alarmtimer, PM: suspend: Expose a function from
PATCH 1/1 alarmtimer, PM: suspend: Expose a function from
Hi!
I am proposing a patch in which I want to return the errno code ETIME instead of EBUSY in enter_state() in the kernel suspend flow. Currently, EBUSY is returned when an imminent alarm is pending which is checked in alarmtimer_suspend() in alarmtimer.c. The proposed patch series moves the check to enter_state() in suspend.c to catch a potential suspend failure early in the suspend flow. I want to replace EBUSY with ETIME to make it more diagnosable in userspace, and may be more appropriate considering a timer is about to expire.
I am reaching out to get an opinion from the
suspend maintainers if this would act as any potential risk in the suspend flow which only has EBUSY, EAGAIN, and EINVAL as return error codes currently. This has been developed as part of a patch series, and only the patch of interest is below this message. Any feedback or insights would be greatly appreciated.
Thank you,
Pranav Prasad
The alarmtimer driver currently fails suspend attempts when there is an alarm pending within the next suspend_check_duration_ns nanoseconds, since
the system is expected to wake up soon anyway. The entire suspend process
is initiated even though the system will immediately awaken. This process includes substantial work before the suspend fails and additional work afterwards to undo the failed suspend that was attempted. Therefore on battery-powered devices that initiate suspend attempts from userspace, it may be advantageous to be able to fail the suspend earlier in the suspend flow to avoid power consumption instead of unnecessarily doing extra work. As one data point, an analysis of a subset of Android devices showed that imminent alarms account for roughly 40% of all suspend failures on average leading to unnecessary power wastage.
To facilitate this, expose
function time_check_suspend_fail() from alarmtimer to be used by the power subsystem to perform the check earlier in the suspend flow. Perform the check in enter_state() and return early if an alarm is to be fired in the next suspend_check_duration_ns nanoseconds, failing suspend.
--- include/linux/time.h | 1 + kernel/power/suspend.c | 3 ++ kernel/time/alarmtimer.c | 113 ++++++++++++++++++++++++++++----------- 3 files changed, 87 insertions(+), 30 deletions(-)
include/linux/time.h +1 -0
@@ -56,6 +56,7 @@ struct tm { };void time64_to_tm (time64_t totalsecs,int offset,struct tm *result); +int time_check_suspend_fail (void );# include <linux/time32.h>
kernel/power/suspend.c +3 -0
@@ -26,6 +26,7 @@#include <linux/suspend.h> #include <linux/syscore_ops.h> #include <linux/swait.h> +#include <linux/time.h> #include <linux/ftrace.h> #include <trace/events/power.h> #include <linux/compiler.h>
@@ -564,6 +565,8 @@ static int enter_state(suspend_state_t state)#endif }else if (!valid_state (state)) {return -EINVAL ; + }else if (time_check_suspend_fail ()) { +return -ETIME ; }if (!mutex_trylock (&system_transition_mutex))return -EBUSY ;
kernel/time/alarmtimer.c +83 -30
138 lines not on this page · raw message
Signature · 1 line
2.43.0.429.g432eaa2c6b-goog
Rafael J. Wysocki
maintainerQuoted 74 lines · Pranav Prasad
Hi!
I am proposing a patch in which I want to return the errno code ETIME instead of EBUSY in
enter_state()in the kernel suspend flow. Currently, EBUSY is returned when an imminent alarm is pending which is checked inalarmtimer_suspend()inalarmtimer.c. The proposed patch series moves the check toenter_state()insuspend.cto catch a potential suspend failure early in the suspend flow. I want to replace EBUSY with ETIME to make it more diagnosable in userspace, and may be more appropriate considering a timer is about to expire.I am reaching out to get an opinion from the
suspend maintainers if this would act as any potential risk in the suspend flow which only has EBUSY, EAGAIN, and EINVAL as return error codes currently. This has been developed as part of a patch series, and only the patch of interest is below this message. Any feedback or insights would be greatly appreciated.Thank you,
Pranav PrasadThe alarmtimer driver currently fails suspend attempts when there is an alarm pending within the next
suspend_check_duration_nsnanoseconds, since
the system is expected to wake up soon anyway. The entire suspend process
is initiated even though the system will immediately awaken. This process includes substantial work before the suspend fails and additional work afterwards to undo the failed suspend that was attempted. Therefore on battery-powered devices that initiate suspend attempts from userspace, it may be advantageous to be able to fail the suspend earlier in the suspend flow to avoid power consumption instead of unnecessarily doing extra work. As one data point, an analysis of a subset of Android devices showed that imminent alarms account for roughly 40% of all suspend failures on average leading to unnecessary power wastage.To facilitate this, expose
functiontime_check_suspend_fail()from alarmtimer to be used by the power subsystem to perform the check earlier in the suspend flow. Perform the check inenter_state()and return early if an alarm is to be fired in the nextsuspend_check_duration_nsnanoseconds, failing suspend.Signed-off-by Pranav Prasad <[email protected]>Signed-off-by Kelly Rossmoyer <[email protected]>--- include/linux/time.h | 1 + kernel/power/suspend.c | 3 ++ kernel/time/alarmtimer.c | 113 ++++++++++++++++++++++++++++----------- 3 files changed, 87 insertions(+), 30 deletions(-)include/linux/time.h +1
-0@@ -56,6 +56,7 @@ struct tm { };void time64_to_tm (time64_t totalsecs,int offset,struct tm *result); +int time_check_suspend_fail (void );# include <linux/time32.h> kernel/power/suspend.c +1
-0@@ -26,6 +26,7 @@#include <linux/suspend.h> #include <linux/syscore_ops.h> #include <linux/swait.h> +#include <linux/time.h> #include <linux/ftrace.h> #include <trace/events/power.h> #include <linux/compiler.h>
kernel/power/suspend.c static int enter_state(suspend_state_t state) · line 569
Quoted 5 lines
kernel/power/suspend.c +2
-0@@ -564,6 +565,8 @@ static int enter_state(suspend_state_t state)#endif }else if (!valid_state (state)) {return -EINVAL ; + }else if (time_check_suspend_fail ()) { +return -ETIME ;
This causes a function defined in modular code to be called from non-modular code which is an obvious mistake.
It also makes the generic suspend code call a function defined in a random driver, which is a total no-go as far as I am concerned.
Why don't you instead define a PM notifier in the alarmtimer driver and check if it is going to trigger shortly from there? PM notifiers run before the tasks freezer, so there would be a little difference timing-wise and you can return whatever error code you like from there. As an additional benefit, you'd be able to handle hibernation in the same way.
Thanks!
Quoted 11 lines · Rafael J. Wysocki
kernel/power/suspend.c static int enter_state(suspend_state_t state) · line 569
Quoted 6 lines · Pranav Prasad
kernel/power/suspend.c +2
-0@@ -564,6 +565,8 @@ static int enter_state(suspend_state_t state)#endif }else if (!valid_state (state)) {return -EINVAL ; + }else if (time_check_suspend_fail ()) { +return -ETIME ;This causes a function defined in modular code to be called from non-modular code which is an obvious mistake.
Quoted 2 lines
It also makes the generic suspend code call a function defined in a random driver, which is a total no-go as far as I am concerned.
Alarmtimers is built-in core infrastructure and not a random modular driver, but nevertheless:
Quoted 6 lines
Why don't you instead define a PM notifier in the alarmtimer driver and check if it is going to trigger shortly from there? PM notifiers run before the tasks freezer, so there would be a little difference timing-wise and you can return whatever error code you like from there. As an additional benefit, you'd be able to handle hibernation in the same way.
Makes sense.
Thanks,
tglx
Quoted 18 lines · Pranav Prasad
Hi!
I am proposing a patch in which I want to return the errno code ETIME instead of EBUSY in
enter_state()in the kernel suspend flow. Currently, EBUSY is returned when an imminent alarm is pending which is checked inalarmtimer_suspend()inalarmtimer.c. The proposed patch series moves the check toenter_state()insuspend.cto catch a potential suspend failure early in the suspend flow. I want to replace EBUSY with ETIME to make it more diagnosable in userspace, and may be more appropriate considering a timer is about to expire.I am reaching out to get an opinion from the
suspend maintainers if this would act as any potential risk in the suspend flow which only has EBUSY, EAGAIN, and EINVAL as return error codes currently. This has been developed as part of a patch series, and only the patch of interest is below this message. Any feedback or insights would be greatly appreciated.
Quoted 2 lines
Thank you,
Pranav Prasad
Can you please use a cover letter instead of putting random stuff into the changelong?
Quoted 13 lines
The alarmtimer driver currently fails suspend attempts when there is an alarm pending within the next
suspend_check_duration_nsnanoseconds, since
the system is expected to wake up soon anyway. The entire suspend process
is initiated even though the system will immediately awaken. This process includes substantial work before the suspend fails and additional work afterwards to undo the failed suspend that was attempted. Therefore on battery-powered devices that initiate suspend attempts from userspace, it may be advantageous to be able to fail the suspend earlier in the suspend flow to avoid power consumption instead of unnecessarily doing extra work. As one data point, an analysis of a subset of Android devices showed that imminent alarms account for roughly 40% of all suspend failures on average leading to unnecessary power wastage.
Quoted 8 lines
To facilitate this, expose
functiontime_check_suspend_fail()from alarmtimer to be used by the power subsystem to perform the check earlier in the suspend flow. Perform the check inenter_state()and return early if an alarm is to be fired in the nextsuspend_check_duration_nsnanoseconds, failing suspend.Signed-off-by Pranav Prasad <[email protected]>Signed-off-by Kelly Rossmoyer <[email protected]>
This Signed-off-by chain is bogus.
Quoted 20 lines
kernel/time/alarmtimer.c +20
-0+/** +* alarmtimer_init_soonest - Initializes parameters to find soonest alarm. +* @min: ptr to relative time to the soonest alarm to expire +* @expires: ptr to absolute time of the soonest alarm to expire +* @type: ptr to alarm type +* +*/ +static void alarmtimer_init_soonest (ktime_t *min,ktime_t *expires,int *type) +{ +unsigned long flags; + +spin_lock_irqsave (&freezer_delta_lock, flags); + *min = freezer_delta; + *expires = freezer_expires; + *type = freezer_alarmtype; + freezer_delta =0 ; +spin_unlock_irqrestore (&freezer_delta_lock, flags); +} + +/**
kernel/time/alarmtimer.c static int alarmtimer_sysfs_add(void) · line 147
Quoted 10 lines
kernel/time/alarmtimer.c +10
-0+* alarmtimer_get_soonest - Finds the soonest alarm to expire among the alarm bases. +* @min: ptr to relative time to the soonest alarm to expire +* @expires: ptr to absolute time of the soonest alarm to expire +* @type: ptr to alarm type +* +*/ +static void alarmtimer_get_soonest (ktime_t *min,ktime_t *expires,int *type) +{ +int i; +unsigned long flags;
https://www.kernel.org/doc/html/latest/process/maintainer-tip.html#variable-declarations
Aside of that 'flags' wants to be in the loop scope.
Quoted 4 lines
kernel/time/alarmtimer.c +4
-0+ +/* Find the soonest timer to expire */ +for (i =0 ; i <ALARM_NUMTYPE ; i++) { +struct alarm_base *base = &alarm_bases[i];
kernel/time/alarmtimer.c static int alarmtimer_sysfs_add(void) · line 161
Quoted 10 lines
kernel/time/alarmtimer.c +10
-0+struct timerqueue_node *next; +ktime_t delta; + +spin_lock_irqsave (&base->lock, flags); + next =timerqueue_getnext (&base->timerqueue); +spin_unlock_irqrestore (&base->lock, flags); +if (!next) +continue ; + delta =ktime_sub (next->expires, base->get_ktime ()); +if (!(*min) || (delta < *min)) {
The inner brackets are pointless
Quoted 7 lines
kernel/time/alarmtimer.c +7
-0+ *expires = next->expires; + *min = delta; + *type = i; + } + } +} +
kernel/time/alarmtimer.c static int alarmtimer_sysfs_add(void) · line 178
Quoted 10 lines
kernel/time/alarmtimer.c +10
-0+/** +* time_check_suspend_fail - Check if suspend should be failed due to an +* alarm within the next suspend_check_duration nanoseconds. +* +* Returns error if suspend should be failed, else returns 0. +*/ +int time_check_suspend_fail (void ) +{ +ktime_t min, expires; +int type;
Why is this unconditional and not checking RTC dev?
Review
Quoted 2 lines
(quoted diff) +2
-0+ /* Initialize parameters to find soonest timer */ + alarmtimer_init_soonest(&min, &expires, &type);
How does that make sense? That function evaluates the freezer state, but there is nothing frozen when this is invoked.
kernel/time/alarmtimer.c static int alarmtimer_sysfs_add(void) · line 194
Quoted 12 lines
kernel/time/alarmtimer.c +12
-0+/* Find the soonest timer to expire */ +alarmtimer_get_soonest (&min, &expires, &type); + +if (min ==0 ) +return 0 ; + +if (ktime_to_ns (min) < suspend_check_duration_ns) +return -EBUSY ; + +return 0 ; +} +EXPORT_SYMBOL (time_check_suspend_fail);
What is this export for?
Quoted 7 lines
kernel/time/alarmtimer.c +1
-0+/** * alarmtimer_get_rtcdev - Return selected rtcdevice * @@ -296,49 +374,24 @@ EXPORT_SYMBOL_GPL(alarm_expires_remaining);static int alarmtimer_suspend (struct device *dev) {
...
Review
Quoted 2 lines
(quoted diff) +2
-0+ /* Initialize parameters to find soonest timer */ + alarmtimer_init_soonest(&min, &expires, &type);
This wants to be _after_ the RTC dev check, no?
kernel/time/alarmtimer.c EXPORT_SYMBOL_GPL(alarm_expires_remaining); · line 338
Quoted 10 lines
kernel/time/alarmtimer.c +2
-2rtc =alarmtimer_get_rtcdev ();/* If we have no rtcdev, just return */ if (!rtc)return 0 ; +/* Find the soonest timer to expire */ +alarmtimer_get_soonest (&min, &expires, &type); -if (ktime_to_ns (min) < suspend_check_duration_ns) { -pm_wakeup_event (dev, suspend_check_duration_ns/NSEC_PER_MSEC );
What injects the pm_wakeup_event after this change?
Thanks,
tglx