On Wed, Feb 28, 2024 at 4:25 PM Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
>
> On 2/28/24 3:55 PM, Andrii Nakryiko wrote:
> > On Tue, Feb 27, 2024 at 6:11 PM Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
> >>
> >> On 2/27/24 12:45 PM, Eduard Zingerman wrote:
> >>> Make bpf_map__set_autocreate() for struct_ops maps toggle autoload
> >>> state for referenced programs.
> >>>
> >>> E.g. for the BPF code below:
> >>>
> >>> SEC("struct_ops/test_1") int BPF_PROG(foo) { ... }
> >>> SEC("struct_ops/test_2") int BPF_PROG(bar) { ... }
> >>>
> >>> SEC(".struct_ops.link")
> >>> struct test_ops___v1 A = {
> >>> .foo = (void *)foo
> >>> };
> >>>
> >>> SEC(".struct_ops.link")
> >>> struct test_ops___v2 B = {
> >>> .foo = (void *)foo,
> >>> .bar = (void *)bar,
> >>> };
> >>>
> >>> And the following libbpf API calls:
> >>>
> >>> bpf_map__set_autocreate(skel->maps.A, true);
> >>> bpf_map__set_autocreate(skel->maps.B, false);
> >>>
> >>> The autoload would be enabled for program 'foo' and disabled for
> >>> program 'bar'.
> >>>
> >>> Do not apply such toggling if program autoload state is set by a call
> >>> to bpf_program__set_autoload().
> >>>
> >>> Signed-off-by: Eduard Zingerman <eddyz87@xxxxxxxxx>
> >>> ---
> >>> tools/lib/bpf/libbpf.c | 35 ++++++++++++++++++++++++++++++++++-
> >>> 1 file changed, 34 insertions(+), 1 deletion(-)
> >>>
> >>> diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
> >>> index b39d3f2898a1..1ea3046724f8 100644
> >>> --- a/tools/lib/bpf/libbpf.c
> >>> +++ b/tools/lib/bpf/libbpf.c
> >>> @@ -446,13 +446,18 @@ struct bpf_program {
> >>> struct bpf_object *obj;
> >>>
> >>> int fd;
> >>> - bool autoload;
> >>> + bool autoload:1;
> >>> + bool autoload_user_set:1;
> >>> bool autoattach;
> >>> bool sym_global;
> >>> bool mark_btf_static;
> >>> enum bpf_prog_type type;
> >>> enum bpf_attach_type expected_attach_type;
> >>> int exception_cb_idx;
> >>> + /* total number of struct_ops maps with autocreate == true
> >>> + * that reference this program
> >>> + */
> >>> + __u32 struct_ops_refs;
> >>
> >> Instead of adding struct_ops_refs and autoload_user_set,
> >>
> >> for BPF_PROG_TYPE_STRUCT_OPS, how about deciding to load it or not by checking
> >> prog->attach_btf_id (non zero) alone. The prog->attach_btf_id is now decided at
> >> load time and is only set if it is used by at least one autocreate map, if I
> >> read patch 2 & 3 correctly.
> >>
> >> Meaning ignore prog->autoload*. Load the struct_ops prog as long as it is used
> >> by one struct_ops map with autocreate == true.
> >>
> >> If the struct_ops prog is not used in any struct_ops map, the bpf prog cannot be
> >> loaded even the autoload is set. If bpf prog is used in a struct_ops map and its
> >> autoload is set to false, the struct_ops map will be in broken state. Thus,
> >
> > We can easily detect this condition and report meaningful error.
> >
> >> prog->autoload does not fit very well with struct_ops prog and may as well
> >> depend on whether the struct_ops prog is used by a struct_ops map alone?
> >
> > I think it's probably fine from a usability standpoint to disable
> > loading the BPF program if its struct_ops map was explicitly set to
> > not auto-create. It's a bit of deviation from other program types, but
> > in practice this logic will make it easier for users.
> >
> > One question I have, though, is whether we should report as an error a
> > stand-alone struct_ops BPF program that is not used from any
> > struct_ops map? Or should we load it nevertheless? Or should we
> > silently not load it?
>
> I think the libbpf could report an error if the prog is not used in any
> struct_ops map at the source code level, not sure if it is useful.
>
> However, it probably should not report error if that struct_ops map (where the
> prog is resided) does not have autocreate set to true.
>
> If a BPF program is not used in any struct_ops map, it cannot be loaded anyway
> because the prog->attach_btf_id is not set. If libbpf tries to load the prog,
> the kernel will reject it also. I think it may be a question on whether it is
> the user intention of not loading the prog if the prog is not used in any
> struct_ops map. I tend to think it is the user intention of not loading it in
> this case.
>
> SEC("struct_ops/test1")
> int BPF_PROG(test1) { ... }
>
> SEC("struct_ops/test2")
> int BPF_PROG(test2) { ... }
>
> SEC("?.struct_ops.link") struct some_ops___v1 a = { .test1 = test1 }
> SEC("?.struct_ops.link") struct some_ops___v2 b = { .test1 = test1,
> .test2 = test2, }
>
> In the above, the userspace may try to load with a newer some_ops___v2 first,
> failed and then try a lower version some_ops___v1 and then succeeded. The test2
> prog will not be used and not expected to be loaded.
>
Yes, it's all sane in the above example. But imagine a stand-alone
struct_ops program with no SEC(".struct_ops") at all:
SEC("struct_ops/test1")
int BPF_PROG(test1) { ... }
/* nothing else */
Currently this will fail, right?
And with your proposal it will succeed without actually even
attempting to load the BPF program. Or am I misunderstanding?
> >
> > I feel like silently not loading is the worst behavior here. So either
> > loading it anyway or reporting an error would be my preference,
> > probably.
>
