bpf: convert explored_states to hash table
authorAlexei Starovoitov <ast@kernel.org>
Wed, 22 May 2019 03:17:07 +0000 (20:17 -0700)
committerDaniel Borkmann <daniel@iogearbox.net>
Thu, 23 May 2019 23:46:22 +0000 (01:46 +0200)
All prune points inside a callee bpf function most likely will have
different callsites. For example, if function foo() is called from
two callsites the half of explored states in all prune points in foo()
will be useless for subsequent walking of one of those callsites.
Fortunately explored_states pruning heuristics keeps the number of states
per prune point small, but walking these states is still a waste of cpu
time when the callsite of the current state is different from the callsite
of the explored state.

To improve pruning logic convert explored_states into hash table and
use simple insn_idx ^ callsite hash to select hash bucket.
This optimization has no effect on programs without bpf2bpf calls
and drastically improves programs with calls.
In the later case it reduces total memory consumption in 1M scale tests
by almost 3 times (peak_states drops from 5752 to 2016).

Care should be taken when comparing the states for equivalency.
Since the same hash bucket can now contain states with different indices
the insn_idx has to be part of verifier_state and compared.

Different hash table sizes and different hash functions were explored,
but the results were not significantly better vs this patch.
They can be improved in the future.

Hit/miss heuristic is not counting index miscompare as a miss.
Otherwise verifier stats become unstable when experimenting
with different hash functions.

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
include/linux/bpf_verifier.h
kernel/bpf/verifier.c

index 02bba09a0ea113f1dbd7a8a70b6388b82786b49d..405b502283c557f8b3f6a3b2be9394a65986d4af 100644 (file)
@@ -187,6 +187,7 @@ struct bpf_func_state {
 struct bpf_verifier_state {
        /* call stack tracking */
        struct bpf_func_state *frame[MAX_CALL_FRAMES];
+       u32 insn_idx;
        u32 curframe;
        u32 active_spin_lock;
        bool speculative;
index 6a3e69ba891eff858e7460e0a996d0e18b55ada8..550091c7a46a40741462ca6f0abd4da49acbe453 100644 (file)
@@ -5436,11 +5436,19 @@ enum {
        BRANCH = 2,
 };
 
+static u32 state_htab_size(struct bpf_verifier_env *env)
+{
+       return env->prog->len;
+}
+
 static struct bpf_verifier_state_list **explored_state(
                                        struct bpf_verifier_env *env,
                                        int idx)
 {
-       return &env->explored_states[idx];
+       struct bpf_verifier_state *cur = env->cur_state;
+       struct bpf_func_state *state = cur->frame[cur->curframe];
+
+       return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
 }
 
 static void init_explored_state(struct bpf_verifier_env *env, int idx)
@@ -6018,7 +6026,8 @@ static void clean_live_states(struct bpf_verifier_env *env, int insn,
 
        sl = *explored_state(env, insn);
        while (sl) {
-               if (sl->state.curframe != cur->curframe)
+               if (sl->state.insn_idx != insn ||
+                   sl->state.curframe != cur->curframe)
                        goto next;
                for (i = 0; i <= cur->curframe; i++)
                        if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
@@ -6384,6 +6393,9 @@ static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
        clean_live_states(env, insn_idx, cur);
 
        while (sl) {
+               states_cnt++;
+               if (sl->state.insn_idx != insn_idx)
+                       goto next;
                if (states_equal(env, &sl->state, cur)) {
                        sl->hit_cnt++;
                        /* reached equivalent register/stack state,
@@ -6401,7 +6413,6 @@ static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
                                return err;
                        return 1;
                }
-               states_cnt++;
                sl->miss_cnt++;
                /* heuristic to determine whether this state is beneficial
                 * to keep checking from state equivalence point of view.
@@ -6428,6 +6439,7 @@ static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
                        sl = *pprev;
                        continue;
                }
+next:
                pprev = &sl->next;
                sl = *pprev;
        }
@@ -6459,6 +6471,7 @@ static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
                kfree(new_sl);
                return err;
        }
+       new->insn_idx = insn_idx;
        new_sl->next = *explored_state(env, insn_idx);
        *explored_state(env, insn_idx) = new_sl;
        /* connect new state to parentage chain. Current frame needs all
@@ -8138,7 +8151,7 @@ static void free_states(struct bpf_verifier_env *env)
        if (!env->explored_states)
                return;
 
-       for (i = 0; i < env->prog->len; i++) {
+       for (i = 0; i < state_htab_size(env); i++) {
                sl = env->explored_states[i];
 
                while (sl) {
@@ -8246,7 +8259,7 @@ int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
                        goto skip_full_check;
        }
 
-       env->explored_states = kvcalloc(env->prog->len,
+       env->explored_states = kvcalloc(state_htab_size(env),
                                       sizeof(struct bpf_verifier_state_list *),
                                       GFP_USER);
        ret = -ENOMEM;