Messages in this thread |  | | Subject | Re: [RFC PATCH v7 5/7] libbpf: Implement BTF type sorting validation for binary search optimization | | From | Eduard Zingerman <> | | Date | Mon, 24 Nov 2025 10:16:48 -0800 |
| |
On Sat, 2025-11-22 at 23:45 +0800, Donglin Peng wrote: > On Sat, Nov 22, 2025 at 5:05 PM Eduard Zingerman <eddyz87@gmail.com> wrote: > > > > On Sat, 2025-11-22 at 00:50 -0800, Eduard Zingerman wrote: > > > > [...] > > > > > > Thanks. I’ve looked into find_btf_percpu_datasec and we can’t use > > > > btf_find_by_name_kind here because the search scope differs. For > > > > a module BTF, find_btf_percpu_datasec only searches within the > > > > module’s own BTF, whereas btf_find_by_name_kind prioritizes > > > > searching the base BTF first. Thus, placing named types ahead is > > > > more effective here. Besides, I found that the '.data..percpu' named > > > > type will be placed at [1] for vmlinux BTF because the prefix '.' is > > > > smaller than any letter, so the linear search only requires one loop to > > > > locate it. However, if we put named types at the end, it will need more > > > > than 60,000 loops.. > > > > > > But this can be easily fixed if a variant of btf_find_by_name_kind() > > > is provided that looks for a match only in a specific BTF. Or accepts > > > a start id parameter. > > > > Also, I double checked, and for my vmlinux the id for '.data..percpu' > > section is 110864, the last id of all. So, having all anonymous types > > in front does not change status-quo compared to current implementation. > > Yes. If types are sorted alphabetically, the '.data..percpu' section will > have ID 1 in vmlinux BTF. In this case, linear search performance is > optimal when named types are placed ahead of anonymous types. > > I would like to understand the benefits of having anonymous types at the > front of named types.
This will allow using strcmp() instead of btf_compare_type_names(), which we have to copy both in kernel and in libbpf. Reducing the code size and cognitive load.
|  |