linux_dsm_epyc7002/drivers/cpufreq/cpufreq_userspace.c

159 lines
4.0 KiB
C
Raw Normal View History

/*
* linux/drivers/cpufreq/cpufreq_userspace.c
*
* Copyright (C) 2001 Russell King
* (C) 2002 - 2004 Dominik Brodowski <linux@brodo.de>
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 as
* published by the Free Software Foundation.
*
*/
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/cpufreq.h>
#include <linux/init.h>
#include <linux/module.h>
#include <linux/mutex.h>
#include <linux/slab.h>
static DEFINE_PER_CPU(unsigned int, cpu_is_managed);
static DEFINE_MUTEX(userspace_mutex);
/**
* cpufreq_set - set the CPU frequency
[CPUFREQ] Eliminate cpufreq_userspace scaling_setspeed deadlock Eliminate cpufreq_userspace scaling_setspeed deadlock. Luming Yu recently uncovered yet another cpufreq related deadlock. One thread that continuously switches the governors and the other thread that repeatedly cats the contents of cpufreq directory causes both these threads to go into a deadlock. Detailed examination of the deadlock showed the exact flow before the deadlock as: Thread 1 Thread 2 ________ ________ cats files under /sys/devices/.../cpufreq/ Set governor to userspace Adds a new sysfs entry for scaling_setspeed cats files under /sys/devices/.../cpufreq/ Set governor to performance Holds cpufreq_rw_sem in write mode Sends a STOP notify to userspace governor cat /sys/devices/.../cpufreq/scaling_setspeed Gets a handle on the above sysfs entry with sysfs_get_active Blocks while trying to get cpufreq_rw_sem in read mode Remove a sysfs entry for scaling_setspeed Blocks on sysfs_deactivate while waiting for earlier get_active (on other thread) to drain At this point both threads go into deadlock and any other thread that tries to do anything with sysfs cpufreq will also block. There seems to be no easy way to avoid this deadlock as long as cpufreq_userspace adds/removes the sysfs entry under same kobject as cpufreq. Below patch moves scaling_setspeed to cpufreq.c, keeping it always and calling back the governor on read/write. This is the cleanest fix I could think of, even though adding two callbacks in governor structure just for this seems unnecessary. Note that the change makes scaling_setspeed under /sys/.../cpufreq permanent and returns <unsupported> when governor is not userspace. Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com> Signed-off-by: Dave Jones <davej@redhat.com>
2007-10-27 00:18:21 +07:00
* @policy: pointer to policy struct where freq is being set
* @freq: target frequency in kHz
*
* Sets the CPU frequency to freq.
*/
[CPUFREQ] Eliminate cpufreq_userspace scaling_setspeed deadlock Eliminate cpufreq_userspace scaling_setspeed deadlock. Luming Yu recently uncovered yet another cpufreq related deadlock. One thread that continuously switches the governors and the other thread that repeatedly cats the contents of cpufreq directory causes both these threads to go into a deadlock. Detailed examination of the deadlock showed the exact flow before the deadlock as: Thread 1 Thread 2 ________ ________ cats files under /sys/devices/.../cpufreq/ Set governor to userspace Adds a new sysfs entry for scaling_setspeed cats files under /sys/devices/.../cpufreq/ Set governor to performance Holds cpufreq_rw_sem in write mode Sends a STOP notify to userspace governor cat /sys/devices/.../cpufreq/scaling_setspeed Gets a handle on the above sysfs entry with sysfs_get_active Blocks while trying to get cpufreq_rw_sem in read mode Remove a sysfs entry for scaling_setspeed Blocks on sysfs_deactivate while waiting for earlier get_active (on other thread) to drain At this point both threads go into deadlock and any other thread that tries to do anything with sysfs cpufreq will also block. There seems to be no easy way to avoid this deadlock as long as cpufreq_userspace adds/removes the sysfs entry under same kobject as cpufreq. Below patch moves scaling_setspeed to cpufreq.c, keeping it always and calling back the governor on read/write. This is the cleanest fix I could think of, even though adding two callbacks in governor structure just for this seems unnecessary. Note that the change makes scaling_setspeed under /sys/.../cpufreq permanent and returns <unsupported> when governor is not userspace. Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com> Signed-off-by: Dave Jones <davej@redhat.com>
2007-10-27 00:18:21 +07:00
static int cpufreq_set(struct cpufreq_policy *policy, unsigned int freq)
{
int ret = -EINVAL;
unsigned int *setspeed = policy->governor_data;
pr_debug("cpufreq_set for cpu %u, freq %u kHz\n", policy->cpu, freq);
mutex_lock(&userspace_mutex);
if (!per_cpu(cpu_is_managed, policy->cpu))
goto err;
*setspeed = freq;
ret = __cpufreq_driver_target(policy, freq, CPUFREQ_RELATION_L);
err:
mutex_unlock(&userspace_mutex);
return ret;
}
[CPUFREQ] Eliminate cpufreq_userspace scaling_setspeed deadlock Eliminate cpufreq_userspace scaling_setspeed deadlock. Luming Yu recently uncovered yet another cpufreq related deadlock. One thread that continuously switches the governors and the other thread that repeatedly cats the contents of cpufreq directory causes both these threads to go into a deadlock. Detailed examination of the deadlock showed the exact flow before the deadlock as: Thread 1 Thread 2 ________ ________ cats files under /sys/devices/.../cpufreq/ Set governor to userspace Adds a new sysfs entry for scaling_setspeed cats files under /sys/devices/.../cpufreq/ Set governor to performance Holds cpufreq_rw_sem in write mode Sends a STOP notify to userspace governor cat /sys/devices/.../cpufreq/scaling_setspeed Gets a handle on the above sysfs entry with sysfs_get_active Blocks while trying to get cpufreq_rw_sem in read mode Remove a sysfs entry for scaling_setspeed Blocks on sysfs_deactivate while waiting for earlier get_active (on other thread) to drain At this point both threads go into deadlock and any other thread that tries to do anything with sysfs cpufreq will also block. There seems to be no easy way to avoid this deadlock as long as cpufreq_userspace adds/removes the sysfs entry under same kobject as cpufreq. Below patch moves scaling_setspeed to cpufreq.c, keeping it always and calling back the governor on read/write. This is the cleanest fix I could think of, even though adding two callbacks in governor structure just for this seems unnecessary. Note that the change makes scaling_setspeed under /sys/.../cpufreq permanent and returns <unsupported> when governor is not userspace. Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com> Signed-off-by: Dave Jones <davej@redhat.com>
2007-10-27 00:18:21 +07:00
static ssize_t show_speed(struct cpufreq_policy *policy, char *buf)
{
return sprintf(buf, "%u\n", policy->cur);
}
static int cpufreq_userspace_policy_init(struct cpufreq_policy *policy)
{
unsigned int *setspeed;
setspeed = kzalloc(sizeof(*setspeed), GFP_KERNEL);
if (!setspeed)
return -ENOMEM;
policy->governor_data = setspeed;
return 0;
}
static void cpufreq_userspace_policy_exit(struct cpufreq_policy *policy)
{
mutex_lock(&userspace_mutex);
kfree(policy->governor_data);
policy->governor_data = NULL;
mutex_unlock(&userspace_mutex);
}
static int cpufreq_userspace_policy_start(struct cpufreq_policy *policy)
{
unsigned int *setspeed = policy->governor_data;
BUG_ON(!policy->cur);
pr_debug("started managing cpu %u\n", policy->cpu);
mutex_lock(&userspace_mutex);
per_cpu(cpu_is_managed, policy->cpu) = 1;
*setspeed = policy->cur;
mutex_unlock(&userspace_mutex);
return 0;
}
static void cpufreq_userspace_policy_stop(struct cpufreq_policy *policy)
{
unsigned int *setspeed = policy->governor_data;
pr_debug("managing cpu %u stopped\n", policy->cpu);
mutex_lock(&userspace_mutex);
per_cpu(cpu_is_managed, policy->cpu) = 0;
*setspeed = 0;
mutex_unlock(&userspace_mutex);
}
static void cpufreq_userspace_policy_limits(struct cpufreq_policy *policy)
{
unsigned int *setspeed = policy->governor_data;
mutex_lock(&userspace_mutex);
pr_debug("limit event for cpu %u: %u - %u kHz, currently %u kHz, last set to %u kHz\n",
policy->cpu, policy->min, policy->max, policy->cur, *setspeed);
if (policy->max < *setspeed)
__cpufreq_driver_target(policy, policy->max, CPUFREQ_RELATION_H);
else if (policy->min > *setspeed)
__cpufreq_driver_target(policy, policy->min, CPUFREQ_RELATION_L);
else
__cpufreq_driver_target(policy, *setspeed, CPUFREQ_RELATION_L);
mutex_unlock(&userspace_mutex);
}
static struct cpufreq_governor cpufreq_gov_userspace = {
.name = "userspace",
.init = cpufreq_userspace_policy_init,
.exit = cpufreq_userspace_policy_exit,
.start = cpufreq_userspace_policy_start,
.stop = cpufreq_userspace_policy_stop,
.limits = cpufreq_userspace_policy_limits,
[CPUFREQ] Eliminate cpufreq_userspace scaling_setspeed deadlock Eliminate cpufreq_userspace scaling_setspeed deadlock. Luming Yu recently uncovered yet another cpufreq related deadlock. One thread that continuously switches the governors and the other thread that repeatedly cats the contents of cpufreq directory causes both these threads to go into a deadlock. Detailed examination of the deadlock showed the exact flow before the deadlock as: Thread 1 Thread 2 ________ ________ cats files under /sys/devices/.../cpufreq/ Set governor to userspace Adds a new sysfs entry for scaling_setspeed cats files under /sys/devices/.../cpufreq/ Set governor to performance Holds cpufreq_rw_sem in write mode Sends a STOP notify to userspace governor cat /sys/devices/.../cpufreq/scaling_setspeed Gets a handle on the above sysfs entry with sysfs_get_active Blocks while trying to get cpufreq_rw_sem in read mode Remove a sysfs entry for scaling_setspeed Blocks on sysfs_deactivate while waiting for earlier get_active (on other thread) to drain At this point both threads go into deadlock and any other thread that tries to do anything with sysfs cpufreq will also block. There seems to be no easy way to avoid this deadlock as long as cpufreq_userspace adds/removes the sysfs entry under same kobject as cpufreq. Below patch moves scaling_setspeed to cpufreq.c, keeping it always and calling back the governor on read/write. This is the cleanest fix I could think of, even though adding two callbacks in governor structure just for this seems unnecessary. Note that the change makes scaling_setspeed under /sys/.../cpufreq permanent and returns <unsupported> when governor is not userspace. Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com> Signed-off-by: Dave Jones <davej@redhat.com>
2007-10-27 00:18:21 +07:00
.store_setspeed = cpufreq_set,
.show_setspeed = show_speed,
.owner = THIS_MODULE,
};
static int __init cpufreq_gov_userspace_init(void)
{
return cpufreq_register_governor(&cpufreq_gov_userspace);
}
static void __exit cpufreq_gov_userspace_exit(void)
{
cpufreq_unregister_governor(&cpufreq_gov_userspace);
}
MODULE_AUTHOR("Dominik Brodowski <linux@brodo.de>, "
"Russell King <rmk@arm.linux.org.uk>");
MODULE_DESCRIPTION("CPUfreq policy governor 'userspace'");
MODULE_LICENSE("GPL");
#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_USERSPACE
struct cpufreq_governor *cpufreq_default_governor(void)
{
return &cpufreq_gov_userspace;
}
fs_initcall(cpufreq_gov_userspace_init);
#else
module_init(cpufreq_gov_userspace_init);
#endif
module_exit(cpufreq_gov_userspace_exit);