[PATCH] kprobes: Fix permanent hang when flushing the kprobe optimizer

[Andrea Parri <parri.andrea@gmail.com>]

Writing 0 to /proc/sys/debug/kprobes-optimization while a kprobe is
jump-optimized never returns. The writer sleeps in D state forever with
kprobe_sysctl_mutex held, so any later read or write of that sysctl
hangs as well. For example, with vfs_read+9 as an optimizable address
in this build:

  # cd /sys/kernel/tracing
  # echo 'p:myprobe vfs_read+9' >> kprobe_events
  # echo 1 > events/kprobes/myprobe/enable
  # # wait until /sys/kernel/debug/kprobes/list shows [OPTIMIZED]
  # echo 0 > /proc/sys/debug/kprobes-optimization

  INFO: task sh:246 blocked for more than 10 seconds.
  Call Trace:
   <TASK>
   __schedule+0x1176/0x4f70
   schedule+0xdc/0x2c0
   schedule_timeout+0x17b/0x260
   wait_for_completion+0x173/0x3c0
   wait_for_kprobe_optimizer_locked+0xbc/0x130
   proc_kprobes_optimization_handler+0x156/0x1b0
   proc_sys_call_handler+0x324/0x490
   vfs_write+0x52d/0xfe0
   ksys_write+0xff/0x200
   do_syscall_64+0x106/0x630
   entry_SYSCALL_64_after_hwframe+0x77/0x7f
   </TASK>
  ...
  INFO: task cat:265 is blocked on a mutex likely owned by task sh:246.

wait_for_kprobe_optimizer_locked() reinitializes optimizer_completion,
asks the optimizer thread to flush and sleeps in wait_for_completion().
The thread drains the (un)optimizing lists, but calls complete() only
if completion_done() is true, i.e. if the completion is already done,
which never happens while someone waits. disarm_all_kprobes() and
kprobe_trace_self_tests_init() wait the same way.

Calling complete() unconditionally would not be enough: the waiter
drops kprobe_mutex while it sleeps, and nothing else serializes the
sysctl handler against the debugfs "enabled" file. A second flusher
that still finds the lists non-empty, e.g. because a disabled probe is
queued for unoptimizing, reinitializes the completion under the first:

  sysctl write                      debugfs "enabled" write
  unoptimize_all_kprobes()
    wait_for_kprobe_optimizer_locked()
      init_completion(c)
      mutex_unlock(&kprobe_mutex)
      wait_for_completion(c)
                                    disarm_all_kprobes()
                                      wait_for_kprobe_optimizer_locked()
                                        init_completion(c)
                                          // c->wait is reset, the first
                                          // waiter is off the queue
                                        mutex_unlock(&kprobe_mutex)
                                        wait_for_completion(c)
  kprobe_optimizer()
    complete(c)
      // wakes the debugfs writer only

where c is &optimizer_completion. Lining up the two writes during an
optimizer pass loses the sysctl writer this way.

Replace the completion with a counter of optimizer passes, bumped at the
end of each pass and signalled with wake_up_var_locked(), both under
kprobe_mutex. A flusher samples the count and waits with
wait_var_event_mutex(), which drops kprobe_mutex only while sleeping, so
a new count means a whole pass ran in the meantime. Nothing is
reinitialized, so several flushers can sleep in the wait at once.

Fixes: 73c12f209462 ("kprobes: Use dedicated kthread for kprobe optimizer")
Cc: stable@vger.kernel.org
Assisted-by: LLM
Signed-off-by: Andrea Parri <parri.andrea@gmail.com>
---
This overlaps with two patches under discussion in the "rcu-tasks: build
Tasks RCU on Tasks Trace readers in trampolines" thread:

 - Masami's RFC "kprobes: Make optprobe optimizer multi-generational and
   asynchronous" [1] rewrites the same code.  It drops the
   completion_done() check, which fixes the hang reported here, but it
   keeps init_completion() in wait_for_kprobe_optimizer_locked(), so the
   concurrent-flusher case described above remains.  The two patches
   conflict textually.

 - Josef's "kprobes: Expose the optprobe jump window to Tasks RCU" [2]
   changes kprobe_optimizer() around synchronize_rcu_tasks(); its
   kernel/kprobes.c hunks still apply on top of this patch.

[1] https://lore.kernel.org/all/179017148080.466588.9116221556625712980.stgit@devnote2/
[2] https://lore.kernel.org/all/20260922-b4-rcu-tasks-preempt-qs-v5-3-410f57770bad@toxicpanda.com/

 kernel/kprobes.c | 22 ++++++++++++++--------
 1 file changed, 14 insertions(+), 8 deletions(-)

diff --git a/kernel/kprobes.c b/kernel/kprobes.c
index 6337da5cab9e7..4edd8ca5c6578 100644
--- a/kernel/kprobes.c
+++ b/kernel/kprobes.c
@@ -42,6 +42,7 @@
 #include <linux/execmem.h>
 #include <linux/cleanup.h>
 #include <linux/wait.h>
+#include <linux/wait_bit.h>
 
 #include <asm/sections.h>
 #include <asm/cacheflush.h>
@@ -526,7 +527,8 @@ enum {
 	OPTIMIZER_ST_FLUSHING = 2,
 };
 
-static DECLARE_COMPLETION(optimizer_completion);
+/* Bumped at the end of each kprobe_optimizer() pass, under 'kprobe_mutex' */
+static unsigned long optimizer_passes;
 
 #define OPTIMIZE_DELAY 5
 
@@ -654,9 +656,9 @@ static void kprobe_optimizer(void)
 		do_free_cleaned_kprobes();
 	}
 
-	/* Step 5: Kick optimizer again if needed. But if there is a flush requested, */
-	if (completion_done(&optimizer_completion))
-		complete(&optimizer_completion);
+	/* Step 5: Wake up flushers, and kick optimizer again if needed. */
+	optimizer_passes++;
+	wake_up_var_locked(&optimizer_passes, &kprobe_mutex);
 
 	if (!list_empty(&optimizing_list) || !list_empty(&unoptimizing_list))
 		kick_kprobe_optimizer();	/*normal kick*/
@@ -708,7 +710,8 @@ static void wait_for_kprobe_optimizer_locked(void)
 	lockdep_assert_held(&kprobe_mutex);
 
 	while (!list_empty(&optimizing_list) || !list_empty(&unoptimizing_list)) {
-		init_completion(&optimizer_completion);
+		unsigned long passes = optimizer_passes;
+
 		/*
 		 * Set state to OPTIMIZER_ST_FLUSHING and wake up the thread if it's
 		 * idle. If it's already kicked, it will see the state change.
@@ -717,9 +720,12 @@ static void wait_for_kprobe_optimizer_locked(void)
 			OPTIMIZER_ST_FLUSHING) != OPTIMIZER_ST_FLUSHING)
 			wake_up(&kprobe_optimizer_wait);
 
-		mutex_unlock(&kprobe_mutex);
-		wait_for_completion(&optimizer_completion);
-		mutex_lock(&kprobe_mutex);
+		/*
+		 * kprobe_optimizer() holds 'kprobe_mutex' for a whole pass, which
+		 * this drops while sleeping, so a new count means a full pass ran.
+		 */
+		wait_var_event_mutex(&optimizer_passes,
+				     optimizer_passes != passes, &kprobe_mutex);
 	}
 }
 
-- 
2.53.0