Scheduler Profiling - Use Immediate Values

Scheduler Profiling - Use Immediate Values

Use immediate values with lower d-cache hit in optimized version as a condition for scheduler profiling call.

Depends on the Immediate values patches, applies to 2.6.23-rc4-mm1.

Signed-off-by Mathieu Desnoyers <[email protected]>

---
 drivers/kvm/svm.c       |    2 +-
 drivers/kvm/vmx.c       |    2 +-
 include/linux/profile.h |   10 ++++------
 kernel/profile.c        |   38 ++++++++++++++++++++++++++------------
 kernel/sched.c          |    3 ++-
 5 files changed, 34 insertions(+), 21 deletions(-)

Index: linux-2.6-lttng/kernel/profile.c
===================================================================

kernel/profile.c +8 -3

@@ -42,9 +42,6 @@ static int (*timer_hook)(struct pt_regs 
 static atomic_t *prof_buffer;
 static unsigned long prof_len, prof_shift;
 
-int prof_on __read_mostly;
-EXPORT_SYMBOL_GPL(prof_on);
-
 static cpumask_t prof_cpu_mask = CPU_MASK_ALL;
 #ifdef CONFIG_SMP
 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);

@@ -52,6 +49,14 @@ static DEFINE_PER_CPU(int, cpu_profile_f
 static DEFINE_MUTEX(profile_flip_mutex);
 #endif /* CONFIG_SMP */
 
+/* Immediate values */
+immediate_char_t sleep_profiling __read_mostly,
+			sched_profiling __read_mostly,
+			kvm_profiling __read_mostly,
+			cpu_profiling __read_mostly;
+EXPORT_SYMBOL_GPL(kvm_profiling);
+EXPORT_SYMBOL_GPL(cpu_profiling);
+
 static int __init profile_setup(char * str)
 {
 	static char __initdata schedstr[] = "schedule";

kernel/profile.c +6 -3

 		return;
 
 	/* only text is profiled */

@@ -288,7 +296,7 @@ void profile_hits(int type, void *__pc, 
 	int i, j, cpu;
 	struct profile_hit *hits;
 
-	if (prof_on != type || !prof_buffer)
+	if (!prof_buffer)
 		return;
 	pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
 	i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;

@@ -398,7 +406,7 @@ void profile_hits(int type, void *__pc, 
 {
 	unsigned long pc;
 
-	if (prof_on != type || !prof_buffer)
+	if (!prof_buffer)
 		return;
 	pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
 	atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);

@@ -555,7 +563,10 @@ static int __init create_hash_tables(voi
 	}
 	return 0;
 out_cleanup:
-	prof_on = 0;
+	immediate_set_early(&sleep_profiling, 0);
+	immediate_set_early(&sched_profiling, 0);
+	immediate_set_early(&kvm_profiling, 0);
+	immediate_set_early(&cpu_profiling, 0);
 	smp_mb();
 	on_each_cpu(profile_nop, NULL, 0, 1);
 	for_each_online_cpu(cpu) {

@@ -582,7 +593,10 @@ static int __init create_proc_profile(vo
 {

Quoted 46 lines · On Thu, 6 Sep 2007 16:05:37 -0400 Mathieu Desnoyers wrote:

Scheduler Profiling - Use Immediate Values

Use immediate values with lower d-cache hit in optimized version as a condition for scheduler profiling call.

Depends on the Immediate values patches, applies to 2.6.23-rc4-mm1.

Signed-off-by Mathieu Desnoyers <[email protected]>

---
 drivers/kvm/svm.c       |    2 +-
 drivers/kvm/vmx.c       |    2 +-
 include/linux/profile.h |   10 ++++------
 kernel/profile.c        |   38 ++++++++++++++++++++++++++------------
 kernel/sched.c          |    3 ++-
 5 files changed, 34 insertions(+), 21 deletions(-)

Index: linux-2.6-lttng/kernel/profile.c
===================================================================

kernel/profile.c +8 -3

@@ -42,9 +42,6 @@ static int (*timer_hook)(struct pt_regs 
 static atomic_t *prof_buffer;
 static unsigned long prof_len, prof_shift;
 
-int prof_on __read_mostly;
-EXPORT_SYMBOL_GPL(prof_on);
-
 static cpumask_t prof_cpu_mask = CPU_MASK_ALL;
 #ifdef CONFIG_SMP
 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);

@@ -52,6 +49,14 @@ static DEFINE_PER_CPU(int, cpu_profile_f
 static DEFINE_MUTEX(profile_flip_mutex);
 #endif /* CONFIG_SMP */
 
+/* Immediate values */
+immediate_char_t sleep_profiling __read_mostly,
+			sched_profiling __read_mostly,
+			kvm_profiling __read_mostly,
+			cpu_profiling __read_mostly;
+EXPORT_SYMBOL_GPL(kvm_profiling);
+EXPORT_SYMBOL_GPL(cpu_profiling);
+
 static int __init profile_setup(char * str)
 {
 	static char __initdata schedstr[] = "schedule";

kernel/profile.c static int __init profile_setup(char * s · line 68

Quoted 5 lines

kernel/profile.c +1 -1

@@ -60,7 +65,7 @@ static int __init profile_setup(char * s
 	int par;
 
 	if (!strncmp(str, sleepstr, strlen(sleepstr))) {
-		prof_on = SLEEP_PROFILING;
+		immediate_set_early(&sleep_profiling, 1);

These are booleans, eh? so one could use , true);

etc.

Quoted 5 lines

kernel/profile.c +0 -0

 		if (str[strlen(sleepstr)] == ',')
 			str += strlen(sleepstr) + 1;
 		if (get_option(&str, &par))

@@ -99,7 +104,10 @@ __setup("profile=", profile_setup);
 

kernel/profile.c __setup("profile=", profile_setup); · line 110

Quoted 7 lines

kernel/profile.c +4 -1

 void __init profile_init(void)
 {
-	if (!prof_on) 
+	if (!_immediate_read(&sleep_profiling) &&
+		!_immediate_read(&sched_profiling) &&
+		!_immediate_read(&kvm_profiling) &&
+		!_immediate_read(&cpu_profiling))

Looks like this (in multiple places) needs to use prof_on, where prof_on is set earlier as the logical 'or' of all (4 now, could be more) profiling flags.

Or below, sometimes you just check prof_buffer. Is that sufficient here also?

Quoted 35 lines

kernel/profile.c +6 -3

 		return;
 
 	/* only text is profiled */

@@ -288,7 +296,7 @@ void profile_hits(int type, void *__pc, 
 	int i, j, cpu;
 	struct profile_hit *hits;
 
-	if (prof_on != type || !prof_buffer)
+	if (!prof_buffer)
 		return;
 	pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
 	i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;

@@ -398,7 +406,7 @@ void profile_hits(int type, void *__pc, 
 {
 	unsigned long pc;
 
-	if (prof_on != type || !prof_buffer)
+	if (!prof_buffer)
 		return;
 	pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
 	atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);

@@ -555,7 +563,10 @@ static int __init create_hash_tables(voi
 	}
 	return 0;
 out_cleanup:
-	prof_on = 0;
+	immediate_set_early(&sleep_profiling, 0);
+	immediate_set_early(&sched_profiling, 0);
+	immediate_set_early(&kvm_profiling, 0);
+	immediate_set_early(&cpu_profiling, 0);
 	smp_mb();
 	on_each_cpu(profile_nop, NULL, 0, 1);
 	for_each_online_cpu(cpu) {

@@ -582,7 +593,10 @@ static int __init create_proc_profile(vo
 {

kernel/profile.c static int __init create_proc_profile(vo · line 602

Quoted 10 lines

kernel/profile.c +4 -1

 	struct proc_dir_entry *entry;
 
-	if (!prof_on)
+	if (!_immediate_read(&sleep_profiling) &&
+		!_immediate_read(&sched_profiling) &&
+		!_immediate_read(&kvm_profiling) &&
+		!_immediate_read(&cpu_profiling))
 		return 0;
 	if (create_hash_tables())
 		return -1;

---
~Randy
*** Remember to use Documentation/SubmitChecklist when testing your code ***

kernel/sched.c static inline void schedule_debug(struct · line 3461

Quoted 3 lines · Mathieu Desnoyers

kernel/sched.c +1 -1

-	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
+	immediate_if (&sched_profiling)

I must say I really dislike immediate_if(). You complained earlier that something breaks coloring, but adding such macros will definitely break a lot of editors (especially if you use it without {} like here)

It would be much nicer and readable to just use if (unlikely(immediate_read(&x)) or if you prefer to hide the unlikely if (immediate_bool_test(&x)) with an implicit unlikely().

-Andi

The idea is not to hide the unlikely, but to leave the opportunity to make this primitive evolve in something that won't depend on a load immediate and only require patching of a jump, given the appropriate gcc support (yet to come).