[1/2] arm64: fix unnecessary tlb flushes

Commit Message

Mark Salter May 2, 2014, 3:20 p.m. UTC

The __cpu_flush_user_tlb_range() and __cpu_flush_user_tlb_range()
functions loop through an address range by page to flush tlb entries.
However, these functions assume a 4K page size. If the kernel is
configured for 64k page sizes, these functions would execute the
tlbi instruction 16 times per page rather than once. This patch
uses the PAGE_SHIFT definition to ensure one tlb flush for any
given page in the range.

Signed-off-by: Mark Salter <[e-mail]>
---
 arch/arm64/mm/tlb.S | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

Patch

diff --git a/arch/arm64/mm/tlb.S b/arch/arm64/mm/tlb.S
index 19da91e..b818073 100644
--- a/arch/arm64/mm/tlb.S
+++ b/arch/arm64/mm/tlb.S
@@ -42,7 +42,7 @@  ENTRY(__cpu_flush_user_tlb_range)
 	bfi	x0, x3, #48, #16		// start VA and ASID
 	bfi	x1, x3, #48, #16		// end VA and ASID
 1:	tlbi	vae1is, x0			// TLB invalidate by address and ASID
-	add	x0, x0, #1
+	add	x0, x0, #(1 << (PAGE_SHIFT - 12))
 	cmp	x0, x1
 	b.lo	1b
 	dsb	sy
@@ -62,7 +62,7 @@  ENTRY(__cpu_flush_kern_tlb_range)
 	lsr	x0, x0, #12			// align address
 	lsr	x1, x1, #12
 1:	tlbi	vaae1is, x0			// TLB invalidate by address
-	add	x0, x0, #1
+	add	x0, x0, #(1 << (PAGE_SHIFT - 12))
 	cmp	x0, x1
 	b.lo	1b
 	dsb	sy