Re: [RFC PATCH v3 10/10] kernel/time: Add /dev/vmclock_host miscdev

[Wen Gu]

On 2026/5/20 21:33, David Woodhouse wrote:

> +/*
> + * Called from pvclock_gtod_notify on every timekeeping update.
> + * Only does real work when ntp_tick or skew_delta changes.
> + */
> +static int vmclock_host_notify(struct notifier_block *nb,
> +			       unsigned long was_set, void *data)
> +{

<...>

> +
> +	/* Compute time tuple: C = A + ntp_error + time_offset */
> +	ns = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
> +	sec = tk->xtime_sec + tk->tai_offset;
> +	ns += tk->ntp_error >> (tk->tkr_mono.shift + tk->ntp_error_shift);
> +	ns += ntp_get_time_offset_ns(tk->id);
> +
> +	while (ns < 0) {
> +		ns += NSEC_PER_SEC;
> +		sec--;
> +	}
> +	while (ns >= NSEC_PER_SEC) {
> +		ns -= NSEC_PER_SEC;
> +		sec++;
> +	}
> +
> +	counter_value = tk->tkr_mono.cycle_last;
> +	hi = div64_u64_rem((u64)ns << 32, 1000000000ULL, &rem);
> +	time_frac = (hi << 32) | div64_u64(rem << 32, 1000000000ULL);
> +
> +	clock_status = !(ntp_get_status() & STA_UNSYNC) ?
> +		VMCLOCK_STATUS_SYNCHRONIZED : VMCLOCK_STATUS_FREERUNNING;
> +
> +	/* Prepare le values */
> +	le_counter_value = cpu_to_le64(counter_value);
> +	le_time_sec = cpu_to_le64(sec);
> +	le_time_frac = cpu_to_le64(time_frac);
> +	le_period_frac = cpu_to_le64(cached_period_frac);
> +	period_shift = cached_period_shift;
> +
> +	/* Update page under seqcount */
> +	WRITE_ONCE(clk->seq_count, cpu_to_le32(
> +		le32_to_cpu(READ_ONCE(clk->seq_count)) + 1));
> +	smp_wmb();
> +
> +	clk->counter_id = counter_id;
> +	clk->counter_value = le_counter_value;
> +	clk->time_sec = le_time_sec;
> +	clk->time_frac_sec = le_time_frac;
> +	if (period_changed) {
> +		clk->counter_period_frac_sec = le_period_frac;
> +		clk->counter_period_shift = period_shift;
> +	}
> +	clk->clock_status = clock_status;
> +
> +	/* Set leap second indicator from NTP time_state */
> +	switch (ntp_get_time_state()) {
> +	case TIME_INS:
> +		clk->leap_indicator = VMCLOCK_LEAP_PRE_POS;
> +		break;
> +	case TIME_DEL:
> +		clk->leap_indicator = VMCLOCK_LEAP_PRE_NEG;
> +		break;
> +	case TIME_OOP:
> +		clk->leap_indicator = VMCLOCK_LEAP_POS;
> +		break;
> +	case TIME_WAIT:
> +		clk->leap_indicator = (ntp_get_status() & STA_DEL) ?
> +			VMCLOCK_LEAP_POST_NEG : VMCLOCK_LEAP_POST_POS;
> +		break;
> +	default:
> +		clk->leap_indicator = VMCLOCK_LEAP_NONE;
> +		break;
> +	}
> +
> +	/* Export as TAI if tai_offset is known, otherwise UTC */
> +	if (tk->tai_offset) {
> +		clk->time_type = VMCLOCK_TIME_TAI;
> +		clk->tai_offset_sec = cpu_to_le16((s16)tk->tai_offset);

I think this should be cpu_to_le16(-(s16)tk->tai_offset)?

vmclock_set_tk_reference() in Patch 9 treats clk->tai_offset_sec as
a (UTC - TAI) offset, so the sign here probably needs to be inverted.

> +		clk->flags = cpu_to_le64(VMCLOCK_FLAG_TAI_OFFSET_VALID |
> +					 VMCLOCK_FLAG_TIME_MONOTONIC |
> +					 VMCLOCK_FLAG_NOTIFICATION_PRESENT);
> +	} else {
> +		clk->time_type = VMCLOCK_TIME_UTC;
> +		clk->tai_offset_sec = 0;
> +		clk->flags = cpu_to_le64(VMCLOCK_FLAG_TIME_MONOTONIC |
> +					 VMCLOCK_FLAG_NOTIFICATION_PRESENT);
> +	}
> +
> +	smp_wmb();
> +	WRITE_ONCE(clk->seq_count, cpu_to_le32(
> +		le32_to_cpu(READ_ONCE(clk->seq_count)) + 1));
> +
> +	wake_up_interruptible(&vmclock_wait);
> +	return NOTIFY_DONE;
> +}