On Fri, 2026-01-09 at 10:56 +0800, Wen Gu wrote:
>
> Hi all,
>
> This is an RFC to discuss the appropriate upstream home and maintenance
> model for a class of devices/drivers which expose a high-precision clock
> to userspace via the existing PHC interface (/dev/ptpX + standard PTP_*
> ioctls), but are not tied to a traditional NIC/IEEE1588 packet
> timestamping pipeline.
>
> Examples already in the tree include (non-exhaustive):
>
> - drivers/ptp/ptp_kvm.c [1]
> - drivers/ptp/ptp_vmw.c [2]
> - drivers/ptp/ptp_s390.c [3]
>
> There are also examples living in their respective subsystem (out of
> scope for this RFC),
> e.g. drivers/hv/hv_util.c [4] and drivers/virtio/virtio_rtc_ptp.c [5].
>
> We (Alibaba Cloud) also posted a driver for a CIPU-provided high-precision
> clock for review [6]. Based on existing in-tree precedent, we placed it
> under drivers/ptp/ and sent it to the netdev list.
>
> During review, concerns were raised that such "non-NIC / pure" PHC drivers
> are not a good fit for netdev maintainership [7], since they are primarily
> time/clock devices rather than networking protocol features.
>
> As a result, I’m sending this RFC to align on a consistent upstream home
> and maintainer model for this class of drivers, both for the existing ones
> and future additions.
>
> #
> ## Observation 1: PHC core/API are already not bound to NIC/IEEE1588
> #
>
> Although PHC support is original associated with NIC-based IEEE 1588
> timestamping, the kernel tree already contains multiple non-NIC PHC
> implementations (examples above), including long-standing and recently
> added drivers. This reflects the reality that the PHC interface is no
> longer tightly bound to NIC/IEEE1588 implementations.
>
> This is enabled by the PHC interface's clean design, it provides a
> well-scoped, layered abstraction that separates the userspace access
> mechanism (/dev/ptpX + standard ioctl semantics) from the underlying
> clock implementation and discipline method (NIC/IEEE1588 packet timestamping
> pipeline, virtualization-provided clocks, platform/firmware time services,
> etc.). The interface defines only generic clock-operation semantics, without
> baking in assumptions about how the clock is produced or synchronized.
>
> Because of this elegant decoupling, the PHC API naturally fits
> "pure time source" devices as long as they can provide a stable, precise
> hardware clock. In practice, PHC has effectively become Linux’s common
> API for high-precision device clocks, rather than inherently bound to
> an IEEE1588 NIC implementation.
>
> #
> ## Observation 2: the PHC (/dev/ptpX) has an established userspace ecosystem
> #
>
> The PHC character device interface (/dev/ptpX + standard PTP_* ioctls) is
> a mature, stable, and widely deployed userspace API for accessing
> high-precision clocks on Linux. It is already the common interface consumed
> by existing software stacks (e.g. chrony, and other applications built around
> PHC devices)
>
> Introducing a new clock type or a new userspace API (e.g. /dev/XXX) would
> require widespread userspace changes, duplicated tooling, and long-term
> fragmentation. This RFC is explicitly NOT proposing a new userspace API.
>
> #
> ## Goal
> #
>
> Establish an appropriate upstream home and maintainer model for "pure time
> source" PHC drivers. If they are not suitable to be maintained under netdev,
> we need a clear place and maintainer(s) for them, and a consistent policy
> for accepting new ones.
>
> #
> ## Proposal
> #
>
> 1. Reorganize drivers/ptp/ to make the interface/implementation split
> explicit,
>
> * drivers/ptp/core : PTP core infrastructure and API.
> (e.g. ptp_chardev.c, ptp_clock.c,
> ptp_sysfs.c, etc.)
>
> * drivers/ptp/pure : Non-network ("pure clock") implementation,
> they are typically platform/architecture/
> virtualization-provided time sources.
> (e.g. ptp_kvm, ptp_vmw, ptp_vmclock,
> ptp_s390, etc.)
>
> * drivers/ptp/* : Network timestamping oriented implementation,
> they primarily used together with IEEE1588
> over the network.
> (e.g. ptp_qoriq, ptp_pch, ptp_dp83640,
> ptp_idt82p33 etc.)
>
> 2. Transition drivers/ptp/pure from netdev maintainership to
> clock/time maintainership (with an appropriate MAINTAINERS entry,
> e.g. PURE TIME PHC), since these PHC implementations are primarily
> clock devices and not network-oriented. New similar drivers can be
> added under drivers/ptp/pure as well.
>
>
> Possible alternatives (please suggest others):
>
> - Move/align "pure time source" PHC drivers under an existing
> timekeeping/clocksource/virt area, without changing the userspace API.
>
>
> I’d like to drive this discussion towards consensus, and I’m happy to
> adapt our series to match whatever direction is agreed upon.
>
> [1]
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=a0e136d436ded817c0aade72efdefa56a00b4e5e
> [2]
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=7d10001e20e46ad6ad95622164686bc2cbfc9802
> [3]
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=2d7de7a3010d713fb89b7ba99e6fdc14475ad106
> [4]
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=3716a49a81ba19dda7202633a68b28564ba95eb5
> [5]
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=9a17125a18f9ae1e1233a8e2d919059445b9d6fd
> [6]
> https://lore.kernel.org/netdev/[email protected]/
> [7] https://lore.kernel.org/netdev/[email protected]/
>
> Thanks for any input.Thanks for starting this discussion. I agree that the existing 'pure' PHC drivers should keep the option of presenting themselves as PTP devices to userspace. I would probably have suggested moving them out of drivers/ptp/… to somewhere else under drivers/ entirely, but that's bikeshedding. I do think we have slightly different requirements for the pure PHCs too though, particularly the virt-focused ones. It would be good if we could set a guest's clock from them at startup, and the primary focus of them is for calibrating the guest's CLOCK_REALTIME. Which means we do actually care about consuming UTC offset and leap second information from them, not just TAI. I'd also like microvms to be able to consume time directly, especially from vmclock, without needing a full-blown NTP-capable userspace. I've experimented with simulating 1PPS support to feed into the kernel's timekeeping, although I don't think that's the best answer: https://lore.kernel.org/all/[email protected]/ We could do with harmonising the workarounds for kvmclock too. I made sure the vmclock driver reports its timestamp pairs in terms of either CSID_X86_TSC or CSID_X86_KVM_CLK as appropriate, but ptp_kvm *only* uses kvmclock (which is daft, since anyone who cares about accurate time will be using tsc). I was thinking that interface of the pure PHC drivers could be really simple, and our new infrastructure could provide the ptp_clock_info glue, including the kvmclock conversion if needed. And *also* hook them in for setting the clock at startup, and even calibrating CLOCK_REALTIME.
smime.p7s
Description: S/MIME cryptographic signature
