On 30/07/2025 2:33, Jacob Keller wrote:

On 7/18/2025 4:29 PM, Jakub Kicinski wrote:

quoted

On Tue, 15 Jul 2025 08:15:30 +0300 Tariq Toukan wrote:

quoted

This patch series introduces support for exposing the raw free-running
cycle counter of PTP hardware clocks. Some telemetry and low-level
logging use cycle counter timestamps rather than nanoseconds.
Currently, there is no generic interface to correlate these raw values
with system time.

To address this, the series introduces two new ioctl commands that
allow userspace to query the device's raw cycle counter together with
host time:

  - PTP_SYS_OFFSET_PRECISE_CYCLES

  - PTP_SYS_OFFSET_EXTENDED_CYCLES

These commands work like their existing counterparts but return the
device timestamp in cycle units instead of real-time nanoseconds.

This can also be useful in the XDP fast path: if a driver inserts the
raw cycle value into metadata instead of a real-time timestamp, it can
avoid the overhead of converting cycles to time in the kernel. Then
userspace can resolve the cycle-to-time mapping using this ioctl when
needed.

Adds the new PTP ioctls and integrates support in ptp_ioctl():
- ptp: Add ioctl commands to expose raw cycle counter values

Support for exposing raw cycles in mlx5:
- net/mlx5: Extract MTCTR register read logic into helper function
- net/mlx5: Support getcyclesx and getcrosscycles

It'd be great to an Ack from Thomas or Richard on this (or failing that
at least other vendors?) Seems like we have a number of parallel
efforts to extend the PTP uAPI, I'm not sure how they all square
against each other, TBH.

Full thread for folks I CCed in:
https://lore.kernel.org/all/1752556533-39218-1-git-send-email-tariqt@nvidia.com/ (local)

I agree with Jakub about the need to properly explain the use cases and
goals in the commit and cover letter. AFAIK there are no current public
APIs for reporting cycles to userspace, so this really only makes sense
with something like DPDK. Even the XDP related helpers expect nanosecond
units now. Its unclear if we will need other parts of the APIs to also
handle cycles, or if simple ability to get the current cycles is sufficient.

The API also doesn't directly provide a way to query the expected or
nominal relationship between cycles and clock time.

If you try to just use PTP_SYS_OFFSET_EXTENDED_CYCLES to compare a
cycles value to a clock value to adjust a timestamp, that requires that
some other process is keeping CLOCK_REALTIME and the PHC clock
synchronized. When handled within the driver, the software typically has
an assumption about the relationship based on expected frequencies.
Thus, a conversion from cycles to time uses this relationship.

You don't appear to expose that relationship through the API, which
means you can only infer it either by knowing the device, or by assuming
CLOCK_REALTIME is already synchronized with the PHC?

I guess userspace could also simply build its own equivalent of the
struct timecounter using this API.. hmm.

Hi Jacob,

You’re right I’m not trying to reason about the nominal frequency.
The goal is to collect (cycle, system time) pairs and use them to 
correlate raw device timestamps with host time. This doesn’t require the 
PHC to be synchronized to CLOCK_REALTIME, but it does assume the user 
can estimate the drift or nominal frequency from the ioctl data.

I’ll clarify this in v2.

Thanks,
Carolina