From: Vladimir Riabchun <ferr.lambarginio@gmail.com>

[ Upstream commit 4099b98203d6b33d990586542fa5beee408032a3 ]

A soft lockup was observed when loading amdgpu module.
If a module has a lot of tracable functions, multiple calls
to kallsyms_lookup can spend too much time in RCU critical
section and with disabled preemption, causing kernel panic.
This is the same issue that was fixed in
commit d0b24b4e91fc ("ftrace: Prevent RCU stall on PREEMPT_VOLUNTARY
kernels") and commit 42ea22e754ba ("ftrace: Add cond_resched() to
ftrace_graph_set_hash()").

Fix it the same way by adding cond_resched() in ftrace_module_enable.

Link: https://lore.kernel.org/aMQD9_lxYmphT-up@vova-pc
Signed-off-by: Vladimir Riabchun <ferr.lambarginio@gmail.com>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
---

LLM Generated explanations, may be completely bogus:

YES
- The change inserts `cond_resched()` inside the inner iteration over
  every ftrace record (`kernel/trace/ftrace.c:7538`). That loop holds
  the ftrace mutex and, for each record, invokes heavy helpers like
  `test_for_valid_rec()` which in turn calls `kallsyms_lookup()`
  (`kernel/trace/ftrace.c:4289`). On huge modules (e.g. amdgpu) this can
  run for tens of milliseconds with preemption disabled, triggering the
  documented soft lockup/panic during module load.
- `ftrace_module_enable()` runs only in process context via
  `prepare_coming_module()` (`kernel/module/main.c:3279`), so adding a
  voluntary reschedule point is safe; the same pattern already exists in
  other long-running ftrace loops (see commits d0b24b4e91fc and
  42ea22e754ba), so this brings consistency without changing control
  flow or semantics.
- No data structures or interfaces change, and the code still executes
  under the same locking (`ftrace_lock`, `text_mutex` when the arch
  overrides `ftrace_arch_code_modify_prepare()`), so the risk of
  regression is minimal: the new call simply yields CPU if needed while
  keeping the locks held, preventing watchdog-induced crashes but
  otherwise behaving identically.

Given it fixes a real, user-visible soft lockup with a contained and
well-understood tweak, this is an excellent candidate for stable
backporting.

 kernel/trace/ftrace.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
index a69067367c296..42bd2ba68a821 100644
--- a/kernel/trace/ftrace.c
+++ b/kernel/trace/ftrace.c
@@ -7535,6 +7535,8 @@ void ftrace_module_enable(struct module *mod)
 		if (!within_module(rec->ip, mod))
 			break;
 
+		cond_resched();
+
 		/* Weak functions should still be ignored */
 		if (!test_for_valid_rec(rec)) {
 			/* Clear all other flags. Should not be enabled anyway */
-- 
2.51.0

On Sat, 25 Oct 2025 12:00:16 -0400
Sasha Levin <sashal@kernel.org> wrote:

> - The change inserts `cond_resched()` inside the inner iteration over
>   every ftrace record (`kernel/trace/ftrace.c:7538`). That loop holds
>   the ftrace mutex and, for each record, invokes heavy helpers like
>   `test_for_valid_rec()` which in turn calls `kallsyms_lookup()`
>   (`kernel/trace/ftrace.c:4289`). On huge modules (e.g. amdgpu) this can
>   run for tens of milliseconds with preemption disabled, triggering the

It got the "preemption disabled" wrong. Well maybe when running
PREEMPT_NONE it is, but the description doesn't imply that.

-- Steve


>   documented soft lockup/panic during module load.
> - `ftrace_module_enable()` runs only in process context via
>   `prepare_coming_module()` (`kernel/module/main.c:3279`), so adding a
>   voluntary reschedule point is safe; the same pattern already exists in
>   other long-running ftrace loops (see commits d0b24b4e91fc and
>   42ea22e754ba), so this brings consistency without changing control
>   flow or semantics.
> - No data structures or interfaces change, and the code still executes
>   under the same locking (`ftrace_lock`, `text_mutex` when the arch
>   overrides `ftrace_arch_code_modify_prepare()`), so the risk of
>   regression is minimal: the new call simply yields CPU if needed while
>   keeping the locks held, preventing watchdog-induced crashes but
>   otherwise behaving identically.

On Sat, Oct 25, 2025 at 03:25:45PM -0400, Steven Rostedt wrote:
>On Sat, 25 Oct 2025 12:00:16 -0400
>Sasha Levin <sashal@kernel.org> wrote:
>
>> - The change inserts `cond_resched()` inside the inner iteration over
>>   every ftrace record (`kernel/trace/ftrace.c:7538`). That loop holds
>>   the ftrace mutex and, for each record, invokes heavy helpers like
>>   `test_for_valid_rec()` which in turn calls `kallsyms_lookup()`
>>   (`kernel/trace/ftrace.c:4289`). On huge modules (e.g. amdgpu) this can
>>   run for tens of milliseconds with preemption disabled, triggering the
>
>It got the "preemption disabled" wrong. Well maybe when running
>PREEMPT_NONE it is, but the description doesn't imply that.

Thanks for the review! I've been trying a new LLM for part of this series, and
it seems to underperform the one I was previously using.

-- 
Thanks,
Sasha