powerpc/64s: Fix local irq disable when PMIs are disabled

When PMI interrupts are soft-masked, local_irq_save() will clear the PMI
mask bit, allowing PMIs in and causing a race condition. This causes a
deadlock in native_hpte_insert via hash_preload, which depends on PMIs
being disabled since commit 8b91cee ("powerpc/64s/hash: Make hash
faults work in NMI context"). native_hpte_insert calls local_irq_save().
It's possible the lpar hash code is also affected when tracing is
enabled because __trace_hcall_entry() calls local_irq_save().

Fix this by making arch_local_irq_save() _or_ the IRQS_DISABLED bit into
the mask.

This was found with the stress_hpt option with a kbuild workload running
together with `perf record -g`.

Fixes: f442d00 ("powerpc/64s: Add support to mask perf interrupts and replay them")
Fixes: 8b91cee ("powerpc/64s/hash: Make hash faults work in NMI context")
Signed-off-by: Nicholas Piggin <[email protected]>
[mpe: Just take the fix without the new warning]
Signed-off-by: Michael Ellerman <[email protected]>
Link: https://lore.kernel.org/r/[email protected]

Author: npiggin

arch/powerpc/include/asm/hw_irq.h

@@ -192,7 +192,7 @@ static inline void arch_local_irq_enable(void)
 
 static inline unsigned long arch_local_irq_save(void)
 {
-	return irq_soft_mask_set_return(IRQS_DISABLED);
+	return irq_soft_mask_or_return(IRQS_DISABLED);
 }
 
 static inline bool arch_irqs_disabled_flags(unsigned long flags)