I found a bug in libasound. When both the client and the slave (in my case, the JACK plugin) try to use the float format, but the sample rate or channel count does not match, libasound *should* insert linear conversion plugins to convert from float to linear, then resample/remap channels, and then convert back to float (because apparently the resamplers and channel remapper don't support floating point, only 'linear' i.e. integers). Currently this doesn't work, snd_pcm_plug_change_format doesn't know what to do and simply returns EINVAL. As a result, snd_pcm_hw_params fails even though the HW params were perfectly valid (it indicates that both the float format and any sample rate are supported).
In my test, this broke audio for WINE (and any other application that tries to use float, such as aplay with the right settings) when I wanted to use the JACK plugin (which only supports the float format).
This patch fixes this bug by doing a conversion to S16 and back when resampling or remapping is needed. And while I was at it, I also removed a check that had no effect because the exact same check is already being done at the start of the function.
I still think it's a bit silly that libasound requires integers for resampling, because both libsamplerate and libspeex use float internally for resampling. So now ALSA is literally doing a float-to-s16-to-float-to-s16-to-float conversion. But changing that would have been a lot harder.
Maarten Baert