arm: vexpress: tc2: fix hotplug/idle/kexec race on cluster power down
Subsystems (3)
PATCH v2 1/1 arm: vexpress: tc2: fix hotplug/idle/kexec race on cluster power down
On the TC2 testchip, when all CPUs in a cluster enter standbywfi and commit a power down request, the power controller will wait for standbywfil2 coming from L2 cache controller to shut the cluster down.
By the time all CPUs in a cluster commit a power down request and enter wfi, the power controller cannot backtrack, or put it another way, a CPU must not be allowed to complete execution independently of the power controller, the only way for it to resume properly must be upon wake-up IRQ pending and subsequent reset triggered from the power controller.
Current MCPM back-end for TC2 disables the GIC CPU IF only when power down is committed through the tc2_pm_suspend() method, that makes sense since a suspended CPU is still online and can receive interrupts whereas a hotplugged CPU, since it is offline, migrated all IRQs and shutdown the per-CPU peripherals, hence their PPIs.
The flaw with this reasoning is the following. If all CPUs in a clusters are entering a power down state either through CPU idle or CPU hotplug, when the last man successfully completes the MCPM power down sequence (and executes wfi), power controller waits for L2 wfi signal to quiesce the cluster and shut it down. If, when all CPUs are sitting in wfi, an online CPU hotplugs back in one of the CPUs in the cluster being shutdown, that CPU receives an IPI that causes wfi to complete (since tc2_pm_down() method does not disable the GIC CPU IF in that case - CPU being hotplugged out, not idle) and the power controller will never see the stanbywfil2 signal coming from L2 that is required for shutdown to happen and the system deadlocks.
Further to this issue, kexec hotplugs secondary CPUs out during kernel reload/restart.
Because kexec may (deliberately) trash the old kernel text, it is not OK for CPUs to follow the MCPM soft reboot path, since instructions after the WFI may have been replaced by kexec.
If tc2_pm_down() does not disable the GIC cpu interface, there is a race between CPU powerdown in the old kernel and the IPI from the new kernel that triggers secondary boot, particularly if the powerdown is slow (due to L2 cache cleaning for example). If the new kernel wins the race, the affected CPU(s) will not really be reset and may execute garbage after the WFI.
The only solution to this problem consists in disabling the GIC CPU IF on a CPU committed to power down regardless of the power down entry method (CPU hotplug or CPU idle). This way, CPU wake-up is under power controller control, which prevents unexpected wfi exit caused by a pending IRQ.
This patch moves the GIC CPU IF disable call in the TC2 MCPM implementation from the tc2_pm_suspend() method to the tc2_pm_down() method to fix the mentioned race condition(s).
--- arch/arm/mach-vexpress/tc2_pm.c | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-)
arch/arm/mach-vexpress/tc2_pm.c +10 -1
@@ -131,6 +131,16 @@ static void tc2_pm_down(u64 residency) }else BUG (); +/* +* If the CPU is committed to power down, make sure +* the power controller will be in charge of waking it +* up upon IRQ, ie IRQ lines are cut from GIC CPU IF +* to the CPU by disabling the GIC CPU IF to prevent wfi +* from completing execution behind power controller back +*/ +if (!skip_wfi) +gic_cpu_if_down (); +if (last_man &&__mcpm_outbound_enter_critical (cpu, cluster)) {arch_spin_unlock (&tc2_pm_lock);
@@ -231,7 +241,6 @@ static void tc2_pm_suspend(u64 residency) cpu =MPIDR_AFFINITY_LEVEL (mpidr,0 ); cluster =MPIDR_AFFINITY_LEVEL (mpidr,1 );ve_spc_set_resume_addr (cluster, cpu,virt_to_phys (mcpm_entry_point)); -gic_cpu_if_down ();tc2_pm_down (residency); }
Signature · 1 line
1.8.2.2
PATCH v2 1/1 arm: vexpress: tc2: fix hotplug/idle/kexec race on cluster power down
Quoted 50 lines · Lorenzo Pieralisi
On the TC2 testchip, when all CPUs in a cluster enter standbywfi and commit a power down request, the power controller will wait for standbywfil2 coming from L2 cache controller to shut the cluster down.
By the time all CPUs in a cluster commit a power down request and enter wfi, the power controller cannot backtrack, or put it another way, a CPU must not be allowed to complete execution independently of the power controller, the only way for it to resume properly must be upon wake-up IRQ pending and subsequent reset triggered from the power controller.Current MCPM back-end for TC2 disables the GIC CPU IF only when power down is committed through the
tc2_pm_suspend()method, that makes sense since a suspended CPU is still online and can receive interrupts whereas a hotplugged CPU, since it is offline, migrated all IRQs and shutdown the per-CPU peripherals, hence their PPIs.The flaw with this reasoning is the following. If all CPUs in a clusters are entering a power down state either through CPU idle or CPU hotplug, when the last man successfully completes the MCPM power down sequence (and executes wfi), power controller waits for L2 wfi signal to quiesce the cluster and shut it down. If, when all CPUs are sitting in wfi, an online CPU hotplugs back in one of the CPUs in the cluster being shutdown, that CPU receives an IPI that causes wfi to complete (since
tc2_pm_down()method does not disable the GIC CPU IF in that case - CPU being hotplugged out, not idle) and the power controller will never see the stanbywfil2 signal coming from L2 that is required for shutdown to happen and the system deadlocks.Further to this issue, kexec hotplugs secondary CPUs out during kernel reload/restart.
Because kexec may (deliberately) trash the old kernel text, it is not OK for CPUs to follow the MCPM soft reboot path, since instructions after the WFI may have been replaced by kexec.If
tc2_pm_down()does not disable the GIC cpu interface, there is a race between CPU powerdown in the old kernel and the IPI from the new kernel that triggers secondary boot, particularly if the powerdown is slow (due to L2 cache cleaning for example). If the new kernel wins the race, the affected CPU(s) will not really be reset and may execute garbage after the WFI.The only solution to this problem consists in disabling the GIC CPU IF on a CPU committed to power down regardless of the power down entry method (CPU hotplug or CPU idle). This way, CPU wake-up is under power controller control, which prevents unexpected wfi exit caused by a pending IRQ.
Quoted 8 lines
This patch moves the GIC CPU IF disable call in the TC2 MCPM implementation from the
tc2_pm_suspend()method to thetc2_pm_down()method to fix the mentioned race condition(s).Reviewed-by Dave Martin <[email protected]>Signed-off-by Lorenzo Pieralisi <[email protected]>Signed-off-by Sudeep KarkadaNagesha <[email protected]>
ARM-SOC dudes: would you please pick this up?
Quoted 37 lines
--- arch/arm/mach-vexpress/tc2_pm.c | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-)arch/arm/mach-vexpress/tc2_pm.c +10
-1@@ -131,6 +131,16 @@ static void tc2_pm_down(u64 residency) }else BUG (); +/* +* If the CPU is committed to power down, make sure +* the power controller will be in charge of waking it +* up upon IRQ, ie IRQ lines are cut from GIC CPU IF +* to the CPU by disabling the GIC CPU IF to prevent wfi +* from completing execution behind power controller back +*/ +if (!skip_wfi) +gic_cpu_if_down (); +if (last_man &&__mcpm_outbound_enter_critical (cpu, cluster)) {arch_spin_unlock (&tc2_pm_lock);@@ -231,7 +241,6 @@ static void tc2_pm_suspend(u64 residency) cpu =MPIDR_AFFINITY_LEVEL (mpidr,0 ); cluster =MPIDR_AFFINITY_LEVEL (mpidr,1 );ve_spc_set_resume_addr (cluster, cpu,virt_to_phys (mcpm_entry_point)); -gic_cpu_if_down ();tc2_pm_down (residency); }Signature · 1 line
1.8.2.2
PATCH v2 1/1 arm: vexpress: tc2: fix hotplug/idle/kexec race on cluster power down
Quoted 3 lines · Nicolas Pitre
Acked-by Nicolas Pitre <[email protected]>ARM-SOC dudes: would you please pick this up?
I can't tell if this is actually needed for 3.12, since kexec isn't supported without the set of patches that Dave Martin just posted.
Anyway, the fix is tiny so I applied it to the for-3.12 fixes branch anyway.
-Olof
PATCH v2 1/1 arm: vexpress: tc2: fix hotplug/idle/kexec race on cluster power down
Quoted 7 lines · Olof Johansson
Quoted 3 lines · Nicolas Pitre
Acked-by Nicolas Pitre <[email protected]>ARM-SOC dudes: would you please pick this up?
I can't tell if this is actually needed for 3.12, since kexec isn't supported without the set of patches that Dave Martin just posted.
I believe some problems have been seen when mixing hotplug with idle, even if kexec is out of the picture. This was a known issue before I started playing with kexec.
This should be considered a separate patch.
Cheers
---Dave
Quoted 10 lines
Anyway, the fix is tiny so I applied it to the for-3.12 fixes branch anyway.
-Olof
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel