On Wed, 13 Jul 2022 18:20:09 +0800
Zheng Yejian <zhengyejian1@xxxxxxxxxx> wrote:
> This patch and problem analysis is based on v4.19 LTS, but v5.4 LTS
> and below seem to be involved.
>
> Hulk Robot reports a softlockup problem, see following logs:
> [ 41.463870] watchdog: BUG: soft lockup - CPU#0 stuck for 22s! [ksoftirqd/0:9]
This detects something that is spinning with preemption disabled but
interrupts enabled.
> Look into above call stack, there is a recursive call in
> 'ftrace_graph_call', and the direct cause of above recursion is that
> 'rcu_dynticks_curr_cpu_in_eqs' is traced, see following snippet:
> __read_once_size_nocheck.constprop.0
> ftrace_graph_call <-- 1. first call
> ......
> rcu_dynticks_curr_cpu_in_eqs
> ftrace_graph_call <-- 2. recursive call here!!!
This is not the bug. That code can handle a recursion:
ftrace_graph_call is assembly that is converted to call
void prepare_ftrace_return(unsigned long ip, unsigned long *parent,
unsigned long frame_pointer)
{
[..]
bit = ftrace_test_recursion_trylock(ip, *parent);
if (bit < 0)
return;
This will stop the code as "bit" will be < 0 on the second call to
ftrace_graph_call. If it was a real recursion issue, it would crash the
machine when the recursion runs out of stack space.
>
> Comparing with mainline kernel, commit ff5c4f5cad33 ("rcu/tree:
> Mark the idle relevant functions noinstr") mark related functions as
> 'noinstr' which implies notrace, noinline and sticks things in the
> .noinstr.text section.
> Link: https://lore.kernel.org/all/20200416114706.625340212@xxxxxxxxxxxxx/
>
> But we cannot directly backport that commit, because there seems to be
> many prepatches. Instead, marking the functions as 'notrace' where it is
> 'noinstr' in that commit and mark 'rcu_dynticks_curr_cpu_in_eqs' as
> inline look like it resolves the problem.
That will not fix your problem.
>
> Reported-by: Hulk Robot <hulkci@xxxxxxxxxx>
> Signed-off-by: Zheng Yejian <zhengyejian1@xxxxxxxxxx>
Can you reproduce this consistently without this patch, and then not so
with this patch?
Or are you just assuming that this fixes a bug because you observed a
recursion?
Please explain to me why this would cause the hang?
-- Steve
