bpf: Implement verifier support for validation of async callbacks.
authorAlexei Starovoitov <ast@kernel.org>
Thu, 15 Jul 2021 00:54:14 +0000 (17:54 -0700)
committerDaniel Borkmann <daniel@iogearbox.net>
Thu, 15 Jul 2021 20:31:10 +0000 (22:31 +0200)
bpf_for_each_map_elem() and bpf_timer_set_callback() helpers are relying on
PTR_TO_FUNC infra in the verifier to validate addresses to subprograms
and pass them into the helpers as function callbacks.
In case of bpf_for_each_map_elem() the callback is invoked synchronously
and the verifier treats it as a normal subprogram call by adding another
bpf_func_state and new frame in __check_func_call().
bpf_timer_set_callback() doesn't invoke the callback directly.
The subprogram will be called asynchronously from bpf_timer_cb().
Teach the verifier to validate such async callbacks as special kind
of jump by pushing verifier state into stack and let pop_stack() process it.

Special care needs to be taken during state pruning.
The call insn doing bpf_timer_set_callback has to be a prune_point.
Otherwise short timer callbacks might not have prune points in front of
bpf_timer_set_callback() which means is_state_visited() will be called
after this call insn is processed in __check_func_call(). Which means that
another async_cb state will be pushed to be walked later and the verifier
will eventually hit BPF_COMPLEXITY_LIMIT_JMP_SEQ limit.
Since push_async_cb() looks like another push_stack() branch the
infinite loop detection will trigger false positive. To recognize
this case mark such states as in_async_callback_fn.
To distinguish infinite loop in async callback vs the same callback called
with different arguments for different map and timer add async_entry_cnt
to bpf_func_state.

Enforce return zero from async callbacks.

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/bpf/20210715005417.78572-9-alexei.starovoitov@gmail.com
include/linux/bpf_verifier.h
kernel/bpf/helpers.c
kernel/bpf/verifier.c

index 5d3169b..242d0b1 100644 (file)
@@ -208,12 +208,19 @@ struct bpf_func_state {
         * zero == main subprog
         */
        u32 subprogno;
+       /* Every bpf_timer_start will increment async_entry_cnt.
+        * It's used to distinguish:
+        * void foo(void) { for(;;); }
+        * void foo(void) { bpf_timer_set_callback(,foo); }
+        */
+       u32 async_entry_cnt;
+       bool in_callback_fn;
+       bool in_async_callback_fn;
 
        /* The following fields should be last. See copy_func_state() */
        int acquired_refs;
        struct bpf_reference_state *refs;
        int allocated_stack;
-       bool in_callback_fn;
        struct bpf_stack_state *stack;
 };
 
index 74b1659..9fe846e 100644 (file)
@@ -1043,7 +1043,6 @@ static enum hrtimer_restart bpf_timer_cb(struct hrtimer *hrtimer)
        void *callback_fn;
        void *key;
        u32 idx;
-       int ret;
 
        callback_fn = rcu_dereference_check(t->callback_fn, rcu_read_lock_bh_held());
        if (!callback_fn)
@@ -1066,10 +1065,9 @@ static enum hrtimer_restart bpf_timer_cb(struct hrtimer *hrtimer)
                key = value - round_up(map->key_size, 8);
        }
 
-       ret = BPF_CAST_CALL(callback_fn)((u64)(long)map,
-                                        (u64)(long)key,
-                                        (u64)(long)value, 0, 0);
-       WARN_ON(ret != 0); /* Next patch moves this check into the verifier */
+       BPF_CAST_CALL(callback_fn)((u64)(long)map, (u64)(long)key,
+                                  (u64)(long)value, 0, 0);
+       /* The verifier checked that return value is zero. */
 
        this_cpu_write(hrtimer_running, NULL);
 out:
index 1cb1b35..ab06256 100644 (file)
@@ -735,6 +735,10 @@ static void print_verifier_state(struct bpf_verifier_env *env,
                        if (state->refs[i].id)
                                verbose(env, ",%d", state->refs[i].id);
        }
+       if (state->in_callback_fn)
+               verbose(env, " cb");
+       if (state->in_async_callback_fn)
+               verbose(env, " async_cb");
        verbose(env, "\n");
 }
 
@@ -1527,6 +1531,54 @@ static void init_func_state(struct bpf_verifier_env *env,
        init_reg_state(env, state);
 }
 
+/* Similar to push_stack(), but for async callbacks */
+static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
+                                               int insn_idx, int prev_insn_idx,
+                                               int subprog)
+{
+       struct bpf_verifier_stack_elem *elem;
+       struct bpf_func_state *frame;
+
+       elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
+       if (!elem)
+               goto err;
+
+       elem->insn_idx = insn_idx;
+       elem->prev_insn_idx = prev_insn_idx;
+       elem->next = env->head;
+       elem->log_pos = env->log.len_used;
+       env->head = elem;
+       env->stack_size++;
+       if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
+               verbose(env,
+                       "The sequence of %d jumps is too complex for async cb.\n",
+                       env->stack_size);
+               goto err;
+       }
+       /* Unlike push_stack() do not copy_verifier_state().
+        * The caller state doesn't matter.
+        * This is async callback. It starts in a fresh stack.
+        * Initialize it similar to do_check_common().
+        */
+       elem->st.branches = 1;
+       frame = kzalloc(sizeof(*frame), GFP_KERNEL);
+       if (!frame)
+               goto err;
+       init_func_state(env, frame,
+                       BPF_MAIN_FUNC /* callsite */,
+                       0 /* frameno within this callchain */,
+                       subprog /* subprog number within this prog */);
+       elem->st.frame[0] = frame;
+       return &elem->st;
+err:
+       free_verifier_state(env->cur_state, true);
+       env->cur_state = NULL;
+       /* pop all elements and return */
+       while (!pop_stack(env, NULL, NULL, false));
+       return NULL;
+}
+
+
 enum reg_arg_type {
        SRC_OP,         /* register is used as source operand */
        DST_OP,         /* register is used as destination operand */
@@ -5704,6 +5756,30 @@ static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn
                }
        }
 
+       if (insn->code == (BPF_JMP | BPF_CALL) &&
+           insn->imm == BPF_FUNC_timer_set_callback) {
+               struct bpf_verifier_state *async_cb;
+
+               /* there is no real recursion here. timer callbacks are async */
+               async_cb = push_async_cb(env, env->subprog_info[subprog].start,
+                                        *insn_idx, subprog);
+               if (!async_cb)
+                       return -EFAULT;
+               callee = async_cb->frame[0];
+               callee->async_entry_cnt = caller->async_entry_cnt + 1;
+
+               /* Convert bpf_timer_set_callback() args into timer callback args */
+               err = set_callee_state_cb(env, caller, callee, *insn_idx);
+               if (err)
+                       return err;
+
+               clear_caller_saved_regs(env, caller->regs);
+               mark_reg_unknown(env, caller->regs, BPF_REG_0);
+               caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
+               /* continue with next insn after call */
+               return 0;
+       }
+
        callee = kzalloc(sizeof(*callee), GFP_KERNEL);
        if (!callee)
                return -ENOMEM;
@@ -5856,6 +5932,7 @@ static int set_timer_callback_state(struct bpf_verifier_env *env,
        /* unused */
        __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
        __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
+       callee->in_async_callback_fn = true;
        return 0;
 }
 
@@ -9224,7 +9301,8 @@ static int check_return_code(struct bpf_verifier_env *env)
        struct tnum range = tnum_range(0, 1);
        enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
        int err;
-       const bool is_subprog = env->cur_state->frame[0]->subprogno;
+       struct bpf_func_state *frame = env->cur_state->frame[0];
+       const bool is_subprog = frame->subprogno;
 
        /* LSM and struct_ops func-ptr's return type could be "void" */
        if (!is_subprog &&
@@ -9249,6 +9327,22 @@ static int check_return_code(struct bpf_verifier_env *env)
        }
 
        reg = cur_regs(env) + BPF_REG_0;
+
+       if (frame->in_async_callback_fn) {
+               /* enforce return zero from async callbacks like timer */
+               if (reg->type != SCALAR_VALUE) {
+                       verbose(env, "In async callback the register R0 is not a known value (%s)\n",
+                               reg_type_str[reg->type]);
+                       return -EINVAL;
+               }
+
+               if (!tnum_in(tnum_const(0), reg->var_off)) {
+                       verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
+                       return -EINVAL;
+               }
+               return 0;
+       }
+
        if (is_subprog) {
                if (reg->type != SCALAR_VALUE) {
                        verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
@@ -9496,6 +9590,13 @@ static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
                return DONE_EXPLORING;
 
        case BPF_CALL:
+               if (insns[t].imm == BPF_FUNC_timer_set_callback)
+                       /* Mark this call insn to trigger is_state_visited() check
+                        * before call itself is processed by __check_func_call().
+                        * Otherwise new async state will be pushed for further
+                        * exploration.
+                        */
+                       init_explored_state(env, t);
                return visit_func_call_insn(t, insn_cnt, insns, env,
                                            insns[t].src_reg == BPF_PSEUDO_CALL);
 
@@ -10503,9 +10604,25 @@ static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
                states_cnt++;
                if (sl->state.insn_idx != insn_idx)
                        goto next;
+
                if (sl->state.branches) {
-                       if (states_maybe_looping(&sl->state, cur) &&
-                           states_equal(env, &sl->state, cur)) {
+                       struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
+
+                       if (frame->in_async_callback_fn &&
+                           frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
+                               /* Different async_entry_cnt means that the verifier is
+                                * processing another entry into async callback.
+                                * Seeing the same state is not an indication of infinite
+                                * loop or infinite recursion.
+                                * But finding the same state doesn't mean that it's safe
+                                * to stop processing the current state. The previous state
+                                * hasn't yet reached bpf_exit, since state.branches > 0.
+                                * Checking in_async_callback_fn alone is not enough either.
+                                * Since the verifier still needs to catch infinite loops
+                                * inside async callbacks.
+                                */
+                       } else if (states_maybe_looping(&sl->state, cur) &&
+                                  states_equal(env, &sl->state, cur)) {
                                verbose_linfo(env, insn_idx, "; ");
                                verbose(env, "infinite loop detected at insn %d\n", insn_idx);
                                return -EINVAL;