Re: debug annotations for bpf progs. Was: [PATCH bpf-next 1/3] bpf: preserve command of the process that loaded the program

[Stanislav Fomichev <sdf@fomichev.me>]

On 10/16, Daniel Borkmann wrote:
> On Tue, Oct 15, 2019 at 02:21:50PM -0700, Alexei Starovoitov wrote:
> > On Fri, Oct 11, 2019 at 5:38 PM Stanislav Fomichev <sdf@fomichev.me> wrote:
> > > On 10/11, Alexei Starovoitov wrote:
> > > > On Fri, Oct 11, 2019 at 9:21 AM Stanislav Fomichev <sdf@google.com> wrote:
> > > > >
> > > > > Even though we have the pointer to user_struct and can recover
> > > > > uid of the user who has created the program, it usually contains
> > > > > 0 (root) which is not very informative. Let's store the comm of the
> > > > > calling process and export it via bpf_prog_info. This should help
> > > > > answer the question "which process loaded this particular program".
> > > > >
> > > > > Signed-off-by: Stanislav Fomichev <sdf@google.com>
> > > > > ---
> > > > >  include/linux/bpf.h      | 1 +
> > > > >  include/uapi/linux/bpf.h | 2 ++
> > > > >  kernel/bpf/syscall.c     | 4 ++++
> > > > >  3 files changed, 7 insertions(+)
> > > > >
> > > > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > > > > index 5b9d22338606..b03ea396afe5 100644
> > > > > --- a/include/linux/bpf.h
> > > > > +++ b/include/linux/bpf.h
> > > > > @@ -421,6 +421,7 @@ struct bpf_prog_aux {
> > > > >                 struct work_struct work;
> > > > >                 struct rcu_head rcu;
> > > > >         };
> > > > > +       char created_by_comm[BPF_CREATED_COMM_LEN];
> > > > >  };
> > > > >
> > > > >  struct bpf_array {
> > > > > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> > > > > index a65c3b0c6935..4e883ecbba1e 100644
> > > > > --- a/include/uapi/linux/bpf.h
> > > > > +++ b/include/uapi/linux/bpf.h
> > > > > @@ -326,6 +326,7 @@ enum bpf_attach_type {
> > > > >  #define BPF_F_NUMA_NODE                (1U << 2)
> > > > >
> > > > >  #define BPF_OBJ_NAME_LEN 16U
> > > > > +#define BPF_CREATED_COMM_LEN   16U
> > > >
> > > > Nack.
> > > > 16 bytes is going to be useless.
> > > > We found it the hard way with prog_name.
> > > > If you want to embed additional debug information
> > > > please use BTF for that.
> > > BTF was my natural choice initially, but then I saw created_by_uid and
> > > thought created_by_comm might have a chance :-)
> > >
> > > To clarify, by BTF you mean creating some unused global variable
> > > and use its name as the debugging info? Or there is some better way?
> > 
> > I was thinking about adding new section to .btf.ext with this extra data,
> > but global variable is a better idea indeed.
> > We'd need to standardize such variables names, so that
> > bpftool can parse and print it while doing 'bpftool prog show'.
> 
> +1, much better indeed.
> 
> > We see more and more cases where services use more than
> > one program in single .c file to accomplish their goals.
> > Tying such debug info (like 'created_by_comm') to each program
> > individually isn't quite right.
> > In that sense global variables are better, since they cover the
> > whole .c file.
> > Beyond 'created_by_comm' there are others things that people
> > will likely want to know.
> > Like which version of llvm was used to compile this .o file.
> > Which unix user name compiled it.
> > The name of service/daemon that will be using this .o
> > and so on.
> 
> Also latest git sha of the source repo, for example.
> 
> > May be some standard prefix to such global variables will do?
> > Like "bpftool prog show" can scan global data for
> > "__annotate_#name" and print both name and string contents ?
> > For folks who regularly ssh into servers to debug bpf progs
> > that will help a lot.
> > May be some annotations llvm can automatically add to .o.
> > Thoughts?
> 
> One thing that might be less clear is how information such as comm
> or comm args would be stuffed into BTF here, but perhaps these two
> wouldn't necessarily need to be part of it since these can be retrieved
> today (as in: "which program is currently holding a reference via fd
> to a certain prog/map"). For that bpftool could simply walk procfs
> once and correlate via fdinfo on unique prog/map id, so we could list
> comms in the dump which should be trivial to add:
> 
>   # ls -la /proc/30651/fd/10
>   lrwx------ 1 root root 64 Oct 16 15:53 /proc/30651/fd/10 -> anon_inode:bpf-map
>   # cat /proc/30651/fdinfo/10
>   pos:	0
>   flags:	02000002
>   mnt_id:	15
>   map_type:	1
>   key_size:	24
>   value_size:	12
>   max_entries:	65536
>   map_flags:	0x0
>   memlock:	6819840
>   map_id:	384          <---
>   frozen:	0
>   # cat /proc/30651/comm 
>   cilium-agent
>   # cat /proc/30651/cmdline 
>   ./daemon/cilium-agent--ipv4-range10.11.0.0/16[...]--enable-node-port=true
> 
> ... and similar for progs. Getting the cmdline from kernel side seems
> rather annoying from looking into what detour procfs needs to perform.
> 
> But aside from these, such annotations via BTF would be really useful.
Tried to do the following:

1. Add: static volatile const char __annotate_source1[] = __FILE__;
   to test_rdonly_maps.c and I think it got optimized away :-/
   At least I don't see it in the 'bpftool btf dump' output.

2. Add: char __annotate_source2[] SEC(".meta") = __FILE__;
   to test_rdonly_maps.c and do all the required plumbing in libbpf
   to treat .meta like .rodata. I think it works, but the map
   disappears after bpftool exits because this data is not referenced
   in the prog and the refcount drops to zero :-(

Am I missing something?