[PATCH 0/2] irqchip/gic-v3: Support pseudo-NMIs when SCR_EL3.FIQ == 0

Alexandru Elisei alexandru.elisei at arm.com
Thu Jun 25 11:00:24 EDT 2020

Trusted Firmware-A's default interrupt routing model is to clear
SCR_EL3.FIQ, which is the only case that GICv3 doesn't support. This series
tries to fix that by detecting it at runtime and using a different priority
value for ICC_PMR_EL1 when masking regular interrupts. As a result, we will
be able to support pseudo-NMIs on all combinations of GIC security states
and firmware configurations.

The series is based on v5.8-rc1, the same as the PMU NMI patches [1] which
I used for testing.

The first patch was there because when I started working on the PMU NMI
patches I found it confusing that there was no clear message stating that
NMIs were successfully enabled. The second patch is the main patch of the
series, where pseudo-NMIs are enabled even if SCR_EL3.FIQ == 0.

The are still some things I'm not 100% sure about regarding the last patch:

- From my very limited experience of trying pseudo-NMIs on 3 machines
  (rockpro64, espressobin-v5 and v7), all of them had SCR_EL3.FIQ zero.
  I tend to believe that since this is the default on TFA, this will also
  be the common case on hardware. However, when Linux is a KVM guest, the
  original set of priorities is used because GIC security is disabled.
  I erred on the side of caution and chose the original set of priorities
  as the common case.

- Most of the changes to arch_local_irq_enable() might seem pointless, but
  I chose to make them so the function looks similar to
  arch_local_irq_disable(). The generated code is identical if the static
  branch is not taken. I tried changing only arch_local_irq_disable(), but
  the enable function ended up looking strangely asymmetrical. If someone
  suggests a better way of doing things, I'd be happy to implement it.

As an aside, the set of priorities that I've added would work in all cases
if there was no need to use GIC_PRIO_PSR_I_SET, but that is a much more
intrusive change and I'm not comfortable attempting it. I'm pretty sure I
will end up breaking things really badly.

I've tested the series using PMU NMIs [1] on the model and on

1. On the model:
- Host with SCR_EL3.FIQ == 1 (so using the original priorities), ran
  perf record -a -- iperf3 -c -t 30.
- On a KVM guest (security disabled, so using the original priorities),
  ran the same command as above.

2. On an espressobin-v7:
- Host with SCR_EL3.FIQ == 0 (using the priority added by the series),
  ran perf record -a -- iperf3 -c -t 60.
- On a KVM guest (security disabled, so using the original priorities),
  ran the same command.
- Stress test for an hour with CONFIG_ARM64_DEBUG_PRIORITY_MASKING set for
  the host and guest. On the host, I ran in parallel
  perf record -a -- iperf3 -c -t 3600  and
  perf record -ae L1-dcache-loads -a -- sleep 3600. On the guest, I
  ran the same iperf3 command as on the host.

[1] https://www.spinics.net/lists/kernel/msg3554236.html

Alexandru Elisei (2):
  irqchip/gicv3: Spell out when pseudo-NMIs are enabled
  irqchip/gic-v3: Support pseudo-NMIs when SCR_EL3.FIQ == 0

 arch/arm64/include/asm/arch_gicv3.h |  8 ++++-
 arch/arm64/include/asm/daifflags.h  |  4 +--
 arch/arm64/include/asm/irqflags.h   | 14 +++++---
 arch/arm64/include/asm/ptrace.h     | 12 +++++++
 arch/arm64/kernel/entry.S           |  2 +-
 arch/arm64/kvm/hyp/switch.c         |  2 +-
 drivers/irqchip/irq-gic-v3.c        | 56 +++++++++++++++++++++--------
 7 files changed, 75 insertions(+), 23 deletions(-)


More information about the linux-arm-kernel mailing list