Re: [PATCH net-next V1 3/4] net/mlx5e: Add HW timestamping (TS) support

[Richard Cochran <richardcochran@gmail.com>]

On Thu, Dec 17, 2015 at 02:35:34PM +0200, Saeed Mahameed wrote:
> @@ -63,6 +65,7 @@
>  #define MLX5E_TX_CQ_POLL_BUDGET        128
>  #define MLX5E_UPDATE_STATS_INTERVAL    200 /* msecs */
>  #define MLX5E_SQ_BF_BUDGET             16
> +#define MLX5E_SERVICE_TASK_DELAY       (HZ / 4)

Hm...
  
> +void mlx5e_timestamp_overflow_check(struct mlx5e_priv *priv)
> +{
> +	bool timeout = time_is_before_jiffies(priv->tstamp.last_overflow_check +
> +					      priv->tstamp.overflow_period);
> +	unsigned long flags;
> +
> +	if (timeout) {
> +		write_lock_irqsave(&priv->tstamp.lock, flags);
> +		timecounter_read(&priv->tstamp.clock);
> +		write_unlock_irqrestore(&priv->tstamp.lock, flags);
> +		priv->tstamp.last_overflow_check = jiffies;

Here you have extra book keeping, because the rate of the work
callbacks is not the same as the rate of the overflow checks.

> +	}
> +}

> +void mlx5e_timestamp_init(struct mlx5e_priv *priv)
> +{
> +	struct mlx5e_tstamp *tstamp = &priv->tstamp;
> +	u64 ns;
> +	u64 frac = 0;
> +	u32 dev_freq;
> +
> +	mlx5e_timestamp_init_config(tstamp);
> +	dev_freq = MLX5_CAP_GEN(priv->mdev, device_frequency_khz);
> +	if (!dev_freq) {
> +		mlx5_core_warn(priv->mdev, "invalid device_frequency_khz. %s failed\n",
> +			       __func__);
> +		return;
> +	}
> +	rwlock_init(&tstamp->lock);
> +	memset(&tstamp->cycles, 0, sizeof(tstamp->cycles));
> +	tstamp->cycles.read = mlx5e_read_clock;
> +	tstamp->cycles.shift = MLX5E_CYCLES_SHIFT;
> +	tstamp->cycles.mult = clocksource_khz2mult(dev_freq,
> +						   tstamp->cycles.shift);
> +	tstamp->nominal_c_mult = tstamp->cycles.mult;
> +	tstamp->cycles.mask = CLOCKSOURCE_MASK(41);
> +
> +	timecounter_init(&tstamp->clock, &tstamp->cycles,
> +			 ktime_to_ns(ktime_get_real()));
> +
> +	/* Calculate period in seconds to call the overflow watchdog - to make
> +	 * sure counter is checked at least once every wrap around.
> +	 */
> +	ns = cyclecounter_cyc2ns(&tstamp->cycles, tstamp->cycles.mask, frac,
> +				 &frac);
> +	do_div(ns, NSEC_PER_SEC / 2 / HZ);
> +	tstamp->overflow_period = ns;
> +}

And here you take great pains to calculate the rate of overflow checks...

> +/* mlx5e_service_task - Run service task for tasks that needed to be done
> + * periodically
> + */
> +static void mlx5e_service_task(struct work_struct *work)
> +{
> +	struct delayed_work *dwork = to_delayed_work(work);
> +	struct mlx5e_priv *priv = container_of(dwork, struct mlx5e_priv,
> +					       service_task);
> +
> +	mutex_lock(&priv->state_lock);
> +	if (test_bit(MLX5E_STATE_OPENED, &priv->state) &&
> +	    !test_bit(MLX5E_STATE_DESTROYING, &priv->state)) {
> +		if (MLX5_CAP_GEN(priv->mdev, device_frequency_khz)) {
> +			mlx5e_timestamp_overflow_check(priv);
> +			/* Only mlx5e_timestamp_overflow_check is called from
> +			 * this service task. schedule a new task only if clock
> +			 * is initialized. if changed, move the scheduler.
> +			 */
> +			schedule_delayed_work(dwork, MLX5E_SERVICE_TASK_DELAY);

Why not simply use the rate you calculated, rather than some hard
coded value?

Consider What happens if MLX5E_SERVICE_TASK_DELAY is too long or way
too short.

> +		}
> +	}
> +	mutex_unlock(&priv->state_lock);
> +}