From: Brian Norris Date: Thu, 12 Jun 2025 18:32:51 +0000 (-0700) Subject: genirq/cpuhotplug: Rebalance managed interrupts across multi-CPU hotplug X-Git-Tag: v6.16-rc3~3^2~3 X-Git-Url: http://git.ipfire.org/?a=commitdiff_plain;h=2b32fc8ff08deac3aa509f321a28e21b1eea5525;p=thirdparty%2Flinux.git genirq/cpuhotplug: Rebalance managed interrupts across multi-CPU hotplug Commit 788019eb559f ("genirq: Retain disable depth for managed interrupts across CPU hotplug") intended to only decrement the disable depth once per managed shutdown, but instead it decrements for each CPU hotplug in the affinity mask, until its depth reaches a point where it finally gets re-started. For example, consider: 1. Interrupt is affine to CPU {M,N} 2. disable_irq() -> depth is 1 3. CPU M goes offline -> interrupt migrates to CPU N / depth is still 1 4. CPU N goes offline -> irq_shutdown() / depth is 2 5. CPU N goes online -> irq_restore_affinity_of_irq() -> irqd_is_managed_and_shutdown()==true -> irq_startup_managed() -> depth is 1 6. CPU M goes online -> irq_restore_affinity_of_irq() -> irqd_is_managed_and_shutdown()==true -> irq_startup_managed() -> depth is 0 *** BUG: driver expects the interrupt is still disabled *** -> irq_startup() -> irqd_clr_managed_shutdown() 7. enable_irq() -> depth underflow / unbalanced enable_irq() warning This should clear the managed-shutdown flag at step 6, so that further hotplugs don't cause further imbalance. Note: It might be cleaner to also remove the irqd_clr_managed_shutdown() invocation from __irq_startup_managed(). But this is currently not possible because of irq_update_affinity_desc() as it sets IRQD_MANAGED_SHUTDOWN and expects irq_startup() to clear it. Fixes: 788019eb559f ("genirq: Retain disable depth for managed interrupts across CPU hotplug") Reported-by: Aleksandrs Vinarskis Signed-off-by: Brian Norris Signed-off-by: Thomas Gleixner Tested-by: Aleksandrs Vinarskis Link: https://lore.kernel.org/all/20250612183303.3433234-2-briannorris@chromium.org --- diff --git a/kernel/irq/chip.c b/kernel/irq/chip.c index b0e0a7332993c..2b274007e8bab 100644 --- a/kernel/irq/chip.c +++ b/kernel/irq/chip.c @@ -205,6 +205,14 @@ __irq_startup_managed(struct irq_desc *desc, const struct cpumask *aff, void irq_startup_managed(struct irq_desc *desc) { + struct irq_data *d = irq_desc_get_irq_data(desc); + + /* + * Clear managed-shutdown flag, so we don't repeat managed-startup for + * multiple hotplugs, and cause imbalanced disable depth. + */ + irqd_clr_managed_shutdown(d); + /* * Only start it up when the disable depth is 1, so that a disable, * hotunplug, hotplug sequence does not end up enabling it during