Re: [RFC PATCH bpf-next 04/11] bpf: Add rbtree map

[Alexei Starovoitov <ast@fb.com>]

On 7/22/22 11:34 AM, Dave Marchevsky wrote:
> +
> +static struct rb_node *rbtree_map_alloc_node(struct bpf_map *map, size_t sz)
> +{
> +	struct rb_node *node;
> +
> +	node = bpf_map_kmalloc_node(map, sz, GFP_KERNEL, map->numa_node);

As Yonghong pointed out this should be GFP_NOWAIT for now.
Later we can convert this to bpf_mem_alloc to make sure it's safe from 
any context.

> +	if (!node)
> +		return NULL;
> +	RB_CLEAR_NODE(node);
> +	return node;
> +}
> +
> +BPF_CALL_2(bpf_rbtree_alloc_node, struct bpf_map *, map, u32, sz)
> +{
> +	struct rb_node *node;
> +
> +	if (map->map_type != BPF_MAP_TYPE_RBTREE)
> +		return (u64)NULL;
> +
> +	if (sz < sizeof(*node))
> +		return (u64)NULL;
> +
> +	node = rbtree_map_alloc_node(map, (size_t)sz);
> +	if (!node)
> +		return (u64)NULL;
> +
> +	return (u64)node;
> +}
> +
> +const struct bpf_func_proto bpf_rbtree_alloc_node_proto = {
> +	.func = bpf_rbtree_alloc_node,
> +	.gpl_only = true,
> +	.ret_type = RET_PTR_TO_BTF_ID_OR_NULL,
> +	.ret_btf_id = &bpf_rbtree_btf_ids[0],

since the btf_id is unused please use
.ret_btf_id   = BPF_PTR_POISON
as bpf_kptr_xchg_proto() is doing.

> +
> +BPF_CALL_2(bpf_rbtree_remove, struct bpf_map *, map, void *, value)
> +{
> +	struct bpf_rbtree *tree = container_of(map, struct bpf_rbtree, map);
> +	struct rb_node *node = (struct rb_node *)value;
> +
> +	if (WARN_ON_ONCE(RB_EMPTY_NODE(node)))
> +		return (u64)NULL;
> +
> +	rb_erase_cached(node, &tree->root);
> +	RB_CLEAR_NODE(node);
> +	return (u64)node;
> +}
> +
> +const struct bpf_func_proto bpf_rbtree_remove_proto = {
> +	.func = bpf_rbtree_remove,
> +	.gpl_only = true,
> +	.ret_type = RET_PTR_TO_BTF_ID_OR_NULL,
> +	.ret_btf_id = &bpf_rbtree_btf_ids[0],
> +	.arg1_type = ARG_CONST_MAP_PTR,
> +	.arg2_type = ARG_PTR_TO_BTF_ID,
> +	.arg2_btf_id = &bpf_rbtree_btf_ids[0],

same for args.

> +
> +BTF_ID_LIST_SINGLE(bpf_rbtree_map_btf_ids, struct, bpf_rbtree)

can be removed?

> +const struct bpf_map_ops rbtree_map_ops = {
> +	.map_meta_equal = bpf_map_meta_equal,
> +	.map_alloc_check = rbtree_map_alloc_check,
> +	.map_alloc = rbtree_map_alloc,
> +	.map_free = rbtree_map_free,
> +	.map_get_next_key = rbtree_map_get_next_key,
> +	.map_push_elem = rbtree_map_push_elem,
> +	.map_peek_elem = rbtree_map_peek_elem,
> +	.map_pop_elem = rbtree_map_pop_elem,
> +	.map_lookup_elem = rbtree_map_lookup_elem,
> +	.map_update_elem = rbtree_map_update_elem,
> +	.map_delete_elem = rbtree_map_delete_elem,
> +	.map_check_btf = rbtree_map_check_btf,
> +	.map_btf_id = &bpf_rbtree_map_btf_ids[0],
> +};
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index 1f50becce141..535f673882cd 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -481,7 +481,9 @@ static bool is_acquire_function(enum bpf_func_id func_id,
>   	    func_id == BPF_FUNC_sk_lookup_udp ||
>   	    func_id == BPF_FUNC_skc_lookup_tcp ||
>   	    func_id == BPF_FUNC_ringbuf_reserve ||
> -	    func_id == BPF_FUNC_kptr_xchg)
> +	    func_id == BPF_FUNC_kptr_xchg ||
> +	    func_id == BPF_FUNC_rbtree_alloc_node ||
> +	    func_id == BPF_FUNC_rbtree_remove)
>   		return true;
>   
>   	if (func_id == BPF_FUNC_map_lookup_elem &&
> @@ -531,6 +533,20 @@ static bool is_cmpxchg_insn(const struct bpf_insn *insn)
>   	       insn->imm == BPF_CMPXCHG;
>   }
>   
> +static bool function_manipulates_rbtree_node(enum bpf_func_id func_id)
> +{
> +	return func_id == BPF_FUNC_rbtree_add ||
> +		func_id == BPF_FUNC_rbtree_remove ||
> +		func_id == BPF_FUNC_rbtree_free_node;
> +}
> +
> +static bool function_returns_rbtree_node(enum bpf_func_id func_id)
> +{
> +	return func_id == BPF_FUNC_rbtree_alloc_node ||
> +		func_id == BPF_FUNC_rbtree_add ||
> +		func_id == BPF_FUNC_rbtree_remove;
> +}
> +
>   /* string representation of 'enum bpf_reg_type'
>    *
>    * Note that reg_type_str() can not appear more than once in a single verbose()
> @@ -3784,6 +3800,13 @@ static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
>   	return 0;
>   }
>

 >    * [ TODO: Existing logic prevents any writes to PTR_TO_BTF_ID. This
 >      broadly turned off in this patch and replaced with "no writes to
 >      struct rb_node is PTR_TO_BTF_ID struct has one". This is a hack and
 >      needs to be replaced. ]

..

> +static bool access_may_touch_field(u32 access_off, size_t access_sz,

can_write is more accurate.
There is no ambiguity here. atype == BPF_WRITE.

> +				   u32 field_off, size_t field_sz)
> +{
> +	return access_off < field_off + field_sz &&
> +		field_off < access_off + access_sz;
> +}
> +
>   /* if any part of struct field can be touched by
>    * load/store reject this program.
>    * To check that [x1, x2) overlaps with [y1, y2)
> @@ -4490,7 +4513,7 @@ static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
>   	const char *tname = btf_name_by_offset(reg->btf, t->name_off);
>   	enum bpf_type_flag flag = 0;
>   	u32 btf_id;
> -	int ret;
> +	int ret, rb_node_off;
>   
>   	if (off < 0) {
>   		verbose(env,
> @@ -4527,8 +4550,13 @@ static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
>   						  off, size, atype, &btf_id, &flag);
>   	} else {
>   		if (atype != BPF_READ) {
> -			verbose(env, "only read is supported\n");
> -			return -EACCES;
> +			rb_node_off = btf_find_rb_node(reg->btf, t);
> +			if (rb_node_off < 0 ||
> +			    access_may_touch_field(off, size, rb_node_off,
> +						   sizeof(struct rb_node))) {
> +				verbose(env, "only read is supported\n");
> +				return -EACCES;
> +			}

Allowing writes into ptr_to_btf_id probably should be a separate patch.
It's a big change.
btf_find_rb_node() alone is not enough.
Otherwise bpf progs will be able to write into any struct that has 
'rb_node'.
Maybe check that reg->btf == this prog's btf ?
Also allow writes into scalars only?
All pointers in prog's struct should be __kptr anyway to be safe.


>   		}
>   
>   		ret = btf_struct_access(&env->log, reg->btf, t, off, size,
> @@ -5764,6 +5792,17 @@ static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
>   		if (meta->func_id == BPF_FUNC_kptr_xchg) {
>   			if (map_kptr_match_type(env, meta->kptr_off_desc, reg, regno))
>   				return -EACCES;
> +		} else if (function_manipulates_rbtree_node(meta->func_id)) {
> +			if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
> +						  meta->map_ptr->btf,
> +						  meta->map_ptr->btf_value_type_id,
> +						  strict_type_match)) {
> +				verbose(env, "rbtree: R%d is of type %s but %s is expected\n",
> +					regno, kernel_type_name(reg->btf, reg->btf_id),
> +					kernel_type_name(meta->map_ptr->btf,
> +							 meta->map_ptr->btf_value_type_id));
> +				return -EACCES;
> +			}
>   		} else if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
>   						 btf_vmlinux, *arg_btf_id,
>   						 strict_type_match)) {
> @@ -6369,10 +6408,17 @@ static int check_map_func_compatibility(struct bpf_verifier_env *env,
>   		break;
>   	case BPF_FUNC_map_pop_elem:
>   		if (map->map_type != BPF_MAP_TYPE_QUEUE &&
> +		    map->map_type != BPF_MAP_TYPE_RBTREE &&
>   		    map->map_type != BPF_MAP_TYPE_STACK)
>   			goto error;
>   		break;
>   	case BPF_FUNC_map_peek_elem:
> +		if (map->map_type != BPF_MAP_TYPE_QUEUE &&
> +		    map->map_type != BPF_MAP_TYPE_STACK &&
> +		    map->map_type != BPF_MAP_TYPE_RBTREE &&
> +		    map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
> +			goto error;
> +		break;
>   	case BPF_FUNC_map_push_elem:
>   		if (map->map_type != BPF_MAP_TYPE_QUEUE &&
>   		    map->map_type != BPF_MAP_TYPE_STACK &&
> @@ -6828,6 +6874,57 @@ static int set_loop_callback_state(struct bpf_verifier_env *env,
>   	return 0;
>   }
>   
> +static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
> +					 struct bpf_func_state *caller,
> +					 struct bpf_func_state *callee,
> +					 int insn_idx)
> +{
> +	struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
> +
> +	/* bpf_rbtree_add(struct bpf_map *map, void *value, void *cb)
> +	 * cb(struct rb_node *a, const struct rb_node *b);
> +	 */
> +	callee->regs[BPF_REG_1].type = PTR_TO_MAP_VALUE;
> +	__mark_reg_known_zero(&callee->regs[BPF_REG_1]);
> +	callee->regs[BPF_REG_1].map_ptr = map_ptr;
> +
> +	callee->regs[BPF_REG_2].type = PTR_TO_MAP_VALUE;
> +	__mark_reg_known_zero(&callee->regs[BPF_REG_2]);
> +	callee->regs[BPF_REG_2].map_ptr = map_ptr;
> +
> +	__mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
> +	__mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
> +	__mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
> +	callee->in_callback_fn = true;
> +	return 0;
> +}
> +
> +static int set_rbtree_find_callback_state(struct bpf_verifier_env *env,
> +					  struct bpf_func_state *caller,
> +					  struct bpf_func_state *callee,
> +					  int insn_idx)
> +{
> +	struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
> +
> +	/* bpf_rbtree_find(struct bpf_map *map, void *key, void *cb)
> +	 * cb(void *key, const struct rb_node *b);
> +	 */
> +	callee->regs[BPF_REG_1].type = PTR_TO_MAP_VALUE;
> +	__mark_reg_known_zero(&callee->regs[BPF_REG_1]);
> +	callee->regs[BPF_REG_1].map_ptr = map_ptr;
> +
> +	callee->regs[BPF_REG_2].type = PTR_TO_MAP_VALUE;
> +	__mark_reg_known_zero(&callee->regs[BPF_REG_2]);
> +	callee->regs[BPF_REG_2].map_ptr = map_ptr;
> +
> +	__mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
> +	__mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
> +	__mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
> +	callee->in_callback_fn = true;

add and find looks the same until this point.
Reuse set_rbtree_add_callback_state here?

> +	callee->callback_ret_range = tnum_range(0, U64_MAX);

that's to enforce that add's cb can only return 0 or 1 ?
But that would require bpf prog to have different cb-s for add and find.
Is this ok?

> +	return 0;
> +}
> +
>   static int set_timer_callback_state(struct bpf_verifier_env *env,
>   				    struct bpf_func_state *caller,
>   				    struct bpf_func_state *callee,
> @@ -7310,6 +7407,14 @@ static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn
>   		err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
>   					set_loop_callback_state);
>   		break;
> +	case BPF_FUNC_rbtree_add:
> +		err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
> +					set_rbtree_add_callback_state);
> +		break;
> +	case BPF_FUNC_rbtree_find:
> +		err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
> +					set_rbtree_find_callback_state);
> +		break;
>   	case BPF_FUNC_dynptr_from_mem:
>   		if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
>   			verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
> @@ -7424,6 +7529,9 @@ static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn
>   		if (func_id == BPF_FUNC_kptr_xchg) {
>   			ret_btf = meta.kptr_off_desc->kptr.btf;
>   			ret_btf_id = meta.kptr_off_desc->kptr.btf_id;
> +		} else if (function_returns_rbtree_node(func_id)) {
> +			ret_btf = meta.map_ptr->btf;
> +			ret_btf_id = meta.map_ptr->btf_value_type_id;
>   		} else {
>   			ret_btf = btf_vmlinux;
>   			ret_btf_id = *fn->ret_btf_id;
> @@ -13462,8 +13570,10 @@ static int convert_ctx_accesses(struct bpf_verifier_env *env)
>   					BPF_SIZE((insn)->code);
>   				env->prog->aux->num_exentries++;
>   			} else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
> +				/*TODO: Not sure what to do here
>   				verbose(env, "Writes through BTF pointers are not allowed\n");
>   				return -EINVAL;
> +				*/

Not sure whether it's worth defining PTR_TO_BTF_ID | PROGS_BTF
for writeable ptr_to_btf_id as return value from rb_alloc/find/add.
It may help here and earlier ?

All ptr_to_btf_id were kernel's or module's BTF so far.
With this change the verifier will see prog's ptr_to_btf_id that point
to prog's BTF.
PROGS_BTF might be a useful flag to have ?

Then instead of
 > +		} else if (function_returns_rbtree_node(func_id)) {
 > +			ret_btf = meta.map_ptr->btf;
 > +			ret_btf_id = meta.map_ptr->btf_value_type_id;

it will check PROGS_BTF flag in proto->ret_type ?
Still not fully generic, since it takes btf and btf_id from the map.
But a bit cleaner than switch() by func_id ?