bpf: fix PerfMap::output() API when appending context data

map.output(&ctx, &data, nr_bytes) can be used to append nr_bytes from
the current context to the &data event.

bpf/aya-bpf/src/maps/perf_map.rs

@@ -32,8 +32,8 @@ impl<T> PerfMap<T> {
         }
     }
 
-    pub fn output<C: BpfContext>(&mut self, ctx: &C, data: &T) {
+    pub fn output<C: BpfContext>(&mut self, ctx: &C, data: &T, flags: u32) {
-        self.output_at_index(ctx, None, data, 0)
+        self.output_at_index(ctx, None, data, flags)
     }
 
     pub fn output_at_index<C: BpfContext>(
@@ -43,10 +43,8 @@ impl<T> PerfMap<T> {
         data: &T,
         flags: u32,
     ) {
-        let index = index
+        let index = index.unwrap_or(BPF_F_CURRENT_CPU.into()) as u64;
-            .map(|i| (i as u64) << 32)
+        let flags = (flags as u64) << 32 | index;
-            .unwrap_or(BPF_F_CURRENT_CPU.into());
-        let flags = index | flags as u64;
         unsafe {
             bpf_perf_event_output(
                 ctx.as_ptr(),