bperf: Do not use BPF_CORE_READ (facebookincubator#196)

Summary:

Use direct BTF access to be faster.

We need bpf_rdonly_cast() here, to cast "struct cgroup_subsys_state *"
to "struct cgroup *"

Reviewed By: anakryiko

Differential Revision: D51837046

hbt/src/perf_event/bpf/bperf_leader_cgroup.bpf.c

@@ -51,9 +51,11 @@ int cpu_cnt = 0;
 
 #define PF_IDLE 0x00000002
 
+void *bpf_rdonly_cast(void *, __u32) __ksym __weak;
+
 static void update_cgroup_output(struct bpf_perf_event_value* diff_val) {
   struct task_struct* task = bpf_get_current_task_btf();
-  struct cgroup* cgrp = BPF_CORE_READ(task, cgroups, dfl_cgrp);
+  struct cgroup* cgrp = task->cgroups->dfl_cgrp;
   __u32 i, level;
 
   /* sched_switch tracepoint is triggered before current task is updated.
@@ -65,10 +67,11 @@ static void update_cgroup_output(struct bpf_perf_event_value* diff_val) {
     return;
 
   for (level = 0; level < MAX_CGROUP_LEVELS; level++) {
-    __u64 id = BPF_CORE_READ(cgrp, kn, id);
+    __u64 id = cgrp->kn->id;
     struct bpf_perf_event_value* val;
 
-    cgrp = (struct cgroup*)BPF_CORE_READ(cgrp, self.parent);
+    cgrp = bpf_rdonly_cast(cgrp->self.parent,
+                           bpf_core_type_id_kernel(struct cgroup));
     if (cgrp == NULL)
       break;