Re: [PATCH bpf-next] bpftool: specify XDP Hints ifname when loading program

[Larysa Zaremba]

On Fri, May 12, 2023 at 11:23:00AM +0100, Quentin Monnet wrote:
> 2023-05-11 17:13 UTC+0200 ~ Larysa Zaremba <larysa.zaremba@xxxxxxxxx>
> > Add ability to specify a network interface used to resolve
> > XDP Hints kfuncs when loading program through bpftool.
> > 
> > Usage:
> > bpftool prog load <bpf_obj_path> <pin_path> dev xdpmeta <ifname>
> 
> Thanks for this patch!
> 
> Regarding the command-line syntax, I'm not a big fan of the optional
> sub-keyword for the device for XDP hints. I must admit I had not
> anticipated other another use for the "dev" keyword. Instead, have you
> considered one of the following:
> 
> 1) Adding a different keyword ("xdpmeta_dev"?) and making it
> incompatible with "dev"
> 
> 2) Another alternative would be adding a sub-keyword for offload too:
> 
>     bpftool p l [...] dev <[offload <ifname> | xdpmeta <ifname>]>
> 
> If the ifname is provided with no sub-keyword, we would consider it for
> offload for legacy support, possibly warn that the syntax is deprecated.
> 
> What do you think?
> 

I think first option would look a little bit nicer, but I like the idea to 
deprecate "dev <ifname>". In my current version, forgetting to add "xdpmeta" 
resulted in not very descriptive errors, this may confuse new users. So what 
about:

bpftool prog load [...] xdpmeta_dev/offload_dev <ifname>

"dev <ifname>" syntax would still work, but with a big warning, like this:

  'bpftool prog [...] dev <ifname>' syntax is deprecated. Going further, please 
  use 'offload_dev <ifname>' to offload program to device. For XDP hints 
  applications, use 'xdpmeta_dev <ifname>'.

> > 
> > Signed-off-by: Larysa Zaremba <larysa.zaremba@xxxxxxxxx>
> > ---
> >  tools/bpf/bpftool/prog.c | 28 +++++++++++++++++++++-------
> >  1 file changed, 21 insertions(+), 7 deletions(-)
> > 
> > diff --git a/tools/bpf/bpftool/prog.c b/tools/bpf/bpftool/prog.c
> > index 91b6075b2db3..a9cb96d99277 100644
> > --- a/tools/bpf/bpftool/prog.c
> > +++ b/tools/bpf/bpftool/prog.c
> > @@ -1517,12 +1517,13 @@ static int load_with_options(int argc, char **argv, bool first_prog_only)
> >  	struct bpf_program *prog = NULL, *pos;
> >  	unsigned int old_map_fds = 0;
> >  	const char *pinmaps = NULL;
> > +	__u32 offload_ifindex = 0;
> >  	bool auto_attach = false;
> > +	__u32 meta_ifindex = 0;
> >  	struct bpf_object *obj;
> >  	struct bpf_map *map;
> >  	const char *pinfile;
> >  	unsigned int i, j;
> > -	__u32 ifindex = 0;
> >  	const char *file;
> >  	int idx, err;
> >  
> > @@ -1614,17 +1615,25 @@ static int load_with_options(int argc, char **argv, bool first_prog_only)
> >  			map_replace[old_map_fds].fd = fd;
> >  			old_map_fds++;
> >  		} else if (is_prefix(*argv, "dev")) {
> > +			__u32 *cur_ifindex;
> > +
> >  			NEXT_ARG();
> >  
> > -			if (ifindex) {
> > -				p_err("offload device already specified");
> > +			if (offload_ifindex || meta_ifindex) {
> > +				p_err("device already specified");
> 
> I'd add a note to tell somehow that offload and XDP hints are mutually
> exclusive. I suspect not all users understand these well.

Ok, will do.

> 
> >  				goto err_free_reuse_maps;
> >  			}
> > +			if (is_prefix(*argv, "xdpmeta")) {
> > +				cur_ifindex = &meta_ifindex;
> > +				NEXT_ARG();
> > +			} else {
> > +				cur_ifindex = &offload_ifindex;
> > +			}
> >  			if (!REQ_ARGS(1))
> >  				goto err_free_reuse_maps;
> >  
> > -			ifindex = if_nametoindex(*argv);
> > -			if (!ifindex) {
> > +			*cur_ifindex = if_nametoindex(*argv);
> > +			if (!(*cur_ifindex)) {
> >  				p_err("unrecognized netdevice '%s': %s",
> >  				      *argv, strerror(errno));
> >  				goto err_free_reuse_maps;
> > @@ -1671,7 +1680,12 @@ static int load_with_options(int argc, char **argv, bool first_prog_only)
> >  				goto err_close_obj;
> >  		}
> >  
> > -		bpf_program__set_ifindex(pos, ifindex);
> > +		if (prog_type == BPF_PROG_TYPE_XDP && meta_ifindex) {
> > +			bpf_program__set_flags(pos, BPF_F_XDP_DEV_BOUND_ONLY);
> > +			bpf_program__set_ifindex(pos, meta_ifindex);
> > +		} else {
> > +			bpf_program__set_ifindex(pos, offload_ifindex);
> > +		}
> >  		if (bpf_program__type(pos) != prog_type)
> >  			bpf_program__set_type(pos, prog_type);
> >  		bpf_program__set_expected_attach_type(pos, expected_attach_type);
> > @@ -1709,7 +1723,7 @@ static int load_with_options(int argc, char **argv, bool first_prog_only)
> >  	idx = 0;
> >  	bpf_object__for_each_map(map, obj) {
> >  		if (bpf_map__type(map) != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
> > -			bpf_map__set_ifindex(map, ifindex);
> > +			bpf_map__set_ifindex(map, offload_ifindex);
> >  
> >  		if (j < old_map_fds && idx == map_replace[j].idx) {
> >  			err = bpf_map__reuse_fd(map, map_replace[j++].fd);
> 
> Could you please also update the following items:
> 
> - The usage message for "bpftool prog load" near the end of prog.c
> 
> - The related doc in Documentation/bpftool-prog.rst (command summary
> near the top, and detailed description later in the doc)
> 
> - Bash completion (for keyword "dev", look for "_sysfs_get_netdevs" in
> bash-completion/bpftool). I'm happy to help with this one if necessary.

Will do all the above in v2.

> 
> Thanks,
> Quentin