Re: [chrony-dev] Support for Multiple PPS Inputs on single PHC

[Matt Corallo <ntp-lists@mattcorallo.com>]

On 2/16/23 12:19 AM, Miroslav Lichvar wrote:
> On Wed, Feb 15, 2023 at 10:27:15PM -0800, Matt Corallo wrote:
>> My naive solution from skimming the code would be to shove
>> formerly-discarded samples into a global limited queue and check for
>> available timestamps in `phc_poll`. However, I have no idea if the time
>> difference between when the sample was taken by the hardware and when the
>> `HCL_CookTime` call is done would impact accuracy (or maybe the opposite,
>> since we'd then be cooking time with the hardware clock right after taking
>> the HCL sample rather than when the PHC timestamp happens), or if such a
>> patch would simply be rejected as a dirty, dirty hack rather than unifying
>> the PHC read sockets across the devices into one socket (via some global
>> tracking the device -> socket mapping?) and passing the samples out
>> appropriately. Let me know what makes the most sense here.
> 
> My first thought is that this should be addressed in the kernel, so
> even different processes having open the PHC device can receive all
> extts samples. If it turns out it's too difficult to do for the
> character device (I'm not very familiar with that subsystem), maybe it
> could be done at least in sysfs (/sys/class/ptp/ptp*/fifo or a new
> file showing the last event like the PPS assert and clear).

I mean my first thought seeing an ioctl on a socket that gives an explicit channel and then receives 
crap from other channels on the same socket was "wtf" so I went and read the kernel to figure out 
why first to see if its a driver bug. I can't seem to find *any* documentation for how these ioctls 
are supposed to work, but it seems the "request" here is kinda misnomer, its really a "configure 
hardware" request, and is unrelated to future reads on the socket, or really the specific socket at all.

As for duplicating the output across sockets, ptp_chardev.c's `ptp_read` is pretty trivial - just 
pop the next sample off the queue and return it. Tweaking that to copy the sample into every reader 
is probably above my paygrade (and has a whole host of leak risk I'd probably screw up). 
`extts_fifo_show` appears to be functionally identical.

I've CC'd the MAINTAINERs for ptp to see what they think about this, though it won't let chrony 
support this without a kernel upgrade - not sure if that's an issue for chrony or not.

Matt