On Wed, Mar 10, 2021 at 6:34 PM Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > > On Tue, Mar 09, 2021 at 08:04:26PM -0800, Andrii Nakryiko wrote: > > + > > + struct btf *strtab_btf; /* we use struct btf to manage strings */ > ... > > + str_off = btf__add_str(linker->strtab_btf, sec->sec_name); > > + sec->shdr->sh_name = str_off; > > That bit took me an hour to grok. > That single line comment above is far far from obvious. Heh, I guess I've been working with BTF, ELF and pahole for too long to notice that it's so non-obvious. pahole wraps `struct btf` in a similar fashion for deduplicated string management. > What the logic is relying on is that string section in BTF format > has the same zero terminated set of strings as ELF's .strtab section. > There is no BTF anywhere here in this 'strtab_btf'. > The naming choice made it double hard. Right. strtab_strs would probably be a slightly better choice. > My understanding that you're using that instead of renaming btf_add_mem() > into something generic to rely on string hashmap for string dedup? It's not about renaming btf_add_mem(). btf_add_mem() just implements memory re-allocation (with exponential increase). But here we want to not add a new string if it's already present. So it's much more complicated logic than btf_add_mem(). > > The commit log in patch 2 that introduces btf_raw_strs() sort of talks about > this code puzzle, but I would never guessed that's what you meant based > on patch 2 alone. > > Did you consider some renaming/generalizing of string management to > avoid btf__add_str() through out the patch 5? > The "btf_" prefix makes things challenging to read. > Especially when patch 6 is using btf__add_str() to add to real BTF. Right. I guess we can extract the "set of strings" data structure out of `struct btf` into libbpf-internal data structure. Then use it from struct btf and separately (and directly) from struct bpf_linker. I'll see what that would involve in terms of refactoring. > > Mainly pointing it out for others who might be looking at the patches. That's a good point, I should have probably at least mentioned that bit more explicitly.
