This is a long standing problem (or a surprising feature) in our implementation
of get_user_pages() (used by direct IO). Since several attempts to fix it
failed (e.g.
http://linux.derkeiler.com/Mailing-Lists/Kernel/2009-04/msg06542.html, or
http://lkml.indiana.edu/hypermail/linux/kernel/0903.1/01498.html refused in
http://comments.gmane.org/gmane.linux.kernel.mm/31569) and it's not completely
clear whether we really want to fix it given the costs, let's at least document
it.

CC: mgorman@suse.de
CC: Jeff Moyer <redacted>
Signed-off-by: Jan Kara <jack@suse.cz>
---

--- a/man2/open.2       2012-04-27 00:07:51.736883092 +0200
+++ b/man2/open.2       2012-04-27 00:29:59.489892980 +0200

@@ -769,7 +769,12 @@

 and the file offset must all be multiples of the logical block size
 of the file system.
 Under Linux 2.6, alignment to 512-byte boundaries
-suffices.
+suffices. However, if the user buffer is not page aligned and direct read
+runs in parallel with a
+.BR fork (2)
+of the reader process, it may happen that the read data is split between
+pages owned by the original process and its child. Thus effectively read
+data is corrupted.
 .LP
 The
 .B O_DIRECT