21.04.2014 07:49, David Henningsson wrote: > On 2014-04-20 21:26, Alexander E. Patrakov wrote: >> Thus, it is not possible to tell the hardware device (that can use >> rewinds) from a properly wrapped software encoder (that can't rewind and >> doesn't pretend to be able to rewind), because for both cases >> snd_pcm_rewindable() would return 0 at the moment PulseAudio needs to >> make a decision. > > The moment PulseAudio needs to make a decision is when a rewind is > requested. No. The decision definitely needs to be at the device-open time. Otherwise this will happen: """ Now I need to rewind in order to accommodate a new low-latency client. Oops, I can't, and I have so much wrong data in my hardware buffer! I should not have created such a big buffer, but now it too late to change anything. """ And indeed, the current code already has logic to choose different buffer sizes for tstamp and irq-driven modes: http://cgit.freedesktop.org/pulseaudio/pulseaudio/tree/src/modules/alsa/alsa-util.c#n298 On my hardware, the buffer sizes for these two modes differ by a factor of 1000. So what I want is really not related to tsched. "Don't choose a big buffer size and high latency, and don't try to rewind, if we know in advance that ALSA cannot rewind or only pretends to be able to rewind" would be a better description of my patch. > Whether or not to enable tsched should not matter in this case, unless > I'm missing something. (And this is probably what Raymond is trying to > say too.) > Or, put in another way, why would it be better for the ALSA device to be > in interrupt driven mode just because it can't rewind? I have two slightly-conflicting answers to this. First answer: Rewinds and timestamp-driven scheduling are only the means to get dynamically reconfigurable latency, which is useful for less dropouts when there are no low-latency clients, lower power usage, and possibly other good things. Due to the inability to do rewinds, the "dynamic client-driven latency" goal becomes unachievable, so there is simply no good point to use timestamp-based scheduling in this case. Of course timestamp-based scheduling will work without rewinds, but, as PulseAudio would then need (due to inability to do rewinds) to lock into the constant minimum latency, the wakeup points will be evenly spaced in time. And that's almost equivalent to the IRQ-based scheduling (with a small exception listed in the second answer). Or to put it another way. Currently, PulseAudio supports two models: "big buffer + timestamp-based scheduling + rewinds" and "small buffer + IRQ-driven scheduling + no rewinds". Intermediate models such as "small buffer + timestamp-based scheduling + no rewinds" are possible, but they would IMHO only unnecessarily inflate the test matrix. Second answer: Well, it is not better. In timestamp-based scheduling mode, we can dynamically adjust latency. The limitation is that, without rewinds, our decisions to reduce latency (e.g. due to a new client) would apply too late. But even with this limitation, it means that we can try to keep as low latency as it actually works on the given hardware (similar to the current watermark logic), disregarding any client-specified latency. The problem is that, if one wants to use timestamp-based scheduling without rewinds, one needs to decouple the current watermark logic, the buffer size choice logic, and the "don't use latency lower than requested by any client" logic, because the later only makes sense when rewinds are possible. -- Alexander E. Patrakov
