[PATCH] perf: fix dwarf unwind using libunwind.

[PATCH] perf: fix dwarf unwind using libunwind.

[Wang Nan]

Original perf tool fails to unwind user stack if the event is issued in
a shared object:

 $ perf record -e syscalls:sys_enter_write -g --call-graph=dwarf babeltrace
 $ perf report --stdio

    # To display the perf.data header info, please use --header/--header-only options.
    #
    # Samples: 37  of event 'syscalls:sys_enter_write'
    # Event count (approx.): 37
    #
    # Children      Self  Command     Shared Object  Symbol
    # ........  ........  ..........  .............  .....................
    #
       100.00%   100.00%  babeltrace  libc-2.18.so   [.] __GI___libc_write
                |
                ---__GI___libc_write

By debugging libunwind I found that there is a bug in unwind-libunwind:
it always passes 0 as segbase to libunwind, cause libunwind unable to
locate debug_frame entry fir first level ip address (I add some more
debugging output into libunwind to make things clear):

               >_Uarm_dwarf_find_debug_frame: start_ip = 10be98, end_ip = 10c2a4
               >_Uarm_dwarf_find_debug_frame: found debug_frame table `/lib/libc-2.18.so': segbase=0x0, len=7, gp=0x0, table_data=0x449388
               >_Uarm_dwarf_search_unwind_table: call lookup:ip = b6cd3bcc, segbase = 0, rel_ip = b6cd3bcc
               >lookup: e->start_ip_offset = bcf18 (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 6d314 (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 33d0c (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 2ad6c (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 23004 (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 18f28 (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 184e8 (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 167cc (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 16294 (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 15d88 (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 15d0c (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 15c40 (rel_ip = b6cd3bcc)
 >_Uarm_dwarf_search_unwind_table: IP b6cd3bcc inside range b6c12000-b6d4c000, but no explicit unwind info found
                >put_rs_cache: unmasking signals/interrupts and releasing lock
               >_Uarm_dwarf_step: returning -10
 >_Uarm_step: dwarf_step()=-10

This patch passes map->start as segbase to dwarf_find_debug_frame(), so
di will be initialized correctly.

In addition, dso and executable are different when setting segbase. This
patch first check whether the elf is executable, and pass segbase only
for shared object.

Signed-off-by: Wang Nan <wangnan0@huawei.com>
---
 tools/perf/util/unwind-libunwind.c | 27 ++++++++++++++++++++++++++-
 1 file changed, 26 insertions(+), 1 deletion(-)

diff --git a/tools/perf/util/unwind-libunwind.c b/tools/perf/util/unwind-libunwind.c
index 371219a..32040fa 100644
--- a/tools/perf/util/unwind-libunwind.c
+++ b/tools/perf/util/unwind-libunwind.c
@@ -185,6 +185,27 @@ static u64 elf_section_offset(int fd, const char *name)
 	return offset;
 }
 
+static int elf_is_exec(int fd, const char *name)
+{
+	Elf *elf;
+	GElf_Ehdr ehdr;
+	int retval = 0;
+
+	elf = elf_begin(fd, PERF_ELF_C_READ_MMAP, NULL);
+	if (elf == NULL)
+		return 0;
+	if (gelf_getehdr(elf, &ehdr) == NULL)
+		goto out;
+
+	retval = (ehdr.e_type == ET_EXEC);
+
+out:
+	elf_end(elf);
+	pr_debug("unwind: elf_is_exec(%s): %d\n", name, retval);
+	return retval;
+
+}
+
 struct table_entry {
 	u32 start_ip_offset;
 	u32 fde_offset;
@@ -322,8 +343,12 @@ find_proc_info(unw_addr_space_t as, unw_word_t ip, unw_proc_info_t *pi,
 #ifndef NO_LIBUNWIND_DEBUG_FRAME
 	/* Check the .debug_frame section for unwinding info */
 	if (!read_unwind_spec_debug_frame(map->dso, ui->machine, &segbase)) {
+		int fd = dso__data_fd(map->dso, ui->machine);
+		int is_exec = elf_is_exec(fd, map->dso->name);
+		unw_word_t base = is_exec ? 0 : map->start;
+
 		memset(&di, 0, sizeof(di));
-		if (dwarf_find_debug_frame(0, &di, ip, 0, map->dso->name,
+		if (dwarf_find_debug_frame(0, &di, ip, base, map->dso->name,
 					   map->start, map->end))
 			return dwarf_search_unwind_table(as, ip, &di, pi,
 							 need_unwind_info, arg);
-- 
1.8.4

Re: [PATCH] perf: fix dwarf unwind using libunwind.

[Jiri Olsa]

On Tue, Jan 13, 2015 at 10:54:37AM +0800, Wang Nan wrote:

SNIP

>  
> +static int elf_is_exec(int fd, const char *name)
> +{
> +	Elf *elf;
> +	GElf_Ehdr ehdr;
> +	int retval = 0;
> +
> +	elf = elf_begin(fd, PERF_ELF_C_READ_MMAP, NULL);
> +	if (elf == NULL)
> +		return 0;
> +	if (gelf_getehdr(elf, &ehdr) == NULL)
> +		goto out;
> +
> +	retval = (ehdr.e_type == ET_EXEC);
> +
> +out:
> +	elf_end(elf);
> +	pr_debug("unwind: elf_is_exec(%s): %d\n", name, retval);
> +	return retval;
> +
> +}

yuou need to add stub for 'NO_LIBUNWIND_DEBUG_FRAME=1', otherwise:

  CC       util/dwarf-aux.o
  CC       util/unwind-libunwind.o
util/unwind-libunwind.c:188:12: error: ‘elf_is_exec’ defined but not used [-Werror=unused-function]
 static int elf_is_exec(int fd, const char *name)
            ^
cc1: all warnings being treated as errors
make[1]: *** [util/unwind-libunwind.o] Error 1
make[1]: *** Waiting for unfinished jobs....
make: *** [all] Error 2

jirka

> +
>  struct table_entry {
>  	u32 start_ip_offset;
>  	u32 fde_offset;
> @@ -322,8 +343,12 @@ find_proc_info(unw_addr_space_t as, unw_word_t ip, unw_proc_info_t *pi,
>  #ifndef NO_LIBUNWIND_DEBUG_FRAME
>  	/* Check the .debug_frame section for unwinding info */
>  	if (!read_unwind_spec_debug_frame(map->dso, ui->machine, &segbase)) {
> +		int fd = dso__data_fd(map->dso, ui->machine);
> +		int is_exec = elf_is_exec(fd, map->dso->name);
> +		unw_word_t base = is_exec ? 0 : map->start;
> +
>  		memset(&di, 0, sizeof(di));
> -		if (dwarf_find_debug_frame(0, &di, ip, 0, map->dso->name,
> +		if (dwarf_find_debug_frame(0, &di, ip, base, map->dso->name,
>  					   map->start, map->end))
>  			return dwarf_search_unwind_table(as, ip, &di, pi,
>  							 need_unwind_info, arg);
> -- 
> 1.8.4
> 

Re: [PATCH] perf: fix dwarf unwind using libunwind.

[Wang Nan]

On 2015/1/13 15:10, Jiri Olsa wrote:
> On Tue, Jan 13, 2015 at 10:54:37AM +0800, Wang Nan wrote:
> 
> SNIP
> 
>>  
>> +static int elf_is_exec(int fd, const char *name)
>> +{
>> +	Elf *elf;
>> +	GElf_Ehdr ehdr;
>> +	int retval = 0;
>> +
>> +	elf = elf_begin(fd, PERF_ELF_C_READ_MMAP, NULL);
>> +	if (elf == NULL)
>> +		return 0;
>> +	if (gelf_getehdr(elf, &ehdr) == NULL)
>> +		goto out;
>> +
>> +	retval = (ehdr.e_type == ET_EXEC);
>> +
>> +out:
>> +	elf_end(elf);
>> +	pr_debug("unwind: elf_is_exec(%s): %d\n", name, retval);
>> +	return retval;
>> +
>> +}
> 
> yuou need to add stub for 'NO_LIBUNWIND_DEBUG_FRAME=1', otherwise:
> 
>   CC       util/dwarf-aux.o
>   CC       util/unwind-libunwind.o
> util/unwind-libunwind.c:188:12: error: ‘elf_is_exec’ defined but not used [-Werror=unused-function]
>  static int elf_is_exec(int fd, const char *name)
>             ^
> cc1: all warnings being treated as errors
> make[1]: *** [util/unwind-libunwind.o] Error 1
> make[1]: *** Waiting for unfinished jobs....
> make: *** [all] Error 2
> 
> jirka
> 

Thanks. I haven't test NO_LIBUNWIND_DEBUG_FRAME=1 case.

I found that it is possible to utilize map->dso->adjust_symbols for this propose so
we don't need to introduce another function to check ehdr.e_type. However
map->dso->adjust_symbols checks not only ET_EXEC but also ET_REL. What do you think?

I'd like to check it and post another patch.

>> +
>>  struct table_entry {
>>  	u32 start_ip_offset;
>>  	u32 fde_offset;
>> @@ -322,8 +343,12 @@ find_proc_info(unw_addr_space_t as, unw_word_t ip, unw_proc_info_t *pi,
>>  #ifndef NO_LIBUNWIND_DEBUG_FRAME
>>  	/* Check the .debug_frame section for unwinding info */
>>  	if (!read_unwind_spec_debug_frame(map->dso, ui->machine, &segbase)) {
>> +		int fd = dso__data_fd(map->dso, ui->machine);
>> +		int is_exec = elf_is_exec(fd, map->dso->name);
>> +		unw_word_t base = is_exec ? 0 : map->start;
>> +
>>  		memset(&di, 0, sizeof(di));
>> -		if (dwarf_find_debug_frame(0, &di, ip, 0, map->dso->name,
>> +		if (dwarf_find_debug_frame(0, &di, ip, base, map->dso->name,
>>  					   map->start, map->end))
>>  			return dwarf_search_unwind_table(as, ip, &di, pi,
>>  							 need_unwind_info, arg);
>> -- 
>> 1.8.4
>>

Re: [PATCH] perf: fix dwarf unwind using libunwind.

[Jiri Olsa]

On Tue, Jan 13, 2015 at 04:36:53PM +0800, Wang Nan wrote:
> On 2015/1/13 15:10, Jiri Olsa wrote:
> > On Tue, Jan 13, 2015 at 10:54:37AM +0800, Wang Nan wrote:
> > 
> > SNIP
> > 
> >>  
> >> +static int elf_is_exec(int fd, const char *name)
> >> +{
> >> +	Elf *elf;
> >> +	GElf_Ehdr ehdr;
> >> +	int retval = 0;
> >> +
> >> +	elf = elf_begin(fd, PERF_ELF_C_READ_MMAP, NULL);
> >> +	if (elf == NULL)
> >> +		return 0;
> >> +	if (gelf_getehdr(elf, &ehdr) == NULL)
> >> +		goto out;
> >> +
> >> +	retval = (ehdr.e_type == ET_EXEC);
> >> +
> >> +out:
> >> +	elf_end(elf);
> >> +	pr_debug("unwind: elf_is_exec(%s): %d\n", name, retval);
> >> +	return retval;
> >> +
> >> +}
> > 
> > yuou need to add stub for 'NO_LIBUNWIND_DEBUG_FRAME=1', otherwise:
> > 
> >   CC       util/dwarf-aux.o
> >   CC       util/unwind-libunwind.o
> > util/unwind-libunwind.c:188:12: error: ‘elf_is_exec’ defined but not used [-Werror=unused-function]
> >  static int elf_is_exec(int fd, const char *name)
> >             ^
> > cc1: all warnings being treated as errors
> > make[1]: *** [util/unwind-libunwind.o] Error 1
> > make[1]: *** Waiting for unfinished jobs....
> > make: *** [all] Error 2
> > 
> > jirka
> > 
> 
> Thanks. I haven't test NO_LIBUNWIND_DEBUG_FRAME=1 case.
> 
> I found that it is possible to utilize map->dso->adjust_symbols for this propose so
> we don't need to introduce another function to check ehdr.e_type. However
> map->dso->adjust_symbols checks not only ET_EXEC but also ET_REL. What do you think?
> 
> I'd like to check it and post another patch.

please also update tests/mate

thanks,
jirka

[PATCH v2] perf: fix dwarf unwind using libunwind.

[Wang Nan]

Perf tool fails to unwind user stack if the event raises in a shared
object. This patch improves tests/dwarf-unwind.c to demonstrate the
problem by utilizing commonly used glibc function "bsearch". If perf
is not statically linked, the testcase will try to unwind a mixed call
trace.

By debugging libunwind I found that there is a bug in unwind-libunwind:
it always passes 0 as segbase to libunwind, cause libunwind unable to
locate debug_frame entry fir first level ip address (I add some more
debugging output into libunwind to make things clear):

               >_Uarm_dwarf_find_debug_frame: start_ip = 10be98, end_ip = 10c2a4
               >_Uarm_dwarf_find_debug_frame: found debug_frame table `/lib/libc-2.18.so': segbase=0x0, len=7, gp=0x0, table_data=0x449388
               >_Uarm_dwarf_search_unwind_table: call lookup:ip = b6cd3bcc, segbase = 0, rel_ip = b6cd3bcc
               >lookup: e->start_ip_offset = bcf18 (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 6d314 (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 33d0c (rel_ip = b6cd3bcc)
                ...
               >lookup: e->start_ip_offset = 15d0c (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 15c40 (rel_ip = b6cd3bcc)
 >_Uarm_dwarf_search_unwind_table: IP b6cd3bcc inside range b6c12000-b6d4c000, but no explicit unwind info found
                >put_rs_cache: unmasking signals/interrupts and releasing lock
               >_Uarm_dwarf_step: returning -10
 >_Uarm_step: dwarf_step()=-10

This patch passes map->start as segbase to dwarf_find_debug_frame(), so
di will be initialized correctly.

In addition, dso and executable are different when setting segbase. This
patch first check whether the elf is executable, and pass segbase only
for shared object.

Signed-off-by: Wang Nan <wangnan0@huawei.com>
---
 tools/perf/tests/dwarf-unwind.c    | 36 ++++++++++++++++++++++++++++++++++--
 tools/perf/util/unwind-libunwind.c | 28 +++++++++++++++++++++++++++-
 2 files changed, 61 insertions(+), 3 deletions(-)

diff --git a/tools/perf/tests/dwarf-unwind.c b/tools/perf/tests/dwarf-unwind.c
index ab28cca..0bf06be 100644
--- a/tools/perf/tests/dwarf-unwind.c
+++ b/tools/perf/tests/dwarf-unwind.c
@@ -11,6 +11,9 @@
 #include "thread.h"
 #include "callchain.h"
 
+/* For bsearch. We try to unwind functions in shared object. */
+#include <stdlib.h>
+
 static int mmap_handler(struct perf_tool *tool __maybe_unused,
 			union perf_event *event,
 			struct perf_sample *sample __maybe_unused,
@@ -28,7 +31,7 @@ static int init_live_machine(struct machine *machine)
 						  mmap_handler, machine, true);
 }
 
-#define MAX_STACK 6
+#define MAX_STACK 8
 
 static int unwind_entry(struct unwind_entry *entry, void *arg)
 {
@@ -37,6 +40,8 @@ static int unwind_entry(struct unwind_entry *entry, void *arg)
 	static const char *funcs[MAX_STACK] = {
 		"test__arch_unwind_sample",
 		"unwind_thread",
+		"compare",
+		"bsearch",
 		"krava_3",
 		"krava_2",
 		"krava_1",
@@ -88,10 +93,37 @@ static int unwind_thread(struct thread *thread)
 	return err;
 }
 
+static int global_unwind_retval = -INT_MAX;
+
+__attribute__ ((noinline))
+static int compare(void *p1, void *p2)
+{
+	/* Any possible value should be 'thread' */
+	struct thread *thread = *(struct thread **)p1;
+
+	if (global_unwind_retval == -INT_MAX)
+		global_unwind_retval = unwind_thread(thread);
+
+	return p1 - p2;
+}
+
 __attribute__ ((noinline))
 static int krava_3(struct thread *thread)
 {
-	return unwind_thread(thread);
+	struct thread *array[2] = {thread, thread};
+	void *fp = &bsearch;
+	/*
+	 * make _bsearch a volatile function pointer to
+	 * prevent potential optimization, which may expand
+	 * bsearch and call compare directly from this function,
+	 * instead of libc shared object.
+	 */
+	void *(*volatile _bsearch)(void *, void *, size_t,
+			size_t, int (*)(void *, void *));
+
+	_bsearch = fp;
+	_bsearch(array, &thread, 2, sizeof(struct thread **), compare);
+	return global_unwind_retval;
 }
 
 __attribute__ ((noinline))
diff --git a/tools/perf/util/unwind-libunwind.c b/tools/perf/util/unwind-libunwind.c
index 371219a..6edf535 100644
--- a/tools/perf/util/unwind-libunwind.c
+++ b/tools/perf/util/unwind-libunwind.c
@@ -185,6 +185,28 @@ static u64 elf_section_offset(int fd, const char *name)
 	return offset;
 }
 
+#ifndef NO_LIBUNWIND_DEBUG_FRAME
+static int elf_is_exec(int fd, const char *name)
+{
+	Elf *elf;
+	GElf_Ehdr ehdr;
+	int retval = 0;
+
+	elf = elf_begin(fd, PERF_ELF_C_READ_MMAP, NULL);
+	if (elf == NULL)
+		return 0;
+	if (gelf_getehdr(elf, &ehdr) == NULL)
+		goto out;
+
+	retval = (ehdr.e_type == ET_EXEC);
+
+out:
+	elf_end(elf);
+	pr_debug("unwind: elf_is_exec(%s): %d\n", name, retval);
+	return retval;
+}
+#endif
+
 struct table_entry {
 	u32 start_ip_offset;
 	u32 fde_offset;
@@ -322,8 +344,12 @@ find_proc_info(unw_addr_space_t as, unw_word_t ip, unw_proc_info_t *pi,
 #ifndef NO_LIBUNWIND_DEBUG_FRAME
 	/* Check the .debug_frame section for unwinding info */
 	if (!read_unwind_spec_debug_frame(map->dso, ui->machine, &segbase)) {
+		int fd = dso__data_fd(map->dso, ui->machine);
+		int is_exec = elf_is_exec(fd, map->dso->name);
+		unw_word_t base = is_exec ? 0 : map->start;
+
 		memset(&di, 0, sizeof(di));
-		if (dwarf_find_debug_frame(0, &di, ip, 0, map->dso->name,
+		if (dwarf_find_debug_frame(0, &di, ip, base, map->dso->name,
 					   map->start, map->end))
 			return dwarf_search_unwind_table(as, ip, &di, pi,
 							 need_unwind_info, arg);
-- 
1.8.4

Re: [PATCH v2] perf: fix dwarf unwind using libunwind.

[Jiri Olsa]

On Wed, Jan 14, 2015 at 10:36:47AM +0800, Wang Nan wrote:
> Perf tool fails to unwind user stack if the event raises in a shared

SNIP

> +
>  __attribute__ ((noinline))
>  static int krava_3(struct thread *thread)
>  {
> -	return unwind_thread(thread);
> +	struct thread *array[2] = {thread, thread};
> +	void *fp = &bsearch;
> +	/*
> +	 * make _bsearch a volatile function pointer to
> +	 * prevent potential optimization, which may expand
> +	 * bsearch and call compare directly from this function,
> +	 * instead of libc shared object.
> +	 */
> +	void *(*volatile _bsearch)(void *, void *, size_t,
> +			size_t, int (*)(void *, void *));
> +
> +	_bsearch = fp;
> +	_bsearch(array, &thread, 2, sizeof(struct thread **), compare);
> +	return global_unwind_retval;
>  }

ah, I've got confused with the NO_LIBUNWIND_DEBUG_FRAME name
and got the impression that we could use it in the tests/make
as another compile option test..

but your change is even better ;-) thanks for updating this test


Acked-by: Jiri Olsa <jolsa@kernel.org>


thanks,
jirka

Re: [PATCH v2] perf: fix dwarf unwind using libunwind.

[Arnaldo Carvalho de Melo]

Em Wed, Jan 14, 2015 at 12:57:49PM +0100, Jiri Olsa escreveu:
> On Wed, Jan 14, 2015 at 10:36:47AM +0800, Wang Nan wrote:
> > Perf tool fails to unwind user stack if the event raises in a shared
> 
> SNIP
> 
> > +
> >  __attribute__ ((noinline))
> >  static int krava_3(struct thread *thread)
> >  {
> > -	return unwind_thread(thread);
> > +	struct thread *array[2] = {thread, thread};
> > +	void *fp = &bsearch;
> > +	/*
> > +	 * make _bsearch a volatile function pointer to
> > +	 * prevent potential optimization, which may expand
> > +	 * bsearch and call compare directly from this function,
> > +	 * instead of libc shared object.
> > +	 */
> > +	void *(*volatile _bsearch)(void *, void *, size_t,
> > +			size_t, int (*)(void *, void *));
> > +
> > +	_bsearch = fp;
> > +	_bsearch(array, &thread, 2, sizeof(struct thread **), compare);
> > +	return global_unwind_retval;
> >  }
> 
> ah, I've got confused with the NO_LIBUNWIND_DEBUG_FRAME name
> and got the impression that we could use it in the tests/make
> as another compile option test..
> 
> but your change is even better ;-) thanks for updating this test
> 
> 
> Acked-by: Jiri Olsa <jolsa@kernel.org>

Thanks, applied to perf/urgent.

- Arnaldo

[tip:perf/urgent] perf test: Fix dwarf unwind using libunwind.

[tip-bot for Wang Nan]

Commit-ID:  b93b0967826a4a00297dde1cbbda8df673c1689e
Gitweb:     http://git.kernel.org/tip/b93b0967826a4a00297dde1cbbda8df673c1689e
Author:     Wang Nan <wangnan0@huawei.com>
AuthorDate: Wed, 14 Jan 2015 10:36:47 +0800
Committer:  Arnaldo Carvalho de Melo <acme@redhat.com>
CommitDate: Fri, 16 Jan 2015 17:49:29 -0300

perf test: Fix dwarf unwind using libunwind.

Perf tool fails to unwind user stack if the event raises in a shared
object. This patch improves tests/dwarf-unwind.c to demonstrate the
problem by utilizing commonly used glibc function "bsearch". If perf is
not statically linked, the testcase will try to unwind a mixed call
trace.

By debugging libunwind I found that there is a bug in unwind-libunwind:
it always passes 0 as segbase to libunwind, cause libunwind unable to
locate debug_frame entry fir first level ip address (I add some more
debugging output into libunwind to make things clear):

               >_Uarm_dwarf_find_debug_frame: start_ip = 10be98, end_ip = 10c2a4
               >_Uarm_dwarf_find_debug_frame: found debug_frame table `/lib/libc-2.18.so': segbase=0x0, len=7, gp=0x0, table_data=0x449388
               >_Uarm_dwarf_search_unwind_table: call lookup:ip = b6cd3bcc, segbase = 0, rel_ip = b6cd3bcc
               >lookup: e->start_ip_offset = bcf18 (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 6d314 (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 33d0c (rel_ip = b6cd3bcc)
                ...
               >lookup: e->start_ip_offset = 15d0c (rel_ip = b6cd3bcc)
               >lookup: e->start_ip_offset = 15c40 (rel_ip = b6cd3bcc)
 >_Uarm_dwarf_search_unwind_table: IP b6cd3bcc inside range b6c12000-b6d4c000, but no explicit unwind info found
                >put_rs_cache: unmasking signals/interrupts and releasing lock
               >_Uarm_dwarf_step: returning -10
 >_Uarm_step: dwarf_step()=-10

This patch passes map->start as segbase to dwarf_find_debug_frame(), so
di will be initialized correctly.

In addition, dso and executable are different when setting segbase. This
patch first check whether the elf is executable, and pass segbase only
for shared object.

Signed-off-by: Wang Nan <wangnan0@huawei.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Li Zefan <lizefan@huawei.com>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1421203007-75799-1-git-send-email-wangnan0@huawei.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
 tools/perf/tests/dwarf-unwind.c    | 36 ++++++++++++++++++++++++++++++++++--
 tools/perf/util/unwind-libunwind.c | 28 +++++++++++++++++++++++++++-
 2 files changed, 61 insertions(+), 3 deletions(-)

diff --git a/tools/perf/tests/dwarf-unwind.c b/tools/perf/tests/dwarf-unwind.c
index ab28cca..0bf06be 100644
--- a/tools/perf/tests/dwarf-unwind.c
+++ b/tools/perf/tests/dwarf-unwind.c
@@ -11,6 +11,9 @@
 #include "thread.h"
 #include "callchain.h"
 
+/* For bsearch. We try to unwind functions in shared object. */
+#include <stdlib.h>
+
 static int mmap_handler(struct perf_tool *tool __maybe_unused,
 			union perf_event *event,
 			struct perf_sample *sample __maybe_unused,
@@ -28,7 +31,7 @@ static int init_live_machine(struct machine *machine)
 						  mmap_handler, machine, true);
 }
 
-#define MAX_STACK 6
+#define MAX_STACK 8
 
 static int unwind_entry(struct unwind_entry *entry, void *arg)
 {
@@ -37,6 +40,8 @@ static int unwind_entry(struct unwind_entry *entry, void *arg)
 	static const char *funcs[MAX_STACK] = {
 		"test__arch_unwind_sample",
 		"unwind_thread",
+		"compare",
+		"bsearch",
 		"krava_3",
 		"krava_2",
 		"krava_1",
@@ -88,10 +93,37 @@ static int unwind_thread(struct thread *thread)
 	return err;
 }
 
+static int global_unwind_retval = -INT_MAX;
+
+__attribute__ ((noinline))
+static int compare(void *p1, void *p2)
+{
+	/* Any possible value should be 'thread' */
+	struct thread *thread = *(struct thread **)p1;
+
+	if (global_unwind_retval == -INT_MAX)
+		global_unwind_retval = unwind_thread(thread);
+
+	return p1 - p2;
+}
+
 __attribute__ ((noinline))
 static int krava_3(struct thread *thread)
 {
-	return unwind_thread(thread);
+	struct thread *array[2] = {thread, thread};
+	void *fp = &bsearch;
+	/*
+	 * make _bsearch a volatile function pointer to
+	 * prevent potential optimization, which may expand
+	 * bsearch and call compare directly from this function,
+	 * instead of libc shared object.
+	 */
+	void *(*volatile _bsearch)(void *, void *, size_t,
+			size_t, int (*)(void *, void *));
+
+	_bsearch = fp;
+	_bsearch(array, &thread, 2, sizeof(struct thread **), compare);
+	return global_unwind_retval;
 }
 
 __attribute__ ((noinline))
diff --git a/tools/perf/util/unwind-libunwind.c b/tools/perf/util/unwind-libunwind.c
index 371219a..6edf535 100644
--- a/tools/perf/util/unwind-libunwind.c
+++ b/tools/perf/util/unwind-libunwind.c
@@ -185,6 +185,28 @@ static u64 elf_section_offset(int fd, const char *name)
 	return offset;
 }
 
+#ifndef NO_LIBUNWIND_DEBUG_FRAME
+static int elf_is_exec(int fd, const char *name)
+{
+	Elf *elf;
+	GElf_Ehdr ehdr;
+	int retval = 0;
+
+	elf = elf_begin(fd, PERF_ELF_C_READ_MMAP, NULL);
+	if (elf == NULL)
+		return 0;
+	if (gelf_getehdr(elf, &ehdr) == NULL)
+		goto out;
+
+	retval = (ehdr.e_type == ET_EXEC);
+
+out:
+	elf_end(elf);
+	pr_debug("unwind: elf_is_exec(%s): %d\n", name, retval);
+	return retval;
+}
+#endif
+
 struct table_entry {
 	u32 start_ip_offset;
 	u32 fde_offset;
@@ -322,8 +344,12 @@ find_proc_info(unw_addr_space_t as, unw_word_t ip, unw_proc_info_t *pi,
 #ifndef NO_LIBUNWIND_DEBUG_FRAME
 	/* Check the .debug_frame section for unwinding info */
 	if (!read_unwind_spec_debug_frame(map->dso, ui->machine, &segbase)) {
+		int fd = dso__data_fd(map->dso, ui->machine);
+		int is_exec = elf_is_exec(fd, map->dso->name);
+		unw_word_t base = is_exec ? 0 : map->start;
+
 		memset(&di, 0, sizeof(di));
-		if (dwarf_find_debug_frame(0, &di, ip, 0, map->dso->name,
+		if (dwarf_find_debug_frame(0, &di, ip, base, map->dso->name,
 					   map->start, map->end))
 			return dwarf_search_unwind_table(as, ip, &di, pi,
 							 need_unwind_info, arg);