Skip to contentSkip to conversation

kallsyms: Fix scheduling with interrupts disabled in self-test

PATCH In review Nicholas Piggin 4 participants 5 messages Jan 12, 2023 – Jan 13, 2023

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 kallsyms: Fix scheduling with interrupts disabled in self-test

kallsyms_on_each* may schedule so must not be called with interrupts disabled. The iteration function could disable interrupts, but this also changes lookup_symbol() to match the change to the other timing code.

---
 kernel/kallsyms_selftest.c | 21 ++++++---------------
 1 file changed, 6 insertions(+), 15 deletions(-)
kernel/kallsyms_selftest.c +6 -15
index f35d9cc1aab1..bfbc12da3326 100644
@@ -157,14 +157,11 @@ static void test_kallsyms_compression_ratio(void)
 static int lookup_name(void *data, const char *name, struct module *mod, unsigned long addr)
 {
 	u64 t0, t1, t;
-	unsigned long flags;
 	struct test_stat *stat = (struct test_stat *)data;
 
-	local_irq_save(flags);
-	t0 = sched_clock();
+	t0 = ktime_get_ns();
 	(void)kallsyms_lookup_name(name);
-	t1 = sched_clock();
-	local_irq_restore(flags);
+	t1 = ktime_get_ns();
 
 	t = t1 - t0;
 	if (t < stat->min)
@@ -234,18 +231,15 @@ static int find_symbol(void *data, const char *name, struct module *mod, unsigne
 static void test_perf_kallsyms_on_each_symbol(void)
 {
 	u64 t0, t1;
-	unsigned long flags;
 	struct test_stat stat;
 
 	memset(&stat, 0, sizeof(stat));
 	stat.max = INT_MAX;
 	stat.name = stub_name;
 	stat.perf = 1;
-	local_irq_save(flags);
-	t0 = sched_clock();
+	t0 = ktime_get_ns();
 	kallsyms_on_each_symbol(find_symbol, &stat);
-	t1 = sched_clock();
-	local_irq_restore(flags);
+	t1 = ktime_get_ns();
 	pr_info("kallsyms_on_each_symbol() traverse all: %lld ns\n", t1 - t0);
 }
 
@@ -270,17 +264,14 @@ static int match_symbol(void *data, unsigned long addr)
 static void test_perf_kallsyms_on_each_match_symbol(void)
 {
 	u64 t0, t1;
-	unsigned long flags;
 	struct test_stat stat;
 
 	memset(&stat, 0, sizeof(stat));
 	stat.max = INT_MAX;
 	stat.name = stub_name;
-	local_irq_save(flags);
-	t0 = sched_clock();
+	t0 = ktime_get_ns();
 	kallsyms_on_each_match_symbol(match_symbol, stat.name, &stat);
-	t1 = sched_clock();
-	local_irq_restore(flags);
+	t1 = ktime_get_ns();
 	pr_info("kallsyms_on_each_match_symbol() traverse all: %lld ns\n", t1 - t0);
 }
 
Signature · 1 line
2.37.2
#1 lore raw
Quoted 4 lines · On Thu, 12 Jan 2023 20:54:26 +1000 Nicholas Piggin <[email protected]> wrote:

kallsyms_on_each* may schedule so must not be called with interrupts disabled. The iteration function could disable interrupts, but this also changes lookup_symbol() to match the change to the other timing code.

Just tested your patch. It fixes the problem on the G5. Thanks!

Regards,
Erhard

#2 lore raw
In reply to Nicholas Piggin
Quoted 5 lines · Nicholas Piggin

kallsyms_on_each* may schedule so must not be called with interrupts disabled. The iteration function could disable interrupts, but this also changes lookup_symbol() to match the change to the other timing code.

Quoted 7 lines

---

Thanks Nicholas!

Petr had just suggested removing this aspect of the selftests, the performance test as its specific to the config, it doesn't run many times to get an average and odd things on a system can create different metrics. Zhen Lei had given up on fixing it and has a patch to instead remove this part of the selftest.

I still find value in keeping it, but Petr, would like your opinion on this fix, if we were to keep it.

Luis

Quoted 68 lines
 kernel/kallsyms_selftest.c | 21 ++++++---------------
 1 file changed, 6 insertions(+), 15 deletions(-)
kernel/kallsyms_selftest.c +6 -15
index f35d9cc1aab1..bfbc12da3326 100644
@@ -157,14 +157,11 @@ static void test_kallsyms_compression_ratio(void)
 static int lookup_name(void *data, const char *name, struct module *mod, unsigned long addr)
 {
 	u64 t0, t1, t;
-	unsigned long flags;
 	struct test_stat *stat = (struct test_stat *)data;
 
-	local_irq_save(flags);
-	t0 = sched_clock();
+	t0 = ktime_get_ns();
 	(void)kallsyms_lookup_name(name);
-	t1 = sched_clock();
-	local_irq_restore(flags);
+	t1 = ktime_get_ns();
 
 	t = t1 - t0;
 	if (t < stat->min)
@@ -234,18 +231,15 @@ static int find_symbol(void *data, const char *name, struct module *mod, unsigne
 static void test_perf_kallsyms_on_each_symbol(void)
 {
 	u64 t0, t1;
-	unsigned long flags;
 	struct test_stat stat;
 
 	memset(&stat, 0, sizeof(stat));
 	stat.max = INT_MAX;
 	stat.name = stub_name;
 	stat.perf = 1;
-	local_irq_save(flags);
-	t0 = sched_clock();
+	t0 = ktime_get_ns();
 	kallsyms_on_each_symbol(find_symbol, &stat);
-	t1 = sched_clock();
-	local_irq_restore(flags);
+	t1 = ktime_get_ns();
 	pr_info("kallsyms_on_each_symbol() traverse all: %lld ns\n", t1 - t0);
 }
 
@@ -270,17 +264,14 @@ static int match_symbol(void *data, unsigned long addr)
 static void test_perf_kallsyms_on_each_match_symbol(void)
 {
 	u64 t0, t1;
-	unsigned long flags;
 	struct test_stat stat;
 
 	memset(&stat, 0, sizeof(stat));
 	stat.max = INT_MAX;
 	stat.name = stub_name;
-	local_irq_save(flags);
-	t0 = sched_clock();
+	t0 = ktime_get_ns();
 	kallsyms_on_each_match_symbol(match_symbol, stat.name, &stat);
-	t1 = sched_clock();
-	local_irq_restore(flags);
+	t1 = ktime_get_ns();
 	pr_info("kallsyms_on_each_match_symbol() traverse all: %lld ns\n", t1 - t0);
 }
 
Signature · 1 line
2.37.2
#3 lore raw
Quoted 16 lines · Luis Chamberlain
Quoted 5 lines · Nicholas Piggin

kallsyms_on_each* may schedule so must not be called with interrupts disabled. The iteration function could disable interrupts, but this also changes lookup_symbol() to match the change to the other timing code.

Quoted 7 lines

---

Thanks Nicholas!

Quoted 8 lines

Petr had just suggested removing this aspect of the selftests, the performance test as its specific to the config, it doesn't run many times to get an average and odd things on a system can create different metrics. Zhen Lei had given up on fixing it and has a patch to instead remove this part of the selftest.

I still find value in keeping it, but Petr, would like your opinion on this fix, if we were to keep it.

I am fine with this fix.

It increases a risk of possible inaccuracy of the measured time. It would count also time spent on unrelated interrupts and eventual rescheduling.

Anyway, it is safe at least. I was against the previous attempts to fix this problem because they might have caused problems for the rest of the system.

Best Regards,
Petr

#4 lore raw
Quoted 23 lines · Petr Mladek
Quoted 22 lines · Luis Chamberlain
Quoted 5 lines · Nicholas Piggin

kallsyms_on_each* may schedule so must not be called with interrupts disabled. The iteration function could disable interrupts, but this also changes lookup_symbol() to match the change to the other timing code.

Quoted 7 lines

---

Thanks Nicholas!

Petr had just suggested removing this aspect of the selftests, the performance test as its specific to the config, it doesn't run many times to get an average and odd things on a system can create different metrics. Zhen Lei had given up on fixing it and has a patch to instead remove this part of the selftest.

Quoted 4 lines
Quoted 2 lines

I still find value in keeping it, but Petr, would like your opinion on this fix, if we were to keep it.

I am fine with this fix.

Merged the fix. I'll push to Linus for 6.2-rc4

Luis