[PATCH] kernel/trace: fix possible deadlock in trie_delete_elem

[Wojciech Gładysz]

On bpf syscall map operations the bpf_disable_instrumentation function
is called for the reason described in the comment to the function.
The description matches the bug case. The function increments a per CPU
integer variable bpf_prog_active. The variable is not processed in the
bpf trace path. The fix implements a similar processing as for kprobe
handling. The fix degrades the bpf tracing by skipping some eBPF trace
sequences that otherwise might yield deadlock.

Reported-by: syzbot+9d95beb2a3c260622518@xxxxxxxxxxxxxxxxxxxxxxxxx
Closes: https://syzkaller.appspot.com/bug?extid=9d95beb2a3c260622518
Link: https://lore.kernel.org/all/000000000000adb08b061413919e@xxxxxxxxxx/T/
Signed-off-by: Wojciech Gładysz <wojciech.gladysz@xxxxxxxxxxxx>
---
 kernel/trace/bpf_trace.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
index 6249dac61701..8de2e084b162 100644
--- a/kernel/trace/bpf_trace.c
+++ b/kernel/trace/bpf_trace.c
@@ -2391,7 +2391,9 @@ void __bpf_trace_run(struct bpf_raw_tp_link *link, u64 *args)
 	struct bpf_trace_run_ctx run_ctx;
 
 	cant_sleep();
-	if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
+
+	// if the instrumentation is not disabled disable recurrence and go
+	if (unlikely(__this_cpu_inc_return(bpf_prog_active) != 1)) {
 		bpf_prog_inc_misses_counter(prog);
 		goto out;
 	}
@@ -2405,7 +2407,7 @@ void __bpf_trace_run(struct bpf_raw_tp_link *link, u64 *args)
 
 	bpf_reset_run_ctx(old_run_ctx);
 out:
-	this_cpu_dec(*(prog->active));
+	__this_cpu_dec(bpf_prog_active);
 }
 
 #define UNPACK(...)			__VA_ARGS__
-- 
2.35.3