public inbox for linux-kernel@vger.kernel.org
 help / color / mirror / Atom feed
* Re: [PATCH] perf probe: Add fastpath to do lookup by function name
  2011-03-24 15:38 [PATCH] perf probe: Add fastpath to do lookup by function name Lin Ming
@ 2011-03-24  7:58 ` Ingo Molnar
  2011-03-24  8:38   ` Lin Ming
  2011-03-24  9:08 ` Masami Hiramatsu
  1 sibling, 1 reply; 11+ messages in thread
From: Ingo Molnar @ 2011-03-24  7:58 UTC (permalink / raw)
  To: Lin Ming
  Cc: Arnaldo Carvalho de Melo, Masami Hiramatsu, Peter Zijlstra,
	linux-kernel


* Lin Ming <ming.m.lin@intel.com> wrote:

> The vmlinux file may have thousands of CUs.
> We can lookup function name from .debug_pubnames section
> to avoid the slow loop on CUs.

Mind including before/after perf stat --repeat 10 results in the changelog, of 
an affected command which got faster?

Thanks,

	Ingo

^ permalink raw reply	[flat|nested] 11+ messages in thread

* Re: [PATCH] perf probe: Add fastpath to do lookup by function name
  2011-03-24  7:58 ` Ingo Molnar
@ 2011-03-24  8:38   ` Lin Ming
  2011-03-24  8:47     ` Ingo Molnar
  0 siblings, 1 reply; 11+ messages in thread
From: Lin Ming @ 2011-03-24  8:38 UTC (permalink / raw)
  To: Ingo Molnar
  Cc: Arnaldo Carvalho de Melo, Masami Hiramatsu, Peter Zijlstra,
	linux-kernel

On Thu, 2011-03-24 at 15:58 +0800, Ingo Molnar wrote:
> * Lin Ming <ming.m.lin@intel.com> wrote:
> 
> > The vmlinux file may have thousands of CUs.
> > We can lookup function name from .debug_pubnames section
> > to avoid the slow loop on CUs.
> 
> Mind including before/after perf stat --repeat 10 results in the changelog, of 
> an affected command which got faster?


>From 611a823a34e655d47a80f92994b004391e2b244c Mon Sep 17 00:00:00 2001
From: Lin Ming <ming.m.lin@intel.com>
Date: Thu, 24 Mar 2011 23:22:24 +0800
Subject: [PATCH] perf probe: Add fastpath to do lookup by function name

The vmlinux file may have thousands of CUs.
We can lookup function name from .debug_pubnames section
to avoid the slow loop on CUs.

./perf stat -r 10 -- ./perf probe -k /home/mlin/vmlinux \
        -s /home/mlin/linux-2.6 \
	--line csum_partial_copy_to_user > tmp.log

before patch applied
=====================
        364.535892 task-clock-msecs         #      0.997 CPUs  
                 0 context-switches         #      0.000 M/sec 
                 0 CPU-migrations           #      0.000 M/sec 
            29,993 page-faults              #      0.082 M/sec 
       865,862,109 cycles                   #   2375.245 M/sec 
     1,255,259,630 instructions             #      1.450 IPC  
       252,400,884 branches                 #    692.390 M/sec 
         3,429,376 branch-misses            #      1.359 %     
         1,386,990 cache-references         #      3.805 M/sec 
           687,188 cache-misses             #      1.885 M/sec 

        0.365792170  seconds time elapsed  


after patch applied
=====================
         89.896405 task-clock-msecs         #      0.991 CPUs
                 1 context-switches         #      0.000 M/sec
                 0 CPU-migrations           #      0.000 M/sec
            10,145 page-faults              #      0.113 M/sec
       214,553,875 cycles                   #   2386.679 M/sec
       226,915,559 instructions             #      1.058 IPC
        44,536,614 branches                 #    495.422 M/sec
           613,074 branch-misses            #      1.377 %
           860,787 cache-references         #      9.575 M/sec
           442,380 cache-misses             #      4.921 M/sec

        0.090716032  seconds time elapsed


Signed-off-by: Lin Ming <ming.m.lin@intel.com>
---
 tools/perf/util/probe-finder.c |   38 ++++++++++++++++++++++++++++++++++++++
 tools/perf/util/probe-finder.h |    1 +
 2 files changed, 39 insertions(+), 0 deletions(-)

diff --git a/tools/perf/util/probe-finder.c b/tools/perf/util/probe-finder.c
index 194f9e2..b2034c2 100644
--- a/tools/perf/util/probe-finder.c
+++ b/tools/perf/util/probe-finder.c
@@ -1876,6 +1876,30 @@ static int find_line_range_by_func(struct line_finder *lf)
 	return param.retval;
 }
 
+static int pubname_search_cb(Dwarf *dbg, Dwarf_Global *gl, void *data)
+{
+	struct line_finder *lf = data;
+	struct line_range *lr = lf->lr;
+
+	if (dwarf_offdie(dbg, gl->die_offset, &lf->sp_die)) {
+		if (dwarf_tag(&lf->sp_die) != DW_TAG_subprogram)
+			return DWARF_CB_OK;
+
+		if (die_compare_name(&lf->sp_die, lr->function)) {
+			if (!dwarf_offdie(dbg, gl->cu_offset, &lf->cu_die))
+				return DWARF_CB_OK;
+
+			if (lr->file && !cu_find_realpath(&lf->cu_die, lr->file))
+				return DWARF_CB_OK;
+
+			lf->found = 1;
+			return DWARF_CB_ABORT;
+		}
+	}
+
+	return DWARF_CB_OK;
+}
+
 int find_line_range(int fd, struct line_range *lr)
 {
 	struct line_finder lf = {.lr = lr, .found = 0};
@@ -1895,6 +1919,19 @@ int find_line_range(int fd, struct line_range *lr)
 		return -EBADF;
 	}
 
+	/* Fastpath: lookup by function name from .debug_pubnames section */
+	if (lr->function) {
+		struct dwarf_callback_param param = {.data = (void *)&lf, .retval = 0};
+
+		dwarf_getpubnames(dbg, pubname_search_cb, &lf, 0);
+		if (lf.found) {
+			lf.found = 0;
+			line_range_search_cb(&lf.sp_die, &param);
+			if (lf.found)
+				goto found;
+		}
+	}
+
 	/* Loop on CUs (Compilation Unit) */
 	while (!lf.found && ret >= 0) {
 		if (dwarf_nextcu(dbg, off, &noff, &cuhl, NULL, NULL, NULL) != 0)
@@ -1923,6 +1960,7 @@ int find_line_range(int fd, struct line_range *lr)
 		off = noff;
 	}
 
+found:
 	/* Store comp_dir */
 	if (lf.found) {
 		comp_dir = cu_get_comp_dir(&lf.cu_die);
diff --git a/tools/perf/util/probe-finder.h b/tools/perf/util/probe-finder.h
index beaefc3..4bc56a4 100644
--- a/tools/perf/util/probe-finder.h
+++ b/tools/perf/util/probe-finder.h
@@ -83,6 +83,7 @@ struct line_finder {
 	int			lno_s;		/* Start line number */
 	int			lno_e;		/* End line number */
 	Dwarf_Die		cu_die;		/* Current CU */
+	Dwarf_Die		sp_die;
 	int			found;
 };
 
-- 
1.7.2.3




^ permalink raw reply related	[flat|nested] 11+ messages in thread

* Re: [PATCH] perf probe: Add fastpath to do lookup by function name
  2011-03-24  8:38   ` Lin Ming
@ 2011-03-24  8:47     ` Ingo Molnar
  0 siblings, 0 replies; 11+ messages in thread
From: Ingo Molnar @ 2011-03-24  8:47 UTC (permalink / raw)
  To: Lin Ming
  Cc: Arnaldo Carvalho de Melo, Masami Hiramatsu, Peter Zijlstra,
	linux-kernel


* Lin Ming <ming.m.lin@intel.com> wrote:

> ./perf stat -r 10 -- ./perf probe -k /home/mlin/vmlinux \
>         -s /home/mlin/linux-2.6 \
> 	--line csum_partial_copy_to_user > tmp.log
> 
> before patch applied
> =====================
>         364.535892 task-clock-msecs         #      0.997 CPUs  
>        865,862,109 cycles                   #   2375.245 M/sec 
>      1,255,259,630 instructions             #      1.450 IPC  

> after patch applied
> =====================
>          89.896405 task-clock-msecs         #      0.991 CPUs
>        214,553,875 cycles                   #   2386.679 M/sec
>        226,915,559 instructions             #      1.058 IPC

That's a very nice speedup :-)

Thanks,

	Ingo

^ permalink raw reply	[flat|nested] 11+ messages in thread

* Re: [PATCH] perf probe: Add fastpath to do lookup by function name
  2011-03-24 15:38 [PATCH] perf probe: Add fastpath to do lookup by function name Lin Ming
  2011-03-24  7:58 ` Ingo Molnar
@ 2011-03-24  9:08 ` Masami Hiramatsu
  2011-03-24 13:47   ` Lin Ming
  1 sibling, 1 reply; 11+ messages in thread
From: Masami Hiramatsu @ 2011-03-24  9:08 UTC (permalink / raw)
  To: Lin Ming
  Cc: Arnaldo Carvalho de Melo, Peter Zijlstra, Ingo Molnar,
	linux-kernel

(2011/03/25 0:38), Lin Ming wrote:
> The vmlinux file may have thousands of CUs.
> We can lookup function name from .debug_pubnames section
> to avoid the slow loop on CUs.
> 
> Signed-off-by: Lin Ming <ming.m.lin@intel.com>
> ---
>  tools/perf/util/probe-finder.c |   38 ++++++++++++++++++++++++++++++++++++++
>  tools/perf/util/probe-finder.h |    1 +
>  2 files changed, 39 insertions(+), 0 deletions(-)
> 
> diff --git a/tools/perf/util/probe-finder.c b/tools/perf/util/probe-finder.c
> index 194f9e2..b2034c2 100644
> --- a/tools/perf/util/probe-finder.c
> +++ b/tools/perf/util/probe-finder.c
> @@ -1876,6 +1876,30 @@ static int find_line_range_by_func(struct line_finder *lf)
>  	return param.retval;
>  }
>  
> +static int pubname_search_cb(Dwarf *dbg, Dwarf_Global *gl, void *data)
> +{
> +	struct line_finder *lf = data;
> +	struct line_range *lr = lf->lr;
> +
> +	if (dwarf_offdie(dbg, gl->die_offset, &lf->sp_die)) {
> +		if (dwarf_tag(&lf->sp_die) != DW_TAG_subprogram)
> +			return DWARF_CB_OK;
> +
> +		if (die_compare_name(&lf->sp_die, lr->function)) {
> +			if (!dwarf_offdie(dbg, gl->cu_offset, &lf->cu_die))
> +				return DWARF_CB_OK;
> +

Just one comment.
Could you ensure that the decl_file of sp_die matches lr->file (by strtailcmp) here?

Other parts look good to me:)

Thanks!

> +			if (lr->file && !cu_find_realpath(&lf->cu_die, lr->file))
> +				return DWARF_CB_OK;
> +
> +			lf->found = 1;
> +			return DWARF_CB_ABORT;
> +		}
> +	}
> +
> +	return DWARF_CB_OK;
> +}
> +
>  int find_line_range(int fd, struct line_range *lr)
>  {
>  	struct line_finder lf = {.lr = lr, .found = 0};
> @@ -1895,6 +1919,19 @@ int find_line_range(int fd, struct line_range *lr)
>  		return -EBADF;
>  	}
>  
> +	/* Fastpath: lookup by function name from .debug_pubnames section */
> +	if (lr->function) {
> +		struct dwarf_callback_param param = {.data = (void *)&lf, .retval = 0};
> +
> +		dwarf_getpubnames(dbg, pubname_search_cb, &lf, 0);
> +		if (lf.found) {
> +			lf.found = 0;
> +			line_range_search_cb(&lf.sp_die, &param);
> +			if (lf.found)
> +				goto found;
> +		}
> +	}
> +
>  	/* Loop on CUs (Compilation Unit) */
>  	while (!lf.found && ret >= 0) {
>  		if (dwarf_nextcu(dbg, off, &noff, &cuhl, NULL, NULL, NULL) != 0)
> @@ -1923,6 +1960,7 @@ int find_line_range(int fd, struct line_range *lr)
>  		off = noff;
>  	}
>  
> +found:
>  	/* Store comp_dir */
>  	if (lf.found) {
>  		comp_dir = cu_get_comp_dir(&lf.cu_die);
> diff --git a/tools/perf/util/probe-finder.h b/tools/perf/util/probe-finder.h
> index beaefc3..4bc56a4 100644
> --- a/tools/perf/util/probe-finder.h
> +++ b/tools/perf/util/probe-finder.h
> @@ -83,6 +83,7 @@ struct line_finder {
>  	int			lno_s;		/* Start line number */
>  	int			lno_e;		/* End line number */
>  	Dwarf_Die		cu_die;		/* Current CU */
> +	Dwarf_Die		sp_die;
>  	int			found;
>  };
>  


-- 
Masami HIRAMATSU
2nd Dept. Linux Technology Center
Hitachi, Ltd., Systems Development Laboratory
E-mail: masami.hiramatsu.pt@hitachi.com

^ permalink raw reply	[flat|nested] 11+ messages in thread

* Re: [PATCH] perf probe: Add fastpath to do lookup by function name
  2011-03-24  9:08 ` Masami Hiramatsu
@ 2011-03-24 13:47   ` Lin Ming
  2011-03-24 14:09     ` [PATCH v2 -tip] " Lin Ming
  0 siblings, 1 reply; 11+ messages in thread
From: Lin Ming @ 2011-03-24 13:47 UTC (permalink / raw)
  To: Masami Hiramatsu
  Cc: Arnaldo Carvalho de Melo, Peter Zijlstra, Ingo Molnar,
	linux-kernel

On Thu, 2011-03-24 at 17:08 +0800, Masami Hiramatsu wrote:
> (2011/03/25 0:38), Lin Ming wrote:
> > The vmlinux file may have thousands of CUs.
> > We can lookup function name from .debug_pubnames section
> > to avoid the slow loop on CUs.
> > 
> > Signed-off-by: Lin Ming <ming.m.lin@intel.com>
> > ---
> >  tools/perf/util/probe-finder.c |   38 ++++++++++++++++++++++++++++++++++++++
> >  tools/perf/util/probe-finder.h |    1 +
> >  2 files changed, 39 insertions(+), 0 deletions(-)
> > 
> > diff --git a/tools/perf/util/probe-finder.c b/tools/perf/util/probe-finder.c
> > index 194f9e2..b2034c2 100644
> > --- a/tools/perf/util/probe-finder.c
> > +++ b/tools/perf/util/probe-finder.c
> > @@ -1876,6 +1876,30 @@ static int find_line_range_by_func(struct line_finder *lf)
> >  	return param.retval;
> >  }
> >  
> > +static int pubname_search_cb(Dwarf *dbg, Dwarf_Global *gl, void *data)
> > +{
> > +	struct line_finder *lf = data;
> > +	struct line_range *lr = lf->lr;
> > +
> > +	if (dwarf_offdie(dbg, gl->die_offset, &lf->sp_die)) {
> > +		if (dwarf_tag(&lf->sp_die) != DW_TAG_subprogram)
> > +			return DWARF_CB_OK;
> > +
> > +		if (die_compare_name(&lf->sp_die, lr->function)) {
> > +			if (!dwarf_offdie(dbg, gl->cu_offset, &lf->cu_die))
> > +				return DWARF_CB_OK;
> > +
> 
> Just one comment.
> Could you ensure that the decl_file of sp_die matches lr->file (by strtailcmp) here?

OK, so the file name compare with cu_find_realpath(..) can be removed,
as below.

This makes the lookup a bit more faster again.

Thanks, I'll post a new version with below changes.

diff --git a/tools/perf/util/probe-finder.c b/tools/perf/util/probe-finder.c
index b2034c2..38e4a05 100644
--- a/tools/perf/util/probe-finder.c
+++ b/tools/perf/util/probe-finder.c
@@ -1880,6 +1880,7 @@ static int pubname_search_cb(Dwarf *dbg, Dwarf_Global *gl, void *data)
 {
 	struct line_finder *lf = data;
 	struct line_range *lr = lf->lr;
+	const char *file;
 
 	if (dwarf_offdie(dbg, gl->die_offset, &lf->sp_die)) {
 		if (dwarf_tag(&lf->sp_die) != DW_TAG_subprogram)
@@ -1889,8 +1890,12 @@ static int pubname_search_cb(Dwarf *dbg, Dwarf_Global *gl, void *data)
 			if (!dwarf_offdie(dbg, gl->cu_offset, &lf->cu_die))
 				return DWARF_CB_OK;
 
-			if (lr->file && !cu_find_realpath(&lf->cu_die, lr->file))
-				return DWARF_CB_OK;
+			if (lr->file) {
+				file = dwarf_decl_file(&lf->sp_die);
+
+				if (file && strtailcmp(file, lr->file))
+					return DWARF_CB_OK;
+			}
 
 			lf->found = 1;
 			return DWARF_CB_ABORT;





^ permalink raw reply related	[flat|nested] 11+ messages in thread

* [PATCH v2 -tip] perf probe: Add fastpath to do lookup by function name
  2011-03-24 13:47   ` Lin Ming
@ 2011-03-24 14:09     ` Lin Ming
  2011-03-25  1:14       ` Masami Hiramatsu
  0 siblings, 1 reply; 11+ messages in thread
From: Lin Ming @ 2011-03-24 14:09 UTC (permalink / raw)
  To: Masami Hiramatsu, Arnaldo Carvalho de Melo
  Cc: Peter Zijlstra, Ingo Molnar, linux-kernel

v2 -> v1:
- Don't compare file names with cu_find_realpath(...), instead, compare them
  with the name returned by dwarf_decl_file(sp_die) 

The vmlinux file may have thousands of CUs.
We can lookup function name from .debug_pubnames section
to avoid the slow loop on CUs.

./perf stat -r 10 -- ./perf probe -k /home/mlin/vmlinux \
        -s /home/mlin/linux-2.6 \
	--line csum_partial_copy_to_user > tmp.log

before patch applied
=====================
        364.535892 task-clock-msecs         #      0.997 CPUs
                 0 context-switches         #      0.000 M/sec
                 0 CPU-migrations           #      0.000 M/sec
            29,993 page-faults              #      0.082 M/sec
       865,862,109 cycles                   #   2375.245 M/sec
     1,255,259,630 instructions             #      1.450 IPC
       252,400,884 branches                 #    692.390 M/sec
         3,429,376 branch-misses            #      1.359 %
         1,386,990 cache-references         #      3.805 M/sec
           687,188 cache-misses             #      1.885 M/sec

        0.365792170  seconds time elapsed

after patch applied
=====================
         89.896405 task-clock-msecs         #      0.991 CPUs
                 1 context-switches         #      0.000 M/sec
                 0 CPU-migrations           #      0.000 M/sec
            10,145 page-faults              #      0.113 M/sec
       214,553,875 cycles                   #   2386.679 M/sec
       226,915,559 instructions             #      1.058 IPC
        44,536,614 branches                 #    495.422 M/sec
           613,074 branch-misses            #      1.377 %
           860,787 cache-references         #      9.575 M/sec
           442,380 cache-misses             #      4.921 M/sec

        0.090716032  seconds time elapsed

Signed-off-by: Lin Ming <ming.m.lin@intel.com>
---
 tools/perf/util/probe-finder.c |   39 +++++++++++++++++++++++++++++++++++++++
 tools/perf/util/probe-finder.h |    1 +
 2 files changed, 40 insertions(+), 0 deletions(-)

diff --git a/tools/perf/util/probe-finder.c b/tools/perf/util/probe-finder.c
index 194f9e2..5cf044c 100644
--- a/tools/perf/util/probe-finder.c
+++ b/tools/perf/util/probe-finder.c
@@ -1876,6 +1876,31 @@ static int find_line_range_by_func(struct line_finder *lf)
 	return param.retval;
 }
 
+static int pubname_search_cb(Dwarf *dbg, Dwarf_Global *gl, void *data)
+{
+	struct line_finder *lf = data;
+	struct line_range *lr = lf->lr;
+
+	if (dwarf_offdie(dbg, gl->die_offset, &lf->sp_die)) {
+		if (dwarf_tag(&lf->sp_die) != DW_TAG_subprogram)
+			return DWARF_CB_OK;
+
+		if (die_compare_name(&lf->sp_die, lr->function)) {
+			if (!dwarf_offdie(dbg, gl->cu_offset, &lf->cu_die))
+				return DWARF_CB_OK;
+
+			if (lr->file &&
+			    strtailcmp(lr->file, dwarf_decl_file(&lf->sp_die)))
+				return DWARF_CB_OK;
+
+			lf->found = 1;
+			return DWARF_CB_ABORT;
+		}
+	}
+
+	return DWARF_CB_OK;
+}
+
 int find_line_range(int fd, struct line_range *lr)
 {
 	struct line_finder lf = {.lr = lr, .found = 0};
@@ -1895,6 +1920,19 @@ int find_line_range(int fd, struct line_range *lr)
 		return -EBADF;
 	}
 
+	/* Fastpath: lookup by function name from .debug_pubnames section */
+	if (lr->function) {
+		struct dwarf_callback_param param = {.data = (void *)&lf, .retval = 0};
+
+		dwarf_getpubnames(dbg, pubname_search_cb, &lf, 0);
+		if (lf.found) {
+			lf.found = 0;
+			line_range_search_cb(&lf.sp_die, &param);
+			if (lf.found)
+				goto found;
+		}
+	}
+
 	/* Loop on CUs (Compilation Unit) */
 	while (!lf.found && ret >= 0) {
 		if (dwarf_nextcu(dbg, off, &noff, &cuhl, NULL, NULL, NULL) != 0)
@@ -1923,6 +1961,7 @@ int find_line_range(int fd, struct line_range *lr)
 		off = noff;
 	}
 
+found:
 	/* Store comp_dir */
 	if (lf.found) {
 		comp_dir = cu_get_comp_dir(&lf.cu_die);
diff --git a/tools/perf/util/probe-finder.h b/tools/perf/util/probe-finder.h
index beaefc3..4bc56a4 100644
--- a/tools/perf/util/probe-finder.h
+++ b/tools/perf/util/probe-finder.h
@@ -83,6 +83,7 @@ struct line_finder {
 	int			lno_s;		/* Start line number */
 	int			lno_e;		/* End line number */
 	Dwarf_Die		cu_die;		/* Current CU */
+	Dwarf_Die		sp_die;
 	int			found;
 };
 
-- 
1.7.2.3




^ permalink raw reply related	[flat|nested] 11+ messages in thread

* [PATCH] perf probe: Add fastpath to do lookup by function name
@ 2011-03-24 15:38 Lin Ming
  2011-03-24  7:58 ` Ingo Molnar
  2011-03-24  9:08 ` Masami Hiramatsu
  0 siblings, 2 replies; 11+ messages in thread
From: Lin Ming @ 2011-03-24 15:38 UTC (permalink / raw)
  To: Arnaldo Carvalho de Melo, Masami Hiramatsu
  Cc: Peter Zijlstra, Ingo Molnar, linux-kernel

The vmlinux file may have thousands of CUs.
We can lookup function name from .debug_pubnames section
to avoid the slow loop on CUs.

Signed-off-by: Lin Ming <ming.m.lin@intel.com>
---
 tools/perf/util/probe-finder.c |   38 ++++++++++++++++++++++++++++++++++++++
 tools/perf/util/probe-finder.h |    1 +
 2 files changed, 39 insertions(+), 0 deletions(-)

diff --git a/tools/perf/util/probe-finder.c b/tools/perf/util/probe-finder.c
index 194f9e2..b2034c2 100644
--- a/tools/perf/util/probe-finder.c
+++ b/tools/perf/util/probe-finder.c
@@ -1876,6 +1876,30 @@ static int find_line_range_by_func(struct line_finder *lf)
 	return param.retval;
 }
 
+static int pubname_search_cb(Dwarf *dbg, Dwarf_Global *gl, void *data)
+{
+	struct line_finder *lf = data;
+	struct line_range *lr = lf->lr;
+
+	if (dwarf_offdie(dbg, gl->die_offset, &lf->sp_die)) {
+		if (dwarf_tag(&lf->sp_die) != DW_TAG_subprogram)
+			return DWARF_CB_OK;
+
+		if (die_compare_name(&lf->sp_die, lr->function)) {
+			if (!dwarf_offdie(dbg, gl->cu_offset, &lf->cu_die))
+				return DWARF_CB_OK;
+
+			if (lr->file && !cu_find_realpath(&lf->cu_die, lr->file))
+				return DWARF_CB_OK;
+
+			lf->found = 1;
+			return DWARF_CB_ABORT;
+		}
+	}
+
+	return DWARF_CB_OK;
+}
+
 int find_line_range(int fd, struct line_range *lr)
 {
 	struct line_finder lf = {.lr = lr, .found = 0};
@@ -1895,6 +1919,19 @@ int find_line_range(int fd, struct line_range *lr)
 		return -EBADF;
 	}
 
+	/* Fastpath: lookup by function name from .debug_pubnames section */
+	if (lr->function) {
+		struct dwarf_callback_param param = {.data = (void *)&lf, .retval = 0};
+
+		dwarf_getpubnames(dbg, pubname_search_cb, &lf, 0);
+		if (lf.found) {
+			lf.found = 0;
+			line_range_search_cb(&lf.sp_die, &param);
+			if (lf.found)
+				goto found;
+		}
+	}
+
 	/* Loop on CUs (Compilation Unit) */
 	while (!lf.found && ret >= 0) {
 		if (dwarf_nextcu(dbg, off, &noff, &cuhl, NULL, NULL, NULL) != 0)
@@ -1923,6 +1960,7 @@ int find_line_range(int fd, struct line_range *lr)
 		off = noff;
 	}
 
+found:
 	/* Store comp_dir */
 	if (lf.found) {
 		comp_dir = cu_get_comp_dir(&lf.cu_die);
diff --git a/tools/perf/util/probe-finder.h b/tools/perf/util/probe-finder.h
index beaefc3..4bc56a4 100644
--- a/tools/perf/util/probe-finder.h
+++ b/tools/perf/util/probe-finder.h
@@ -83,6 +83,7 @@ struct line_finder {
 	int			lno_s;		/* Start line number */
 	int			lno_e;		/* End line number */
 	Dwarf_Die		cu_die;		/* Current CU */
+	Dwarf_Die		sp_die;
 	int			found;
 };
 
-- 
1.7.2.3


^ permalink raw reply related	[flat|nested] 11+ messages in thread

* Re: [PATCH v2 -tip] perf probe: Add fastpath to do lookup by function name
  2011-03-24 14:09     ` [PATCH v2 -tip] " Lin Ming
@ 2011-03-25  1:14       ` Masami Hiramatsu
  2011-03-25  2:57         ` Arnaldo Carvalho de Melo
  0 siblings, 1 reply; 11+ messages in thread
From: Masami Hiramatsu @ 2011-03-25  1:14 UTC (permalink / raw)
  To: Lin Ming
  Cc: Arnaldo Carvalho de Melo, Peter Zijlstra, Ingo Molnar,
	linux-kernel

(2011/03/24 23:09), Lin Ming wrote:
> v2 -> v1:
> - Don't compare file names with cu_find_realpath(...), instead, compare them
>   with the name returned by dwarf_decl_file(sp_die) 
> 
> The vmlinux file may have thousands of CUs.
> We can lookup function name from .debug_pubnames section
> to avoid the slow loop on CUs.
> 
> ./perf stat -r 10 -- ./perf probe -k /home/mlin/vmlinux \
>         -s /home/mlin/linux-2.6 \
> 	--line csum_partial_copy_to_user > tmp.log
> 
> before patch applied
> =====================
>         364.535892 task-clock-msecs         #      0.997 CPUs
>                  0 context-switches         #      0.000 M/sec
>                  0 CPU-migrations           #      0.000 M/sec
>             29,993 page-faults              #      0.082 M/sec
>        865,862,109 cycles                   #   2375.245 M/sec
>      1,255,259,630 instructions             #      1.450 IPC
>        252,400,884 branches                 #    692.390 M/sec
>          3,429,376 branch-misses            #      1.359 %
>          1,386,990 cache-references         #      3.805 M/sec
>            687,188 cache-misses             #      1.885 M/sec
> 
>         0.365792170  seconds time elapsed
> 
> after patch applied
> =====================
>          89.896405 task-clock-msecs         #      0.991 CPUs
>                  1 context-switches         #      0.000 M/sec
>                  0 CPU-migrations           #      0.000 M/sec
>             10,145 page-faults              #      0.113 M/sec
>        214,553,875 cycles                   #   2386.679 M/sec
>        226,915,559 instructions             #      1.058 IPC
>         44,536,614 branches                 #    495.422 M/sec
>            613,074 branch-misses            #      1.377 %
>            860,787 cache-references         #      9.575 M/sec
>            442,380 cache-misses             #      4.921 M/sec
> 
>         0.090716032  seconds time elapsed

Thanks! Looks very good :)

Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>

> 
> Signed-off-by: Lin Ming <ming.m.lin@intel.com>
> ---
>  tools/perf/util/probe-finder.c |   39 +++++++++++++++++++++++++++++++++++++++
>  tools/perf/util/probe-finder.h |    1 +
>  2 files changed, 40 insertions(+), 0 deletions(-)
> 
> diff --git a/tools/perf/util/probe-finder.c b/tools/perf/util/probe-finder.c
> index 194f9e2..5cf044c 100644
> --- a/tools/perf/util/probe-finder.c
> +++ b/tools/perf/util/probe-finder.c
> @@ -1876,6 +1876,31 @@ static int find_line_range_by_func(struct line_finder *lf)
>  	return param.retval;
>  }
>  
> +static int pubname_search_cb(Dwarf *dbg, Dwarf_Global *gl, void *data)
> +{
> +	struct line_finder *lf = data;
> +	struct line_range *lr = lf->lr;
> +
> +	if (dwarf_offdie(dbg, gl->die_offset, &lf->sp_die)) {
> +		if (dwarf_tag(&lf->sp_die) != DW_TAG_subprogram)
> +			return DWARF_CB_OK;
> +
> +		if (die_compare_name(&lf->sp_die, lr->function)) {
> +			if (!dwarf_offdie(dbg, gl->cu_offset, &lf->cu_die))
> +				return DWARF_CB_OK;
> +
> +			if (lr->file &&
> +			    strtailcmp(lr->file, dwarf_decl_file(&lf->sp_die)))
> +				return DWARF_CB_OK;
> +
> +			lf->found = 1;
> +			return DWARF_CB_ABORT;
> +		}
> +	}
> +
> +	return DWARF_CB_OK;
> +}
> +
>  int find_line_range(int fd, struct line_range *lr)
>  {
>  	struct line_finder lf = {.lr = lr, .found = 0};
> @@ -1895,6 +1920,19 @@ int find_line_range(int fd, struct line_range *lr)
>  		return -EBADF;
>  	}
>  
> +	/* Fastpath: lookup by function name from .debug_pubnames section */
> +	if (lr->function) {
> +		struct dwarf_callback_param param = {.data = (void *)&lf, .retval = 0};
> +
> +		dwarf_getpubnames(dbg, pubname_search_cb, &lf, 0);
> +		if (lf.found) {
> +			lf.found = 0;
> +			line_range_search_cb(&lf.sp_die, &param);
> +			if (lf.found)
> +				goto found;
> +		}
> +	}
> +
>  	/* Loop on CUs (Compilation Unit) */
>  	while (!lf.found && ret >= 0) {
>  		if (dwarf_nextcu(dbg, off, &noff, &cuhl, NULL, NULL, NULL) != 0)
> @@ -1923,6 +1961,7 @@ int find_line_range(int fd, struct line_range *lr)
>  		off = noff;
>  	}
>  
> +found:
>  	/* Store comp_dir */
>  	if (lf.found) {
>  		comp_dir = cu_get_comp_dir(&lf.cu_die);
> diff --git a/tools/perf/util/probe-finder.h b/tools/perf/util/probe-finder.h
> index beaefc3..4bc56a4 100644
> --- a/tools/perf/util/probe-finder.h
> +++ b/tools/perf/util/probe-finder.h
> @@ -83,6 +83,7 @@ struct line_finder {
>  	int			lno_s;		/* Start line number */
>  	int			lno_e;		/* End line number */
>  	Dwarf_Die		cu_die;		/* Current CU */
> +	Dwarf_Die		sp_die;
>  	int			found;
>  };
>  


-- 
Masami HIRAMATSU
2nd Dept. Linux Technology Center
Hitachi, Ltd., Systems Development Laboratory
E-mail: masami.hiramatsu.pt@hitachi.com

^ permalink raw reply	[flat|nested] 11+ messages in thread

* Re: [PATCH v2 -tip] perf probe: Add fastpath to do lookup by function name
  2011-03-25  1:14       ` Masami Hiramatsu
@ 2011-03-25  2:57         ` Arnaldo Carvalho de Melo
  2011-03-25  6:33           ` Lin Ming
  0 siblings, 1 reply; 11+ messages in thread
From: Arnaldo Carvalho de Melo @ 2011-03-25  2:57 UTC (permalink / raw)
  To: Masami Hiramatsu; +Cc: Lin Ming, Peter Zijlstra, Ingo Molnar, linux-kernel

Em Fri, Mar 25, 2011 at 10:14:25AM +0900, Masami Hiramatsu escreveu:
> (2011/03/24 23:09), Lin Ming wrote:
> > v2 -> v1:
> > - Don't compare file names with cu_find_realpath(...), instead, compare them
> >   with the name returned by dwarf_decl_file(sp_die) 
> > 
> > The vmlinux file may have thousands of CUs.
> > We can lookup function name from .debug_pubnames section
> > to avoid the slow loop on CUs.
> > 
> > ./perf stat -r 10 -- ./perf probe -k /home/mlin/vmlinux \
> >         -s /home/mlin/linux-2.6 \
> > 	--line csum_partial_copy_to_user > tmp.log
> > 
> > before patch applied
> > =====================
> >         364.535892 task-clock-msecs         #      0.997 CPUs
> >                  0 context-switches         #      0.000 M/sec
> >                  0 CPU-migrations           #      0.000 M/sec
> >             29,993 page-faults              #      0.082 M/sec
> >        865,862,109 cycles                   #   2375.245 M/sec
> >      1,255,259,630 instructions             #      1.450 IPC
> >        252,400,884 branches                 #    692.390 M/sec
> >          3,429,376 branch-misses            #      1.359 %
> >          1,386,990 cache-references         #      3.805 M/sec
> >            687,188 cache-misses             #      1.885 M/sec
> > 
> >         0.365792170  seconds time elapsed
> > 
> > after patch applied
> > =====================
> >          89.896405 task-clock-msecs         #      0.991 CPUs
> >                  1 context-switches         #      0.000 M/sec
> >                  0 CPU-migrations           #      0.000 M/sec
> >             10,145 page-faults              #      0.113 M/sec
> >        214,553,875 cycles                   #   2386.679 M/sec
> >        226,915,559 instructions             #      1.058 IPC
> >         44,536,614 branches                 #    495.422 M/sec
> >            613,074 branch-misses            #      1.377 %
> >            860,787 cache-references         #      9.575 M/sec
> >            442,380 cache-misses             #      4.921 M/sec
> > 
> >         0.090716032  seconds time elapsed
> 
> Thanks! Looks very good :)
> 
> Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>

Indeed, I'll try and process this one tomorrow,

Thanks a lot!

- Arnaldo

^ permalink raw reply	[flat|nested] 11+ messages in thread

* Re: [PATCH v2 -tip] perf probe: Add fastpath to do lookup by function name
  2011-03-25  2:57         ` Arnaldo Carvalho de Melo
@ 2011-03-25  6:33           ` Lin Ming
  2011-03-25  8:30             ` Lin Ming
  0 siblings, 1 reply; 11+ messages in thread
From: Lin Ming @ 2011-03-25  6:33 UTC (permalink / raw)
  To: Arnaldo Carvalho de Melo
  Cc: Masami Hiramatsu, Peter Zijlstra, Ingo Molnar, linux-kernel

在 2011-03-25五的 10:57 +0800,Arnaldo Carvalho de Melo写道:
> Em Fri, Mar 25, 2011 at 10:14:25AM +0900, Masami Hiramatsu escreveu:
> > (2011/03/24 23:09), Lin Ming wrote:
> > > v2 -> v1:
> > > - Don't compare file names with cu_find_realpath(...), instead, compare them
> > >   with the name returned by dwarf_decl_file(sp_die) 
> > > 
> > > The vmlinux file may have thousands of CUs.
> > > We can lookup function name from .debug_pubnames section
> > > to avoid the slow loop on CUs.
> > > 
> > > ./perf stat -r 10 -- ./perf probe -k /home/mlin/vmlinux \
> > >         -s /home/mlin/linux-2.6 \
> > > 	--line csum_partial_copy_to_user > tmp.log
> > > 
> > > before patch applied
> > > =====================
> > >         364.535892 task-clock-msecs         #      0.997 CPUs
> > >                  0 context-switches         #      0.000 M/sec
> > >                  0 CPU-migrations           #      0.000 M/sec
> > >             29,993 page-faults              #      0.082 M/sec
> > >        865,862,109 cycles                   #   2375.245 M/sec
> > >      1,255,259,630 instructions             #      1.450 IPC
> > >        252,400,884 branches                 #    692.390 M/sec
> > >          3,429,376 branch-misses            #      1.359 %
> > >          1,386,990 cache-references         #      3.805 M/sec
> > >            687,188 cache-misses             #      1.885 M/sec
> > > 
> > >         0.365792170  seconds time elapsed
> > > 
> > > after patch applied
> > > =====================
> > >          89.896405 task-clock-msecs         #      0.991 CPUs
> > >                  1 context-switches         #      0.000 M/sec
> > >                  0 CPU-migrations           #      0.000 M/sec
> > >             10,145 page-faults              #      0.113 M/sec
> > >        214,553,875 cycles                   #   2386.679 M/sec
> > >        226,915,559 instructions             #      1.058 IPC
> > >         44,536,614 branches                 #    495.422 M/sec
> > >            613,074 branch-misses            #      1.377 %
> > >            860,787 cache-references         #      9.575 M/sec
> > >            442,380 cache-misses             #      4.921 M/sec
> > > 
> > >         0.090716032  seconds time elapsed
> > 
> > Thanks! Looks very good :)
> > 
> > Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
> 
> Indeed, I'll try and process this one tomorrow,

Except find_line_range, I just realized that the same optimization maybe
added for find_probes.

Thanks,
Lin Ming

> 
> Thanks a lot!
> 
> - Arnaldo



^ permalink raw reply	[flat|nested] 11+ messages in thread

* Re: [PATCH v2 -tip] perf probe: Add fastpath to do lookup by function name
  2011-03-25  6:33           ` Lin Ming
@ 2011-03-25  8:30             ` Lin Ming
  0 siblings, 0 replies; 11+ messages in thread
From: Lin Ming @ 2011-03-25  8:30 UTC (permalink / raw)
  To: Arnaldo Carvalho de Melo
  Cc: Masami Hiramatsu, Peter Zijlstra, Ingo Molnar, linux-kernel

On Fri, 2011-03-25 at 14:33 +0800, Lin Ming wrote:
> 在 2011-03-25五的 10:57 +0800,Arnaldo Carvalho de Melo写道:
> > Em Fri, Mar 25, 2011 at 10:14:25AM +0900, Masami Hiramatsu escreveu:
> > > (2011/03/24 23:09), Lin Ming wrote:
> > > > v2 -> v1:
> > > > - Don't compare file names with cu_find_realpath(...), instead, compare them
> > > >   with the name returned by dwarf_decl_file(sp_die) 
> > > > 
> > > > The vmlinux file may have thousands of CUs.
> > > > We can lookup function name from .debug_pubnames section
> > > > to avoid the slow loop on CUs.
> > > > 
> > > > ./perf stat -r 10 -- ./perf probe -k /home/mlin/vmlinux \
> > > >         -s /home/mlin/linux-2.6 \
> > > > 	--line csum_partial_copy_to_user > tmp.log
> > > > 
> > > > before patch applied
> > > > =====================
> > > >         364.535892 task-clock-msecs         #      0.997 CPUs
> > > >                  0 context-switches         #      0.000 M/sec
> > > >                  0 CPU-migrations           #      0.000 M/sec
> > > >             29,993 page-faults              #      0.082 M/sec
> > > >        865,862,109 cycles                   #   2375.245 M/sec
> > > >      1,255,259,630 instructions             #      1.450 IPC
> > > >        252,400,884 branches                 #    692.390 M/sec
> > > >          3,429,376 branch-misses            #      1.359 %
> > > >          1,386,990 cache-references         #      3.805 M/sec
> > > >            687,188 cache-misses             #      1.885 M/sec
> > > > 
> > > >         0.365792170  seconds time elapsed
> > > > 
> > > > after patch applied
> > > > =====================
> > > >          89.896405 task-clock-msecs         #      0.991 CPUs
> > > >                  1 context-switches         #      0.000 M/sec
> > > >                  0 CPU-migrations           #      0.000 M/sec
> > > >             10,145 page-faults              #      0.113 M/sec
> > > >        214,553,875 cycles                   #   2386.679 M/sec
> > > >        226,915,559 instructions             #      1.058 IPC
> > > >         44,536,614 branches                 #    495.422 M/sec
> > > >            613,074 branch-misses            #      1.377 %
> > > >            860,787 cache-references         #      9.575 M/sec
> > > >            442,380 cache-misses             #      4.921 M/sec
> > > > 
> > > >         0.090716032  seconds time elapsed
> > > 
> > > Thanks! Looks very good :)
> > > 
> > > Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
> > 
> > Indeed, I'll try and process this one tomorrow,
> 
> Except find_line_range, I just realized that the same optimization maybe
> added for find_probes.

I have send out a v3 patch to add fastpath for find_probes.

Thanks,
Lin Ming

> 
> Thanks,
> Lin Ming
> 
> > 
> > Thanks a lot!
> > 
> > - Arnaldo
> 



^ permalink raw reply	[flat|nested] 11+ messages in thread

end of thread, other threads:[~2011-03-25  8:30 UTC | newest]

Thread overview: 11+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2011-03-24 15:38 [PATCH] perf probe: Add fastpath to do lookup by function name Lin Ming
2011-03-24  7:58 ` Ingo Molnar
2011-03-24  8:38   ` Lin Ming
2011-03-24  8:47     ` Ingo Molnar
2011-03-24  9:08 ` Masami Hiramatsu
2011-03-24 13:47   ` Lin Ming
2011-03-24 14:09     ` [PATCH v2 -tip] " Lin Ming
2011-03-25  1:14       ` Masami Hiramatsu
2011-03-25  2:57         ` Arnaldo Carvalho de Melo
2011-03-25  6:33           ` Lin Ming
2011-03-25  8:30             ` Lin Ming

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox