Re: [PATCH bpf-next v9 04/23] bpf/verifier: allow kfunc to return an allocated mem

From: Kumar Kartikeya Dwivedi
Date: Wed Aug 31 2022 - 07:07:24 EST

On Wed, 31 Aug 2022 at 07:50, Benjamin Tissoires
<benjamin.tissoires@xxxxxxxxxx> wrote:
>
> On Fri, Aug 26, 2022 at 3:25 AM Kumar Kartikeya Dwivedi
> <memxor@xxxxxxxxx> wrote:
> >
> > On Wed, 24 Aug 2022 at 15:41, Benjamin Tissoires
> > <benjamin.tissoires@xxxxxxxxxx> wrote:
> > >
> > > For drivers (outside of network), the incoming data is not statically
> > > defined in a struct. Most of the time the data buffer is kzalloc-ed
> > > and thus we can not rely on eBPF and BTF to explore the data.
> > >
> > > This commit allows to return an arbitrary memory, previously allocated by
> > > the driver.
> > > An interesting extra point is that the kfunc can mark the exported
> > > memory region as read only or read/write.
> > >
> > > So, when a kfunc is not returning a pointer to a struct but to a plain
> > > type, we can consider it is a valid allocated memory assuming that:
> > > - one of the arguments is either called rdonly_buf_size or
> > > rdwr_buf_size
> > > - and this argument is a const from the caller point of view
> > >
> > > We can then use this parameter as the size of the allocated memory.
> > >
> > > The memory is either read-only or read-write based on the name
> > > of the size parameter.
> > >
> > > Acked-by: Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx>
> > > Signed-off-by: Benjamin Tissoires <benjamin.tissoires@xxxxxxxxxx>
> > >
> > > ---
> > >
> > > changes in v9:
> > > - updated to match upstream (replaced kfunc_flag by a field in
> > > kfunc_meta)
> > >
> > > no changes in v8
> > >
> > > changes in v7:
> > > - ensures btf_type_is_struct_ptr() checks for a ptr first
> > > (squashed from next commit)
> > > - remove multiple_ref_obj_id need
> > > - use btf_type_skip_modifiers instead of manually doing it in
> > > btf_type_is_struct_ptr()
> > > - s/strncmp/strcmp/ in btf_is_kfunc_arg_mem_size()
> > > - check for tnum_is_const when retrieving the size value
> > > - have only one check for "Ensure only one argument is referenced
> > > PTR_TO_BTF_ID"
> > > - add some more context to the commit message
> > >
> > > changes in v6:
> > > - code review from Kartikeya:
> > > - remove comment change that had no reasons to be
> > > - remove handling of PTR_TO_MEM with kfunc releases
> > > - introduce struct bpf_kfunc_arg_meta
> > > - do rdonly/rdwr_buf_size check in btf_check_kfunc_arg_match
> > > - reverted most of the changes in verifier.c
> > > - make sure kfunc acquire is using a struct pointer, not just a plain
> > > pointer
> > > - also forward ref_obj_id to PTR_TO_MEM in kfunc to not use after free
> > > the allocated memory
> > >
> > > changes in v5:
> > > - updated PTR_TO_MEM comment in btf.c to match upstream
> > > - make it read-only or read-write based on the name of size
> > >
> > > new in v4
> > >
> > > change btf.h
> > >
> > > fix allow kfunc to return an allocated mem
> > > ---
> > > include/linux/bpf.h | 9 +++-
> > > include/linux/btf.h | 10 +++++
> > > kernel/bpf/btf.c | 98 ++++++++++++++++++++++++++++++++++---------
> > > kernel/bpf/verifier.c | 43 +++++++++++++------
> > > 4 files changed, 128 insertions(+), 32 deletions(-)
> > >
> > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > > index 39bd36359c1e..90dd218e0199 100644
> > > --- a/include/linux/bpf.h
> > > +++ b/include/linux/bpf.h
> > > @@ -1932,13 +1932,20 @@ int btf_distill_func_proto(struct bpf_verifier_log *log,
> > > const char *func_name,
> > > struct btf_func_model *m);
> > > [...]
> > > +
> > > static int btf_check_func_arg_match(struct bpf_verifier_env *env,
> > > const struct btf *btf, u32 func_id,
> > > struct bpf_reg_state *regs,
> > > bool ptr_to_mem_ok,
> > > - u32 kfunc_flags)
> > > + struct bpf_kfunc_arg_meta *kfunc_meta)
> > > {
> > > enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
> > > bool rel = false, kptr_get = false, trusted_arg = false;
> > > @@ -6207,12 +6232,12 @@ static int btf_check_func_arg_match(struct bpf_verifier_env *env,
> > > return -EINVAL;
> > > }
> > >
> > > - if (is_kfunc) {
> > > + if (is_kfunc && kfunc_meta) {
> > > /* Only kfunc can be release func */
> > > - rel = kfunc_flags & KF_RELEASE;
> > > - kptr_get = kfunc_flags & KF_KPTR_GET;
> > > - trusted_arg = kfunc_flags & KF_TRUSTED_ARGS;
> > > - sleepable = kfunc_flags & KF_SLEEPABLE;
> > > + rel = kfunc_meta->flags & KF_RELEASE;
> > > + kptr_get = kfunc_meta->flags & KF_KPTR_GET;
> > > + trusted_arg = kfunc_meta->flags & KF_TRUSTED_ARGS;
> > > + sleepable = kfunc_meta->flags & KF_SLEEPABLE;
> > > }
> > >
> > > /* check that BTF function arguments match actual types that the
> > > @@ -6225,6 +6250,35 @@ static int btf_check_func_arg_match(struct bpf_verifier_env *env,
> > >
> > > t = btf_type_skip_modifiers(btf, args[i].type, NULL);
> > > if (btf_type_is_scalar(t)) {
> > > + if (is_kfunc && kfunc_meta) {
> > > + bool is_buf_size = false;
> > > +
> > > + /* check for any const scalar parameter of name "rdonly_buf_size"
> > > + * or "rdwr_buf_size"
> > > + */
> > > + if (btf_is_kfunc_arg_mem_size(btf, &args[i], reg,
> > > + "rdonly_buf_size")) {
> > > + kfunc_meta->r0_rdonly = true;
> > > + is_buf_size = true;
> > > + } else if (btf_is_kfunc_arg_mem_size(btf, &args[i], reg,
> > > + "rdwr_buf_size"))
> > > + is_buf_size = true;
> > > +
> > > + if (is_buf_size) {
> > > + if (kfunc_meta->r0_size) {
> > > + bpf_log(log, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
> > > + return -EINVAL;
> > > + }
> > > +
> > > + if (!tnum_is_const(reg->var_off)) {
> > > + bpf_log(log, "R%d is not a const\n", regno);
> > > + return -EINVAL;
> > > + }
> > > +
> > > + kfunc_meta->r0_size = reg->var_off.value;
> >
> > Sorry for not pointing it out before, but you will need a call to
> > mark_chain_precision here after this, since the value of the scalar is
> > being used to decide the size of the returned pointer.
>
> No worries.
>
> I do however have a couple of questions (I have strictly no idea what
> mark_chain_precision does):

See this patch for some background:
https://lore.kernel.org/bpf/20220823185300.406-2-memxor@xxxxxxxxx
Same case here, it is setting the size of r0 PTR_TO_MEM.

> - which register number should I call mark_chain_precision() as
> parameter? r0 or regno (the one with the constant)?

Yes, regno, i.e. the one with the constant.

> - mark_chain_precision() is declared static in verifier.c. Should I
> export it so btf.c can have access to it, or can I delay the call to
> mark_chain_precision() in verifier.c when I set
> regs[BPF_REG_0].mem_size?
>

Yes, but then you have to remember the regno you have to call it for.
So it might be easier to just make it non-static and call in btf.c.

>
> >
> > > + }
> > > + }
> > > +
> > > if (reg->type == SCALAR_VALUE)
> > > continue;
> > > bpf_log(log, "R%d is not a scalar\n", regno);
> > > @@ -6255,6 +6309,19 @@ static int btf_check_func_arg_match(struct bpf_verifier_env *env,
> > > if (ret < 0)
> > > return ret;
> > >
> > > + if (is_kfunc && reg->type == PTR_TO_BTF_ID) {
> >
> > I think you can drop this extra check 'reg->type == PTR_TO_BTF_ID),
> > this condition of only one ref_obj_id should hold regardless of the
> > type.
>
> Ack.
>
> Cheers,
> Benjamin
>
> >
> > > [...]
> >
>