When comparing current and cached states verifier should consider
bpf_func_state->callback_depth. Current state cannot be pruned against
cached state, when current states has more iterations left compared to
cached state. Current state has more iterations left when it's
callback_depth is smaller.

Below is an example illustrating this bug, minimized from mailing list
discussion [0] (assume that BPF_F_TEST_STATE_FREQ is set).
The example is not a safe program: if loop_cb point (1) is followed by
loop_cb point (2), then division by zero is possible at point (4).

    struct ctx {
    	__u64 a;
    	__u64 b;
    	__u64 c;
    };

    static void loop_cb(int i, struct ctx *ctx)
    {
    	/* assume that generated code is "fallthrough-first":
    	 * if ... == 1 goto
    	 * if ... == 2 goto
    	 * <default>
    	 */
    	switch (bpf_get_prandom_u32()) {
    	case 1:  /* 1 */ ctx->a = 42; return 0; break;
    	case 2:  /* 2 */ ctx->b = 42; return 0; break;
    	default: /* 3 */ ctx->c = 42; return 0; break;
    	}
    }

    SEC("tc")
    __failure
    __flag(BPF_F_TEST_STATE_FREQ)
    int test(struct __sk_buff *skb)
    {
    	struct ctx ctx = { 7, 7, 7 };

    	bpf_loop(2, loop_cb, &ctx, 0);              /* 0 */
    	/* assume generated checks are in-order: .a first */
    	if (ctx.a == 42 && ctx.b == 42 && ctx.c == 7)
    		asm volatile("r0 /= 0;":::"r0");    /* 4 */
    	return 0;
    }

Prior to this commit verifier built the following checkpoint tree for
this example:

 .------------------------------------- Checkpoint / State name
 |    .-------------------------------- Code point number
 |    |   .---------------------------- Stack state {ctx.a,ctx.b,ctx.c}
 |    |   |        .------------------- Callback depth in frame #0
 v    v   v        v
   - (0) {7P,7P,7},depth=0
     - (3) {7P,7P,7},depth=1
       - (0) {7P,7P,42},depth=1
         - (3) {7P,7,42},depth=2
           - (0) {7P,7,42},depth=2      loop terminates because of depth limit
             - (4) {7P,7,42},depth=0    predicted false, ctx.a marked precise
             - (6) exit
(a)      - (2) {7P,7,42},depth=2
           - (0) {7P,42,42},depth=2     loop terminates because of depth limit
             - (4) {7P,42,42},depth=0   predicted false, ctx.a marked precise
             - (6) exit
(b)      - (1) {7P,7P,42},depth=2
           - (0) {42P,7P,42},depth=2    loop terminates because of depth limit
             - (4) {42P,7P,42},depth=0  predicted false, ctx.{a,b} marked precise
             - (6) exit
     - (2) {7P,7,7},depth=1             considered safe, pruned using checkpoint (a)
(c)  - (1) {7P,7P,7},depth=1            considered safe, pruned using checkpoint (b)

Here checkpoint (b) has callback_depth of 2, meaning that it would
never reach state {42,42,7}.
While checkpoint (c) has callback_depth of 1, and thus
could yet explore the state {42,42,7} if not pruned prematurely.
This commit makes forbids such premature pruning,
allowing verifier to explore states sub-tree starting at (c):

(c)  - (1) {7,7,7P},depth=1
       - (0) {42P,7,7P},depth=1
         ...
         - (2) {42,7,7},depth=2
           - (0) {42,42,7},depth=2      loop terminates because of depth limit
             - (4) {42,42,7},depth=0    predicted true, ctx.{a,b,c} marked precise
               - (5) division by zero

[0] https://lore.kernel.org/bpf/9b251840-7cb8-4d17-bd23-1fc8071d8eef@linux.dev/

Fixes: bb124da6 ("bpf: keep track of max number of bpf_loop callback iterations")
Suggested-by: Yonghong Song <yonghong.song@linux.dev>
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
Acked-by: Yonghong Song <yonghong.song@linux.dev>
Link: https://lore.kernel.org/r/20240222154121.6991-2-eddyz87@gmail.comSigned-off-by: Alexei Starovoitov <ast@kernel.org>