Re: [PATCH bpf-next 2/3] bpf: add bpffs pretty print for cgroup local storage maps

[Roman Gushchin <guro@fb.com>]

On Sat, Dec 08, 2018 at 05:56:30PM -0800, Martin Lau wrote:
> On Fri, Dec 07, 2018 at 04:53:14PM -0800, Roman Gushchin wrote:
> > Implement bpffs pretty printing for cgroup local storage maps
> > (both shared and per-cpu).
> > Output example (captured for tools/testing/selftests/bpf/netcnt_prog.c):
> > 
> > Shared:
> >   $ cat /sys/fs/bpf/map_2
> >   # WARNING!! The output is for debug purpose only
> >   # WARNING!! The output format will change
> >   {4294968594,1}: {9999,1039896}
> > 
> > Per-cpu:
> >   $ cat /sys/fs/bpf/map_1
> >   # WARNING!! The output is for debug purpose only
> >   # WARNING!! The output format will change
> >   {4294968594,1}: {
> >   	cpu0: {0,0,0,0,0}
> >   	cpu1: {0,0,0,0,0}
> >   	cpu2: {1,104,0,0,0}
> >   	cpu3: {0,0,0,0,0}
> >   }
> > 
> > Signed-off-by: Roman Gushchin <guro@fb.com>
> > Cc: Alexei Starovoitov <ast@kernel.org>
> > Cc: Daniel Borkmann <daniel@iogearbox.net>
> > ---
> >  include/linux/btf.h        | 10 +++++
> >  kernel/bpf/local_storage.c | 90 +++++++++++++++++++++++++++++++++++++-
> >  2 files changed, 99 insertions(+), 1 deletion(-)
> > 
> > diff --git a/include/linux/btf.h b/include/linux/btf.h
> > index 8c2199b5d250..ac67bc4cbfd9 100644
> > --- a/include/linux/btf.h
> > +++ b/include/linux/btf.h
> > @@ -5,6 +5,7 @@
> >  #define _LINUX_BTF_H 1
> >  
> >  #include <linux/types.h>
> > +#include <uapi/linux/btf.h>
> >  
> >  struct btf;
> >  struct btf_type;
> > @@ -63,4 +64,13 @@ static inline const char *btf_name_by_offset(const struct btf *btf,
> >  }
> >  #endif
> >  
> > +static inline const struct btf_type *btf_orig_type(const struct btf *btf,
> > +						   const struct btf_type *t)
> > +{
> > +	while (t && BTF_INFO_KIND(t->info) == BTF_KIND_TYPEDEF)
> > +		t = btf_type_by_id(btf, t->type);
> Only typedef is allowed?  Not even const?
> If that is not the case, can btf_type_id_size() be reused?
> The "type following" has already been done once and then remembered
> in the verification time.
> 
> If cgroup_storage_check_btf() can only allow typedef, please
> move "btf_orig_type()" to the local_storage.c.
> 
> > +
> > +	return t;
> > +}
> > +
> >  #endif
> > diff --git a/kernel/bpf/local_storage.c b/kernel/bpf/local_storage.c
> > index b65017dead44..7b51fe1aba3c 100644
> > --- a/kernel/bpf/local_storage.c
> > +++ b/kernel/bpf/local_storage.c
> > @@ -1,11 +1,13 @@
> >  //SPDX-License-Identifier: GPL-2.0
> >  #include <linux/bpf-cgroup.h>
> >  #include <linux/bpf.h>
> > +#include <linux/btf.h>
> >  #include <linux/bug.h>
> >  #include <linux/filter.h>
> >  #include <linux/mm.h>
> >  #include <linux/rbtree.h>
> >  #include <linux/slab.h>
> > +#include <uapi/linux/btf.h>
> >  
> >  DEFINE_PER_CPU(struct bpf_cgroup_storage*, bpf_cgroup_storage[MAX_BPF_CGROUP_STORAGE_TYPE]);
> >  
> > @@ -308,6 +310,91 @@ static int cgroup_storage_delete_elem(struct bpf_map *map, void *key)
> >  	return -EINVAL;
> >  }
> >  
> > +static int cgroup_storage_check_btf(const struct bpf_map *map,
> > +				    const struct btf *btf,
> > +				    const struct btf_type *key_type,
> Actually, in "map_check_btf()" (just before cgroup_storage_check_btf()
> is called), btf_type_id_size() has already been called
> to get the true size and the resolved type (i.e. BTF_INFO_KIND_STRUCT here)
> in order to reject "key_size != map->key_size".  Hence, the key_type
> passed to cgroup_storage_check_btf() here will not be KIND_TYPEDEF or
> KIND_CONST.

So, the type here is a structure, and its fields are typedefs of ints.
Looks like reusing btf_type_id_size() is the best approach.

> 
> > +				    const struct btf_type *value_type)
> > +{
> > +	const struct btf_type *st, *t;
> > +	struct btf_member *m;
> > +
> > +	/* Key is expected to be of struct bpf_cgroup_storage_key type,
> > +	 * which is:
> > +	 * struct bpf_cgroup_storage_key {
> > +	 *	__u64	cgroup_inode_id;
> > +	 *	__u32	attach_type;
> > +	 * };
> > +	 */
> > +
> > +	/*
> > +	 * Key_type must be a structure (or a typedef of a structure) with
> > +	 * two members.
> > +	 */
> > +	st = btf_orig_type(btf, key_type);
> > +	if (BTF_INFO_KIND(st->info) != BTF_KIND_STRUCT ||
> > +	    BTF_INFO_VLEN(st->info) != 2)
> > +		return -EINVAL;
> > +
> > +	/*
> > +	 * The first field must be a 64 bit integer at 0 offset.
> > +	 */
> > +	m = (struct btf_member *)(st + 1);
> > +	t = btf_orig_type(btf, btf_type_by_id(btf, m->type));
> > +	if (!t || BTF_INFO_KIND(t->info) != BTF_KIND_INT || m->offset ||
> > +	    t->size !=
> > +	    FIELD_SIZEOF(struct bpf_cgroup_storage_key, cgroup_inode_id))
> Instead of t->size,
> BTF_INT_BITS() and BTF_INT_OFFSET() need to be checked (please refer to the
> key_type check in array_map_check_btf()).

Gotcha.

> 
> I think exposing "btf_type_int_is_regular()" from btf.c will be easier here.
> Actually, add "btf_type_is_reg_int(t, expected_size)" to btf.h and btf.c,
> like (uncompiled and untested code):

> 
> /*
>  * Not a bit field and it must be the expected size.
>  */
> bool btf_type_is_reg_int(const struct btf_type *t, u32 expected_size)
> {
>  	u8 nr_bits, nr_bytes;
>         u32 int_data;
> 
> 	if (!btf_type_is_int(t))
> 		return false;
> 
> 	int_data = btf_type_int(t);
> 	nr_bits = BTF_INT_BITS(int_data);
> 	nr_bytes = BITS_ROUNDUP_BYTES(nr_bits);
> 	if (BITS_PER_BYTE_MASKED(nr_bits) ||
> 	    BTF_INT_OFFSET(int_data) ||
> 	    nr_bytes != expected_size)
>                 return false;
> 
> 	return true;
> }

Looks good to me. Will implement in v2.

Thanks!