We need to issue a DSB before doing TLB invalidation to make sure that the
table walker sees the new VA mapping after the TLBI finishes. For
flush_tlb_page, we do a DSB ISHST (synchronization barrier for writes in
the Inner Shareable domain) because translation table walks are now
coherent for arm. For local_flush_tlb_all, we only need to affect the
Non-shareable domain, and we do a DSB NSHST. We need a synchronization
barrier here, and not a memory ordering barrier, because a table walk is
not a memory operation and therefore not affected by the DMB.
For the same reasons, we downgrade the full system DSB after the TLBI to a
DSB ISH (synchronization barrier for reads and writes in the Inner
Shareable domain), and, respectively, DSB NSH (in the Non-shareable
domain).
With these two changes, our TLB maintenance functions now match what Linux
does in __flush_tlb_kernel_page, and, respectively, in local_flush_tlb_all.
A similar change was implemented in Linux commit 62cbbc42e001 ("ARM: tlb:
reduce scope of barrier domains for TLB invalidation").
Signed-off-by: Alexandru Elisei <alexandru.elisei@xxxxxxx>
---
lib/arm/asm/mmu.h | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/lib/arm/asm/mmu.h b/lib/arm/asm/mmu.h
index 361f3cdcc3d5..2bf8965ed35e 100644
--- a/lib/arm/asm/mmu.h
+++ b/lib/arm/asm/mmu.h
@@ -17,9 +17,10 @@
static inline void local_flush_tlb_all(void)
{
+ dsb(nshst);
/* TLBIALL */
asm volatile("mcr p15, 0, %0, c8, c7, 0" :: "r" (0));
- dsb();
+ dsb(nsh);
isb();
}
@@ -31,9 +32,10 @@ static inline void flush_tlb_all(void)
static inline void flush_tlb_page(unsigned long vaddr)
{
+ dsb(ishst);
/* TLBIMVAAIS */
asm volatile("mcr p15, 0, %0, c8, c3, 3" :: "r" (vaddr));
- dsb();
+ dsb(ish);
isb();
}
--
2.7.4
