[RFC PATCH 2/2] zap_pte_range: fix partial TLB flushing in response to a dirty pte

[Will Deacon <will.deacon@arm.com>]

When we encounter a dirty page during unmap, we force a TLB invalidation
to avoid a race with pte_mkclean and stale, dirty TLB entries in the
CPU.

This uses the same force_flush logic as the batch failure code, but
since we don't break out of the loop when finding a dirty pte, tlb->end
can be < addr as we only batch for present ptes. This can result in a
negative range being passed to subsequent TLB invalidation calls,
potentially leading to massive over-invalidation of the TLB (observed
in practice running firefox on arm64).

This patch fixes the issue by restricting the use of addr in the TLB
range calculations. The first range then ends up covering tlb->start to
min(tlb->end, addr), which corresponds to the currently batched range.
The second range then covers anything remaining, which may still lead to
a (much reduced) over-invalidation of the TLB.

Signed-off-by: Will Deacon <will.deacon@arm.com>
---
 mm/memory.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/mm/memory.c b/mm/memory.c
index 3e503831e042..ea41508d41f3 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -1194,11 +1194,10 @@ again:
 		 * then update the range to be the remaining
 		 * TLB range.
 		 */
-		old_end = tlb->end;
-		tlb->end = addr;
+		tlb->end = old_end = min(tlb->end, addr);
 		tlb_flush_mmu_tlbonly(tlb);
-		tlb->start = addr;
-		tlb->end = old_end;
+		tlb->start = old_end;
+		tlb->end = end;
 	}
 	pte_unmap_unlock(start_pte, ptl);
 
-- 
2.1.1