On Mon, Feb 26, 2024 at 1:48 PM Bryce Kahle <bryce.kahle@xxxxxxxxxxxxx> wrote: > > On Fri, Feb 2, 2024, at 2:10 PM, Andrii Nakryiko wrote: > > But yes, you'd have to specify both vmlinux and all the module > > BTFs at the same time (which bpftool allows you to do easily with its > > CLI interface, so not really a problem) > > I didn't see a way to specify a directory for vmlinux and all the > modules BTFs. Are you suggesting I specify the path to each > individually? I didn't see a way to do that with the current CLI api. > It assumes that the input is only a single path. so right now we have bpftool min_core_btf <input-btf> <output-btf> <input1.bpf.o> ... <inputN.bpf.o> so we'd have to either add a flag and do bpftool min_core_btf <input-btf> -E <extra-btf1> -E <extra-btf2> <output-btf> ... or define special key/value pair (we do that for other commands to specify extra options): bpftool min_core_btf <input-btf> extra <extra-btf-1> extra <extra-btf-2> <output-btf> .... This has a tiny chance that user used "extra" as a name of one of input object file (we can probably disregard). Yet another option is to introduce new command, something like `bpftool min_core_btf_multi ...` and define new convention. OR. We can pivot and say that we do what you want as two steps: 1) generate one large combined BTF from multiple BTFs, something along the lines of `bpftool btf merge <btf1> ... <btfN>`. We'd need to specify how split BTF should be handled. 2) then use existing min_core_btf command with this merged BTF I don't know what's best. > > On Mon, Feb 5, 2024 at 10:21 AM Andrii Nakryiko > <andrii.nakryiko@xxxxxxxxx> wrote: > > 2) append each module BTF to that instance (we have btf__add_btf() API > > already for that). This will rewrite type IDs for each module > > (shifting them by some constant number) > > It looks like btf__add_btf() doesn't support split BTF, which the > module BTF has to be in order for it to parse correctly. Any > suggestions for how to proceed? Do I need to add support for split BTF > to btf__add_btf() to libbpf? If so, any thoughts on how that should > work would be appreciated. Yep, seems like it is explicitly not supported. I think one of the problems with split BTF is that you don't want to append shared types from base BTF, because that would be a waste. So you need a variant of btf__add_btf() that allows you to specify a range of types to append. But at that point what to do if some of the added types reference types that were not appended? A bunch of unpleasant issues to be dealt with. So perhaps instead of using btf__add_btf(), bpftool can just do btf__add_type() API, which remaps strings properly, but doesn't touch type IDs. Libbpf has internal btf_type_visit_type_ids() function that calls provided callback for each field that contains type ID. bpftool can do its own remapping. If we assume that we are merging vmlinux BTF and a bunch of module BTFs, then this remapping is actually pretty straightforward: if type ID belongs to base BTF, don't touch it. Otherwise shift it by some amount, common for each type within the module's BTF.
