Enabling tstamp in proc status file externally

Enabling tstamp in proc status file externally

[Pavel Hofman]

Hi,

Please is there any way to enable the tstamp in stream status without 
modifying the client calling the alsa-lib API? I wanted to measure 
samplerate ratio between soundcards using data in their status proc 
files (comparing advancement of tstamp vs. hw_ptr). The method seems to 
work quite good, but some clients enable the stream status tstamp (e.g. 
pulseaudio) and some don't (e.g. sox, aplay), resulting in zeros in the 
status proc file.

Thanks a lot for any help or hint.

Best regards,

Pavel.

Re: Enabling tstamp in proc status file externally

[Takashi Sakamoto]

Hi Pavel,

On Thu, Jun 09, 2022 at 02:38:58PM +0200, Pavel Hofman wrote:
> Hi,
> 
> Please is there any way to enable the tstamp in stream status without
> modifying the client calling the alsa-lib API? I wanted to measure
> samplerate ratio between soundcards using data in their status proc files
> (comparing advancement of tstamp vs. hw_ptr). The method seems to work quite
> good, but some clients enable the stream status tstamp (e.g. pulseaudio) and
> some don't (e.g. sox, aplay), resulting in zeros in the status proc file.
> 
> Thanks a lot for any help or hint.

One night sleep after posting my comment to your patch for aplay[1] brings
me an idea to use tracepoints events for your purpose (it's 5:00 am at
UTC+07:00).

ALSA PCM core supports some kinds of tracepoints events[2]. They are
categorized to two parts; the history of hwptr/applptr and hardware
parameters of PCM substream. I think the former category of tracepoints
events are available for your work to invent diagnostics tools since all
of tracepoints events can be retrieved by user space application with
system time stamp. I think the type of time stamp is selectable by
options when retrieving records of tracepoints events. Furthermore the
time stamp is essentially the same as the ones of trigger/current/driver
time stamps in ALSA PCM interface.

I did not add enough description about the category of tracepoints when
committed to document [2], but roughly describe here:

- hwptr
 - the position for audio frame transmission (e.g. DMA).
- applptr
 - the position for user space application to read/write audio frame
   except for operations over mmapped buffer (but depending on audio
   hardware)

This is call graph when operating the procfs node:

(sound/core/pcm.c)
->snd_pcm_substream_proc_status_read()
  (sound/core/pcm_native.c)
  ->snd_pcm_status64()
    (sound/core/pcm_lib.c)
    ->snd_pcm_update_hw_ptr()
      (sound/core/pcm_trace.h)
      ->trace_hwptr()

You can see hwptr event is triggered as well. Actually, trace_hwptr() is
called more frequently by usual ALSA PCM applications; e.g. ioctl(2)
with PCM hwptr request.

We have some ways to retrieve the tracepoints events in user space:
 - tracefs
 - perf system call
 - bpf

You can easily find many good documentations about them; e.g.
  - BPF Performance Tools (Brendan Gregg 2020, Addison Wesley,
    ISBN-13 78-0136554820 )


I prefer your intention diagnose the runtime of PCM substream. I think
every one almost certainly love the work. So it's nice to go forward
without discouraged.


[1] Re: [PATCH] aplay: Support setting timestamp
https://lore.kernel.org/alsa-devel/Yq2Lfn+RJx96Qqvh@workstation/
[2] Documentation/sound/designs/tracepoints.rst in source of Linux kernel
https://www.kernel.org/doc/html/latest/sound/designs/tracepoints.html


Cheerss

Takashi Sakamoto

Re: Enabling tstamp in proc status file externally

[Pavel Hofman]

Hi Takashi,

Dne 18. 06. 22 v 22:52 Takashi Sakamoto napsal(a):
> Hi Pavel,
> 
> On Thu, Jun 09, 2022 at 02:38:58PM +0200, Pavel Hofman wrote:
>> Hi,
>>
>> Please is there any way to enable the tstamp in stream status without
>> modifying the client calling the alsa-lib API? I wanted to measure
>> samplerate ratio between soundcards using data in their status proc files
>> (comparing advancement of tstamp vs. hw_ptr). The method seems to work quite
>> good, but some clients enable the stream status tstamp (e.g. pulseaudio) and
>> some don't (e.g. sox, aplay), resulting in zeros in the status proc file.
>>
>> Thanks a lot for any help or hint.
> 
> One night sleep after posting my comment to your patch for aplay[1] brings
> me an idea to use tracepoints events for your purpose (it's 5:00 am at
> UTC+07:00).
> 
> ALSA PCM core supports some kinds of tracepoints events[2]. They are
> categorized to two parts; the history of hwptr/applptr and hardware
> parameters of PCM substream. I think the former category of tracepoints
> events are available for your work to invent diagnostics tools since all
> of tracepoints events can be retrieved by user space application with
> system time stamp. I think the type of time stamp is selectable by
> options when retrieving records of tracepoints events. Furthermore the
> time stamp is essentially the same as the ones of trigger/current/driver
> time stamps in ALSA PCM interface.
> 
> I did not add enough description about the category of tracepoints when
> committed to document [2], but roughly describe here:
> 
> - hwptr
>   - the position for audio frame transmission (e.g. DMA).
> - applptr
>   - the position for user space application to read/write audio frame
>     except for operations over mmapped buffer (but depending on audio
>     hardware)
> 
> This is call graph when operating the procfs node:
> 
> (sound/core/pcm.c)
> ->snd_pcm_substream_proc_status_read()
>    (sound/core/pcm_native.c)
>    ->snd_pcm_status64()
>      (sound/core/pcm_lib.c)
>      ->snd_pcm_update_hw_ptr()
>        (sound/core/pcm_trace.h)
>        ->trace_hwptr()
> 
> You can see hwptr event is triggered as well. Actually, trace_hwptr() is
> called more frequently by usual ALSA PCM applications; e.g. ioctl(2)
> with PCM hwptr request.
> 
> We have some ways to retrieve the tracepoints events in user space:
>   - tracefs
>   - perf system call
>   - bpf


Thanks a lot for your detailed explanation. Please correct me if I am 
wrong but IIUC the snd_pcm_update_hw_ptr does not update the timestamp 
if it's not enabled. I already have access to the timestamp via the 
procfs status file, but if the client does not enable the timestamp 
specifically, the struct field will not be updated. That's why I added 
the timestamp-enable code to the alsa clients aplay/axfer.

Can the tracepoints modify the status struct and enable the timestamping 
from aside?

Thanks a lot,

Pavel.

Re: Enabling tstamp in proc status file externally

[Takashi Sakamoto]

On Sun, Jun 19, 2022 at 10:03:24AM +0200, Pavel Hofman wrote:
> Hi Takashi,
> 
> Dne 18. 06. 22 v 22:52 Takashi Sakamoto napsal(a):
> > Hi Pavel,
> > 
> > On Thu, Jun 09, 2022 at 02:38:58PM +0200, Pavel Hofman wrote:
> > > Hi,
> > > 
> > > Please is there any way to enable the tstamp in stream status without
> > > modifying the client calling the alsa-lib API? I wanted to measure
> > > samplerate ratio between soundcards using data in their status proc files
> > > (comparing advancement of tstamp vs. hw_ptr). The method seems to work quite
> > > good, but some clients enable the stream status tstamp (e.g. pulseaudio) and
> > > some don't (e.g. sox, aplay), resulting in zeros in the status proc file.
> > > 
> > > Thanks a lot for any help or hint.
> > 
> > One night sleep after posting my comment to your patch for aplay[1] brings
> > me an idea to use tracepoints events for your purpose (it's 5:00 am at
> > UTC+07:00).
> > 
> > ALSA PCM core supports some kinds of tracepoints events[2]. They are
> > categorized to two parts; the history of hwptr/applptr and hardware
> > parameters of PCM substream. I think the former category of tracepoints
> > events are available for your work to invent diagnostics tools since all
> > of tracepoints events can be retrieved by user space application with
> > system time stamp. I think the type of time stamp is selectable by
> > options when retrieving records of tracepoints events. Furthermore the
> > time stamp is essentially the same as the ones of trigger/current/driver
> > time stamps in ALSA PCM interface.
> > 
> > I did not add enough description about the category of tracepoints when
> > committed to document [2], but roughly describe here:
> > 
> > - hwptr
> >   - the position for audio frame transmission (e.g. DMA).
> > - applptr
> >   - the position for user space application to read/write audio frame
> >     except for operations over mmapped buffer (but depending on audio
> >     hardware)
> > 
> > This is call graph when operating the procfs node:
> > 
> > (sound/core/pcm.c)
> > ->snd_pcm_substream_proc_status_read()
> >    (sound/core/pcm_native.c)
> >    ->snd_pcm_status64()
> >      (sound/core/pcm_lib.c)
> >      ->snd_pcm_update_hw_ptr()
> >        (sound/core/pcm_trace.h)
> >        ->trace_hwptr()
> > 
> > You can see hwptr event is triggered as well. Actually, trace_hwptr() is
> > called more frequently by usual ALSA PCM applications; e.g. ioctl(2)
> > with PCM hwptr request.
> > 
> > We have some ways to retrieve the tracepoints events in user space:
> >   - tracefs
> >   - perf system call
> >   - bpf
> 
> 
> Thanks a lot for your detailed explanation. Please correct me if I am wrong
> but IIUC the snd_pcm_update_hw_ptr does not update the timestamp if it's not
> enabled. I already have access to the timestamp via the procfs status file,
> but if the client does not enable the timestamp specifically, the struct
> field will not be updated. That's why I added the timestamp-enable code to
> the alsa clients aplay/axfer.
> 
> Can the tracepoints modify the status struct and enable the timestamping
> from aside?

The configuration of runtime of PCM bustream is immutable from
tracepoints, thus we have no way to enable it unless changing code of
PCM application.

In my understanding, your aim is to estimate samplerate ratio by
computing the history of hw_ptr with time stamp in running PCM substream.
Both of them are provided by the tracepoints events, In the case, Linux
tracepoints framework record time stamp. I think the time stamp is mostly
equivalent to the time stamp which ALSA PCM core put into the structure
and expose to user space, since both of them are sampled from system time
within atomic context. The clock_id is selectable; e.g. perf_event_open
system call receive 'struct perf_event_attr' which includes clockid field
since Linux kernel 4.1 or later[1].


[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=34f439278cef

Regards

Takashi Sakamoto