On Tue, Apr 07, 2015 at 07:30:57PM -0700, Lori Hikichi wrote:
On 15-04-06 09:19 AM, Mark Brown wrote:
OK, so that seems fragile - what happens if we're slightly late processing an interrupt or miss one entirely? Most hardware has some way to read back the current position which tends to be more reliable, is that not an option here?
The hardware updates a read pointer (rdaddr) which we feed to ALSA via the ".pointer" hook. So yes, the hardware does have a register that tells us its progress. This routine (ringbuf_inc) actually updates a write pointer (wraddr) which is a misnomer. The write pointer doesn’t actually tell us where we are writing to – ALSA keeps track of that. The wraddr only prevents the hardware from reading past it. We just use it, along with a low water mark configuration register, to keep the periodic interrupts firing. The hardware is tracking the distance between rdaddr and wraddr and comparing that to the low water mark.
The code has handling for both read and write so it's not just updating a write pointer. Is there no flexibility in the hardware regarding interrupt generation?
Being late processing the interrupt is okay since there are more samples to read. That is, it was only a low water mark interrupt and we have one period of valid data still (we configure low water to be one period). Missing an interrupt is okay since the hardware will just stop reading from buffer when rdaddr has hit wraddr.
Stopping if we miss an interrupt is precisely the sort of situation we want to avoid if we can - if the application is sufficiently far ahead of the hardware everything should continue to work fine. The minimal period size appears to be very small so this is a potential issue, if an application tries to use many very small periods it's both more vulnerable to some other interrupt taking longer than might be desirable and likely that things would be fine as the application is hopefully more than one period ahead of where the hardware is at.
If the hardware is always going to halt at the end of the period there's not a huge amount we can do about this except possibly raise the minimum period if systems are running into trouble but if there's a way to avoid doing that then that would be even better.