On Thu, 2014-08-28 at 11:00 -0700, Ben Hutchings wrote: > On Thu, 2014-08-28 at 09:38 -0300, Mauricio Faria de Oliveira wrote: > > Package: src:linux > > > > Hi Ben, > > > > On 08/26/2014 05:05 AM, Ben Hutchings wrote: > > > The first build failure (requiring __SANE_USERSPACE_TYPES__) appears to > > > be fixed upstream in 3.16. [...] > > > > I'm still hitting that w/ 3.16 (linux-tools from trunk, plus .orig > > from 3.16~rc7~exp1 or upstream 3.16.1 -- not usual, but I believe > > it was supposed to work). > > > > May you please point where/commit/excerpt you see this fixed? I could > > only find a related commit for MIPS [1], but nothing for powerpc. > > I thought this would do it: > > commit d944c4eebcf4c0d5e5d9728fec110cbf0047ad7f > Author: Borislav Petkov <[email protected]> > Date: Fri Apr 25 21:31:02 2014 +0200 > > tools: Consolidate types.h
OK, I forgot to apply this fix for 3.16-1. So coming back to this, I
tried to understand what __SANE_USERSPACE_TYPES__ has to do with
defining 'u32' (which the kernel headers should never define for
userland), and why we would include include/linux/types.h (rather than
the corresponding UAPI header) at all.
The answer is that devicetable-offsets.c is supposed to be built like
kernel code, but we're fudging it - we're also trying to make modpost
support cross-builds and not depend on any kernel configuration. As we
don't want it to depend on the host architecture, our search path is
just:
include
So the chain of included headers is:
<linux/types.h> => include/linux/types.h
<asm/types.h> => /usr/include/.../asm/types.h
either:
<asm-generic/int-ll64.h> => include/asm-generic/int-ll64.h
or:
<asm-generic/int-l64.h> => /usr/include/asm-generic/int-l64.h
Since there is no include/asm-generic/int-l64.h, whether or not we
define __SANE_USERSPACE_TYPES__ determines whether we include a kernel
or userland header. So that's why your fix works, but we really
shouldn't be including any userland headers at all.
I think we should avoid <linux/types.h> altogether, and make sure we
don't accidentally include UAPI headers:
Index: linux-tools/debian/build/scripts/mod/Makefile.real
===================================================================
--- linux-tools/debian/build/scripts/mod/Makefile.real (revision 21797)
+++ linux-tools/debian/build/scripts/mod/Makefile.real (working copy)
@@ -2,8 +2,6 @@
top_srcdir = ../..
-CFLAGS += -I$(top_srcdir)/include
-
include $(top_srcdir)/debian/build/Makefile.inc
modpost.real-$(TYPE): file2alias.real-$(TYPE).o modpost.real-$(TYPE).o
sumversion.real-$(TYPE).o
@@ -13,7 +11,7 @@
$(CC) -I real-$(TYPE) $(CFLAGS) -c -o $@ $<
real-$(TYPE)/devicetable-offsets.s: $(SOURCEDIR)/devicetable-offsets.c
- $(CC) -include real-$(TYPE)/types.h $(CFLAGS) -S -o $@ $<
+ $(CC) -include real-$(TYPE)/types.h $(CFLAGS) -nostdinc
-I$(top_srcdir)/include -S -o $@ $<
real-$(TYPE)/devicetable-offsets.h: real-$(TYPE)/devicetable-offsets.s
echo >$@ "#define __DEVICEVTABLE_OFFSETS_H__"
Index: linux-tools/debian/build/scripts/mod/real-lsb-32/types.h
===================================================================
--- linux-tools/debian/build/scripts/mod/real-lsb-32/types.h (revision 21797)
+++ linux-tools/debian/build/scripts/mod/real-lsb-32/types.h (working copy)
@@ -1,2 +1,3 @@
-#include <linux/types.h>
+#include "../types.h"
typedef __u32 kernel_ulong_t;
+#define BITS_PER_LONG 32
Index: linux-tools/debian/build/scripts/mod/real-lsb-64/types.h
===================================================================
--- linux-tools/debian/build/scripts/mod/real-lsb-64/types.h (revision 21797)
+++ linux-tools/debian/build/scripts/mod/real-lsb-64/types.h (working copy)
@@ -1,2 +1,3 @@
-#include <linux/types.h>
+#include "../types.h"
typedef __u64 __attribute__((aligned(8))) kernel_ulong_t;
+#define BITS_PER_LONG 64
Index: linux-tools/debian/build/scripts/mod/real-msb-32/types.h
===================================================================
--- linux-tools/debian/build/scripts/mod/real-msb-32/types.h (revision 21797)
+++ linux-tools/debian/build/scripts/mod/real-msb-32/types.h (working copy)
@@ -1,2 +1,3 @@
-#include <linux/types.h>
+#include "../types.h"
typedef __u32 kernel_ulong_t;
+#define BITS_PER_LONG 32
Index: linux-tools/debian/build/scripts/mod/real-msb-64/types.h
===================================================================
--- linux-tools/debian/build/scripts/mod/real-msb-64/types.h (revision 21797)
+++ linux-tools/debian/build/scripts/mod/real-msb-64/types.h (working copy)
@@ -1,2 +1,3 @@
-#include <linux/types.h>
+#include "../types.h"
typedef __u64 __attribute__((aligned(8))) kernel_ulong_t;
+#define BITS_PER_LONG 64
Index: linux-tools/debian/build/scripts/mod/types.h
===================================================================
--- linux-tools/debian/build/scripts/mod/types.h (revision 0)
+++ linux-tools/debian/build/scripts/mod/types.h (working copy)
@@ -0,0 +1,6 @@
+/* Minimal definitions for mod_devicetable.h and devicetable-offsets.c */
+typedef unsigned char __u8;
+typedef unsigned short __u16;
+typedef unsigned int __u32;
+typedef unsigned long long __u64;
+#define offsetof(a,b) __builtin_offsetof(a,b)
Index: linux-tools/debian/changelog
===================================================================
--- linux-tools/debian/changelog (revision 21797)
+++ linux-tools/debian/changelog (working copy)
@@ -1,3 +1,13 @@
+linux-tools (3.16-2) UNRELEASED; urgency=medium
+
+ [ Ben Hutchings ]
+ * linux-kbuild: Change the type headers used for devicetable-offsets.c
+ to avoid depending on UAPI headers or <linux/types.h>. This really
+ closes: #754213. It also fixes modpost handling of input device IDs
+ when host and target have differing word size.
+
+ -- Ben Hutchings <[email protected]> Tue, 09 Sep 2014 12:18:29 +0100
+
linux-tools (3.16-1) unstable; urgency=medium
* New upstream release
--- END ---
Avoiding <linux/types.h> forced me to define BITS_PER_LONG, which
resulted in the second fix mentioned above. However, a cross-modpost
still doesn't correctly handle various other device ID types where the
structure is defined to have pointer fields (not kernel_ulong_t).
Ben.
--
Ben Hutchings
If God had intended Man to program,
we'd have been born with serial I/O ports.
signature.asc
Description: This is a digitally signed message part
