[PATCH] libbpf: avoid inline hint definition from 'linux/stddef.h'

Linux headers might pull 'linux/stddef.h' which defines
'__always_inline' as the following:

   #ifndef __always_inline
   #define __always_inline __inline__
   #endif

This becomes an issue if the program picks up the 'linux/stddef.h'
definition as the macro now just hints inline to clang.

This change now enforces the proper definition for BPF programs
regardless of the include order.

Signed-off-by: Pedro Tammela <redacted>
---
 tools/lib/bpf/bpf_helpers.h | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/tools/lib/bpf/bpf_helpers.h b/tools/lib/bpf/bpf_helpers.h
index ae6c975e0b87..5fa483c0b508 100644
--- a/tools/lib/bpf/bpf_helpers.h
+++ b/tools/lib/bpf/bpf_helpers.h

@@ -29,9 +29,12 @@
  */
 #define SEC(NAME) __attribute__((section(NAME), used))
 
-#ifndef __always_inline
+/*
+ * Avoid 'linux/stddef.h' definition of '__always_inline'.
+ */
+#undef __always_inline
 #define __always_inline inline __attribute__((always_inline))
-#endif
+
 #ifndef __noinline
 #define __noinline __attribute__((noinline))
 #endif

-- 
2.25.1

On 3/14/21 6:38 PM, Pedro Tammela wrote:

Linux headers might pull 'linux/stddef.h' which defines
'__always_inline' as the following:

    #ifndef __always_inline
    #define __always_inline __inline__
    #endif

This becomes an issue if the program picks up the 'linux/stddef.h'
definition as the macro now just hints inline to clang.

How did the program end up including linux/stddef.h ? Would be good to
also have some more details on how we got here for the commit desc.

I think the comment should have more details on 'why' we undef it as in
few months looking at it again, the next question to dig into would be
what was wrong with linux/stddef.h. Providing a better rationale would
be nice for readers here.

+#undef __always_inline
  #define __always_inline inline __attribute__((always_inline))
-#endif
+
  #ifndef __noinline
  #define __noinline __attribute__((noinline))
  #endif

On Tue, Mar 16, 2021 at 2:01 PM Daniel Borkmann [off-list ref] wrote:

On 3/14/21 6:38 PM, Pedro Tammela wrote:

quoted

Linux headers might pull 'linux/stddef.h' which defines
'__always_inline' as the following:

    #ifndef __always_inline
    #define __always_inline __inline__
    #endif

This becomes an issue if the program picks up the 'linux/stddef.h'
definition as the macro now just hints inline to clang.

How did the program end up including linux/stddef.h ? Would be good to
also have some more details on how we got here for the commit desc.

It's an UAPI header, so why not? Is there anything special about
stddef.h that makes it unsuitable to be included?

quoted

This change now enforces the proper definition for BPF programs
regardless of the include order.

Signed-off-by: Pedro Tammela <redacted>
---
  tools/lib/bpf/bpf_helpers.h | 7 +++++--
  1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/tools/lib/bpf/bpf_helpers.h b/tools/lib/bpf/bpf_helpers.h
index ae6c975e0b87..5fa483c0b508 100644
--- a/tools/lib/bpf/bpf_helpers.h
+++ b/tools/lib/bpf/bpf_helpers.h

@@ -29,9 +29,12 @@
   */
  #define SEC(NAME) __attribute__((section(NAME), used))

-#ifndef __always_inline
+/*
+ * Avoid 'linux/stddef.h' definition of '__always_inline'.
+ */

I think the comment should have more details on 'why' we undef it as in
few months looking at it again, the next question to dig into would be
what was wrong with linux/stddef.h. Providing a better rationale would
be nice for readers here.

So for whatever reason commit bot didn't send notification, but I've
already landed this yesterday. To me, with #undef + #define it's
pretty clear that we "force-define" __always_inline exactly how we
want it, but we can certainly add clarifying comment in the follow up,
if you think it's needed.

quoted

+#undef __always_inline
  #define __always_inline inline __attribute__((always_inline))
-#endif
+
  #ifndef __noinline
  #define __noinline __attribute__((noinline))
  #endif

On 3/16/21 10:34 PM, Andrii Nakryiko wrote:

On Tue, Mar 16, 2021 at 2:01 PM Daniel Borkmann [off-list ref] wrote:

quoted

On 3/14/21 6:38 PM, Pedro Tammela wrote:

quoted

Linux headers might pull 'linux/stddef.h' which defines
'__always_inline' as the following:

     #ifndef __always_inline
     #define __always_inline __inline__
     #endif

This becomes an issue if the program picks up the 'linux/stddef.h'
definition as the macro now just hints inline to clang.

How did the program end up including linux/stddef.h ? Would be good to
also have some more details on how we got here for the commit desc.

It's an UAPI header, so why not? Is there anything special about
stddef.h that makes it unsuitable to be included?

Hm, fair enough, looks like linux/types.h already pulls it in, so no. We
defined our own stddef.h longer time ago, so looks like we never ran into
this issue.

quoted

quoted

This change now enforces the proper definition for BPF programs
regardless of the include order.

Signed-off-by: Pedro Tammela <redacted>
---
   tools/lib/bpf/bpf_helpers.h | 7 +++++--
   1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/tools/lib/bpf/bpf_helpers.h b/tools/lib/bpf/bpf_helpers.h
index ae6c975e0b87..5fa483c0b508 100644
--- a/tools/lib/bpf/bpf_helpers.h
+++ b/tools/lib/bpf/bpf_helpers.h

@@ -29,9 +29,12 @@
    */
   #define SEC(NAME) __attribute__((section(NAME), used))

-#ifndef __always_inline
+/*
+ * Avoid 'linux/stddef.h' definition of '__always_inline'.
+ */

I think the comment should have more details on 'why' we undef it as in
few months looking at it again, the next question to dig into would be
what was wrong with linux/stddef.h. Providing a better rationale would
be nice for readers here.

So for whatever reason commit bot didn't send notification, but I've
already landed this yesterday. To me, with #undef + #define it's
pretty clear that we "force-define" __always_inline exactly how we
want it, but we can certainly add clarifying comment in the follow up,
if you think it's needed.

Up to you, but given you applied it it's probably not worth the trouble;
missed it earlier given I didn't see the patchbot message in the thread
initially. :/

quoted

quoted

+#undef __always_inline
   #define __always_inline inline __attribute__((always_inline))
-#endif
+
   #ifndef __noinline
   #define __noinline __attribute__((noinline))
   #endif