Messages in this thread | | | Date | Fri, 1 Apr 2022 13:46:18 +0200 | From | Peter Zijlstra <> | Subject | Re: [PATCH] sched/core: Fix forceidle balancing |
| |
On Thu, Mar 31, 2022 at 03:00:40PM -0400, Joel Fernandes wrote: > Hi, > > By the way, might be slightly related - we still see crashes with > pick_task_fair() in our kernel even with this change: > https://lkml.org/lkml/2020/11/17/2137
Please as to not use lkml.org. Please use something with a MsgID in like lore.
> Is it possible that when doing pick_task_fair() especially on a remote > CPU, both the "cfs_rq->curr" and the rbtree's "left" be NULL with core > scheduling? In this case, se will be NULL and can cause crashes right? > I think the code assumes this can never happen. > > +Guenter Roeck kindly debugged pick_task_fair() in a crash as > follows. Copying some details he mentioned in a bug report: > > Assembler/source: > > 25: e8 4f 11 00 00 call 0x1179 ; se = > pick_next_entity(cfs_rq, curr); > 2a:* 48 8b 98 60 01 00 00 mov 0x160(%rax),%rbx ; trapping > instruction [cfs_rq = group_cfs_rq(se);] > 31: 48 85 db test %rbx,%rbx > 34: 75 d1 jne 0x7 > 36: 48 89 c7 mov %rax,%rdi > > At 2a: RAX = se == NULL after pick_next_entity(). Looking closely into > pick_next_entity(), it can indeed return NULL if curr is NULL and if > left in pick_next_entity() is NULL. Per line 7:, curr is in %r14 and > indeed 0. > > Thoughts?
It is possible for ->curr and ->leftmost to be NULL, but then we should also be having ->nr_running == 0 and not call pick in the first place. Because picking a task from no tasks doesn't make much sense.
| |