lkml.org 
[lkml]   [2022]   [Dec]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH] rcu: Fix kernel stack overflow caused by kprobe on rcu_irq_enter_check_tick()
Date
On Mon, 5 Dec 2022 14:23:53 +0100, Frederic Weisbecker wrote:
> On Mon, Nov 21, 2022 at 11:57:03AM -0800, Paul E. McKenney wrote:
> > On Sat, Nov 19, 2022 at 12:00:49PM +0800, Zheng Yejian wrote:
> > > Register kprobe on __rcu_irq_enter_check_tick() can cause kernel stack
> > > overflow [1]. This issue is first found in v5.10 and can be reproduced
> > > by enabling CONFIG_NO_HZ_FULL and doing like:
> > > # cd /sys/kernel/debug/tracing/
> > > # echo 'p:mp1 __rcu_irq_enter_check_tick' >> kprobe_events
> > > # echo 1 > events/kprobes/enable
> > >
> > > So __rcu_irq_enter_check_tick() should not be kprobed, mark it as noinstr.
> >
> > Good catch!
> >
> > I am inclined to queue this, but noticed that one of its callers need
> > it to be noinstr but that the others do not.
> >
> > Need noinstr:
> >
> > o enter_from_kernel_mode() -> __enter_from_kernel_mode() ->
> > rcu_irq_enter_check_tick() -> __rcu_irq_enter_check_tick()
> >
> > Doesn't need noinstr:
> >
> > o ct_nmi_enter() -> rcu_irq_enter_check_tick() ->
> > __rcu_irq_enter_check_tick(), courtesy of the call to
> > instrumentation_begin() in ct_nmi_enter() that precedes the call
> > to rcu_irq_enter_check_tick().
> >
> > o irqentry_enter() -> rcu_irq_enter_check_tick() ->
> > __rcu_irq_enter_check_tick(), courtesy of the call to
> > instrumentation_begin() in irqentry_enter() that precedes the
> > call to rcu_irq_enter_check_tick().
> >
> > Is tagging __rcu_irq_enter_check_tick() with noinstr as
> > proposed in this patch the right thing to do, or should there
> > be calls to instrumentation_begin() and instrumentation_end() in
> > enter_from_kernel_mode()? Or something else entirely?
>
> Tagging as noinstr doesn't look right as there are functions in
> __rcu_irq_enter_check_tick() that can be traced anyway. Also that
> function has the constraint that it can't be called while RCU is idle
> so it's up to the caller to call instrumentation_begin()/end().

This problem is due to __rcu_irq_enter_check_tick() being kprobe-ed, so
how about adding __rcu_irq_enter_check_tick() into kprobe blacklist by
tagging with NOKPROBE_SYMBOL:

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index cf34a961821a..41606d3ed083 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -659,6 +659,7 @@ void __rcu_irq_enter_check_tick(void)
}
raw_spin_unlock_rcu_node(rdp->mynode);
}
+NOKPROBE_SYMBOL(__rcu_irq_enter_check_tick);
#endif /* CONFIG_NO_HZ_FULL */

+Cc: <mhiramat@kernel.org>

--
Best regards,
Zheng Yejian

\
 
 \ /
  Last update: 2023-03-26 23:22    [W:0.065 / U:0.288 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site