It was [PATCHv2 1/5] arm64/entry-common: push the judgement of nmi ahead [1], but as discussion, I found it hit a deadlock, so change subject in v3. Linux kernel places strict semantics between NMI and maskable interrupt. So does the RCU component, else deadlock may happen. The current arm64 entry code can partially breach this rule through calling rcu_nmi_enter() in pNMI conext without calling __nmi_enter(). Finally, it calls __rcu_irq_enter_check_tick() trying to hold raw_spin_lock_rcu_node(rdp->mynode), but this spinlock can already been held by a task running note_gp_changes()->raw_spin_trylock_rcu_node(rnp). Then deadlock happens. The organization of this series: [1/4] aims to fix this bug [2-3/4] ease the implementation of [1/4] [4/4] is a cleanup To do: Does arm need to handle pNMI as arm64? v2 -> v3: drop handle_arch_nmi interface [1]: https://lore.kernel.org/linux-arm-kernel/20210924132837.45994-2-kernelfans@xxxxxxxxx/ Cc: Mark Rutland <mark.rutland@xxxxxxx> Cc: Paul E. McKenney <paulmck@xxxxxxxxxx> Cc: Catalin Marinas <catalin.marinas@xxxxxxx> Cc: Will Deacon <will@xxxxxxxxxx> Cc: Marc Zyngier <maz@xxxxxxxxxx> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx> Cc: Joey Gouly <joey.gouly@xxxxxxx> Cc: Sami Tolvanen <samitolvanen@xxxxxxxxxx> Cc: Julien Thierry <julien.thierry@xxxxxxx> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx> Cc: Yuichi Ito <ito-yuichi@xxxxxxxxxxx> Cc: rcu@xxxxxxxxxxxxxxx To: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx Pingfan Liu (4): arm64: entry: judge nmi ealier to avoid deadlock in RCU arm64: entry: distinguish pNMI earlier in el0 interrupt irqchip: GICv3: expose pNMI discriminator arm64: entry: remove pNMI judgement in __el1_interrupt() path arch/arm64/kernel/entry-common.c | 55 +++++++++++++++++++++----------- arch/arm64/kernel/irq.c | 15 +++++++++ drivers/irqchip/irq-gic-v3.c | 16 ++++++---- 3 files changed, 62 insertions(+), 24 deletions(-) -- 2.31.1
