Merge branch 'check-bpf_func_state-callback_depth-when-pruning-states'

Eduard Zingerman says:

====================
check bpf_func_state->callback_depth when pruning states

This patch-set fixes bug in states pruning logic hit in mailing list
discussion [0]. The details of the fix are in patch #1.

The main idea for the fix belongs to Yonghong Song,
mine contribution is merely in review and test cases.

There are some changes in verification performance:

File                       Program        Insns    (DIFF)  States  (DIFF)
-------------------------  -------------  ---------------  --------------
pyperf600_bpf_loop.bpf.o   on_event          +15 (+0.42%)     +0 (+0.00%)
strobemeta_bpf_loop.bpf.o  on_event        +857 (+37.95%)   +60 (+38.96%)
xdp_synproxy_kern.bpf.o    syncookie_tc   +2892 (+30.39%)  +109 (+36.33%)
xdp_synproxy_kern.bpf.o    syncookie_xdp  +2892 (+30.01%)  +109 (+36.09%)

(when tested on a subset of selftests identified by
 selftests/bpf/veristat.cfg and Cilium bpf object files from [4])

Changelog:
v2 [2] -> v3:
- fixes for verifier.c commit message as suggested by Yonghong;
- patch-set re-rerouted to 'bpf' tree as suggested in [2];
- patch for test_tcp_custom_syncookie is sent separately to 'bpf-next' [3].
- veristat results updated using 'bpf' tree as baseline and clang 16.

v1 [1] -> v2:
- patch #2 commit message updated to better reflect verifier behavior
  with regards to checkpoints tree (suggested by Yonghong);
- veristat results added (suggested by Andrii).

[0] https://lore.kernel.org/bpf/[email protected]/
[1] https://lore.kernel.org/bpf/[email protected]/
[2] https://lore.kernel.org/bpf/[email protected]/
[3] https://lore.kernel.org/bpf/[email protected]/
[4] https://github.com/anakryiko/cilium
====================

Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Alexei Starovoitov <[email protected]>

Author: Alexei Starovoitov

kernel/bpf/verifier.c

@@ -16602,6 +16602,9 @@ static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_stat
 {
 	int i;
 
+	if (old->callback_depth > cur->callback_depth)
+		return false;
+
 	for (i = 0; i < MAX_BPF_REG; i++)
 		if (!regsafe(env, &old->regs[i], &cur->regs[i],
 			     &env->idmap_scratch, exact))

tools/testing/selftests/bpf/progs/verifier_iterating_callbacks.c

@@ -239,4 +239,74 @@ int bpf_loop_iter_limit_nested(void *unused)
 	return 1000 * a + b + c;
 }
 
+struct iter_limit_bug_ctx {
+	__u64 a;
+	__u64 b;
+	__u64 c;
+};
+
+static __naked void iter_limit_bug_cb(void)
+{
+	/* This is the same as C code below, but written
+	 * in assembly to control which branches are fall-through.
+	 *
+	 *   switch (bpf_get_prandom_u32()) {
+	 *   case 1:  ctx->a = 42; break;
+	 *   case 2:  ctx->b = 42; break;
+	 *   default: ctx->c = 42; break;
+	 *   }
+	 */
+	asm volatile (
+	"r9 = r2;"
+	"call %[bpf_get_prandom_u32];"
+	"r1 = r0;"
+	"r2 = 42;"
+	"r0 = 0;"
+	"if r1 == 0x1 goto 1f;"
+	"if r1 == 0x2 goto 2f;"
+	"*(u64 *)(r9 + 16) = r2;"
+	"exit;"
+	"1: *(u64 *)(r9 + 0) = r2;"
+	"exit;"
+	"2: *(u64 *)(r9 + 8) = r2;"
+	"exit;"
+	:
+	: __imm(bpf_get_prandom_u32)
+	: __clobber_all
+	);
+}
+
+SEC("tc")
+__failure
+__flag(BPF_F_TEST_STATE_FREQ)
+int iter_limit_bug(struct __sk_buff *skb)
+{
+	struct iter_limit_bug_ctx ctx = { 7, 7, 7 };
+
+	bpf_loop(2, iter_limit_bug_cb, &ctx, 0);
+
+	/* This is the same as C code below,
+	 * written in assembly to guarantee checks order.
+	 *
+	 *   if (ctx.a == 42 && ctx.b == 42 && ctx.c == 7)
+	 *     asm volatile("r1 /= 0;":::"r1");
+	 */
+	asm volatile (
+	"r1 = *(u64 *)%[ctx_a];"
+	"if r1 != 42 goto 1f;"
+	"r1 = *(u64 *)%[ctx_b];"
+	"if r1 != 42 goto 1f;"
+	"r1 = *(u64 *)%[ctx_c];"
+	"if r1 != 7 goto 1f;"
+	"r1 /= 0;"
+	"1:"
+	:
+	: [ctx_a]"m"(ctx.a),
+	  [ctx_b]"m"(ctx.b),
+	  [ctx_c]"m"(ctx.c)
+	: "r1"
+	);
+	return 0;
+}
+
 char _license[] SEC("license") = "GPL";