Wander reported that performing a SR-IOV setup on PREEMPT_RT can fail/ timeout. The reason is that during the setup the VF device performs a reset (igbvf_reset()) and polls for an ACK (e1000_reset_hw_vf() -> e1000_check_for_ack_vf()) with disabled bottom halves. For the ACK to complete it is required for the igb_msix_other() interrupt handler to run. The interrupt handler is forced-threaded on PREEMPT_RT and therefore delayed until after bottom halves are enabled again. This happens only after e1000_reset_hw_vf() times out. This scenario requires that the interrupt handler and the reset handler run on the same CPU. This scenario is not limited to PREEMPT_RT but can also happen without PREEMPT_RT if the interrupts are forced threaded via `threadirqs'. Setups without forced threaded interrupts are not affected.
The interrupt handler (igb_msix_other()) does not require bottom halves to be disabled. It does not call into the network stack which would mandate it. Requesting the handler explicit as a threaded interrupt will not disable bottom halves prior invocation of the handler thus avoiding the scenario. Request igb_msix_other as a threaded interrupt handler. Reported-by: Wander Lairson Costa <wan...@redhat.com> Closes: https://lore.kernel.org/all/20240920185918.616302-2-wan...@redhat.com/ Signed-off-by: Sebastian Andrzej Siewior <bige...@linutronix.de> --- I've been sitting on this one for a while. While this avoids the timeout on a PREEMPT_RT setup, the !PREEMPT_RT + threadirqs setup remains affected. The difference is that PREEMPT_RT allows a context switch within a local_bh_disable() section while !PREEMPT_RT does not. Allowing e1000_reset_hw_vf() to run/ wait/ poll without e1000_hw::mbx_lock, which disable BH, should fix both setups. drivers/net/ethernet/intel/igb/igb_main.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/drivers/net/ethernet/intel/igb/igb_main.c b/drivers/net/ethernet/intel/igb/igb_main.c index c646c71915f03..0827e8dcd9de7 100644 --- a/drivers/net/ethernet/intel/igb/igb_main.c +++ b/drivers/net/ethernet/intel/igb/igb_main.c @@ -912,8 +912,8 @@ static int igb_request_msix(struct igb_adapter *adapter) struct net_device *netdev = adapter->netdev; int i, err = 0, vector = 0, free_vector = 0; - err = request_irq(adapter->msix_entries[vector].vector, - igb_msix_other, 0, netdev->name, adapter); + err = request_threaded_irq(adapter->msix_entries[vector].vector, + igb_msix_other, 0, netdev->name, adapter); if (err) goto err_out; -- 2.49.0
