On 11. 04. 23 15:57, Oswald Buddenhagen wrote:
On Tue, Apr 11, 2023 at 01:09:59PM +0200, Jaroslav Kysela wrote:
On 08. 04. 23 9:24, Oswald Buddenhagen wrote:
Also, silencing touches the DMA buffer which may not be desired.
hypothetically, yes. but practically? [...]
The buffers can be mmaped so used directly for application and
hardware.
yes, and they are owned by the hardware/driver. an application would
know better than doing with them anything they were not designed for.
And lastly drivers can handle draining correctly (stop at the exact
position - see substream->ops->trigger with SNDRV_PCM_TRIGGER_DRAIN
argument).
yeah. hypothetically. afaict, there is exactly one driver which supports
this. most (older) hardware wouldn't even have the capability to do such
precise timing without external help.
Most hardware has FIFO and most drivers uses the DMA position, so I think that
the interrupt -> stop DMA latency may be covered with this FIFO in most cases.
on most hardware it would be quite a stunt to re-program the buffer
pointers on the fly to enable a mid-period interrupt. and given the
reliability problems insisted on by takashi in the other thread, the
approach seems questionable at best. and that's still ignoring the
effort of migrating tens (hundreds?) of drivers.
I said to not change things at the moment. Drivers may be revised at some
point. If we have large buffers, the silencing may consume many CPU ticks. If
the driver already behaves nicely for the drain operation, this is an extra
step which should be avoided. This can be handled using new snd_pcm_ops, of
course.
You're using a hammer to fix a little issue.
Again, I would improve the documentation.
no amount of documentation will fix a bad api. it's just not how humans
work.
Which code does not fill the last period? Alsa utilities do it. We can
eventually handle this in alsa-lib.
the silencing is controlled using sw_params, so applications may
request the silencing before drain.
yeah, they could, but they don't, and most won't ever.
you're arguing for not doing a very practical and simple change that
will fix a lot of user code at once, for the sake of preventing an
entirely hypothetical and implausible problem. that is not a good
trade-off.
I'm arguing that we should not do anything extra with the buffers until the
application requests that. That's the clear API context.
Lastly, I think that you cannot call updated snd_pcm_playback_silence()
function with runtime->silence_size == 0.
if (runtime->silence_size < runtime->boundary) {
you missed the hunk that adjusts the code accordingly.
Ohh, yes. You're right.
If we allow modification of the PCM buffer, I think that we should:
- Do not modify the buffer for drivers already working with the
appl_ptr data (end position) only.
- Handle the situation with the large buffer; it may make sense
to change the "wait" operation from the end-of-period interrupt to time
scheduler and stop the drain more early when the end-of-valid data condition
is fulfilled.
- Increase the protocol version.
But as I wrote, I would make those extensions configurable
(SNDRV_PCM_HW_PARAMS_DRAIN_ALLOW_SILENCE). It can be turned on by default.
Jaroslav
--
Jaroslav Kysela <perex@xxxxxxxx>
Linux Sound Maintainer; ALSA Project; Red Hat, Inc.
