On Fri, Oct 11, 2019 at 11:31 AM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Wed, Oct 9, 2019 at 9:17 PM Alexei Starovoitov <ast@xxxxxxxxxx> wrote:
> >
> > libbpf analyzes bpf C program, searches in-kernel BTF for given type name
> > and stores it into expected_attach_type.
> > The kernel verifier expects this btf_id to point to something like:
> > typedef void (*btf_trace_kfree_skb)(void *, struct sk_buff *skb, void *loc);
> > which represents signature of raw_tracepoint "kfree_skb".
> >
> > Then btf_ctx_access() matches ctx+0 access in bpf program with 'skb'
> > and 'ctx+8' access with 'loc' arguments of "kfree_skb" tracepoint.
> > In first case it passes btf_id of 'struct sk_buff *' back to the verifier core
> > and 'void *' in second case.
> >
> > Then the verifier tracks PTR_TO_BTF_ID as any other pointer type.
> > Like PTR_TO_SOCKET points to 'struct bpf_sock',
> > PTR_TO_TCP_SOCK points to 'struct bpf_tcp_sock', and so on.
> > PTR_TO_BTF_ID points to in-kernel structs.
> > If 1234 is btf_id of 'struct sk_buff' in vmlinux's BTF
> > then PTR_TO_BTF_ID#1234 points to one of in kernel skbs.
> >
> > When PTR_TO_BTF_ID#1234 is dereferenced (like r2 = *(u64 *)r1 + 32)
> > the btf_struct_access() checks which field of 'struct sk_buff' is
> > at offset 32. Checks that size of access matches type definition
> > of the field and continues to track the dereferenced type.
> > If that field was a pointer to 'struct net_device' the r2's type
> > will be PTR_TO_BTF_ID#456. Where 456 is btf_id of 'struct net_device'
> > in vmlinux's BTF.
> >
> > Such verifier analysis prevents "cheating" in BPF C program.
> > The program cannot cast arbitrary pointer to 'struct sk_buff *'
> > and access it. C compiler would allow type cast, of course,
> > but the verifier will notice type mismatch based on BPF assembly
> > and in-kernel BTF.
> >
> > Signed-off-by: Alexei Starovoitov <ast@xxxxxxxxxx>
> > ---
> > include/linux/bpf.h | 17 +++-
> > include/linux/bpf_verifier.h | 4 +
> > kernel/bpf/btf.c | 186 +++++++++++++++++++++++++++++++++++
> > kernel/bpf/verifier.c | 86 +++++++++++++++-
> > kernel/trace/bpf_trace.c | 2 +-
> > 5 files changed, 290 insertions(+), 5 deletions(-)
> >
>
> [...]
>
> > +int btf_struct_access(struct bpf_verifier_log *log,
> > + const struct btf_type *t, int off, int size,
> > + enum bpf_access_type atype,
> > + u32 *next_btf_id)
> > +{
> > + const struct btf_member *member;
> > + const struct btf_type *mtype;
> > + const char *tname, *mname;
> > + int i, moff = 0, msize;
> > +
> > +again:
> > + tname = __btf_name_by_offset(btf_vmlinux, t->name_off);
> > + if (!btf_type_is_struct(t)) {
> > + bpf_log(log, "Type '%s' is not a struct", tname);
> > + return -EINVAL;
> > + }
> > +
> > + for_each_member(i, t, member) {
> > + /* offset of the field in bits */
> > + moff = btf_member_bit_offset(t, member);
>
> This whole logic of offset/size checking doesn't work for bitfields.
> Your moff % 8 might be non-zero (most probably, actually, for
> bitfield). Also, msize of underlying integer type is not the same as
> member's bit size. So probably just check that it's a bitfield and
> skip it?
>
> The check is surprisingly subtle and not straightforward, btw. You
Well, this part is not true, checking btf_member_bitfield_size(t,
member) for non-zero is enough to derive it's bitfield.
> need to get btf_member_bitfield_size(t, member) and check that it's
> not equal to underlying type's size (which is in bytes, so * 8). It's
> unfortunate it's so non-straightforward. But if you don't filter that,
> all those `moff / 8` and `msize` checks are bogus.
>
> > +
> > + if (off + size <= moff / 8)
> > + /* won't find anything, field is already too far */
> > + break;
> > +
> > + /* type of the field */
> > + mtype = btf_type_by_id(btf_vmlinux, member->type);
> > + mname = __btf_name_by_offset(btf_vmlinux, member->name_off);
> > +
> > + /* skip modifiers */
> > + while (btf_type_is_modifier(mtype))
> > + mtype = btf_type_by_id(btf_vmlinux, mtype->type);
> > +
> > + if (btf_type_is_array(mtype))
> > + /* array deref is not supported yet */
> > + continue;
> > +
> > + if (!btf_type_has_size(mtype) && !btf_type_is_ptr(mtype)) {
> > + bpf_log(log, "field %s doesn't have size\n", mname);
> > + return -EFAULT;
> > + }
> > + if (btf_type_is_ptr(mtype))
> > + msize = 8;
> > + else
> > + msize = mtype->size;
> > + if (off >= moff / 8 + msize)
> > + /* no overlap with member, keep iterating */
> > + continue;
> > + /* the 'off' we're looking for is either equal to start
> > + * of this field or inside of this struct
> > + */
> > + if (btf_type_is_struct(mtype)) {
> > + /* our field must be inside that union or struct */
> > + t = mtype;
> > +
> > + /* adjust offset we're looking for */
> > + off -= moff / 8;
> > + goto again;
> > + }
> > + if (msize != size) {
> > + /* field access size doesn't match */
> > + bpf_log(log,
> > + "cannot access %d bytes in struct %s field %s that has size %d\n",
> > + size, tname, mname, msize);
> > + return -EACCES;
>
> Are you sure this has to be an error? Why not just default to
> SCALAR_VALUE here? E.g., if compiler generated one read for few
> smaller fields, or user wants to read lower 1 byte of int field, etc.
> I think if you move this size check into the following ptr check, it
> should be fine. Pointer is the only case where you care about correct
> read/value, isn't it?
>
> > + }
> > +
> > + if (btf_type_is_ptr(mtype)) {
> > + const struct btf_type *stype;
> > +
> > + stype = btf_type_by_id(btf_vmlinux, mtype->type);
> > + /* skip modifiers */
> > + while (btf_type_is_modifier(stype))
> > + stype = btf_type_by_id(btf_vmlinux, stype->type);
> > + if (btf_type_is_struct(stype)) {
> > + *next_btf_id = mtype->type;
> > + return PTR_TO_BTF_ID;
> > + }
> > + }
> > + /* all other fields are treated as scalars */
> > + return SCALAR_VALUE;
> > + }
> > + bpf_log(log, "struct %s doesn't have field at offset %d\n", tname, off);
> > + return -EINVAL;
> > +}
> > +
>
> [...]
