On 09/08/2010 09:47 PM, Andrew Morton wrote:

On Tue,  7 Sep 2010 13:45:34 -0700
Gary King [off-list ref] wrote:

quoted

I have been seeing problems on Tegra 2 (ARMv7 SMP) systems with HIGHMEM
enabled on 2.6.35 (plus some patches targetted at 2.6.36 to perform
cache maintenance lazily), and the root cause appears to be that the
mm bouncing code is calling flush_dcache_page before it copies the
bounce buffer into the bio.

The patch below reorders these two operations, and eliminates numerous
arbitrary application crashes on my dev system.

Gary

--

quoted

From 678c9bca8d8a8f254f28af91e69fad3aa1be7593 Mon Sep 17 00:00:00 2001

From: Gary King <redacted>
Date: Mon, 6 Sep 2010 15:37:12 -0700
Subject: bounce: call flush_dcache_page after bounce_copy_vec

the bounced page needs to be flushed after data is copied into it,
to ensure that architecture implementations can synchronize
instruction and data caches if necessary.

Signed-off-by: Gary King <redacted>
---
 mm/bounce.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/mm/bounce.c b/mm/bounce.c
index 13b6dad..1481de6 100644
--- a/mm/bounce.c
+++ b/mm/bounce.c

@@ -116,8 +116,8 @@ static void copy_to_high_bio_irq(struct bio *to, struct bio *from)
 		 */
 		vfrom = page_address(fromvec->bv_page) + tovec->bv_offset;
 
-		flush_dcache_page(tovec->bv_page);
 		bounce_copy_vec(tovec, vfrom);
+		flush_dcache_page(tovec->bv_page);
 	}
 }

Oh my, that was bad.

Indeed, I wonder how that could have been wrong for so long.

I queued your fix for 2.6.36 and tagged it for -stable backporting,
thanks.

Thanks, you can add my acked-by.

-- 
Jens Axboe