Skip to contentSkip to conversation

perfcounter: Ignore the nmi call frames in the x86-64 backtraces

PATCH In review Frederic Weisbecker 2 participants 5 messages Jul 1, 2009 – Jul 7, 2009

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 perfcounter: Ignore the nmi call frames in the x86-64 backtraces

About every callchains recorded with perf record are filled up including the internal perfcounter nmi frame:

perf_callchain
perf_counter_overflow
intel_pmu_handle_irq
perf_counter_nmi_handler
notifier_call_chain
atomic_notifier_call_chain
notify_die
do_nmi
nmi

We want ignore this frame as it's not interesting for instrumentation. To solve this, we simply ignore every frames from nmi context.

New example of "perf report -s sym -c" after this patch:

9.59% [k] search_by_key

             4.88%
                search_by_key
                reiserfs_read_locked_inode
                reiserfs_iget
                reiserfs_lookup
                do_lookup
                __link_path_walk
                path_walk
                do_path_lookup
                user_path_at
                vfs_fstatat
                vfs_lstat
                sys_newlstat
                system_call_fastpath
                __lxstat
                0x406fb1

             3.19%
                search_by_key
                search_by_entry_key
                reiserfs_find_entry
                reiserfs_lookup
                do_lookup
                __link_path_walk
                path_walk
                do_path_lookup
                user_path_at
                vfs_fstatat
                vfs_lstat
                sys_newlstat
                system_call_fastpath
                __lxstat
                0x406fb1

[...]

For now this patch only solves the problem in x86-64.

---
 arch/x86/include/asm/stacktrace.h  |    2 ++
 arch/x86/kernel/cpu/perf_counter.c |    8 +++++++-
 arch/x86/kernel/dumpstack_32.c     |    6 ++++++
 arch/x86/kernel/dumpstack_64.c     |   22 +++++++++++++++-------
 4 files changed, 30 insertions(+), 8 deletions(-)
arch/x86/include/asm/stacktrace.h +2 -0
index f517944..cf86a5e 100644
@@ -3,6 +3,8 @@
 
 extern int kstack_depth_to_print;
 
+int x86_is_stack_id(int id, char *name);
+
 /* Generic stack tracer with callbacks */
 
 struct stacktrace_ops {
arch/x86/kernel/cpu/perf_counter.c +7 -1
index d4cf4ce..36c3dc7 100644
@@ -1561,6 +1561,7 @@ void callchain_store(struct perf_callchain_entry *entry, u64 ip)
 
 static DEFINE_PER_CPU(struct perf_callchain_entry, irq_entry);
 static DEFINE_PER_CPU(struct perf_callchain_entry, nmi_entry);
+static DEFINE_PER_CPU(int, in_nmi_frame);
 
 
 static void
@@ -1576,7 +1577,9 @@ static void backtrace_warning(void *data, char *msg)
 
 static int backtrace_stack(void *data, char *name)
 {
-	/* Process all stacks: */
+	per_cpu(in_nmi_frame, smp_processor_id()) =
+			x86_is_stack_id(NMI_STACK, name);
+
 	return 0;
 }
 
@@ -1584,6 +1587,9 @@ static void backtrace_address(void *data, unsigned long addr, int reliable)
 {
 	struct perf_callchain_entry *entry = data;
 
+	if (per_cpu(in_nmi_frame, smp_processor_id()))
+		return;
+
 	if (reliable)
 		callchain_store(entry, addr);
 }
arch/x86/kernel/dumpstack_32.c +6 -0
index d593cd1..bca5fba 100644
@@ -19,6 +19,12 @@
 
 #include "dumpstack.h"
 
+/* Just a stub for now */
+int x86_is_stack_id(int id, char *name)
+{
+	return 0;
+}
+
 void dump_trace(struct task_struct *task, struct pt_regs *regs,
 		unsigned long *stack, unsigned long bp,
 		const struct stacktrace_ops *ops, void *data)
arch/x86/kernel/dumpstack_64.c +15 -7
index d35db59..54b0a32 100644
@@ -19,10 +19,8 @@
 
 #include "dumpstack.h"
 
-static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
-					unsigned *usedp, char **idp)
-{
-	static char ids[][8] = {
+
+static char x86_stack_ids[][8] = {
 		[DEBUG_STACK - 1] = "#DB",
 		[NMI_STACK - 1] = "NMI",
 		[DOUBLEFAULT_STACK - 1] = "#DF",
@@ -33,6 +31,15 @@ static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
 			N_EXCEPTION_STACKS + DEBUG_STKSZ / EXCEPTION_STKSZ - 2] = "#DB[?]"
 #endif
 	};
+
+int x86_is_stack_id(int id, char *name)
+{
+	return x86_stack_ids[id - 1] == name;
+}
+
+static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
+					unsigned *usedp, char **idp)
+{
 	unsigned k;
 
 	/*
@@ -61,7 +68,7 @@ static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
 			if (*usedp & (1U << k))
 				break;
 			*usedp |= 1U << k;
-			*idp = ids[k];
+			*idp = x86_stack_ids[k];
 			return (unsigned long *)end;
 		}
 		/*
@@ -81,12 +88,13 @@ static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
 			do {
 				++j;
 				end -= EXCEPTION_STKSZ;
-				ids[j][4] = '1' + (j - N_EXCEPTION_STACKS);
+				x86_stack_ids[j][4] = '1' +
+						(j - N_EXCEPTION_STACKS);
 			} while (stack < end - EXCEPTION_STKSZ);
 			if (*usedp & (1U << j))
 				break;
 			*usedp |= 1U << j;
-			*idp = ids[j];
+			*idp = x86_stack_ids[j];
 			return (unsigned long *)end;
 		}
 #endif
Signature · 1 line
1.6.2.3
#1 lore raw

PATCH 2/2 perfcounter: Handle pipe read failures in perf stat

Building builtin-stat.c reports the following errors:

cc1: warnings being treated as errors
builtin-stat.c: In function ‘run_perf_stat’:
builtin-stat.c:242: erreur: ignoring return value of ‘read’, declared with attribute warn_unused_result builtin-stat.c:255: erreur: ignoring return value of ‘read’, declared with attribute warn_unused_result
make: *** [builtin-stat.o] Erreur 1

This patch handles the possible pipe read failures.

---
 tools/perf/builtin-stat.c |    6 ++++--
 1 files changed, 4 insertions(+), 2 deletions(-)
tools/perf/builtin-stat.c +4 -2
index 01cc07e..27921a8 100644
@@ -239,7 +239,8 @@ static int run_perf_stat(int argc __used, const char **argv)
 		/*
 		 * Wait until the parent tells us to go.
 		 */
-		read(go_pipe[0], &buf, 1);
+		if (read(go_pipe[0], &buf, 1) == -1)
+			perror("unable to read pipe");
 
 		execvp(argv[0], (char **)argv);
 
@@ -252,7 +253,8 @@ static int run_perf_stat(int argc __used, const char **argv)
 	 */
 	close(child_ready_pipe[1]);
 	close(go_pipe[0]);
-	read(child_ready_pipe[0], &buf, 1);
+	if (read(child_ready_pipe[0], &buf, 1) == -1)
+		perror("unable to read pipe");
 	close(child_ready_pipe[0]);
 
 	for (counter = 0; counter < nr_counters; counter++)
Signature · 1 line
1.6.2.3
#2 lore raw

[tip:perfcounters/urgent] perf_counter: Ignore the nmi call frames in the x86-64 backtraces

In reply to Frederic Weisbecker
Commit-ID:  0406ca6d8e849d9dd027c8cb6791448e81411aef
Gitweb:     http://git.kernel.org/tip/0406ca6d8e849d9dd027c8cb6791448e81411aef
Author:     Frederic Weisbecker <[email protected]>
AuthorDate: Wed, 1 Jul 2009 21:02:09 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Wed, 1 Jul 2009 22:37:23 +0200

perf_counter: Ignore the nmi call frames in the x86-64 backtraces

About every callchains recorded with perf record are filled up including the internal perfcounter nmi frame:

perf_callchain
perf_counter_overflow
intel_pmu_handle_irq
perf_counter_nmi_handler
notifier_call_chain
atomic_notifier_call_chain
notify_die
do_nmi
nmi

We want ignore this frame as it's not interesting for instrumentation. To solve this, we simply ignore every frames from nmi context.

New example of "perf report -s sym -c" after this patch:

9.59% [k] search_by_key

             4.88%
                search_by_key
                reiserfs_read_locked_inode
                reiserfs_iget
                reiserfs_lookup
                do_lookup
                __link_path_walk
                path_walk
                do_path_lookup
                user_path_at
                vfs_fstatat
                vfs_lstat
                sys_newlstat
                system_call_fastpath
                __lxstat
                0x406fb1

             3.19%
                search_by_key
                search_by_entry_key
                reiserfs_find_entry
                reiserfs_lookup
                do_lookup
                __link_path_walk
                path_walk
                do_path_lookup
                user_path_at
                vfs_fstatat
                vfs_lstat
                sys_newlstat
                system_call_fastpath
                __lxstat
                0x406fb1

[...]

For now this patch only solves the problem in x86-64.

Cc Peter Zijlstra <[email protected]>
Cc Mike Galbraith <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>

LKML-Reference: <[email protected]>

---
 arch/x86/include/asm/stacktrace.h  |    2 ++
 arch/x86/kernel/cpu/perf_counter.c |    8 +++++++-
 arch/x86/kernel/dumpstack_32.c     |    6 ++++++
 arch/x86/kernel/dumpstack_64.c     |   22 +++++++++++++++-------
 4 files changed, 30 insertions(+), 8 deletions(-)
arch/x86/include/asm/stacktrace.h +2 -0
index f517944..cf86a5e 100644
@@ -3,6 +3,8 @@
 
 extern int kstack_depth_to_print;
 
+int x86_is_stack_id(int id, char *name);
+
 /* Generic stack tracer with callbacks */
 
 struct stacktrace_ops {
arch/x86/kernel/cpu/perf_counter.c +7 -1
index d4cf4ce..36c3dc7 100644
@@ -1561,6 +1561,7 @@ void callchain_store(struct perf_callchain_entry *entry, u64 ip)
 
 static DEFINE_PER_CPU(struct perf_callchain_entry, irq_entry);
 static DEFINE_PER_CPU(struct perf_callchain_entry, nmi_entry);
+static DEFINE_PER_CPU(int, in_nmi_frame);
 
 
 static void
@@ -1576,7 +1577,9 @@ static void backtrace_warning(void *data, char *msg)
 
 static int backtrace_stack(void *data, char *name)
 {
-	/* Process all stacks: */
+	per_cpu(in_nmi_frame, smp_processor_id()) =
+			x86_is_stack_id(NMI_STACK, name);
+
 	return 0;
 }
 
@@ -1584,6 +1587,9 @@ static void backtrace_address(void *data, unsigned long addr, int reliable)
 {
 	struct perf_callchain_entry *entry = data;
 
+	if (per_cpu(in_nmi_frame, smp_processor_id()))
+		return;
+
 	if (reliable)
 		callchain_store(entry, addr);
 }
arch/x86/kernel/dumpstack_32.c +6 -0
index d593cd1..bca5fba 100644
@@ -19,6 +19,12 @@
 
 #include "dumpstack.h"
 
+/* Just a stub for now */
+int x86_is_stack_id(int id, char *name)
+{
+	return 0;
+}
+
 void dump_trace(struct task_struct *task, struct pt_regs *regs,
 		unsigned long *stack, unsigned long bp,
 		const struct stacktrace_ops *ops, void *data)
arch/x86/kernel/dumpstack_64.c +15 -7
index d35db59..54b0a32 100644
@@ -19,10 +19,8 @@
 
 #include "dumpstack.h"
 
-static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
-					unsigned *usedp, char **idp)
-{
-	static char ids[][8] = {
+
+static char x86_stack_ids[][8] = {
 		[DEBUG_STACK - 1] = "#DB",
 		[NMI_STACK - 1] = "NMI",
 		[DOUBLEFAULT_STACK - 1] = "#DF",
@@ -33,6 +31,15 @@ static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
 			N_EXCEPTION_STACKS + DEBUG_STKSZ / EXCEPTION_STKSZ - 2] = "#DB[?]"
 #endif
 	};
+
+int x86_is_stack_id(int id, char *name)
+{
+	return x86_stack_ids[id - 1] == name;
+}
+
+static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
+					unsigned *usedp, char **idp)
+{
 	unsigned k;
 
 	/*
@@ -61,7 +68,7 @@ static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
 			if (*usedp & (1U << k))
 				break;
 			*usedp |= 1U << k;
-			*idp = ids[k];
+			*idp = x86_stack_ids[k];
 			return (unsigned long *)end;
 		}
 		/*
@@ -81,12 +88,13 @@ static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
 			do {
 				++j;
 				end -= EXCEPTION_STKSZ;
-				ids[j][4] = '1' + (j - N_EXCEPTION_STACKS);
+				x86_stack_ids[j][4] = '1' +
+						(j - N_EXCEPTION_STACKS);
 			} while (stack < end - EXCEPTION_STKSZ);
 			if (*usedp & (1U << j))
 				break;
 			*usedp |= 1U << j;
-			*idp = ids[j];
+			*idp = x86_stack_ids[j];
 			return (unsigned long *)end;
 		}
 #endif
#3 lore raw
In reply to Frederic Weisbecker

arch/x86/kernel/dumpstack_64.c static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack, · line 38

Quoted 4 lines · Frederic Weisbecker
arch/x86/kernel/dumpstack_64.c +4 -0
+int x86_is_stack_id(int id, char *name)
+{
+	return x86_stack_ids[id - 1] == name;
+}

OK, this bit would need a bit of a comment explaining why this works, as I don't think the C language guarantees de-duplication of constant strings.

Therefore the above only works correctly if its passed a pointer from the x86_stack_ids[] array -- as it indeed is, since the string passed is obtained through the below method:

Quoted 15 lines
arch/x86/kernel/dumpstack_64.c +4 -1
+static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
+					unsigned *usedp, char **idp)
+{
 	unsigned k;
 
 	/*
@@ -61,7 +68,7 @@ static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
 			if (*usedp & (1U << k))
 				break;
 			*usedp |= 1U << k;
-			*idp = ids[k];
+			*idp = x86_stack_ids[k];
 			return (unsigned long *)end;
 		}
 		/*
#4 lore raw
Quoted 11 lines · Peter Zijlstra

arch/x86/kernel/dumpstack_64.c static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack, · line 38

Quoted 4 lines · Frederic Weisbecker
arch/x86/kernel/dumpstack_64.c +4 -0
+int x86_is_stack_id(int id, char *name)
+{
+	return x86_stack_ids[id - 1] == name;
+}

OK, this bit would need a bit of a comment explaining why this works, as I don't think the C language guarantees de-duplication of constant strings.

Quoted 3 lines

Therefore the above only works correctly if its passed a pointer from the x86_stack_ids[] array -- as it indeed is, since the string passed is obtained through the below method:

Right, that need some comments, I will do that with x86-32 support at the same time.

Thanks.

Quoted 17 lines
Quoted 15 lines
arch/x86/kernel/dumpstack_64.c +4 -1
+static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
+					unsigned *usedp, char **idp)
+{
 	unsigned k;
 
 	/*
@@ -61,7 +68,7 @@ static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
 			if (*usedp & (1U << k))
 				break;
 			*usedp |= 1U << k;
-			*idp = ids[k];
+			*idp = x86_stack_ids[k];
 			return (unsigned long *)end;
 		}
 		/*