> @Daiwei Li, I don't have a 82580 handy, please confirm that the patch fixes the issue you are having.
Thank you for the patch! I can confirm it fixes my issue. Below I offer a patch that also works in response to Paul's feedback. > Please also add a description of the test case I am running ptp4l to serve PTP to a client device attached to the NIC. To test, I am rebuilding igb.ko and reloading it. Without this patch, I see repeatedly in the output of ptp4l: > timed out while polling for tx timestamp increasing tx_timestamp_timeout or > increasing kworker priority may correct this issue, but a driver bug likely > causes it as well as my client device failing to sync time. > and maybe the PCI vendor and device code of your network device. % lspci -nn | grep Network 17:00.0 Ethernet controller [0200]: Intel Corporation 82580 Gigabit Network Connection [8086:150e] (rev 01) 17:00.1 Ethernet controller [0200]: Intel Corporation 82580 Gigabit Network Connection [8086:150e] (rev 01) 17:00.2 Ethernet controller [0200]: Intel Corporation 82580 Gigabit Network Connection [8086:150e] (rev 01) 17:00.3 Ethernet controller [0200]: Intel Corporation 82580 Gigabit Network Connection [8086:150e] (rev 01) > Bug, or was it a feature? According to https://lore.kernel.org/all/cdcb8be0.1ec2c%25matthew.v...@intel.com/ it was a bug. It looks like the datasheet was not updated to acknowledge this bug: https://www.intel.com/content/www/us/en/content-details/333167/intel-82580-eb-82580-db-gbe-controller-datasheet.html (section 8.17.28.1). > Is there a nicer way to write this, so `ack` is only assigned in case > for the 82580? diff --git a/drivers/net/ethernet/intel/igb/igb_main.c b/drivers/net/ethernet/intel/igb/igb_main.c index ada42ba63549..87ec1258e22a 100644 --- a/drivers/net/ethernet/intel/igb/igb_main.c +++ b/drivers/net/ethernet/intel/igb/igb_main.c @@ -6986,6 +6986,10 @@ static void igb_tsync_interrupt(struct igb_adapter *adapter) struct e1000_hw *hw = &adapter->hw; u32 tsicr = rd32(E1000_TSICR); struct ptp_clock_event event; + const u32 mask = (TSINTR_SYS_WRAP | E1000_TSICR_TXTS | + TSINTR_TT0 | TSINTR_TT1 | + TSINTR_AUTT0 | TSINTR_AUTT1); + if (tsicr & TSINTR_SYS_WRAP) { event.type = PTP_CLOCK_PPS; @@ -7009,6 +7013,13 @@ static void igb_tsync_interrupt(struct igb_adapter *adapter) if (tsicr & TSINTR_AUTT1) igb_extts(adapter, 1); + + if (hw->mac.type == e1000_82580) { + /* 82580 has a hardware bug that requires a explicit + * write to clear the TimeSync interrupt cause. + */ + wr32(E1000_TSICR, tsicr & mask); + } } On Fri, Aug 9, 2024 at 10:04 PM Richard Cochran <richardcoch...@gmail.com> wrote: > > On Fri, Aug 09, 2024 at 05:23:02PM -0700, Vinicius Costa Gomes wrote: > > It was reported that 82580 NICs have a hardware bug that makes it > > necessary to write into the TSICR (TimeSync Interrupt Cause) register > > to clear it. > > Bug, or was it a feature? > > Or IOW, maybe i210 changed the semantics of the TSICR? > > And what about the 82576? > > Thanks, > Richard