| Message ID | 20171226072626.GA4153@gaurav.jindal |
|---|---|
| State | New |
| Headers | show |
| Series | cpuidle: preventive check in cpuidle_select against crash | expand |
On Tue, Dec 26, 2017 at 8:26 AM, gaurav jindal <gauravjindal1104@gmail.com> wrote: > When selecting the idle state using cpuidle_select, there is no > check on cpuidle_curr_governor. In cpuidle_switch_governor, > cpuidle_currr_governor can be set to NULL to specify "disabled". How exactly?
On Wed, Dec 27, 2017 at 01:42:58AM +0100, Rafael J. Wysocki wrote: > On Tue, Dec 26, 2017 at 8:26 AM, gaurav jindal > <gauravjindal1104@gmail.com> wrote: > > When selecting the idle state using cpuidle_select, there is no > > check on cpuidle_curr_governor. In cpuidle_switch_governor, > > cpuidle_currr_governor can be set to NULL to specify "disabled". > > How exactly? In cpuidle_switch_governor: /** * cpuidle_switch_governor - changes the governor * @gov: the new target governor * * NOTE: "gov" can be NULL to specify disabled * Must be called with cpuidle_lock acquired. */ int cpuidle_switch_governor(struct cpuidle_governor *gov) { struct cpuidle_device *dev; if (gov == cpuidle_curr_governor) return 0; cpuidle_uninstall_idle_handler(); if (cpuidle_curr_governor) { list_for_each_entry(dev, &cpuidle_detected_devices, device_list) cpuidle_disable_device(dev); } cpuidle_curr_governor = gov; This allows to set the cpuidle_switch_governor as NULL. Although there is no current code flow leading here, but it has a potential for bug in future. So may be better to have prevention.
On Wed, Dec 27, 2017 at 2:57 AM, gaurav jindal <gauravjindal1104@gmail.com> wrote: > On Wed, Dec 27, 2017 at 01:42:58AM +0100, Rafael J. Wysocki wrote: >> On Tue, Dec 26, 2017 at 8:26 AM, gaurav jindal >> <gauravjindal1104@gmail.com> wrote: >> > When selecting the idle state using cpuidle_select, there is no >> > check on cpuidle_curr_governor. In cpuidle_switch_governor, >> > cpuidle_currr_governor can be set to NULL to specify "disabled". >> >> How exactly? > > In cpuidle_switch_governor: > > /** > * cpuidle_switch_governor - changes the governor > * @gov: the new target governor > * > * NOTE: "gov" can be NULL to specify disabled > * Must be called with cpuidle_lock acquired. > */ > int cpuidle_switch_governor(struct cpuidle_governor *gov) > { > struct cpuidle_device *dev; > > if (gov == cpuidle_curr_governor) > return 0; > > cpuidle_uninstall_idle_handler(); > > if (cpuidle_curr_governor) { > list_for_each_entry(dev, &cpuidle_detected_devices, device_list) > cpuidle_disable_device(dev); > } > > cpuidle_curr_governor = gov; > > This allows to set the cpuidle_switch_governor as NULL. Although there is no > current code flow leading here, but it has a potential for bug in future. So > may be better to have prevention. Or maybe not. Why don't you make cpuidle_switch_governor() check the argument against NULL instead?
On Wed, Dec 27, 2017 at 03:30:02AM +0100, Rafael J. Wysocki wrote: > On Wed, Dec 27, 2017 at 2:57 AM, gaurav jindal > <gauravjindal1104@gmail.com> wrote: > > On Wed, Dec 27, 2017 at 01:42:58AM +0100, Rafael J. Wysocki wrote: > >> On Tue, Dec 26, 2017 at 8:26 AM, gaurav jindal > >> <gauravjindal1104@gmail.com> wrote: > >> > When selecting the idle state using cpuidle_select, there is no > >> > check on cpuidle_curr_governor. In cpuidle_switch_governor, > >> > cpuidle_currr_governor can be set to NULL to specify "disabled". > >> > >> How exactly? > > > > In cpuidle_switch_governor: > > > > /** > > * cpuidle_switch_governor - changes the governor > > * @gov: the new target governor > > * > > * NOTE: "gov" can be NULL to specify disabled > > * Must be called with cpuidle_lock acquired. > > */ > > int cpuidle_switch_governor(struct cpuidle_governor *gov) > > { > > struct cpuidle_device *dev; > > > > if (gov == cpuidle_curr_governor) > > return 0; > > > > cpuidle_uninstall_idle_handler(); > > > > if (cpuidle_curr_governor) { > > list_for_each_entry(dev, &cpuidle_detected_devices, device_list) > > cpuidle_disable_device(dev); > > } > > > > cpuidle_curr_governor = gov; > > > > This allows to set the cpuidle_switch_governor as NULL. Although there is no > > current code flow leading here, but it has a potential for bug in future. So > > may be better to have prevention. > > Or maybe not. > > Why don't you make cpuidle_switch_governor() check the argument > against NULL instead? If we check gov (argument passed in cpuidle_switch_governor())against NULL in cpuidle_switch_governor, can be a problem in a case where it is called as cpuidle_switch_governor(NULL); If cpuidle_curr_governor is not NULL, first the device is disabled. if (cpuidle_curr_governor) { list_for_each_entry(dev, &cpuidle_detected_devices, device_list) cpuidle_disable_device(dev); } after this cpuidle_curr_governor is set to gov, which is NULL in this case. cpuidle_curr_governor = gov; /* if is not updated by inserting a check, it will have an oudated value*/ Now, if gov is not NULL (which it is in this case), cpuidle device is enabled if (gov) { list_for_each_entry(dev, &cpuidle_detected_devices, device_list) cpuidle_enable_device(dev); cpuidle_install_idle_handler(); printk(KERN_INFO "cpuidle: using governor %s\n", gov->name); } If we check for gov against NULL in this function, it will produce dangling pointers and resource leaks.
On Friday, December 29, 2017 7:45:22 PM CET gaurav jindal wrote: > On Wed, Dec 27, 2017 at 03:30:02AM +0100, Rafael J. Wysocki wrote: > > On Wed, Dec 27, 2017 at 2:57 AM, gaurav jindal > > <gauravjindal1104@gmail.com> wrote: > > > On Wed, Dec 27, 2017 at 01:42:58AM +0100, Rafael J. Wysocki wrote: > > >> On Tue, Dec 26, 2017 at 8:26 AM, gaurav jindal > > >> <gauravjindal1104@gmail.com> wrote: > > >> > When selecting the idle state using cpuidle_select, there is no > > >> > check on cpuidle_curr_governor. In cpuidle_switch_governor, > > >> > cpuidle_currr_governor can be set to NULL to specify "disabled". > > >> > > >> How exactly? > > > > > > In cpuidle_switch_governor: > > > > > > /** > > > * cpuidle_switch_governor - changes the governor > > > * @gov: the new target governor > > > * > > > * NOTE: "gov" can be NULL to specify disabled > > > * Must be called with cpuidle_lock acquired. > > > */ > > > int cpuidle_switch_governor(struct cpuidle_governor *gov) > > > { > > > struct cpuidle_device *dev; > > > > > > if (gov == cpuidle_curr_governor) > > > return 0; > > > > > > cpuidle_uninstall_idle_handler(); > > > > > > if (cpuidle_curr_governor) { > > > list_for_each_entry(dev, &cpuidle_detected_devices, device_list) > > > cpuidle_disable_device(dev); > > > } > > > > > > cpuidle_curr_governor = gov; > > > > > > This allows to set the cpuidle_switch_governor as NULL. Although there is no > > > current code flow leading here, but it has a potential for bug in future. So > > > may be better to have prevention. > > > > Or maybe not. > > > > Why don't you make cpuidle_switch_governor() check the argument > > against NULL instead? > > If we check gov (argument passed in cpuidle_switch_governor())against > NULL in cpuidle_switch_governor, can be a problem in a case where it > is called as > cpuidle_switch_governor(NULL); > > If cpuidle_curr_governor is not NULL, first the device is disabled. > > if (cpuidle_curr_governor) { > list_for_each_entry(dev, &cpuidle_detected_devices, device_list) > cpuidle_disable_device(dev); > } > > after this cpuidle_curr_governor is set to gov, which is NULL in this case. > > cpuidle_curr_governor = gov; > /* if is not updated by inserting a check, it will have an oudated value*/ > > Now, if gov is not NULL (which it is in this case), cpuidle device is enabled > > if (gov) { > list_for_each_entry(dev, &cpuidle_detected_devices, device_list) > cpuidle_enable_device(dev); > cpuidle_install_idle_handler(); > printk(KERN_INFO "cpuidle: using governor %s\n", gov->name); > } > If we check for gov against NULL in this function, it will produce > dangling pointers and resource leaks. I didn't recommend you to introduce bugs. Just return -EINVAL if gov is NULL before checking if gov is equal to cpuidle_curr_governor. Thanks, Rafael
On Wed, Jan 03, 2018 at 12:16:26PM +0100, Rafael J. Wysocki wrote: > On Friday, December 29, 2017 7:45:22 PM CET gaurav jindal wrote: > > On Wed, Dec 27, 2017 at 03:30:02AM +0100, Rafael J. Wysocki wrote: > > > On Wed, Dec 27, 2017 at 2:57 AM, gaurav jindal > > > <gauravjindal1104@gmail.com> wrote: > > > > On Wed, Dec 27, 2017 at 01:42:58AM +0100, Rafael J. Wysocki wrote: > > > >> On Tue, Dec 26, 2017 at 8:26 AM, gaurav jindal > > > >> <gauravjindal1104@gmail.com> wrote: > > > >> > When selecting the idle state using cpuidle_select, there is no > > > >> > check on cpuidle_curr_governor. In cpuidle_switch_governor, > > > >> > cpuidle_currr_governor can be set to NULL to specify "disabled". > > > >> > > > >> How exactly? > > > > > > > > In cpuidle_switch_governor: > > > > > > > > /** > > > > * cpuidle_switch_governor - changes the governor > > > > * @gov: the new target governor > > > > * > > > > * NOTE: "gov" can be NULL to specify disabled > > > > * Must be called with cpuidle_lock acquired. > > > > */ > > > > int cpuidle_switch_governor(struct cpuidle_governor *gov) > > > > { > > > > struct cpuidle_device *dev; > > > > > > > > if (gov == cpuidle_curr_governor) > > > > return 0; > > > > > > > > cpuidle_uninstall_idle_handler(); > > > > > > > > if (cpuidle_curr_governor) { > > > > list_for_each_entry(dev, &cpuidle_detected_devices, device_list) > > > > cpuidle_disable_device(dev); > > > > } > > > > > > > > cpuidle_curr_governor = gov; > > > > > > > > This allows to set the cpuidle_switch_governor as NULL. Although there is no > > > > current code flow leading here, but it has a potential for bug in future. So > > > > may be better to have prevention. > > > > > > Or maybe not. > > > > > > Why don't you make cpuidle_switch_governor() check the argument > > > against NULL instead? > > > > If we check gov (argument passed in cpuidle_switch_governor())against > > NULL in cpuidle_switch_governor, can be a problem in a case where it > > is called as > > cpuidle_switch_governor(NULL); > > > > If cpuidle_curr_governor is not NULL, first the device is disabled. > > > > if (cpuidle_curr_governor) { > > list_for_each_entry(dev, &cpuidle_detected_devices, device_list) > > cpuidle_disable_device(dev); > > } > > > > after this cpuidle_curr_governor is set to gov, which is NULL in this case. > > > > cpuidle_curr_governor = gov; > > /* if is not updated by inserting a check, it will have an oudated value*/ > > > > Now, if gov is not NULL (which it is in this case), cpuidle device is enabled > > > > if (gov) { > > list_for_each_entry(dev, &cpuidle_detected_devices, device_list) > > cpuidle_enable_device(dev); > > cpuidle_install_idle_handler(); > > printk(KERN_INFO "cpuidle: using governor %s\n", gov->name); > > } > > If we check for gov against NULL in this function, it will produce > > dangling pointers and resource leaks. > > I didn't recommend you to introduce bugs. > I did not intend to do so. I am really sorry it got expressed in that way :(. > Just return -EINVAL if gov is NULL before checking if gov is equal to > cpuidle_curr_governor. > Okay > Thanks, > Rafael > this patch checks if the new governor is NULL before updating the cupidle_curr_governor. Signed-off-by: gaurav jindal<gauravjindal1104@gmail.com> --- diff --git a/drivers/cpuidle/governor.c b/drivers/cpuidle/governor.c index 4e78263..5d359af 100644 --- a/drivers/cpuidle/governor.c +++ b/drivers/cpuidle/governor.c @@ -36,14 +36,15 @@ static struct cpuidle_governor * __cpuidle_find_governor(const char *str) /** * cpuidle_switch_governor - changes the governor * @gov: the new target governor - * - * NOTE: "gov" can be NULL to specify disabled * Must be called with cpuidle_lock acquired. */ int cpuidle_switch_governor(struct cpuidle_governor *gov) { struct cpuidle_device *dev; + if (!gov) + return -EINVAL; + if (gov == cpuidle_curr_governor) return 0;
diff --git a/drivers/cpuidle/cpuidle.c b/drivers/cpuidle/cpuidle.c index 68a1682..bf08e3a 100644 --- a/drivers/cpuidle/cpuidle.c +++ b/drivers/cpuidle/cpuidle.c @@ -268,6 +268,19 @@ int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv, */ int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev) { + + /* Since negative return is not allowed + * we have to return 0 even if the + * framework cannot select the idle state + */ + if (!cpuidle_curr_governor) { + pr_err("idle governor is disabled\n"); + return 0; + } + if (!cpuidle_curr_governor->select) { + pr_err("idle governor select is NULL\n"); + return 0; + } return cpuidle_curr_governor->select(drv, dev); }
When selecting the idle state using cpuidle_select, there is no check on cpuidle_curr_governor. In cpuidle_switch_governor, cpuidle_currr_governor can be set to NULL to specify "disabled". Since cpuidle_select cannot return negative value, it has to return 0 in case of error. Printing logs and returning can help in debugging and preventing possible kernel crash scenarios. Signed-off-by: Gaurav Jindal<gauravjindal1104@gmail.com> ---