Skip to contentSkip to conversation

net: dsa: sja1105: Ensure PTP time for rxtstamp reconstruction is not in the past

PATCH In review Vladimir Oltean 2 participants 2 messages Sep 28, 2019 – Oct 1, 2019

Subsystems (4)
lore
Share & tools
lore

Vladimir Oltean

author maintainer
#0 lore raw

PATCH 1/1 net: dsa: sja1105: Ensure PTP time for rxtstamp reconstruction is not in the past

Sometimes the PTP synchronization on the switch 'jumps':

  ptp4l[11241.155]: rms    8 max   16 freq -21732 +/-  11 delay   742 +/-   0
  ptp4l[11243.157]: rms    7 max   17 freq -21731 +/-  10 delay   744 +/-   0
  ptp4l[11245.160]: rms 33592410 max 134217731 freq +192422 +/- 8530253 delay   743 +/-   0
  ptp4l[11247.163]: rms 811631 max 964131 freq +10326 +/- 557785 delay   743 +/-   0
  ptp4l[11249.166]: rms 261936 max 533876 freq -304323 +/- 126371 delay   744 +/-   0
  ptp4l[11251.169]: rms 48700 max 57740 freq -20218 +/- 30532 delay   744 +/-   0
  ptp4l[11253.171]: rms 14570 max 30163 freq  -5568 +/- 7563 delay   742 +/-   0
  ptp4l[11255.174]: rms 2914 max 3440 freq -22001 +/- 1667 delay   744 +/-   1
  ptp4l[11257.177]: rms  811 max 1710 freq -22653 +/- 451 delay   744 +/-   1
  ptp4l[11259.180]: rms  177 max  218 freq -21695 +/-  89 delay   741 +/-   0
  ptp4l[11261.182]: rms   45 max   92 freq -21677 +/-  32 delay   742 +/-   0
  ptp4l[11263.186]: rms   14 max   32 freq -21733 +/-  11 delay   742 +/-   0
  ptp4l[11265.188]: rms    9 max   14 freq -21725 +/-  12 delay   742 +/-   0
  ptp4l[11267.191]: rms    9 max   16 freq -21727 +/-  13 delay   742 +/-   0
  ptp4l[11269.194]: rms    6 max   15 freq -21726 +/-   9 delay   743 +/-   0
  ptp4l[11271.197]: rms    8 max   15 freq -21728 +/-  11 delay   743 +/-   0
  ptp4l[11273.200]: rms    6 max   12 freq -21727 +/-   8 delay   743 +/-   0
  ptp4l[11275.202]: rms    9 max   17 freq -21720 +/-  11 delay   742 +/-   0
  ptp4l[11277.205]: rms    9 max   18 freq -21725 +/-  12 delay   742 +/-   0

Background: the switch only offers partial RX timestamps (24 bits) and it is up to the driver to read the PTP clock to fill those timestamps up to 64 bits. But the PTP clock readout needs to happen quickly enough (in 0.135 seconds, in fact), otherwise the PTP clock will wrap around 24 bits, condition which cannot be detected.

Looking at the 'max 134217731' value on output line 3, one can see that in hex it is 0x8000003. Because the PTP clock resolution is 8 ns, that means 0x1000000 in ticks, which is exactly 2^24. So indeed this is a PTP clock wraparound, but the reason might be surprising.

What is going on is that sja1105_tstamp_reconstruct(priv, now, ts) expects a "now" time that is later than the "ts" was snapshotted at. This, of course, is obvious: we read the PTP time _after_ the partial RX timestamp was received. However, the workqueue is processing frames from a skb queue and reuses the same PTP time, read once at the beginning. Normally the skb queue only contains one frame and all goes well. But when the skb queue contains two frames, the second frame that gets dequeued might have been partially timestamped by the RX MAC _after_ we had read our PTP time initially.

The code was originally like that due to concerns that SPI access for PTP time readout is a slow process, and we are time-constrained anyway (aka: premature optimization). But some timing analysis reveals that the time spent until the RX timestamp is completely reconstructed is 1 order of magnitude lower than the 0.135 s deadline even under worst-case conditions. So we can afford to read the PTP time for each frame in the RX timestamping queue, which of course ensures that the full PTP time is in the partial timestamp's future.

Fixes f3097be21bf1 ("net: dsa: sja1105: Add a state machine for RX timestamping")
---
 drivers/net/dsa/sja1105/sja1105_main.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/net/dsa/sja1105/sja1105_main.c +2 -2
index b9def744bcb3..f3d38ff144c4 100644
@@ -2005,12 +2005,12 @@ static void sja1105_rxtstamp_work(struct work_struct *work)
 
 	mutex_lock(&priv->ptp_lock);
 
-	now = priv->tstamp_cc.read(&priv->tstamp_cc);
-
 	while ((skb = skb_dequeue(&data->skb_rxtstamp_queue)) != NULL) {
 		struct skb_shared_hwtstamps *shwt = skb_hwtstamps(skb);
 		u64 ts;
 
+		now = priv->tstamp_cc.read(&priv->tstamp_cc);
+
 		*shwt = (struct skb_shared_hwtstamps) {0};
 
 		ts = SJA1105_SKB_CB(skb)->meta_tstamp;
Signature · 1 line
2.17.1
maintainer
#1 lore raw

From: Vladimir Oltean <[email protected]>
Date: Sun, 29 Sep 2019 01:08:17 +0300

Quoted 1 line

Sometimes the PTP synchronization on the switch 'jumps':

...

Quoted 30 lines

Background: the switch only offers partial RX timestamps (24 bits) and it is up to the driver to read the PTP clock to fill those timestamps up to 64 bits. But the PTP clock readout needs to happen quickly enough (in 0.135 seconds, in fact), otherwise the PTP clock will wrap around 24 bits, condition which cannot be detected.

Looking at the 'max 134217731' value on output line 3, one can see that in hex it is 0x8000003. Because the PTP clock resolution is 8 ns, that means 0x1000000 in ticks, which is exactly 2^24. So indeed this is a PTP clock wraparound, but the reason might be surprising.

What is going on is that sja1105_tstamp_reconstruct(priv, now, ts) expects a "now" time that is later than the "ts" was snapshotted at. This, of course, is obvious: we read the PTP time _after_ the partial RX timestamp was received. However, the workqueue is processing frames from a skb queue and reuses the same PTP time, read once at the beginning. Normally the skb queue only contains one frame and all goes well. But when the skb queue contains two frames, the second frame that gets dequeued might have been partially timestamped by the RX MAC _after_ we had read our PTP time initially.

The code was originally like that due to concerns that SPI access for PTP time readout is a slow process, and we are time-constrained anyway (aka: premature optimization). But some timing analysis reveals that the time spent until the RX timestamp is completely reconstructed is 1 order of magnitude lower than the 0.135 s deadline even under worst-case conditions. So we can afford to read the PTP time for each frame in the RX timestamping queue, which of course ensures that the full PTP time is in the partial timestamp's future.

Quoted 2 lines
Fixes f3097be21bf1 ("net: dsa: sja1105: Add a state machine for RX timestamping")

Applied and queued up for -stable, thanks.