| Message ID | 1527144234-96396-1-git-send-email-kevin.wangtao@hisilicon.com (mailing list archive) |
|---|---|
| State | Not Applicable, archived |
| Headers | show |
On Thu, May 24, 2018 at 8:43 AM, Kevin Wangtao <kevin.wangtao@hisilicon.com> wrote: > consider such situation, current user_policy.min is 1000000, > current user_policy.max is 1200000, in cpufreq_set_policy, > other driver may update policy.min to 1200000, policy.max to > 1300000. After that, If we input "echo 1300000 > scaling_min_freq", > then user_policy.min will be 1300000, and user_policy.max is > still 1200000, because the input value is checked with policy.max > not user_policy.max. if we get all related cpus offline and > online again, it will cause cpufreq_init_policy fail because > user_policy.min is higher than user_policy.max. How do you reproduce this, exactly? > The solution is when user space tries to write scaling_(max|min)_freq, > the min/max of new_policy should be reinitialized with min/max > of user_policy, like what cpufreq_update_policy does. > > Signed-off-by: Kevin Wangtao <kevin.wangtao@hisilicon.com> > --- > drivers/cpufreq/cpufreq.c | 2 ++ > 1 file changed, 2 insertions(+) > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c > index b79c532..8b33e08 100644 > --- a/drivers/cpufreq/cpufreq.c > +++ b/drivers/cpufreq/cpufreq.c > @@ -697,6 +697,8 @@ static ssize_t store_##file_name \ > struct cpufreq_policy new_policy; \ > \ > memcpy(&new_policy, policy, sizeof(*policy)); \ > + new_policy->min = policy->user_policy.min; \ > + new_policy->max = policy->user_policy.max; \ It looks like you haven't even tried to build this, have you? > \ > ret = sscanf(buf, "%u", &new_policy.object); \ > if (ret != 1) \ > -- > 2.8.1 >
在 2018/5/24 15:45, Rafael J. Wysocki 写道: > On Thu, May 24, 2018 at 8:43 AM, Kevin Wangtao > <kevin.wangtao@hisilicon.com> wrote: >> consider such situation, current user_policy.min is 1000000, >> current user_policy.max is 1200000, in cpufreq_set_policy, >> other driver may update policy.min to 1200000, policy.max to >> 1300000. After that, If we input "echo 1300000 > scaling_min_freq", >> then user_policy.min will be 1300000, and user_policy.max is >> still 1200000, because the input value is checked with policy.max >> not user_policy.max. if we get all related cpus offline and >> online again, it will cause cpufreq_init_policy fail because >> user_policy.min is higher than user_policy.max. > > How do you reproduce this, exactly? I write a driver register CPUFREQ_POLICY_NOTIFIER, and when event is CPUFREQ_ADJUST, it will modify policy's min/max according to some conditions, I test it with writing scaling_(max|min)_freq to traverse all frequencies repeatly, and also repeat hotplug as background. > >> The solution is when user space tries to write scaling_(max|min)_freq, >> the min/max of new_policy should be reinitialized with min/max >> of user_policy, like what cpufreq_update_policy does. >> >> Signed-off-by: Kevin Wangtao <kevin.wangtao@hisilicon.com> >> --- >> drivers/cpufreq/cpufreq.c | 2 ++ >> 1 file changed, 2 insertions(+) >> >> diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c >> index b79c532..8b33e08 100644 >> --- a/drivers/cpufreq/cpufreq.c >> +++ b/drivers/cpufreq/cpufreq.c >> @@ -697,6 +697,8 @@ static ssize_t store_##file_name \ >> struct cpufreq_policy new_policy; \ >> \ >> memcpy(&new_policy, policy, sizeof(*policy)); \ >> + new_policy->min = policy->user_policy.min; \ new_policy.min = policy->user_policy.min; >> + new_policy->max = policy->user_policy.max; \ new_policy.max = policy->user_policy.max > > It looks like you haven't even tried to build this, have you? sorry for that, I test it on another branch and write this patch manually without build > >> \ >> ret = sscanf(buf, "%u", &new_policy.object); \ >> if (ret != 1) \ >> -- >> 2.8.1 >> > > . >
在 2018/5/24 15:45, Rafael J. Wysocki 写道: > On Thu, May 24, 2018 at 8:43 AM, Kevin Wangtao > <kevin.wangtao@hisilicon.com> wrote: >> consider such situation, current user_policy.min is 1000000, >> current user_policy.max is 1200000, in cpufreq_set_policy, >> other driver may update policy.min to 1200000, policy.max to >> 1300000. After that, If we input "echo 1300000 > scaling_min_freq", >> then user_policy.min will be 1300000, and user_policy.max is >> still 1200000, because the input value is checked with policy.max >> not user_policy.max. if we get all related cpus offline and >> online again, it will cause cpufreq_init_policy fail because >> user_policy.min is higher than user_policy.max. > > How do you reproduce this, exactly? I can also reproduce this issue with upstream code, write max frequency to scaling_max_freq and scaling_min_freq, run benchmark to let cpu cooling take effect to clip freq, then write the cliped freq to scaling_max_freq, thus user_policy.min is still max frequency but user_policy.max is cliped freq which is lower than max frequency. > >> The solution is when user space tries to write scaling_(max|min)_freq, >> the min/max of new_policy should be reinitialized with min/max >> of user_policy, like what cpufreq_update_policy does. >> >> Signed-off-by: Kevin Wangtao <kevin.wangtao@hisilicon.com> >> --- >> drivers/cpufreq/cpufreq.c | 2 ++ >> 1 file changed, 2 insertions(+) >> >> diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c >> index b79c532..8b33e08 100644 >> --- a/drivers/cpufreq/cpufreq.c >> +++ b/drivers/cpufreq/cpufreq.c >> @@ -697,6 +697,8 @@ static ssize_t store_##file_name \ >> struct cpufreq_policy new_policy; \ >> \ >> memcpy(&new_policy, policy, sizeof(*policy)); \ >> + new_policy->min = policy->user_policy.min; \ >> + new_policy->max = policy->user_policy.max; \ > > It looks like you haven't even tried to build this, have you? > >> \ >> ret = sscanf(buf, "%u", &new_policy.object); \ >> if (ret != 1) \ >> -- >> 2.8.1 >> > > . >
Hi Kevin,
Thank you for the patch! Yet something to improve:
[auto build test ERROR on pm/linux-next]
[also build test ERROR on v4.17-rc6]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system]
url: https://github.com/0day-ci/linux/commits/Kevin-Wangtao/cpufreq-reinitialize-new-policy-min-max-when-writing-scaling_-max-min-_freq/20180527-132510
base: https://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm.git linux-next
config: i386-randconfig-x079-201821 (attached as .config)
compiler: gcc-7 (Debian 7.3.0-16) 7.3.0
reproduce:
# save the attached .config to linux build tree
make ARCH=i386
All errors (new ones prefixed by >>):
drivers/cpufreq/cpufreq.c: In function 'store_scaling_min_freq':
>> drivers/cpufreq/cpufreq.c:700:12: error: invalid type argument of '->' (have 'struct cpufreq_policy')
new_policy->min = policy->user_policy.min; \
^
drivers/cpufreq/cpufreq.c:715:1: note: in expansion of macro 'store_one'
store_one(scaling_min_freq, min);
^~~~~~~~~
drivers/cpufreq/cpufreq.c:701:12: error: invalid type argument of '->' (have 'struct cpufreq_policy')
new_policy->max = policy->user_policy.max; \
^
drivers/cpufreq/cpufreq.c:715:1: note: in expansion of macro 'store_one'
store_one(scaling_min_freq, min);
^~~~~~~~~
drivers/cpufreq/cpufreq.c: In function 'store_scaling_max_freq':
>> drivers/cpufreq/cpufreq.c:700:12: error: invalid type argument of '->' (have 'struct cpufreq_policy')
new_policy->min = policy->user_policy.min; \
^
drivers/cpufreq/cpufreq.c:716:1: note: in expansion of macro 'store_one'
store_one(scaling_max_freq, max);
^~~~~~~~~
drivers/cpufreq/cpufreq.c:701:12: error: invalid type argument of '->' (have 'struct cpufreq_policy')
new_policy->max = policy->user_policy.max; \
^
drivers/cpufreq/cpufreq.c:716:1: note: in expansion of macro 'store_one'
store_one(scaling_max_freq, max);
^~~~~~~~~
vim +700 drivers/cpufreq/cpufreq.c
685
686 static int cpufreq_set_policy(struct cpufreq_policy *policy,
687 struct cpufreq_policy *new_policy);
688
689 /**
690 * cpufreq_per_cpu_attr_write() / store_##file_name() - sysfs write access
691 */
692 #define store_one(file_name, object) \
693 static ssize_t store_##file_name \
694 (struct cpufreq_policy *policy, const char *buf, size_t count) \
695 { \
696 int ret, temp; \
697 struct cpufreq_policy new_policy; \
698 \
699 memcpy(&new_policy, policy, sizeof(*policy)); \
> 700 new_policy->min = policy->user_policy.min; \
701 new_policy->max = policy->user_policy.max; \
702 \
703 ret = sscanf(buf, "%u", &new_policy.object); \
704 if (ret != 1) \
705 return -EINVAL; \
706 \
707 temp = new_policy.object; \
708 ret = cpufreq_set_policy(policy, &new_policy); \
709 if (!ret) \
710 policy->user_policy.object = temp; \
711 \
712 return ret ? ret : count; \
713 }
714
---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation
On Friday, May 25, 2018 4:54:04 AM CEST Wangtao (Kevin, Kirin) wrote: > > 在 2018/5/24 15:45, Rafael J. Wysocki 写道: > > On Thu, May 24, 2018 at 8:43 AM, Kevin Wangtao > > <kevin.wangtao@hisilicon.com> wrote: > >> consider such situation, current user_policy.min is 1000000, > >> current user_policy.max is 1200000, in cpufreq_set_policy, > >> other driver may update policy.min to 1200000, policy.max to > >> 1300000. After that, If we input "echo 1300000 > scaling_min_freq", > >> then user_policy.min will be 1300000, and user_policy.max is > >> still 1200000, because the input value is checked with policy.max > >> not user_policy.max. if we get all related cpus offline and > >> online again, it will cause cpufreq_init_policy fail because > >> user_policy.min is higher than user_policy.max. > > > > How do you reproduce this, exactly? > > I write a driver register CPUFREQ_POLICY_NOTIFIER, and when event is CPUFREQ_ADJUST, > it will modify policy's min/max according to some conditions, I test it with writing > scaling_(max|min)_freq to traverse all frequencies repeatly, and also repeat hotplug > as background. You are expected to use cpufreq_update_policy() to update the limits in policy notifiers. Do you use it in your driver?
On Saturday, May 26, 2018 8:50:46 AM CEST Wangtao (Kevin, Kirin) wrote: > > 在 2018/5/24 15:45, Rafael J. Wysocki 写道: > > On Thu, May 24, 2018 at 8:43 AM, Kevin Wangtao > > <kevin.wangtao@hisilicon.com> wrote: > >> consider such situation, current user_policy.min is 1000000, > >> current user_policy.max is 1200000, in cpufreq_set_policy, > >> other driver may update policy.min to 1200000, policy.max to > >> 1300000. After that, If we input "echo 1300000 > scaling_min_freq", > >> then user_policy.min will be 1300000, and user_policy.max is > >> still 1200000, because the input value is checked with policy.max > >> not user_policy.max. if we get all related cpus offline and > >> online again, it will cause cpufreq_init_policy fail because > >> user_policy.min is higher than user_policy.max. > > > > How do you reproduce this, exactly? > > I can also reproduce this issue with upstream code, write max frequency to scaling_max_freq > and scaling_min_freq, run benchmark to let cpu cooling take effect to clip freq, then write > the cliped freq to scaling_max_freq, thus user_policy.min is still max frequency but user_policy.max > is cliped freq which is lower than max frequency. OK, this is a bit more convincing. It looks like bad interaction between cpufreq_update_policy() and updates of the limits via sysfs.
diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c index b79c532..8b33e08 100644 --- a/drivers/cpufreq/cpufreq.c +++ b/drivers/cpufreq/cpufreq.c @@ -697,6 +697,8 @@ static ssize_t store_##file_name \ struct cpufreq_policy new_policy; \ \ memcpy(&new_policy, policy, sizeof(*policy)); \ + new_policy->min = policy->user_policy.min; \ + new_policy->max = policy->user_policy.max; \ \ ret = sscanf(buf, "%u", &new_policy.object); \ if (ret != 1) \
consider such situation, current user_policy.min is 1000000, current user_policy.max is 1200000, in cpufreq_set_policy, other driver may update policy.min to 1200000, policy.max to 1300000. After that, If we input "echo 1300000 > scaling_min_freq", then user_policy.min will be 1300000, and user_policy.max is still 1200000, because the input value is checked with policy.max not user_policy.max. if we get all related cpus offline and online again, it will cause cpufreq_init_policy fail because user_policy.min is higher than user_policy.max. The solution is when user space tries to write scaling_(max|min)_freq, the min/max of new_policy should be reinitialized with min/max of user_policy, like what cpufreq_update_policy does. Signed-off-by: Kevin Wangtao <kevin.wangtao@hisilicon.com> --- drivers/cpufreq/cpufreq.c | 2 ++ 1 file changed, 2 insertions(+)