Re: [PATCH net v2 2/2] ptp: vclock: preserve time across failed physical clock samples

[Paolo Abeni <pabeni@redhat.com>]

On 10/5/26 20:45, James Hilliard wrote:
> The cyclecounter read callback cannot report errors. Passing an
> unsuccessful PHC read through it consumes an invalid sample, and a zero
> sample followed by the real counter can add an extra 32-bit wrap to
> virtual time. This was identified by code inspection of reset-time PHC
> access and the virtual clock callers.
> 
> Read the parent clock before updating or initializing the virtual
> counter, propagate failures from clock operations, and leave its state
> unchanged on failure. Initialize it before publishing a new virtual
> clock.
> 
> Simply skipping failed reads is insufficient: the next successful read
> can arrive after the 32-bit counter has wrapped, while timestamp
> conversion has an even shorter half-wrap window. Use the full physical
> nanosecond sample with a wide frequency-adjustment multiply and retain
> fractional nanoseconds. This also allows delayed packet timestamps to be
> converted without truncating their distance from the last sample.
> Serialize extended and cross-timestamp sampling with adjustments and
> advance the anchor on each successful clock read.
> 
> The initial sample can also fail partway through a sysfs request to create
> multiple virtual clocks. Unregister any clocks created by that request
> and clear their index entries, preserving the previously installed clocks
> and count. Otherwise the failed request leaves registered children that
> are not included in n_vclocks. This cleanup also handles existing
> allocation and registration failure paths.
> 
> Fixes: 5d43f951b1ac ("ptp: add ptp virtual clock driver framework")
> Fixes: 73f37068d540 ("ptp: support ptp physical/virtual clocks conversion")
> Signed-off-by: James Hilliard <james.hilliard1@gmail.com>

This addresses multiple issues, should instead be split accordingly.

Also sashiko remarks hints at possible regression wrt the new wrap schema.

/P