Skip to contentSkip to conversation

sched/rt: fix the case where sched_rt_period_us is negative

PATCH Superseded Yajun Deng 2 participants 4 messages May 12, 2022 – May 17, 2022 Scheduler

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Yajun Deng

author
#0 lore raw

PATCH v2 1/1 sched/rt: fix the case where sched_rt_period_us is negative

The proc_dointvec() is for integer, but sysctl_sched_rt_period is a unsigned integer, proc_dointvec() would convert negative number into positive number. So both proc_dointvec() and sched_rt_global_validate() aren't return error even if we set a negative number.

Use proc_dointvec_minmax() instead of proc_dointvec() and use extra1 limit the minimum value for sched_rt_period_us/sched_rt_runtime_us.

Fixes 391e43da797a ("sched: Move all scheduler bits into kernel/sched/")
---
 kernel/sched/rt.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
kernel/sched/rt.c +4 -4
index b491a0f8c25d..3add32679885 100644
@@ -37,6 +37,7 @@ static struct ctl_table sched_rt_sysctls[] = {
 		.maxlen         = sizeof(unsigned int),
 		.mode           = 0644,
 		.proc_handler   = sched_rt_handler,
+		.extra1		= SYSCTL_ONE,
 	},
 	{
 		.procname       = "sched_rt_runtime_us",
@@ -44,6 +45,8 @@ static struct ctl_table sched_rt_sysctls[] = {
 		.maxlen         = sizeof(int),
 		.mode           = 0644,
 		.proc_handler   = sched_rt_handler,
+		.extra1		= SYSCTL_NEG_ONE,
+		.extra2		= (void *)&sysctl_sched_rt_period,
 	},
 	{
 		.procname       = "sched_rr_timeslice_ms",
@@ -2959,9 +2962,6 @@ static int sched_rt_global_constraints(void)
 #ifdef CONFIG_SYSCTL
 static int sched_rt_global_validate(void)
 {
-	if (sysctl_sched_rt_period <= 0)
-		return -EINVAL;
-
 	if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
 		((sysctl_sched_rt_runtime > sysctl_sched_rt_period) ||
 		 ((u64)sysctl_sched_rt_runtime *
@@ -2992,7 +2992,7 @@ static int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
 	old_period = sysctl_sched_rt_period;
 	old_runtime = sysctl_sched_rt_runtime;
 
-	ret = proc_dointvec(table, write, buffer, lenp, ppos);
+	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 
 	if (!ret && write) {
 		ret = sched_rt_global_validate();
Signature · 1 line
2.25.1
#1 lore raw
Quoted 9 lines · Yajun Deng

The proc_dointvec() is for integer, but sysctl_sched_rt_period is a unsigned integer, proc_dointvec() would convert negative number into positive number. So both proc_dointvec() and sched_rt_global_validate() aren't return error even if we set a negative number.

Use proc_dointvec_minmax() instead of proc_dointvec() and use extra1 limit the minimum value for sched_rt_period_us/sched_rt_runtime_us.

Fixes 391e43da797a ("sched: Move all scheduler bits into kernel/sched/")

That's just the last apparent change of the incriminated variable. AFAICT the issue stems from:

- sysctl_sched_rt_period being unsigned int
- the ctl entry using proc_dointvec()
- the bounds check on sysctl_sched_rt_period being just <= 0 which doesn't actually respect the [1, INT_MAX] stated in Documentation/scheduler/sched-rt-group.rst

The <= thing was added by:

ec5d498991e8 ("sched: fix deadlock in setting scheduler parameter to zero")

but AFAICT the unsigned int vs proc_dointvec() thing dates back to:

d0b27fa77854 ("sched: rt-group: synchonised bandwidth period")

Quoted 2 lines

---

In the absence of a cover letter (e.g. single-patch submission), this is where you should write patch version changelogs (see Documentation/process).

kernel/sched/rt.c static struct ctl_table sched_rt_sysctls[] = { · line 40

Quoted 12 lines
 kernel/sched/rt.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
kernel/sched/rt.c +1 -0
index b491a0f8c25d..3add32679885 100644
@@ -37,6 +37,7 @@ static struct ctl_table sched_rt_sysctls[] = {
              .maxlen         = sizeof(unsigned int),
              .mode           = 0644,
              .proc_handler   = sched_rt_handler,
+		.extra1		= SYSCTL_ONE,

Per earlier comment, the Documentation says that this needs to be within [1, INT_MAX], which you do get by excluding negative values when casting to int...

How about we make sysctl_sched_rt_period int on top of this, then all variables modified by the sched_rt_handler() proc_dointvec() are *actually* int, and the upper bound requires less mental gymnastics to be figured out?

kernel/sched/rt.c static struct ctl_table sched_rt_sysctls[] = { · line 49

Quoted 9 lines
kernel/sched/rt.c +2 -0
      },
      {
              .procname       = "sched_rt_runtime_us",
@@ -44,6 +45,8 @@ static struct ctl_table sched_rt_sysctls[] = {
              .maxlen         = sizeof(int),
              .mode           = 0644,
              .proc_handler   = sched_rt_handler,
+		.extra1		= SYSCTL_NEG_ONE,
+		.extra2		= (void *)&sysctl_sched_rt_period,

Per this, you could also remove the

                ((sysctl_sched_rt_runtime > sysctl_sched_rt_period) ||

from sched_rt_global_validate(), no?

Quoted 24 lines
kernel/sched/rt.c +1 -5
      },
      {
              .procname       = "sched_rr_timeslice_ms",
@@ -2959,9 +2962,6 @@ static int sched_rt_global_constraints(void)
 #ifdef CONFIG_SYSCTL
 static int sched_rt_global_validate(void)
 {
-	if (sysctl_sched_rt_period <= 0)
-		return -EINVAL;
-
      if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
              ((sysctl_sched_rt_runtime > sysctl_sched_rt_period) ||
               ((u64)sysctl_sched_rt_runtime *
@@ -2992,7 +2992,7 @@ static int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
      old_period = sysctl_sched_rt_period;
      old_runtime = sysctl_sched_rt_runtime;
 
-	ret = proc_dointvec(table, write, buffer, lenp, ppos);
+	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 
      if (!ret && write) {
              ret = sched_rt_global_validate();
--

2.25.1

author
#2 lore raw
Quoted 21 lines · Valentin Schneider
Quoted 9 lines · Yajun Deng

The proc_dointvec() is for integer, but sysctl_sched_rt_period is a unsigned integer, proc_dointvec() would convert negative number into positive number. So both proc_dointvec() and sched_rt_global_validate() aren't return error even if we set a negative number.

Use proc_dointvec_minmax() instead of proc_dointvec() and use extra1 limit the minimum value for sched_rt_period_us/sched_rt_runtime_us.

Fixes 391e43da797a ("sched: Move all scheduler bits into kernel/sched/")

That's just the last apparent change of the incriminated variable. AFAICT the issue stems from:

- sysctl_sched_rt_period being unsigned int
- the ctl entry using proc_dointvec()
- the bounds check on sysctl_sched_rt_period being just <= 0 which doesn't actually respect the [1, INT_MAX] stated in Documentation/scheduler/sched-rt-group.rst

Quoted 8 lines

The <= thing was added by:

ec5d498991e8 ("sched: fix deadlock in setting scheduler parameter to zero")

but AFAICT the unsigned int vs proc_dointvec() thing dates back to:

d0b27fa77854 ("sched: rt-group: synchonised bandwidth period")

I know that, but I didn't find out the source. Thank you for helping me find out it.

Quoted 7 lines
Quoted 2 lines

---

In the absence of a cover letter (e.g. single-patch submission), this is where you should write patch version changelogs (see Documentation/process).

Got it, I will add it to the next version.

Quoted 17 lines

kernel/sched/rt.c static struct ctl_table sched_rt_sysctls[] = { · line 40

Quoted 12 lines

kernel/sched/rt.c | 8 ++++----

1 file changed, 4 insertions(+), 4 deletions(-)
kernel/sched/rt.c +1 -0
index b491a0f8c25d..3add32679885 100644
@@ -37,6 +37,7 @@ static struct ctl_table sched_rt_sysctls[] = {
 .maxlen = sizeof(unsigned int),
 .mode = 0644,
 .proc_handler = sched_rt_handler,
+ .extra1 = SYSCTL_ONE,

Per earlier comment, the Documentation says that this needs to be within [1, INT_MAX], which you do get by excluding negative values when casting to int...

Quoted 4 lines

How about we make sysctl_sched_rt_period int on top of this, then all variables modified by the sched_rt_handler() proc_dointvec() are *actually* int, and the upper bound requires less mental gymnastics to be figured out?

Yes, we can make sysctl_sched_rt_period int.

Quoted 12 lines

kernel/sched/rt.c static struct ctl_table sched_rt_sysctls[] = { · line 49

Quoted 9 lines
kernel/sched/rt.c +2 -0
 },
 {
 .procname = "sched_rt_runtime_us",
@@ -44,6 +45,8 @@ static struct ctl_table sched_rt_sysctls[] = {
 .maxlen = sizeof(int),
 .mode = 0644,
 .proc_handler = sched_rt_handler,
+ .extra1 = SYSCTL_NEG_ONE,
+ .extra2 = (void *)&sysctl_sched_rt_period,

Per this, you could also remove the

Quoted 4 lines

((sysctl_sched_rt_runtime > sysctl_sched_rt_period) ||

from sched_rt_global_validate(), no?

No, the extra2 just limit the maximum value of sysctl_sched_rt_runtime is sysctl_sched_rt_period, but not limit the minimum value of sysctl_sched_rt_period is sysctl_sched_rt_runtime. (sysctl_sched_rt_runtime > sysctl_sched_rt_period) can do both. Its purpose is to return error earlier. Perhaps I should remove extra2 to avoid ambiguity.

Quoted 24 lines
Quoted 24 lines
kernel/sched/rt.c +0 -3
 }, 
 {
 .procname = "sched_rr_timeslice_ms",
@@ -2959,9 +2962,6 @@ static int sched_rt_global_constraints(void)
 #ifdef CONFIG_SYSCTL
 static int sched_rt_global_validate(void)
 {
- if (sysctl_sched_rt_period <= 0)
- return -EINVAL;
-

if ((sysctl_sched_rt_runtime != RUNTIME_INF) && ((sysctl_sched_rt_runtime > sysctl_sched_rt_period) || ((u64)sysctl_sched_rt_runtime *

kernel/sched/rt.c +1 -2
@@ -2992,7 +2992,7 @@ static int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
 old_period = sysctl_sched_rt_period;
 old_runtime = sysctl_sched_rt_runtime;
 
- ret = proc_dointvec(table, write, buffer, lenp, ppos);
+ ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 
 if (!ret && write) {
 ret = sched_rt_global_validate();
--

2.25.1

#3 lore raw
Quoted 13 lines · Yajun Deng
Quoted 12 lines · Valentin Schneider

kernel/sched/rt.c static struct ctl_table sched_rt_sysctls[] = { · line 49

Quoted 9 lines

},
{
.procname = "sched_rt_runtime_us",

kernel/sched/rt.c +2 -0
@@ -44,6 +45,8 @@ static struct ctl_table sched_rt_sysctls[] = {
 .maxlen = sizeof(int),
 .mode = 0644,
 .proc_handler = sched_rt_handler,
+ .extra1 = SYSCTL_NEG_ONE,
+ .extra2 = (void *)&sysctl_sched_rt_period,

Per this, you could also remove the

Quoted 6 lines
Quoted 4 lines

((sysctl_sched_rt_runtime > sysctl_sched_rt_period) ||

from sched_rt_global_validate(), no?

No, the extra2 just limit the maximum value of sysctl_sched_rt_runtime is sysctl_sched_rt_period, but not limit the minimum value of sysctl_sched_rt_period is sysctl_sched_rt_runtime. (sysctl_sched_rt_runtime > sysctl_sched_rt_period) can do both.

Gotcha.

Quoted 2 lines

Its purpose is to return error earlier. Perhaps I should remove extra2 to avoid ambiguity.

It's probably better to only have the "pure" bounds in there yes.