Commit 07016151 authored by Daniel Borkmann's avatar Daniel Borkmann Committed by David S. Miller

bpf, verifier: further improve search pruning

The verifier needs to go through every path of the program in
order to check that it terminates safely, which can be quite a
lot of instructions that need to be processed f.e. in cases with
more branchy programs. With search pruning from f1bca824 ("bpf:
add search pruning optimization to verifier") the search space can
already be reduced significantly when the verifier detects that
a previously walked path with same register and stack contents
terminated already (see verifier's states_equal()), so the search
can skip walking those states.

When working with larger programs of > ~2000 (out of max 4096)
insns, we found that the current limit of 32k instructions is easily
hit. For example, a case we ran into is that the search space cannot
be pruned due to branches at the beginning of the program that make
use of certain stack space slots (STACK_MISC), which are never used
in the remaining program (STACK_INVALID). Therefore, the verifier
needs to walk paths for the slots in STACK_INVALID state, but also
all remaining paths with a stack structure, where the slots are in
STACK_MISC, which can nearly double the search space needed. After
various experiments, we find that a limit of 64k processed insns is
a more reasonable choice when dealing with larger programs in practice.
This still allows to reject extreme crafted cases that can have a
much higher complexity (f.e. > ~300k) within the 4096 insns limit
due to search pruning not being able to take effect.

Furthermore, we found that a lot of states can be pruned after a
call instruction, f.e. we were able to reduce the search state by
~35% in some cases with this heuristic, trade-off is to keep a bit
more states in env->explored_states. Usually, call instructions
have a number of preceding register assignments and/or stack stores,
where search pruning has a better chance to suceed in states_equal()
test. The current code marks the branch targets with STATE_LIST_MARK
in case of conditional jumps, and the next (t + 1) instruction in
case of unconditional jump so that f.e. a backjump will walk it. We
also did experiments with using t + insns[t].off + 1 as a marker in
the unconditionally jump case instead of t + 1 with the rationale
that these two branches of execution that converge after the label
might have more potential of pruning. We found that it was a bit
better, but not necessarily significantly better than the current
state, perhaps also due to clang not generating back jumps often.
Hence, we left that as is for now.
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Acked-by: default avatarAlexei Starovoitov <ast@kernel.org>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent 1fc2257e
......@@ -202,6 +202,9 @@ struct verifier_env {
bool allow_ptr_leaks;
};
#define BPF_COMPLEXITY_LIMIT_INSNS 65536
#define BPF_COMPLEXITY_LIMIT_STACK 1024
/* verbose verifier prints what it's seeing
* bpf_check() is called under lock, so no race to access these global vars
*/
......@@ -454,7 +457,7 @@ static struct verifier_state *push_stack(struct verifier_env *env, int insn_idx,
elem->next = env->head;
env->head = elem;
env->stack_size++;
if (env->stack_size > 1024) {
if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
verbose("BPF program is too complex\n");
goto err;
}
......@@ -1543,6 +1546,8 @@ static int check_cfg(struct verifier_env *env)
goto peek_stack;
else if (ret < 0)
goto err_free;
if (t + 1 < insn_cnt)
env->explored_states[t + 1] = STATE_LIST_MARK;
} else if (opcode == BPF_JA) {
if (BPF_SRC(insns[t].code) != BPF_K) {
ret = -EINVAL;
......@@ -1747,7 +1752,7 @@ static int do_check(struct verifier_env *env)
insn = &insns[insn_idx];
class = BPF_CLASS(insn->code);
if (++insn_processed > 32768) {
if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
verbose("BPF program is too large. Proccessed %d insn\n",
insn_processed);
return -E2BIG;
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment