When an IPI is generated by a CPU, the pattern looks roughly like:

  <write shared data>
  smp_wmb();
  <write to GIC to signal SGI>

On the receiving CPU we rely on the fact that, once we've taken the
interrupt, then the freshly written shared data must be visible to us.
Put another way, the CPU isn't going to speculate taking an interrupt.

Unfortunately, this assumption turns out to be broken.

Consider that CPUx wants to send an IPI to CPUy, which will cause CPUy
to read some shared_data. Before CPUx has done anything, a random
peripheral raises an IRQ to the GIC and the IRQ line on CPUy is raised.
CPUy then takes the IRQ and starts executing the entry code, heading
towards gic_handle_irq. Furthermore, let's assume that a bunch of the
previous interrupts handled by CPUy were SGIs, so the branch predictor
kicks in and speculates that irqnr will be <16 and we're likely to
head into handle_IPI. The prefetcher then grabs a speculative copy of
shared_data which contains a stale value.

Meanwhile, CPUx gets round to updating shared_data and asking the GIC
to send an SGI to CPUy. Internally, the GIC decides that the SGI is
more important than the peripheral interrupt (which hasn't yet been
ACKed) but doesn't need to do anything to CPUy, because the IRQ line
is already raised.

CPUy then reads the ACK register on the GIC, sees the SGI value which
confirms the branch prediction and we end up with a stale shared_data
value.

This patch fixes the problem by adding an smp_rmb() to the IPI entry
code in gic_handle_irq. As it turns out, the combination of a control
dependency and an ISB instruction from the EOI in the GICv3 driver is
enough to provide the ordering we need, so we add a comment there
justifying the absence of an explicit smp_rmb().

Cc: stable@vger.kernel.org
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>