[alsa-devel] [Alsa-user] Ticks when playing to USB DAC at high sample rates

[Daniel Mack]

Hi Jeff,

On 05.11.2012 02:53, Jeffrey Barish wrote:
> On Sun 04 November 2012 11:39:48 Daniel Mack wrote:
>> [cc alsa-devel]
>>
>> On 04.11.2012 02:25, Jeffrey Barish wrote:
>>> On Fri 26 October 2012 15:21:06 Jeffrey Barish wrote:
>>>> On Fri 26 October 2012 21:47:05 Daniel Mack wrote:
>>>>> On 26.10.2012 21:43, Jeffrey Barish wrote:
>>>>>> On Thu 25 October 2012 19:10:45 Daniel Mack wrote:
>>>>>>> On 25.10.2012 17:18, Jeffrey Barish wrote:
>>>>>>>> I found something in the snd_usb_audio code (in endpoint.c) that
>>>>>>>> could
>>>>>>>> explain one of the problems I have observed (the ticks). I would
>>>>>>>> normally test my theory by modifying the code. In this case, I would
>>>>>>>> like to stick in a print statement to see what values are being
>>>>>>>> assigned
>>>>>>>> to certain variables. Unfortunately, I am too ignorant to do
>>>>>>>> something
>>>>>>>> even this trivial as I have never worked on kernel code. I think I am
>>>>>>>> supposed to use printk,
>>>>>>>
>>>>>>> printk is nice for simple debugging, yes. But note that this call is
>>>>>>> timing critical and should not be used in "fast path" code.
>>>>>>> Introducing
>>>>>>> a printk for each received packet for example will almost certainly
>>>>>>> make
>>>>>>> the driver behave quite differently.
>>>>>>>
>>>>>>>> but beyond that I am lost. Can someone provide
>>>>>>>> me with some directions? I need to know how to make the driver. To
>>>>>>>> that
>>>>>>>> end, I probably will have to install additional packages. After
>>>>>>>> making
>>>>>>>> the driver, I need to know how to install it over the existing
>>>>>>>> driver.
>>>>>>>
>>>>>>> Here's one way to do it:
>>>>>>>
>>>>>>> 1. git clone
>>>>>>> git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound.git (your
>>>>>>> patch should apply on top of this tree eventually)
>>>>>>> 2. cd sound
>>>>>>> 3. zcat /proc/config.gz >.config
>>>>>>> 4. build and install the kernel image. How that is done depends on the
>>>>>>> distribution you're using. For Ubuntu follow the docs at [1] (start at
>>>>>>> point #5). For Fedora and others, something like "make && make
>>>>>>> install"
>>>>>>> should do
>>>>>>> 5. reboot and check that the new kernel is running
>>>>>>> 6. hack on sound/usb
>>>>>>> 7. make M=sound/usb
>>>>>>> 8. reload the module with "sudo rmmod snd_usb_audio; sudo insmod
>>>>>>> sound/usb/snd-usb-audio.ko" (better plug out the device before so you
>>>>>>> always have the same defined point of start)
>>>>>>>
>>>>>>>
>>>>>>> Hope that works for you.
>>>>>>>
>>>>>>>
>>>>>>> Daniel
>>>>>>>
>>>>>>> [1] https://wiki.ubuntu.com/KernelTeam/GitKernelBuild
>>>>>>
>>>>>> Your directions were almost perfect, so even I was able to build the
>>>>>> kernel. I made a discovery using the new kernel that might help someone
>>>>>> more familiar with the code than I am to localize the problem.  I am
>>>>>> still hearing the blip when I play audio sampled at 88.2 kHz, but I
>>>>>> just
>>>>>> noticed that the blip is perfectly periodic, with a period of about
>>>>>> 16.4
>>>>>> seconds.  I am playing a sine wave synthesized using GStreamer using
>>>>>> the
>>>>>> following command:
>>>>>>
>>>>>> gst-launch audiotestsrc volume=0.01 ! audio/x-raw-float, width=64,
>>>>>> rate=88200, channels=2, endianness=1234 ! audioconvert ! alsasink
>>>>>>
>>>>>> A sine wave makes it easier to hear the blip.  Does this clue suggest
>>>>>> anything?
>>>>>>
>>>>>> I also want to mention that when I use the new kernel, I do not get the
>>>>>> ticks at either 88.2 or 96 kHz even when I do not use the external USB
>>>>>> hub.  I plan next to back up to the 3.6.2 kernel to see whether I still
>>>>>> get ticks there.
>>>>>
>>>>> Which kernel did you use when you heard the 'blibs'?
>>>>
>>>> The latest news is bad.  I am on 3.2.0 now.  The USB DAC is working
>>>> perfectly at this moment at both 96 kHz and 88.2 kHz without the external
>>>> USB hub (imagine calling that bad news).  If I set the srate to 88.2 kHz
>>>> and stop and start the sine wave, sometimes I get the blip.  Forget about
>>>> its being periodic.  It was definitely periodic before lunch; now I
>>>> usually
>>>> get random intervals if I get any blips at all.  As I am typing this
>>>> message, I can't get blips at all.  There was some correlation between
>>>> changing sample rates and blips, but I can't reproduce that behavior now.
>>>> What is most weird is that I haven't gotten any ticks since lunch with
>>>> any
>>>> kernel or with either sample rate, yet they were reliable earlier today
>>>> unless I used the external USB hub. I obviously need to experiment some
>>>> more to see whether I can observe a pattern.
>>>
>>> To conclude (?) this thread, I am now convinced that the anomalies I
>>> observed were unrelated to the device driver.  I had two theories
>>> remaining.  One was that the problem was somehow related to an
>>> overheating problem.  I used a heat gun to convince myself that the
>>> theory was wrong.  The other was that the problem had something to do
>>> with services running in the background that interfered with the device
>>> driver.  I removed or disabled all services that I could identify as
>>> superfluous.  Removed services include zeitgeist, apparmor, modemmanager,
>>> mdadm, and bluetooth.  Disabled services include atd, dns-clean, and
>>> pppd-dns.  The system has been running perfectly with an 88.2 kHz sample
>>> rate and no external USB hub for 2 days.  However, I tried booting the
>>> system with a copy of the OS from before I removed superfluous services. 
>>> It still ran perfectly.  Thus, I cannot convince myself that removing
>>> superfluous services actually solved the problem.  I suppose it's
>>> possible that some service was running sporadically, which could explain
>>> why the problem seemed to come and go and also why the problem did not
>>> occur when I booted the copy of the OS. As long as the problem remains
>>> dormant I have no alternative but to move on and see whether the problem
>>> eventually recurs.  Only time will tell.
>>>
>>> Thanks for all the help.  Hey, I built a kernel and got my own version of
>>> the device driver to run.  That was exciting.
>>>
>>> --------------------------------------------------------------------------
>>> ----
>>>
>>> I'm back.  The USB DAC continued to work for two more days after my last
>>> message and then the blips resumed.  I've been digging into the driver
>>> code
>>> for two days.  I spent today trying to convince myself that the driver is
>>> actually using an asynchronous protocol.  The blips are almost certainly
>>> buffer underruns, which I could account for if the driver were not
>>> responding to asynchronous feedback from the DAC.  I expected to see a
>>> deviation from the pattern of one frame with 89 samples followed by 4
>>> frames with 88 (which results in 88,200 samples per second).  That
>>> sequence would work perfectly if the clocks in the computer and in the
>>> USB DAC were exactly the same frequency, but of course they are not. 
>>> Consequently, I would expect to see requests for one sample more or less
>>> once in a while, depending on whether the clock in the computer is slower
>>> or faster than the one in the DAC.  What I see is 89 + 4 * 88 almost all
>>> the time, as expected .  Once in a while there is a request for an
>>> additional frame of 88 and then the pattern of 89 + 4 * 88 resumes.  I
>>> presume that the break in the usual pattern indicates that the device
>>> driver is responding to a request from the DAC for fewer samples, which
>>> confirms that the driver is using the asynchronous protocol.  That's
>>> good, if I am right. However, sticking in an extra packet of only 88
>>> would be the expected deviation if the host were getting a little ahead
>>> of the DAC, yet the blip is almost certainly a buffer underrun, which
>>> means that the host is actually falling behind.  Thus, an extra packet of
>>> 89 would make more sense.  I should also mention that there is no
>>> correlation between the timing of the extra packet of 88 and the audible
>>> blip.  Something seems wrong here, but I'm not sure where to go next.
>>>
>>> On a related topic, to confirm that the blip is a buffer underrun, I would
>>> like to synthesize a waveform that is DC.  With DC, an overrun would not
>>> produce any audible anomalies whereas an underrun would be audible.  I am
>>> using GStreamer to synthesize the test waveform.  Does anyone know of a
>>> way to synthesize a DC waveform using GStreamer?  If not, is there any
>>> other tool available in Linux?
>>>
>>> By the way, I bought another USB DAC from a different manufacturer to be
>>> sure that the anomalies I am observing were not due to a defect in the
>>> DAC.  They both display the same problem.
>>>
>>> -----------------------------------------------------
>>>
>>> I modified the device driver to output a constant instead of the incoming
>>> waveform.  I get blips, as I expected, so I am now certain that the blips
>>> are caused by a buffer underrun.  The blips occur periodically again --
>>> period of about 16.5 seconds.  I presume that the period is related to
>>> the difference in clock rates between the host and DAC.  Roughly every
>>> 3-4 blips, I get an extra packet of 88 samples.  To prevent buffer
>>> underrun, I ought to be getting an extra packet of 89 samples every 16.5
>>> seconds, but I never do.
>>
>> Thanks for all your investigation on this! Sending computed data instead
>> of the payload delivered by the ALSA stack is in fact a very good way to
>> isolate the issue.
>>
>> The question is: if the algorithm in the USB driver miscalculates the
>> number of samples to send over time, what't the reason for that? Does
>> the device report bogus values or is there this a bug in the driver?
> 
> Well, I have two DACs, each from a different manufacturer.  Both experience 
> underruns (though not at the same rate).  It's possible that they are both 
> reporting bogus values, but unlikely, so I'm leaning toward the bug theory.
> 
>> Did you read chapter 5.4.12.1 of the USB2.0 spec, and can you explain
>> where you suspect the bug to be? The current implementation in the
>> driver is quite straight forward, mapping all feedback values to Q16.16
>> and keeping the lower 16bits around for phase compensation between the
>> calculation of the packet sizes.
> 
> Do you perhaps mean chapter 5.12.4.1 (Synchronization Type)?  I'm not finding a 
> 5.4.12.1 in usb_20.pdf from a zip dated 102512.

Yes, sorry.

> Please tell me whether I have this right:
> 
> You start transferring the sound in snd_usb_endpoint_start using 
> usb_submit_urb to submit 2 urbs to the USB core.  The urbs specify that on 
> completion, the USB core should call snd_complete_urb.  snd_complete_urb 
> resubmits the urb after calling retire_playback_urb (by way of 
> retire_outbound_urb which calls retire_data_urb which is retire_playback_urb), 
> which updates its estimate of delay (what's that?  It seems to be 0 always) 

That's just an information that is passed to userspace for latency
compensation matters. As you don't stream data that ALSA passes up the
stack anymore in your tests (but static data), you can ignore this
detail for now.

> and prepare_playback_urb (by way of prepare_outbound_urb which calls 
> prepare_data_urb which is prepare_playback_urb), which fills the transfer_buffer 
> in the urb with the audio data (from a DMA buffer, I guess).
> 
> Before filling transfer_buffer, prepare_playback_urb calls 
> snd_usb_endpoint_next_packet_size to compute the number of samples to send.  
> This code appears to be the crux of what I need to understand because the 
> value returned from this function is what is supposed to adjust to prevent 
> buffer underruns.

That's correct.

> It isn't clear to me how this function gets information 
> about the state of the buffer in the device.

Via the callback of the feedback data endpoint - the momentary frequency
is updated there and referred to again in
snd_usb_endpoint_next_packet_size().

> The only argument to 
> snd_usb_endpoint_next_packet_size is ep, which is a pointer to the 
> snd_usb_endpoint structure.

Yes, the endpoint structure hold all runtime streaming details. Once a
sync urb is given back from the host controller, retire_inbound_urb() is
called and snd_usb_handle_sync_urb() cares for updating the momentary
frequency of the slave of the sync endpoint, which is the data endpoint
that is later on used by next_packet_size(). Makes sense?

> I see that ep->phase is the Q16.16 value you 
> mentioned, so the lower 16 bits are for "phase compensation between the 
> calculation of the packet sizes".  I don't completely understand that phrase.  
> freqm is an unchanging value (5780275)

freqm shouldn't be unchanging. Are you sure it really is? If so, is
snd_usb_handle_sync_urb() ever called in your setup?

> and datainterval is an unchanging value 
> (0).  phase fluctuates in a narrow range (roughly 5792282 to 5832702), but it 
> is fluctuating only as a consequence of this calculation: we take the bottom 16 
> bits of its value and add a fixed value.  If I understand this calculation 
> correctly, I can almost see how it could provide the sequence of packet sizes 
> (the 89 + 4 * 88 sequence) necessary to produce the correct number of samples 
> per second, but I don't see how it adjusts that number to the slightly 
> different concept of time at the device.  However, even in that understanding I 
> am perplexed because I occasionally see an extra packet of 88.

The device reports its buffer state by providing feedback information in
a perioc manner, as letting the host know about what the device think is
its correct data rate is the only way to compensate clock drifts.

> Also, I think I understand why you start the transfer with 2 urbs:  When the 
> device finishes with the first urb, the next urb has to be there already to 
> avoid a period when no urb is present.  However, right away an alarm goes off.  
> When the device finishes processing the first urb, it provides feedback about 
> the state of the buffer.  However, the host cannot respond to this feedback 
> until after the second urb is processed. Have we convinced ourselves that the
> lag doesn't matter?  I don't know because I don't understand what happens to 
> the urb after we send it to the USB core. What do you know about the nature
> of the buffering in the device?  Is there a place for the USB hardware to store 
> the contents of transfer_buffer independently of a buffer that feeds the actual 
> DAC?  When does the USB hardware transfer the second urb to the device? 

Isochronous transfers are scheduled on the bus and sent in the
configured time slot (or as soon as possible, if the ASAP flag is set).
Data payload should be distributed as evenly as possible over the
microframes to not blast the device's internal buffers due to bursts.

> Who 
> tells the DAC to start playing the contents of the buffer?

Ususally, once it has its buffers filled with data.

>  I would like to 
> hear that the device has a FIFO whose size is sufficient to hold the contents of 
> at least two transfer_buffers.  The first two urbs would fill it before 
> activating the DAC.  Every time the DAC finishes outputting the contents of one 
> urb, it provides feedback with the amount of space remaining in the FIFO.

Not quite. It does report the momentary frequency in periodic packets
(like every 16ms), not after each packet.

>  We 
> would send that many samples in the next urb.  I am fantasizing this solution.  
> Tell me how it really works.

It's described in 5.12.4.1 (Synchronization Type). The driver aims to be
as close the the actual data rate that is reported by the device every
couple of milliseconds.

Does that answer your questions?


Daniel