Since the backport commit eea46baf1451 ("ftrace: Fix possible
use-after-free issue in ftrace_location()") on linux-5.4.y branch, the
old ftrace_int3_handler()->ftrace_location() path has included
rcu_read_lock(), which has mcount location inside and leads to potential
double fault.

Replace rcu_read_lock/unlock with preempt_enable/disable notrace macros
so that the mcount location does not appear on the int3 handler path.

This fix is specific to linux-5.4.y branch, the only branch still using
ftrace_int3_handler with commit e60b613df8b6 ("ftrace: Fix possible
use-after-free issue in ftrace_location()") backported. It also avoids
the need to backport the code conversion to text_poke() on this branch.

Reported-by: Koichiro Den <koichiro.den@canonical.com>
Closes: https://lore.kernel.org/all/74gjhwxupvozwop7ndhrh7t5qeckomt7yqvkkbm5j2tlx6dkfk@rgv7sijvry2k
Fixes: eea46baf1451 ("ftrace: Fix possible use-after-free issue in ftrace_location()") # linux-5.4.y
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Signed-off-by: Koichiro Den <koichiro.den@canonical.com>
---
 kernel/trace/ftrace.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
index 380032a27f98..2eb1a8ec5755 100644
--- a/kernel/trace/ftrace.c
+++ b/kernel/trace/ftrace.c
@@ -1554,7 +1554,7 @@ unsigned long ftrace_location_range(unsigned long start, unsigned long end)
 	struct dyn_ftrace key;
 	unsigned long ip = 0;
 
-	rcu_read_lock();
+	preempt_disable_notrace();
 	key.ip = start;
 	key.flags = end;	/* overload flags, as it is unsigned long */
 
@@ -1572,7 +1572,7 @@ unsigned long ftrace_location_range(unsigned long start, unsigned long end)
 			break;
 		}
 	}
-	rcu_read_unlock();
+	preempt_enable_notrace();
 	return ip;
 }
 
-- 
2.43.0

On Wed, Jan 08, 2025 at 12:17:36PM +0900, Koichiro Den wrote:
> Since the backport commit eea46baf1451 ("ftrace: Fix possible
> use-after-free issue in ftrace_location()") on linux-5.4.y branch, the
> old ftrace_int3_handler()->ftrace_location() path has included
> rcu_read_lock(), which has mcount location inside and leads to potential
> double fault.
> 
> Replace rcu_read_lock/unlock with preempt_enable/disable notrace macros
> so that the mcount location does not appear on the int3 handler path.
> 
> This fix is specific to linux-5.4.y branch, the only branch still using
> ftrace_int3_handler with commit e60b613df8b6 ("ftrace: Fix possible
> use-after-free issue in ftrace_location()") backported. It also avoids
> the need to backport the code conversion to text_poke() on this branch.
> 
> Reported-by: Koichiro Den <koichiro.den@canonical.com>
> Closes: https://lore.kernel.org/all/74gjhwxupvozwop7ndhrh7t5qeckomt7yqvkkbm5j2tlx6dkfk@rgv7sijvry2k
> Fixes: eea46baf1451 ("ftrace: Fix possible use-after-free issue in ftrace_location()") # linux-5.4.y
> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
> Signed-off-by: Koichiro Den <koichiro.den@canonical.com>
> ---
>  kernel/trace/ftrace.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)

Thanks, now queued up.

greg k-h