BPI-R64 fully-tickless hardpps soak — 5 s GPS timepulse

Cortex-A53, 12.5 MHz counter, NO_HZ_IDLE, all periodic wakeup sources quiesced (vmstat, runtime-PM autosuspend, fair dl_server, ethernet pollers) — the only wakeup is the GPS pulse every 5 s. hardpps disciplines CLOCK_REALTIME via STA_PPSTIME with phase delivery scaled to 1/5 rate (matching the 5 s delivery window — the unscaled rate is unstable, gain −5). Kernel carries the timekeeping series; snapshot ntp_error captured in the PPS hardirq at every pulse.

Per-pulse snapshot ntp_error

Pulse arrival phase (corrected ts_real)

Discipline state

Tickless gap verification

ntp_error drain lifecycle

The proportional drain engages when accumulated ntp_error exceeds one minute's worth of ±1-dither delivery (~45 ns here), repays over ~10 s re-sized each second, hands back to the dither below threshold, and aborts mid-second if it overshoots zero.

Methodology

This run is the test asked for on the list: hardpps on a tickless kernel driven by a real PPS source, with its real jitter and IRQ latency, and the CPU genuinely idle between pulses.

• Hardware: Banana Pi R64 (MT7622, 2×Cortex-A53), 12.5 MHz arch counter, u-blox GNSS HAT. Timepulse reconfigured to one pulse per 5 s (UBX-CFG-TP5, GNSS-locked, aligned to top of second) so idle gaps are bounded by the pulse, not the pulse by the gaps. PPS on a GPIO via pps-gpio; timestamp taken in the hardirq handler.
• Kernel: the timekeeping series on 7.3-rc1, CONFIG_NO_HZ_IDLE. For this run STA_PPSTIME phase delivery is scaled to 1/5 rate to match the 5 s delivery window (the stock undamped rate is computed for a 1 s window; across a 5 s tickless gap it over-delivers 5× and the phase loop diverges geometrically — found by this bench, fix to be part of the series before NTP_PPS loses its !NO_HZ_COMMON dependency).
• Quiescence: no distro userspace — a single static PID 1 binary sleeps between once-a-minute dumps. All periodic wakeup sources found by tracing were disabled: vmstat interval raised, runtime-PM autosuspend clamped, per-CPU fair dl_server off, ethernet (mtk_eth_soc + mt7530 1 s pollers) not probed. The per-minute log dump and its mmc/console I/O are issued immediately after a pulse edge so they never perturb a timestamp. The gap histogram (above) verifies the result: the timekeeper routinely advances in single 5 s bursts (1250 ticks at HZ=250) — the only wakeup between dumps is the pulse itself.
• Instrumentation: at each pulse the hardirq records snapshot_ntp_error() — the divergence between the ideal NTP-disciplined time and the sanitized clock_gettime() line that the series corrects in pps_get_ts() — plus the corrected ts_real phase against the second boundary. The kernel discipline state (hardpps corrections, second_overflow decisions, every mult step) is logged via printk to the kmsg ring and drained to disk per dump.
• Caveats: the ∼30 µs correction stripe in the slew plot is bench self-noise (a recurring IRQ-latency event, one-sided: latency can only make a pulse look late), and this kernel still carries the trace printks; a quiet-kernel confirmation run and the tickful (nohz=off) A/B comparison on identical hardware are the follow-ups.

Data

• pulses.csv — per-pulse snapshot ntp_error, tk ledger, cycle_delta, corrected ts_real phase
• hardpps_phase.csv — per-pulse raw error, median-filtered correction, jitter, injected time_offset
• adjtimex.csv — PLL state, 1/min
• hist.csv — per-minute log2 histogram of timekeeping-advance gap sizes + max
• ntp_so.csv, tk_mult.csv, tk_cancel.csv — kernel discipline traces (second_overflow chunks, every mult step, mid-tick cancels)
• bench.log.gz — the raw on-board log (dumps + kmsg), from which the CSVs are extracted