[PATCH 0/3] perf tools: Add libdw DWARF post unwind support for ARM

[PATCH 0/3] perf tools: Add libdw DWARF post unwind support for ARM

[Jean Pihet]

Adding libdw DWARF post unwind support, which is part
of elfutils-devel/libdw-dev package from version 0.158.

Also includes the test suite for dwarf unwinding, by adding the
arch specific test code and the perf_regs_load function.

Jean Pihet (3):
  perf tests: Introduce perf_regs_load function on ARM
  perf tests: Add dwarf unwind test on ARM
  perf tools: Add libdw DWARF post unwind support for ARM

 tools/perf/Makefile.perf                 |  2 +-
 tools/perf/arch/arm/Makefile             |  7 ++++
 tools/perf/arch/arm/include/perf_regs.h  |  5 +++
 tools/perf/arch/arm/tests/dwarf-unwind.c | 59 +++++++++++++++++++++++++++++++
 tools/perf/arch/arm/tests/regs_load.S    | 60 ++++++++++++++++++++++++++++++++
 tools/perf/arch/arm/util/unwind-libdw.c  | 36 +++++++++++++++++++
 tools/perf/tests/builtin-test.c          |  2 +-
 tools/perf/tests/tests.h                 |  2 +-
 8 files changed, 170 insertions(+), 3 deletions(-)
 create mode 100644 tools/perf/arch/arm/tests/dwarf-unwind.c
 create mode 100644 tools/perf/arch/arm/tests/regs_load.S
 create mode 100644 tools/perf/arch/arm/util/unwind-libdw.c

---
Rebased on the latest jolsa/perf/core

-- 
1.7.11.7

[PATCH 1/3] perf tests: Introduce perf_regs_load function on ARM

[Jean Pihet]

Introducing perf_regs_load function, which is going
to be used for dwarf unwind test in following patches.

It takes single argument as a pointer to the regs dump
buffer and populates it with current registers values.

Signed-off-by: Jean Pihet <jean.pihet@linaro.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
---
 tools/perf/arch/arm/Makefile            |  1 +
 tools/perf/arch/arm/include/perf_regs.h |  2 ++
 tools/perf/arch/arm/tests/regs_load.S   | 60 +++++++++++++++++++++++++++++++++
 3 files changed, 63 insertions(+)
 create mode 100644 tools/perf/arch/arm/tests/regs_load.S

diff --git a/tools/perf/arch/arm/Makefile b/tools/perf/arch/arm/Makefile
index 67e9b3d..9b8f87e 100644
--- a/tools/perf/arch/arm/Makefile
+++ b/tools/perf/arch/arm/Makefile
@@ -4,4 +4,5 @@ LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/dwarf-regs.o
 endif
 ifndef NO_LIBUNWIND
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/unwind-libunwind.o
+LIB_OBJS += $(OUTPUT)arch/$(ARCH)/tests/regs_load.o
 endif
diff --git a/tools/perf/arch/arm/include/perf_regs.h b/tools/perf/arch/arm/include/perf_regs.h
index 2a1cfde..1476ae7 100644
--- a/tools/perf/arch/arm/include/perf_regs.h
+++ b/tools/perf/arch/arm/include/perf_regs.h
@@ -5,6 +5,8 @@
 #include "../../util/types.h"
 #include <asm/perf_regs.h>
 
+void perf_regs_load(u64 *regs);
+
 #define PERF_REGS_MASK	((1ULL << PERF_REG_ARM_MAX) - 1)
 #define PERF_REG_IP	PERF_REG_ARM_PC
 #define PERF_REG_SP	PERF_REG_ARM_SP
diff --git a/tools/perf/arch/arm/tests/regs_load.S b/tools/perf/arch/arm/tests/regs_load.S
new file mode 100644
index 0000000..66a10b6
--- /dev/null
+++ b/tools/perf/arch/arm/tests/regs_load.S
@@ -0,0 +1,60 @@
+#include <linux/linkage.h>
+#include <asm/assembler.h>
+
+#define R0 0x00
+#define R1 0x08
+#define R2 0x10
+#define R3 0x18
+#define R4 0x20
+#define R5 0x28
+#define R6 0x30
+#define R7 0x38
+#define R8 0x40
+#define R9 0x48
+#define SL 0x50
+#define FP 0x58
+#define IP 0x60
+#define SP 0x68
+#define LR 0x70
+#define PC 0x78
+
+/*
+ * Implementation of void perf_regs_load(u64 *regs);
+ *
+ * This functions fills in the 'regs' buffer from the actual registers values,
+ * in the way the perf built-in unwinding test expects them:
+ * - the return values (i.e. caller values) of fp, sp and lr are retrieved
+ *   and stored from the fp, in order to skip the call to this function.
+ * The built-in unwinding test then unwinds the call stack from the dwarf
+ * information in unwind__get_entries.
+ *
+ * Notes:
+ * - the 8 bytes stride in the registers offsets comes from the fact
+ * that the registers are stored in an u64 array (u64 *regs),
+ * - since perf is built with -fno-omit-frame-pointer, the fp value can
+ * be used to retrieve the caller values of fp, sp, lr.
+ */
+
+.text
+ENTRY(perf_regs_load)
+	str r0, [r0, #R0]
+	str r1, [r0, #R1]
+	str r2, [r0, #R2]
+	str r3, [r0, #R3]
+	str r4, [r0, #R4]
+	str r5, [r0, #R5]
+	str r6, [r0, #R6]
+	str r7, [r0, #R7]
+	str r8, [r0, #R8]
+	str r9, [r0, #R9]
+	str sl, [r0, #SL]
+	ldr r2, [fp, #-12]	@ retrieve and save return fp
+	str r2, [r0, #FP]
+	str ip, [r0, #IP]
+	ldr r2, [fp, #-8]	@ retrieve and save return sp
+	str r2, [r0, #SP]
+	ldr r2, [fp, #-4]	@ retrieve and save return lr
+	str r2, [r0, #LR]
+	str pc, [r0, #PC]
+	mov pc, lr
+ENDPROC(perf_regs_load)
-- 
1.7.11.7

[PATCH 2/3] perf tests: Add dwarf unwind test on ARM

[Jean Pihet]

Adding dwarf unwind test, that setups live machine data over
the perf test thread and does the remote unwind.

Need to use -fno-optimize-sibling-calls for test compilation,
otherwise 'krava_*' function calls are optimized into jumps
and ommited from the stack unwind.

So far it was enabled only for x86.

Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: David Ahern <dsahern@gmail.com>
Signed-off-by: Jean Pihet <jean.pihet@linaro.org>
---
 tools/perf/Makefile.perf                 |  2 +-
 tools/perf/arch/arm/Makefile             |  1 +
 tools/perf/arch/arm/include/perf_regs.h  |  3 ++
 tools/perf/arch/arm/tests/dwarf-unwind.c | 59 ++++++++++++++++++++++++++++++++
 tools/perf/tests/builtin-test.c          |  2 +-
 tools/perf/tests/tests.h                 |  2 +-
 6 files changed, 66 insertions(+), 3 deletions(-)
 create mode 100644 tools/perf/arch/arm/tests/dwarf-unwind.c

diff --git a/tools/perf/Makefile.perf b/tools/perf/Makefile.perf
index 2baf61c..dea2d633 100644
--- a/tools/perf/Makefile.perf
+++ b/tools/perf/Makefile.perf
@@ -411,7 +411,7 @@ LIB_OBJS += $(OUTPUT)tests/code-reading.o
 LIB_OBJS += $(OUTPUT)tests/sample-parsing.o
 LIB_OBJS += $(OUTPUT)tests/parse-no-sample-id-all.o
 ifndef NO_DWARF_UNWIND
-ifeq ($(ARCH),x86)
+ifeq ($(ARCH),$(filter $(ARCH),x86 arm))
 LIB_OBJS += $(OUTPUT)tests/dwarf-unwind.o
 endif
 endif
diff --git a/tools/perf/arch/arm/Makefile b/tools/perf/arch/arm/Makefile
index 9b8f87e..221f21d 100644
--- a/tools/perf/arch/arm/Makefile
+++ b/tools/perf/arch/arm/Makefile
@@ -5,4 +5,5 @@ endif
 ifndef NO_LIBUNWIND
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/unwind-libunwind.o
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/tests/regs_load.o
+LIB_OBJS += $(OUTPUT)arch/$(ARCH)/tests/dwarf-unwind.o
 endif
diff --git a/tools/perf/arch/arm/include/perf_regs.h b/tools/perf/arch/arm/include/perf_regs.h
index 1476ae7..c2cefff 100644
--- a/tools/perf/arch/arm/include/perf_regs.h
+++ b/tools/perf/arch/arm/include/perf_regs.h
@@ -8,6 +8,9 @@
 void perf_regs_load(u64 *regs);
 
 #define PERF_REGS_MASK	((1ULL << PERF_REG_ARM_MAX) - 1)
+#define PERF_REGS_MAX	PERF_REG_ARM_MAX
+#define PERF_SAMPLE_REGS_ABI	PERF_SAMPLE_REGS_ABI_32
+
 #define PERF_REG_IP	PERF_REG_ARM_PC
 #define PERF_REG_SP	PERF_REG_ARM_SP
 
diff --git a/tools/perf/arch/arm/tests/dwarf-unwind.c b/tools/perf/arch/arm/tests/dwarf-unwind.c
new file mode 100644
index 0000000..d618f5f
--- /dev/null
+++ b/tools/perf/arch/arm/tests/dwarf-unwind.c
@@ -0,0 +1,59 @@
+#include <string.h>
+#include "perf_regs.h"
+#include "thread.h"
+#include "map.h"
+#include "event.h"
+#include "tests/tests.h"
+
+#define STACK_SIZE 8192
+
+static int sample_ustack(struct perf_sample *sample,
+			 struct thread *thread, u64 *regs)
+{
+	struct stack_dump *stack = &sample->user_stack;
+	struct map *map;
+	unsigned long sp;
+	u64 stack_size, *buf;
+
+	buf = malloc(STACK_SIZE);
+	if (!buf) {
+		pr_debug("failed to allocate sample uregs data\n");
+		return -1;
+	}
+
+	sp = (unsigned long) regs[PERF_REG_ARM_SP];
+
+	map = map_groups__find(&thread->mg, MAP__FUNCTION, (u64) sp);
+	if (!map) {
+		pr_debug("failed to get stack map\n");
+		return -1;
+	}
+
+	stack_size = map->end - sp;
+	stack_size = stack_size > STACK_SIZE ? STACK_SIZE : stack_size;
+
+	memcpy(buf, (void *) sp, stack_size);
+	stack->data = (char *) buf;
+	stack->size = stack_size;
+	return 0;
+}
+
+int test__arch_unwind_sample(struct perf_sample *sample,
+			     struct thread *thread)
+{
+	struct regs_dump *regs = &sample->user_regs;
+	u64 *buf;
+
+	buf = malloc(sizeof(u64) * PERF_REGS_MAX);
+	if (!buf) {
+		pr_debug("failed to allocate sample uregs data\n");
+		return -1;
+	}
+
+	perf_regs_load(buf);
+	regs->abi  = PERF_SAMPLE_REGS_ABI;
+	regs->regs = buf;
+	regs->mask = PERF_REGS_MASK;
+
+	return sample_ustack(sample, thread, buf);
+}
diff --git a/tools/perf/tests/builtin-test.c b/tools/perf/tests/builtin-test.c
index 0d5afaf..5e0764b 100644
--- a/tools/perf/tests/builtin-test.c
+++ b/tools/perf/tests/builtin-test.c
@@ -115,7 +115,7 @@ static struct test {
 		.desc = "Test parsing with no sample_id_all bit set",
 		.func = test__parse_no_sample_id_all,
 	},
-#if defined(__x86_64__) || defined(__i386__)
+#if defined(__x86_64__) || defined(__i386__) || defined(__arm__)
 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 	{
 		.desc = "Test dwarf unwind",
diff --git a/tools/perf/tests/tests.h b/tools/perf/tests/tests.h
index a9d7cb0..8f91fb0 100644
--- a/tools/perf/tests/tests.h
+++ b/tools/perf/tests/tests.h
@@ -45,7 +45,7 @@ int test__hists_filter(void);
 int test__mmap_thread_lookup(void);
 int test__thread_mg_share(void);
 
-#if defined(__x86_64__) || defined(__i386__)
+#if defined(__x86_64__) || defined(__i386__) || defined(__arm__)
 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 struct thread;
 struct perf_sample;
-- 
1.7.11.7

[PATCH 2/3] perf tests: Add dwarf unwind test on ARM

[Jiri Olsa]

On Tue, May 06, 2014 at 05:26:18PM +0200, Jean Pihet wrote:

SNIP

> diff --git a/tools/perf/arch/arm/tests/dwarf-unwind.c b/tools/perf/arch/arm/tests/dwarf-unwind.c
> new file mode 100644
> index 0000000..d618f5f
> --- /dev/null
> +++ b/tools/perf/arch/arm/tests/dwarf-unwind.c
> @@ -0,0 +1,59 @@
> +#include <string.h>
> +#include "perf_regs.h"
> +#include "thread.h"
> +#include "map.h"
> +#include "event.h"
> +#include "tests/tests.h"
> +
> +#define STACK_SIZE 8192
> +
> +static int sample_ustack(struct perf_sample *sample,
> +			 struct thread *thread, u64 *regs)
> +{
> +	struct stack_dump *stack = &sample->user_stack;
> +	struct map *map;
> +	unsigned long sp;
> +	u64 stack_size, *buf;
> +
> +	buf = malloc(STACK_SIZE);
> +	if (!buf) {
> +		pr_debug("failed to allocate sample uregs data\n");
> +		return -1;
> +	}
> +
> +	sp = (unsigned long) regs[PERF_REG_ARM_SP];
> +
> +	map = map_groups__find(&thread->mg, MAP__FUNCTION, (u64) sp);
> +	if (!map) {
> +		pr_debug("failed to get stack map\n");
> +		return -1;
> +	}

there's a memory leak of 'buf' already fixed fox x86:

  perf tests x86: Fix memory leak in sample_ustack()
  commit 763d7f5f2718f085bab5a9e63308349728f3ad12
  Author: Masanari Iida <standby24x7@gmail.com>
  Date:   Sun Apr 20 00:16:41 2014 +0900

jirka

[PATCH 2/3] perf tests: Add dwarf unwind test on ARM

[Jean Pihet]

Hi Jiri,

On 7 May 2014 14:06, Jiri Olsa <jolsa@redhat.com> wrote:
> On Tue, May 06, 2014 at 05:26:18PM +0200, Jean Pihet wrote:
>
SNIP
>
> there's a memory leak of 'buf' already fixed fox x86:
>
>   perf tests x86: Fix memory leak in sample_ustack()
>   commit 763d7f5f2718f085bab5a9e63308349728f3ad12
>   Author: Masanari Iida <standby24x7@gmail.com>
>   Date:   Sun Apr 20 00:16:41 2014 +0900
>
> jirka

Ok

Here is the diff between the x86 and the ARM implementations:
$ diff -urN tools/perf/arch/arm64/tests/dwarf-unwind.c
tools/perf/arch/x86/tests/dwarf-unwind.c
--- tools/perf/arch/arm64/tests/dwarf-unwind.c    2014-05-06
17:31:17.507961045 +0200
+++ tools/perf/arch/x86/tests/dwarf-unwind.c    2014-05-06
16:52:00.589776839 +0200
@@ -21,11 +21,12 @@
         return -1;
     }

-    sp = (unsigned long) regs[PERF_REG_ARM64_SP];
+    sp = (unsigned long) regs[PERF_REG_X86_SP];

-    map = map_groups__find(&thread->mg, MAP__FUNCTION, (u64) sp);
+    map = map_groups__find(thread->mg, MAP__VARIABLE, (u64) sp);
     if (!map) {
         pr_debug("failed to get stack map\n");
+        free(buf);
         return -1;
     }

Which leads to a few questions:
- the map_groups__find parameters need to be fixed too, right?
- the free(buf) needs to be fixed,
- given that the remaining difference in the file is just a register
macro, it is worth to factor the code in a single file. Does that make
sense? If worthwhile I can do that once the ARM and ARM64 support is
merged in.

What do you think?

Regards,
Jean

[PATCH 2/3] perf tests: Add dwarf unwind test on ARM

[Jiri Olsa]

On Wed, May 07, 2014 at 02:24:53PM +0200, Jean Pihet wrote:
> Hi Jiri,
> 
> On 7 May 2014 14:06, Jiri Olsa <jolsa@redhat.com> wrote:
> > On Tue, May 06, 2014 at 05:26:18PM +0200, Jean Pihet wrote:
> >
> SNIP
> >
> > there's a memory leak of 'buf' already fixed fox x86:
> >
> >   perf tests x86: Fix memory leak in sample_ustack()
> >   commit 763d7f5f2718f085bab5a9e63308349728f3ad12
> >   Author: Masanari Iida <standby24x7@gmail.com>
> >   Date:   Sun Apr 20 00:16:41 2014 +0900
> >
> > jirka
> 
> Ok
> 
> Here is the diff between the x86 and the ARM implementations:
> $ diff -urN tools/perf/arch/arm64/tests/dwarf-unwind.c
> tools/perf/arch/x86/tests/dwarf-unwind.c
> --- tools/perf/arch/arm64/tests/dwarf-unwind.c    2014-05-06
> 17:31:17.507961045 +0200
> +++ tools/perf/arch/x86/tests/dwarf-unwind.c    2014-05-06
> 16:52:00.589776839 +0200
> @@ -21,11 +21,12 @@
>          return -1;
>      }
> 
> -    sp = (unsigned long) regs[PERF_REG_ARM64_SP];
> +    sp = (unsigned long) regs[PERF_REG_X86_SP];
> 
> -    map = map_groups__find(&thread->mg, MAP__FUNCTION, (u64) sp);
> +    map = map_groups__find(thread->mg, MAP__VARIABLE, (u64) sp);
>      if (!map) {
>          pr_debug("failed to get stack map\n");
> +        free(buf);
>          return -1;
>      }
> 
> Which leads to a few questions:
> - the map_groups__find parameters need to be fixed too, right?

the reason for this is following commit:
  6392b4e perf x86: Fix perf to use non-executable stack, again

which also adds global link flags: -Wl,-z,noexecstack 
so I'm guessing arm is affected too

> - the free(buf) needs to be fixed,
> - given that the remaining difference in the file is just a register
> macro, it is worth to factor the code in a single file. Does that make
> sense? If worthwhile I can do that once the ARM and ARM64 support is
> merged in.

we can do the code factoring later here, np

thanks,
jirka

[PATCH 3/3] perf tools: Add libdw DWARF post unwind support for ARM

[Jean Pihet]

Adding libdw DWARF post unwind support, which is part
of elfutils-devel/libdw-dev package from version 0.158.

The new code is contained in unwin-libdw.c object, and
implements unwind__get_entries unwind interface function.

Signed-off-by: Jean Pihet <jean.pihet@linaro.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: David Ahern <dsahern@gmail.com>
---
 tools/perf/arch/arm/Makefile            |  5 +++++
 tools/perf/arch/arm/util/unwind-libdw.c | 36 +++++++++++++++++++++++++++++++++
 2 files changed, 41 insertions(+)
 create mode 100644 tools/perf/arch/arm/util/unwind-libdw.c

diff --git a/tools/perf/arch/arm/Makefile b/tools/perf/arch/arm/Makefile
index 221f21d..09d6215 100644
--- a/tools/perf/arch/arm/Makefile
+++ b/tools/perf/arch/arm/Makefile
@@ -4,6 +4,11 @@ LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/dwarf-regs.o
 endif
 ifndef NO_LIBUNWIND
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/unwind-libunwind.o
+endif
+ifndef NO_LIBDW_DWARF_UNWIND
+LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/unwind-libdw.o
+endif
+ifndef NO_DWARF_UNWIND
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/tests/regs_load.o
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/tests/dwarf-unwind.o
 endif
diff --git a/tools/perf/arch/arm/util/unwind-libdw.c b/tools/perf/arch/arm/util/unwind-libdw.c
new file mode 100644
index 0000000..b4176c6
--- /dev/null
+++ b/tools/perf/arch/arm/util/unwind-libdw.c
@@ -0,0 +1,36 @@
+#include <elfutils/libdwfl.h>
+#include "../../util/unwind-libdw.h"
+#include "../../util/perf_regs.h"
+
+bool libdw__arch_set_initial_registers(Dwfl_Thread *thread, void *arg)
+{
+	struct unwind_info *ui = arg;
+	struct regs_dump *user_regs = &ui->sample->user_regs;
+	Dwarf_Word dwarf_regs[PERF_REG_ARM_MAX];
+
+#define REG(r) ({						\
+	Dwarf_Word val = 0;					\
+	perf_reg_value(&val, user_regs, PERF_REG_ARM_##r);	\
+	val;							\
+})
+
+	dwarf_regs[0]  = REG(R0);
+	dwarf_regs[1]  = REG(R1);
+	dwarf_regs[2]  = REG(R2);
+	dwarf_regs[3]  = REG(R3);
+	dwarf_regs[4]  = REG(R4);
+	dwarf_regs[5]  = REG(R5);
+	dwarf_regs[6]  = REG(R6);
+	dwarf_regs[7]  = REG(R7);
+	dwarf_regs[8]  = REG(R8);
+	dwarf_regs[9]  = REG(R9);
+	dwarf_regs[10] = REG(R10);
+	dwarf_regs[11] = REG(FP);
+	dwarf_regs[12] = REG(IP);
+	dwarf_regs[13] = REG(SP);
+	dwarf_regs[14] = REG(LR);
+	dwarf_regs[15] = REG(PC);
+
+	return dwfl_thread_state_registers(thread, 0, PERF_REG_ARM_MAX,
+					   dwarf_regs);
+}
-- 
1.7.11.7

[PATCH 3/3] perf tools: Add libdw DWARF post unwind support for ARM

[Will Deacon]

On Tue, May 06, 2014 at 04:26:19PM +0100, Jean Pihet wrote:
> Adding libdw DWARF post unwind support, which is part
> of elfutils-devel/libdw-dev package from version 0.158.
> 
> The new code is contained in unwin-libdw.c object, and
> implements unwind__get_entries unwind interface function.

  Reviewed-by: Will Deacon <will.deacon@arm.com>

Will

[PATCH 3/3] perf tools: Add libdw DWARF post unwind support for ARM

[Jean Pihet]

On 6 May 2014 19:56, Will Deacon <will.deacon@arm.com> wrote:
> On Tue, May 06, 2014 at 04:26:19PM +0100, Jean Pihet wrote:
>> Adding libdw DWARF post unwind support, which is part
>> of elfutils-devel/libdw-dev package from version 0.158.
>>
>> The new code is contained in unwin-libdw.c object, and
>> implements unwind__get_entries unwind interface function.
>
>   Reviewed-by: Will Deacon <will.deacon@arm.com>

Ok, thx!

Jiri, Arnaldo,
Can you take the series for ARM and ARM64?

Jean
>
> Will

[PATCH 3/3] perf tools: Add libdw DWARF post unwind support for ARM

[Will Deacon]

On Wed, May 07, 2014 at 10:52:05AM +0100, Jean Pihet wrote:
> On 6 May 2014 19:56, Will Deacon <will.deacon@arm.com> wrote:
> > On Tue, May 06, 2014 at 04:26:19PM +0100, Jean Pihet wrote:
> >> Adding libdw DWARF post unwind support, which is part
> >> of elfutils-devel/libdw-dev package from version 0.158.
> >>
> >> The new code is contained in unwin-libdw.c object, and
> >> implements unwind__get_entries unwind interface function.
> >
> >   Reviewed-by: Will Deacon <will.deacon@arm.com>
> 
> Ok, thx!
> 
> Jiri, Arnaldo,
> Can you take the series for ARM and ARM64?

Wait: I thought you were going to add compat support for arm64? I haven't
acked that patch yet, because it's lacking those changes.

The 32-bit ARM patches look fine though.

Will

[PATCH 3/3] perf tools: Add libdw DWARF post unwind support for ARM

[Jean Pihet]

On 7 May 2014 12:00, Will Deacon <will.deacon@arm.com> wrote:
> On Wed, May 07, 2014 at 10:52:05AM +0100, Jean Pihet wrote:
>> On 6 May 2014 19:56, Will Deacon <will.deacon@arm.com> wrote:
>> > On Tue, May 06, 2014 at 04:26:19PM +0100, Jean Pihet wrote:
>> >> Adding libdw DWARF post unwind support, which is part
>> >> of elfutils-devel/libdw-dev package from version 0.158.
>> >>
>> >> The new code is contained in unwin-libdw.c object, and
>> >> implements unwind__get_entries unwind interface function.
>> >
>> >   Reviewed-by: Will Deacon <will.deacon@arm.com>
>>
>> Ok, thx!
>>
>> Jiri, Arnaldo,
>> Can you take the series for ARM and ARM64?
>
> Wait: I thought you were going to add compat support for arm64? I haven't
> acked that patch yet, because it's lacking those changes.
The compat mode is supported by libdw and _should_ work in principle.
I am following it up with Mark Wielaard on elfutils-devel.
Cf. [1] for the discussion and [2] for the compat mode support in libdw.

[1] https://lists.fedorahosted.org/pipermail/elfutils-devel/2014-March/003905.html
[2] https://lists.fedorahosted.org/pipermail/elfutils-devel/2014-March/003910.html

So in summary the kernel changes should be fine (it works ok with
libunwind), the problem lies somewhere in the libdw registers setup
and usage for compat mode.

>
> The 32-bit ARM patches look fine though.
Thx!

>
> Will

Jean

[PATCH 1/4] tools: perf: consolidate types.h for ARM and ARM64

[Jean Pihet]

Prevents a build breakage since commit d944c4eebcf4c0d5e5d9728fec110cbf0047ad7f
'tools: Consolidate types.h'

Signed-off-by: Jean Pihet <jean.pihet@linaro.org>
---
 tools/perf/arch/arm/include/perf_regs.h   | 2 +-
 tools/perf/arch/arm64/include/perf_regs.h | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/tools/perf/arch/arm/include/perf_regs.h b/tools/perf/arch/arm/include/perf_regs.h
index 2a1cfde..bcca511 100644
--- a/tools/perf/arch/arm/include/perf_regs.h
+++ b/tools/perf/arch/arm/include/perf_regs.h
@@ -2,7 +2,7 @@
 #define ARCH_PERF_REGS_H
 
 #include <stdlib.h>
-#include "../../util/types.h"
+#include <linux/types.h>
 #include <asm/perf_regs.h>
 
 #define PERF_REGS_MASK	((1ULL << PERF_REG_ARM_MAX) - 1)
diff --git a/tools/perf/arch/arm64/include/perf_regs.h b/tools/perf/arch/arm64/include/perf_regs.h
index 2359546..e9441b9 100644
--- a/tools/perf/arch/arm64/include/perf_regs.h
+++ b/tools/perf/arch/arm64/include/perf_regs.h
@@ -2,7 +2,7 @@
 #define ARCH_PERF_REGS_H
 
 #include <stdlib.h>
-#include "../../util/types.h"
+#include <linux/types.h>
 #include <asm/perf_regs.h>
 
 #define PERF_REGS_MASK	((1ULL << PERF_REG_ARM64_MAX) - 1)
-- 
1.7.11.7

[PATCH 2/4] perf tests: Introduce perf_regs_load function on ARM

[Jean Pihet]

Introducing perf_regs_load function, which is going
to be used for dwarf unwind test in following patches.

It takes single argument as a pointer to the regs dump
buffer and populates it with current registers values.

Signed-off-by: Jean Pihet <jean.pihet@linaro.org>
Reviewed-by: Will Deacon <will.deacon@arm.com>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
---
 tools/perf/arch/arm/Makefile            |  1 +
 tools/perf/arch/arm/include/perf_regs.h |  2 ++
 tools/perf/arch/arm/tests/regs_load.S   | 58 +++++++++++++++++++++++++++++++++
 3 files changed, 61 insertions(+)
 create mode 100644 tools/perf/arch/arm/tests/regs_load.S

diff --git a/tools/perf/arch/arm/Makefile b/tools/perf/arch/arm/Makefile
index 67e9b3d..9b8f87e 100644
--- a/tools/perf/arch/arm/Makefile
+++ b/tools/perf/arch/arm/Makefile
@@ -4,4 +4,5 @@ LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/dwarf-regs.o
 endif
 ifndef NO_LIBUNWIND
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/unwind-libunwind.o
+LIB_OBJS += $(OUTPUT)arch/$(ARCH)/tests/regs_load.o
 endif
diff --git a/tools/perf/arch/arm/include/perf_regs.h b/tools/perf/arch/arm/include/perf_regs.h
index bcca511..33abcfa 100644
--- a/tools/perf/arch/arm/include/perf_regs.h
+++ b/tools/perf/arch/arm/include/perf_regs.h
@@ -5,6 +5,8 @@
 #include <linux/types.h>
 #include <asm/perf_regs.h>
 
+void perf_regs_load(u64 *regs);
+
 #define PERF_REGS_MASK	((1ULL << PERF_REG_ARM_MAX) - 1)
 #define PERF_REG_IP	PERF_REG_ARM_PC
 #define PERF_REG_SP	PERF_REG_ARM_SP
diff --git a/tools/perf/arch/arm/tests/regs_load.S b/tools/perf/arch/arm/tests/regs_load.S
new file mode 100644
index 0000000..e09e983
--- /dev/null
+++ b/tools/perf/arch/arm/tests/regs_load.S
@@ -0,0 +1,58 @@
+#include <linux/linkage.h>
+
+#define R0 0x00
+#define R1 0x08
+#define R2 0x10
+#define R3 0x18
+#define R4 0x20
+#define R5 0x28
+#define R6 0x30
+#define R7 0x38
+#define R8 0x40
+#define R9 0x48
+#define SL 0x50
+#define FP 0x58
+#define IP 0x60
+#define SP 0x68
+#define LR 0x70
+#define PC 0x78
+
+/*
+ * Implementation of void perf_regs_load(u64 *regs);
+ *
+ * This functions fills in the 'regs' buffer from the actual registers values,
+ * in the way the perf built-in unwinding test expects them:
+ * - the PC at the time at the call to this function. Since this function
+ *   is called using a bl instruction, the PC value is taken from LR.
+ * The built-in unwinding test then unwinds the call stack from the dwarf
+ * information in unwind__get_entries.
+ *
+ * Notes:
+ * - the 8 bytes stride in the registers offsets comes from the fact
+ * that the registers are stored in an u64 array (u64 *regs),
+ * - the regs buffer needs to be zeroed before the call to this function,
+ * in this case using a calloc in dwarf-unwind.c.
+ */
+
+.text
+.type perf_regs_load,%function
+ENTRY(perf_regs_load)
+	str r0, [r0, #R0]
+	str r1, [r0, #R1]
+	str r2, [r0, #R2]
+	str r3, [r0, #R3]
+	str r4, [r0, #R4]
+	str r5, [r0, #R5]
+	str r6, [r0, #R6]
+	str r7, [r0, #R7]
+	str r8, [r0, #R8]
+	str r9, [r0, #R9]
+	str sl, [r0, #SL]
+	str fp, [r0, #FP]
+	str ip, [r0, #IP]
+	str sp, [r0, #SP]
+	str lr, [r0, #LR]
+	str lr, [r0, #PC]	// store pc as lr in order to skip the call
+	                        //  to this function
+	mov pc, lr
+ENDPROC(perf_regs_load)
-- 
1.7.11.7

[PATCH 3/4] perf tests: Add dwarf unwind test on ARM

[Jean Pihet]

Adding dwarf unwind test, that setups live machine data over
the perf test thread and does the remote unwind.

Need to use -fno-optimize-sibling-calls for test compilation,
otherwise 'krava_*' function calls are optimized into jumps
and ommited from the stack unwind.

So far it was enabled only for x86.

Reviewed-by: Will Deacon <will.deacon@arm.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Cc: David Ahern <dsahern@gmail.com>
Signed-off-by: Jean Pihet <jean.pihet@linaro.org>
---
 tools/perf/Makefile.perf                 |  2 +-
 tools/perf/arch/arm/Makefile             |  1 +
 tools/perf/arch/arm/include/perf_regs.h  |  3 ++
 tools/perf/arch/arm/tests/dwarf-unwind.c | 60 ++++++++++++++++++++++++++++++++
 tools/perf/config/Makefile               |  4 +--
 tools/perf/tests/builtin-test.c          |  2 +-
 tools/perf/tests/tests.h                 |  2 +-
 7 files changed, 69 insertions(+), 5 deletions(-)
 create mode 100644 tools/perf/arch/arm/tests/dwarf-unwind.c

diff --git a/tools/perf/Makefile.perf b/tools/perf/Makefile.perf
index 2baf61c..dea2d633 100644
--- a/tools/perf/Makefile.perf
+++ b/tools/perf/Makefile.perf
@@ -411,7 +411,7 @@ LIB_OBJS += $(OUTPUT)tests/code-reading.o
 LIB_OBJS += $(OUTPUT)tests/sample-parsing.o
 LIB_OBJS += $(OUTPUT)tests/parse-no-sample-id-all.o
 ifndef NO_DWARF_UNWIND
-ifeq ($(ARCH),x86)
+ifeq ($(ARCH),$(filter $(ARCH),x86 arm))
 LIB_OBJS += $(OUTPUT)tests/dwarf-unwind.o
 endif
 endif
diff --git a/tools/perf/arch/arm/Makefile b/tools/perf/arch/arm/Makefile
index 9b8f87e..221f21d 100644
--- a/tools/perf/arch/arm/Makefile
+++ b/tools/perf/arch/arm/Makefile
@@ -5,4 +5,5 @@ endif
 ifndef NO_LIBUNWIND
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/unwind-libunwind.o
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/tests/regs_load.o
+LIB_OBJS += $(OUTPUT)arch/$(ARCH)/tests/dwarf-unwind.o
 endif
diff --git a/tools/perf/arch/arm/include/perf_regs.h b/tools/perf/arch/arm/include/perf_regs.h
index 33abcfa..f619c9c 100644
--- a/tools/perf/arch/arm/include/perf_regs.h
+++ b/tools/perf/arch/arm/include/perf_regs.h
@@ -8,6 +8,9 @@
 void perf_regs_load(u64 *regs);
 
 #define PERF_REGS_MASK	((1ULL << PERF_REG_ARM_MAX) - 1)
+#define PERF_REGS_MAX	PERF_REG_ARM_MAX
+#define PERF_SAMPLE_REGS_ABI	PERF_SAMPLE_REGS_ABI_32
+
 #define PERF_REG_IP	PERF_REG_ARM_PC
 #define PERF_REG_SP	PERF_REG_ARM_SP
 
diff --git a/tools/perf/arch/arm/tests/dwarf-unwind.c b/tools/perf/arch/arm/tests/dwarf-unwind.c
new file mode 100644
index 0000000..9f870d2
--- /dev/null
+++ b/tools/perf/arch/arm/tests/dwarf-unwind.c
@@ -0,0 +1,60 @@
+#include <string.h>
+#include "perf_regs.h"
+#include "thread.h"
+#include "map.h"
+#include "event.h"
+#include "tests/tests.h"
+
+#define STACK_SIZE 8192
+
+static int sample_ustack(struct perf_sample *sample,
+			 struct thread *thread, u64 *regs)
+{
+	struct stack_dump *stack = &sample->user_stack;
+	struct map *map;
+	unsigned long sp;
+	u64 stack_size, *buf;
+
+	buf = malloc(STACK_SIZE);
+	if (!buf) {
+		pr_debug("failed to allocate sample uregs data\n");
+		return -1;
+	}
+
+	sp = (unsigned long) regs[PERF_REG_ARM_SP];
+
+	map = map_groups__find(thread->mg, MAP__VARIABLE, (u64) sp);
+	if (!map) {
+		pr_debug("failed to get stack map\n");
+		free(buf);
+		return -1;
+	}
+
+	stack_size = map->end - sp;
+	stack_size = stack_size > STACK_SIZE ? STACK_SIZE : stack_size;
+
+	memcpy(buf, (void *) sp, stack_size);
+	stack->data = (char *) buf;
+	stack->size = stack_size;
+	return 0;
+}
+
+int test__arch_unwind_sample(struct perf_sample *sample,
+			     struct thread *thread)
+{
+	struct regs_dump *regs = &sample->user_regs;
+	u64 *buf;
+
+	buf = calloc(1, sizeof(u64) * PERF_REGS_MAX);
+	if (!buf) {
+		pr_debug("failed to allocate sample uregs data\n");
+		return -1;
+	}
+
+	perf_regs_load(buf);
+	regs->abi  = PERF_SAMPLE_REGS_ABI;
+	regs->regs = buf;
+	regs->mask = PERF_REGS_MASK;
+
+	return sample_ustack(sample, thread, buf);
+}
diff --git a/tools/perf/config/Makefile b/tools/perf/config/Makefile
index f2edc59..729bbdf 100644
--- a/tools/perf/config/Makefile
+++ b/tools/perf/config/Makefile
@@ -40,11 +40,11 @@ ifeq ($(ARCH),arm64)
   LIBUNWIND_LIBS = -lunwind -lunwind-aarch64
 endif
 
-# So far there's only x86 libdw unwind support merged in perf.
+# So far there's only x86 and arm libdw unwind support merged in perf.
 # Disable it on all other architectures in case libdw unwind
 # support is detected in system. Add supported architectures
 # to the check.
-ifneq ($(ARCH),x86)
+ifneq ($(ARCH),$(filter $(ARCH),x86 arm))
   NO_LIBDW_DWARF_UNWIND := 1
 endif
 
diff --git a/tools/perf/tests/builtin-test.c b/tools/perf/tests/builtin-test.c
index 0d5afaf..5e0764b 100644
--- a/tools/perf/tests/builtin-test.c
+++ b/tools/perf/tests/builtin-test.c
@@ -115,7 +115,7 @@ static struct test {
 		.desc = "Test parsing with no sample_id_all bit set",
 		.func = test__parse_no_sample_id_all,
 	},
-#if defined(__x86_64__) || defined(__i386__)
+#if defined(__x86_64__) || defined(__i386__) || defined(__arm__)
 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 	{
 		.desc = "Test dwarf unwind",
diff --git a/tools/perf/tests/tests.h b/tools/perf/tests/tests.h
index a9d7cb0..8f91fb0 100644
--- a/tools/perf/tests/tests.h
+++ b/tools/perf/tests/tests.h
@@ -45,7 +45,7 @@ int test__hists_filter(void);
 int test__mmap_thread_lookup(void);
 int test__thread_mg_share(void);
 
-#if defined(__x86_64__) || defined(__i386__)
+#if defined(__x86_64__) || defined(__i386__) || defined(__arm__)
 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 struct thread;
 struct perf_sample;
-- 
1.7.11.7

[PATCH 4/4] perf tools: Add libdw DWARF post unwind support for ARM

[Jean Pihet]

Adding libdw DWARF post unwind support, which is part
of elfutils-devel/libdw-dev package from version 0.158.

The new code is contained in unwin-libdw.c object, and
implements unwind__get_entries unwind interface function.

Signed-off-by: Jean Pihet <jean.pihet@linaro.org>
Reviewed-by: Will Deacon <will.deacon@arm.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Cc: David Ahern <dsahern@gmail.com>
---
 tools/perf/arch/arm/Makefile            |  5 +++++
 tools/perf/arch/arm/util/unwind-libdw.c | 36 +++++++++++++++++++++++++++++++++
 2 files changed, 41 insertions(+)
 create mode 100644 tools/perf/arch/arm/util/unwind-libdw.c

diff --git a/tools/perf/arch/arm/Makefile b/tools/perf/arch/arm/Makefile
index 221f21d..09d6215 100644
--- a/tools/perf/arch/arm/Makefile
+++ b/tools/perf/arch/arm/Makefile
@@ -4,6 +4,11 @@ LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/dwarf-regs.o
 endif
 ifndef NO_LIBUNWIND
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/unwind-libunwind.o
+endif
+ifndef NO_LIBDW_DWARF_UNWIND
+LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/unwind-libdw.o
+endif
+ifndef NO_DWARF_UNWIND
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/tests/regs_load.o
 LIB_OBJS += $(OUTPUT)arch/$(ARCH)/tests/dwarf-unwind.o
 endif
diff --git a/tools/perf/arch/arm/util/unwind-libdw.c b/tools/perf/arch/arm/util/unwind-libdw.c
new file mode 100644
index 0000000..b4176c6
--- /dev/null
+++ b/tools/perf/arch/arm/util/unwind-libdw.c
@@ -0,0 +1,36 @@
+#include <elfutils/libdwfl.h>
+#include "../../util/unwind-libdw.h"
+#include "../../util/perf_regs.h"
+
+bool libdw__arch_set_initial_registers(Dwfl_Thread *thread, void *arg)
+{
+	struct unwind_info *ui = arg;
+	struct regs_dump *user_regs = &ui->sample->user_regs;
+	Dwarf_Word dwarf_regs[PERF_REG_ARM_MAX];
+
+#define REG(r) ({						\
+	Dwarf_Word val = 0;					\
+	perf_reg_value(&val, user_regs, PERF_REG_ARM_##r);	\
+	val;							\
+})
+
+	dwarf_regs[0]  = REG(R0);
+	dwarf_regs[1]  = REG(R1);
+	dwarf_regs[2]  = REG(R2);
+	dwarf_regs[3]  = REG(R3);
+	dwarf_regs[4]  = REG(R4);
+	dwarf_regs[5]  = REG(R5);
+	dwarf_regs[6]  = REG(R6);
+	dwarf_regs[7]  = REG(R7);
+	dwarf_regs[8]  = REG(R8);
+	dwarf_regs[9]  = REG(R9);
+	dwarf_regs[10] = REG(R10);
+	dwarf_regs[11] = REG(FP);
+	dwarf_regs[12] = REG(IP);
+	dwarf_regs[13] = REG(SP);
+	dwarf_regs[14] = REG(LR);
+	dwarf_regs[15] = REG(PC);
+
+	return dwfl_thread_state_registers(thread, 0, PERF_REG_ARM_MAX,
+					   dwarf_regs);
+}
-- 
1.7.11.7

[PATCH 0/3] perf tools: Add libdw DWARF post unwind support for ARM

[Jean Pihet]

Adding libdw DWARF post unwind support, which is part
of elfutils-devel/libdw-dev package from version 0.158.

Also includes the test suite for dwarf unwinding, by adding the
arch specific test code and the perf_regs_load function.

Jean Pihet (3):
  perf tests: Introduce perf_regs_load function on ARM
  perf tests: Add dwarf unwind test on ARM
  perf tools: Add libdw DWARF post unwind support for ARM

 tools/perf/Makefile.perf                 |  2 +-
 tools/perf/arch/arm/Makefile             |  7 ++++
 tools/perf/arch/arm/include/perf_regs.h  |  5 +++
 tools/perf/arch/arm/tests/dwarf-unwind.c | 59 ++++++++++++++++++++++++++++++++
 tools/perf/arch/arm/tests/regs_load.S    | 51 +++++++++++++++++++++++++++
 tools/perf/arch/arm/util/unwind-libdw.c  | 36 +++++++++++++++++++
 tools/perf/tests/builtin-test.c          |  2 +-
 tools/perf/tests/tests.h                 |  2 +-
 8 files changed, 161 insertions(+), 3 deletions(-)
 create mode 100644 tools/perf/arch/arm/tests/dwarf-unwind.c
 create mode 100644 tools/perf/arch/arm/tests/regs_load.S
 create mode 100644 tools/perf/arch/arm/util/unwind-libdw.c

---

- Rebased on latest acme/perf/core git tree,
- Tested on quad-core ARMv7 machine

-- 
1.7.11.7

[PATCH 0/3] perf tools: Add libdw DWARF post unwind support for ARM

[Jean Pihet]

Hi,

Here are some notes about the performance improvement and the test usage.

Jiri, do you need these notes in the one of commit description?

1. Using libdw vs libunwind on ARMv7

The performance gain is (more than) significant: >800%.

I did not profile perf itself ;-p just did a timing measurement.
The benchmark has been run multiple times with different perf.data
sizes. The results are consistent across the tests runs.

Usage:
./tools/perf/perf record --call-graph dwarf --
../../libunwind/test_app/stress_bt
time ./tools/perf/perf report --stdio > /dev/null 2>&1

Platform:
Quad-core marvell XP370. perf runs on 1 cpu

Perf data size:
304MB    libunwind    libdw        improvement
real    9m31.577s    1m13.052s    782%
user    5m33.020s    1m2.910s    529%
sys        3m57.770s    0m10.090s    2356%

2. unwind test usage
perf test list gives the list of supported tests. In this case the
test #23 is dwarf unwinding:

./tools/perf/perf test 23
23: Test dwarf unwind                                      : Ok


Regards,
Jean


On 3 March 2014 10:53, Jean Pihet <jean.pihet@linaro.org> wrote:
> Adding libdw DWARF post unwind support, which is part
> of elfutils-devel/libdw-dev package from version 0.158.
>
> Also includes the test suite for dwarf unwinding, by adding the
> arch specific test code and the perf_regs_load function.
>
> Jean Pihet (3):
>   perf tests: Introduce perf_regs_load function on ARM
>   perf tests: Add dwarf unwind test on ARM
>   perf tools: Add libdw DWARF post unwind support for ARM
>
>  tools/perf/Makefile.perf                 |  2 +-
>  tools/perf/arch/arm/Makefile             |  7 ++++
>  tools/perf/arch/arm/include/perf_regs.h  |  5 +++
>  tools/perf/arch/arm/tests/dwarf-unwind.c | 59 ++++++++++++++++++++++++++++++++
>  tools/perf/arch/arm/tests/regs_load.S    | 51 +++++++++++++++++++++++++++
>  tools/perf/arch/arm/util/unwind-libdw.c  | 36 +++++++++++++++++++
>  tools/perf/tests/builtin-test.c          |  2 +-
>  tools/perf/tests/tests.h                 |  2 +-
>  8 files changed, 161 insertions(+), 3 deletions(-)
>  create mode 100644 tools/perf/arch/arm/tests/dwarf-unwind.c
>  create mode 100644 tools/perf/arch/arm/tests/regs_load.S
>  create mode 100644 tools/perf/arch/arm/util/unwind-libdw.c
>
> ---
>
> - Rebased on latest acme/perf/core git tree,
> - Tested on quad-core ARMv7 machine
>
> --
> 1.7.11.7
>

[PATCH 0/3] perf tools: Add libdw DWARF post unwind support for ARM

[Jiri Olsa]

On Tue, Mar 04, 2014 at 02:24:37AM +0100, Jean Pihet wrote:
> Hi,
> 
> Here are some notes about the performance improvement and the test usage.
> 
> Jiri, do you need these notes in the one of commit description?
> 
> 1. Using libdw vs libunwind on ARMv7
> 
> The performance gain is (more than) significant: >800%.

I dont mind.. but 800% would look just great in changelog ;-)

jirka

> 
> I did not profile perf itself ;-p just did a timing measurement.
> The benchmark has been run multiple times with different perf.data
> sizes. The results are consistent across the tests runs.
> 
> Usage:
> ./tools/perf/perf record --call-graph dwarf --
> ../../libunwind/test_app/stress_bt
> time ./tools/perf/perf report --stdio > /dev/null 2>&1
> 
> Platform:
> Quad-core marvell XP370. perf runs on 1 cpu
> 
> Perf data size:
> 304MB    libunwind    libdw        improvement
> real    9m31.577s    1m13.052s    782%
> user    5m33.020s    1m2.910s    529%
> sys        3m57.770s    0m10.090s    2356%
> 
> 2. unwind test usage
> perf test list gives the list of supported tests. In this case the
> test #23 is dwarf unwinding:
> 
> ./tools/perf/perf test 23
> 23: Test dwarf unwind                                      : Ok
> 
> 
> Regards,
> Jean
> 
> 
> On 3 March 2014 10:53, Jean Pihet <jean.pihet@linaro.org> wrote:
> > Adding libdw DWARF post unwind support, which is part
> > of elfutils-devel/libdw-dev package from version 0.158.
> >
> > Also includes the test suite for dwarf unwinding, by adding the
> > arch specific test code and the perf_regs_load function.
> >
> > Jean Pihet (3):
> >   perf tests: Introduce perf_regs_load function on ARM
> >   perf tests: Add dwarf unwind test on ARM
> >   perf tools: Add libdw DWARF post unwind support for ARM
> >
> >  tools/perf/Makefile.perf                 |  2 +-
> >  tools/perf/arch/arm/Makefile             |  7 ++++
> >  tools/perf/arch/arm/include/perf_regs.h  |  5 +++
> >  tools/perf/arch/arm/tests/dwarf-unwind.c | 59 ++++++++++++++++++++++++++++++++
> >  tools/perf/arch/arm/tests/regs_load.S    | 51 +++++++++++++++++++++++++++
> >  tools/perf/arch/arm/util/unwind-libdw.c  | 36 +++++++++++++++++++
> >  tools/perf/tests/builtin-test.c          |  2 +-
> >  tools/perf/tests/tests.h                 |  2 +-
> >  8 files changed, 161 insertions(+), 3 deletions(-)
> >  create mode 100644 tools/perf/arch/arm/tests/dwarf-unwind.c
> >  create mode 100644 tools/perf/arch/arm/tests/regs_load.S
> >  create mode 100644 tools/perf/arch/arm/util/unwind-libdw.c
> >
> > ---
> >
> > - Rebased on latest acme/perf/core git tree,
> > - Tested on quad-core ARMv7 machine
> >
> > --
> > 1.7.11.7
> >

[PATCH 0/3] perf tools: Add libdw DWARF post unwind support for ARM

[Jiri Olsa]

On Tue, Mar 04, 2014 at 02:24:37AM +0100, Jean Pihet wrote:
> Hi,
> 
> Here are some notes about the performance improvement and the test usage.
> 
> Jiri, do you need these notes in the one of commit description?
> 
> 1. Using libdw vs libunwind on ARMv7
> 
> The performance gain is (more than) significant: >800%.
> 
> I did not profile perf itself ;-p just did a timing measurement.
> The benchmark has been run multiple times with different perf.data
> sizes. The results are consistent across the tests runs.
> 
> Usage:
> ./tools/perf/perf record --call-graph dwarf --
> ../../libunwind/test_app/stress_bt
> time ./tools/perf/perf report --stdio > /dev/null 2>&1
> 
> Platform:
> Quad-core marvell XP370. perf runs on 1 cpu
> 
> Perf data size:
> 304MB    libunwind    libdw        improvement
> real    9m31.577s    1m13.052s    782%
> user    5m33.020s    1m2.910s    529%
> sys        3m57.770s    0m10.090s    2356%
> 
> 2. unwind test usage
> perf test list gives the list of supported tests. In this case the
> test #23 is dwarf unwinding:
> 
> ./tools/perf/perf test 23
> 23: Test dwarf unwind                                      : Ok
> 
> 
> Regards,
> Jean
> 
> 

I can't speak for the arm specific stuff, but generally it fits
and looks ok (and does not break the x86 build)

also.. who wouldn't ack 800% speed improvement ;-)

Acked-by: Jiri Olsa <jolsa@redhat.com>

thanks,
jirka