On Tue, Apr 27, 2021 at 6:55 PM Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > > On Tue, Apr 27, 2021 at 10:45:38AM -0700, Andrii Nakryiko wrote: > > > > > > > > > + * If btf_fd is zero look for the name in vmlinux BTF and in module's BTFs. > > > > > + * Return > > > > > + * Returns btf_id and btf_obj_fd in lower and upper 32 bits. > > > > > > > > Mention that for vmlinux BTF btf_obj_fd will be zero? Also who "owns" > > > > the FD? If the BPF program doesn't close it, when are they going to be > > > > cleaned up? > > > > > > just like bpf_sys_bpf. Who owns returned FD? The program that called > > > the helper, of course. > > > > "program" as in the user-space process that did bpf_prog_test_run(), > > right? In the cover letter you mentioned that BPF_PROG_TYPE_SYSCALL > > might be called on syscall entry in the future, for that case there is > > no clear "owning" process, so would be curious to see how that problem > > gets solved. > > well, there is always an owner process. When syscall progs is attached > to syscall such FDs will be in the process that doing syscall. > It's kinda 'random', but that's the job of the prog to make 'non random'. > If it's doing syscalls that will install FDs it should have a reason > to do so. Likely there will be limitations on what bpf helpers such syscall > prog can do if it's attached to this or that syscall. > Currently it's test_run only. > > I'm not sure whether you're hinting that it all should be FD-less or I'm > putting a question in your mouth, but I've considered doing that and > figured that it's an overkill. It's possible to convert .*bpf.* do deal > with FDs and with some other temporary handle. Instead of map_fd the > loader prog would create a map and get a handle back that it will use > later in prog_load, etc. > But amount of refactoring looks excessive. > The generated loader prog should be correct by construction and > clean up after itself instead of burdening the kernel cleaning > those extra handles. That's not really the suggestion or question I had in mind. I was contemplating how the FD handling will happen if such BPF program is running from some other process's context and it seemed (and still seems) very surprising if new FD will just be added to a "random" process. Ignoring all the technical difficulties, I'd say ideally those FDs should be owned by BPF program itself, and when it gets unloaded, just like at the process exit, all still open FDs should be closed. How technically feasible that is is entirely different question. But basically, I wanted to confirm I understand where those new FDs are attached to.
