Re: [Xenomai-core] [BUG] scheduling order of dying shadow threads
Jan Kiszka wrote: Philippe Gerum wrote: Jan Kiszka wrote: Hi Philippe, I think this one is for you: ;) Sebastian got almost mad with his CAN driver while tracing a strange scheduling behaviour during shadow thread deletion for several days(!) - and I was right on the way to follow him yesterday evening. Attached is a simplified demonstration of the effect, consisting of a RTDM driver and both a kernel and user space application to trigger it. I've spotted the issue in nucleus/shadow.c. Basically, the root thread priority boost was leaking to a non-shadow thread due to a missing priority reset in the lostage APC handler, whilst a shadow was in the process of relaxing. Really funky bug, thanks! :o> Fixed in the repo hopefully for good. The scheduling sequence is now correct with your demo app on my box. Yep, looks good here as well. Great and quick work! Just don't expect that someone can follow your explanations easily. :) Well, sorry. Here is a more useful explanation : The demo thread in your code calls sleep(1) before exiting, which causes the underlying shadow thread to relax. The same would happen without sleeping, since a terminating thread is silently relaxed by the nucleus in any case as needed. When relaxing the current thread, xnshadow_relax() first boosts the priority of the root thread (i.e. the placeholder for Linux in the Xenomai scheduler) right before suspending itself. Before that, a wake up request has been scheduled (using an APC), so that lostage_handler will be called, which will in turn invoke wake_up_process() for the relaxing thread. This is needed because shadows running in primary mode are seen as suspended in the Linux sense in TASK_INTERRUPTIBLE state. The reason for this is that both Xenomai and Linux schedulers must have a mutually exclusive control over a shadow, they should not be allowed to both fiddle concurrently with a single thread context; conversely, relaxed thread operating in secondary mode are seen as suspended on the XNRELAX condition by the nucleus. IOW, what we want to do here is some kind of transition from the Xenomai to the Linux scheduler for the relaxing shadow thread. This way, we make sure that the Linux scheduler will get back in control for the awaken shadow thread, which ends up running in secondary mode once it has been resumed by wake_up_process(). Problem is that the unless we actually reset the root thread priority to the lowest one in lostage_handler in order to revert the priority boost done in xnshadow_relax, there is a short window of time during which a normal Linux task that has been preempted by the APC request that runs lostage_handler could run and wreck the scheduling sequence (e.g. your main() context). The fix is about downgrading the root thread priority and waking the relaxed shadow up in the same move, so that the priority scheme is kept intact. Now the question is: why does the root thread priority need to be upgraded while relaxing a shadow? The answer is simple: when relaxing a shadow, you are not expected to change tasks in a Xenomai or Linux sense, you are only changing the Xenomai exec mode for a shadow, which means that we must ensure that giving control back to the Linux kernel just for the purpose of changing the current exec mode won't cause the current priority level of the relaxing thread to be lost and spuriously downgraded to the lowest one of the system. So we just boost it to be equal to the one of the relaxing thread; this way, the Linux kernel code undergoes a Xenomai RT priority boost so that Linux cannot be preempted by lower priority Xenomai threads. When a shadow thread running in secondary mode is switched in, the root thread priority always inherits the Xenomai priority level for that thread; conversely, when a non-Xenomai/regular Linux task is scheduled in, the root thread priority is downgraded to the lowest Xenomai priority. If one thinks a bit ahead now, having this scheme in place, we should be able to benefit from every improvement in the vanilla Linux kernel granularity toward real-time guarantees. Because we don't break the priority scheme moving in and out of the Linux domain, a Xenomai scheduling decision remains consistent with the Linux priority scheme, which is a necessary condition for providing a high integration level between Xeno and the vanilla kernel. I think this issue has some similarity with the one I once stumbled over regarding non-RT signalling to Linux. I'm not going to repeat my general concerns regarding the priority boosting of the root thread now... ;) Each time you spot a bug like this, your stack of concerns should lose at least one element, isn'it? :o> Until Linux is really able to provide a fine-grained, non-disruptive and not easily disrupted (e.g. locking semantics in drivers), and low-overhead core implementation for RT support, Xeno will need to provide its own primary scheduler for latency-cr
Re: [Xenomai-core] [BUG] scheduling order of dying shadow threads
On Tue, 8 Nov 2005, Philippe Gerum wrote: > Jan Kiszka wrote: > > Hi Philippe, > > > > I think this one is for you: ;) > > > > Sebastian got almost mad with his CAN driver while tracing a strange > > scheduling behaviour during shadow thread deletion for several days(!) - > > and I was right on the way to follow him yesterday evening. Attached is > > a simplified demonstration of the effect, consisting of a RTDM driver > > and both a kernel and user space application to trigger it. > > > > I've spotted the issue in nucleus/shadow.c. Basically, the root thread > priority > boost was leaking to a non-shadow thread due to a missing priority reset in > the > lostage APC handler, whilst a shadow was in the process of relaxing. Really > funky bug, thanks! :o> Fixed in the repo hopefully for good. The scheduling > sequence is now correct with your demo app on my box. Thank you, problem has disappeared here, too! :-) Sebastian
Re: [Xenomai-core] [BUG] scheduling order of dying shadow threads
Philippe Gerum wrote: > Jan Kiszka wrote: >> Hi Philippe, >> >> I think this one is for you: ;) >> >> Sebastian got almost mad with his CAN driver while tracing a strange >> scheduling behaviour during shadow thread deletion for several days(!) - >> and I was right on the way to follow him yesterday evening. Attached is >> a simplified demonstration of the effect, consisting of a RTDM driver >> and both a kernel and user space application to trigger it. >> > > I've spotted the issue in nucleus/shadow.c. Basically, the root thread > priority boost was leaking to a non-shadow thread due to a missing > priority reset in the lostage APC handler, whilst a shadow was in the > process of relaxing. Really funky bug, thanks! :o> Fixed in the repo > hopefully for good. The scheduling sequence is now correct with your > demo app on my box. > Yep, looks good here as well. Great and quick work! Just don't expect that someone can follow your explanations easily. :) I think this issue has some similarity with the one I once stumbled over regarding non-RT signalling to Linux. I'm not going to repeat my general concerns regarding the priority boosting of the root thread now... ;) Jan signature.asc Description: OpenPGP digital signature
Re: [Xenomai-core] [BUG] scheduling order of dying shadow threads
Jan Kiszka wrote: Hi Philippe, I think this one is for you: ;) Sebastian got almost mad with his CAN driver while tracing a strange scheduling behaviour during shadow thread deletion for several days(!) - and I was right on the way to follow him yesterday evening. Attached is a simplified demonstration of the effect, consisting of a RTDM driver and both a kernel and user space application to trigger it. I've spotted the issue in nucleus/shadow.c. Basically, the root thread priority boost was leaking to a non-shadow thread due to a missing priority reset in the lostage APC handler, whilst a shadow was in the process of relaxing. Really funky bug, thanks! :o> Fixed in the repo hopefully for good. The scheduling sequence is now correct with your demo app on my box. -- Philippe.
Re: [Xenomai-core] [BUG] scheduling order of dying shadow threads
Jan Kiszka wrote: Hi Philippe, I think this one is for you: ;) Sebastian got almost mad with his CAN driver while tracing a strange scheduling behaviour during shadow thread deletion for several days(!) - and I was right on the way to follow him yesterday evening. Attached is a simplified demonstration of the effect, consisting of a RTDM driver and both a kernel and user space application to trigger it. Assume two or more user space RT-threads blocking on the same RTDM semaphore inside a driver (I was not yet able to reproduce this with a simply native user space application :/). All get then woken up on rtdm_sem_destroy during device closure. They increment a global counter, save the current value in a per-thread variable, and then terminate. They had also passed another per-thread variable to the RTDM driver which was updated in the kernel using the same(!) counter. /* application */ void demo(void *arg) { rt_dev_read(dev, &value_k[(int)arg], 0); value_u[(int)arg] = ++counter; } /* driver */ int demo_read_rt(struct rtdm_dev_context*context, rtdm_user_info_t *user_info, void *buf, size_t nbyte) { struct demodrv_context *my_context; int ret; my_context = (struct demodrv_context *)context->dev_private; ret = rtdm_sem_down(&my_context->read_sem); *(int *)buf = ++(*counter); return ret; } That global counter is also incremented during closure to visualise the call order: int demo_close_rt(struct rtdm_dev_context *context, rtdm_user_info_t *user_info) { struct demodrv_context *my_context; my_context = (struct demodrv_context *)context->dev_private; printk("close 1: %d\n", xnpod_current_thread()->cprio); rtdm_sem_destroy(&my_context->read_sem); printk("close 2: %d\n", xnpod_current_thread()->cprio); (*counter)++; return 0; } Now one would expect the following content of the involved variables when running 3 threads e.g.: thread 1 (prio 99) / thread 2(prio 98) | / thread 3 (prio 97) | | / value_k: 1, 3, 5 value_u: 2, 4, 6 counter: 7 This is indeed what we get when the application locates in kernel space, i.e. does not use shadow threads. But when it is a user space application, the result looks like this: thread 1 / thread 2 | / thread 3 | | / value_k: 1, 4, 6 value_u: 2, 5, 7 counter: 7 Which means that first thread returns from kernel to user space and terminates, then the close handler gets executed again, and only afterwards the remaining threads! The reason is also displayed by demodrv: close 1: 0 - prio of root thread before rtdm_sem_destroy close 2: 99 - ... and after rtdm_sem_destroy Which means that the non-RT thread calling rt_dev_close gets lifted to prio 99 on calling rtdm_sem_destroy, the prio of the thread first woken up. It seems to loose this prio quite soon again, but not soon enough to avoid the inversion - very strange. Any ideas? Not yet, but your analysis looks right, the main thread seems to spuriously recycle the priority level of task1 as the latter enters the secondary mode. The good news is that it's likely that such issue only impacts the deletion process which involves some specific transitions. I'll try to reproduce this bug asap and let you know. Thanks. Jan ___ Xenomai-core mailing list Xenomai-core@gna.org https://mail.gna.org/listinfo/xenomai-core -- Philippe.
Re: [Xenomai-core] [BUG] scheduling order of dying shadow threads
Jan Kiszka wrote: Philippe Gerum wrote: Jan Kiszka wrote: Hi Philippe, I think this one is for you: ;) Sebastian got almost mad with his CAN driver while tracing a strange scheduling behaviour during shadow thread deletion for several days(!) - and I was right on the way to follow him yesterday evening. Attached is a simplified demonstration of the effect, consisting of a RTDM driver and both a kernel and user space application to trigger it. I've spotted the issue in nucleus/shadow.c. Basically, the root thread priority boost was leaking to a non-shadow thread due to a missing priority reset in the lostage APC handler, whilst a shadow was in the process of relaxing. Really funky bug, thanks! :o> Fixed in the repo hopefully for good. The scheduling sequence is now correct with your demo app on my box. Yep, looks good here as well. Great and quick work! Just don't expect that someone can follow your explanations easily. :) Well, sorry. Here is a more useful explanation : The demo thread in your code calls sleep(1) before exiting, which causes the underlying shadow thread to relax. The same would happen without sleeping, since a terminating thread is silently relaxed by the nucleus in any case as needed. When relaxing the current thread, xnshadow_relax() first boosts the priority of the root thread (i.e. the placeholder for Linux in the Xenomai scheduler) right before suspending itself. Before that, a wake up request has been scheduled (using an APC), so that lostage_handler will be called, which will in turn invoke wake_up_process() for the relaxing thread. This is needed because shadows running in primary mode are seen as suspended in the Linux sense in TASK_INTERRUPTIBLE state. The reason for this is that both Xenomai and Linux schedulers must have a mutually exclusive control over a shadow, they should not be allowed to both fiddle concurrently with a single thread context; conversely, relaxed thread operating in secondary mode are seen as suspended on the XNRELAX condition by the nucleus. IOW, what we want to do here is some kind of transition from the Xenomai to the Linux scheduler for the relaxing shadow thread. This way, we make sure that the Linux scheduler will get back in control for the awaken shadow thread, which ends up running in secondary mode once it has been resumed by wake_up_process(). Problem is that the unless we actually reset the root thread priority to the lowest one in lostage_handler in order to revert the priority boost done in xnshadow_relax, there is a short window of time during which a normal Linux task that has been preempted by the APC request that runs lostage_handler could run and wreck the scheduling sequence (e.g. your main() context). The fix is about downgrading the root thread priority and waking the relaxed shadow up in the same move, so that the priority scheme is kept intact. Now the question is: why does the root thread priority need to be upgraded while relaxing a shadow? The answer is simple: when relaxing a shadow, you are not expected to change tasks in a Xenomai or Linux sense, you are only changing the Xenomai exec mode for a shadow, which means that we must ensure that giving control back to the Linux kernel just for the purpose of changing the current exec mode won't cause the current priority level of the relaxing thread to be lost and spuriously downgraded to the lowest one of the system. So we just boost it to be equal to the one of the relaxing thread; this way, the Linux kernel code undergoes a Xenomai RT priority boost so that Linux cannot be preempted by lower priority Xenomai threads. When a shadow thread running in secondary mode is switched in, the root thread priority always inherits the Xenomai priority level for that thread; conversely, when a non-Xenomai/regular Linux task is scheduled in, the root thread priority is downgraded to the lowest Xenomai priority. If one thinks a bit ahead now, having this scheme in place, we should be able to benefit from every improvement in the vanilla Linux kernel granularity toward real-time guarantees. Because we don't break the priority scheme moving in and out of the Linux domain, a Xenomai scheduling decision remains consistent with the Linux priority scheme, which is a necessary condition for providing a high integration level between Xeno and the vanilla kernel. I think this issue has some similarity with the one I once stumbled over regarding non-RT signalling to Linux. I'm not going to repeat my general concerns regarding the priority boosting of the root thread now... ;) Each time you spot a bug like this, your stack of concerns should lose at least one element, isn'it? :o> Until Linux is really able to provide a fine-grained, non-disruptive and not easily disrupted (e.g. locking semantics in drivers), and low-overhead core implementation for RT support, Xeno will need to provide its own primary scheduler for latency-cr
Re: [Xenomai-core] [BUG] scheduling order of dying shadow threads
On Tue, 8 Nov 2005, Philippe Gerum wrote: > Jan Kiszka wrote: > > Hi Philippe, > > > > I think this one is for you: ;) > > > > Sebastian got almost mad with his CAN driver while tracing a strange > > scheduling behaviour during shadow thread deletion for several days(!) - > > and I was right on the way to follow him yesterday evening. Attached is > > a simplified demonstration of the effect, consisting of a RTDM driver > > and both a kernel and user space application to trigger it. > > > > I've spotted the issue in nucleus/shadow.c. Basically, the root thread > priority > boost was leaking to a non-shadow thread due to a missing priority reset in > the > lostage APC handler, whilst a shadow was in the process of relaxing. Really > funky bug, thanks! :o> Fixed in the repo hopefully for good. The scheduling > sequence is now correct with your demo app on my box. Thank you, problem has disappeared here, too! :-) Sebastian ___ Xenomai-core mailing list Xenomai-core@gna.org https://mail.gna.org/listinfo/xenomai-core
Re: [Xenomai-core] [BUG] scheduling order of dying shadow threads
Philippe Gerum wrote: > Jan Kiszka wrote: >> Hi Philippe, >> >> I think this one is for you: ;) >> >> Sebastian got almost mad with his CAN driver while tracing a strange >> scheduling behaviour during shadow thread deletion for several days(!) - >> and I was right on the way to follow him yesterday evening. Attached is >> a simplified demonstration of the effect, consisting of a RTDM driver >> and both a kernel and user space application to trigger it. >> > > I've spotted the issue in nucleus/shadow.c. Basically, the root thread > priority boost was leaking to a non-shadow thread due to a missing > priority reset in the lostage APC handler, whilst a shadow was in the > process of relaxing. Really funky bug, thanks! :o> Fixed in the repo > hopefully for good. The scheduling sequence is now correct with your > demo app on my box. > Yep, looks good here as well. Great and quick work! Just don't expect that someone can follow your explanations easily. :) I think this issue has some similarity with the one I once stumbled over regarding non-RT signalling to Linux. I'm not going to repeat my general concerns regarding the priority boosting of the root thread now... ;) Jan signature.asc Description: OpenPGP digital signature ___ Xenomai-core mailing list Xenomai-core@gna.org https://mail.gna.org/listinfo/xenomai-core
Re: [Xenomai-core] [BUG] scheduling order of dying shadow threads
Jan Kiszka wrote: Hi Philippe, I think this one is for you: ;) Sebastian got almost mad with his CAN driver while tracing a strange scheduling behaviour during shadow thread deletion for several days(!) - and I was right on the way to follow him yesterday evening. Attached is a simplified demonstration of the effect, consisting of a RTDM driver and both a kernel and user space application to trigger it. I've spotted the issue in nucleus/shadow.c. Basically, the root thread priority boost was leaking to a non-shadow thread due to a missing priority reset in the lostage APC handler, whilst a shadow was in the process of relaxing. Really funky bug, thanks! :o> Fixed in the repo hopefully for good. The scheduling sequence is now correct with your demo app on my box. -- Philippe. ___ Xenomai-core mailing list Xenomai-core@gna.org https://mail.gna.org/listinfo/xenomai-core
Re: [Xenomai-core] [BUG] scheduling order of dying shadow threads
Jan Kiszka wrote: Hi Philippe, I think this one is for you: ;) Sebastian got almost mad with his CAN driver while tracing a strange scheduling behaviour during shadow thread deletion for several days(!) - and I was right on the way to follow him yesterday evening. Attached is a simplified demonstration of the effect, consisting of a RTDM driver and both a kernel and user space application to trigger it. Assume two or more user space RT-threads blocking on the same RTDM semaphore inside a driver (I was not yet able to reproduce this with a simply native user space application :/). All get then woken up on rtdm_sem_destroy during device closure. They increment a global counter, save the current value in a per-thread variable, and then terminate. They had also passed another per-thread variable to the RTDM driver which was updated in the kernel using the same(!) counter. /* application */ void demo(void *arg) { rt_dev_read(dev, &value_k[(int)arg], 0); value_u[(int)arg] = ++counter; } /* driver */ int demo_read_rt(struct rtdm_dev_context*context, rtdm_user_info_t *user_info, void *buf, size_t nbyte) { struct demodrv_context *my_context; int ret; my_context = (struct demodrv_context *)context->dev_private; ret = rtdm_sem_down(&my_context->read_sem); *(int *)buf = ++(*counter); return ret; } That global counter is also incremented during closure to visualise the call order: int demo_close_rt(struct rtdm_dev_context *context, rtdm_user_info_t *user_info) { struct demodrv_context *my_context; my_context = (struct demodrv_context *)context->dev_private; printk("close 1: %d\n", xnpod_current_thread()->cprio); rtdm_sem_destroy(&my_context->read_sem); printk("close 2: %d\n", xnpod_current_thread()->cprio); (*counter)++; return 0; } Now one would expect the following content of the involved variables when running 3 threads e.g.: thread 1 (prio 99) / thread 2(prio 98) | / thread 3 (prio 97) | | / value_k: 1, 3, 5 value_u: 2, 4, 6 counter: 7 This is indeed what we get when the application locates in kernel space, i.e. does not use shadow threads. But when it is a user space application, the result looks like this: thread 1 / thread 2 | / thread 3 | | / value_k: 1, 4, 6 value_u: 2, 5, 7 counter: 7 Which means that first thread returns from kernel to user space and terminates, then the close handler gets executed again, and only afterwards the remaining threads! The reason is also displayed by demodrv: close 1: 0 - prio of root thread before rtdm_sem_destroy close 2: 99 - ... and after rtdm_sem_destroy Which means that the non-RT thread calling rt_dev_close gets lifted to prio 99 on calling rtdm_sem_destroy, the prio of the thread first woken up. It seems to loose this prio quite soon again, but not soon enough to avoid the inversion - very strange. Any ideas? Not yet, but your analysis looks right, the main thread seems to spuriously recycle the priority level of task1 as the latter enters the secondary mode. The good news is that it's likely that such issue only impacts the deletion process which involves some specific transitions. I'll try to reproduce this bug asap and let you know. Thanks. Jan ___ Xenomai-core mailing list Xenomai-core@gna.org https://mail.gna.org/listinfo/xenomai-core -- Philippe. ___ Xenomai-core mailing list Xenomai-core@gna.org https://mail.gna.org/listinfo/xenomai-core