Messages in this thread | | | Date | Thu, 4 May 2023 09:49:50 +0800 | Subject | Re: [Question] softlockup in run_timer_softirq | From | "liujian (CE)" <> |
| |
On 2023/5/2 11:06, John Stultz wrote: > On Wed, Feb 15, 2023 at 12:34 AM liujian (CE) <liujian56@huawei.com> wrote: >>> On Fri, Feb 10, 2023 at 1:51 AM liujian (CE) <liujian56@huawei.com> wrote: >>>> >>>> During the syz test, we encountered many problems with various timer >>>> handler functions softlockup. >>>> >>>> We analyze __run_timers() and find the following problem. >>>> >>>> In the while loop of __run_timers(), because there are too many timers >>>> or improper timer handler functions, if the processing time of the >>>> expired timers is always greater than the time wheel's next_expiry, >>>> the function will loop infinitely. >>>> >>>> The following extreme test case can be used to reproduce the problem. >>>> An extreme test case[1] is constructed to reproduce the problem. >>> >>> Thanks for reporting and sending out this data: >>> >>> First, any chance you might submit this as a in-kernel-stress test? >>> Maybe utilizing the kernel/torture.c framework? >>> >> Okay, I'll learn this framework and do this thing. >>> (Though the test may need to occasionally take a break so the system can >>> eventually catch up) >>> >>>> Is this a problem or an unreasonable use? >>>> >>>> Can we limit the running time of __run_timers() [2]? >>>> >>>> Does anyone have a good idea to solve this problem? >>> >>> So your patch reminds me of Peter's softirq_needs_break() logic: >>> >>> https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/log/?h=co >>> re/softirq >>> >>> Maybe it could extend that series for the timer softirq as well? >>> >> Thank you. Yes. >> Base on the patchset and the extended patch for timer [1], the soft lockup problem does not occur. >> >> By the way, I see this is a very old patchset? Will this patchset push the main line? @John @Peter >> >> >> [1] >> Author: Liu Jian <liujian56@huawei.com> >> Date: Tue Feb 14 09:53:46 2023 +0800 >> >> softirq, timer: Use softirq_needs_break() >> >> In the while loop of __run_timers(), because there are too many timers or >> improper timer handler functions, if the processing time of the expired >> timers is always greater than the time wheel's next_expiry, the function >> will loop infinitely. >> >> To prevent this, use the timeout/break logic provided by SoftIRQs.If the >> running time exceeds the limit, break the loop and an additional >> TIMER_SOFTIRQ is triggered. >> >> Signed-off-by: Liu Jian <liujian56@huawei.com> >> >> diff --git a/kernel/time/timer.c b/kernel/time/timer.c >> index 63a8ce7177dd..70744a469a39 100644 >> --- a/kernel/time/timer.c >> +++ b/kernel/time/timer.c >> @@ -1992,7 +1992,7 @@ void timer_clear_idle(void) >> * __run_timers - run all expired timers (if any) on this CPU. >> * @base: the timer vector to be processed. >> */ >> -static inline void __run_timers(struct timer_base *base) >> +static inline void __run_timers(struct timer_base *base, struct softirq_action *h) >> { >> struct hlist_head heads[LVL_DEPTH]; >> int levels; >> @@ -2020,6 +2020,12 @@ static inline void __run_timers(struct timer_base *base) >> >> while (levels--) >> expire_timers(base, heads + levels); >> + >> + if (softirq_needs_break(h)) { >> + if (time_after_eq(jiffies, base->next_expiry)) >> + __raise_softirq_irqoff(TIMER_SOFTIRQ); >> + break; >> + } >> } >> raw_spin_unlock_irq(&base->lock); >> timer_base_unlock_expiry(base); >> @@ -2032,9 +2038,9 @@ static __latent_entropy void run_timer_softirq(struct softirq_action *h) >> { >> struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]); >> >> - __run_timers(base); >> + __run_timers(base, h); >> if (IS_ENABLED(CONFIG_NO_HZ_COMMON)) >> - __run_timers(this_cpu_ptr(&timer_bases[BASE_DEF])); >> + __run_timers(this_cpu_ptr(&timer_bases[BASE_DEF]), h); >> } >> >> /* > > So I wanted to revive this old thread, as Frank Woo mentioned his team > has seen a similar issue as well. > > Liujian: I'm curious if you've made any further progress with your > adapted patch ontop of PeterZ's softirq_needs_break patch series? > Hi John, Only the commit ("softirq, timer: Use softirq_needs_break()") is added to the patchset of Peter, and no other modification is made. > Might it be worth re-submitting the whole series for consideration upstream? > I agree very much and expect, because we often encounter similar problems when doing fuzzy tests (especially when the test machine is poor). > thanks > -john
| |