base/drivers/arch_topology: Replace mutex with READ_ONCE / WRITE_ONCE

In the case of asymmetric SoC with the same micro-architecture, we have a group of CPUs with smaller OPPs than the other group. One example is the 96boards dragonboard 820c. There is no dmips/MHz difference between both groups, so no need to specify the values in the DT. Unfortunately, without these defined, there is no scaling capacity computation triggered, so we need to write 'capacity-dmips-mhz' for each CPU with the same value in order to force the scaled capacity computation.

In order to fix this situation, allocate 'raw_capacity' so the pointer is set and the init_cpu_capacity_callback() function can be called.

This was tested on db820c:
- specified values in the DT (correct results)
- partial values defined in the DT (error + fallback to defaults)
- no specified values in the DT (correct results)

correct results are:

   cat /sys/devices/system/cpu/cpu*/cpu_capacity
    758
    758
   1024
   1024

   ... respectively for CPU0, CPU1, CPU2 and CPU3.

That reflects the capacity for the max frequencies 1593600 and 2150400.

Hi Daniel,

+cc: Russell King <[email protected]>

Quoted 24 lines · Daniel Lezcano

In the case of asymmetric SoC with the same micro-architecture, we have a group of CPUs with smaller OPPs than the other group. One example is the 96boards dragonboard 820c. There is no dmips/MHz difference between both groups, so no need to specify the values in the DT. Unfortunately, without these defined, there is no scaling capacity computation triggered, so we need to write 'capacity-dmips-mhz' for each CPU with the same value in order to force the scaled capacity computation.

In order to fix this situation, allocate 'raw_capacity' so the pointer is set and the init_cpu_capacity_callback() function can be called.

This was tested on db820c:
  - specified values in the DT (correct results)
  - partial values defined in the DT (error + fallback to defaults)
  - no specified values in the DT (correct results)

correct results are:
   cat /sys/devices/system/cpu/cpu*/cpu_capacity
    758
    758
   1024
   1024

Quoted 3 lines

   ... respectively for CPU0, CPU1, CPU2 and CPU3.

That reflects the capacity for the max frequencies 1593600 and 2150400.

[...]

I'm afraid that this change is incompatible with the still existing cpu_efficiency interface we have in Arm32 for A15/A7 systems like Arm TC2:

In case you specify clock-frequency dt properties per cpu for such a system, the cpu_capacity values are determined via the cpu_efficiency code in arch/arm/kernel/topology.c.

So on Arm TC2 with clock-frequency = <1000000000> [A15] and <800000000> [A7] you get:

root@linaro-nano:~# cat /sys/devices/system/cpu/cpu*/cpu_capacity
606
1441
1441
606
606

With your patches on top (cpu_capacity functionality in drivers/base/arch_topology.c does not have to be switched on by specifying capacity-dmips-mhz dt properties anymore) we end up scaling by max frequency again:

root@linaro-nano:~# cat /sys/devices/system/cpu/cpu*/cpu_capacity
358
1024
1024
358
358

I tried to remove the cpu_efficiency based API a year ago but Russell pointed out that the compatibility has to be maintained for longer:

https://lore.kernel.org/lkml/[email protected]/

I assume that the capacity-dmips-mhz dt property is like a switch to turn this functionality on for big.Little and so called gold/silver platforms, which have cores with the same uArch but in frequency domains with different max frequency values.

So what's wrong with specifying capacity-dmips-mhz = <1024> for all cores for those gold/silver platforms?

I don't expect that there will be
so many of them. And normal SMP platforms (w/o frequency domains w/o different max frequency values) don't have to execute this code.

The mutex protects a per_cpu variable access. The potential race can happen only when the cpufreq governor module is loaded and at the same time the cpu capacity is changed in the sysfs.

There is no real interest of using a mutex to protect a variable assignation when there is no situation where a task can take the lock and block.

Replace the mutex by READ_ONCE / WRITE_ONCE.

Signed-off-by Daniel Lezcano <[email protected]>

Cc Sudeep Holla <[email protected]>

Reviewed-by Viresh Kumar <[email protected]>

---
 drivers/base/arch_topology.c  | 7 +------
 include/linux/arch_topology.h | 2 +-
 2 files changed, 2 insertions(+), 7 deletions(-)

drivers/base/arch_topology.c +1 -2

index edfcf8d..fd5325b 100644
@@ -31,12 +31,11 @@ void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq,
 		per_cpu(freq_scale, i) = scale;
 }
 
-static DEFINE_MUTEX(cpu_scale_mutex);
 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
 
 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
 {
-	per_cpu(cpu_scale, cpu) = capacity;
+	WRITE_ONCE(per_cpu(cpu_scale, cpu), capacity);
 }
 
 static ssize_t cpu_capacity_show(struct device *dev,

Hi Daniel,

Quoted 40 lines · Daniel Lezcano

The mutex protects a per_cpu variable access. The potential race can happen only when the cpufreq governor module is loaded and at the same time the cpu capacity is changed in the sysfs.

There is no real interest of using a mutex to protect a variable assignation when there is no situation where a task can take the lock and block.

Replace the mutex by READ_ONCE / WRITE_ONCE.

Signed-off-by Daniel Lezcano <[email protected]>

Cc Sudeep Holla <[email protected]>

Reviewed-by Viresh Kumar <[email protected]>

---
 drivers/base/arch_topology.c  | 7 +------
 include/linux/arch_topology.h | 2 +-
 2 files changed, 2 insertions(+), 7 deletions(-)

drivers/base/arch_topology.c +1 -2

index edfcf8d..fd5325b 100644
@@ -31,12 +31,11 @@ void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq,
 		per_cpu(freq_scale, i) = scale;
 }
 
-static DEFINE_MUTEX(cpu_scale_mutex);
 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
 
 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
 {
-	per_cpu(cpu_scale, cpu) = capacity;
+	WRITE_ONCE(per_cpu(cpu_scale, cpu), capacity);
 }
 
 static ssize_t cpu_capacity_show(struct device *dev,

@@ -71,10 +70,8 @@ static ssize_t cpu_capacity_store(struct device *dev,
 	if (new_capacity > SCHED_CAPACITY_SCALE)
 		return -EINVAL;
 

drivers/base/arch_topology.c static ssize_t cpu_capacity_store(struct device *dev, · line 77

Quoted 4 lines

drivers/base/arch_topology.c +0 -2

-	mutex_lock(&cpu_scale_mutex);
 	for_each_cpu(i, &cpu_topology[this_cpu].core_sibling)
 		topology_set_cpu_scale(i, new_capacity);
-	mutex_unlock(&cpu_scale_mutex);

IIRC this was meant to ensure atomic updates of all siblings with the new capacity value. I actually now wonder if readers should not grab the mutex as well (cpu_capacity_show()). Can't we get into a situation where a reader might see siblings with intermediate values (while the loop above is performing an update)?

BTW, please update my email address. :-)

Quoted 43 lines · Juri Lelli

Hi Daniel,

Quoted 40 lines · Daniel Lezcano

The mutex protects a per_cpu variable access. The potential race can happen only when the cpufreq governor module is loaded and at the same time the cpu capacity is changed in the sysfs.

There is no real interest of using a mutex to protect a variable assignation when there is no situation where a task can take the lock and block.

Replace the mutex by READ_ONCE / WRITE_ONCE.

Signed-off-by Daniel Lezcano <[email protected]>

Cc Sudeep Holla <[email protected]>

Reviewed-by Viresh Kumar <[email protected]>

---
 drivers/base/arch_topology.c  | 7 +------
 include/linux/arch_topology.h | 2 +-
 2 files changed, 2 insertions(+), 7 deletions(-)

drivers/base/arch_topology.c +1 -2

index edfcf8d..fd5325b 100644
@@ -31,12 +31,11 @@ void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq,
 		per_cpu(freq_scale, i) = scale;
 }
 
-static DEFINE_MUTEX(cpu_scale_mutex);
 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
 
 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
 {
-	per_cpu(cpu_scale, cpu) = capacity;
+	WRITE_ONCE(per_cpu(cpu_scale, cpu), capacity);
 }
 
 static ssize_t cpu_capacity_show(struct device *dev,

@@ -71,10 +70,8 @@ static ssize_t cpu_capacity_store(struct device *dev,
 	if (new_capacity > SCHED_CAPACITY_SCALE)
 		return -EINVAL;
 

Quoted 10 lines

drivers/base/arch_topology.c static ssize_t cpu_capacity_store(struct device *dev, · line 77

Quoted 4 lines

drivers/base/arch_topology.c +0 -2

-	mutex_lock(&cpu_scale_mutex);
 	for_each_cpu(i, &cpu_topology[this_cpu].core_sibling)
 		topology_set_cpu_scale(i, new_capacity);
-	mutex_unlock(&cpu_scale_mutex);

IIRC this was meant to ensure atomic updates of all siblings with the new capacity value. I actually now wonder if readers should not grab the mutex as well (cpu_capacity_show()). Can't we get into a situation where a reader might see siblings with intermediate values (while the loop above is performing an update)?

With or without this patch, it is the case:

task1 task2
| |
read("/sys/.../cpu1/cpu_capacity) |

                  |                  write("/sys/.../cpu1/cpu_capacity")

read("/sys/.../cpu2/cpu_capacity) |

There is no guarantee userspace can have a consistent view of the capacity. As soon as it reads a capacity, it can be changed in its back.

Quoted 9 lines

With or without this patch, it is the case:

task1 task2
| |
read("/sys/.../cpu1/cpu_capacity) |

                  |                  write("/sys/.../cpu1/cpu_capacity")

read("/sys/.../cpu2/cpu_capacity) |

Quoted 2 lines

There is no guarantee userspace can have a consistent view of the capacity. As soon as it reads a capacity, it can be changed in its back.

True, but w/o the mutex task1 could read different cpu_capacity values for a cluster (it actually can also with current implementation, we should grab the mutex in the read path as well if we want to avoid this).

On 29/11/2018 08:04, Juri Lelli wrote:

[ ... ]

Quoted 9 lines

Quoted 9 lines

With or without this patch, it is the case:

task1 task2
| |
read("/sys/.../cpu1/cpu_capacity) |

                  |                  write("/sys/.../cpu1/cpu_capacity")

read("/sys/.../cpu2/cpu_capacity) |

Quoted 7 lines

Quoted 2 lines

There is no guarantee userspace can have a consistent view of the capacity. As soon as it reads a capacity, it can be changed in its back.

True, but w/o the mutex task1 could read different cpu_capacity values for a cluster (it actually can also with current implementation, we should grab the mutex in the read path as well if we want to avoid this).

Even if the mutex is on the read path, the userspace can see different capacities because it will read the cpu_capacity per cpu directory.

The mutex will be take when reading cpu0/cpu_capacity, not for cpu[0-9]/cpu_capacity. Between two reads, a write can happen because the lock is released in between.

Do you agree with the patch ? Or do you want me to drop it ?

Quoted 28 lines · Daniel Lezcano

On 29/11/2018 08:04, Juri Lelli wrote:

[ ... ]

Quoted 9 lines

Quoted 9 lines

With or without this patch, it is the case:

task1 task2
| |
read("/sys/.../cpu1/cpu_capacity) |

                  |                  write("/sys/.../cpu1/cpu_capacity")

read("/sys/.../cpu2/cpu_capacity) |

Quoted 7 lines

Quoted 2 lines

There is no guarantee userspace can have a consistent view of the capacity. As soon as it reads a capacity, it can be changed in its back.

True, but w/o the mutex task1 could read different cpu_capacity values for a cluster (it actually can also with current implementation, we should grab the mutex in the read path as well if we want to avoid this).

Even if the mutex is on the read path, the userspace can see different capacities because it will read the cpu_capacity per cpu directory.

The mutex will be take when reading cpu0/cpu_capacity, not for cpu[0-9]/cpu_capacity. Between two reads, a write can happen because the lock is released in between.

Quoted 1 line

Do you agree with the patch ? Or do you want me to drop it ?

I don't actually have cases at hand that are showing regression with it, I was just trying to understand if we might potentially hit problems in the future. So, I'm not against this patch. :-)

Quoted 29 lines · Juri Lelli

Quoted 28 lines · Daniel Lezcano

On 29/11/2018 08:04, Juri Lelli wrote:

[ ... ]

Quoted 9 lines

Quoted 9 lines

With or without this patch, it is the case:

task1 task2
| |
read("/sys/.../cpu1/cpu_capacity) |

                  |                  write("/sys/.../cpu1/cpu_capacity")

read("/sys/.../cpu2/cpu_capacity) |

Quoted 7 lines

Quoted 2 lines

There is no guarantee userspace can have a consistent view of the capacity. As soon as it reads a capacity, it can be changed in its back.

True, but w/o the mutex task1 could read different cpu_capacity values for a cluster (it actually can also with current implementation, we should grab the mutex in the read path as well if we want to avoid this).

Even if the mutex is on the read path, the userspace can see different capacities because it will read the cpu_capacity per cpu directory.

The mutex will be take when reading cpu0/cpu_capacity, not for cpu[0-9]/cpu_capacity. Between two reads, a write can happen because the lock is released in between.

Quoted 5 lines

Quoted 1 line

Do you agree with the patch ? Or do you want me to drop it ?

I don't actually have cases at hand that are showing regression with it, I was just trying to understand if we might potentially hit problems in the future. So, I'm not against this patch. :-)

not-not-acked-by ? :)