On Mon, Apr 26, 2021 at 8:00 PM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Mon, Apr 26, 2021 at 10:29:09AM -0700, Andrii Nakryiko wrote:
> > On Thu, Apr 22, 2021 at 5:27 PM Alexei Starovoitov
> > <alexei.starovoitov@xxxxxxxxx> wrote:
> > >
> > > From: Alexei Starovoitov <ast@xxxxxxxxxx>
> > >
> > > In order to be able to generate loader program in the later
> > > patches change the order of data and text relocations.
> > > Also improve the test to include data relos.
> > >
> > > If the kernel supports "FD array" the map_fd relocations can be processed
> > > before text relos since generated loader program won't need to manually
> > > patch ld_imm64 insns with map_fd.
> > > But ksym and kfunc relocations can only be processed after all calls
> > > are relocated, since loader program will consist of a sequence
> > > of calls to bpf_btf_find_by_name_kind() followed by patching of btf_id
> > > and btf_obj_fd into corresponding ld_imm64 insns. The locations of those
> > > ld_imm64 insns are specified in relocations.
> > > Hence process all data relocations (maps, ksym, kfunc) together after call relos.
> > >
> > > Signed-off-by: Alexei Starovoitov <ast@xxxxxxxxxx>
> > > ---
> > > tools/lib/bpf/libbpf.c | 86 +++++++++++++++----
> > > .../selftests/bpf/progs/test_subprogs.c | 13 +++
> > > 2 files changed, 80 insertions(+), 19 deletions(-)
> > >
> > > diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
> > > index 17cfc5b66111..c73a85b97ca5 100644
> > > --- a/tools/lib/bpf/libbpf.c
> > > +++ b/tools/lib/bpf/libbpf.c
> > > @@ -6379,11 +6379,15 @@ bpf_object__relocate_data(struct bpf_object *obj, struct bpf_program *prog)
> > > insn[0].imm = ext->ksym.kernel_btf_id;
> > > break;
> > > case RELO_SUBPROG_ADDR:
> > > - insn[0].src_reg = BPF_PSEUDO_FUNC;
> > > - /* will be handled as a follow up pass */
> > > + if (insn[0].src_reg != BPF_PSEUDO_FUNC) {
> > > + pr_warn("prog '%s': relo #%d: bad insn\n",
> > > + prog->name, i);
> > > + return -EINVAL;
> > > + }
> >
> > given SUBPROG_ADDR is now handled similarly to RELO_CALL in a
> > different place, I'd probably drop this error check and just combine
> > RELO_SUBPROG_ADDR and RELO_CALL cases with just a /* handled already
> > */ comment.
>
> I prefer to keep them separate. I've hit this pr_warn couple times
> while messing with relos and it saved my time.
> I bet it will save time to the next developer too.
hmm.. ok, not critical to me
>
> > > + /* handled already */
> > > break;
> > > case RELO_CALL:
> > > - /* will be handled as a follow up pass */
> > > + /* handled already */
> > > break;
> > > default:
> > > pr_warn("prog '%s': relo #%d: bad relo type %d\n",
> > > @@ -6552,6 +6556,31 @@ static struct reloc_desc *find_prog_insn_relo(const struct bpf_program *prog, si
> > > sizeof(*prog->reloc_desc), cmp_relo_by_insn_idx);
> > > }
> > >
> > > +static int append_subprog_relos(struct bpf_program *main_prog, struct bpf_program *subprog)
> > > +{
> > > + int new_cnt = main_prog->nr_reloc + subprog->nr_reloc;
> > > + struct reloc_desc *relos;
> > > + size_t off = subprog->sub_insn_off;
> > > + int i;
> > > +
> > > + if (main_prog == subprog)
> > > + return 0;
> > > + relos = libbpf_reallocarray(main_prog->reloc_desc, new_cnt, sizeof(*relos));
> > > + if (!relos)
> > > + return -ENOMEM;
> > > + memcpy(relos + main_prog->nr_reloc, subprog->reloc_desc,
> > > + sizeof(*relos) * subprog->nr_reloc);
> > > +
> > > + for (i = main_prog->nr_reloc; i < new_cnt; i++)
> > > + relos[i].insn_idx += off;
> >
> > nit: off is used only here, so there is little point in having it as a
> > separate var, inline?
>
> sure.
>
> > > + /* After insn_idx adjustment the 'relos' array is still sorted
> > > + * by insn_idx and doesn't break bsearch.
> > > + */
> > > + main_prog->reloc_desc = relos;
> > > + main_prog->nr_reloc = new_cnt;
> > > + return 0;
> > > +}
> > > +
> > > static int
> > > bpf_object__reloc_code(struct bpf_object *obj, struct bpf_program *main_prog,
> > > struct bpf_program *prog)
> > > @@ -6560,18 +6589,32 @@ bpf_object__reloc_code(struct bpf_object *obj, struct bpf_program *main_prog,
> > > struct bpf_program *subprog;
> > > struct bpf_insn *insns, *insn;
> > > struct reloc_desc *relo;
> > > - int err;
> > > + int err, i;
> > >
> > > err = reloc_prog_func_and_line_info(obj, main_prog, prog);
> > > if (err)
> > > return err;
> > >
> > > + for (i = 0; i < prog->nr_reloc; i++) {
> > > + relo = &prog->reloc_desc[i];
> > > + insn = &main_prog->insns[prog->sub_insn_off + relo->insn_idx];
> > > +
> > > + if (relo->type == RELO_SUBPROG_ADDR)
> > > + /* mark the insn, so it becomes insn_is_pseudo_func() */
> > > + insn[0].src_reg = BPF_PSEUDO_FUNC;
> > > + }
> > > +
> >
> > This will do the same work over and over each time we append a subprog
> > to main_prog. This should logically follow append_subprog_relos(), but
> > you wanted to do it for main_prog with the same code, right?
>
> It cannot follow append_subprog_relos.
> It has to be done before the loop below.
> Otherwise !insn_is_pseudo_func() won't catch it and all ld_imm64 insns
> will be considered which will make the loop below more complex and slower.
> The find_prog_insn_relo() will be called a lot more times.
> !relo condition would be treated different ld_imm64 vs call insn, etc.
if you process main_prog->insns first all the calls to subprogs would
be updated, then it recursively would do this right before a new
subprog is added (initial call is bpf_object__reloc_code(obj, prog,
prog) where prog is entry-point program). But either way I'm not
suggesting doing this and splitting this logic into two places.
>
> > How about instead doing this before we start appending subprogs to
> > main_progs? I.e., do it explicitly in bpf_object__relocate() before
> > you start code relocation loop.
>
> Not sure I follow.
> Do another loop:
> for (i = 0; i < obj->nr_programs; i++)
> for (i = 0; i < prog->nr_reloc; i++)
> if (relo->type == RELO_SUBPROG_ADDR)
> ?
> That's an option too.
> I can do that if you prefer.
> It felt cleaner to do this mark here right before the loop below that needs it.
Yes, I'm proposing to do another loop in bpf_object__relocate() before
we start adding subprogs to main_progs. The reason is that
bpf_object__reloc_code() is called recursively many times for the same
main_prog, so doing that here is O(N^2) in the number of total
instructions in main_prog. It processes the same (already processed)
instructions many times unnecessarily. It's wasteful and unclean.
>
> > > for (insn_idx = 0; insn_idx < prog->sec_insn_cnt; insn_idx++) {
> > > insn = &main_prog->insns[prog->sub_insn_off + insn_idx];
> > > if (!insn_is_subprog_call(insn) && !insn_is_pseudo_func(insn))
> > > continue;
> > >
> > > relo = find_prog_insn_relo(prog, insn_idx);
> > > + if (relo && relo->type == RELO_EXTERN_FUNC)
> > > + /* kfunc relocations will be handled later
> > > + * in bpf_object__relocate_data()
> > > + */
> > > + continue;
> > > if (relo && relo->type != RELO_CALL && relo->type != RELO_SUBPROG_ADDR) {
> > > pr_warn("prog '%s': unexpected relo for insn #%zu, type %d\n",
> > > prog->name, insn_idx, relo->type);
> >
> > [...]
>
> --
