On Sat, Dec 18, 2021 at 7:18 PM Kumar Kartikeya Dwivedi
[off-list ref] wrote:

On Sun, Dec 19, 2021 at 07:58:39AM IST, Alexei Starovoitov wrote:

quoted

On Fri, Dec 17, 2021 at 07:20:27AM +0530, Kumar Kartikeya Dwivedi wrote:

quoted

diff --git a/include/linux/bpf_verifier.h b/include/linux/bpf_verifier.h
index b80fe5bf2a02..a6ef11db6823 100644
--- a/include/linux/bpf_verifier.h
+++ b/include/linux/bpf_verifier.h

@@ -128,6 +128,16 @@ struct bpf_reg_state {
     * allowed and has the same effect as bpf_sk_release(sk).
     */
    u32 ref_obj_id;
+   /* This is set for pointers which are derived from referenced
+    * pointer (e.g. PTR_TO_BTF_ID pointer walking), so that the
+    * pointers obtained by walking referenced PTR_TO_BTF_ID
+    * are appropriately invalidated when the lifetime of their
+    * parent object ends.
+    *
+    * Only one of ref_obj_id and parent_ref_obj_id can be set,
+    * never both at once.
+    */
+   u32 parent_ref_obj_id;

How would it handle parent of parent?

When you do:

r1 = acquire();

it gets ref_obj_id as N, then when you load r1->next, it does mark_btf_ld_reg
with reg->ref_obj_id ?: reg->parent_ref_obj_id, the latter is zero so it copies
ref, but into parent_ref_obj_id.

r2 = r1->next;

From here on, parent_ref_obj_id is propagated into all further mark_btf_ld_reg,
so if we do since ref_obj_id will be zero from previous mark_btf_ld_reg:

r3 = r2->next; // it will copy parent_ref_obj_id

I think it even works fine when you reach it indirectly, like foo->bar->foo,
if first foo is referenced.

... but maybe I missed some detail, do you see a problem in this approach?

quoted

Did you consider map_uid approach ?
Similar uid can be added for PTR_TO_BTF_ID.
Then every such pointer will be unique. Each deref will get its own uid.

I'll look into it, I didn't consider it before. My idea was to invalidate
pointers obtained from a referenced ptr_to_btf_id so I copied the same
ref_obj_id into parent_ref_obj_id, so that it can be matched during release. How
would that work in the btf_uid approach if they are unique? Do we copy the same
ref_obj_id into btf_uid? Then it's not very different except being btf_id ptr
specific state, right?

Or we can copy ref_obj_id and also set uid to disallow it from being released,
but still allow invalidation.

The goal is to disallow:
struct foo { struct foo *next; };

r1 = acquire(...); // BTF ID of struct foo
if (r1) {
        r2 = r1->next;
        release(r2);
}

right?
With btf_uid approach each deref gets its own uid.
r2 = r1->next
and
r3 = r1->next
will get different uids.
When type == PTR_TO_BTF_ID its reg->ref_obj_id will be considered
together with btf_uid.
Both ref_obj_id and btf_uid need to be the same.

But let's go back a bit.
Why ref_obj_id is copied on deref?
Shouldn't r2 get a different ref_obj_id after r2 = r1->next ?