On Wed, Mar 5, 2025 at 6:59 PM Menglong Dong [off-list ref] wrote:

I'm not sure if it works. However, indirect call is also used
in function graph, so we still have better performance. Isn't it?

Let me have a look at the code of the function graph first :/

Menglong,

Function graph infra isn't going to help.
"call foo" isn't a problem either.

But we have to step back.
per-function metadata is an optimization and feels like
we're doing a premature optimization here without collecting
performance numbers first.

Let's implement multi-fentry with generic get_metadata_by_ip() first.
get_metadata_by_ip() will be a hashtable in such a case and
then we can compare its performance when it's implemented as
a direct lookup from ip-4 (this patch) vs hash table
(that does 'ip' to 'metadata' lookup).

If/when we decide to do this per-function metadata we can also
punt to generic hashtable for cfi, IBT, FineIBT, etc configs.
When mitigations are enabled the performance suffers anyway,
so hashtable lookup vs direct ip-4 lookup won't make much difference.
So we can enable per-function metadata only on non-mitigation configs
when FUNCTION_ALIGNMENT=16.
There will be some number of bytes available before every function
and if we can tell gcc/llvm to leave at least 5 bytes there
the growth of vmlinux .text will be within a noise.

So let's figure out the design of multi-fenty first with a hashtable
for metadata and decide next steps afterwards.