bpf: bpf_event_entry_gen's alloc needs to be in atomic context

Should have been obvious, only called from bpf() syscall via map_update_elem()
that calls bpf_fd_array_map_update_elem() under RCU read lock and thus this
must also be in GFP_ATOMIC, of course.

Fixes: 3b1efb1 ("bpf, maps: flush own entries on perf map release")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

kernel/bpf/arraymap.c

@@ -430,7 +430,7 @@ static struct bpf_event_entry *bpf_event_entry_gen(struct file *perf_file,
 {
 	struct bpf_event_entry *ee;
 
-	ee = kzalloc(sizeof(*ee), GFP_KERNEL);
+	ee = kzalloc(sizeof(*ee), GFP_ATOMIC);
 	if (ee) {
 		ee->event = perf_file->private_data;
 		ee->perf_file = perf_file;