[LTP] [PATCH/RFC] profil01: don't use address of function

[LTP] [PATCH/RFC] profil01: don't use address of function

[Jan Stancek]

This testcase is failing on ppc64 BE system, because
address of &test_profil does not reference beginning of code,
but entry in .opd section:
  "This section contains the official procedure descriptors.
   A pointer to a function shall reference a procedure descriptor
   in this section."
which was at address larger than all .text of test_profil.

This patch is replacing address of function approach with gcc's
__builtin_return_address() to get current pc value.

Signed-off-by: Jan Stancek <jstancek@redhat.com>
---
 testcases/kernel/syscalls/profil/profil01.c |   15 +++++++++++----
 1 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/testcases/kernel/syscalls/profil/profil01.c b/testcases/kernel/syscalls/profil/profil01.c
index 5511f68..521ef97 100644
--- a/testcases/kernel/syscalls/profil/profil01.c
+++ b/testcases/kernel/syscalls/profil/profil01.c
@@ -31,9 +31,11 @@
 
 #define PROFIL_TIME 5
 
-/* should be large enough to hold data for test_profil() .text
- * on x86_64 this is ~600 bytes, so 16k should enough for all arches */
-#define PROFIL_BUFLEN (16*1024)
+/* Should be large enough to hold data for test_profil() .text,
+ * on x86_64 this is ~600 bytes, so 16k should enough for all arches.
+ * We will monitor 16k on each side around current pc value,
+ * just in case compiler put call to get_pc() below "data shuffling" code */
+#define PROFIL_BUFLEN (32*1024)
 
 char *TCID = "profil01";
 int TST_TOTAL = 1;
@@ -46,11 +48,16 @@ static void alrm_handler(int sig)
 	profil_done = 1;
 }
 
+static void __attribute__ ((noinline)) *get_pc(void)
+{
+	return __builtin_return_address(0);
+}
+
 static void test_profil(void)
 {
 	unsigned short buf[PROFIL_BUFLEN] = { 0 };
 	volatile int data[8] = { 0 };
-	size_t offset = (size_t) &test_profil, count = 0;
+	size_t offset = (size_t) get_pc() - PROFIL_BUFLEN/2, count = 0;
 	int ret, i;
 
 	/* reset for test looping */
-- 
1.7.1


------------------------------------------------------------------------------
Slashdot TV.  
Video for Nerds.  Stuff that matters.
http://tv.slashdot.org/
_______________________________________________
Ltp-list mailing list
Ltp-list@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/ltp-list

Re: [LTP] [PATCH/RFC] profil01: don't use address of function

[chrubis]

Hi!
> This testcase is failing on ppc64 BE system, because
> address of &test_profil does not reference beginning of code,
> but entry in .opd section:
>   "This section contains the official procedure descriptors.
>    A pointer to a function shall reference a procedure descriptor
>    in this section."
> which was at address larger than all .text of test_profil.
> 
> This patch is replacing address of function approach with gcc's
> __builtin_return_address() to get current pc value.

How is it with compatibility? In which gcc version was this building
introduced? Does it work with clang?

If we are in doubt, we should add a configure check and use it only when
available. And probably to default to &test_profil in all cases but
ppc64 BE which is known to fail.

-- 
Cyril Hrubis
chrubis@suse.cz

------------------------------------------------------------------------------
Slashdot TV.  
Video for Nerds.  Stuff that matters.
http://tv.slashdot.org/
_______________________________________________
Ltp-list mailing list
Ltp-list@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/ltp-list

Re: [LTP] [PATCH/RFC] profil01: don't use address of function

[chrubis]

Hi!
> I think it was introduced in 90s, though not every architecture supported
> it at first.
> 
> I see it in very first release of clang 1.0 sources:
> ./src/gcc/builtins.def:DEF_GCC_BUILTIN
>         (BUILT_IN_RETURN_ADDRESS, "return_address", BT_FN_PTR_UNSIGNED, ATTR_NULL)
> but I didn't try to compile it with that.
> 
> > 
> > If we are in doubt, we should add a configure check and use it only when
> > available. And probably to default to &test_profil in all cases but
> > ppc64 BE which is known to fail.
> 
> I'm trying one other idea:
> 
> +static inline void *get_pc(void)
> +{
> +       extern void *profil01_label;
> +
> +       asm("profil01_label:");
> +       return &profil01_label;
> +}

It looks like I've accidentally deleted your reply to this email, sorry.

As both methods (the one with builtin and the one with asm()) seems to
work fine, feel free to proceed with either one of them.

-- 
Cyril Hrubis
chrubis@suse.cz

------------------------------------------------------------------------------
Slashdot TV.  
Video for Nerds.  Stuff that matters.
http://tv.slashdot.org/
_______________________________________________
Ltp-list mailing list
Ltp-list@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/ltp-list

Re: [LTP] [PATCH/RFC] profil01: don't use address of function

[chrubis]

Hi!
> > It looks like I've accidentally deleted your reply to this email, sorry.
> > 
> > As both methods (the one with builtin and the one with asm()) seems to
> > work fine, feel free to proceed with either one of them.
> 
> Pushed. (I went with the gcc builtin)

Tested that it fixed the failure I've been seeing on ia64.

-- 
Cyril Hrubis
chrubis@suse.cz

------------------------------------------------------------------------------
Slashdot TV.  
Video for Nerds.  Stuff that matters.
http://tv.slashdot.org/
_______________________________________________
Ltp-list mailing list
Ltp-list@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/ltp-list