[PATCH net v2 0/4][pull request] igc: TX timestamping fixes

[PATCH net v2 0/4][pull request] igc: TX timestamping fixes

[Tony Nguyen]

This is the fixes part of the series intended to add support for using
the 4 timestamp registers present in i225/i226.

Moving the timestamp handling to be inline with the interrupt handling
has the advantage of improving the TX timestamping retrieval latency,
here are some numbers using ntpperf:

Before:

$ sudo ./ntpperf -i enp3s0 -m 10:22:22:22:22:21 -d 192.168.1.3 -s 172.18.0.0/16 -I -H -o -37
               |          responses            |     TX timestamp offset (ns)
rate   clients |  lost invalid   basic  xleave |    min    mean     max stddev
1000       100   0.00%   0.00%   0.00% 100.00%      -56      +9     +52     19
1500       150   0.00%   0.00%   0.00% 100.00%      -40     +30     +75     22
2250       225   0.00%   0.00%   0.00% 100.00%      -11     +29     +72     15
3375       337   0.00%   0.00%   0.00% 100.00%      -18     +40     +88     22
5062       506   0.00%   0.00%   0.00% 100.00%      -19     +23     +77     15
7593       759   0.00%   0.00%   0.00% 100.00%       +7     +47   +5168     43
11389     1138   0.00%   0.00%   0.00% 100.00%      -11     +41   +5240     39
17083     1708   0.00%   0.00%   0.00% 100.00%      +19     +60   +5288     50
25624     2562   0.00%   0.00%   0.00% 100.00%       +1     +56   +5368     58
38436     3843   0.00%   0.00%   0.00% 100.00%      -84     +12   +8847     66
57654     5765   0.00%   0.00% 100.00%   0.00%
86481     8648   0.00%   0.00% 100.00%   0.00%
129721   12972   0.00%   0.00% 100.00%   0.00%
194581   16384   0.00%   0.00% 100.00%   0.00%
291871   16384  27.35%   0.00%  72.65%   0.00%
437806   16384  50.05%   0.00%  49.95%   0.00%

After:

$ sudo ./ntpperf -i enp3s0 -m 10:22:22:22:22:21 -d 192.168.1.3 -s 172.18.0.0/16 -I -H -o -37
               |          responses            |     TX timestamp offset (ns)
rate   clients |  lost invalid   basic  xleave |    min    mean     max stddev
1000       100   0.00%   0.00%   0.00% 100.00%      -44      +0     +61     19
1500       150   0.00%   0.00%   0.00% 100.00%       -6     +39     +81     16
2250       225   0.00%   0.00%   0.00% 100.00%      -22     +25     +69     15
3375       337   0.00%   0.00%   0.00% 100.00%      -28     +15     +56     14
5062       506   0.00%   0.00%   0.00% 100.00%       +7     +78    +143     27
7593       759   0.00%   0.00%   0.00% 100.00%      -54     +24    +144     47
11389     1138   0.00%   0.00%   0.00% 100.00%      -90     -33     +28     21
17083     1708   0.00%   0.00%   0.00% 100.00%      -50      -2     +35     14
25624     2562   0.00%   0.00%   0.00% 100.00%      -62      +7     +66     23
38436     3843   0.00%   0.00%   0.00% 100.00%      -33     +30   +5395     36
57654     5765   0.00%   0.00% 100.00%   0.00%
86481     8648   0.00%   0.00% 100.00%   0.00%
129721   12972   0.00%   0.00% 100.00%   0.00%
194581   16384  19.50%   0.00%  80.50%   0.00%
291871   16384  35.81%   0.00%  64.19%   0.00%
437806   16384  55.40%   0.00%  44.60%   0.00%

During this series, and to show that as is always the case, things are
never easy as they should be, a hardware issue was found, and it took
some time to find the workaround(s). The bug and workaround are better
explained in patch 4/4.

Note: the workaround has a simpler alternative, but it would involve
adding support for the other timestamp registers, and only using the
TXSTMP{H/L}_0 as a way to clear the interrupt. But I feel bad about
throwing this kind of resources away. Didn't test this extensively but
it should work.

Also, as Marc Kleine-Budde suggested, after some consensus is reached
on this series, most parts of it will be proposed for igb.
---
v2:
 - Fixed possible race condition when disabling TX timestamping, added
   a per queue flag, should make the hot path (no timestamps enabled)
   a bit nicer
 - Renamed the igc_ptp_tx_work() to something more sensible (it's no
   longer called in a workqueue)
 - Improved commit message, added more details about how to trigger
   the hardware issue, and more details about the alternative
   workaround to the cover letter
 - Added some numbers, from the cover letter, to the commit message
   itself

v1: https://lore.kernel.org/netdev/20230530174928.2516291-1-anthony.l.nguyen@intel.com/

The following are changes since commit 2ba7e7ebb6a71407cbe25cd349c9b05d40520bf0:
  Merge tag 'nf-23-06-21' of git://git.kernel.org/pub/scm/linux/kernel/git/netfilter/nf
and are available in the git repository at:
  git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/net-queue 1GbE

Vinicius Costa Gomes (4):
  igc: Fix race condition in PTP tx code
  igc: Check if hardware TX timestamping is enabled earlier
  igc: Retrieve TX timestamp during interrupt handling
  igc: Work around HW bug causing missing timestamps

 drivers/net/ethernet/intel/igc/igc.h      |   8 +-
 drivers/net/ethernet/intel/igc/igc_main.c |  14 ++-
 drivers/net/ethernet/intel/igc/igc_ptp.c  | 142 ++++++++++++++++------
 3 files changed, 117 insertions(+), 47 deletions(-)

-- 
2.38.1

[PATCH net v2 1/4] igc: Fix race condition in PTP tx code

[Tony Nguyen]

From: Vinicius Costa Gomes <vinicius.gomes@intel.com>

Currently, the igc driver supports timestamping only one tx packet at a
time. During the transmission flow, the skb that requires hardware
timestamping is saved in adapter->ptp_tx_skb. Once hardware has the
timestamp, an interrupt is delivered, and adapter->ptp_tx_work is
scheduled. In igc_ptp_tx_work(), we read the timestamp register, update
adapter->ptp_tx_skb, and notify the network stack.

While the thread executing the transmission flow (the user process
running in kernel mode) and the thread executing ptp_tx_work don't
access adapter->ptp_tx_skb concurrently, there are two other places
where adapter->ptp_tx_skb is accessed: igc_ptp_tx_hang() and
igc_ptp_suspend().

igc_ptp_tx_hang() is executed by the adapter->watchdog_task worker
thread which runs periodically so it is possible we have two threads
accessing ptp_tx_skb at the same time. Consider the following scenario:
right after __IGC_PTP_TX_IN_PROGRESS is set in igc_xmit_frame_ring(),
igc_ptp_tx_hang() is executed. Since adapter->ptp_tx_start hasn't been
written yet, this is considered a timeout and adapter->ptp_tx_skb is
cleaned up.

This patch fixes the issue described above by adding the ptp_tx_lock to
protect access to ptp_tx_skb and ptp_tx_start fields from igc_adapter.
Since igc_xmit_frame_ring() called in atomic context by the networking
stack, ptp_tx_lock is defined as a spinlock, and the irq safe variants
of lock/unlock are used.

With the introduction of the ptp_tx_lock, the __IGC_PTP_TX_IN_PROGRESS
flag doesn't provide much of a use anymore so this patch gets rid of it.

Fixes: 2c344ae24501 ("igc: Add support for TX timestamping")
Signed-off-by: Andre Guedes <andre.guedes@intel.com>
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Reviewed-by: Kurt Kanzenbach <kurt@linutronix.de>
Tested-by: Naama Meir <naamax.meir@linux.intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
---
 drivers/net/ethernet/intel/igc/igc.h      |  5 +-
 drivers/net/ethernet/intel/igc/igc_main.c |  9 ++--
 drivers/net/ethernet/intel/igc/igc_ptp.c  | 57 ++++++++++++-----------
 3 files changed, 41 insertions(+), 30 deletions(-)

diff --git a/drivers/net/ethernet/intel/igc/igc.h b/drivers/net/ethernet/intel/igc/igc.h
index 34aebf00a512..7da0657ea48f 100644
--- a/drivers/net/ethernet/intel/igc/igc.h
+++ b/drivers/net/ethernet/intel/igc/igc.h
@@ -229,6 +229,10 @@ struct igc_adapter {
 	struct ptp_clock *ptp_clock;
 	struct ptp_clock_info ptp_caps;
 	struct work_struct ptp_tx_work;
+	/* Access to ptp_tx_skb and ptp_tx_start are protected by the
+	 * ptp_tx_lock.
+	 */
+	spinlock_t ptp_tx_lock;
 	struct sk_buff *ptp_tx_skb;
 	struct hwtstamp_config tstamp_config;
 	unsigned long ptp_tx_start;
@@ -401,7 +405,6 @@ enum igc_state_t {
 	__IGC_TESTING,
 	__IGC_RESETTING,
 	__IGC_DOWN,
-	__IGC_PTP_TX_IN_PROGRESS,
 };
 
 enum igc_tx_flags {
diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
index fa764190f270..9fcb263bd3a7 100644
--- a/drivers/net/ethernet/intel/igc/igc_main.c
+++ b/drivers/net/ethernet/intel/igc/igc_main.c
@@ -1590,9 +1590,10 @@ static netdev_tx_t igc_xmit_frame_ring(struct sk_buff *skb,
 		 * the other timer registers before skipping the
 		 * timestamping request.
 		 */
-		if (adapter->tstamp_config.tx_type == HWTSTAMP_TX_ON &&
-		    !test_and_set_bit_lock(__IGC_PTP_TX_IN_PROGRESS,
-					   &adapter->state)) {
+		unsigned long flags;
+
+		spin_lock_irqsave(&adapter->ptp_tx_lock, flags);
+		if (adapter->tstamp_config.tx_type == HWTSTAMP_TX_ON && !adapter->ptp_tx_skb) {
 			skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
 			tx_flags |= IGC_TX_FLAGS_TSTAMP;
 
@@ -1601,6 +1602,8 @@ static netdev_tx_t igc_xmit_frame_ring(struct sk_buff *skb,
 		} else {
 			adapter->tx_hwtstamp_skipped++;
 		}
+
+		spin_unlock_irqrestore(&adapter->ptp_tx_lock, flags);
 	}
 
 	if (skb_vlan_tag_present(skb)) {
diff --git a/drivers/net/ethernet/intel/igc/igc_ptp.c b/drivers/net/ethernet/intel/igc/igc_ptp.c
index 4e10ced736db..56128e55f5c0 100644
--- a/drivers/net/ethernet/intel/igc/igc_ptp.c
+++ b/drivers/net/ethernet/intel/igc/igc_ptp.c
@@ -603,6 +603,7 @@ static int igc_ptp_set_timestamp_mode(struct igc_adapter *adapter,
 	return 0;
 }
 
+/* Requires adapter->ptp_tx_lock held by caller. */
 static void igc_ptp_tx_timeout(struct igc_adapter *adapter)
 {
 	struct igc_hw *hw = &adapter->hw;
@@ -610,7 +611,6 @@ static void igc_ptp_tx_timeout(struct igc_adapter *adapter)
 	dev_kfree_skb_any(adapter->ptp_tx_skb);
 	adapter->ptp_tx_skb = NULL;
 	adapter->tx_hwtstamp_timeouts++;
-	clear_bit_unlock(__IGC_PTP_TX_IN_PROGRESS, &adapter->state);
 	/* Clear the tx valid bit in TSYNCTXCTL register to enable interrupt. */
 	rd32(IGC_TXSTMPH);
 	netdev_warn(adapter->netdev, "Tx timestamp timeout\n");
@@ -618,20 +618,20 @@ static void igc_ptp_tx_timeout(struct igc_adapter *adapter)
 
 void igc_ptp_tx_hang(struct igc_adapter *adapter)
 {
-	bool timeout = time_is_before_jiffies(adapter->ptp_tx_start +
-					      IGC_PTP_TX_TIMEOUT);
+	unsigned long flags;
 
-	if (!test_bit(__IGC_PTP_TX_IN_PROGRESS, &adapter->state))
-		return;
+	spin_lock_irqsave(&adapter->ptp_tx_lock, flags);
 
-	/* If we haven't received a timestamp within the timeout, it is
-	 * reasonable to assume that it will never occur, so we can unlock the
-	 * timestamp bit when this occurs.
-	 */
-	if (timeout) {
-		cancel_work_sync(&adapter->ptp_tx_work);
-		igc_ptp_tx_timeout(adapter);
-	}
+	if (!adapter->ptp_tx_skb)
+		goto unlock;
+
+	if (time_is_after_jiffies(adapter->ptp_tx_start + IGC_PTP_TX_TIMEOUT))
+		goto unlock;
+
+	igc_ptp_tx_timeout(adapter);
+
+unlock:
+	spin_unlock_irqrestore(&adapter->ptp_tx_lock, flags);
 }
 
 /**
@@ -641,6 +641,8 @@ void igc_ptp_tx_hang(struct igc_adapter *adapter)
  * If we were asked to do hardware stamping and such a time stamp is
  * available, then it must have been for this skb here because we only
  * allow only one such packet into the queue.
+ *
+ * Context: Expects adapter->ptp_tx_lock to be held by caller.
  */
 static void igc_ptp_tx_hwtstamp(struct igc_adapter *adapter)
 {
@@ -676,13 +678,7 @@ static void igc_ptp_tx_hwtstamp(struct igc_adapter *adapter)
 	shhwtstamps.hwtstamp =
 		ktime_add_ns(shhwtstamps.hwtstamp, adjust);
 
-	/* Clear the lock early before calling skb_tstamp_tx so that
-	 * applications are not woken up before the lock bit is clear. We use
-	 * a copy of the skb pointer to ensure other threads can't change it
-	 * while we're notifying the stack.
-	 */
 	adapter->ptp_tx_skb = NULL;
-	clear_bit_unlock(__IGC_PTP_TX_IN_PROGRESS, &adapter->state);
 
 	/* Notify the stack and free the skb after we've unlocked */
 	skb_tstamp_tx(skb, &shhwtstamps);
@@ -693,24 +689,33 @@ static void igc_ptp_tx_hwtstamp(struct igc_adapter *adapter)
  * igc_ptp_tx_work
  * @work: pointer to work struct
  *
- * This work function polls the TSYNCTXCTL valid bit to determine when a
- * timestamp has been taken for the current stored skb.
+ * This work function checks the TSYNCTXCTL valid bit to determine when
+ * a timestamp has been taken for the current stored skb.
  */
 static void igc_ptp_tx_work(struct work_struct *work)
 {
 	struct igc_adapter *adapter = container_of(work, struct igc_adapter,
 						   ptp_tx_work);
 	struct igc_hw *hw = &adapter->hw;
+	unsigned long flags;
 	u32 tsynctxctl;
 
-	if (!test_bit(__IGC_PTP_TX_IN_PROGRESS, &adapter->state))
-		return;
+	spin_lock_irqsave(&adapter->ptp_tx_lock, flags);
+
+	if (!adapter->ptp_tx_skb)
+		goto unlock;
 
 	tsynctxctl = rd32(IGC_TSYNCTXCTL);
-	if (WARN_ON_ONCE(!(tsynctxctl & IGC_TSYNCTXCTL_TXTT_0)))
-		return;
+	tsynctxctl &= IGC_TSYNCTXCTL_TXTT_0;
+	if (!tsynctxctl) {
+		WARN_ONCE(1, "Received a TSTAMP interrupt but no TSTAMP is ready.\n");
+		goto unlock;
+	}
 
 	igc_ptp_tx_hwtstamp(adapter);
+
+unlock:
+	spin_unlock_irqrestore(&adapter->ptp_tx_lock, flags);
 }
 
 /**
@@ -959,6 +964,7 @@ void igc_ptp_init(struct igc_adapter *adapter)
 		return;
 	}
 
+	spin_lock_init(&adapter->ptp_tx_lock);
 	spin_lock_init(&adapter->tmreg_lock);
 	INIT_WORK(&adapter->ptp_tx_work, igc_ptp_tx_work);
 
@@ -1023,7 +1029,6 @@ void igc_ptp_suspend(struct igc_adapter *adapter)
 	cancel_work_sync(&adapter->ptp_tx_work);
 	dev_kfree_skb_any(adapter->ptp_tx_skb);
 	adapter->ptp_tx_skb = NULL;
-	clear_bit_unlock(__IGC_PTP_TX_IN_PROGRESS, &adapter->state);
 
 	if (pci_device_is_present(adapter->pdev)) {
 		igc_ptp_time_save(adapter);
-- 
2.38.1

[PATCH net v2 2/4] igc: Check if hardware TX timestamping is enabled earlier

[Tony Nguyen]

From: Vinicius Costa Gomes <vinicius.gomes@intel.com>

Before requesting a packet transmission to be hardware timestamped,
check if the user has TX timestamping enabled. Fixes an issue that if
a packet was internally forwarded to the NIC, and it had the
SKBTX_HW_TSTAMP flag set, the driver would mark that timestamp as
skipped.

In reality, that timestamp was "not for us", as TX timestamp could
never be enabled in the NIC.

Checking if the TX timestamping is enabled earlier has a secondary
effect that when TX timestamping is disabled, there's no need to check
for timestamp timeouts.

We should only take care to free any pending timestamp when TX
timestamping is disabled, as that skb would never be released
otherwise.

Fixes: 2c344ae24501 ("igc: Add support for TX timestamping")
Suggested-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Tested-by: Naama Meir <naamax.meir@linux.intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
---
 drivers/net/ethernet/intel/igc/igc.h      |  1 +
 drivers/net/ethernet/intel/igc/igc_main.c |  5 +--
 drivers/net/ethernet/intel/igc/igc_ptp.c  | 42 +++++++++++++++++++++--
 3 files changed, 43 insertions(+), 5 deletions(-)

diff --git a/drivers/net/ethernet/intel/igc/igc.h b/drivers/net/ethernet/intel/igc/igc.h
index 7da0657ea48f..66fb67c17e4f 100644
--- a/drivers/net/ethernet/intel/igc/igc.h
+++ b/drivers/net/ethernet/intel/igc/igc.h
@@ -581,6 +581,7 @@ enum igc_ring_flags_t {
 	IGC_RING_FLAG_TX_CTX_IDX,
 	IGC_RING_FLAG_TX_DETECT_HANG,
 	IGC_RING_FLAG_AF_XDP_ZC,
+	IGC_RING_FLAG_TX_HWTSTAMP,
 };
 
 #define ring_uses_large_buffer(ring) \
diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
index 9fcb263bd3a7..eb50bfd5b867 100644
--- a/drivers/net/ethernet/intel/igc/igc_main.c
+++ b/drivers/net/ethernet/intel/igc/igc_main.c
@@ -1585,7 +1585,8 @@ static netdev_tx_t igc_xmit_frame_ring(struct sk_buff *skb,
 		}
 	}
 
-	if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)) {
+	if (unlikely(test_bit(IGC_RING_FLAG_TX_HWTSTAMP, &tx_ring->flags) &&
+		     skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)) {
 		/* FIXME: add support for retrieving timestamps from
 		 * the other timer registers before skipping the
 		 * timestamping request.
@@ -1593,7 +1594,7 @@ static netdev_tx_t igc_xmit_frame_ring(struct sk_buff *skb,
 		unsigned long flags;
 
 		spin_lock_irqsave(&adapter->ptp_tx_lock, flags);
-		if (adapter->tstamp_config.tx_type == HWTSTAMP_TX_ON && !adapter->ptp_tx_skb) {
+		if (!adapter->ptp_tx_skb) {
 			skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
 			tx_flags |= IGC_TX_FLAGS_TSTAMP;
 
diff --git a/drivers/net/ethernet/intel/igc/igc_ptp.c b/drivers/net/ethernet/intel/igc/igc_ptp.c
index 56128e55f5c0..42f622ceb64b 100644
--- a/drivers/net/ethernet/intel/igc/igc_ptp.c
+++ b/drivers/net/ethernet/intel/igc/igc_ptp.c
@@ -536,9 +536,36 @@ static void igc_ptp_enable_rx_timestamp(struct igc_adapter *adapter)
 	wr32(IGC_TSYNCRXCTL, val);
 }
 
+static void igc_ptp_clear_tx_tstamp(struct igc_adapter *adapter)
+{
+	unsigned long flags;
+
+	cancel_work_sync(&adapter->ptp_tx_work);
+
+	spin_lock_irqsave(&adapter->ptp_tx_lock, flags);
+
+	dev_kfree_skb_any(adapter->ptp_tx_skb);
+	adapter->ptp_tx_skb = NULL;
+
+	spin_unlock_irqrestore(&adapter->ptp_tx_lock, flags);
+}
+
 static void igc_ptp_disable_tx_timestamp(struct igc_adapter *adapter)
 {
 	struct igc_hw *hw = &adapter->hw;
+	int i;
+
+	/* Clear the flags first to avoid new packets to be enqueued
+	 * for TX timestamping.
+	 */
+	for (i = 0; i < adapter->num_tx_queues; i++) {
+		struct igc_ring *tx_ring = adapter->tx_ring[i];
+
+		clear_bit(IGC_RING_FLAG_TX_HWTSTAMP, &tx_ring->flags);
+	}
+
+	/* Now we can clean the pending TX timestamp requests. */
+	igc_ptp_clear_tx_tstamp(adapter);
 
 	wr32(IGC_TSYNCTXCTL, 0);
 }
@@ -546,12 +573,23 @@ static void igc_ptp_disable_tx_timestamp(struct igc_adapter *adapter)
 static void igc_ptp_enable_tx_timestamp(struct igc_adapter *adapter)
 {
 	struct igc_hw *hw = &adapter->hw;
+	int i;
 
 	wr32(IGC_TSYNCTXCTL, IGC_TSYNCTXCTL_ENABLED | IGC_TSYNCTXCTL_TXSYNSIG);
 
 	/* Read TXSTMP registers to discard any timestamp previously stored. */
 	rd32(IGC_TXSTMPL);
 	rd32(IGC_TXSTMPH);
+
+	/* The hardware is ready to accept TX timestamp requests,
+	 * notify the transmit path.
+	 */
+	for (i = 0; i < adapter->num_tx_queues; i++) {
+		struct igc_ring *tx_ring = adapter->tx_ring[i];
+
+		set_bit(IGC_RING_FLAG_TX_HWTSTAMP, &tx_ring->flags);
+	}
+
 }
 
 /**
@@ -1026,9 +1064,7 @@ void igc_ptp_suspend(struct igc_adapter *adapter)
 	if (!(adapter->ptp_flags & IGC_PTP_ENABLED))
 		return;
 
-	cancel_work_sync(&adapter->ptp_tx_work);
-	dev_kfree_skb_any(adapter->ptp_tx_skb);
-	adapter->ptp_tx_skb = NULL;
+	igc_ptp_clear_tx_tstamp(adapter);
 
 	if (pci_device_is_present(adapter->pdev)) {
 		igc_ptp_time_save(adapter);
-- 
2.38.1

[PATCH net v2 3/4] igc: Retrieve TX timestamp during interrupt handling

[Tony Nguyen]

From: Vinicius Costa Gomes <vinicius.gomes@intel.com>

When the interrupt is handled, the TXTT_0 bit in the TSYNCTXCTL
register should already be set and the timestamp value already loaded
in the appropriate register.

This simplifies the handling, and reduces the latency for retrieving
the TX timestamp, which increase the amount of TX timestamps that can
be handled in a given time period.

As the "work" function doesn't run in a workqueue anymore, rename it
to something more sensible, a event handler.

Using ntpperf[1] we can see the following performance improvements:

Before:

$ sudo ./ntpperf -i enp3s0 -m 10:22:22:22:22:21 -d 192.168.1.3 -s 172.18.0.0/16 -I -H -o -37
               |          responses            |     TX timestamp offset (ns)
rate   clients |  lost invalid   basic  xleave |    min    mean     max stddev
1000       100   0.00%   0.00%   0.00% 100.00%      -56      +9     +52     19
1500       150   0.00%   0.00%   0.00% 100.00%      -40     +30     +75     22
2250       225   0.00%   0.00%   0.00% 100.00%      -11     +29     +72     15
3375       337   0.00%   0.00%   0.00% 100.00%      -18     +40     +88     22
5062       506   0.00%   0.00%   0.00% 100.00%      -19     +23     +77     15
7593       759   0.00%   0.00%   0.00% 100.00%       +7     +47   +5168     43
11389     1138   0.00%   0.00%   0.00% 100.00%      -11     +41   +5240     39
17083     1708   0.00%   0.00%   0.00% 100.00%      +19     +60   +5288     50
25624     2562   0.00%   0.00%   0.00% 100.00%       +1     +56   +5368     58
38436     3843   0.00%   0.00%   0.00% 100.00%      -84     +12   +8847     66
57654     5765   0.00%   0.00% 100.00%   0.00%
86481     8648   0.00%   0.00% 100.00%   0.00%
129721   12972   0.00%   0.00% 100.00%   0.00%
194581   16384   0.00%   0.00% 100.00%   0.00%
291871   16384  27.35%   0.00%  72.65%   0.00%
437806   16384  50.05%   0.00%  49.95%   0.00%

After:

$ sudo ./ntpperf -i enp3s0 -m 10:22:22:22:22:21 -d 192.168.1.3 -s 172.18.0.0/16 -I -H -o -37
               |          responses            |     TX timestamp offset (ns)
rate   clients |  lost invalid   basic  xleave |    min    mean     max stddev
1000       100   0.00%   0.00%   0.00% 100.00%      -44      +0     +61     19
1500       150   0.00%   0.00%   0.00% 100.00%       -6     +39     +81     16
2250       225   0.00%   0.00%   0.00% 100.00%      -22     +25     +69     15
3375       337   0.00%   0.00%   0.00% 100.00%      -28     +15     +56     14
5062       506   0.00%   0.00%   0.00% 100.00%       +7     +78    +143     27
7593       759   0.00%   0.00%   0.00% 100.00%      -54     +24    +144     47
11389     1138   0.00%   0.00%   0.00% 100.00%      -90     -33     +28     21
17083     1708   0.00%   0.00%   0.00% 100.00%      -50      -2     +35     14
25624     2562   0.00%   0.00%   0.00% 100.00%      -62      +7     +66     23
38436     3843   0.00%   0.00%   0.00% 100.00%      -33     +30   +5395     36
57654     5765   0.00%   0.00% 100.00%   0.00%
86481     8648   0.00%   0.00% 100.00%   0.00%
129721   12972   0.00%   0.00% 100.00%   0.00%
194581   16384  19.50%   0.00%  80.50%   0.00%
291871   16384  35.81%   0.00%  64.19%   0.00%
437806   16384  55.40%   0.00%  44.60%   0.00%

[1] https://github.com/mlichvar/ntpperf

Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Reviewed-by: Kurt Kanzenbach <kurt@linutronix.de>
Tested-by: Naama Meir <naamax.meir@linux.intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
---
 drivers/net/ethernet/intel/igc/igc.h      |  2 +-
 drivers/net/ethernet/intel/igc/igc_main.c |  2 +-
 drivers/net/ethernet/intel/igc/igc_ptp.c  | 15 +++++----------
 3 files changed, 7 insertions(+), 12 deletions(-)

diff --git a/drivers/net/ethernet/intel/igc/igc.h b/drivers/net/ethernet/intel/igc/igc.h
index 66fb67c17e4f..a00738bf6b19 100644
--- a/drivers/net/ethernet/intel/igc/igc.h
+++ b/drivers/net/ethernet/intel/igc/igc.h
@@ -228,7 +228,6 @@ struct igc_adapter {
 
 	struct ptp_clock *ptp_clock;
 	struct ptp_clock_info ptp_caps;
-	struct work_struct ptp_tx_work;
 	/* Access to ptp_tx_skb and ptp_tx_start are protected by the
 	 * ptp_tx_lock.
 	 */
@@ -638,6 +637,7 @@ int igc_ptp_set_ts_config(struct net_device *netdev, struct ifreq *ifr);
 int igc_ptp_get_ts_config(struct net_device *netdev, struct ifreq *ifr);
 void igc_ptp_tx_hang(struct igc_adapter *adapter);
 void igc_ptp_read(struct igc_adapter *adapter, struct timespec64 *ts);
+void igc_ptp_tx_tstamp_event(struct igc_adapter *adapter);
 
 #define igc_rx_pg_size(_ring) (PAGE_SIZE << igc_rx_pg_order(_ring))
 
diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
index eb50bfd5b867..eb4f0e562f60 100644
--- a/drivers/net/ethernet/intel/igc/igc_main.c
+++ b/drivers/net/ethernet/intel/igc/igc_main.c
@@ -5223,7 +5223,7 @@ static void igc_tsync_interrupt(struct igc_adapter *adapter)
 
 	if (tsicr & IGC_TSICR_TXTS) {
 		/* retrieve hardware timestamp */
-		schedule_work(&adapter->ptp_tx_work);
+		igc_ptp_tx_tstamp_event(adapter);
 		ack |= IGC_TSICR_TXTS;
 	}
 
diff --git a/drivers/net/ethernet/intel/igc/igc_ptp.c b/drivers/net/ethernet/intel/igc/igc_ptp.c
index 42f622ceb64b..cf963a12a92f 100644
--- a/drivers/net/ethernet/intel/igc/igc_ptp.c
+++ b/drivers/net/ethernet/intel/igc/igc_ptp.c
@@ -540,8 +540,6 @@ static void igc_ptp_clear_tx_tstamp(struct igc_adapter *adapter)
 {
 	unsigned long flags;
 
-	cancel_work_sync(&adapter->ptp_tx_work);
-
 	spin_lock_irqsave(&adapter->ptp_tx_lock, flags);
 
 	dev_kfree_skb_any(adapter->ptp_tx_skb);
@@ -724,16 +722,14 @@ static void igc_ptp_tx_hwtstamp(struct igc_adapter *adapter)
 }
 
 /**
- * igc_ptp_tx_work
- * @work: pointer to work struct
+ * igc_ptp_tx_tstamp_event
+ * @adapter: board private structure
  *
- * This work function checks the TSYNCTXCTL valid bit to determine when
- * a timestamp has been taken for the current stored skb.
+ * Called when a TX timestamp interrupt happens to retrieve the
+ * timestamp and send it up to the socket.
  */
-static void igc_ptp_tx_work(struct work_struct *work)
+void igc_ptp_tx_tstamp_event(struct igc_adapter *adapter)
 {
-	struct igc_adapter *adapter = container_of(work, struct igc_adapter,
-						   ptp_tx_work);
 	struct igc_hw *hw = &adapter->hw;
 	unsigned long flags;
 	u32 tsynctxctl;
@@ -1004,7 +1000,6 @@ void igc_ptp_init(struct igc_adapter *adapter)
 
 	spin_lock_init(&adapter->ptp_tx_lock);
 	spin_lock_init(&adapter->tmreg_lock);
-	INIT_WORK(&adapter->ptp_tx_work, igc_ptp_tx_work);
 
 	adapter->tstamp_config.rx_filter = HWTSTAMP_FILTER_NONE;
 	adapter->tstamp_config.tx_type = HWTSTAMP_TX_OFF;
-- 
2.38.1

[PATCH net v2 4/4] igc: Work around HW bug causing missing timestamps

[Tony Nguyen]

From: Vinicius Costa Gomes <vinicius.gomes@intel.com>

There's an hardware issue that can cause missing timestamps. The bug
is that the interrupt is only cleared if the IGC_TXSTMPH_0 register is
read.

The bug can cause a race condition if a timestamp is captured at the
wrong time, and we will miss that timestamp. To reduce the time window
that the problem is able to happen, in case no timestamp was ready, we
read the "previous" value of the timestamp registers, and we compare
with the "current" one, if it didn't change we can be reasonably sure
that no timestamp was captured. If they are different, we use the new
value as the captured timestamp.

The HW bug is not easy to reproduce, got to reproduce it when smashing
the NIC with timestamping requests from multiple applications (e.g.
multiple ntpperf instances + ptp4l), after 10s of minutes.

This workaround has more impact when multiple timestamp registers are
used, and the IGC_TXSTMPH_0 register always need to be read, so the
interrupt is cleared.

Fixes: 2c344ae24501 ("igc: Add support for TX timestamping")
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Tested-by: Naama Meir <naamax.meir@linux.intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
---
 drivers/net/ethernet/intel/igc/igc_ptp.c | 48 ++++++++++++++++++------
 1 file changed, 37 insertions(+), 11 deletions(-)

diff --git a/drivers/net/ethernet/intel/igc/igc_ptp.c b/drivers/net/ethernet/intel/igc/igc_ptp.c
index cf963a12a92f..32ef112f8291 100644
--- a/drivers/net/ethernet/intel/igc/igc_ptp.c
+++ b/drivers/net/ethernet/intel/igc/igc_ptp.c
@@ -685,14 +685,49 @@ static void igc_ptp_tx_hwtstamp(struct igc_adapter *adapter)
 	struct sk_buff *skb = adapter->ptp_tx_skb;
 	struct skb_shared_hwtstamps shhwtstamps;
 	struct igc_hw *hw = &adapter->hw;
+	u32 tsynctxctl;
 	int adjust = 0;
 	u64 regval;
 
 	if (WARN_ON_ONCE(!skb))
 		return;
 
-	regval = rd32(IGC_TXSTMPL);
-	regval |= (u64)rd32(IGC_TXSTMPH) << 32;
+	tsynctxctl = rd32(IGC_TSYNCTXCTL);
+	tsynctxctl &= IGC_TSYNCTXCTL_TXTT_0;
+	if (tsynctxctl) {
+		regval = rd32(IGC_TXSTMPL);
+		regval |= (u64)rd32(IGC_TXSTMPH) << 32;
+	} else {
+		/* There's a bug in the hardware that could cause
+		 * missing interrupts for TX timestamping. The issue
+		 * is that for new interrupts to be triggered, the
+		 * IGC_TXSTMPH_0 register must be read.
+		 *
+		 * To avoid discarding a valid timestamp that just
+		 * happened at the "wrong" time, we need to confirm
+		 * that there was no timestamp captured, we do that by
+		 * assuming that no two timestamps in sequence have
+		 * the same nanosecond value.
+		 *
+		 * So, we read the "low" register, read the "high"
+		 * register (to latch a new timestamp) and read the
+		 * "low" register again, if "old" and "new" versions
+		 * of the "low" register are different, a valid
+		 * timestamp was captured, we can read the "high"
+		 * register again.
+		 */
+		u32 txstmpl_old, txstmpl_new;
+
+		txstmpl_old = rd32(IGC_TXSTMPL);
+		rd32(IGC_TXSTMPH);
+		txstmpl_new = rd32(IGC_TXSTMPL);
+
+		if (txstmpl_old == txstmpl_new)
+			return;
+
+		regval = txstmpl_new;
+		regval |= (u64)rd32(IGC_TXSTMPH) << 32;
+	}
 	if (igc_ptp_systim_to_hwtstamp(adapter, &shhwtstamps, regval))
 		return;
 
@@ -730,22 +765,13 @@ static void igc_ptp_tx_hwtstamp(struct igc_adapter *adapter)
  */
 void igc_ptp_tx_tstamp_event(struct igc_adapter *adapter)
 {
-	struct igc_hw *hw = &adapter->hw;
 	unsigned long flags;
-	u32 tsynctxctl;
 
 	spin_lock_irqsave(&adapter->ptp_tx_lock, flags);
 
 	if (!adapter->ptp_tx_skb)
 		goto unlock;
 
-	tsynctxctl = rd32(IGC_TSYNCTXCTL);
-	tsynctxctl &= IGC_TSYNCTXCTL_TXTT_0;
-	if (!tsynctxctl) {
-		WARN_ONCE(1, "Received a TSTAMP interrupt but no TSTAMP is ready.\n");
-		goto unlock;
-	}
-
 	igc_ptp_tx_hwtstamp(adapter);
 
 unlock:
-- 
2.38.1

Re: [PATCH net v2 0/4][pull request] igc: TX timestamping fixes

[patchwork-bot+netdevbpf]

Hello:

This series was applied to netdev/net.git (main)
by Tony Nguyen <anthony.l.nguyen@intel.com>:

On Thu, 22 Jun 2023 09:52:40 -0700 you wrote:
> This is the fixes part of the series intended to add support for using
> the 4 timestamp registers present in i225/i226.
> 
> Moving the timestamp handling to be inline with the interrupt handling
> has the advantage of improving the TX timestamping retrieval latency,
> here are some numbers using ntpperf:
> 
> [...]

Here is the summary with links:
  - [net,v2,1/4] igc: Fix race condition in PTP tx code
    https://git.kernel.org/netdev/net/c/9c50e2b150c8
  - [net,v2,2/4] igc: Check if hardware TX timestamping is enabled earlier
    https://git.kernel.org/netdev/net/c/ce58c7cc8b99
  - [net,v2,3/4] igc: Retrieve TX timestamp during interrupt handling
    https://git.kernel.org/netdev/net/c/afa141583d82
  - [net,v2,4/4] igc: Work around HW bug causing missing timestamps
    https://git.kernel.org/netdev/net/c/c789ad7cbebc

You are awesome, thank you!
-- 
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html