Re: [PATCH v3 bpf-next 3/5] bpf: Introduce kptr_rcu.

[David Vernet <void@manifault.com>]

On Tue, Feb 28, 2023 at 11:49:23AM -0800, Alexei Starovoitov wrote:
> On Tue, Feb 28, 2023 at 10:45:16AM -0600, David Vernet wrote:
> > On Mon, Feb 27, 2023 at 08:01:19PM -0800, Alexei Starovoitov wrote:
> > > From: Alexei Starovoitov <ast@kernel.org>
> > > 
> > > The life time of certain kernel structures like 'struct cgroup' is protected by RCU.
> > > Hence it's safe to dereference them directly from __kptr tagged pointers in bpf maps.
> > > The resulting pointer is MEM_RCU and can be passed to kfuncs that expect KF_RCU.
> > > Derefrence of other kptr-s returns PTR_UNTRUSTED.
> > > 
> > > For example:
> > > struct map_value {
> > >    struct cgroup __kptr *cgrp;
> > > };
> > > 
> > > SEC("tp_btf/cgroup_mkdir")
> > > int BPF_PROG(test_cgrp_get_ancestors, struct cgroup *cgrp_arg, const char *path)
> > > {
> > >   struct cgroup *cg, *cg2;
> > > 
> > >   cg = bpf_cgroup_acquire(cgrp_arg); // cg is PTR_TRUSTED and ref_obj_id > 0
> > >   bpf_kptr_xchg(&v->cgrp, cg);
> > > 
> > >   cg2 = v->cgrp; // This is new feature introduced by this patch.
> > >   // cg2 is PTR_MAYBE_NULL | MEM_RCU.
> > >   // When cg2 != NULL, it's a valid cgroup, but its percpu_ref could be zero
> > > 
> > >   bpf_cgroup_ancestor(cg2, level); // safe to do.
> > > }
> > > 
> > > Signed-off-by: Alexei Starovoitov <ast@kernel.org>
> > > ---
> > >  Documentation/bpf/kfuncs.rst                  | 11 ++++---
> > >  include/linux/bpf.h                           | 15 ++++++---
> > >  include/linux/btf.h                           |  2 +-
> > >  kernel/bpf/btf.c                              | 16 +++++++++
> > >  kernel/bpf/helpers.c                          |  7 ++--
> > >  kernel/bpf/syscall.c                          |  4 +++
> > >  kernel/bpf/verifier.c                         | 33 ++++++++++++-------
> > >  net/bpf/test_run.c                            |  3 +-
> > >  .../selftests/bpf/progs/map_kptr_fail.c       |  4 +--
> > >  tools/testing/selftests/bpf/verifier/calls.c  |  2 +-
> > >  .../testing/selftests/bpf/verifier/map_kptr.c |  2 +-
> > >  11 files changed, 69 insertions(+), 30 deletions(-)
> > > 
> > > diff --git a/Documentation/bpf/kfuncs.rst b/Documentation/bpf/kfuncs.rst
> > > index 7d7c1144372a..49c5cb6f46e7 100644
> > > --- a/Documentation/bpf/kfuncs.rst
> > > +++ b/Documentation/bpf/kfuncs.rst
> > > @@ -232,11 +232,12 @@ added later.
> > >  2.4.8 KF_RCU flag
> > >  -----------------
> > >  
> > > -The KF_RCU flag is used for kfuncs which have a rcu ptr as its argument.
> > > -When used together with KF_ACQUIRE, it indicates the kfunc should have a
> > > -single argument which must be a trusted argument or a MEM_RCU pointer.
> > > -The argument may have reference count of 0 and the kfunc must take this
> > > -into consideration.
> > > +The KF_RCU flag is a weaker version of KF_TRUSTED_ARGS. The kfuncs marked with
> > > +KF_RCU expect either PTR_TRUSTED or MEM_RCU arguments. The verifier guarantees
> > > +that the objects are valid and there is no use-after-free, but the pointers
> > > +maybe NULL and pointee object's reference count could have reached zero, hence
> > 
> > s/maybe/may be
> > 
> > > +kfuncs must do != NULL check and consider refcnt==0 case when accessing such
> > > +arguments.
> > 
> > Hmmm, given that it's only necessary to check refcnt==0 if the kfunc is
> > KF_ACQUIRE, wdyt about addending this paragraph with something like the
> > following (note as well the addition of the KF_RET_NULL suggestion):
> > 
> > ...the pointers may be NULL, and the object's refcount could have
> > reached zero. The kfuncs must therefore do a != NULL check, and if
> > returning a KF_ACQUIRE pointer, also check that refcnt != 0. Note as
> > well that a KF_ACQUIRE kfunc that is KF_RCU should **very** likely also
> > be KF_RET_NULL, for both of these reasons.
> 
> Good suggestion.
> 
> > >  .. _KF_deprecated_flag:
> > >  
> > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > > index 520b238abd5a..d4b5faa0a777 100644
> > > --- a/include/linux/bpf.h
> > > +++ b/include/linux/bpf.h
> > > @@ -178,11 +178,12 @@ enum btf_field_type {
> > >  	BPF_TIMER      = (1 << 1),
> > >  	BPF_KPTR_UNREF = (1 << 2),
> > >  	BPF_KPTR_REF   = (1 << 3),
> > > -	BPF_KPTR       = BPF_KPTR_UNREF | BPF_KPTR_REF,
> > > -	BPF_LIST_HEAD  = (1 << 4),
> > > -	BPF_LIST_NODE  = (1 << 5),
> > > -	BPF_RB_ROOT    = (1 << 6),
> > > -	BPF_RB_NODE    = (1 << 7),
> > > +	BPF_KPTR_RCU   = (1 << 4), /* kernel internal. not exposed to bpf prog */
> > > +	BPF_KPTR       = BPF_KPTR_UNREF | BPF_KPTR_REF | BPF_KPTR_RCU,
> > > +	BPF_LIST_HEAD  = (1 << 5),
> > > +	BPF_LIST_NODE  = (1 << 6),
> > > +	BPF_RB_ROOT    = (1 << 7),
> > > +	BPF_RB_NODE    = (1 << 8),
> > >  	BPF_GRAPH_NODE_OR_ROOT = BPF_LIST_NODE | BPF_LIST_HEAD |
> > >  				 BPF_RB_NODE | BPF_RB_ROOT,
> > >  };
> > > @@ -284,6 +285,8 @@ static inline const char *btf_field_type_name(enum btf_field_type type)
> > >  	case BPF_KPTR_UNREF:
> > >  	case BPF_KPTR_REF:
> > >  		return "kptr";
> > > +	case BPF_KPTR_RCU:
> > > +		return "kptr_rcu";
> > >  	case BPF_LIST_HEAD:
> > >  		return "bpf_list_head";
> > >  	case BPF_LIST_NODE:
> > > @@ -307,6 +310,7 @@ static inline u32 btf_field_type_size(enum btf_field_type type)
> > >  		return sizeof(struct bpf_timer);
> > >  	case BPF_KPTR_UNREF:
> > >  	case BPF_KPTR_REF:
> > > +	case BPF_KPTR_RCU:
> > >  		return sizeof(u64);
> > >  	case BPF_LIST_HEAD:
> > >  		return sizeof(struct bpf_list_head);
> > > @@ -331,6 +335,7 @@ static inline u32 btf_field_type_align(enum btf_field_type type)
> > >  		return __alignof__(struct bpf_timer);
> > >  	case BPF_KPTR_UNREF:
> > >  	case BPF_KPTR_REF:
> > > +	case BPF_KPTR_RCU:
> > >  		return __alignof__(u64);
> > >  	case BPF_LIST_HEAD:
> > >  		return __alignof__(struct bpf_list_head);
> > > diff --git a/include/linux/btf.h b/include/linux/btf.h
> > > index 49e0fe6d8274..556b3e2e7471 100644
> > > --- a/include/linux/btf.h
> > > +++ b/include/linux/btf.h
> > > @@ -70,7 +70,7 @@
> > >  #define KF_TRUSTED_ARGS (1 << 4) /* kfunc only takes trusted pointer arguments */
> > >  #define KF_SLEEPABLE    (1 << 5) /* kfunc may sleep */
> > >  #define KF_DESTRUCTIVE  (1 << 6) /* kfunc performs destructive actions */
> > > -#define KF_RCU          (1 << 7) /* kfunc only takes rcu pointer arguments */
> > > +#define KF_RCU          (1 << 7) /* kfunc takes either rcu or trusted pointer arguments */
> > >  
> > >  /*
> > >   * Tag marking a kernel function as a kfunc. This is meant to minimize the
> > > diff --git a/kernel/bpf/btf.c b/kernel/bpf/btf.c
> > > index 01dee7d48e6d..a44ea1f6164b 100644
> > > --- a/kernel/bpf/btf.c
> > > +++ b/kernel/bpf/btf.c
> > > @@ -3552,6 +3552,18 @@ static int btf_find_field(const struct btf *btf, const struct btf_type *t,
> > >  	return -EINVAL;
> > >  }
> > >  
> > 
> > Could you please add a comment here that once gcc has tag support, we
> > can replace this mechanism with just checking the type's BTF tag? I like
> > this a lot in the interim though -- it's a very easy way to add kfuncs
> > for new RCU-protected types, and will be trivially easy to remove and
> > cleanup later.
> 
> +1
> 
> > > +BTF_SET_START(rcu_protected_types)
> > > +BTF_ID(struct, prog_test_ref_kfunc)
> > > +BTF_ID(struct, cgroup)
> > > +BTF_SET_END(rcu_protected_types)
> > > +
> > > +static bool rcu_protected_object(const struct btf *btf, u32 btf_id)
> > > +{
> > > +	if (!btf_is_kernel(btf))
> > > +		return false;
> > > +	return btf_id_set_contains(&rcu_protected_types, btf_id);
> > > +}
> > > +
> > >  static int btf_parse_kptr(const struct btf *btf, struct btf_field *field,
> > >  			  struct btf_field_info *info)
> > >  {
> > > @@ -3615,6 +3627,10 @@ static int btf_parse_kptr(const struct btf *btf, struct btf_field *field,
> > >  		field->kptr.dtor = (void *)addr;
> > >  	}
> > >  
> > > +	if (info->type == BPF_KPTR_REF && rcu_protected_object(kernel_btf, id))
> > > +		/* rcu dereference of this field will return MEM_RCU instead of PTR_UNTRUSTED */
> > > +		field->type = BPF_KPTR_RCU;
> > 
> > Can you move this into the if block above, and update the conditional to
> > just be:
> > 
> > if (rcu_protected_object(kernel_btf, id))
> 
> good idea.
> 
> > Also, outside the scope of your patch and subjective, but IMO it's a bit
> > confusing that we're looking at info->type, when field->type already ==
> > info->type. When reading the code it looks like field->type is unset
> > unless we set it to BPF_KPTR_RCU, but in reality we're just overwriting
> > it from being BPF_KPTR_REF. Might be worth tidying up at some point (I
> > can do that in a follow-on patch once this series lands).
> 
> The caller of btf_parse_kptr() provided temporary btf_field_info array.
> Since there is only one caller it's easy to see. Not sure what clean up you have in mind.

I was thinking of checking field->type instead of info->type in the if
statements to make it more clear that field->type is already set. This
is really just a subjective point about readability. Probably wasn't
worth the additional noise on the patch.

> > >  	field->kptr.btf_id = id;
> > >  	field->kptr.btf = kernel_btf;
> > >  	field->kptr.module = mod;
> > > diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c
> > > index a784be6f8bac..fed74afd45d1 100644
> > > --- a/kernel/bpf/helpers.c
> > > +++ b/kernel/bpf/helpers.c
> > > @@ -2094,11 +2094,12 @@ __bpf_kfunc struct cgroup *bpf_cgroup_ancestor(struct cgroup *cgrp, int level)
> > >  {
> > >  	struct cgroup *ancestor;
> > >  
> > > -	if (level > cgrp->level || level < 0)
> > > +	if (!cgrp || level > cgrp->level || level < 0)
> > >  		return NULL;
> > >  
> > >  	ancestor = cgrp->ancestors[level];
> > > -	cgroup_get(ancestor);
> > > +	if (!cgroup_tryget(ancestor))
> > > +		return NULL;
> > >  	return ancestor;
> > >  }
> > >  
> > > @@ -2183,7 +2184,7 @@ BTF_ID_FLAGS(func, bpf_rbtree_first, KF_RET_NULL)
> > >  BTF_ID_FLAGS(func, bpf_cgroup_acquire, KF_ACQUIRE | KF_TRUSTED_ARGS)
> > >  BTF_ID_FLAGS(func, bpf_cgroup_kptr_get, KF_ACQUIRE | KF_KPTR_GET | KF_RET_NULL)
> > >  BTF_ID_FLAGS(func, bpf_cgroup_release, KF_RELEASE)
> > > -BTF_ID_FLAGS(func, bpf_cgroup_ancestor, KF_ACQUIRE | KF_TRUSTED_ARGS | KF_RET_NULL)
> > > +BTF_ID_FLAGS(func, bpf_cgroup_ancestor, KF_ACQUIRE | KF_RCU | KF_RET_NULL)
> > >  BTF_ID_FLAGS(func, bpf_cgroup_from_id, KF_ACQUIRE | KF_RET_NULL)
> > >  #endif
> > >  BTF_ID_FLAGS(func, bpf_task_from_pid, KF_ACQUIRE | KF_RET_NULL)
> > > diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> > > index e3fcdc9836a6..2e730918911c 100644
> > > --- a/kernel/bpf/syscall.c
> > > +++ b/kernel/bpf/syscall.c
> > > @@ -539,6 +539,7 @@ void btf_record_free(struct btf_record *rec)
> > >  		switch (rec->fields[i].type) {
> > >  		case BPF_KPTR_UNREF:
> > >  		case BPF_KPTR_REF:
> > > +		case BPF_KPTR_RCU:
> > >  			if (rec->fields[i].kptr.module)
> > >  				module_put(rec->fields[i].kptr.module);
> > >  			btf_put(rec->fields[i].kptr.btf);
> > > @@ -584,6 +585,7 @@ struct btf_record *btf_record_dup(const struct btf_record *rec)
> > >  		switch (fields[i].type) {
> > >  		case BPF_KPTR_UNREF:
> > >  		case BPF_KPTR_REF:
> > > +		case BPF_KPTR_RCU:
> > >  			btf_get(fields[i].kptr.btf);
> > >  			if (fields[i].kptr.module && !try_module_get(fields[i].kptr.module)) {
> > >  				ret = -ENXIO;
> > > @@ -669,6 +671,7 @@ void bpf_obj_free_fields(const struct btf_record *rec, void *obj)
> > >  			WRITE_ONCE(*(u64 *)field_ptr, 0);
> > >  			break;
> > >  		case BPF_KPTR_REF:
> > > +		case BPF_KPTR_RCU:
> > 
> > The fact that we're adding this case is IMO a sign that we're arguably
> > breaking abstractions a bit. BPF_KPTR_REF should really be the kptr type
> > that holds a reference and for which we should be firing the destructor,
> > and RCU protection should ideally be something we could just derive
> > later in the verifier. 
> 
> I've considered keeping BPF_KPTR_REF as-is and just add a "bool is_kptr_rcu;"
> to indicate it's a BPF_KPTR_REF with extra RCU properties, but they are different
> enough. So it's cleaner to make them stand out.
> With BPF_KPTR_RCU being different type it's impossible for other bits
> in the verifier to silently accept BPF_KPTR_REF that shouldn't have RCU property.

Makes sense, will reply below.

> > Not a huge problem given that this complexity is
> > completely hidden from the user, but I'm not fully understanding why
> > the extra complexity of BPF_KPTR_RCU is necessary. See below in another
> > comment in verifier.c.
> > 
> > >  			field->kptr.dtor((void *)xchg((unsigned long *)field_ptr, 0));
> > 
> > Also completely unrelated to your patch set, but we should probably only
> > invoke field->kptr.dtor() if the value in field_ptr ends up being
> > non-NULL after the xchg. Otherwise, all KF_RELEASE kfuncs have to check
> > for NULL, even though they expect inherently trusted args. I can also do
> > that in a follow-on patch.
> 
> Good point. The verifier forces bpf progs to do if (ptr != NULL) bpf_..__release(ptr);
> but we still have duplicated !=NULL check inside dtor-s,
> because both BPF_KPTR_RCU and BPF_KPTR_REF can be NULL here.
> It would be good to clean up indeed.

Great, I'll send a patch after this set lands.

> > 
> > >  			break;
> > >  		case BPF_LIST_HEAD:
> > > @@ -1058,6 +1061,7 @@ static int map_check_btf(struct bpf_map *map, const struct btf *btf,
> > >  				break;
> > >  			case BPF_KPTR_UNREF:
> > >  			case BPF_KPTR_REF:
> > > +			case BPF_KPTR_RCU:
> > >  				if (map->map_type != BPF_MAP_TYPE_HASH &&
> > >  				    map->map_type != BPF_MAP_TYPE_LRU_HASH &&
> > >  				    map->map_type != BPF_MAP_TYPE_ARRAY &&
> > > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > > index e4234266e76d..0b728ce0dde9 100644
> > > --- a/kernel/bpf/verifier.c
> > > +++ b/kernel/bpf/verifier.c
> > > @@ -4183,7 +4183,7 @@ static int map_kptr_match_type(struct bpf_verifier_env *env,
> > >  			       struct bpf_reg_state *reg, u32 regno)
> > >  {
> > >  	const char *targ_name = kernel_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
> > > -	int perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED;
> > > +	int perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU;
> > >  	const char *reg_name = "";
> > >  
> > >  	/* Only unreferenced case accepts untrusted pointers */
> > > @@ -4230,12 +4230,12 @@ static int map_kptr_match_type(struct bpf_verifier_env *env,
> > >  	 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
> > >  	 * is zero. We must also ensure that btf_struct_ids_match does not walk
> > >  	 * the struct to match type against first member of struct, i.e. reject
> > > -	 * second case from above. Hence, when type is BPF_KPTR_REF, we set
> > > +	 * second case from above. Hence, when type is BPF_KPTR_REF | BPF_KPTR_RCU, we set
> > >  	 * strict mode to true for type match.
> > >  	 */
> > >  	if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
> > >  				  kptr_field->kptr.btf, kptr_field->kptr.btf_id,
> > > -				  kptr_field->type == BPF_KPTR_REF))
> > > +				  kptr_field->type == BPF_KPTR_REF || kptr_field->type == BPF_KPTR_RCU))
> > >  		goto bad_type;
> > >  	return 0;
> > >  bad_type:
> > > @@ -4250,6 +4250,14 @@ static int map_kptr_match_type(struct bpf_verifier_env *env,
> > >  	return -EINVAL;
> > >  }
> > >  
> > > +/* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock()
> > > + * can dereference RCU protected pointers and result is PTR_TRUSTED.
> > > + */
> > > +static bool in_rcu_cs(struct bpf_verifier_env *env)
> > > +{
> > > +	return env->cur_state->active_rcu_lock || !env->prog->aux->sleepable;
> > > +}
> > > +
> > >  static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
> > >  				 int value_regno, int insn_idx,
> > >  				 struct btf_field *kptr_field)
> > > @@ -4273,7 +4281,7 @@ static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
> > >  	/* We only allow loading referenced kptr, since it will be marked as
> > >  	 * untrusted, similar to unreferenced kptr.
> > >  	 */
> > > -	if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) {
> > > +	if (class != BPF_LDX && kptr_field->type != BPF_KPTR_UNREF) {
> > >  		verbose(env, "store to referenced kptr disallowed\n");
> > >  		return -EACCES;
> > >  	}
> > > @@ -4284,7 +4292,10 @@ static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
> > >  		 * value from map as PTR_TO_BTF_ID, with the correct type.
> > >  		 */
> > >  		mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf,
> > > -				kptr_field->kptr.btf_id, PTR_MAYBE_NULL | PTR_UNTRUSTED);
> > > +				kptr_field->kptr.btf_id,
> > > +				kptr_field->type == BPF_KPTR_RCU && in_rcu_cs(env) ?
> > 
> > If we replaced this kptr_field->type == BPF_KPTR_RCU check with
> > something like btf_rcu_safe_kptr(kptr_field), corresponding to:
> > 
> > bool btf_rcu_safe_kptr(const struct btf_field *field)
> > {
> > 	const struct btf_field_kptr *kptr = &field->kptr;
> > 
> > 	return field->type == BPF_KPTR_REF && rcu_protected_object(kptr->btf, kptr->btf_id);
> > }
> > 
> > Wouldn't that allow us to avoid having to define BPF_KPTR_RCU at all?
> > Given that BPF_KPTR_RCU is really just an instance of BPF_KPTR_REF which
> > may also derive safety from RCU protection, this seems both simpler and
> > more thematic. Or am I missing something?
> 
> See my earlier reply. It felt cleaner to keep them separate so that
> BPF_KPTR_RCU won't be accepted in placed where only BPF_KPTR_REF is ok.
> I'm probably overthinking.
> Looking at the code again all places with BPF_KPTR_REF were appended with BPF_KPTR_RCU.
> So, yeah, let's go with your suggestion above. A lot less code to maintain and

Yep, that's what I was thinking. The fact that there are so many places
where we're appending BPF_KPTR_RCU to BPF_KPTR_REF suggests to me
they're more similar than they are different, so either a bool
rcu_protected field as you suggested above, or a dynamic check as I
suggested, seems like the preferable tradeoff.

> if it turns out to be an issue we can go back to separate types.

Sounds good to me.