On 1/7/20 12:46 AM, Alexei Starovoitov wrote:

On Sun, Dec 29, 2019 at 03:37:40PM +0100, Jiri Olsa wrote:

quoted

When unwinding the stack we need to identify each
address to successfully continue. Adding latch tree
to keep trampolines for quick lookup during the
unwind.

Signed-off-by: Jiri Olsa <jolsa@kernel.org>

...

quoted

+bool is_bpf_trampoline(void *addr)
+{
+	return latch_tree_find(addr, &tree, &tree_ops) != NULL;
+}
+
  struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
  {
  	struct bpf_trampoline *tr;

@@ -65,6 +98,7 @@ struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
  	for (i = 0; i < BPF_TRAMP_MAX; i++)
  		INIT_HLIST_HEAD(&tr->progs_hlist[i]);
  	tr->image = image;
+	latch_tree_insert(&tr->tnode, &tree, &tree_ops);

Thanks for the fix. I was thinking to apply it, but then realized that bpf
dispatcher logic has the same issue.
Could you generalize the fix for both?
May be bpf_jit_alloc_exec_page() can do latch_tree_insert() ?
and new version of bpf_jit_free_exec() is needed that will do latch_tree_erase().
Wdyt?

Also this patch is buggy since your latch lookup happens under RCU, but
I don't see anything that waits a grace period once you remove from the
tree. Instead you free the trampoline right away.

On a different question, given we have all the kallsym infrastructure
for BPF already in place, did you look into whether it's feasible to
make it a bit more generic to also cover JITed buffers from trampolines?