ptp: vclock: propagate physical clock sampling errors

The cyclecounter read callback cannot report errors. ptp_vclock_read()
ignores the return from the physical clock and consumes its output even
if the read failed. A zero sample followed by the real counter can add
an extra 32-bit wrap to virtual time. This was found by code inspection
of reset-time PHC access and the virtual clock callers.

Sample the parent before calling timecounter_read() or timecounter_init()
and have the cyclecounter callback return the cached successful sample.
On failure, leave the counter and frequency adjustment unchanged and
return the PHC error to the caller.

Retry the first failed refresh after 10 ms instead of waiting another two
seconds. With the existing 32-bit nanosecond mask, timecounter_cyc2time()
can interpret a fresh timestamp as being in the past once the last
successful refresh is about 2.147 seconds old. An isolated read failure
should not deliberately defer the next attempt until that sample is four
seconds old.

Double the retry delay after consecutive failures, up to the normal
two-second refresh interval, and reset it after a successful refresh.
Each virtual clock has its own worker, and a failed parent read can involve
bus traffic and error logging. Backoff limits that repeated work during a
persistent failure while retaining a prompt first retry.

Limit each retry delay to the next regular refresh deadline. Otherwise,
backoff can skip the next two-second refresh opportunity and lose a full
counter wrap even when the parent has recovered before that refresh was
due. Advance the deadline when an attempt reaches it without resetting the
backoff, so sustained failures still settle at the normal refresh interval.
Keep the retry state private to each clock's auxiliary worker.

Take the initial sample and initialize the counter before publishing the
clock. Return an error pointer from registration and propagate its errno
through n_vclocks, including when the initial sample fails. Queue the first
refresh immediately after registration so registration latency is not
followed by another full refresh interval.

Keep the existing cyclecounter mask and scaling. These scheduling changes
do not guarantee timely samples: prolonged read failures or worker delays
can still exceed the half-wrap conversion window, and a gap of about
4.295 seconds between successful samples can lose a full counter wrap.
Do not assume that getcycles64() exposes a full-width nanosecond counter.

Fixes: 5d43f951b1ac ("ptp: add ptp virtual clock driver framework")
Signed-off-by: James Hilliard <james.hilliard1@gmail.com>
Link: https://patch.msgid.link/20261026-ptp-vclock-sampling-v6-3-e705612541f1@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

0 files changed, 0 insertions, 0 deletions