Skip to contentSkip to conversation

locking/qrwlock: include asm/byteorder.h as needed

PATCH Applied Arnd Bergmann 4 participants 6 messages Feb 2, 2018 – Feb 6, 2018

Subsystems (2)
lore
Share & tools
lore

Arnd Bergmann

author maintainer
#0 lore raw

PATCH 1/2 locking/qrwlock: include asm/byteorder.h as needed

Moving the qrwlock struct definition into a header file introduced a subtle bug on all little-endian machines, where some files in some configurations would see the fields in an incorrect order. This was found by building with an LTO enabled compiler that warns every time we try to link together files with incompatible data structures.

A second patch changes linux/kconfig.h to always define the symbols, but this seems to be the root cause of most of the issues, so I'd suggest we do both.

On a current linux-next kernel, I verified that this header is responsible for all type mismatches as a result from the endianess confusion.

Cc Will Deacon <[email protected]>
Cc Peter Zijlstra <[email protected]>
Fixes e0d02285f16e ("locking/qrwlock: Use 'struct qrwlock' instead of 'struct __qrwlock'")
---
 include/asm-generic/qrwlock_types.h | 1 +
 1 file changed, 1 insertion(+)
include/asm-generic/qrwlock_types.h +1 -0
index 137ecdd16daa..c36f1d5a2572 100644
@@ -3,6 +3,7 @@
 #define __ASM_GENERIC_QRWLOCK_TYPES_H
 
 #include <linux/types.h>
+#include <asm/byteorder.h>
 #include <asm/spinlock_types.h>
 
 /*
Signature · 1 line
2.9.0
author maintainer
#1 lore raw

PATCH 2/2 Kbuild: always define endianess in kconfig.h

Build testing with LTO found a couple of files that get compiled differently depending on whether asm/byteorder.h gets included early enough or not. In particular, include/asm-generic/qrwlock_types.h is affected by this, but there are probably others as well.

The symptom is a series of LTO link time warnings, including these:

net/netlabel/netlabel_unlabeled.h:223: error: type of 'netlbl_unlhsh_add' does not match original declaration [-Werror=lto-type-mismatch]
int netlbl_unlhsh_add(struct net *net,
net/netlabel/netlabel_unlabeled.c:377: note: 'netlbl_unlhsh_add' was previously declared here

include/net/ipv6.h:360: error: type of 'ipv6_renew_options_kern' does not match original declaration [-Werror=lto-type-mismatch]
ipv6_renew_options_kern(struct sock *sk,
net/ipv6/exthdrs.c:1162: note: 'ipv6_renew_options_kern' was previously declared here

net/core/dev.c:761: note: 'dev_get_by_name_rcu' was previously declared here
struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
net/core/dev.c:761: note: code may be misoptimized unless -fno-strict-aliasing is used

drivers/gpu/drm/i915/i915_drv.h:3377: error: type of 'i915_gem_object_set_to_wc_domain' does not match original declaration [-Werror=lto-type-mismatch]
i915_gem_object_set_to_wc_domain(struct drm_i915_gem_object *obj, bool write);
drivers/gpu/drm/i915/i915_gem.c:3639: note: 'i915_gem_object_set_to_wc_domain' was previously declared here

include/linux/debugfs.h:92:9: error: type of 'debugfs_attr_read' does not match original declaration [-Werror=lto-type-mismatch]
ssize_t debugfs_attr_read(struct file *file, char __user *buf,
fs/debugfs/file.c:318: note: 'debugfs_attr_read' was previously declared here

include/linux/rwlock_api_smp.h:30: error: type of '_raw_read_unlock' does not match original declaration [-Werror=lto-type-mismatch]
void __lockfunc _raw_read_unlock(rwlock_t *lock) __releases(lock);
kernel/locking/spinlock.c:246:26: note: '_raw_read_unlock' was previously declared here

include/linux/fs.h:3308:5: error: type of 'simple_attr_open' does not match original declaration [-Werror=lto-type-mismatch]
int simple_attr_open(struct inode *inode, struct file *file,
fs/libfs.c:795: note: 'simple_attr_open' was previously declared here

All of the above are caused by include/asm-generic/qrwlock_types.h failing to include asm/byteorder.h after commit e0d02285f16e ("locking/qrwlock:
Use 'struct qrwlock' instead of 'struct __qrwlock'") in linux-4.15.

Similar bugs may or may not exist in older kernels as well, but there is no easy way to test those with link-time optimizations, and kernels before 4.14 are harder to fix because they don't have Babu's patch series

We had similar issues with CONFIG_ symbols in the past and ended up always including the configuration headers though linux/kconfig.h. This works around the issue through that same file, defining either __BIG_ENDIAN or __LITTLE_ENDIAN depending on CONFIG_CPU_BIG_ENDIAN, which is now always set on all architectures since commit 4c97a0c8fee3 ("arch: define CPU_BIG_ENDIAN for all fixed big endian archs").

---
 include/linux/kconfig.h | 6 ++++++
 1 file changed, 6 insertions(+)
include/linux/kconfig.h +6 -0
index fec5076eda91..cc8fa109cfa3 100644
@@ -4,6 +4,12 @@
 
 #include <generated/autoconf.h>
 
+#ifdef CONFIG_CPU_BIG_ENDIAN
+#define __BIG_ENDIAN 4321
+#else
+#define __LITTLE_ENDIAN 1234
+#endif
+
 #define __ARG_PLACEHOLDER_1 0,
 #define __take_second_arg(__ignored, val, ...) val
 
Signature · 1 line
2.9.0
#2 lore raw

Re: [PATCH 2/2] Kbuild: always define endianess in kconfig.h

Hi Arnd,

2018-02-03 0:40 GMT+09:00 Arnd Bergmann <[email protected]>:

Quoted 35 lines

Build testing with LTO found a couple of files that get compiled differently depending on whether asm/byteorder.h gets included early enough or not. In particular, include/asm-generic/qrwlock_types.h is affected by this, but there are probably others as well.

The symptom is a series of LTO link time warnings, including these:

net/netlabel/netlabel_unlabeled.h:223: error: type of 'netlbl_unlhsh_add' does not match original declaration [-Werror=lto-type-mismatch]
int netlbl_unlhsh_add(struct net *net,
net/netlabel/netlabel_unlabeled.c:377: note: 'netlbl_unlhsh_add' was previously declared here

include/net/ipv6.h:360: error: type of 'ipv6_renew_options_kern' does not match original declaration [-Werror=lto-type-mismatch]
ipv6_renew_options_kern(struct sock *sk,
net/ipv6/exthdrs.c:1162: note: 'ipv6_renew_options_kern' was previously declared here

net/core/dev.c:761: note: 'dev_get_by_name_rcu' was previously declared here
struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
net/core/dev.c:761: note: code may be misoptimized unless -fno-strict-aliasing is used

drivers/gpu/drm/i915/i915_drv.h:3377: error: type of 'i915_gem_object_set_to_wc_domain' does not match original declaration [-Werror=lto-type-mismatch]
i915_gem_object_set_to_wc_domain(struct drm_i915_gem_object *obj, bool write);
drivers/gpu/drm/i915/i915_gem.c:3639: note: 'i915_gem_object_set_to_wc_domain' was previously declared here

include/linux/debugfs.h:92:9: error: type of 'debugfs_attr_read' does not match original declaration [-Werror=lto-type-mismatch]
ssize_t debugfs_attr_read(struct file *file, char __user *buf,
fs/debugfs/file.c:318: note: 'debugfs_attr_read' was previously declared here

include/linux/rwlock_api_smp.h:30: error: type of '_raw_read_unlock' does not match original declaration [-Werror=lto-type-mismatch]
void __lockfunc _raw_read_unlock(rwlock_t *lock) __releases(lock);
kernel/locking/spinlock.c:246:26: note: '_raw_read_unlock' was previously declared here

include/linux/fs.h:3308:5: error: type of 'simple_attr_open' does not match original declaration [-Werror=lto-type-mismatch]
int simple_attr_open(struct inode *inode, struct file *file,
fs/libfs.c:795: note: 'simple_attr_open' was previously declared here

Quoted 7 lines

All of the above are caused by include/asm-generic/qrwlock_types.h failing to include asm/byteorder.h after commit e0d02285f16e ("locking/qrwlock:
Use 'struct qrwlock' instead of 'struct __qrwlock'") in linux-4.15.

Similar bugs may or may not exist in older kernels as well, but there is no easy way to test those with link-time optimizations, and kernels before 4.14 are harder to fix because they don't have Babu's patch series

Is this patch only for stable kernel,
not for the mainline?

Quoted 32 lines

We had similar issues with CONFIG_ symbols in the past and ended up always including the configuration headers though linux/kconfig.h. This works around the issue through that same file, defining either __BIG_ENDIAN or __LITTLE_ENDIAN depending on CONFIG_CPU_BIG_ENDIAN, which is now always set on all architectures since commit 4c97a0c8fee3 ("arch: define CPU_BIG_ENDIAN for all fixed big endian archs").

---
 include/linux/kconfig.h | 6 ++++++
 1 file changed, 6 insertions(+)
include/linux/kconfig.h +6 -1
index fec5076eda91..cc8fa109cfa3 100644
@@ -4,6 +4,12 @@
 
 #include <generated/autoconf.h>
 
+#ifdef CONFIG_CPU_BIG_ENDIAN
+#define __BIG_ENDIAN 4321
+#else
+#define __LITTLE_ENDIAN 1234
+#endif
+
 #define __ARG_PLACEHOLDER_1 0,
 #define __take_second_arg(__ignored, val, ...) val
 
--
Quoted 2 lines

2.9.0

If all architectures define CONFIG_CPU_BIG_ENDIAN or CONFIG_CPU_LITTLE_ENDIAN, is it possible to use this for endian test?

i.e.

Is it possible to replace like this?
#ifdef __BIG_ENDIAN -> #ifdef CONFIG_CPU_BIG_ENDIAN

   #ifdef __LITTLE_ENDIAN     -> #ifdef CONFIG_CPU_LITTLE_ENDIAN
Signature · 2 lines
Best Regards
Masahiro Yamada
author maintainer
#3 lore raw

Re: [PATCH 2/2] Kbuild: always define endianess in kconfig.h

Quoted 31 lines · Masahiro Yamada

2018-02-03 0:40 GMT+09:00 Arnd Bergmann <[email protected]>:

Quoted 22 lines
---
 include/linux/kconfig.h | 6 ++++++
 1 file changed, 6 insertions(+)
include/linux/kconfig.h +6 -1
index fec5076eda91..cc8fa109cfa3 100644
@@ -4,6 +4,12 @@
 
 #include <generated/autoconf.h>
 
+#ifdef CONFIG_CPU_BIG_ENDIAN
+#define __BIG_ENDIAN 4321
+#else
+#define __LITTLE_ENDIAN 1234
+#endif
+
 #define __ARG_PLACEHOLDER_1 0,
 #define __take_second_arg(__ignored, val, ...) val
 
--
Quoted 2 lines

2.9.0

If all architectures define CONFIG_CPU_BIG_ENDIAN or CONFIG_CPU_LITTLE_ENDIAN, is it possible to use this for endian test?

Quoted 6 lines

i.e.

Is it possible to replace like this?
#ifdef __BIG_ENDIAN -> #ifdef CONFIG_CPU_BIG_ENDIAN

   #ifdef __LITTLE_ENDIAN     -> #ifdef CONFIG_CPU_LITTLE_ENDIAN

I believe we don't always define CONFIG_CPU_LITTLE_ENDIAN, but it would work in principle. The main issue here is changing so many files:

$ git grep -wl '__\(BIG\|LITTLE\)_ENDIAN' | wc -l
418
$ git grep -wl 'CONFIG_CPU_\(BIG\|LITTLE\)_ENDIAN' | wc -l
145

Arnd

#5 lore raw
In reply to Arnd Bergmann
locking/qrwlock: include asm/byteorder.h as needed
Commit-ID:  ca66e797120fb09b8138623fb4b563e952586ef5
Gitweb:     https://git.kernel.org/tip/ca66e797120fb09b8138623fb4b563e952586ef5
Author:     Arnd Bergmann <[email protected]>
AuthorDate: Fri, 2 Feb 2018 16:40:44 +0100
Committer:  Ingo Molnar <[email protected]>
CommitDate: Tue, 6 Feb 2018 10:28:58 +0100

locking/qrwlock: include asm/byteorder.h as needed

Moving the qrwlock struct definition into a header file introduced a subtle bug on all little-endian machines, where some files in some configurations would see the fields in an incorrect order. This was found by building with an LTO enabled compiler that warns every time we try to link together files with incompatible data structures.

A second patch changes linux/kconfig.h to always define the symbols, but this seems to be the root cause of most of the issues, so I'd suggest we do both.

On a current linux-next kernel, I verified that this header is responsible for all type mismatches as a result from the endianess confusion.

Cc Andrew Morton <[email protected]>
Cc Babu Moger <[email protected]>
Cc Greg Kroah-Hartman <[email protected]>
Cc Linus Torvalds <[email protected]>
Cc Nicolas Pitre <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Will Deacon <[email protected]>
Fixes e0d02285f16e ("locking/qrwlock: Use 'struct qrwlock' instead of 'struct __qrwlock'")
---
 include/asm-generic/qrwlock_types.h | 1 +
 1 file changed, 1 insertion(+)
include/asm-generic/qrwlock_types.h +1 -0
index 137ecdd..c36f1d5 100644
@@ -3,6 +3,7 @@
 #define __ASM_GENERIC_QRWLOCK_TYPES_H
 
 #include <linux/types.h>
+#include <asm/byteorder.h>
 #include <asm/spinlock_types.h>
 
 /*