Skip to contentSkip to conversation

perf probe: Add fastpath to do lookup by function name

PATCH Superseded Lin Ming 4 participants 11 messages Mar 24, 2011 – Mar 25, 2011 perf events

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Lin Ming

author
#0 lore raw

PATCH 1/1 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.

Signed-off-by Lin Ming <[email protected]>
---
 tools/perf/util/probe-finder.c |   38 ++++++++++++++++++++++++++++++++++++++
 tools/perf/util/probe-finder.h |    1 +
 2 files changed, 39 insertions(+), 0 deletions(-)
tools/perf/util/probe-finder.c +38 -0
index 194f9e2..b2034c2 100644
@@ -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);
tools/perf/util/probe-finder.h +1 -0
index beaefc3..4bc56a4 100644
@@ -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;
 };
Signature · 1 line
1.7.2.3
#1 lore raw
Quoted 3 lines · Lin Ming

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

author
#2 lore raw
Quoted 11 lines · Ingo Molnar
Quoted 3 lines · Lin Ming

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 <[email protected]>
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 <[email protected]>
---
 tools/perf/util/probe-finder.c |   38 ++++++++++++++++++++++++++++++++++++++
 tools/perf/util/probe-finder.h |    1 +
 2 files changed, 39 insertions(+), 0 deletions(-)
tools/perf/util/probe-finder.c +38 -0
index 194f9e2..b2034c2 100644
@@ -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);
tools/perf/util/probe-finder.h +1 -0
index beaefc3..4bc56a4 100644
@@ -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;
 };
Signature · 1 line
1.7.2.3
#3 lore raw
Quoted 10 lines · Lin Ming

./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
Quoted 5 lines

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

#4 lore raw
In reply to Lin Ming
Quoted 21 lines · Lin Ming

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 <[email protected]>
---
 tools/perf/util/probe-finder.c |   38 ++++++++++++++++++++++++++++++++++++++
 tools/perf/util/probe-finder.h |    1 +
 2 files changed, 39 insertions(+), 0 deletions(-)
tools/perf/util/probe-finder.c +3 -0
index 194f9e2..b2034c2 100644
@@ -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;

tools/perf/util/probe-finder.c static int find_line_range_by_func(struct line_finder *lf) · line 1891

Quoted 10 lines
tools/perf/util/probe-finder.c +10 -0
+	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!

Quoted 54 lines
tools/perf/util/probe-finder.c +25 -0
+			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);
tools/perf/util/probe-finder.h +1 -0
index beaefc3..4bc56a4 100644
@@ -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;
 };
Signature · 4 lines
Masami HIRAMATSU
2nd Dept. Linux Technology Center
Hitachi, Ltd., Systems Development Laboratory
E-mail: [email protected]
author
#5 lore raw
Quoted 19 lines · Masami Hiramatsu
Quoted 18 lines · Lin Ming

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 <[email protected]>
---
 tools/perf/util/probe-finder.c |   38 ++++++++++++++++++++++++++++++++++++++
 tools/perf/util/probe-finder.h |    1 +
 2 files changed, 39 insertions(+), 0 deletions(-)
tools/perf/util/probe-finder.c +0 -0
index 194f9e2..b2034c2 100644
@@ -1876,6 +1876,30 @@ static int find_line_range_by_func(struct line_finder *lf)
 	return param.retval;
 }
Quoted 16 lines

tools/perf/util/probe-finder.c static int find_line_range_by_func(struct line_finder *lf) · line 1891

Quoted 13 lines
tools/perf/util/probe-finder.c +13 -0
+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.

tools/perf/util/probe-finder.c +7 -2
index b2034c2..38e4a05 100644
@@ -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;
author
#6 lore raw

PATCH v2 1/1 perf probe: Add fastpath to do lookup by function name

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 <[email protected]>
---
 tools/perf/util/probe-finder.c |   39 +++++++++++++++++++++++++++++++++++++++
 tools/perf/util/probe-finder.h |    1 +
 2 files changed, 40 insertions(+), 0 deletions(-)
tools/perf/util/probe-finder.c +39 -0
index 194f9e2..5cf044c 100644
@@ -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);
tools/perf/util/probe-finder.h +1 -0
index beaefc3..4bc56a4 100644
@@ -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;
 };
Signature · 1 line
1.7.2.3
Acked-by on v2 1/1
#7 lore raw
Quoted 40 lines · Lin Ming

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
Quoted 1 line
        0.090716032  seconds time elapsed

Thanks! Looks very good :)

Acked-by Masami Hiramatsu <[email protected]>
Quoted 83 lines
Signed-off-by Lin Ming <[email protected]>
---
 tools/perf/util/probe-finder.c |   39 +++++++++++++++++++++++++++++++++++++++
 tools/perf/util/probe-finder.h |    1 +
 2 files changed, 40 insertions(+), 0 deletions(-)
tools/perf/util/probe-finder.c +39 -0
index 194f9e2..5cf044c 100644
@@ -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);
tools/perf/util/probe-finder.h +1 -0
index beaefc3..4bc56a4 100644
@@ -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;
 };
Signature · 4 lines
Masami HIRAMATSU
2nd Dept. Linux Technology Center
Hitachi, Ltd., Systems Development Laboratory
E-mail: [email protected]
#8 lore raw

Em Fri, Mar 25, 2011 at 10:14:25AM +0900, Masami Hiramatsu escreveu:

Quoted 41 lines
Quoted 40 lines · Lin Ming

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
Quoted 5 lines
Quoted 1 line
        0.090716032  seconds time elapsed

Thanks! Looks very good :)

Acked-by Masami Hiramatsu <[email protected]>

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

Thanks a lot!

- Arnaldo

author
#9 lore raw
Quoted 42 lines · 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:

Quoted 41 lines
Quoted 40 lines · Lin Ming

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
Quoted 7 lines
Quoted 5 lines
Quoted 1 line
        0.090716032  seconds time elapsed

Thanks! Looks very good :)

Acked-by Masami Hiramatsu <[email protected]>

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

Quoted 4 lines

Thanks a lot!

- Arnaldo

author
#10 lore raw
Quoted 45 lines · Lin Ming
Quoted 44 lines · 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:

Quoted 43 lines
Quoted 41 lines · Lin Ming

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
Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

Thanks! Looks very good :)

Acked-by Masami Hiramatsu <[email protected]>

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

Quoted 9 lines

Thanks,
Lin Ming

Quoted 4 lines

Thanks a lot!

- Arnaldo