dw_wdt watchdog on linux-rt 4.18.5-rt4 not triggering
Hi all!
I'm seeing an issue with the dw_wdt watchdog on the SoCFPGA ARM platform with the latest linux-rt v4.18.5-rt4. Actually I seem to have the same problem, that these patches try to fix:
38a1222ae4f364d5bd5221fe305dbb0889f45d15
Author: Christophe Leroy <[email protected]>
AuthorDate: Fri Dec 8 11:18:35 2017 +0100
Commit: Wim Van Sebroeck <[email protected]>
CommitDate: Thu Dec 28 20:45:57 2017 +0100
Follows: v4.15-rc3 (345)
Precedes: v4.16-rc1 (13997)
watchdog: core: make sure the watchdog worker always works
When running a command like 'chrt -f 50 dd if=/dev/zero of=/dev/null',
the watchdog_worker fails to service the HW watchdog and the HW watchdog fires long before the watchdog soft timeout.
At the moment, the watchdog_worker is invoked as a delayed work. Delayed works are handled by non realtime kernel threads. The WQ_HIGHPRI flag only increases the niceness of that threads.
This patch replaces the delayed work logic by kthread delayed work,
and sets the associated kernel task to SCHED_FIFO with the highest
priority, in order to ensure that the watchdog worker will run as
soon as possible.
1ff688209e2ed23f699269b9733993e2ce123fd2
Author: Christophe Leroy <[email protected]>
AuthorDate: Thu Jan 18 12:11:21 2018 +0100
Commit: Wim Van Sebroeck <[email protected]>
CommitDate: Sun Jan 21 12:44:59 2018 +0100
Follows: v4.15-rc3 (349)
Precedes: v4.16-rc1 (13993)
watchdog: core: make sure the watchdog_worker is not deferred
commit 4cd13c21b207e ("softirq: Let ksoftirqd do its job") has the
effect of deferring timer handling in case of high CPU load, hence
delaying the delayed work allthought the worker is running which high realtime priority.
As hrtimers are not managed by softirqs, this patch replaces the delayed work by a plain work and uses an hrtimer to schedule that work.
If I run the same test or 'chrt 50 hackbench 20 -l 150' or any task where I change the prio with chrt and that runs long enough, I get a system reset from the watchdog because it times out. This only happens if the watchdog is already enabled on boot and CONFIG_PREEMPT_RT_FULL is set.
Any idea if I'm missing something essential? If I understand it correctly, the two commits fix the framework and therefore the dw_wdt driver doesn't need any updates.
Best regards,
Steffen
--
Pengutronix e.K. | Steffen Trumtrar | Industrial Linux Solutions | http://www.pengutronix.de/ | Peiner Str. 6-8, 31137 Hildesheim, Germany| Phone: +49-5121-206917-0 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555|
Quoted 43 lines · Steffen Trumtrar
Hi all!
I'm seeing an issue with the
dw_wdtwatchdog on the SoCFPGA ARM platform with the latest linux-rt v4.18.5-rt4. Actually I seem to have the same problem, that these patches try to fix:
38a1222ae4f364d5bd5221fe305dbb0889f45d15
Author: Christophe Leroy <[email protected]>
AuthorDate: Fri Dec 8 11:18:35 2017 +0100
Commit: Wim Van Sebroeck <[email protected]>
CommitDate: Thu Dec 28 20:45:57 2017 +0100Follows: v4.15-rc3 (345)
Precedes: v4.16-rc1 (13997)watchdog: core: make sure the watchdog worker always works
When running a command like 'chrt -f 50 dd if=/dev/zero of=/dev/null', the
watchdog_workerfails to service the HW watchdog and the HW watchdog fires long before the watchdog soft timeout.At the moment, the
watchdog_workeris invoked as a delayed work. Delayed works are handled by non realtime kernel threads. TheWQ_HIGHPRIflag only increases the niceness of that threads.This patch replaces the delayed work logic by kthread delayed work, and sets the associated kernel task to
SCHED_FIFOwith the highest priority, in order to ensure that the watchdog worker will run as soon as possible.
1ff688209e2ed23f699269b9733993e2ce123fd2
Author: Christophe Leroy <[email protected]>
AuthorDate: Thu Jan 18 12:11:21 2018 +0100
Commit: Wim Van Sebroeck <[email protected]>
CommitDate: Sun Jan 21 12:44:59 2018 +0100Follows: v4.15-rc3 (349)
Precedes: v4.16-rc1 (13993)watchdog: core: make sure the
watchdog_workeris not deferred
Quoted 7 lines
commit
4cd13c21b207e("softirq: Let ksoftirqd do its job") has the effect of deferring timer handling in case of high CPU load, hence delaying the delayed work allthought the worker is running which high realtime priority.As hrtimers are not managed by softirqs, this patch replaces the delayed work by a plain work and uses an hrtimer to schedule that work.
These above two commits are trying very hard to ensure timely wakeup and execution of the watchdogd thread. First by moving moving to kthread delayed work, and secondly to vanilla kthread work + hardirq.
This is sufficient on mainline, because hardirq expiry fns are unconditionally executed in hardirq context. With PREEMPT_RT_FULL, however, the hrtimer expiry functions are executed in softirq context unless explicitly opted out.
...meaning that w/ PREEMPT_RT_FULL the expiry (and therefore the watchdogd wakeup) may be indefinitely starved if there are runnable RT tasks of higher priority than the softirq callback thread (ktimersoftd @ SCHED_FIFO 1 by default).
This is an inversion.
One possible solution is to opt-out of the hrtimer softirq deferral by making use of the HRTIMER_MODE_HARD, however, the expiry function will need to be vetted for use in hardirq context w/ PREEMPT_RT_FULL. From a cursory glance at the kthread_worker locking, it is not hardirq safe. :-\
Julia
Guenter Roeck
maintainerQuoted 46 lines · Steffen Trumtrar
Hi all!
I'm seeing an issue with the
dw_wdtwatchdog on the SoCFPGA ARM platform with the latest linux-rt v4.18.5-rt4. Actually I seem to have the same problem, that these patches try to fix:38a1222ae4f364d5bd5221fe305dbb0889f45d15 Author: Christophe Leroy <[email protected]>
AuthorDate: Fri Dec 8 11:18:35 2017 +0100 Commit: Wim Van Sebroeck <[email protected]>
CommitDate: Thu Dec 28 20:45:57 2017 +0100Follows: v4.15-rc3 (345)
Precedes: v4.16-rc1 (13997)watchdog: core: make sure the watchdog worker always works
When running a command like 'chrt -f 50 dd if=/dev/zero of=/dev/null', the
watchdog_workerfails to service the HW watchdog and the HW watchdog fires long before the watchdog soft timeout.At the moment, the
watchdog_workeris invoked as a delayed work. Delayed works are handled by non realtime kernel threads. TheWQ_HIGHPRIflag only increases the niceness of that threads.This patch replaces the delayed work logic by kthread delayed work, and sets the associated kernel task to
SCHED_FIFOwith the highest priority, in order to ensure that the watchdog worker will run as soon as possible.1ff688209e2ed23f699269b9733993e2ce123fd2 Author: Christophe Leroy <[email protected]>
AuthorDate: Thu Jan 18 12:11:21 2018 +0100 Commit: Wim Van Sebroeck <[email protected]>
CommitDate: Sun Jan 21 12:44:59 2018 +0100Follows: v4.15-rc3 (349)
Precedes: v4.16-rc1 (13993)watchdog: core: make sure the
watchdog_workeris not deferredcommit
4cd13c21b207e("softirq: Let ksoftirqd do its job") has the effect of deferring timer handling in case of high CPU load, hence delaying the delayed work allthought the worker is running which high realtime priority.
Quoted 8 lines
As hrtimers are not managed by softirqs, this patch replaces the delayed work by a plain work and uses an hrtimer to schedule that work.
If I run the same test or 'chrt 50 hackbench 20 -l 150' or any task where I change the prio with chrt and that runs long enough, I get a system reset from the watchdog because it times out. This only happens if the watchdog is already enabled on boot and
CONFIG_PREEMPT_RT_FULLis set.Any idea if I'm missing something essential? If I understand it correctly, the two commits fix the framework and therefore the
dw_wdtdriver doesn't need any updates.
I find your e-mail confusing, sorry. The subject says that the watchdog is not triggering, the description says that it is triggering when it should not.
You also provide no information if the watchdog is active (open from user space) or not. There is some indication in "This only happens if the watchdog is already enabled on boot" but that isn't really precise - it may be enabled on boot but still open. On top of that, your e-mail suggests that the problem may be a regression, since you refer to a specific kernel release, yet you provide no information if the very same test worked with a different kernel version, or what that kernel version would be.
Please not only describe what you are doing, but also provide the complete context. Specifically,
- Did this ever work ? If yes, what are working kernel versions ?
- Is the watchdog device open ?
- Does it make a difference if it is ?
- What is the configured watchdog timeout (both from BIOS/ROMMON and in Linux) ?
Thanks,
Guenter
Quoted 51 lines · Guenter Roeck
Quoted 50 lines · Steffen Trumtrar
Hi all!
I'm seeing an issue with the
dw_wdtwatchdog on the SoCFPGA ARM platform with the latest linux-rt v4.18.5-rt4. Actually I seem to have the same problem, that these patches try to fix:38a1222ae4f364d5bd5221fe305dbb0889f45d15 Author: Christophe Leroy <[email protected]>
AuthorDate: Fri Dec 8 11:18:35 2017 +0100 Commit: Wim Van Sebroeck <[email protected]>
CommitDate: Thu Dec 28 20:45:57 2017 +0100Follows: v4.15-rc3 (345)
Precedes: v4.16-rc1 (13997)watchdog: core: make sure the watchdog worker always works
When running a command like 'chrt -f 50 dd if=/dev/zero of=/dev/null', the
watchdog_workerfails to service the HW watchdog and the HW watchdog fires long before the watchdog soft timeout.At the moment, the
watchdog_workeris invoked as a delayed work. Delayed works are handled by non realtime kernel threads. TheWQ_HIGHPRIflag only increases the niceness of that threads.This patch replaces the delayed work logic by kthread delayed work, and sets the associated kernel task to
SCHED_FIFOwith the highest priority, in order to ensure that the watchdog worker will run as soon as possible.1ff688209e2ed23f699269b9733993e2ce123fd2 Author: Christophe Leroy <[email protected]>
AuthorDate: Thu Jan 18 12:11:21 2018 +0100 Commit: Wim Van Sebroeck <[email protected]>
CommitDate: Sun Jan 21 12:44:59 2018 +0100Follows: v4.15-rc3 (349)
Precedes: v4.16-rc1 (13993)watchdog: core: make sure the
watchdog_workeris not deferredcommit
4cd13c21b207e("softirq: Let ksoftirqd do its job") has the effect of deferring timer handling in case of high CPU load, hence delaying the delayed work allthought the worker is running which high realtime priority.As hrtimers are not managed by softirqs, this patch replaces the delayed work by a plain work and uses an hrtimer to schedule that work.
Quoted 8 lines
Quoted 4 lines
If I run the same test or 'chrt 50 hackbench 20 -l 150' or any task where I change the prio with chrt and that runs long enough, I get a system reset from the watchdog because it times out. This only happens if the watchdog is already enabled on boot and
CONFIG_PREEMPT_RT_FULLis set.Any idea if I'm missing something essential? If I understand it correctly, the two commits fix the framework and therefore the
dw_wdtdriver doesn't need any updates.I find your e-mail confusing, sorry. The subject says that the watchdog is not triggering, the description says that it is triggering when it should not.
Sorry. Let me try again.
The problem I observe, is that the watchdog is trigged, because it doesn't get pinged. The ksoftirqd seems to be blocked although it runs at a much higher priority than the blocking userspace task.
Quoted 11 lines
You also provide no information if the watchdog is active (open from user space) or not. There is some indication in "This only happens if the watchdog is already enabled on boot" but that isn't really precise - it may be enabled on boot but still open. On top of that, your e-mail suggests that the problem may be a regression, since you refer to a specific kernel release, yet you provide no information if the very same test worked with a different kernel version, or what that kernel version would be.
Please not only describe what you are doing, but also provide the complete context. Specifically,
- Did this ever work ? If yes, what are working kernel versions ?
I don't know, if it ever worked or not. This is the first kernel version I tried. According to the two commits mentioned, I assume that it won't work in older versions.
Quoted 2 lines
- Is the watchdog device open ?
- Does it make a difference if it is ?
In my test case, the device is not open. It gets started by the bootloader and than is running. I tried opening the device after it was already running, but it does not make a difference. If the watchdog is put into running state by opening it from userspace, the bug does not occur. If the bootloader starts it and the kernel just continues pinging the watchdog, it does occur, open or not.
Quoted 1 line
- What is the configured watchdog timeout (both from BIOS/ROMMON and in Linux) ?
The watchdog is configured to a timeout of 5 seconds, both in bootloader and kernel. The dw_wdt driver will round it to the nearest value it supports. Higher values do not make the bug go away.
Thanks,
Steffen
Signature · 4 lines
Pengutronix e.K. | | Industrial Linux Solutions | http://www.pengutronix.de/ | Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
Guenter Roeck
maintainerQuoted 57 lines · Steffen Trumtrar
Quoted 56 lines · Guenter Roeck
Quoted 54 lines · Steffen Trumtrar
Hi all!
I'm seeing an issue with the
dw_wdtwatchdog on the SoCFPGA ARM platform with the latest linux-rt v4.18.5-rt4. Actually I seem to have the same problem, that these patches try to fix:38a1222ae4f364d5bd5221fe305dbb0889f45d15 Author: Christophe Leroy <[email protected]>
AuthorDate: Fri Dec 8 11:18:35 2017 +0100 Commit: Wim Van Sebroeck <[email protected]>
CommitDate: Thu Dec 28 20:45:57 2017 +0100Follows: v4.15-rc3 (345)
Precedes: v4.16-rc1 (13997)watchdog: core: make sure the watchdog worker always works
When running a command like 'chrt -f 50 dd if=/dev/zero of=/dev/null', the
watchdog_workerfails to service the HW watchdog and the HW watchdog fires long before the watchdog soft timeout.At the moment, the
watchdog_workeris invoked as a delayed work. Delayed works are handled by non realtime kernel threads. TheWQ_HIGHPRIflag only increases the niceness of that threads.This patch replaces the delayed work logic by kthread delayed work, and sets the associated kernel task to
SCHED_FIFOwith the highest priority, in order to ensure that the watchdog worker will run as soon as possible.1ff688209e2ed23f699269b9733993e2ce123fd2 Author: Christophe Leroy <[email protected]>
AuthorDate: Thu Jan 18 12:11:21 2018 +0100 Commit: Wim Van Sebroeck <[email protected]>
CommitDate: Sun Jan 21 12:44:59 2018 +0100Follows: v4.15-rc3 (349)
Precedes: v4.16-rc1 (13993)watchdog: core: make sure the
watchdog_workeris not deferredcommit
4cd13c21b207e("softirq: Let ksoftirqd do its job") has the effect of deferring timer handling in case of high CPU load, hence delaying the delayed work allthought the worker is running which high realtime priority.As hrtimers are not managed by softirqs, this patch replaces the delayed work by a plain work and uses an hrtimer to schedule that work.
If I run the same test or 'chrt 50 hackbench 20 -l 150' or any task where I change the prio with chrt and that runs long enough, I get a system reset from the watchdog because it times out. This only happens if the watchdog is already enabled on boot and
CONFIG_PREEMPT_RT_FULLis set.Any idea if I'm missing something essential? If I understand it correctly, the two commits fix the framework and therefore the
dw_wdtdriver doesn't need any updates.
Quoted 9 lines
Quoted 3 lines
I find your e-mail confusing, sorry. The subject says that the watchdog is not triggering, the description says that it is triggering when it should not.
Sorry. Let me try again.
The problem I observe, is that the watchdog is trigged, because it doesn't get pinged. The ksoftirqd seems to be blocked although it runs at a much higher priority than the blocking userspace task.
Are you sure about that ? The other email seemed to suggest that the userspace task is running at higher priority.
Quoted 15 lines
Quoted 11 lines
You also provide no information if the watchdog is active (open from user space) or not. There is some indication in "This only happens if the watchdog is already enabled on boot" but that isn't really precise - it may be enabled on boot but still open. On top of that, your e-mail suggests that the problem may be a regression, since you refer to a specific kernel release, yet you provide no information if the very same test worked with a different kernel version, or what that kernel version would be.
Please not only describe what you are doing, but also provide the complete context. Specifically,
- Did this ever work ? If yes, what are working kernel versions ?I don't know, if it ever worked or not. This is the first kernel version I tried. According to the two commits mentioned, I assume that it won't work in older versions.
Quoted 9 lines
Quoted 2 lines
- Is the watchdog device open ?
- Does it make a difference if it is ?In my test case, the device is not open. It gets started by the bootloader and than is running. I tried opening the device after it was already running, but it does not make a difference. If the watchdog is put into running state by opening it from userspace, the bug does not occur. If the bootloader starts it and the kernel just continues pinging the watchdog, it does occur, open or not.
The big question here is if the watchdog daemon that keeps the watchdog open is running at higher priority than the userspace task taking all available CPU time.
If the watchdog daemon runs at lower priority, the observed behavior would be as expected.
Overall, we have a number possibilities to consider:
- The kernel watchdog timer thread is not triggered at all under some circumstances, meaning it is not set properly. So far we have no real indication that this is the case (since the code works fine unless some userspace task takes all available CPU time).
- The watchdog device is closed. The kernel watchdog timer thread is starved and does not get to run. The question is what to do in this situation. In a real time system, this is almost always a fatal condition. Should the system really be kept alive in this situation ?
- The watchdog device is open.
- The watchdog daemon runs at higher priority than the process taking all CPU time. Everything should work as expected.
- The watchdog daemon runs at the same or at a lower priority than the process taking all CPU time. I would argue that, in this case, everything is also working as expected: The watchdog daemon is starved, does not get to run, and the system resets because, per its configuration, it is in bad shape.
Overall, the only real possible problem would be if the watchdog thread in the kernel does not run because of some bug, and that it is not really starved. We would probably have to instrument the kernel to find out if this is the case, unless someone has a better idea.
Am I missing something ?
Thanks,
Guenter
Hello all-
Quoted 2 lines · Guenter Roeck
Quoted 1 line · Steffen Trumtrar
On Tue, Sep 18, 2018 at 06:46:15AM -0700, Guenter Roeck wrote:
[..]
Quoted 6 lines
Quoted 4 lines
The problem I observe, is that the watchdog is trigged, because it doesn't get pinged. The ksoftirqd seems to be blocked although it runs at a much higher priority than the blocking userspace task.
Are you sure about that ? The other email seemed to suggest that the userspace task is running at higher priority.
Also: ksoftirqd is irrelevant on RT for the kernel watchdog thread. The relevant thread is ktimersoftd, which is the thread responsible for invoking hrtimer expiry functions, like what's being used for watchdogd.
[..]
Quoted 6 lines
Overall, we have a number possibilities to consider:
- The kernel watchdog timer thread is not triggered at all under some circumstances, meaning it is not set properly. So far we have no real indication that this is the case (since the code works fine unless some userspace task takes all available CPU time).
What do you mean by "not triggered". Do you mean woken-up/activated from a scheduling perspective? In the case I identified in my other email, the watchdogd thread wakeup doesn't even occur, even when the periodic ping timer expires, because ktimersoftd has been starved.
I suspect that's what's going on for Steffen, but am not yet sure.
Quoted 4 lines
- The watchdog device is closed. The kernel watchdog timer thread is starved and does not get to run. The question is what to do in this situation. In a real time system, this is almost always a fatal condition. Should the system really be kept alive in this situation ?
Sometimes its the right decision, sometimes its not. The only sensible thing to do is to allow the user make the decision that's right for their application needs by allowing the relative prioritization of watchdogd and their application threads.
...which they can do now, but it's not effective on RT because of the timer deferral through ktimersoftd.
The solution, in my mind, and like I mentioned in my other email, is to opt-out of the ktimersoftd-deferral mechanism. This requires some tweaking with the kthread_worker bits to ensure safety in hardirq context, but that seems straightforward. See the below.
Julia
-- 8< --
drivers/watchdog/watchdog_dev.c +1 -1
@@ -945,7 +945,7 @@ static int watchdog_cdev_register(struct watchdog_device *wdd, dev_t devno)return -ENODEV ;kthread_init_work (&wd_data->work, watchdog_ping_work); -hrtimer_init (&wd_data->timer,CLOCK_MONOTONIC ,HRTIMER_MODE_REL ); +hrtimer_init (&wd_data->timer,CLOCK_MONOTONIC ,HRTIMER_MODE_REL_HARD ); wd_data->timer.function = watchdog_timer_expired;if (wdd->id ==0 ) {
include/linux/kthread.h +1 -1
@@ -85,7 +85,7 @@ enum {struct kthread_worker {unsigned int flags; -spinlock_t lock; +raw_spinlock_t lock;struct list_head work_list;struct list_head delayed_work_list;struct task_struct *task;
kernel/kthread.c +21 -21
146 lines not on this page · raw message
Guenter Roeck
maintainerQuoted 25 lines · Julia Cartwright
Hello all-
Quoted 2 lines · Guenter Roeck
Quoted 1 line · Steffen Trumtrar
On Tue, Sep 18, 2018 at 06:46:15AM -0700, Guenter Roeck wrote:
[..]
Quoted 6 lines
Quoted 4 lines
The problem I observe, is that the watchdog is trigged, because it doesn't get pinged. The ksoftirqd seems to be blocked although it runs at a much higher priority than the blocking userspace task.
Are you sure about that ? The other email seemed to suggest that the userspace task is running at higher priority.
Also: ksoftirqd is irrelevant on RT for the kernel watchdog thread. The relevant thread is ktimersoftd, which is the thread responsible for invoking hrtimer expiry functions, like what's being used for watchdogd.
[..]
Quoted 6 lines
Overall, we have a number possibilities to consider:
- The kernel watchdog timer thread is not triggered at all under some circumstances, meaning it is not set properly. So far we have no real indication that this is the case (since the code works fine unless some userspace task takes all available CPU time).
Quoted 5 lines
What do you mean by "not triggered". Do you mean woken-up/activated from a scheduling perspective? In the case I identified in my other email, the watchdogd thread wakeup doesn't even occur, even when the periodic ping timer expires, because ktimersoftd has been starved.
Sorry for not using the correct term. Sometimes I am a bit sloppy. Yes, I meant "woken-up/activated from a scheduling perspective".
Quoted 7 lines
I suspect that's what's going on for Steffen, but am not yet sure.
Quoted 4 lines
- The watchdog device is closed. The kernel watchdog timer thread is starved and does not get to run. The question is what to do in this situation. In a real time system, this is almost always a fatal condition. Should the system really be kept alive in this situation ?
Quoted 5 lines
Sometimes its the right decision, sometimes its not. The only sensible thing to do is to allow the user make the decision that's right for their application needs by allowing the relative prioritization of watchdogd and their application threads.
Agreed, but that doesn't help if the watchdog daemon is not open or if the hardware watchdog interval is too small and the kernel mechanism is needed to ping the watchdog.
Quoted 8 lines
...which they can do now, but it's not effective on RT because of the timer deferral through ktimersoftd.
The solution, in my mind, and like I mentioned in my other email, is to opt-out of the ktimersoftd-deferral mechanism. This requires some tweaking with the
kthread_workerbits to ensure safety in hardirq context, but that seems straightforward. See the below.
Makes sense to me, though I have no idea what it would take to push the necessary changes into the core kernel.
However, I must be missing something: Looking into the kernel code, it seems to me that the spin_lock functions call the respective raw_ spinlock functions right away. With that in mind, why would the kernel code change be necessary ? Also, I don't see HRTIMER_MODE_REL_HARD defined anywhere. Is this RT specific ?
Thanks,
Guenter
Quoted 193 lines
Julia
-- 8< --
drivers/watchdog/watchdog_dev.c +1
-1@@ -945,7 +945,7 @@ static int watchdog_cdev_register(struct watchdog_device *wdd, dev_t devno)return -ENODEV ;kthread_init_work (&wd_data->work, watchdog_ping_work); -hrtimer_init (&wd_data->timer,CLOCK_MONOTONIC ,HRTIMER_MODE_REL ); +hrtimer_init (&wd_data->timer,CLOCK_MONOTONIC ,HRTIMER_MODE_REL_HARD ); wd_data->timer.function = watchdog_timer_expired;if (wdd->id ==0 ) {include/linux/kthread.h +1
-1@@ -85,7 +85,7 @@ enum {struct kthread_worker {unsigned int flags; -spinlock_t lock; +raw_spinlock_t lock;struct list_head work_list;struct list_head delayed_work_list;struct task_struct *task;kernel/kthread.c +21
-21@@ -597,7 +597,7 @@ void __kthread_init_worker(struct kthread_worker *worker, struct lock_class_key *key) {memset (worker,0 ,sizeof (struct kthread_worker )); -spin_lock_init (&worker->lock); +raw_spin_lock_init (&worker->lock);lockdep_set_class_and_name (&worker->lock, key, name);INIT_LIST_HEAD (&worker->work_list);INIT_LIST_HEAD (&worker->delayed_work_list);@@ -639,21 +639,21 @@ int kthread_worker_fn(void *worker_ptr)if (kthread_should_stop ()) {__set_current_state (TASK_RUNNING ); -spin_lock_irq (&worker->lock); +raw_spin_lock_irq (&worker->lock); worker->task =NULL ; -spin_unlock_irq (&worker->lock); +raw_spin_unlock_irq (&worker->lock);return 0 ; } work =NULL ; -spin_lock_irq (&worker->lock); +raw_spin_lock_irq (&worker->lock);if (!list_empty (&worker->work_list)) { work =list_first_entry (&worker->work_list, struct kthread_work, node);list_del_init (&work->node); } worker->current_work = work; -spin_unlock_irq (&worker->lock); +raw_spin_unlock_irq (&worker->lock);if (work) {__set_current_state (TASK_RUNNING );@@ -810,12 +810,12 @@ bool kthread_queue_work(struct kthread_worker *worker,bool ret = false;unsigned long flags; -spin_lock_irqsave (&worker->lock, flags); +raw_spin_lock_irqsave (&worker->lock, flags);if (!queuing_blocked (worker, work)) {kthread_insert_work (worker, work, &worker->work_list); ret = true; } -spin_unlock_irqrestore (&worker->lock, flags); +raw_spin_unlock_irqrestore (&worker->lock, flags);return ret; }EXPORT_SYMBOL_GPL (kthread_queue_work);@@ -841,7 +841,7 @@ void kthread_delayed_work_timer_fn(struct timer_list *t)if (WARN_ON_ONCE (!worker))return ; -spin_lock (&worker->lock); +raw_spin_lock (&worker->lock);/* Work must not be used with >1 worker, see kthread_queue_work(). */ WARN_ON_ONCE (work->worker != worker);@@ -850,7 +850,7 @@ void kthread_delayed_work_timer_fn(struct timer_list *t)list_del_init (&work->node);kthread_insert_work (worker, work, &worker->work_list); -spin_unlock (&worker->lock); +raw_spin_unlock (&worker->lock); }EXPORT_SYMBOL (kthread_delayed_work_timer_fn);@@ -906,14 +906,14 @@ bool kthread_queue_delayed_work(struct kthread_worker *worker,unsigned long flags;bool ret = false; -spin_lock_irqsave (&worker->lock, flags); +raw_spin_lock_irqsave (&worker->lock, flags);if (!queuing_blocked (worker, work)) {__kthread_queue_delayed_work (worker, dwork, delay); ret = true; } -spin_unlock_irqrestore (&worker->lock, flags); +raw_spin_unlock_irqrestore (&worker->lock, flags);return ret; }EXPORT_SYMBOL_GPL (kthread_queue_delayed_work);@@ -949,7 +949,7 @@ void kthread_flush_work(struct kthread_work *work)if (!worker)return ; -spin_lock_irq (&worker->lock); +raw_spin_lock_irq (&worker->lock);/* Work must not be used with >1 worker, see kthread_queue_work(). */ WARN_ON_ONCE (work->worker != worker);@@ -961,7 +961,7 @@ void kthread_flush_work(struct kthread_work *work)else noop = true; -spin_unlock_irq (&worker->lock); +raw_spin_unlock_irq (&worker->lock);if (!noop)wait_for_completion (&fwork.done);@@ -994,9 +994,9 @@ static bool __kthread_cancel_work(struct kthread_work *work, bool is_dwork,* any queuing is blocked by setting the canceling counter. */ work->canceling++; -spin_unlock_irqrestore (&worker->lock, *flags); +raw_spin_unlock_irqrestore (&worker->lock, *flags);del_timer_sync (&dwork->timer); -spin_lock_irqsave (&worker->lock, *flags); +raw_spin_lock_irqsave (&worker->lock, *flags); work->canceling--; }@@ -1043,7 +1043,7 @@ bool kthread_mod_delayed_work(struct kthread_worker *worker,unsigned long flags;int ret = false; -spin_lock_irqsave (&worker->lock, flags); +raw_spin_lock_irqsave (&worker->lock, flags);/* Do not bother with canceling when never queued. */ if (!work->worker)@@ -1060,7 +1060,7 @@ bool kthread_mod_delayed_work(struct kthread_worker *worker, fast_queue:__kthread_queue_delayed_work (worker, dwork, delay); out: -spin_unlock_irqrestore (&worker->lock, flags); +raw_spin_unlock_irqrestore (&worker->lock, flags);return ret; }EXPORT_SYMBOL_GPL (kthread_mod_delayed_work);@@ -1074,7 +1074,7 @@ static bool __kthread_cancel_work_sync(struct kthread_work *work, bool is_dwork)if (!worker) goto out; -spin_lock_irqsave (&worker->lock, flags); +raw_spin_lock_irqsave (&worker->lock, flags);/* Work must not be used with >1 worker, see kthread_queue_work(). */ WARN_ON_ONCE (work->worker != worker);@@ -1088,13 +1088,13 @@ static bool __kthread_cancel_work_sync(struct kthread_work *work, bool is_dwork)* In the meantime, block any queuing by setting the canceling counter. */ work->canceling++; -spin_unlock_irqrestore (&worker->lock, flags); +raw_spin_unlock_irqrestore (&worker->lock, flags);kthread_flush_work (work); -spin_lock_irqsave (&worker->lock, flags); +raw_spin_lock_irqsave (&worker->lock, flags); work->canceling--; out_fast: -spin_unlock_irqrestore (&worker->lock, flags); +raw_spin_unlock_irqrestore (&worker->lock, flags); out:return ret; }
Quoted 2 lines · Guenter Roeck
Quoted 1 line · Julia Cartwright
On Wed, Sep 19, 2018 at 12:43:03PM -0700, Guenter Roeck wrote:
[..]
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
Overall, we have a number possibilities to consider:
- The kernel watchdog timer thread is not triggered at all under some circumstances, meaning it is not set properly. So far we have no real indication that this is the case (since the code works fine unless some userspace task takes all available CPU time).
Quoted 8 lines
Quoted 5 lines
What do you mean by "not triggered". Do you mean woken-up/activated from a scheduling perspective? In the case I identified in my other email, the watchdogd thread wakeup doesn't even occur, even when the periodic ping timer expires, because ktimersoftd has been starved.
Sorry for not using the correct term. Sometimes I am a bit sloppy. Yes, I meant "woken-up/activated from a scheduling perspective".
Thanks for the clarification. I think we're on the same page. :)
Quoted 7 lines
Quoted 7 lines
I suspect that's what's going on for Steffen, but am not yet sure.
Quoted 4 lines
- The watchdog device is closed. The kernel watchdog timer thread is starved and does not get to run. The question is what to do in this situation. In a real time system, this is almost always a fatal condition. Should the system really be kept alive in this situation ?
Quoted 8 lines
Quoted 4 lines
Sometimes its the right decision, sometimes its not. The only sensible thing to do is to allow the user make the decision that's right for their application needs by allowing the relative prioritization of watchdogd and their application threads.
Agreed, but that doesn't help if the watchdog daemon is not open or if the hardware watchdog interval is too small and the kernel mechanism is needed to ping the watchdog.
Makes sense.
Quoted 10 lines
Quoted 7 lines
...which they can do now, but it's not effective on RT because of the timer deferral through ktimersoftd.
The solution, in my mind, and like I mentioned in my other email, is to opt-out of the ktimersoftd-deferral mechanism. This requires some tweaking with the
kthread_workerbits to ensure safety in hardirq context, but that seems straightforward. See the below.Makes sense to me, though I have no idea what it would take to push the necessary changes into the core kernel.
As of now, this bug doesn't exist in mainline because the hrtimer deferral bits haven't landed yet, as you note below.
Quoted 5 lines
However, I must be missing something: Looking into the kernel code, it seems to me that the
spin_lockfunctions call the respectiveraw_spinlock functions right away. With that in mind, why would the kernel code change be necessary ? Also, I don't seeHRTIMER_MODE_REL_HARDdefined anywhere. Is this RT specific ?
Yes, there is no functional difference in mainline currently between a spin_lock_t and a raw_spin_lock_t. There is also no HRTIMER_MODE_REL_HARD like mentioned before. These are features/concepts currently only in the RT tree, but should be making their way into mainline soon.
As far as path forward, I'd like to get some confirmation from Steffen and/or Tim that the proposed patch fixes their issue, then I'll cook some proper patches; the kthread_worker bits could go mainline now because there is no dependency, but the watchdog change will need to be RT-only for now.
Julia
Guenter Roeck
maintainerQuoted 66 lines · Julia Cartwright
Quoted 2 lines · Guenter Roeck
Quoted 1 line · Julia Cartwright
On Wed, Sep 19, 2018 at 12:43:03PM -0700, Guenter Roeck wrote:
[..]
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
Overall, we have a number possibilities to consider:
- The kernel watchdog timer thread is not triggered at all under some circumstances, meaning it is not set properly. So far we have no real indication that this is the case (since the code works fine unless some userspace task takes all available CPU time).
Quoted 8 lines
Quoted 5 lines
What do you mean by "not triggered". Do you mean woken-up/activated from a scheduling perspective? In the case I identified in my other email, the watchdogd thread wakeup doesn't even occur, even when the periodic ping timer expires, because ktimersoftd has been starved.
Sorry for not using the correct term. Sometimes I am a bit sloppy. Yes, I meant "woken-up/activated from a scheduling perspective".
Thanks for the clarification. I think we're on the same page. :)
Quoted 7 lines
Quoted 7 lines
I suspect that's what's going on for Steffen, but am not yet sure.
Quoted 4 lines
- The watchdog device is closed. The kernel watchdog timer thread is starved and does not get to run. The question is what to do in this situation. In a real time system, this is almost always a fatal condition. Should the system really be kept alive in this situation ?
Quoted 8 lines
Quoted 4 lines
Sometimes its the right decision, sometimes its not. The only sensible thing to do is to allow the user make the decision that's right for their application needs by allowing the relative prioritization of watchdogd and their application threads.
Agreed, but that doesn't help if the watchdog daemon is not open or if the hardware watchdog interval is too small and the kernel mechanism is needed to ping the watchdog.
Makes sense.
Quoted 10 lines
Quoted 7 lines
...which they can do now, but it's not effective on RT because of the timer deferral through ktimersoftd.
The solution, in my mind, and like I mentioned in my other email, is to opt-out of the ktimersoftd-deferral mechanism. This requires some tweaking with the
kthread_workerbits to ensure safety in hardirq context, but that seems straightforward. See the below.Makes sense to me, though I have no idea what it would take to push the necessary changes into the core kernel.
As of now, this bug doesn't exist in mainline because the hrtimer deferral bits haven't landed yet, as you note below.
Quoted 5 lines
However, I must be missing something: Looking into the kernel code, it seems to me that the
spin_lockfunctions call the respectiveraw_spinlock functions right away. With that in mind, why would the kernel code change be necessary ? Also, I don't seeHRTIMER_MODE_REL_HARDdefined anywhere. Is this RT specific ?Yes, there is no functional difference in mainline currently between a
spin_lock_tand araw_spin_lock_t. There is also noHRTIMER_MODE_REL_HARDlike mentioned before. These are features/concepts currently only in the RT tree, but should be making their way into mainline soon.
Quoted 6 lines
As far as path forward, I'd like to get some confirmation from Steffen and/or Tim that the proposed patch fixes their issue, then I'll cook some proper patches; the
kthread_workerbits could go mainline now because there is no dependency, but the watchdog change will need to be RT-only for now.
SGTM.
Thanks,
Guenter
Hi!
Quoted 49 lines · Julia Cartwright
Hello all-
Quoted 2 lines · Guenter Roeck
Quoted 1 line · Steffen Trumtrar
On Tue, Sep 18, 2018 at 06:46:15AM -0700, Guenter Roeck wrote:
[..]
Quoted 6 lines
Quoted 4 lines
The problem I observe, is that the watchdog is trigged, because it doesn't get pinged. The ksoftirqd seems to be blocked although it runs at a much higher priority than the blocking userspace task.
Are you sure about that ? The other email seemed to suggest that the userspace task is running at higher priority.
Also: ksoftirqd is irrelevant on RT for the kernel watchdog thread. The relevant thread is ktimersoftd, which is the thread responsible for invoking hrtimer expiry functions, like what's being used for watchdogd.
[..]
Quoted 9 lines
Overall, we have a number possibilities to consider:
- The kernel watchdog timer thread is not triggered at all under some
circumstances, meaning it is not set properly. So far we have no real
indication that this is the case (since the code works fine unless some
userspace task takes all available CPU time).What do you mean by "not triggered". Do you mean woken-up/activated from a scheduling perspective? In the case I identified in my other email, the watchdogd thread wakeup doesn't even occur, even when the periodic ping timer expires, because ktimersoftd has been starved.
I suspect that's what's going on for Steffen, but am not yet sure.
Quoted 8 lines
- The watchdog device is closed. The kernel watchdog timer thread is
starved and does not get to run. The question is what to do in this
situation. In a real time system, this is almost always a fatal
condition. Should the system really be kept alive in this situation ?Sometimes its the right decision, sometimes its not. The only sensible thing to do is to allow the user make the decision that's right for their application needs by allowing the relative prioritization of watchdogd and their application threads.
Quoted 8 lines
...which they can do now, but it's not effective on RT because of the timer deferral through ktimersoftd.
The solution, in my mind, and like I mentioned in my other email, is to opt-out of the ktimersoftd-deferral mechanism. This requires some tweaking with the
kthread_workerbits to ensure safety in hardirq context, but that seems straightforward. See the below.
I just tested your patch and it works for me \o/
Thanks,
Steffen
Signature · 4 lines
Pengutronix e.K. | Steffen Trumtrar | Industrial Linux Solutions | http://www.pengutronix.de/ | Peiner Str. 6-8, 31137 Hildesheim, Germany| Phone: +49-5121-206917-0 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555|
cover Fix watchdogd wakeup deferral on RT
The following two patches solve an issue reported by Steffen Trumtrar and Tim Sander to the linux-rt-users mailing list. Namely, the wakeup of the watchdogd thread being starved by an RT task due to the hrtimer deferral through ktimersoftd (on PREEMPT_RT_FULL).
The first patch adjusts the kthread_worker locking to make use of a raw spinlock, making it suitable for work item queueing from hardirq context on PREEMPT_RT. This patch can be applied directly to mainline without any functional change.
The second patch adjusts the hrtimer used by the watchdog core to be a _HARD timer (and thus not deferred through ktimersoftd w/ PREEMPT_RT). This patch depends on hrtimer patches carried in the RT patch, and so should therefore land there.
Julia Cartwright (2):
kthread: convert worker lock to raw spinlock watchdog, rt: prevent deferral of watchdogd wakeup
drivers/watchdog/watchdog_dev.c | 2 +- include/linux/kthread.h | 2 +- kernel/kthread.c | 42 ++++++++++++++++----------------- 3 files changed, 23 insertions(+), 23 deletions(-)
Signature · 1 line
2.18.0
PATCH 1/2 kthread: convert worker lock to raw spinlock
In order to enable the queuing of kthread work items from hardirq context even when PREEMPT_RT_FULL is enabled, convert the worker spin_lock to a raw_spin_lock.
This is only acceptable to do because the work performed under the lock is well-bounded and minimal.
--- include/linux/kthread.h | 2 +- kernel/kthread.c | 42 ++++++++++++++++++++--------------------- 2 files changed, 22 insertions(+), 22 deletions(-)
include/linux/kthread.h +1 -1
@@ -85,7 +85,7 @@ enum {struct kthread_worker {unsigned int flags; -spinlock_t lock; +raw_spinlock_t lock;struct list_head work_list;struct list_head delayed_work_list;struct task_struct *task;
kernel/kthread.c +21 -21
146 lines not on this page · raw message
Signature · 1 line
2.18.0
Re: [PATCH 1/2] kthread: convert worker lock to raw spinlock
Quoted 4 lines · Julia Cartwright
In order to enable the queuing of kthread work items from hardirq context even when
PREEMPT_RT_FULLis enabled, convert the workerspin_lockto araw_spin_lock.
Quoted 8 lines
This is only acceptable to do because the work performed under the lock is well-bounded and minimal.
Cc Sebastian Andrzej Siewior <[email protected]>Cc Guenter Roeck <[email protected]>Reported-and-tested-by Steffen Trumtrar <[email protected]>Reported-by Tim Sander <[email protected]>Signed-off-by Julia Cartwright <[email protected]>
Sebastian
Re: [PATCH 1/2] kthread: convert worker lock to raw spinlock
In reply to Julia CartwrightHi Julia,
Quoted 6 lines · Julia Cartwright
In order to enable the queuing of kthread work items from hardirq context even when
PREEMPT_RT_FULLis enabled, convert the workerspin_lockto araw_spin_lock.This is only acceptable to do because the work performed under the lock is well-bounded and minimal.
Clearly not my topic..., but out of curiosity: What do you mean by "well-bounded" and "minimal"? Can you maybe point me to some doc.?
Andrea
Quoted 190 lines
Cc Sebastian Andrzej Siewior <[email protected]>Cc Guenter Roeck <[email protected]>Reported-and-tested-by Steffen Trumtrar <[email protected]>Reported-by Tim Sander <[email protected]>Signed-off-by Julia Cartwright <[email protected]>--- include/linux/kthread.h | 2 +- kernel/kthread.c | 42 ++++++++++++++++++++--------------------- 2 files changed, 22 insertions(+), 22 deletions(-)include/linux/kthread.h +1
-1@@ -85,7 +85,7 @@ enum {struct kthread_worker {unsigned int flags; -spinlock_t lock; +raw_spinlock_t lock;struct list_head work_list;struct list_head delayed_work_list;struct task_struct *task;kernel/kthread.c +21
-21@@ -597,7 +597,7 @@ void __kthread_init_worker(struct kthread_worker *worker, struct lock_class_key *key) {memset (worker,0 ,sizeof (struct kthread_worker )); -spin_lock_init (&worker->lock); +raw_spin_lock_init (&worker->lock);lockdep_set_class_and_name (&worker->lock, key, name);INIT_LIST_HEAD (&worker->work_list);INIT_LIST_HEAD (&worker->delayed_work_list);@@ -639,21 +639,21 @@ int kthread_worker_fn(void *worker_ptr)if (kthread_should_stop ()) {__set_current_state (TASK_RUNNING ); -spin_lock_irq (&worker->lock); +raw_spin_lock_irq (&worker->lock); worker->task =NULL ; -spin_unlock_irq (&worker->lock); +raw_spin_unlock_irq (&worker->lock);return 0 ; } work =NULL ; -spin_lock_irq (&worker->lock); +raw_spin_lock_irq (&worker->lock);if (!list_empty (&worker->work_list)) { work =list_first_entry (&worker->work_list, struct kthread_work, node);list_del_init (&work->node); } worker->current_work = work; -spin_unlock_irq (&worker->lock); +raw_spin_unlock_irq (&worker->lock);if (work) {__set_current_state (TASK_RUNNING );@@ -810,12 +810,12 @@ bool kthread_queue_work(struct kthread_worker *worker,bool ret = false;unsigned long flags; -spin_lock_irqsave (&worker->lock, flags); +raw_spin_lock_irqsave (&worker->lock, flags);if (!queuing_blocked (worker, work)) {kthread_insert_work (worker, work, &worker->work_list); ret = true; } -spin_unlock_irqrestore (&worker->lock, flags); +raw_spin_unlock_irqrestore (&worker->lock, flags);return ret; }EXPORT_SYMBOL_GPL (kthread_queue_work);@@ -841,7 +841,7 @@ void kthread_delayed_work_timer_fn(struct timer_list *t)if (WARN_ON_ONCE (!worker))return ; -spin_lock (&worker->lock); +raw_spin_lock (&worker->lock);/* Work must not be used with >1 worker, see kthread_queue_work(). */ WARN_ON_ONCE (work->worker != worker);@@ -850,7 +850,7 @@ void kthread_delayed_work_timer_fn(struct timer_list *t)list_del_init (&work->node);kthread_insert_work (worker, work, &worker->work_list); -spin_unlock (&worker->lock); +raw_spin_unlock (&worker->lock); }EXPORT_SYMBOL (kthread_delayed_work_timer_fn);@@ -906,14 +906,14 @@ bool kthread_queue_delayed_work(struct kthread_worker *worker,unsigned long flags;bool ret = false; -spin_lock_irqsave (&worker->lock, flags); +raw_spin_lock_irqsave (&worker->lock, flags);if (!queuing_blocked (worker, work)) {__kthread_queue_delayed_work (worker, dwork, delay); ret = true; } -spin_unlock_irqrestore (&worker->lock, flags); +raw_spin_unlock_irqrestore (&worker->lock, flags);return ret; }EXPORT_SYMBOL_GPL (kthread_queue_delayed_work);@@ -949,7 +949,7 @@ void kthread_flush_work(struct kthread_work *work)if (!worker)return ; -spin_lock_irq (&worker->lock); +raw_spin_lock_irq (&worker->lock);/* Work must not be used with >1 worker, see kthread_queue_work(). */ WARN_ON_ONCE (work->worker != worker);@@ -961,7 +961,7 @@ void kthread_flush_work(struct kthread_work *work)else noop = true; -spin_unlock_irq (&worker->lock); +raw_spin_unlock_irq (&worker->lock);if (!noop)wait_for_completion (&fwork.done);@@ -994,9 +994,9 @@ static bool __kthread_cancel_work(struct kthread_work *work, bool is_dwork,* any queuing is blocked by setting the canceling counter. */ work->canceling++; -spin_unlock_irqrestore (&worker->lock, *flags); +raw_spin_unlock_irqrestore (&worker->lock, *flags);del_timer_sync (&dwork->timer); -spin_lock_irqsave (&worker->lock, *flags); +raw_spin_lock_irqsave (&worker->lock, *flags); work->canceling--; }@@ -1043,7 +1043,7 @@ bool kthread_mod_delayed_work(struct kthread_worker *worker,unsigned long flags;int ret = false; -spin_lock_irqsave (&worker->lock, flags); +raw_spin_lock_irqsave (&worker->lock, flags);/* Do not bother with canceling when never queued. */ if (!work->worker)@@ -1060,7 +1060,7 @@ bool kthread_mod_delayed_work(struct kthread_worker *worker, fast_queue:__kthread_queue_delayed_work (worker, dwork, delay); out: -spin_unlock_irqrestore (&worker->lock, flags); +raw_spin_unlock_irqrestore (&worker->lock, flags);return ret; }EXPORT_SYMBOL_GPL (kthread_mod_delayed_work);@@ -1074,7 +1074,7 @@ static bool __kthread_cancel_work_sync(struct kthread_work *work, bool is_dwork)if (!worker) goto out; -spin_lock_irqsave (&worker->lock, flags); +raw_spin_lock_irqsave (&worker->lock, flags);/* Work must not be used with >1 worker, see kthread_queue_work(). */ WARN_ON_ONCE (work->worker != worker);@@ -1088,13 +1088,13 @@ static bool __kthread_cancel_work_sync(struct kthread_work *work, bool is_dwork)* In the meantime, block any queuing by setting the canceling counter. */ work->canceling++; -spin_unlock_irqrestore (&worker->lock, flags); +raw_spin_unlock_irqrestore (&worker->lock, flags);kthread_flush_work (work); -spin_lock_irqsave (&worker->lock, flags); +raw_spin_lock_irqsave (&worker->lock, flags); work->canceling--; out_fast: -spin_unlock_irqrestore (&worker->lock, flags); +raw_spin_unlock_irqrestore (&worker->lock, flags); out:return ret; }Signature · 1 line
2.18.0
Re: [PATCH 1/2] kthread: convert worker lock to raw spinlock
Quoted 3 lines · Andrea Parri
Clearly not my topic..., but out of curiosity: What do you mean by "well-bounded" and "minimal"? Can you maybe point me to some doc.?
it means that the lock is not held for an arbitrary amount of time like by processing a list with thousand items. Well-bounded would mean not process more than one or five (or so) at a time.
Quoted 1 line
Andrea
Sebastian
PATCH 2/2 watchdog, rt: prevent deferral of watchdogd wakeup
In reply to Julia CartwrightWhen PREEMPT_RT_FULL is enabled, all hrtimer expiry functions are deferred for execution into the context of ktimersoftd unless otherwise annotated.
Deferring the expiry of the hrtimer used by the watchdog core, however, is a waste, as the callback does nothing but queue a kthread work item and wakeup watchdogd.
It's worst then that, too: the deferral through ktimersoftd also means that for correct behavior a user must adjust the scheduling parameters of both watchdogd _and_ ktimersoftd, which is unnecessary and has other side effects (like causing unrelated expiry functions to execute at potentially elevated priority).
Instead, mark the hrtimer used by the watchdog core as being _HARD to allow it's execution directly from hardirq context. The work done in this expiry function is well-bounded and minimal.
A user still must adjust the scheduling parameters of the watchdogd to be correct w.r.t. their application needs.
--- drivers/watchdog/watchdog_dev.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/watchdog/watchdog_dev.c +1 -1
@@ -945,7 +945,7 @@ static int watchdog_cdev_register(struct watchdog_device *wdd, dev_t devno)return -ENODEV ;kthread_init_work (&wd_data->work, watchdog_ping_work); -hrtimer_init (&wd_data->timer,CLOCK_MONOTONIC ,HRTIMER_MODE_REL ); +hrtimer_init (&wd_data->timer,CLOCK_MONOTONIC ,HRTIMER_MODE_REL_HARD ); wd_data->timer.function = watchdog_timer_expired;if (wdd->id ==0 ) {
Signature · 1 line
2.18.0
kbuild test robot
build botRe: [PATCH RT 2/2] watchdog, rt: prevent deferral of watchdogd wakeup
Thank you for the patch! Yet something to improve:
Hi Julia,
Thank you for the patch! Yet something to improve:
[auto build test ERROR on linus/master]
[also build test ERROR on v4.19-rc5 next-20180928]
[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/Julia-Cartwright/kthread-convert-worker-lock-to-raw-spinlock/20180929-052522
config: i386-randconfig-x008-201838 (attached as .config)
compiler: gcc-7 (Debian 7.3.0-1) 7.3.0 reproduce:
# save the attached .config to linux build tree
make ARCH=i386 All errors (new ones prefixed by >>):
drivers//watchdog/watchdog_dev.c: In function 'watchdog_cdev_register':
>> drivers//watchdog/watchdog_dev.c:948:49: error: 'HRTIMER_MODE_REL_HARD' undeclared (first use in this function); did you mean 'HRTIMER_MODE_REL_SOFT'?
hrtimer_init(&wd_data->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
^~~~~~~~~~~~~~~~~~~~~
HRTIMER_MODE_REL_SOFT drivers//watchdog/watchdog_dev.c:948:49: note: each undeclared identifier is reported only once for each function it appears in
vim +948 drivers//watchdog/watchdog_dev.c
919
920 /*
921 * watchdog_cdev_register: register watchdog character device
922 * @wdd: watchdog device
923 * @devno: character device number
924 *
925 * Register a watchdog character device including handling the legacy
926 * /dev/watchdog node. /dev/watchdog is actually a miscdevice and
927 * thus we set it up like that.
928 */
929
930 static int watchdog_cdev_register(struct watchdog_device *wdd, dev_t devno)
931 {
932 struct watchdog_core_data *wd_data;
933 int err;
934
935 wd_data = kzalloc(sizeof(struct watchdog_core_data), GFP_KERNEL);
936 if (!wd_data)
937 return -ENOMEM;
938 kref_init(&wd_data->kref);
939 mutex_init(&wd_data->lock);
940
941 wd_data->wdd = wdd;
942 wdd->wd_data = wd_data;
943
944 if (IS_ERR_OR_NULL(watchdog_kworker))
945 return -ENODEV;
946
947 kthread_init_work(&wd_data->work, watchdog_ping_work);
> 948 hrtimer_init(&wd_data->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
949 wd_data->timer.function = watchdog_timer_expired;
950
951 if (wdd->id == 0) {
952 old_wd_data = wd_data;
953 watchdog_miscdev.parent = wdd->parent;
954 err = misc_register(&watchdog_miscdev);
955 if (err != 0) {
956 pr_err("%s: cannot register miscdev on minor=%d (err=%d).\n",
957 wdd->info->identity, WATCHDOG_MINOR, err);
958 if (err == -EBUSY)
959 pr_err("%s: a legacy watchdog module is probably present.\n",
960 wdd->info->identity);
961 old_wd_data = NULL;
962 kfree(wd_data);
963 return err;
964 }
965 }
966
967 /* Fill in the data structures */
968 cdev_init(&wd_data->cdev, &watchdog_fops);
969 wd_data->cdev.owner = wdd->ops->owner;
970
971 /* Add the device */
972 err = cdev_add(&wd_data->cdev, devno, 1);
973 if (err) {
974 pr_err("watchdog%d unable to add device %d:%d\n",
975 wdd->id, MAJOR(watchdog_devt), wdd->id);
976 if (wdd->id == 0) {
977 misc_deregister(&watchdog_miscdev);
978 old_wd_data = NULL;
979 kref_put(&wd_data->kref, watchdog_core_data_release);
980 }
981 return err;
982 }
983
984 /* Record time of most recent heartbeat as 'just before now'. */
985 wd_data->last_hw_keepalive = ktime_sub(ktime_get(), 1);
986
987 /*
988 * If the watchdog is running, prevent its driver from being unloaded,
989 * and schedule an immediate ping.
990 */
991 if (watchdog_hw_running(wdd)) {
992 __module_get(wdd->ops->owner);
993 kref_get(&wd_data->kref);
994 if (handle_boot_enabled)
995 hrtimer_start(&wd_data->timer, 0, HRTIMER_MODE_REL);
996 else
997 pr_info("watchdog%d running and kernel based pre-userspace handler disabled\n",
998 wdd->id);
999 }
1000
1001 return 0;
1002 }
1003
---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation
Re: [PATCH RT 2/2] watchdog, rt: prevent deferral of watchdogd wakeup
Quoted 3 lines · kbuild test robot
[also build test ERROR on v4.19-rc5 next-20180928]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system]
It's against the RT tree, so it won't work against next or upstream. I think it would be good to have a machine parseable tag to describe which tree/branch/commit patches are applicable to. If that tag is missing, the bot can assume it's against upstream/next.
Thanks,
tglx
Re: [PATCH RT 2/2] watchdog, rt: prevent deferral of watchdogd wakeup
Quoted 9 lines · Thomas Gleixner
Quoted 3 lines · kbuild test robot
[also build test ERROR on v4.19-rc5 next-20180928]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system]It's against the RT tree, so it won't work against next or upstream. I think it would be good to have a machine parseable tag to describe which tree/branch/commit patches are applicable to. If that tag is missing, the bot can assume it's against upstream/next.
I asked the testing team to ignore patches or test against the RT tree if the patch is tagged [PATCH RT]. I think it worked since I haven't seen those mails.
Quoted 3 lines
Thanks,
tglx
Sebastian
RE: [kbuild-all] [PATCH RT 2/2] watchdog, rt: prevent deferral of watchdogd wakeup
Quoted 10 lines
Subject: Re: [kbuild-all] [PATCH RT 2/2] watchdog, rt: prevent deferral of watchdogd wakeup
Quoted 3 lines · Thomas Gleixner
Quoted 2 lines · kbuild test robot
[also build test ERROR on v4.19-rc5 next-20180928]
[if your patch is applied to the wrong git tree, please drop us a note to helpimprove the system]
Quoted 2 lines
Quoted 1 line
Quoted 8 lines
Quoted 4 lines
It's against the RT tree, so it won't work against next or upstream. I think it would be good to have a machine parseable tag to describe which tree/branch/commit patches are applicable to. If that tag is missing, the bot can assume it's against upstream/next.
I asked the testing team to ignore patches or test against the RT tree if the patch is tagged [PATCH RT]. I think it worked since I haven't seen those mails.
thanks Sebastian for the input, we will consider this to add to our TODO.
Quoted 10 lines
Quoted 4 lines
Thanks,
tglx
Sebastian
_______________________________________________
kbuild-all mailing list
[email protected]
https://lists.01.org/mailman/listinfo/kbuild-all
kbuild test robot
build botRe: [PATCH RT 2/2] watchdog, rt: prevent deferral of watchdogd wakeup
In reply to Julia CartwrightThank you for the patch! Yet something to improve:
Hi Julia,
Thank you for the patch! Yet something to improve:
[auto build test ERROR on linus/master]
[also build test ERROR on v4.19-rc5 next-20180928]
[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/Julia-Cartwright/kthread-convert-worker-lock-to-raw-spinlock/20180929-052522
config: i386-randconfig-s0-201838 (attached as .config)
compiler: gcc-6 (Debian 6.4.0-9) 6.4.0 20171026 reproduce:
# save the attached .config to linux build tree
make ARCH=i386 All errors (new ones prefixed by >>):
drivers//watchdog/watchdog_dev.c: In function 'watchdog_cdev_register':
>> drivers//watchdog/watchdog_dev.c:948:49: error: 'HRTIMER_MODE_REL_HARD' undeclared (first use in this function)
hrtimer_init(&wd_data->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
^~~~~~~~~~~~~~~~~~~~~ drivers//watchdog/watchdog_dev.c:948:49: note: each undeclared identifier is reported only once for each function it appears in
vim +/HRTIMER_MODE_REL_HARD +948 drivers//watchdog/watchdog_dev.c
919
920 /*
921 * watchdog_cdev_register: register watchdog character device
922 * @wdd: watchdog device
923 * @devno: character device number
924 *
925 * Register a watchdog character device including handling the legacy
926 * /dev/watchdog node. /dev/watchdog is actually a miscdevice and
927 * thus we set it up like that.
928 */
929
930 static int watchdog_cdev_register(struct watchdog_device *wdd, dev_t devno)
931 {
932 struct watchdog_core_data *wd_data;
933 int err;
934
935 wd_data = kzalloc(sizeof(struct watchdog_core_data), GFP_KERNEL);
936 if (!wd_data)
937 return -ENOMEM;
938 kref_init(&wd_data->kref);
939 mutex_init(&wd_data->lock);
940
941 wd_data->wdd = wdd;
942 wdd->wd_data = wd_data;
943
944 if (IS_ERR_OR_NULL(watchdog_kworker))
945 return -ENODEV;
946
947 kthread_init_work(&wd_data->work, watchdog_ping_work);
> 948 hrtimer_init(&wd_data->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
949 wd_data->timer.function = watchdog_timer_expired;
950
951 if (wdd->id == 0) {
952 old_wd_data = wd_data;
953 watchdog_miscdev.parent = wdd->parent;
954 err = misc_register(&watchdog_miscdev);
955 if (err != 0) {
956 pr_err("%s: cannot register miscdev on minor=%d (err=%d).\n",
957 wdd->info->identity, WATCHDOG_MINOR, err);
958 if (err == -EBUSY)
959 pr_err("%s: a legacy watchdog module is probably present.\n",
960 wdd->info->identity);
961 old_wd_data = NULL;
962 kfree(wd_data);
963 return err;
964 }
965 }
966
967 /* Fill in the data structures */
968 cdev_init(&wd_data->cdev, &watchdog_fops);
969 wd_data->cdev.owner = wdd->ops->owner;
970
971 /* Add the device */
972 err = cdev_add(&wd_data->cdev, devno, 1);
973 if (err) {
974 pr_err("watchdog%d unable to add device %d:%d\n",
975 wdd->id, MAJOR(watchdog_devt), wdd->id);
976 if (wdd->id == 0) {
977 misc_deregister(&watchdog_miscdev);
978 old_wd_data = NULL;
979 kref_put(&wd_data->kref, watchdog_core_data_release);
980 }
981 return err;
982 }
983
984 /* Record time of most recent heartbeat as 'just before now'. */
985 wd_data->last_hw_keepalive = ktime_sub(ktime_get(), 1);
986
987 /*
988 * If the watchdog is running, prevent its driver from being unloaded,
989 * and schedule an immediate ping.
990 */
991 if (watchdog_hw_running(wdd)) {
992 __module_get(wdd->ops->owner);
993 kref_get(&wd_data->kref);
994 if (handle_boot_enabled)
995 hrtimer_start(&wd_data->timer, 0, HRTIMER_MODE_REL);
996 else
997 pr_info("watchdog%d running and kernel based pre-userspace handler disabled\n",
998 wdd->id);
999 }
1000
1001 return 0;
1002 }
1003
---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation
Re: [PATCH RT 2/2] watchdog, rt: prevent deferral of watchdogd wakeup
In reply to Julia CartwrightQuoted 18 lines · Julia Cartwright
When
PREEMPT_RT_FULLis enabled, all hrtimer expiry functions are deferred for execution into the context of ktimersoftd unless otherwise annotated.Deferring the expiry of the hrtimer used by the watchdog core, however, is a waste, as the callback does nothing but queue a kthread work item and wakeup watchdogd.
It's worst then that, too: the deferral through ktimersoftd also means that for correct behavior a user must adjust the scheduling parameters of both watchdogd _and_ ktimersoftd, which is unnecessary and has other side effects (like causing unrelated expiry functions to execute at potentially elevated priority).
Instead, mark the hrtimer used by the watchdog core as being
_HARDto allow it's execution directly from hardirq context. The work done in this expiry function is well-bounded and minimal.
Quoted 7 lines
A user still must adjust the scheduling parameters of the watchdogd to be correct w.r.t. their application needs.
Cc Guenter Roeck <[email protected]>Reported-and-tested-by Steffen Trumtrar <[email protected]>Reported-by Tim Sander <[email protected]>Signed-off-by Julia Cartwright <[email protected]>
Quoted 18 lines
--- drivers/watchdog/watchdog_dev.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)drivers/watchdog/watchdog_dev.c +1
-1@@ -945,7 +945,7 @@ static int watchdog_cdev_register(struct watchdog_device *wdd, dev_t devno)return -ENODEV ;kthread_init_work (&wd_data->work, watchdog_ping_work); -hrtimer_init (&wd_data->timer,CLOCK_MONOTONIC ,HRTIMER_MODE_REL ); +hrtimer_init (&wd_data->timer,CLOCK_MONOTONIC ,HRTIMER_MODE_REL_HARD ); wd_data->timer.function = watchdog_timer_expired;if (wdd->id ==0 ) {
Re: [PATCH RT 2/2] watchdog, rt: prevent deferral of watchdogd wakeup
In reply to Julia CartwrightQuoted 3 lines · Julia Cartwright
When
PREEMPT_RT_FULLis enabled, all hrtimer expiry functions are deferred for execution into the context of ktimersoftd unless otherwise annotated.
…
Quoted 1 line
Signed-off-by Julia Cartwright <[email protected]>
did s@HRTIMER_MODE_REL@HRTIMER_MODE_REL_HARD@ and then applied.
Thank you Julia.
Sebastian
Hi
I am also seeing this problem.
Am Dienstag, 18. September 2018, 15:46:15 CEST schrieb Guenter Roeck:
Quoted 3 lines
Please not only describe what you are doing, but also provide the complete context. Specifically,
- Did this ever work ?
I have not found a working kernel version.
Quoted 1 line
If yes, what are working kernel versions ?
I have tried different kernel version which where *not* working:
patch-4.11.12-rt15 (with some Altera/Intel driver patches) 4.16.8-rt3
4.18-rc8-rt1
4.18.7-rt5 (with adxl372 iio patches and denali nand race fix reported on mainline)
Quoted 1 line
- Is the watchdog device open ?
Good point: The bootloader arms the watchdog. Then WATCHDOG_HANDLE_BOOT_ENABLED is enabled in kernel. But i forgot to let usermode take over.
Quoted 1 line
- Does it make a difference if it is ?
If i set RuntimeWatchdogSec=30 in /etc/systemd/system.conf (i.e. enable usermode watchdog) the problem seems to go away (at least i didn't had a reboot dony with my short testruns by now).
Quoted 2 lines
- What is the configured watchdog timeout (both from BIOS/ROMMON and in Linux) ?
30s
So to wrap up: i have the suspicion that the option WATCHDOG_HANDLE_BOOT_ENABLED without usermode taking over in Linux leads to some strange behaviour with preempt rt?
Best regards
Tim