March 2018

One release blocker is now resolved: after a couple of iterations, my patches for the crash on port switches (and the associated device state change redesign) are now in master. Other things blocking the release have appeared, however. There are fixes for bluetooth and alsa latency issues from Georg Chini that I need to review, a crash in module-allow-passthrough needs to be investigated and fixed, and what I’m currently working on is converting paprefs from GConf to GSettings, which involves adding a new GSettings module in PulseAudio. Debian is getting rid of GConf, and that means that paprefs will be removed from Debian if it’s not converted. The conversion work is mostly done by Sylvain Baubeau, who submitted patches already a couple of years ago, but the patches were neglected (you can blame me). I’m now ironing out the remaining issues.

I haven’t done any OpenEmbedded work in a while, but that changed when libvorbis got a new release. OpenEmbedded Core is in a feature freeze currently, so I didn’t do a full version upgrade, I just backported some security fixes from the new release. While testing the patches on the stable branches of OpenEmbedded, I ran into an unrelated issue with e2fsprogs failing to build on my machine. The OpenEmbedded stable branches have older versions of e2fsprogs, and those were not compatible with glibc 2.27 that I had installed. e2fsprogs upstream already has a fix for the compatibility problem, so I backported the fix and submitted it to OpenEmbedded.

The rest of my my work was the usual discussion and review stuff. One thing worth mentioning about the reviews is that I accepted patches for removing BlueZ 4 support from PulseAudio (will land in PulseAudio 13.0, not the upcoming 12.0 release). BlueZ 4 has been deprecated for many years now, and I’m not aware of any distributions that are still relying on it, except on old (but still maintained) versions. Those old distribution versions are going to be using old PulseAudio anyway, so the only thing that will break is individual users who are using old distributions and are building a new PulseAudio version themselves from source.

This post was originally written on 2018-04-06, and first made available to my Patreon supporters. Speaking of Patreon – I’m using crowdfunding in an attempt to make it financially sustainable to continue my volunteer work as a PulseAudio maintainer. If you’d like to help, check out my Patreon page (or Liberapay).

February 2018

We have entered the release preparation phase again. The master branch is frozen, and the first PulseAudio 12.0 release candidate will be published once the remaining release blocker bug is fixed. The bug is about a frequent crash that happens when changing ports on alsa devices. I finished the first version of the patch set for fixing the bug, but in the review discussion there were some ideas for improving the patches, so another version will be needed, which is in progress. There’s a fair bit of labour-intensive refactoring involved, which is why writing the patches is taking some time.

I finally got around writing a second version of an old patch for improving the accuracy of alsa card profile availability information. Alsa card profiles typically have one playback device and one capture device associated with them, and previously only those profiles whose both devices were unavailable were themselves marked as unavailable, so it was possible that during startup PulseAudio would choose to activate a profile whose sink is unavailable, resulting in no audio until the user manually changes the output. Now only profiles whose both devices are available are marked as available, and hence only those profiles are candidates for becoming the initial chosen profile.

It was reported that on OpenSuSE (and apparently on Fedora too) a quick relogin after logging out can result in audio not working. The symptoms included there being two PulseAudio instances running, and the two instances conflicted with each other. It turned out that on logout the PulseAudio daemon from the old login session kept running for a while before terminating itself, but on relogin libpulse didn’t detect the old daemon, and therefore spawned a new instance. The reason for this was that logind (a component of systemd) cleared all files from the user’s “runtime directory” on logout. That directory contains PulseAudio’s socket files, and libpulse depends on those files to determine whether there’s a PulseAudio daemon running. I made a fix that makes PulseAudio exit immediately on logout, so there’s no more risk of the ghost of the old login session messing with the new login session (another solution would have been to use systemd to manage PulseAudio in the login session, but that was an uncomfortably big change for the stable releases of OpenSuSE).

As usual, I also reviewed a bunch of patches and had various discussions on the mailing list and bug tracker.

This post was originally written on 2018-03-05, and first made available to my Patreon supporters. Speaking of Patreon – I’m using crowdfunding in an attempt to make it financially sustainable to continue my volunteer work as a PulseAudio maintainer. If you’d like to help, check out my Patreon page (or Liberapay).

January 2018

This month I continued reviewing Georg’s messaging patches (and also had a long discussion thread on the topic). We had some disagreements about how the public API should look like, but those disagreements have been largely resolved now.

It was reported that PulseAudio unnecessarily made dbus-daemon log error messages related to communication with bluetoothd. That was caused by sending a reply to a method call from bluetoothd, when bluetoothd indicated that didn’t want a reply to that call. I made a fix for that.

Raspberry Pi 3 has a bluetooth adapter that apparently is by default configured so that the HSP profile doesn’t work with PulseAudio. I was able to give a hint about similar problems with other adapters, and luckily a workaround for another Broadcom adapter worked also with the adapter in Raspberry Pi 3. Ideally the kernel driver would be fixed to work out-of-the-box, but for now you’ll need to run “sudo hcitool cmd 0x3F 0x01C 0x01 0x02 0x00 0x01 0x01” to make HSP work on Raspberry Pi 3.

The new 2.27 release of glibc added a syscall wrapper (memfd_create) that conflicts with PulseAudio’s internal implementation of that wrapper. I fixed PulseAudio so that it can now be built with both old and new glibc versions.

A recent patch of mine caused a bad case of crashing: whenever ports are changed on an alsa card (e.g. when plugging in headphones), PulseAudio will crash. The patch that introduced this bug is otherwise fine, it just didn’t take into account that the port change callback isn’t always called from the main thread as I assumed, causing things to happen in the wrong thread. I started fixing the issue, but the fix isn’t ready yet. I decided to clean up the core port changing design a bit, since calling the port change callback from different threads is not nice at all.

This post was originally written on 2018-02-06, and first made available to my Patreon supporters. Speaking of Patreon – I’m using crowdfunding in an attempt to make it financially sustainable to continue my volunteer work as a PulseAudio maintainer. If you’d like to help, check out my Patreon page (or Liberapay).

December 2017

The Intel HDMI LPE driver continues to cause problems to PulseAudio despite previous attempts at making it work. To recap, the driver acts in an unusual way when the HDMI cable is not plugged in and PulseAudio tries to play audio to it. Of course, PulseAudio shouldn’t play audio to it in that situation, but there are still cases where that might happen. I wrote a patch that immediately suspends the HDMI sink when the cable is unplugged, which should finally make the worst problems go away (PulseAudio used to get killed due to too much CPU use). Even if the sink is suspended, though, streams may still be connected to it, so I made another patch for automatically moving streams away in this situation, but that was rejected in review. Georg Chini made a proposal for a better scheme, which I will implement. Details will hopefully appear in the January report.

I also made a couple of smaller fixes: A recent patch in module-augment-properties introduced a memory leak that was caught by Coverity, and I fixed that. Another recent patch in the build system had a typo that prevented module-default-device-restore from getting loaded, I fixed that too.

In order to make this report appear less short, I’ll mention some patches that I reviewed:

  • Georg Chini fixed a crash related to profile changes (the root cause for the crash remains a mystery, but the patch adds an extra safety check that should make the crash impossible).
  • Arun Raghavan simplified the build system by changing how module symbols are defined (we used to generate “symdef headers” with an m4 script, the script is now gone).
  • Roliga added a new “sink_input_properties” option to module-ladspa-sink that allows setting e.g. a custom description for the virtual stream that the module creates.
  • Samo Pogačnik added a new “use_system_clock_for_timing” option for module-pipe-sink, which makes it possible to use e.g. the “cat” command to write the sink output to a file. Previously “cat” would consume data so fast that streams connected to the pipe sink couldn’t keep up.
  • Andrius Štikonas ported qpaeq from Qt 4 to Qt 5.
  • I reviewed one of the messaging subsystem patches from Georg Chini that I’ve mentioned before.

In other news, Arun Raghavan has been working on a Meson-based build system for PulseAudio. That may or may not be merged soon. It won’t fully replace the old Autotools-based build system at least in the beginning, but it should be usable in many cases, and improve the build speed in those cases.

This post was originally written on 2018-01-04, and first made available to my Patreon supporters. Speaking of Patreon – I’m using crowdfunding in an attempt to make it financially sustainable to continue my volunteer work as a PulseAudio maintainer. If you’d like to help, check out my Patreon page.

November 2017

It was reported that if a stream has a long underrun (i.e. doesn’t provide any data to play) and then tries to catch up, PulseAudio doesn’t handle that correctly in certain situations, resulting in some amount of silence being played instead of real audio data. I found out why this happens, but I didn’t fix the bug yet.

If a user manually selects a HDMI profile on a sound card that supports both analog and HDMI outputs, module-card-restore will remember that choice. If the HDMI cable is unplugged while PulseAudio is not running, and PulseAudio is then started, module-card-restore used to select the HDMI profile even though it wasn’t plugged in (and PulseAudio knew that it wasn’t plugged in). I fixed module-card-restore so that it doesn’t any more choose unavailable profiles.

Unloading module-echo-cancel crashed sometimes, because messages that the module sent from one thread to another could be processed after the module had been unloaded, and the message handling code didn’t take that into account. I made a fix for that.

Back in February I debugged crashing related to certain assembly code in PulseAudio. The assembly code was used to figure out the processor model in order to enable processor-specific optimizations. At that time my conclusion was that GCC was probably buggy, but I never got around to filing a bug report. Thankfully “EoD” (I don’t know their real name) now filed the bug, and the GCC developers quickly identified the problem: the assembly code in PulseAudio didn’t take x86-64’s feature known as “red zone” into account, so the bug was in PulseAudio after all. The GCC developers also provided several suggestions for fixing the problem, and it was quite straightforward to make a fix (it turned out that no assembly code was required in the first place, since at least GCC and clang provide the __get_cpuid() function that does what the assembly code used to do).

It was reported that the PulseEffects application didn’t work on KDE. Something was moving the application’s stream to a wrong device. After improving the stream moving logging in PulseAudio, the culprit was found to be module-device-manager. That module is only used on KDE to manage stream routing, which is why the problem was specific to KDE. I fixed the module so that it doesn’t mess with streams that have been routed following an explicit request from the owning application.

One recurring problem with bluetooth headsets is that while PulseAudio correctly detects that the headset supports the HSP profile, no audio gets streamed when trying to use the HSP profile. This is typically a problem in the computer’s bluetooth adapter. It may be missing some required firmware or it may be configured in a way that doesn’t work with PulseAudio. It can be hard to figure out what the underlying problem is, but sometimes a fix is found, and last month a user found a fix for the Broadcom BCM4354 bluetooth adapter. I documented the fix on our bluetooth wiki page.

I updated the PulseAudio, ALSA and LAME recipes in OpenEmbedded to the latest releases.

This post was originally written on 2017-12-06, and first made available to my Patreon supporters. Speaking of Patreon – I’m using crowdfunding in an attempt to make it financially sustainable to continue my volunteer work as a PulseAudio maintainer. If you’d like to help, check out my Patreon page.

October 2017

Last month I continued reviewing Georg Chini’s patches for improving the infrastructure for adding new module specific client APIs.

The webrtc backend of module-echo-cancel had locale-dependent parsing code for the beamforming parameters, which meant that parsing failed on those locales that use a comma as the decimal mark. I made the parsing locale-independent.

I found out that the alsa-lib configuration for Intel HDMI LPE contained a definition for the “front” device, which PulseAudio assumes to be analog output. That’s not strictly speaking a valid assumption, but without that assumption it becomes ambiguous whether the sound card has one or two devices if both “front” and “hdmi” exist. I made a patch for alsa-lib that removes the “front” device. With that alsa-lib patch and the earlier pulseaudio patches (of which I made a new version, still pending review) for autodetecting the HDMI device index the Intel HDMI LPE driver reportedly is finally working properly with PulseAudio!

When connecting a new bluetooth headset, PulseAudio used to activate the HSP profile by default. I made a patch that changes the profile priorities so that A2DP has higher priority. A2DP is a better default, because it has better audio quality. HSP is needed with phone applications, but we already have automatic profile switching when a phone application appears.

I updated some outdated volume documentation for libpulse, and also added some documentation about the difference between the sound card clock and the system clock.

I silenced one Coverity warning by adding a couple of assertions in the code. The warning was a false positive, but figuring that out was quite difficult, so I also added comments why the assertions are valid.

It was reported that having module-remap-source on top of a monitor source caused high latency. That was weird, because the processing that module-remap-source does shouldn’t cause any extra latency. I learned that monitor sources cause extra latency if they monitor a sink that supports rewinding. The amount of extra latency changes depending on the latency of streams connected to the monitored sink, but module-remap-source and module-virtual-source didn’t take those changes into account, often causing more latency than necessary. The alsa sink also was overly conservative when setting its “max rewind” value, which made the problem worse. I made fixes for both issues.

This post was originally written on 2017-11-03, and first made available to my Patreon supporters. Speaking of Patreon – I’m using crowdfunding in an attempt to make it financially sustainable to continue my volunteer work as a PulseAudio maintainer. If you’d like to help, check out my Patreon page.

September 2017

Last month PulseAudio 11.0 got released, hooray! The new stuff is documented in the release notes, and I’ve probably linked to the draft version before in these reports, but here’s the link again (this time not a draft):

Quickly after the release some bad regressions were reported, boo! Some filter sinks were crashing in the new version if the flat volumes feature was enabled. Also, the ALSA mixer handling was broken for sound cards that don’t have any configuration in alsa-lib. I made a fix for both of these issues.

Also, the new bluetooth MTU autodetection feature was known to cause regressions, but it was enabled by default anyway in 11.0. That was deemed a bad idea after all, and the feature was disabled in 11.1. Arun also fixed a crash in the MTU autodetection code.

Another change in 11.1 was that I removed a priority boost for devices that report their form factor as “internal”. The 11.0 release notes advertised that bluetooth and USB sound cards are now prioritized over internal sound cards by default, but actually that was true only for bluetooth devices. The change had not been properly tested for USB, and the internal sound card priority boost still made internal sound cards have higher priority than USB sound cards (at least in cases where the USB sound card doesn’t provide any form factor information).

That’s enough about the releases. Now on to other stuff.

It was reported that on a certain sound card PulseAudio didn’t detect the speaker output. That was because ALSA didn’t indicate the existence of that port except by providing a jack detection mixer control named “Speaker Jack”, and PulseAudio didn’t have configuration for that. I made a patch for that.

When bluetooth headsets advertise their capabilities, they can choose from two different identifiers for the HSP Headset role, which was news to me. PulseAudio used to recognize only one of the identifiers, so to PulseAudio it seemed like the headset didn’t support the HSP profile at all, and naturally that prevented users from using that profile. Now both HSP HS identifiers are recognized.

The equalizer GUI “qpaeq” that is shipped with PulseAudio used to be licensed under AGPL, unlike all other PulseAudio code. Back in December I sent emails to all qpaeq contributors asking for a permission to relicense qpaeq under LGPL. It took some time to receive replies from all contributors, and then it took even longer time for me to take the next step, but now the qpaeq license has finally been changed to LGPL.

I improved the initial ALSA card profile selection. When a new sound card appears (either during PulseAudio startup or when hotplugging a card), PulseAudio will try to choose the best available profile for the card. In case of ALSA cards, the profile availability is based on what has been plugged in (assuming that ALSA provides that information). Previously the algorithm for setting the profile availability was very simplistic: only profiles whose all ports have been indicated as unplugged by ALSA were marked as unavailable. If for example a profile has a sink and a source, and only the sink is unavailable, PulseAudio used to mark the profile as available, which meant that PulseAudio could by default activate a profile whose output was not functional. Now all profiles (apart from some rare exceptions – the algorithm is still not perfect) that don’t have all of their devices available are marked as unavailable.

Victor Gaydov published a very long blog post that explains very comprehensively the PulseAudio design. I have read it all, and I can recommend you to have a look as well. The official PulseAudio documentation is severely lacking in providing an overview of all the core concepts, so this was a very welcome post.

This post was originally written on 2017-10-04, and first made available to my Patreon supporters. Speaking of Patreon – I’m using crowdfunding in an attempt to make it financially sustainable to continue my volunteer work as a PulseAudio maintainer. If you’d like to help, check out my Patreon page.