[PATCH] powerpc/64s: Fix lost pending interrupt due to race causing lost update to irq_happened

force_external_irq_replay() can be called in the do_IRQ path with
interrupts hard enabled and soft disabled if may_hard_irq_enable() set
MSR[EE]=1. It updates local_paca->irq_happened with a load, modify,
store sequence. If a maskable interrupt hits during this sequence, it
will go to the masked handler to be marked pending in irq_happened.
This update will be lost when the interrupt returns and the store
instruction executes.  This can result in unpredictable latencies,
timeouts, lockups, etc.

Fix this by ensuring hard interrupts are disabled before modifying
irq_happened.

This could cause any maskable asynchronous interrupt to get lost, but
it was noticed on P9 SMP system doing RDMA NVMe target over 100GbE,
so very high external interrupt rate and high IPI rate. The hang was
bisected down to enabling doorbell interrupts for IPIs. These provided
an interrupt type that could run at high rates in the do_IRQ path,
stressing the race.

Fixes: 1d607bb3bd ("powerpc/irq: Add mechanism to force a replay of interrupts")
Reported-by: Carol L. Soto <redacted>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---

This has survived stress testing quite well so far, may need a little
more testing but I'd like to post it now to get some more comments.

We can optimise the mtmsr a bit more (e.g., skip it if interrupts are
already disabled or EE alrady set), but I've got some other patches
pending which change things there slightly, so I prefer to have this
minimal fix now, then make such changes upstream later.

---
 arch/powerpc/kernel/irq.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/arch/powerpc/kernel/irq.c b/arch/powerpc/kernel/irq.c
index f88038847790..061aa0f47bb1 100644
--- a/arch/powerpc/kernel/irq.c
+++ b/arch/powerpc/kernel/irq.c

@@ -476,6 +476,14 @@ void force_external_irq_replay(void)
 	 */
 	WARN_ON(!arch_irqs_disabled());
 
+	/*
+	 * Interrupts must always be hard disabled before irq_happened is
+	 * modified (to prevent lost update in case of interrupt between
+	 * load and store).
+	 */
+	__hard_irq_disable();
+	local_paca->irq_happened |= PACA_IRQ_HARD_DIS;
+
 	/* Indicate in the PACA that we have an interrupt to replay */
 	local_paca->irq_happened |= PACA_IRQ_EE;
 }

-- 
2.16.1

On Wed, 2018-03-21 at 12:22 +1000, Nicholas Piggin wrote:

force_external_irq_replay() can be called in the do_IRQ path with
interrupts hard enabled and soft disabled if may_hard_irq_enable() set
MSR[EE]=1. It updates local_paca->irq_happened with a load, modify,
store sequence. If a maskable interrupt hits during this sequence, it
will go to the masked handler to be marked pending in irq_happened.
This update will be lost when the interrupt returns and the store
instruction executes.  This can result in unpredictable latencies,
timeouts, lockups, etc.

Fix this by ensuring hard interrupts are disabled before modifying
irq_happened.

This could cause any maskable asynchronous interrupt to get lost, but
it was noticed on P9 SMP system doing RDMA NVMe target over 100GbE,
so very high external interrupt rate and high IPI rate. The hang was
bisected down to enabling doorbell interrupts for IPIs. These provided
an interrupt type that could run at high rates in the do_IRQ path,
stressing the race.

Fixes: 1d607bb3bd ("powerpc/irq: Add mechanism to force a replay of interrupts")
Reported-by: Carol L. Soto <redacted>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---

Nice one. We need that back into the distros asap.

Adding Carol...

On Wed, 21 Mar 2018 13:32:28 +1100
Benjamin Herrenschmidt [off-list ref] wrote:

On Wed, 2018-03-21 at 12:22 +1000, Nicholas Piggin wrote:

quoted

force_external_irq_replay() can be called in the do_IRQ path with
interrupts hard enabled and soft disabled if may_hard_irq_enable() set
MSR[EE]=1. It updates local_paca->irq_happened with a load, modify,
store sequence. If a maskable interrupt hits during this sequence, it
will go to the masked handler to be marked pending in irq_happened.
This update will be lost when the interrupt returns and the store
instruction executes.  This can result in unpredictable latencies,
timeouts, lockups, etc.

Fix this by ensuring hard interrupts are disabled before modifying
irq_happened.

This could cause any maskable asynchronous interrupt to get lost, but
it was noticed on P9 SMP system doing RDMA NVMe target over 100GbE,
so very high external interrupt rate and high IPI rate. The hang was
bisected down to enabling doorbell interrupts for IPIs. These provided
an interrupt type that could run at high rates in the do_IRQ path,
stressing the race.

Fixes: 1d607bb3bd ("powerpc/irq: Add mechanism to force a replay of interrupts")
Reported-by: Carol L. Soto <redacted>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---  

Nice one. We need that back into the distros asap.

Yes Carol is working on this with distros. Backporting should be
simple, so hoping to get this upstream soon to help the process.
We can take that as a Reviewed-by:? :)

Thanks,
Nick

On Wed, 2018-03-21 at 13:16 +1000, Nicholas Piggin wrote:

Yes Carol is working on this with distros. Backporting should be
simple, so hoping to get this upstream soon to help the process.
We can take that as a Reviewed-by:? :)

Ah yup.

Cheers,
Ben.

On Wed, 2018-03-21 at 02:22:28 UTC, Nicholas Piggin wrote:

force_external_irq_replay() can be called in the do_IRQ path with
interrupts hard enabled and soft disabled if may_hard_irq_enable() set
MSR[EE]=1. It updates local_paca->irq_happened with a load, modify,
store sequence. If a maskable interrupt hits during this sequence, it
will go to the masked handler to be marked pending in irq_happened.
This update will be lost when the interrupt returns and the store
instruction executes.  This can result in unpredictable latencies,
timeouts, lockups, etc.

Fix this by ensuring hard interrupts are disabled before modifying
irq_happened.

This could cause any maskable asynchronous interrupt to get lost, but
it was noticed on P9 SMP system doing RDMA NVMe target over 100GbE,
so very high external interrupt rate and high IPI rate. The hang was
bisected down to enabling doorbell interrupts for IPIs. These provided
an interrupt type that could run at high rates in the do_IRQ path,
stressing the race.

Fixes: 1d607bb3bd ("powerpc/irq: Add mechanism to force a replay of interrupts")
Reported-by: Carol L. Soto <redacted>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>

Applied to powerpc fixes, thanks.

https://git.kernel.org/powerpc/c/ff6781fd1bb404d8a551c02c35c70c

cheers