On Fri, Dec 18, 2020 at 03:56:14PM -0800, Andrii Nakryiko wrote:

+
+/* shuffled layout for relocatable (CO-RE) reads */
+struct callback_head___shuffled {
+	void (*func)(struct callback_head___shuffled *head);
+	struct callback_head___shuffled *next;
+};
+
+struct callback_head k_probe_in = {};
+struct callback_head___shuffled k_core_in = {};
+
+struct callback_head *u_probe_in = 0;
+struct callback_head___shuffled *u_core_in = 0;
+
+long k_probe_out = 0;
+long u_probe_out = 0;
+
+long k_core_out = 0;
+long u_core_out = 0;
+
+int my_pid = 0;
+
+SEC("raw_tracepoint/sys_enter")
+int handler(void *ctx)
+{
+	int pid = bpf_get_current_pid_tgid() >> 32;
+
+	if (my_pid != pid)
+		return 0;
+
+	/* next pointers for kernel address space have to be initialized from
+	 * BPF side, user-space mmaped addresses are stil user-space addresses
+	 */
+	k_probe_in.next = &k_probe_in;
+	__builtin_preserve_access_index(({k_core_in.next = &k_core_in;}));
+
+	k_probe_out = (long)BPF_PROBE_READ(&k_probe_in, next, next, func);
+	k_core_out = (long)BPF_CORE_READ(&k_core_in, next, next, func);
+	u_probe_out = (long)BPF_PROBE_READ_USER(u_probe_in, next, next, func);
+	u_core_out = (long)BPF_CORE_READ_USER(u_core_in, next, next, func);

I don't understand what the test suppose to demonstrate.
co-re relocs work for kernel btf only.
Are you saying that 'struct callback_head' happened to be used by user space
process that allocated it in user memory. And that is the same struct as
being used by the kernel? So co-re relocs that apply against the kernel
will sort-of work against the data of user space process because
the user space is using the same struct? That sounds convoluted.
I struggle to see the point of patch 1:
+#define bpf_core_read_user(dst, sz, src)                                   \
+       bpf_probe_read_user(dst, sz, (const void *)__builtin_preserve_access_index(src))

co-re for user structs? Aren't they uapi? No reloc is needed.