mmap.2: Describe treatment of 'offset' for MAP_ANONYMOUS

The manpage claimed that the offset argument is ignored, and when
I interpreted that as "I don't need to set that register when
doing the syscall", I got failures. I was able to spot two reasons
for that:

What I probably ran into:
At least on x86-64, sys_mmap (in arch/x86/kernel/sys_x86_64.c)
always checks that the offset is page-aligned, even for
MAP_ANONYMOUS.

Another one, could probably trigger on 32-bit x86:
In do_mmap(), there is a check to ensure that pgoff together with
the allocation length won't cause an overflow, even for
MAP_ANONYMOUS.

Document that userspace should pass in zero, since that's
probably what everyone is doing already. (It would also be
possible to describe the constraints on the offset more
carefully, but zero works, and nobody should need to pass in
anything else.)

Signed-off-by: Jann Horn <jann@thejh.net>

man2/mmap.2

@@ -164,9 +164,7 @@ The mapping is not backed by any file;
 its contents are initialized to zero.
 The
 .I fd
-and
-.I offset
-arguments are ignored;
+argument is ignored;
 however, some implementations require
 .I fd
 to be \-1 if
@@ -175,6 +173,11 @@ to be \-1 if
 .BR MAP_ANON )
 is specified,
 and portable applications should ensure this.
+The
+.I offset
+argument should be zero.
+.\" See the pgoff overflow check in do_mmap().
+.\" See the offset check in sys_mmap in arch/x86/kernel/sys_x86_64.c.
 The use of
 .B MAP_ANONYMOUS
 in conjunction with