issued, the audio device driver should not send any command to the hardware. After such commands, the audio mixers may change. This is why two tokens are necessary, one for tuner and another for audio.

Because of the above mentioned difference in behavior between dvb and v4l applications when audio is busy, two tokens (one for tuner and another for audio) are necessary and audio token lock should not be held at ALSA open/close.

Special cases:

Dvb applications access tuner in exclusive mode. i.e only one dvb application at a time is allowed to open the device read/write mode. Dvb applications don't use audio node, however, devices can't use audio hardware while in DVB mode. Dvb applications receive data as MPEG-TS, using a separate device node. The same DMA engine that provides video (and, sometimes audio) is used by the DVB device node, making it inaccessible to audio applications while tuner is in DVB mode. Hence, the need to prevent audio applications from accessing audio node when tuner is in DVB mode. As a result, dvb-core will have to hold tuner and audio tokens so v4l2-core and ALSA know that audio is not available. Dvb disables audio hardware so it could be powered-off in some cases.

Audio applications access audio in exclusive mode. i.e only one audio application at a time is allowed to open the device in read/write mode. Audio applications create threads and thread closes and re-opens the audio device. Threads can do this and hence something that higher level construct has to allow. Audio application has to hold audio token so dvb and v4l2 know that it is in use.

V4l2 applications access tuner and audio in shared v4l2 mode. i.e several v4l2 processes and threads could use tuner and audio at the same time. V4L2 core handles concurrency. There's just one file handler with full control to start/stop stream at V4L2 side. The higher level construct should not break the ability of multiple v4l2 applications to access tuner and audio in shared mode, and disallow dvb and audio applications access when they are in use by the V4L2-core.

Dvb-core when it gets the tuner, it should also obtain audio right away. v4l2-core when it gets the tuner, it should get the audio at the same time. When dvb-core has the tuner, v4l2 shouldn't get it and vice versa.

When dvb-core has the audio locked, audio application should detect condition and stop streaming, as part of the hardware can be powered off. It can only return opening the device after DVB releases audio token.

When v4l2-core has audio locked, audio application should detect the condition and stop sending commands to audio hardware. It can only resume audio access after V4L2 releases audio token.

Open issues: During testing, snd_pcm_lib_ioctls are coming from dvb application. It is likely that these are related to the audio output and not audio capture or the application in question is an hybrid one. This issue needs further investigation.

Alternatives: (proposed by Sakari Ailus) Can Media controller be used to solve the problem?

I don't think the MC has anything to do with this. The problem is with existing non-MC drivers and existing applications. Adding MC support to drivers will not help this in any way that I can see.

In the end this is simply a resource ownership problem: who owns the tuner? The only thing special about it is that it crosses subsystems.

The usage of the media controller for this specific usage is that we should not force userspace applications to be aware of the media controller just because of hardware locking.

Currently, media entities may only be entities bound to a given subsystem, but likely need to change media controller for complex embedded DVB device support ...

In case of the Media controller, mutual exclusion of different users is currently performed by adding the entities to a pipeline and incrementing the streaming count once streaming is enabled --- on different interfaces streaming may mean a different thing.

However, we'll still need to find a way for ALSA to prevent it to use the audio demod and DMA engine that will be powered off when DVB is streaming.

The Media controller interface does not handle serializing potential users that may wish to configure the device. Handling serializing is necessary if Media controller is extended instead of pursuing Media Token API to solve the problem.

Reconfiguring the DMA engine and some other registers via V4L2 API should be blocked. The same applies to firmware load, if the device is using tuner input for analog TV.

If we use the media controller, we'll need to add a state to it, to indicate that a block at the pipeline is being reconfigured.

It is dependent on Media Controller adoption on ALSA as well.

Next steps:

1. Review RFC
2. Evaluate the media controller alternative.
3. Proceed with development once the decision is made.

---

Regards,

Hans

On 11/18/2014 02:15 PM, Sakari Ailus wrote:

Hi Shuah,

Many thanks for the RFC, for and being so patient with my terrible review schedule. My comments below.

On Tue, Nov 04, 2014 at 04:08:50PM -0700, Shuah Khan wrote:

...

Hi Mauro,

Here is the RFC as promised. I also included the Media controller as a an alternative and captured the discussion in the thread on that topic. Please review.

-- Shuah

---

RFC Media Token API Specification

Let's start with a diagram of a media device (without IR, eeprom and others):

http://linuxtv.org/downloads/presentations/typical_hybrid_hardware.png

The dot lines represent the parts of the graph that are switched by the tuner, DMA or input select.

Please notice that the DMA engines, together with the stuff needed to control A/V switches is at one single chip. Changing the registers there can affect the other streams, specially on most sophisticated devices like cx231xx, where it even has a power management IP block that validates if a device to be turned on/off won't exceed the maximum drain current of 500mA. That's basically why we need to do a temporary lock alsa, dvb, v4l and IR drivers when doing certain changes.

Also, please notice that I2C buses that can be as slow as 10kbps are used to control for several devices, like: - the tuner - the Digital TV (DTV) demod - Analog and/or Video demod (sometimes embedded at the main chip) - DTV demux (sometimes embedded at the main chip) - The remote controller (sometimes embedded at the main chip)

For some devices, after powered on, or when certain parameters change, a new firmware (and sometimes a hardware reset) is required. The firmware size can be about 64KB or even bigger.

Also, the A/V switch it is actually two independent switches (or one switch for video and one audio mux for audio) that needs to be changed together when the source changes.

There are two components that are shared there between analog and digital: the tuner (where the signal is captured) and the DMA engine used to stream analog and Digital TV (dvb).

PS.: the diagram is over-simplified, as the tuner is just one of the possible inputs for the analog part of the device. Other possible inputs are S-Video, composite, HDMI, etc.

Sometimes, the audio DMA is also shared, e. g. just one stream comes from the hardware. It is up to the driver to split audio and video and send them to the V4L2 and ALSA APIs. This is the case of tm6000 driver.

Those shared components can be used either at analog or digital mode, but not at the same time.

Also, programming the V4L2 analog and audio DMA and demods should be done via V4L2 API, as this API allows the selection of the proper audio/video input (almost all devices have multiple analog inputs).

Please notice that, if the tuner is on digital mode, the entire analog path is disabled, including ALSA output.

If the tuner is on analog mode, both ALSA and V4L2 can work at the same time. However, during the period where the tuner firmware is loaded, and during the DMA configuration and input selection time, neither ALSA or V4L2 can stream. Such configuration/firmware load is commanded via V4L2 API, as ALSA knows nothing about tuner or input selection.

At a higher level the problem description is:

There are 3 different device files that get created to control tuner and audio functions on a media device. 4 drivers (dvb, v4l2, alsa, and the main usb driver for the usb device), and 3 core APIs (dvb-core, v4l-core, audio) that control the tuner and audio hardware and provide user API to these 3 device files.

The above driver model is simplified, there's 4th component for some drivers: the mceusb driver, that handles remote controllers. The mceusb handles the Microsoft Media Center Remote Control protocol. It supports stand alone remote controller devices, but it also supports a few USB devices that use a separate interface for IR.

There are currently some issues on cx231xx and mceusb, as both drivers can be used at the same time, but, when cx231xx sends certain commands, the mceusb IR polls fail. This is out of the scope of the audio lock, but it also needs to be addressed some day.

Most media user applications, drivers and the core have no knowledge of each other. The only thing that is common across all these drivers is the parent device for the main usb device which is controlled by the usb driver.

Some media user applications like MythTV can handle all 3 APIs, however, MythTV doesn't know how to associate ALSA, V4L2 and DVB devnodes that belong to the same device. If MythTV finds, 3 V4L2 nodes, 3 ALSA nodes, and 1 DVB node, it doesn't know which device is associated with the DVB node.

Almost all applications that are aware of V4L2 API are also aware of ALSA API and may associate audio and video, as there is a way to associate it using sysfs. However, several applications don't use it.

The premise for the main design idea in this series is creating a common construct at the parent device structure that is visible to all drivers to act as a master access control (lock). Let's call this media token object with two sub-tokens one for tuner and another for audio.

Each of the APIS evolved separately, hence have their own backwards compatibility to maintain. Starting with v4l2:

V4L2 case: Multiple v4l2 applications are allowed to open /dev/video0 in read/write mode with no restrictions as long as the tuner is in analog mode. V4L2-core handles conflicting requests between v4l2 applications. V4L2-core doesn't have the knowledge that the tuner is in use by a dvb and/or audio is in use. Individual drivers may have this knowledge as, except for one case (bttv driver), they share some data.

As soon as a V4L2 application starts, digital stream glitches and audio glitches.

DVB case: Multiple dvb applications can open the dvb device in read only mode. As soon an application open the device read/write mode a separate kthread is kicked off to handle the request. Only one application can open the device in read/write mode. There's no issue with ALSA in R/O mode, as the application is not allowed to modify anything with the stream. This is used only to monitor an already opened device in R/W mode.

Similar to V4L2-core case, dvb-core doesn't have any knowledge that the tuner is in use by v4l2 and/or audio is in use. As soon as a dvb application starts v4l2 video glitches and audio glitches.

Audio case: Same scenario is applicable to audio application. When a v4l2 or dvb application starts, audio application gets impacted.

Problems to address:

Dvb owns tuner and audio: another dvb, v4l2 application and ALSA application should detect tuner/audio busy right away and exit. Dvb applications don't use audio node, however, devices can't use audio hardware while in DVB mode.

V4l2 owns tuner and audio: dvb should detect tuner/audio busy right away and exit. The V4L2-core should only hold the token for the required time to initialize the device and/or load the firmware. ALSA applications should wait for v4L2-core to finish programming at audio, and should keep working after that.

Audio owns audio: dvb applications should detect audio busy and exit. V4L2 applications should work. However, when certain V4L2 ioctls are issued, the audio device driver should not send any command to the hardware. After such commands, the audio mixers may change. This is why two tokens are necessary, one for tuner and another for audio.

Because of the above mentioned difference in behavior between dvb and v4l applications when audio is busy, two tokens (one for tuner and another for audio) are necessary and audio token lock should not be held at ALSA open/close.

Special cases:

Dvb applications access tuner in exclusive mode. i.e only one dvb application at a time is allowed to open the device read/write mode. Dvb applications don't use audio node, however, devices can't use audio hardware while in DVB mode. Dvb applications receive data as MPEG-TS, using a separate device node. The same DMA engine that provides video (and, sometimes audio) is used by the DVB device node, making it inaccessible to audio applications while tuner is in DVB mode. Hence, the need to prevent audio applications from accessing audio node when tuner is in DVB mode. As a result, dvb-core will have to hold tuner and audio tokens so v4l2-core and ALSA know that audio is not available. Dvb disables audio hardware so it could be powered-off in some cases.

Audio applications access audio in exclusive mode. i.e only one audio application at a time is allowed to open the device in read/write mode. Audio applications create threads and thread closes and re-opens the audio device. Threads can do this and hence something that higher level construct has to allow. Audio application has to hold audio token so dvb and v4l2 know that it is in use.

V4l2 applications access tuner and audio in shared v4l2 mode. i.e several v4l2 processes and threads could use tuner and audio at the same time. V4L2 core handles concurrency. There's just one file handler with full control to start/stop stream at V4L2 side. The higher level construct should not break the ability of multiple v4l2 applications to access tuner and audio in shared mode, and disallow dvb and audio applications access when they are in use by the V4L2-core.

Dvb-core when it gets the tuner, it should also obtain audio right away. v4l2-core when it gets the tuner, it should get the audio at the same time. When dvb-core has the tuner, v4l2 shouldn't get it and vice versa.

When dvb-core has the audio locked, audio application should detect condition and stop streaming, as part of the hardware can be powered off. It can only return opening the device after DVB releases audio token.

When v4l2-core has audio locked, audio application should detect the condition and stop sending commands to audio hardware. It can only resume audio access after V4L2 releases audio token.

Open issues: During testing, snd_pcm_lib_ioctls are coming from dvb application. It is likely that these are related to the audio output and not audio capture or the application in question is an hybrid one. This issue needs further investigation.

Alternatives: (proposed by Sakari Ailus) Can Media controller be used to solve the problem?

The usage of the media controller for this specific usage is that we should not force userspace applications to be aware of the media controller just because of hardware locking.

Currently, media entities may only be entities bound to a given subsystem, but likely need to change media controller for complex embedded DVB device support ...

In case of the Media controller, mutual exclusion of different users is currently performed by adding the entities to a pipeline and incrementing the streaming count once streaming is enabled --- on different interfaces streaming may mean a different thing.

However, we'll still need to find a way for ALSA to prevent it to use the audio demod and DMA engine that will be powered off when DVB is streaming.

The Media controller interface does not handle serializing potential users that may wish to configure the device. Handling serializing is necessary if Media controller is extended instead of pursuing Media Token API to solve the problem.

Reconfiguring the DMA engine and some other registers via V4L2 API should be blocked. The same applies to firmware load, if the device is using tuner input for analog TV.

If we use the media controller, we'll need to add a state to it, to indicate that a block at the pipeline is being reconfigured.

It is dependent on Media Controller adoption on ALSA as well.

Thank you for the detailed description of the problem domain.

Using Media controller for this at this point isn't straightforward, and especially for existing drivers for which the current APIs serve the existing devices well enough, perhaps not the best solution even when the missing pieces were implemented.

More complex devices, though, still need MC in order to control them in a meaningful way. If a tuner is connected into one, the media token framework wouldn't help there. In other words, we'll need something for such devices as well. That'd be proper MC support for DVB and ALSA, but it'll be a separate discussion.

One big upside in this patchset is that it does not change the user space interface.

Right that is the goal.

One concern I have is how generic this framework really is. Do you see potential use cases outside the current one?

I'd move this under drivers/media, and possibly think of the naming of the framework a little bit.

Others gave the same feedback on both of your points on generic and location. I moved it under drivers/media for the patch v3 pre-work I did before we started the RFC discussion. So we are good there.

I'll review the rest of the v2.

Great. Please do and give me feedback. I will wait for that before I get started with patch v3 work.

thanks, -- Shuah

Hi Shuah,

Some notes below...

On 10/21/2014 07:32 PM, Shuah Khan wrote:

On 10/21/2014 10:05 AM, Takashi Iwai wrote:

...

At Tue, 21 Oct 2014 17:42:51 +0200, Hans Verkuil wrote:

...

...

Quite often media apps open the alsa device at the start and then switch between TV, radio or DVB mode. If the alsa device would claim the tuner just by being opened (as opposed to actually using the tuner, which happens when you start streaming),

What about parameter changes? The sound devices have to be configured before using. Don't they influence on others at all, i.e. you can change the PCM sample rate etc during TV, radio or DVB is running?

Yes. kaffeine uses snd_usb_capture_ops ioctl -> snd_pcm_lib_ioctl

Other v4l and vlc (dvb) uses open/close as well as trigger start and stop. trigger start/stop is done by a special audio thread in some cases. open/close happens from the main thread.

...

...

then that would make it impossible for the application to switch tuner mode. In general you want to avoid that open() will start configuring hardware since that can quite often be slow. Tuner configuration in particular can be slow since several common tuners need to load firmware over i2c. You only want to do that when it is really needed, and not when some application (udev!) opens the device just to examine what sort of device it is.

But most apps close the device soon after that, no? Which programs keep the PCM device (not the control) opened without actually using?

...

So claiming the tuner in the trigger seems to be the right place. If returning EBUSY is a poor error code for alsa, then we can use something else for that. EACCES perhaps?

Sorry, I'm not convinced by that. If the device has to be controlled exclusively, the right position is the open/close. Otherwise, the program cannot know when it becomes inaccessible out of sudden during its operation.

Let me share my test matrix for this patch series. Hans pointed out one test case I didn't know about as a result missed testing. Please see if any of the tests miss use-cases or break them: you can scroll down to the proposal at the end, if this is too much detail :)

Digital active and analog starting testing: kaffeine running

• v4l2-ctl --all - works

I would recommend using v4l2-ctl --all --list-inputs, this enumerates inputs as well (which includes reading the tuner status). This would have failed with all these tests with this patch series.

• Changing channels works with the same token hold, even when frequency changes. Tested changing channels that force freq change.
• vlc resource is busy with no disruption to kaffeine
• xawtv - tuner busy when it tries to do ioctls that change tuner settings - snd_usb_pcm_open detects device is busy ( pcm open called from the same thread, trigger gets called from another thread )
• tvtime - tuner busy when it tries to do ioctls that change tuner settings with no disruption to kaffeine ( pcm open called from the same thread, trigger gets called from another thread )
• vlc - audio capture on WinTV HVR-950 - device is busy start vlc with no channels for this test
• arecord to capture on WinTV HVR-950 - device busy

vlc running vlc -v channels.xspf

• v4l2-ctl --all - works
• Changing channels works with the same token hold, even when frequency changes. Tested changing channels that force freq change.
• kaffeine resource is busy with no disruption to vlc
• xawtv - tuner busy when it tries to do ioctls that change tuner settings - snd_usb_pcm_open detects device is busy ( pcm open called from the same thread, trigger gets called from another thread )
• tvtime - tuner busy when it tries to do ioctls that change tuner settings with no disruption to kaffeine ( pcm open called from the same thread, trigger gets called from another thread )
• vlc - audio capture on WinTV HVR-950 - device is busy
• arecord to capture on WinTV HVR-950 - device busy

Analog active and start digital testing: xawtv -noalsa -c /dev/video1

• v4l2-ctl --all - works
• start kaffeine - fails with device busy and no disruption
• start vlc - fails with device busy and no disruption
• tvtime - tuner busy when it tries to do ioctls that change tuner settings with no disruption to kaffeine
• vlc - audio capture on WinTV HVR-950 - device is busy
• arecord to capture on WinTV HVR-950 - device busy

tvtime

• v4l2-ctl --all - works
• start kaffeine - fails with device busy and no disruption
• start vlc - fails with device busy and no disruption
• xawtv - tuner busy when it tries to do ioctls that change tuner settings with no disruption to kaffeine
• vlc - audio capture on WinTV HVR-950 - device is busy
• arecord to capture on WinTV HVR-950 - device busy

The following audio/video start/stop combination tests: ( used arecord as well to test these cases, arecord )

• tvtime start/vlc start/vlc stop/tvtime stop no disruption to tvtime
• tvtime start/vlc start/tvtie stop/vlc stop One tvtime stops, could trigger capture manually
• vlc start/tvtime start/tvtime stop/vlc stop vlc audio capture continues, tvtime detect tuner busy
• vlc start/tvtime start/vlc stop/tvtime start when vlc stops, tvtime could open the tuner device

Repeated the above with kaffeine and vlc and arecord audio capture combinations.

Hans pointed out I am missing:

v4l2-ctl -f 180 --sleep=10 While it is sleeping you must still be able to set the frequency from another console.

And it doesn't matter which of the two v4l2-ctl processes ends first, as long as one has a reference to the tuner you should be blocked from switching the tuner to a different mode (radio or dvb)

I think this above fails because tuner token ownership doesn't span processes. Token should act as a lock between dvb/audio/v4l, and v4l itself has mechanisms to handle the multiple ownership token shouldn't interfere with.

Provided the only thing that is breaking is the v4l case, I think I can get it to work with the bit field approach.

Here is what I propose for patch v3:

• make it a module under media
• collapse tuner and audio tokens into media token
• change names (get rid of abbreviated tkn stuff)
• Make other changes Takashi/Lars pointed out in pcm
• hold token in pcm open/close

I remain skeptical about this. I'm simply not sure what, if anything, might break after this change. You need to test with applications that can switch between V4L and DVB and that can handle an alsa device for audio capture. MythTV is probably one. None of the applications you are testing can switch between V4L and DVB devices dynamically to my knowledge.

Today it is no problem for an application to open the alsa device, start streaming analog video + audio, stop streaming video + audio, switch to a digital input and start streaming the transport stream from the dvb input.

This is all valid. But after this change you will no longer be able to switch back and forth because the alsa open() locks the tuner mode.

I'm afraid that we will break some applications, either open source or not, by making this change.

By locking the tuner mode only while streaming audio apps would be prevented from breaking. The only time the trigger function will return an error is when some other application has the tuner in a different (e.g. DVB) mode and starting to stream would forcefully switch the tuner to a different mode. And we *know* that will cause all sorts of unpredictable behavior which is why we are doing all this effort to prevent that from happening. Returning an error in that case makes perfect sense to me and it is something that should not happen during normal use.

• add a bitfield to struct v4l2_fh to handle the v4l specific multiple ownership cases.
• v4l-core and au0828-videp patches will see changes.
• dvb patch should still be good (crossing fingers)

The remainder of the list looks OK to me.

I would also recommend making more use of the various command line tools: arecord, v4l2-ctl, dvbv5-zap. By combining these you can test scenarios that are hard to test in applications such as this:

- make a small utility that just opens the alsa node without streaming and that keeps it open (i.e. sleep(100000) or something). Let's call it alsaopen. - run alsaopen. - call arecord to start streaming audio (most likely using the analog TV tuner). - stream with v4l2-ctl to see if that works (it should) - stop v4l2-ctl - now use dvbv5-zap to select a DVB channel: I would expect that that fails because you are still streaming audio. - stop arecord. - try dvbv5-zap again: I think this should work, even though the alsaopen utility still has alsa open, but with the current proposal this will fail. - stop alsaopen. - now dvbv5-zap should work.

Regards,

Hans

On 10/25/2014 04:57 AM, Mauro Carvalho Chehab wrote:

Hi,

Sorry to enter into this thread so late. Last week was a full week, due to the 4 conferences I paticiapated, and last week I needed to fill lots of trip reports. Also, I have another trip to give two speeches.

Em Tue, 21 Oct 2014 11:32:10 -0600 Shuah Khan shuahkh@osg.samsung.com escreveu:

...

Here is what I propose for patch v3:

• make it a module under media
• collapse tuner and audio tokens into media token

I'm a little skeptical about this. Merging tuner and audio tokens seems weird on my eyes, as there are actually two different hardware resources we need to lock, and we may be locking them on different places.

I think the suggestion for collapsing came about because, in this patch series, the dvb and v4l use-cases are such that tuner and audio need to be held together. i.e when tuner is held, audio is held as well. With an intent to simplify the usage, I decided to have get_tuner interface to return audio token. It does simplify the logic for callers - all the paces that get tuner will also need to make a call to get audio and handle errors when audio is not available. That made it even more confusing perhaps and raised the question that "why can't we collapse".

Having two tokens will allow the ability hold them independently. Maybe the right approach for patch v3 is don't collapse, but change dvb and v4l to get tuner followed by audio and handle the errors paths themselves.

...

• change names (get rid of abbreviated tkn stuff)
• Make other changes Takashi/Lars pointed out in pcm
• hold token in pcm open/close

-- Shuah

At Wed, 15 Oct 2014 14:21:34 -0600, Shuah Khan wrote:

On 10/15/2014 10:48 AM, Takashi Iwai wrote:

...

At Tue, 14 Oct 2014 08:58:36 -0600, Shuah Khan wrote:

...

Add media token device resource framework to allow sharing resources such as tuner, dma, audio etc. across media drivers and non-media sound drivers that control media hardware. The Media token resource is created at the main struct device that is common to all drivers that claim various pieces of the main media device, which allows them to find the resource using the main struct device. As an example, digital, analog, and snd-usb-audio drivers can use the media token resource API using the main struct device for the interface the media device is attached to.

This patch series consists of media token resource framework and changes to use it in dvb-core, v4l2-core, au0828 driver, and snd-usb-audio driver.

With these changes dvb and v4l2 can share the tuner without disrupting each other. Used tvtime, xawtv, kaffeine, and vlc, vlc audio capture option, arecord/aplay during development to identify v4l2 vb2 and vb1 ioctls and file operations that disrupt the digital stream and would require changes to check tuner ownership prior to changing the tuner configuration. vb2 changes are made in the v4l2-core and vb1 changes are made in the au0828 driver to encourage porting drivers to vb2 to advantage of the new media token resource framework with changes in the core.

In this patch v2 series, fixed problems identified in the patch v1 series. Important ones are changing snd-usb-audio to use media tokens, holding tuner lock in VIDIOC_ENUMINPUT, and VIDIOC_QUERYSTD.

Just took a quick glance over the patches, and my first concern is why this has to be lib/*. This means it's always built-in as long as this config is enabled (and will be so on distro kernel) even if it's not used at all.

Right this module gets built when CONFIG_MEDIA_SUPPORT is enabled and stubs are in place when it is not enabled. The intent is for this feature to be enabled by default when media support is enabled. When a driver doesn't create the resource, it will simply not find it and for drivers like snd-usb-audio that aren't tried to media support, the stubs are in place and feature is essentially disabled.

I picked lib so this module can be included in non-media drivers e.g: snd-usb-audio.

Does this help explain the design? I didn't want to introduce a new config for this feature. If lib isn't right place, could you recommend another one that makes this modules available to non-media drivers? moving isn't a problem.

We can create a small module depending on CONFIG_MEDIA. But it'll be rather a question of the size. If it's reasonably small and generic enough, it's worth to put into lib/*, I think.

Takashi

Em Wed, 29 Oct 2014 11:17:44 +0200 Sakari Ailus sakari.ailus@linux.intel.com escreveu:

Hi Shuah and others,

Shuah Khan wrote:

...

Add media token device resource framework to allow sharing resources such as tuner, dma, audio etc. across media drivers and non-media sound drivers that control media hardware. The Media token resource is created at the main struct device that is common to all drivers that claim various pieces of the main media device, which allows them to find the resource using the main struct device. As an example, digital, analog, and snd-usb-audio drivers can use the media token resource API using the main struct device for the interface the media device is attached to.

This patch series consists of media token resource framework and changes to use it in dvb-core, v4l2-core, au0828 driver, and snd-usb-audio driver.

With these changes dvb and v4l2 can share the tuner without disrupting each other. Used tvtime, xawtv, kaffeine, and vlc, vlc audio capture option, arecord/aplay during development to identify v4l2 vb2 and vb1 ioctls and file operations that disrupt the digital stream and would require changes to check tuner ownership prior to changing the tuner configuration. vb2 changes are made in the v4l2-core and vb1 changes are made in the au0828 driver to encourage porting drivers to vb2 to advantage of the new media token resource framework with changes in the core.

I know this comes quite late after the first patch series has been sent, but I'd like to ask if you have you considered a different approach: rather than implementing something entirely new, the Media controller can almost do this already. It models the physical layout of the device, instead of creating special use case specific Media entity like constructs for tuner and audio. Also the Media token framework does not appear to be as a perfect match for the Media controller framework which is also planned to be used by DVB already:

URL:http://linuxtv.org/news.php?entry=2014-10-21.mchehab; look for "3) DVB API improvements". There have been ALSA MC patches as well but I'm not aware of the status of those at the moment.

The tokens appear much like media entities of specific kind to me.

Yeah, it could be seen as that.

Currently, media entities may only be entities bound to a given subsystem, but I don't think it has to (or perhaps even may) stay that way.

We had some discussions about that with Laurent in San Jose. Yeah, we will likely need to change that at the media controller, for complex embedded DVB devices.

The usage of the media controller for this specific usage is that we should not force userspace apps to be aware of the media controller just because of hardware locking.

In case of the Media controller, mutual exclusion of different users is currently performed by adding the entities to a pipeline and incrementing the streaming count once streaming is enabled --- on different interfaces streaming may mean a different thing.

Well, we'll still need to find a way for ALSA to prevent it to use the audio demod and DMA engine that will be powered off when DVB is streaming.

The Media controller interface does not handle serialising potential users that may wish to configure the device. If that's needed then we'll need to think how to add it.

Yes, this would be needed needed if we take this approach.

Reconfiguring the DMA engine and some other registers via V4L2 API should be blocked. The same applies to firmware load, if the device is using tuner input for analog TV.

If we use the media controller, we'll need to add a state to it, to indicate that a block at the pipeline is being reconfigured.

Takashi,

What's the status of Media Controller adoption on ALSA?

Regards, Mauro