On Thu, Mar 10, 2022 at 3:05 PM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Mon, Mar 7, 2022 at 10:26 PM Kumar Kartikeya Dwivedi
> <memxor@xxxxxxxxx> wrote:
> >
> > On Tue, Mar 08, 2022 at 11:12:13AM IST, Andrii Nakryiko wrote:
> > > On Sun, Mar 6, 2022 at 3:43 PM Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx> wrote:
> > > >
> > > > In the next patch, we will introduce a new helper 'bpf_packet_pointer'
> > > > that takes offset and len and returns a packet pointer. There we want to
> > > > statically enforce offset is in range [0, 0xffff], and that len is a
> > > > constant value, in range [1, 0xffff]. This also helps us avoid a
> > > > pointless runtime check. To make these checks possible, we need to
> > > > ensure we only get a scalar type. Although a lot of other argument types
> > > > take scalars, their intent is different. Hence add general ARG_SCALAR
> > > > and ARG_CONSTANT types, where the latter is also checked to be constant
> > > > in addition to being scalar.
> > > >
> > > > Signed-off-by: Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx>
> > > > ---
> > > > include/linux/bpf.h | 2 ++
> > > > kernel/bpf/verifier.c | 13 +++++++++++++
> > > > 2 files changed, 15 insertions(+)
> > > >
> > > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > > > index 88449fbbe063..7841d90b83df 100644
> > > > --- a/include/linux/bpf.h
> > > > +++ b/include/linux/bpf.h
> > > > @@ -391,6 +391,8 @@ enum bpf_arg_type {
> > > > ARG_PTR_TO_STACK, /* pointer to stack */
> > > > ARG_PTR_TO_CONST_STR, /* pointer to a null terminated read-only string */
> > > > ARG_PTR_TO_TIMER, /* pointer to bpf_timer */
> > > > + ARG_SCALAR, /* a scalar with any value(s) */
> > >
> > > What's the difference between ARG_ANYTHING and ARG_SCALAR?
> > >
> >
> > ARG_SCALAR only accepts reg->type == SCALAR, ARG_ANYTHING accepts anything as
> > long as reg->type != NOT_INIT (due to SRC_OP for check_reg_arg and early return
> > without further checks).
> >
>
> Ah, ok, didn't realize that it's not always scalar for ARG_ANYTHING
>
>
> > > > + ARG_CONSTANT, /* a scalar with constant value */
> > >
> > > This ARG_CONSTANT serves a very similar purpose as
> > > ARG_CONST_ALLOC_SIZE_OR_ZERO, tbh. The only difference is that one is
> > > used to set meta->mem_size and this one is used (through extra func_id
> > > special handling) to set meta->ret_pkt_len. But meta->mem_size and
> > > meta->ret_pkt_len mean the same thing: how many bytes are directly
> > > accessible through a pointer returned from the helper. So I feel like
> > > there is some opportunity to unify and generalize, instead of adding
> > > more custom variants of constants. WDYT?
> > >
> >
> > I see, indeed it would make sense to make both equivalent, since
> > CONST_ALLOC_SIZE must also be a constant. Joanne also mentioned consolidating,
> > but I didn't understand how that would work for ARG_CONSTANT and ARG_CONST_SIZE
> > ones.
> >
> > I'm wondering whether we can take a step back and should go with the following
> > convention:
> >
> > ARG_MEM_SIZE, and two type flags, ARG_ZERO | ARG_CONSTANT
> >
> > Old New (in bpf_func_proto)
> > -------------------------------------------------------------------
> > ARG_CONST_SIZE ARG_MEM_SIZE
> > ARG_CONST_SIZE_OR_ZERO ARG_MEM_SIZE | ARG_ZERO
> > ARG_CONST_ALLOC_SIZE ARG_MEM_SIZE | ARG_CONST
> > ARG_CONST_ALLOC_SIZE_OR_ZERO ARG_MEM_SIZE | ARG_CONST | ARG_ZERO
> > ARG_CONSTANT (mine) ARG_MEM_SIZE | ARG_CONST
> >
>
> I think using "ARG_MEM_SIZE" as part of ARG_CONSTANT is backwards and
> misleading. It makes more sense to me to have ARG_CONSTANT and use
> ARG_ZERO (or rather ARG_MAYBE_ZERO?) and ARG_MEM_SIZE (to specify that
> this constant is describing the size of memory of a pointer that is
> passed in a previous argument).
>
> Basically, something like:
>
> ARG_CONST_SIZE => ARG_CONSTANT | ARG_MEM_SIZE
> ARG_CONST_SIZE_OR_ZERO => ARG_CONSTANT | ARG_MEM_SIZE | ARG_MAYBE_ZERO
>
> Then we can replace ARG_CONST_ALLOC_SIZE and
> ARG_CONST_ALLOC_SIZE_OR_ZERO with ARG_CONSTANT and ARG_CONSTANT |
> ARG_MAYBE_ZERO and we'll have a bit of special case to handle
> bpf_ringbuf_reserve.
>
> For ARG_CONSTANT, verifier will remember the value in
> bpf_call_arg_meta, and then we can use it as necessary (e.g., instead
> of mem_size when ARG_MEM_SIZE is specified) depending on context,
> helper being called, etc.
>
> Adding ARG_CONST just makes no sense as we always want constant value,
> otherwise it might as well be just ARG_ANYTHING, right?
Re-reading this, this paragraph is very confusing (especially taking
into account what I wrote above). My point was that in your table, you
have ARG_MEM_SIZE as a "base type" and ARG_CONST as "modifier". And
that makes little sense to me, because in all cases we have a
constant, but not in all cases we use that constant to describe the
size of memory passed in a previous argument. So I inverted that,
ARG_CONSTANT as "base type", ARG_MEM_SIZE and ARG_MAYBE_ZERO as
modifiers. And we then don't need 5 different resulting types because
"CONST_ALLOC_SIZE" handling is just a custom constant handling for
bpf_ringbuf_reserve. Just like for your use case you wanted to use
plain ARG_CONSTANT and add some extra logic for your
bpf_packet_pointer(). I hope this clarifies it a bit.
>
> I haven't spent much time thinking about this, though, so I'm probably
> missing something.
>
>
> > When we detect ARG_CONST, we always set meta->mem_size, which can be used to
> > refine returned pointer range, otherwise meta->mem_size = -1 by default (so it
> > will be -1 for the !tnum_is_const(reg->var_off) case).
> >
> > if (arg_type & ARG_CONST)
> > meta->mem_size = reg->var_off.value;
> > if (!(arg_type & ARG_ZERO) && !meta->mem_size)
> > // error
> >
> > The check_mem_size_reg call is only made when we see that previous reg was
> > ARG_PTR_TO_MEM. When preceding argument is not ARG_PTR_TO_MEM, we error if
> > ARG_CONST is not set for ARG_MEM_SIZE (so that either the mem size is for
> > previous parameter, or otherwise a constant size for the returned pointer).
> > We can also only allow certain pointer return types for that case.
> >
> > If that is too much automagic, we can also discern using ARG_MEM_SIZE vs
> > ARG_RET_MEM_SIZE, but I think the above is fine.
> >
> > ARG_CONST ofcourse only applies to args taking scalar type.
> >
> > >
> > >
> > > > __BPF_ARG_TYPE_MAX,
> > > >
> > > > /* Extended arg_types. */
> > > > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > > > index ec3a7b6c9515..0373d5bd240f 100644
> > > > --- a/kernel/bpf/verifier.c
> > > > +++ b/kernel/bpf/verifier.c
> > > > @@ -5163,6 +5163,12 @@ static bool arg_type_is_int_ptr(enum bpf_arg_type type)
> > > > type == ARG_PTR_TO_LONG;
> > > > }
> > > >
> > > > +static bool arg_type_is_scalar(enum bpf_arg_type type)
> > > > +{
> > > > + return type == ARG_SCALAR ||
> > > > + type == ARG_CONSTANT;
> > > > +}
> > > > +
> > > > static int int_ptr_type_to_size(enum bpf_arg_type type)
> > > > {
> > > > if (type == ARG_PTR_TO_INT)
> > > > @@ -5302,6 +5308,8 @@ static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
> > > > [ARG_PTR_TO_STACK] = &stack_ptr_types,
> > > > [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
> > > > [ARG_PTR_TO_TIMER] = &timer_types,
> > > > + [ARG_SCALAR] = &scalar_types,
> > > > + [ARG_CONSTANT] = &scalar_types,
> > > > };
> > > >
> > > > static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
> > > > @@ -5635,6 +5643,11 @@ static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
> > > > verbose(env, "string is not zero-terminated\n");
> > > > return -EINVAL;
> > > > }
> > > > + } else if (arg_type_is_scalar(arg_type)) {
> > > > + if (arg_type == ARG_CONSTANT && !tnum_is_const(reg->var_off)) {
> > > > + verbose(env, "R%d is not a known constant\n", regno);
> > > > + return -EACCES;
> > > > + }
> > > > }
> > > >
> > > > return err;
> > > > --
> > > > 2.35.1
> > > >
> >
> > --
> > Kartikeya
