bpf: ensure state checkpointing at iter_next() call sites

State equivalence check and checkpointing performed in is_state_visited()
employs certain heuristics to try to save memory by avoiding state checkpoints
if not enough jumps and instructions happened since last checkpoint. This leads
to unpredictability of whether a particular instruction will be checkpointed
and how regularly. While normally this is not causing much problems (except
inconveniences for predictable verifier tests, which we overcome with
BPF_F_TEST_STATE_FREQ flag), turns out it's not the case for open-coded
iterators.

Checking and saving state checkpoints at iter_next() call is crucial for fast
convergence of open-coded iterator loop logic, so we need to force it. If we
don't do that, is_state_visited() might skip saving a checkpoint, causing
unnecessarily long sequence of not checkpointed instructions and jumps, leading
to exhaustion of jump history buffer, and potentially other undesired outcomes.
It is expected that with correct open-coded iterators convergence will happen
quickly, so we don't run a risk of exhausting memory.

This patch adds, in addition to prune and jump instruction marks, also a
"forced checkpoint" mark, and makes sure that any iter_next() call instruction
is marked as such.
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/r/20230310060149.625887-1-andrii@kernel.orgSigned-off-by: Alexei Starovoitov <ast@kernel.org>

include/linux/bpf_verifier.h

@@ -477,8 +477,12 @@ struct bpf_insn_aux_data {
 
 	/* below fields are initialized once */
 	unsigned int orig_idx; /* original instruction index */
-	bool prune_point;
 	bool jmp_point;
+	bool prune_point;
+	/* ensure we check state equivalence and save state checkpoint and
+	 * this instruction, regardless of any heuristics
+	 */
+	bool force_checkpoint;
 };
 
 #define MAX_USED_MAPS 64 /* max number of maps accessed by one eBPF program */

kernel/bpf/verifier.c

@@ -13865,6 +13865,17 @@ static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
 	return env->insn_aux_data[insn_idx].prune_point;
 }
 
+static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
+{
+	env->insn_aux_data[idx].force_checkpoint = true;
+}
+
+static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
+{
+	return env->insn_aux_data[insn_idx].force_checkpoint;
+}
+
+
 enum {
 	DONE_EXPLORING = 0,
 	KEEP_EXPLORING = 1,
@@ -13984,8 +13995,21 @@ static int visit_insn(int t, struct bpf_verifier_env *env)
 			struct bpf_kfunc_call_arg_meta meta;
 
 			ret = fetch_kfunc_meta(env, insn, &meta, NULL);
-			if (ret == 0 && is_iter_next_kfunc(&meta))
+			if (ret == 0 && is_iter_next_kfunc(&meta)) {
 				mark_prune_point(env, t);
+				/* Checking and saving state checkpoints at iter_next() call
+				 * is crucial for fast convergence of open-coded iterator loop
+				 * logic, so we need to force it. If we don't do that,
+				 * is_state_visited() might skip saving a checkpoint, causing
+				 * unnecessarily long sequence of not checkpointed
+				 * instructions and jumps, leading to exhaustion of jump
+				 * history buffer, and potentially other undesired outcomes.
+				 * It is expected that with correct open-coded iterators
+				 * convergence will happen quickly, so we don't run a risk of
+				 * exhausting memory.
+				 */
+				mark_force_checkpoint(env, t);
+			}
 		}
 		return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL);
 
@@ -15172,7 +15196,8 @@ static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
 	struct bpf_verifier_state_list *sl, **pprev;
 	struct bpf_verifier_state *cur = env->cur_state, *new;
 	int i, j, err, states_cnt = 0;
-	bool add_new_state = env->test_state_freq ? true : false;
+	bool force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx);
+	bool add_new_state = force_new_state;
 
 	/* bpf progs typically have pruning point every 4 instructions
 	 * http://vger.kernel.org/bpfconf2019.html#session-1
@@ -15269,7 +15294,7 @@ static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
 			 * at the end of the loop are likely to be useful in pruning.
 			 */
 skip_inf_loop_check:
-			if (!env->test_state_freq &&
+			if (!force_new_state &&
 			    env->jmps_processed - env->prev_jmps_processed < 20 &&
 			    env->insn_processed - env->prev_insn_processed < 100)
 				add_new_state = false;