From: Andrea Arcangeli <andrea@suse.de>
Subject: Race condition in userspace testcase
References: 46948, LTC11574


Additional Comment #103 From Andrea Arcangeli 2004-10-15 19:41
the last patch I attached is the safest I believe.

I'm not sure if a lock_unlock or lock_unlock is always guaranteed to happen
after the I/O, and this makes sure no race can happen anymore.

---
 fs/bio.c |   10 ++++++++++
 1 file changed, 10 insertions(+)

--- a/fs/bio.c
+++ b/fs/bio.c
@@ -1192,6 +1192,16 @@ void bio_endio(struct bio *bio, int erro
 	else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 		error = -EIO;
 
+	if (bio_data_dir(bio) == READ)
+		/*
+		 * If the current cpu has written to the page by hand
+		 * without dma, we must enforce ordering to be sure
+		 * this written data will be visible before we expose
+		 * the page contents to other cpus (for example with
+		 * a set_pte).
+		 */
+		smp_wmb();
+
 	if (bio->bi_end_io)
 		bio->bi_end_io(bio, error);
 }