On 2026/6/30 18:57, David Woodhouse wrote:

On Tue, 2026-06-30 at 11:15 +0800, Wen Gu wrote:

I was thinking we'd move them to drivers/phc, and simplify them as we do.

Most of them are just a lot of PTP driver boilerplate, wrapping around
one central function like

static int vmclock_get_crosststamp(struct vmclock_state *st,
                                    struct ptp_system_timestamp *sts,
                                    struct system_counterval_t *system_counter,
                                    struct timespec64 *tspec)

...which is called with different permutations of arguments depending
on the actual PTP call.

I was thinking of reducing the duplication and having the PHC drivers
provide *only* that central function. Let the common PHC code provide
the interface to PTP (as well as to core timekeeping, for setting the
clock at boot, for timekeeping_set_reference() in the vmclock case, and
perhaps even for a PPS-like discipline from other clocks).

Here's a *very* hastily thrown together proof of concept; utterly
untested and AI-produced, and I've only given it the bare minimum of
oversight thus far (I have been meaning to do this for weeks but other
things have taken precedence so far)...

https://git.infradead.org/?p=linux-phc.git;a=shortlog;h=refs/heads/next

Thanks for the POC. I've looked over it and agree this is a cleaner
approach overall. The v3 was meant as an interim step, but given the POC
it makes more sense to go straight to drivers/phc.

I'm happy to adapt the Alibaba CIPU PHC driver to the new framework
once it's in place, and glad to help along the way if needed.