[IMPACT] The CPU x doesn't have a chance to exit the xgene_msi_cascade function to unlock desc->lock before Linux scheduce and call xgene_msi_set_affinity (irqbalance is caller) in the same CPU x. In irq_set_affinity, call raw_spin_lock_irqsave(&desc->lock, flags) which cause deadlock to CPU x because it disables preempt.
Use chained_irq_enter and exit as the standard way to cascade interrupt functions. [TEST CASE] Turn off irqbalance Run a single tcp stream Randomly change the affinity of the receiving ring: @ ethtool -S $INTF | grep rx[0-9].*_pac @ - to detect the rx ring @ grep $INTF /proc/interrupts@ - to find it's interrupt @ printf "%x" $(( 2 ** $((RANDOM % 8)) )) > /proc/irq/$IRQ/smp_affinity @ - to change the affinity [Regression Potential] Fix specific to the xgene pci msi code. -- You received this bug notification because you are a member of Kernel Packages, which is subscribed to linux in Ubuntu. https://bugs.launchpad.net/bugs/1359514 Title: pci-xgene-msi: fixed deadlock in irq_set_affinity Status in “linux” package in Ubuntu: Confirmed Bug description: Issue: CPU affinity is changed while irqbalance is running. + Problem explanation: - Old code + Call xgene_msi_cascade function (CPU x) + raw_spin_lock(&desc->lock); (CPU x) + Goto generic_handle_irq (CPU x) + The CPU x doesn't have a chance to exit the xgene_msi_cascade function to unlock desc->lock before Linux scheduce and call xgene_msi_set_affinity (irqbalance is caller) in the same CPU x + In irq_set_affinity, call raw_spin_lock_irqsave(&desc->lock, flags) which cause deadlock to CPU x because it disables preempt - New code + Use chained_irq_enter and exit as the standard way to cascade interrupt functions. To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1359514/+subscriptions -- Mailing list: https://launchpad.net/~kernel-packages Post to : [email protected] Unsubscribe : https://launchpad.net/~kernel-packages More help : https://help.launchpad.net/ListHelp
