Re: [PATCH] tracing: uprobes: Fix slab use-after-free in filter_chain

On Thu, Sep 10, 2026 at 02:48:49AM +0000, sashiko-bot@kernel.org wrote:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [Critical] Race condition in unregister_trace_uprobe leads to use-after-free (UAF) of trace_uprobe during lockless read accesses.
--

commit 04fcafa318e6d04f74e677950139db1dd3bf079c
Author: Zafir Rasyidi Taufik [off-list ref]

tracing: uprobes: Fix slab use-after-free in filter_chain

This commit addresses a use-after-free in uprobe_mmap. It resolves an issue
where unregister_trace_uprobe() would fail to unlink the consumer when a
trace_probe has a sibling event.

quoted

diff --git a/kernel/trace/trace_uprobe.c b/kernel/trace/trace_uprobe.c
index 861d857adadb3..110ec7b78b3dc 100644
--- a/kernel/trace/trace_uprobe.c
+++ b/kernel/trace/trace_uprobe.c

@@ -408,6 +408,11 @@ static int unregister_trace_uprobe(struct trace_uprobe *tu)
 		return ret;
 
 unreg:
+	if (tu->uprobe) {
+		uprobe_unregister_nosync(tu->uprobe, &tu->consumer);
+		tu->uprobe = NULL;
+		uprobe_unregister_sync();
+	}
 	dyn_event_remove(&tu->devent);
 	trace_probe_unlink(&tu->tp);
 	free_trace_uprobe(tu);

[Severity: Critical]
This is a pre-existing issue, but does this code leave a window for a
use-after-free of trace_uprobe during lockless read accesses?

When a sibling probe exists, this code unlinks and immediately frees the
trace probe without synchronizing with trace readers via RCU or shared locks.

For example, when a user queries an attached perf event, it initiates a
lockless traversal:

bpf_task_fd_query()
  bpf_get_perf_event_info()
    bpf_get_uprobe_info()
      trace_uprobe_primary_from_call()

This accesses the trace probe structure without acquiring event_mutex or
trace_event_sem.

Additionally, even if trace_event_sem is held for read, such as during
print_uprobe_event(), unregister_trace_uprobe() bypasses taking
trace_event_sem for write when unregistering a sibling probe (it skips
unregister_uprobe_event() and jumps directly to unreg).

At the same time, if unregister_trace_uprobe() is called, it unlinks the
probe via trace_probe_unlink(). The list removal relies on standard
list_del_init() instead of list_del_rcu(), allowing concurrent lockless
readers to traverse into soon-to-be-freed nodes when free_trace_uprobe(tu)
is called immediately after.

Could this lead to memory corruption when readers access the freed probe?

I will investigate furthur regarding this. If it's correct I will think
of a patch and send it separately.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260910023106.1795811-3-zafir.taufik@starlabs.sg?part=1