[src/trunk]: src/sys/arch/amd64/amd64 I actually came across the solution to ...

[maxv <maxv%NetBSD.org@localhost>]

details:   https://anonhg.NetBSD.org/src/rev/d5645748425c
branches:  trunk
changeset: 348913:d5645748425c
user:      maxv <maxv%NetBSD.org@localhost>
date:      Tue Nov 15 15:26:59 2016 +0000

description:
I actually came across the solution to this issue in the Intel SDM for a
totally unrelated reason a few weeks ago. The reason we need a particular
module_map on amd64 is because gcc makes us use RIP-relative addressing.

The offset field of the associated opcodes is a 32bit signed displacement,
which means we can access only up to 2GB around the current instruction.
And given that kernel_map is too far away from the kernel .text, it is not
RIP-addressable. Hence the module_map embedded into the bootstrap memory,
which is right above the kernel image.

diffstat:

 sys/arch/amd64/amd64/machdep.c |  16 +++++++++-------
 1 files changed, 9 insertions(+), 7 deletions(-)

diffs (38 lines):

diff -r 0cf97e625f5a -r d5645748425c sys/arch/amd64/amd64/machdep.c
--- a/sys/arch/amd64/amd64/machdep.c    Tue Nov 15 15:00:55 2016 +0000
+++ b/sys/arch/amd64/amd64/machdep.c    Tue Nov 15 15:26:59 2016 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: machdep.c,v 1.231 2016/11/15 15:00:56 maxv Exp $       */
+/*     $NetBSD: machdep.c,v 1.232 2016/11/15 15:26:59 maxv Exp $       */
 
 /*-
  * Copyright (c) 1996, 1997, 1998, 2000, 2006, 2007, 2008, 2011
@@ -111,7 +111,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: machdep.c,v 1.231 2016/11/15 15:00:56 maxv Exp $");
+__KERNEL_RCSID(0, "$NetBSD: machdep.c,v 1.232 2016/11/15 15:26:59 maxv Exp $");
 
 /* #define XENDEBUG_LOW  */
 
@@ -377,12 +377,14 @@
        /*
         * Create the module map.
         *
-        * XXX: the module map is taken as what is left of the bootstrap memory
-        * created in locore.S, which is not big enough if we want to load many
-        * modules dynamically. We really should be using kernel_map instead.
+        * The kernel uses RIP-relative addressing with a maximum offset of
+        * 2GB. The problem is, kernel_map is too far away in memory from
+        * the kernel .text. So we cannot use it, and have to create a
+        * special module_map.
         *
-        * But the modules must be located above the kernel image, and that
-        * wouldn't be guaranteed if we were using kernel_map.
+        * The module map is taken as what is left of the bootstrap memory
+        * created in locore.S. This memory is right above the kernel
+        * image, so this is the best place to put our modules.
         */
        uvm_map_setup(&module_map_store, module_start, module_end, 0);
        module_map_store.pmap = pmap_kernel();