On 21.09.2018 16:26, Jussi Laako wrote:
On 21.09.2018 16:06, Takashi Iwai wrote:
On Fri, 21 Sep 2018 14:53:20 +0200, Jussi Laako wrote:
On 21.09.2018 11:43, Takashi Iwai wrote:
On Fri, 21 Sep 2018 10:35:49 +0200, Jussi Laako wrote:
The caching is currently enabled for all elements, but changing it should be trivial. The patch below adds is_volatile flag to the element, and you can set it to true in the quirk somehow for uncached controls.
This patch looks good and what I need! When trying to figure out where it is safe to set the volatile flag I thought that I could set it transparently in add_single_ctl_with_resume() in mixer_quirks.c based on the SNDRV_CTL_ELEM_ACCESS_VOLATILE flag in provided snd_kcontrol_new access-member. However, add_single_ctl_with_resume() is allocating just size of usb_mixer_elem_list unlike snd_create_std_mono_ctl_offset() which in turn is allocating full usb_mixer_elem_info size. However, the mixer code seems to be assuming that the item is always usb_mixer_elem_info instead of just list header item. Is this allocation behavior correct and is the item turned into usb_mixer_elem_info somewhere, or is this some kind of bug? So can I safely turn the allocation in add_single_ctl_with_resume() into zero initialized usb_mixer_elem_info instead and set the flag there while keeping correct behavior, or am I missing something?
We can check SNDRV_CTL_ELEM_ACCESS_VOLATILE instead of introducing a new flag like my patch, yes. It's just a matter of easiness. Since there is no link from usb_mixer_elem_info to the assigned snd_kcontorl, we'd need to add an extra argument in the call path to inform about the volatileness, and it may become messy. We'll find the simplest way.
And, IMO, it'd be better to check the cache behavior at setting, since the cache may be accessed via multiple paths (at reading each element and at resuming the whole system). But it's no requirement, either.
I have tried with attached patch, on top of your patch, but I'm still getting cached values from libasound2. It didn't break anything either... However, I don't see a link between this cache behavior and specific user space mixer descriptor. Because polling the mixer element with snd_mixer_selem_get_capture_volume() keeps always returning the same value. While in parallel checking the value by running "amixer" from command line gives updated correct value... So is there a second layer of cache somewhere in libasound2 to explain this discrepancy? Because one process sees updated value while other one doesn't. Difference being that the ALSA descriptors are closed and reopened every time "amixer" is run...
Erm, I should have looked your *actual* patch before replying. Basically all your newly created elements are independent from the other USB standard things, and they are all read-only. That is, you don't need to use add_single_ctl_with_resume(), but just create each element via snd_ctl_new1() and add it via snd_ctl_add(). (Also you can pass the mixer object directly in private_value.)
If it still doesn't work as expected, it's rather something wrong in the alsa-lib side, or ACCESS_VOLATILE flag isn't passed / set properly.
By the way is the add_single_ctl_with_resume() correct in allocating just snd_mixer_elem_list instead of snd_mixer_elem_info? I got suspicious about this when reading the sources. Because the code is similar to snd_create_std_mono_ctl_offset() but they allocate different size which they go on to pass to the same functions. And in quite many places functions cast the snd_mixer_elem_list pointer to snd_mixer_elem_info pointer to access the entire structure which would of course be wrong if it's just snd_mixer_elem_list size item. The patch I posted in my previous mail that changes add_single_ctl_with_resume() to allocate snd_mixer_elem_info instead at least didn't seem to break things...
OK, I already started going through the alsa-lib code, I suspect something is going wrong there because different processes give different values depending on process lifetime and at the driver level there's no process descriptor context yet...
OK, problem solved! My original patch for the RME mixer quirks is good, no changes needed to that one.
Problematic caching happens inside alsa-lib in simple mixer interface, where snd_mixer_selem_get_capture_volume() always returns cached value for mixer value read and doesn't trigger update from kernel. This is now fixed in my user space by using snd_hctl_elem_read() instead which triggers update from kernel, and as a side effect also seems to trigger update to the cached value so that later call to snd_mixer_selem_get_capture_volume() returns updated value too. I'm not sure if something should be done about snd_mixer_selem_get_capture_volume() behavior or not, but now at least I know how to work around this by using different API layer.
In addition, I've applied attached patch on top of your volatile patch, but I think that part is not strictly necessary but maybe good thing to do. This attached patch fixes (?) the allocation behavior of add_single_ctl_with_resume() I was talking about and adds volatile check on snd_usb_get_cur_mix_value() entry which should be superfluous to your change later in the function.
Thanks gain!
- Jussi