summaryrefslogtreecommitdiffstats
path: root/include/linux
diff options
context:
space:
mode:
authorVenki Pallipadi <venkatesh.pallipadi@intel.com>2007-10-26 10:18:21 -0700
committerDave Jones <davej@redhat.com>2008-02-06 22:57:58 -0500
commit9e76988e9390a4ff4d171f690586d0c58186b47e (patch)
treee033aa977a66b4ea2dc01b0e846eb7f6f8134857 /include/linux
parentb25e75899e449456409cfa1a3b042257c03d4355 (diff)
downloadlinux-9e76988e9390a4ff4d171f690586d0c58186b47e.tar.gz
linux-9e76988e9390a4ff4d171f690586d0c58186b47e.tar.bz2
linux-9e76988e9390a4ff4d171f690586d0c58186b47e.zip
[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>
Diffstat (limited to 'include/linux')
-rw-r--r--include/linux/cpufreq.h4
1 files changed, 4 insertions, 0 deletions
diff --git a/include/linux/cpufreq.h b/include/linux/cpufreq.h
index 23932d7741a9..ddd8652fc3f3 100644
--- a/include/linux/cpufreq.h
+++ b/include/linux/cpufreq.h
@@ -167,6 +167,10 @@ struct cpufreq_governor {
char name[CPUFREQ_NAME_LEN];
int (*governor) (struct cpufreq_policy *policy,
unsigned int event);
+ ssize_t (*show_setspeed) (struct cpufreq_policy *policy,
+ char *buf);
+ int (*store_setspeed) (struct cpufreq_policy *policy,
+ unsigned int freq);
unsigned int max_transition_latency; /* HW must be able to switch to
next freq faster than this value in nano secs or we
will fallback to performance governor */