On Mon, Nov 2, 2020 at 5:26 PM Martin KaFai Lau [off-list ref] wrote:

On Sat, Oct 31, 2020 at 11:52:37AM -0700, Alexander Duyck wrote:
[ ... ]

quoted

+struct tcpbpf_globals global = { 0 };
 int _version SEC("version") = 1;

 SEC("sockops")

@@ -105,29 +72,15 @@ int bpf_testcb(struct bpf_sock_ops *skops)

      op = (int) skops->op;

-     update_event_map(op);
+     global.event_map |= (1 << op);

      switch (op) {
      case BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB:
              /* Test failure to set largest cb flag (assumes not defined) */
-             bad_call_rv = bpf_sock_ops_cb_flags_set(skops, 0x80);
+             global.bad_cb_test_rv = bpf_sock_ops_cb_flags_set(skops, 0x80);
              /* Set callback */
-             good_call_rv = bpf_sock_ops_cb_flags_set(skops,
+             global.good_cb_test_rv = bpf_sock_ops_cb_flags_set(skops,
                                               BPF_SOCK_OPS_STATE_CB_FLAG);
-             /* Update results */
-             {
-                     __u32 key = 0;
-                     struct tcpbpf_globals g, *gp;
-
-                     gp = bpf_map_lookup_elem(&global_map, &key);
-                     if (!gp)
-                             break;
-                     g = *gp;
-                     g.bad_cb_test_rv = bad_call_rv;
-                     g.good_cb_test_rv = good_call_rv;
-                     bpf_map_update_elem(&global_map, &key, &g,
-                                         BPF_ANY);
-             }
              break;
      case BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB:
              skops->sk_txhash = 0x12345f;

@@ -143,10 +96,8 @@ int bpf_testcb(struct bpf_sock_ops *skops)

                              thdr = (struct tcphdr *)(header + offset);
                              v = thdr->syn;
-                             __u32 key = 1;

-                             bpf_map_update_elem(&sockopt_results, &key, &v,
-                                                 BPF_ANY);
+                             global.tcp_saved_syn = v;
                      }
              }
              break;

@@ -156,25 +107,16 @@ int bpf_testcb(struct bpf_sock_ops *skops)
              break;
      case BPF_SOCK_OPS_STATE_CB:
              if (skops->args[1] == BPF_TCP_CLOSE) {
-                     __u32 key = 0;
-                     struct tcpbpf_globals g, *gp;
-
-                     gp = bpf_map_lookup_elem(&global_map, &key);
-                     if (!gp)
-                             break;
-                     g = *gp;
                      if (skops->args[0] == BPF_TCP_LISTEN) {
-                             g.num_listen++;
+                             global.num_listen++;
                      } else {
-                             g.total_retrans = skops->total_retrans;
-                             g.data_segs_in = skops->data_segs_in;
-                             g.data_segs_out = skops->data_segs_out;
-                             g.bytes_received = skops->bytes_received;
-                             g.bytes_acked = skops->bytes_acked;
+                             global.total_retrans = skops->total_retrans;
+                             global.data_segs_in = skops->data_segs_in;
+                             global.data_segs_out = skops->data_segs_out;
+                             global.bytes_received = skops->bytes_received;
+                             global.bytes_acked = skops->bytes_acked;
                      }
-                     g.num_close_events++;
-                     bpf_map_update_elem(&global_map, &key, &g,
-                                         BPF_ANY);

It is interesting that there is no race in the original "g.num_close_events++"
followed by the bpf_map_update_elem().  It seems quite fragile though.

How would it race with the current code though? At this point we are
controlling the sockets in a single thread. As such the close events
should already be serialized shouldn't they? This may have been a
problem with the old code, but even then it was only two sockets so I
don't think there was much risk of them racing against each other
since the two sockets were linked anyway.

quoted

+                     global.num_close_events++;

There is __sync_fetch_and_add().

not sure about the global.event_map though, may be use an individual
variable for each _CB.  Thoughts?

I think this may be overkill for what we actually need. Since we are
closing the sockets in a single threaded application there isn't much
risk of the sockets all racing against each other in the close is
there?