build failure in Linus' tree with gcc 4.4.3

build failure in Linus' tree with gcc 4.4.3

[Stephen Rothwell]

[-- Attachment #1: Type: text/plain, Size: 1037 bytes --]

Hi all,

My x86_64 allmodconfig build fail like this:

  OBJCOPY arch/x86/boot/vmlinux.bin
  AS      arch/x86/boot/header.o
  LD      arch/x86/boot/setup.elf
  OBJCOPY arch/x86/boot/setup.bin
  BUILD   arch/x86/boot/bzImage
Setup is 14768 bytes (padded to 14848 bytes).
System is 4766 kB
/bin/sh: line 1: 20126 Segmentation fault      arch/x86/boot/tools/build arch/x86/boot/setup.bin arch/x86/boot/vmlinux.bin > arch/x86/boot/bzImage

I have bisected it down to commit
291f36325f9f252bd76ef5f603995f37e453fc60 ("x86, efi: EFI boot stub
support").  This may well be a compiler bug that this commit just happens
to poke.

The toolchain is PowerPC hosted cross tools:
x86_64-linux-gcc (GCC) 4.4.3
GNU ld (GNU Binutils) 2.20

I tried an allmodconfig build with my native compiler (gcc (Debian
4.6.2-14) 4.6.2, GNU gold (GNU Binutils for Debian 2.22) 1.11) and it
worked fine.

That is a large commit, so I am not sure where to go from here.
-- 
Cheers,
Stephen Rothwell                    sfr@canb.auug.org.au

[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]

Re: build failure in Linus' tree with gcc 4.4.3

[Stephen Rothwell]

[-- Attachment #1: Type: text/plain, Size: 3000 bytes --]

Hi all,

On Mon, 20 Feb 2012 13:39:36 +1100 Stephen Rothwell <sfr@canb.auug.org.au> wrote:
>
> Hi all,
> 
> My x86_64 allmodconfig build fail like this:
> 
>   OBJCOPY arch/x86/boot/vmlinux.bin
>   AS      arch/x86/boot/header.o
>   LD      arch/x86/boot/setup.elf
>   OBJCOPY arch/x86/boot/setup.bin
>   BUILD   arch/x86/boot/bzImage
> Setup is 14768 bytes (padded to 14848 bytes).
> System is 4766 kB
> /bin/sh: line 1: 20126 Segmentation fault      arch/x86/boot/tools/build arch/x86/boot/setup.bin arch/x86/boot/vmlinux.bin > arch/x86/boot/bzImage
> 
> I have bisected it down to commit
> 291f36325f9f252bd76ef5f603995f37e453fc60 ("x86, efi: EFI boot stub
> support").  This may well be a compiler bug that this commit just happens
> to poke.
> 
> The toolchain is PowerPC hosted cross tools:
> x86_64-linux-gcc (GCC) 4.4.3
> GNU ld (GNU Binutils) 2.20
> 
> I tried an allmodconfig build with my native compiler (gcc (Debian
> 4.6.2-14) 4.6.2, GNU gold (GNU Binutils for Debian 2.22) 1.11) and it
> worked fine.
> 
> That is a large commit, so I am not sure where to go from here.

Actually, that commit does the following:

diff --git a/arch/x86/boot/tools/build.c b/arch/x86/boot/tools/build.c
index fdc60a0..4e9bd6b 100644
--- a/arch/x86/boot/tools/build.c
+++ b/arch/x86/boot/tools/build.c
@@ -135,6 +135,9 @@ static void usage(void)
 
 int main(int argc, char ** argv)
 {
+#ifdef CONFIG_EFI_STUB
+	unsigned int file_sz, pe_header;
+#endif
 	unsigned int i, sz, setup_sectors;
 	int c;
 	u32 sys_size;
@@ -194,6 +197,42 @@ int main(int argc, char ** argv)
 	buf[0x1f6] = sys_size >> 16;
 	buf[0x1f7] = sys_size >> 24;
 
+#ifdef CONFIG_EFI_STUB
+	file_sz = sz + i + ((sys_size * 16) - sz);
+
+	pe_header = *(unsigned int *)&buf[0x3c];
+
+	/* Size of code */
+	*(unsigned int *)&buf[pe_header + 0x1c] = file_sz;
+
+	/* Size of image */
+	*(unsigned int *)&buf[pe_header + 0x50] = file_sz;
+
+#ifdef CONFIG_X86_32
+	/* Address of entry point */
+	*(unsigned int *)&buf[pe_header + 0x28] = i;
+
+	/* .text size */
+	*(unsigned int *)&buf[pe_header + 0xb0] = file_sz;
+
+	/* .text size of initialised data */
+	*(unsigned int *)&buf[pe_header + 0xb8] = file_sz;
+#else
+	/*
+	 * Address of entry point. startup_32 is at the beginning and
+	 * the 64-bit entry point (startup_64) is always 512 bytes
+	 * after.
+	 */
+	*(unsigned int *)&buf[pe_header + 0x28] = i + 512;
+
+	/* .text size */
+	*(unsigned int *)&buf[pe_header + 0xc0] = file_sz;
+
+	/* .text size of initialised data */
+	*(unsigned int *)&buf[pe_header + 0xc8] = file_sz;
+#endif /* CONFIG_X86_32 */
+#endif /* CONFIG_EFI_STUB */
+
 	crc = partial_crc32(buf, i, crc);
 	if (fwrite(buf, 1, i, stdout) != i)
 		die("Writing setup failed");

Which is all endian specific code that will run on the host when building
the kernel ...

-- 
Cheers,
Stephen Rothwell                    sfr@canb.auug.org.au
http://www.canb.auug.org.au/~sfr/

[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]

Re: build failure in Linus' tree with gcc 4.4.3

[Matt Fleming]

On Mon, 2012-02-20 at 14:22 +1100, Stephen Rothwell wrote:
> Which is all endian specific code that will run on the host when building
> the kernel ...

Gah, right. I didn't think about cross-building this code.

Thanks Stephen, I'll fix this up.

Re: build failure in Linus' tree with gcc 4.4.3

[Matt Fleming]

On Mon, 2012-02-20 at 09:43 +0000, Matt Fleming wrote:
> On Mon, 2012-02-20 at 14:22 +1100, Stephen Rothwell wrote:
> > Which is all endian specific code that will run on the host when building
> > the kernel ...
> 
> Gah, right. I didn't think about cross-building this code.
> 
> Thanks Stephen, I'll fix this up.

Looks like the segfault is caused by an unaligned access? How does this
patch look?

>From 54b2707a6a911330d8db2f4ec2fb1baa5e38acf9 Mon Sep 17 00:00:00 2001
From: Matt Fleming <matt.fleming@intel.com>
Date: Mon, 20 Feb 2012 17:32:42 +0000
Subject: [PATCH] x86, efi: Fix segfault caused by unaligned access

We need to use memcpy() instead of directly dereferencing a pointer
because memcpy() correctly handles the case where the
source/destination are unaligned, which can lead to a segfault when
cross-building an x86 kernel on risc architectures.

Stephen Rothwell noticed this bug when he hit a segfault while
cross-building an x86_64 allmodconfig kernel on PowerPC.

Cc: H. Peter Anvin <hpa@zytor.com>
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
---
 arch/x86/boot/tools/build.c |   26 ++++++++++++++------------
 1 files changed, 14 insertions(+), 12 deletions(-)

diff --git a/arch/x86/boot/tools/build.c b/arch/x86/boot/tools/build.c
index 4e9bd6b..b4d85b5 100644
--- a/arch/x86/boot/tools/build.c
+++ b/arch/x86/boot/tools/build.c
@@ -200,36 +200,38 @@ int main(int argc, char ** argv)
 #ifdef CONFIG_EFI_STUB
 	file_sz = sz + i + ((sys_size * 16) - sz);
 
-	pe_header = *(unsigned int *)&buf[0x3c];
+	memcpy(&pe_header, &buf[0x3c], sizeof(pe_header));
 
 	/* Size of code */
-	*(unsigned int *)&buf[pe_header + 0x1c] = file_sz;
+	memcpy(&buf[pe_header + 0x1c], &file_sz, sizeof(file_sz));
 
 	/* Size of image */
-	*(unsigned int *)&buf[pe_header + 0x50] = file_sz;
+	memcpy(&buf[pe_header + 0x50], &file_sz, sizeof(file_sz));
 
 #ifdef CONFIG_X86_32
 	/* Address of entry point */
-	*(unsigned int *)&buf[pe_header + 0x28] = i;
+	memcpy(&buf[pe_header + 0x28], &i, sizeof(i));
 
 	/* .text size */
-	*(unsigned int *)&buf[pe_header + 0xb0] = file_sz;
+	memcpy(&buf[pe_header + 0xb0], &file_sz, sizeof(file_sz));
 
 	/* .text size of initialised data */
-	*(unsigned int *)&buf[pe_header + 0xb8] = file_sz;
+	memcpy(&buf[pe_header + 0xb8], &file_sz, sizeof(file_sz));
 #else
+	/* .text size */
+	memcpy(&buf[pe_header + 0xc0], &file_sz, sizeof(file_sz));
+
+	/* .text size of initialised data */
+	memcpy(&buf[pe_header + 0xc8], &file_sz, sizeof(file_sz));
+
 	/*
 	 * Address of entry point. startup_32 is at the beginning and
 	 * the 64-bit entry point (startup_64) is always 512 bytes
 	 * after.
 	 */
-	*(unsigned int *)&buf[pe_header + 0x28] = i + 512;
+	file_sz = i + 512;
+	memcpy(&buf[pe_header + 0x28], &file_sz, sizeof(file_sz));
 
-	/* .text size */
-	*(unsigned int *)&buf[pe_header + 0xc0] = file_sz;
-
-	/* .text size of initialised data */
-	*(unsigned int *)&buf[pe_header + 0xc8] = file_sz;
 #endif /* CONFIG_X86_32 */
 #endif /* CONFIG_EFI_STUB */
 
-- 
1.7.4.4

Re: build failure in Linus' tree with gcc 4.4.3

[Stephen Rothwell]

[-- Attachment #1: Type: text/plain, Size: 406 bytes --]

Hi Matt,

On Tue, 21 Feb 2012 10:01:57 +0000 Matt Fleming <matt.fleming@intel.com> wrote:
>
> Looks like the segfault is caused by an unaligned access? How does this
> patch look?

I didn't try it, but it won't work.  PowerPC is big endian, the data you
are copying is little endian ....

-- 
Cheers,
Stephen Rothwell                    sfr@canb.auug.org.au
http://www.canb.auug.org.au/~sfr/

[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]

Re: build failure in Linus' tree with gcc 4.4.3

[Matt Fleming]

On Tue, 2012-02-21 at 21:38 +1100, Stephen Rothwell wrote:
> Hi Matt,
> 
> On Tue, 21 Feb 2012 10:01:57 +0000 Matt Fleming <matt.fleming@intel.com> wrote:
> >
> > Looks like the segfault is caused by an unaligned access? How does this
> > patch look?
> 
> I didn't try it, but it won't work.  PowerPC is big endian, the data you
> are copying is little endian ....

Urgh, you must feel like you're talking to a brick wall. I forgot you
mentioned the endian issue in your original post, sorry. Let's try this
one more time...

>From b3885c72be4cb06b2647a5067432da6017ea7902 Mon Sep 17 00:00:00 2001
From: Matt Fleming <matt.fleming@intel.com>
Date: Mon, 20 Feb 2012 17:32:42 +0000
Subject: [PATCH] x86, efi: Fix unaligned access and endian issues

We need to read from and write to 'buf' a byte at a time otherwise
it's possible we'll perform an unaligned access, which can lead to a
segfault when cross-building an x86 kernel on risc architectures.

Also, we may need to convert the endianness of the data we read
from/write to buf, so let's add some helper functions to do that.

Stephen Rothwell noticed this bug when he hit a segfault while
cross-building an x86_64 allmodconfig kernel on PowerPC.

Cc: H. Peter Anvin <hpa@zytor.com>
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
---
 arch/x86/boot/tools/build.c |   44 ++++++++++++++++++++++++++++++------------
 1 files changed, 31 insertions(+), 13 deletions(-)

diff --git a/arch/x86/boot/tools/build.c b/arch/x86/boot/tools/build.c
index 4e9bd6b..56efa6f 100644
--- a/arch/x86/boot/tools/build.c
+++ b/arch/x86/boot/tools/build.c
@@ -133,6 +133,26 @@ static void usage(void)
 	die("Usage: build setup system [> image]");
 }
 
+static inline u32 read32_le(u8 *src)
+{
+	u32 data;
+
+	data = *src++;
+	data |= *src++ >> 8;
+	data |= *src++ >> 16;
+	data |= *src++ >> 24;
+
+	return data;
+}
+
+static inline void write32_le(u8 *dst, u32 data)
+{
+	*dst++ = data;
+	*dst++ = data >> 8;
+	*dst++ = data >> 16;
+	*dst++ = data >> 24;
+}
+
 int main(int argc, char ** argv)
 {
 #ifdef CONFIG_EFI_STUB
@@ -192,44 +212,42 @@ int main(int argc, char ** argv)
 
 	/* Patch the setup code with the appropriate size parameters */
 	buf[0x1f1] = setup_sectors-1;
-	buf[0x1f4] = sys_size;
-	buf[0x1f5] = sys_size >> 8;
-	buf[0x1f6] = sys_size >> 16;
-	buf[0x1f7] = sys_size >> 24;
+	write32_le(&buf[0x1f4], sys_size);
 
 #ifdef CONFIG_EFI_STUB
 	file_sz = sz + i + ((sys_size * 16) - sz);
 
-	pe_header = *(unsigned int *)&buf[0x3c];
+	pe_header = read32_le(&buf[0x3c]);
 
 	/* Size of code */
-	*(unsigned int *)&buf[pe_header + 0x1c] = file_sz;
+	write32_le(&buf[pe_header + 0x1c], file_sz);
 
 	/* Size of image */
-	*(unsigned int *)&buf[pe_header + 0x50] = file_sz;
+	write32_le(&buf[pe_header + 0x50], file_sz);
 
 #ifdef CONFIG_X86_32
 	/* Address of entry point */
-	*(unsigned int *)&buf[pe_header + 0x28] = i;
+	write32_le(&buf[pe_header + 0x28], i);
 
 	/* .text size */
-	*(unsigned int *)&buf[pe_header + 0xb0] = file_sz;
+	write32_le(&buf[pe_header + 0xb0], file_sz);
 
 	/* .text size of initialised data */
-	*(unsigned int *)&buf[pe_header + 0xb8] = file_sz;
+	write32_le(&buf[pe_header + 0xb8], file_sz);
 #else
 	/*
 	 * Address of entry point. startup_32 is at the beginning and
 	 * the 64-bit entry point (startup_64) is always 512 bytes
 	 * after.
 	 */
-	*(unsigned int *)&buf[pe_header + 0x28] = i + 512;
+	write32_le(&buf[pe_header + 0x28], i + 512);
 
 	/* .text size */
-	*(unsigned int *)&buf[pe_header + 0xc0] = file_sz;
+	write32_le(&buf[pe_header + 0xc0], file_sz);
 
 	/* .text size of initialised data */
-	*(unsigned int *)&buf[pe_header + 0xc8] = file_sz;
+	write32_le(&buf[pe_header + 0xc8], file_sz);
+
 #endif /* CONFIG_X86_32 */
 #endif /* CONFIG_EFI_STUB */
 
-- 
1.7.4.4

Re: build failure in Linus' tree with gcc 4.4.3

[Stephen Rothwell]

[-- Attachment #1: Type: text/plain, Size: 401 bytes --]

Hi Matt,

On Tue, 21 Feb 2012 11:32:13 +0000 Matt Fleming <matt.fleming@intel.com> wrote:
>
> +static inline u32 read32_le(u8 *src)
> +{
> +	u32 data;
> +
> +	data = *src++;
> +	data |= *src++ >> 8;
> +	data |= *src++ >> 16;
> +	data |= *src++ >> 24;

Shouldn't these be "<<" ?

-- 
Cheers,
Stephen Rothwell                    sfr@canb.auug.org.au
http://www.canb.auug.org.au/~sfr/

[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]

Re: build failure in Linus' tree with gcc 4.4.3

[Matt Fleming]

On Tue, 2012-02-21 at 22:51 +1100, Stephen Rothwell wrote:
> Hi Matt,
> 
> On Tue, 21 Feb 2012 11:32:13 +0000 Matt Fleming <matt.fleming@intel.com> wrote:
> >
> > +static inline u32 read32_le(u8 *src)
> > +{
> > +	u32 data;
> > +
> > +	data = *src++;
> > +	data |= *src++ >> 8;
> > +	data |= *src++ >> 16;
> > +	data |= *src++ >> 24;
> 
> Shouldn't these be "<<" ?

Oops. Yes.