Switching mm and tinkering with current->active_mm should be done with
irqs disabled. There is a path where exit_lazy_flush_tlb can be called
with irqs enabled:
exit_lazy_flush_tlb
flush_type_needed
__flush_all_mm
tlb_finish_mmu
exit_mmap
Which results in the switching being incorrectly with irqs enabled.
Reported-by: Sachin Sant
Link:
https://lore.kernel.org/linuxppc-dev/87a5xcgopc.fsf@mail.lhotse/T/#m105488939d0cd9f980978ed2fdeeb89bf731e673
Fixes: a665eec0a22e1 ("powerpc/64s/radix: Fix mm_cpumask trimming race vs
kthread_use_mm")
Signed-off-by: Nicholas Piggin
---
This sounds worse than it probably is, radix can likely tolerate an
interrupt hitting in mm switch, and the active_mm update may not be racy
in practice either. Still be good to backport it because I'm not 100%
sure of that.
This path can be stressed by reducing tlb_mm_cpumask_trim_timer (e.g.,
to 3).
Thanks,
Nick
arch/powerpc/mm/book3s64/radix_tlb.c | 10 +-
1 file changed, 9 insertions(+), 1 deletion(-)
diff --git a/arch/powerpc/mm/book3s64/radix_tlb.c
b/arch/powerpc/mm/book3s64/radix_tlb.c
index ce804b7bf84e..0bd4866d9824 100644
--- a/arch/powerpc/mm/book3s64/radix_tlb.c
+++ b/arch/powerpc/mm/book3s64/radix_tlb.c
@@ -795,12 +795,20 @@ void exit_lazy_flush_tlb(struct mm_struct *mm, bool
always_flush)
goto out;
if (current->active_mm == mm) {
+ unsigned long flags;
+
WARN_ON_ONCE(current->mm != NULL);
- /* Is a kernel thread and is using mm as the lazy tlb */
+ /*
+* It is a kernel thread and is using mm as the lazy tlb, so
+* switch it to init_mm. This is not always called from IPI
+* (e.g., flush_type_needed), so must disable irqs.
+*/
+ local_irq_save(flags);
mmgrab_lazy_tlb(_mm);
current->active_mm = _mm;
switch_mm_irqs_off(mm, _mm, current);
mmdrop_lazy_tlb(mm);
+ local_irq_restore(flags);
}
/*
--
2.40.1