Quelle  iters.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/* Copyright (c) 2023 Meta Platforms, Inc. and affiliates. */

#include <stdbool.h>
#include <linux/bpf.h>
#include <bpf/bpf_helpers.h>
#include "bpf_misc.h"
#include "bpf_compiler.h"

static volatile int zero = 0;

int my_pid;
int arr[256];
int small_arr[16] SEC(".data.small_arr");

struct {
 __uint(type, BPF_MAP_TYPE_HASH);
 __uint(max_entries, 10);
 __type(key, int);
 __type(value, int);
} amap SEC(".maps");

#ifdef REAL_TEST
#define MY_PID_GUARD() if (my_pid != (bpf_get_current_pid_tgid() >> 32)) return 0
#else
#define MY_PID_GUARD() ({ })
#endif

SEC("?raw_tp")
__failure __msg("math between map_value pointer and register with unbounded min value is not allowed")
int iter_err_unsafe_c_loop(const void *ctx)
{
 struct bpf_iter_num it;
 int *v, i = zero; /* obscure initial value of i */

 MY_PID_GUARD();

 bpf_iter_num_new(&it, 0, 1000);
 while ((v = bpf_iter_num_next(&it))) {
  i++;
 }
 bpf_iter_num_destroy(&it);

 small_arr[i] = 123; /* invalid */

 return 0;
}

SEC("?raw_tp")
__failure __msg("unbounded memory access")
int iter_err_unsafe_asm_loop(const void *ctx)
{
 struct bpf_iter_num it;

 MY_PID_GUARD();

 asm volatile (
  "r6 = %[zero];" /* iteration counter */
  "r1 = %[it];" /* iterator state */
  "r2 = 0;"
  "r3 = 1000;"
  "r4 = 1;"
  "call %[bpf_iter_num_new];"
 "loop:"
  "r1 = %[it];"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto out;"
  "r6 += 1;"
  "goto loop;"
 "out:"
  "r1 = %[it];"
  "call %[bpf_iter_num_destroy];"
  "r1 = %[small_arr];"
  "r2 = r6;"
  "r2 <<= 2;"
  "r1 += r2;"
  "*(u32 *)(r1 + 0) = r6;" /* invalid */
  :
  : [it]"r"(&it),
    [small_arr]"r"(small_arr),
    [zero]"r"(zero),
    __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_next),
    __imm(bpf_iter_num_destroy)
  : __clobber_common, "r6"
 );

 return 0;
}

SEC("raw_tp")
__success
int iter_while_loop(const void *ctx)
{
 struct bpf_iter_num it;
 int *v;

 MY_PID_GUARD();

 bpf_iter_num_new(&it, 0, 3);
 while ((v = bpf_iter_num_next(&it))) {
  bpf_printk("ITER_BASIC: E1 VAL: v=%d", *v);
 }
 bpf_iter_num_destroy(&it);

 return 0;
}

SEC("raw_tp")
__success
int iter_while_loop_auto_cleanup(const void *ctx)
{
 __attribute__((cleanup(bpf_iter_num_destroy))) struct bpf_iter_num it;
 int *v;

 MY_PID_GUARD();

 bpf_iter_num_new(&it, 0, 3);
 while ((v = bpf_iter_num_next(&it))) {
  bpf_printk("ITER_BASIC: E1 VAL: v=%d", *v);
 }
 /* (!) no explicit bpf_iter_num_destroy() */

 return 0;
}

SEC("raw_tp")
__success
int iter_for_loop(const void *ctx)
{
 struct bpf_iter_num it;
 int *v;

 MY_PID_GUARD();

 bpf_iter_num_new(&it, 5, 10);
 for (v = bpf_iter_num_next(&it); v; v = bpf_iter_num_next(&it)) {
  bpf_printk("ITER_BASIC: E2 VAL: v=%d", *v);
 }
 bpf_iter_num_destroy(&it);

 return 0;
}

SEC("raw_tp")
__success
int iter_bpf_for_each_macro(const void *ctx)
{
 int *v;

 MY_PID_GUARD();

 bpf_for_each(num, v, 5, 10) {
  bpf_printk("ITER_BASIC: E2 VAL: v=%d", *v);
 }

 return 0;
}

SEC("raw_tp")
__success
int iter_bpf_for_macro(const void *ctx)
{
 int i;

 MY_PID_GUARD();

 bpf_for(i, 5, 10) {
  bpf_printk("ITER_BASIC: E2 VAL: v=%d", i);
 }

 return 0;
}

SEC("raw_tp")
__success
int iter_pragma_unroll_loop(const void *ctx)
{
 struct bpf_iter_num it;
 int *v, i;

 MY_PID_GUARD();

 bpf_iter_num_new(&it, 0, 2);
 __pragma_loop_no_unroll
 for (i = 0; i < 3; i++) {
  v = bpf_iter_num_next(&it);
  bpf_printk("ITER_BASIC: E3 VAL: i=%d v=%d", i, v ? *v : -1);
 }
 bpf_iter_num_destroy(&it);

 return 0;
}

SEC("raw_tp")
__success
int iter_manual_unroll_loop(const void *ctx)
{
 struct bpf_iter_num it;
 int *v;

 MY_PID_GUARD();

 bpf_iter_num_new(&it, 100, 200);
 v = bpf_iter_num_next(&it);
 bpf_printk("ITER_BASIC: E4 VAL: v=%d", v ? *v : -1);
 v = bpf_iter_num_next(&it);
 bpf_printk("ITER_BASIC: E4 VAL: v=%d", v ? *v : -1);
 v = bpf_iter_num_next(&it);
 bpf_printk("ITER_BASIC: E4 VAL: v=%d", v ? *v : -1);
 v = bpf_iter_num_next(&it);
 bpf_printk("ITER_BASIC: E4 VAL: v=%d\n", v ? *v : -1);
 bpf_iter_num_destroy(&it);

 return 0;
}

SEC("raw_tp")
__success
int iter_multiple_sequential_loops(const void *ctx)
{
 struct bpf_iter_num it;
 int *v, i;

 MY_PID_GUARD();

 bpf_iter_num_new(&it, 0, 3);
 while ((v = bpf_iter_num_next(&it))) {
  bpf_printk("ITER_BASIC: E1 VAL: v=%d", *v);
 }
 bpf_iter_num_destroy(&it);

 bpf_iter_num_new(&it, 5, 10);
 for (v = bpf_iter_num_next(&it); v; v = bpf_iter_num_next(&it)) {
  bpf_printk("ITER_BASIC: E2 VAL: v=%d", *v);
 }
 bpf_iter_num_destroy(&it);

 bpf_iter_num_new(&it, 0, 2);
 __pragma_loop_no_unroll
 for (i = 0; i < 3; i++) {
  v = bpf_iter_num_next(&it);
  bpf_printk("ITER_BASIC: E3 VAL: i=%d v=%d", i, v ? *v : -1);
 }
 bpf_iter_num_destroy(&it);

 bpf_iter_num_new(&it, 100, 200);
 v = bpf_iter_num_next(&it);
 bpf_printk("ITER_BASIC: E4 VAL: v=%d", v ? *v : -1);
 v = bpf_iter_num_next(&it);
 bpf_printk("ITER_BASIC: E4 VAL: v=%d", v ? *v : -1);
 v = bpf_iter_num_next(&it);
 bpf_printk("ITER_BASIC: E4 VAL: v=%d", v ? *v : -1);
 v = bpf_iter_num_next(&it);
 bpf_printk("ITER_BASIC: E4 VAL: v=%d\n", v ? *v : -1);
 bpf_iter_num_destroy(&it);

 return 0;
}

SEC("raw_tp")
__success
int iter_limit_cond_break_loop(const void *ctx)
{
 struct bpf_iter_num it;
 int *v, i = 0, sum = 0;

 MY_PID_GUARD();

 bpf_iter_num_new(&it, 0, 10);
 while ((v = bpf_iter_num_next(&it))) {
  bpf_printk("ITER_SIMPLE: i=%d v=%d", i, *v);
  sum += *v;

  i++;
  if (i > 3)
   break;
 }
 bpf_iter_num_destroy(&it);

 bpf_printk("ITER_SIMPLE: sum=%d\n", sum);

 return 0;
}

SEC("raw_tp")
__success
int iter_obfuscate_counter(const void *ctx)
{
 struct bpf_iter_num it;
 int *v, sum = 0;
 /* Make i's initial value unknowable for verifier to prevent it from
 * pruning if/else branch inside the loop body and marking i as precise.
 */
 int i = zero;

 MY_PID_GUARD();

 bpf_iter_num_new(&it, 0, 10);
 while ((v = bpf_iter_num_next(&it))) {
  int x;

  i += 1;

  /* If we initialized i as `int i = 0;` above, verifier would
 * track that i becomes 1 on first iteration after increment
 * above, and here verifier would eagerly prune else branch
 * and mark i as precise, ruining open-coded iterator logic
 * completely, as each next iteration would have a different
 * *precise* value of i, and thus there would be no
 * convergence of state. This would result in reaching maximum
 * instruction limit, no matter what the limit is.
 */
  if (i == 1)
   x = 123;
  else
   x = i * 3 + 1;

  bpf_printk("ITER_OBFUSCATE_COUNTER: i=%d v=%d x=%d", i, *v, x);

  sum += x;
 }
 bpf_iter_num_destroy(&it);

 bpf_printk("ITER_OBFUSCATE_COUNTER: sum=%d\n", sum);

 return 0;
}

SEC("raw_tp")
__success
int iter_search_loop(const void *ctx)
{
 struct bpf_iter_num it;
 int *v, *elem = NULL;
 bool found = false;

 MY_PID_GUARD();

 bpf_iter_num_new(&it, 0, 10);

 while ((v = bpf_iter_num_next(&it))) {
  bpf_printk("ITER_SEARCH_LOOP: v=%d", *v);

  if (*v == 2) {
   found = true;
   elem = v;
   barrier_var(elem);
  }
 }

 /* should fail to verify if bpf_iter_num_destroy() is here */

 if (found)
  /* here found element will be wrong, we should have copied
 * value to a variable, but here we want to make sure we can
 * access memory after the loop anyways
 */
  bpf_printk("ITER_SEARCH_LOOP: FOUND IT = %d!\n", *elem);
 else
  bpf_printk("ITER_SEARCH_LOOP: NOT FOUND IT!\n");

 bpf_iter_num_destroy(&it);

 return 0;
}

SEC("raw_tp")
__success
int iter_array_fill(const void *ctx)
{
 int sum, i;

 MY_PID_GUARD();

 bpf_for(i, 0, ARRAY_SIZE(arr)) {
  arr[i] = i * 2;
 }

 sum = 0;
 bpf_for(i, 0, ARRAY_SIZE(arr)) {
  sum += arr[i];
 }

 bpf_printk("ITER_ARRAY_FILL: sum=%d (should be %d)\n", sum, 255 * 256);

 return 0;
}

static int arr2d[4][5];
static int arr2d_row_sums[4];
static int arr2d_col_sums[5];

SEC("raw_tp")
__success
int iter_nested_iters(const void *ctx)
{
 int sum, row, col;

 MY_PID_GUARD();

 bpf_for(row, 0, ARRAY_SIZE(arr2d)) {
  bpf_for( col, 0, ARRAY_SIZE(arr2d[0])) {
   arr2d[row][col] = row * col;
  }
 }

 /* zero-initialize sums */
 sum = 0;
 bpf_for(row, 0, ARRAY_SIZE(arr2d)) {
  arr2d_row_sums[row] = 0;
 }
 bpf_for(col, 0, ARRAY_SIZE(arr2d[0])) {
  arr2d_col_sums[col] = 0;
 }

 /* calculate sums */
 bpf_for(row, 0, ARRAY_SIZE(arr2d)) {
  bpf_for(col, 0, ARRAY_SIZE(arr2d[0])) {
   sum += arr2d[row][col];
   arr2d_row_sums[row] += arr2d[row][col];
   arr2d_col_sums[col] += arr2d[row][col];
  }
 }

 bpf_printk("ITER_NESTED_ITERS: total sum=%d", sum);
 bpf_for(row, 0, ARRAY_SIZE(arr2d)) {
  bpf_printk("ITER_NESTED_ITERS: row #%d sum=%d", row, arr2d_row_sums[row]);
 }
 bpf_for(col, 0, ARRAY_SIZE(arr2d[0])) {
  bpf_printk("ITER_NESTED_ITERS: col #%d sum=%d%s",
      col, arr2d_col_sums[col],
      col == ARRAY_SIZE(arr2d[0]) - 1 ? "\n" : "");
 }

 return 0;
}

SEC("raw_tp")
__success
int iter_nested_deeply_iters(const void *ctx)
{
 int sum = 0;

 MY_PID_GUARD();

 bpf_repeat(10) {
  bpf_repeat(10) {
   bpf_repeat(10) {
    bpf_repeat(10) {
     bpf_repeat(10) {
      sum += 1;
     }
    }
   }
  }
  /* validate that we can break from inside bpf_repeat() */
  break;
 }

 return sum;
}

static __noinline void fill_inner_dimension(int row)
{
 int col;

 bpf_for(col, 0, ARRAY_SIZE(arr2d[0])) {
  arr2d[row][col] = row * col;
 }
}

static __noinline int sum_inner_dimension(int row)
{
 int sum = 0, col;

 bpf_for(col, 0, ARRAY_SIZE(arr2d[0])) {
  sum += arr2d[row][col];
  arr2d_row_sums[row] += arr2d[row][col];
  arr2d_col_sums[col] += arr2d[row][col];
 }

 return sum;
}

SEC("raw_tp")
__success
int iter_subprog_iters(const void *ctx)
{
 int sum, row, col;

 MY_PID_GUARD();

 bpf_for(row, 0, ARRAY_SIZE(arr2d)) {
  fill_inner_dimension(row);
 }

 /* zero-initialize sums */
 sum = 0;
 bpf_for(row, 0, ARRAY_SIZE(arr2d)) {
  arr2d_row_sums[row] = 0;
 }
 bpf_for(col, 0, ARRAY_SIZE(arr2d[0])) {
  arr2d_col_sums[col] = 0;
 }

 /* calculate sums */
 bpf_for(row, 0, ARRAY_SIZE(arr2d)) {
  sum += sum_inner_dimension(row);
 }

 bpf_printk("ITER_SUBPROG_ITERS: total sum=%d", sum);
 bpf_for(row, 0, ARRAY_SIZE(arr2d)) {
  bpf_printk("ITER_SUBPROG_ITERS: row #%d sum=%d",
      row, arr2d_row_sums[row]);
 }
 bpf_for(col, 0, ARRAY_SIZE(arr2d[0])) {
  bpf_printk("ITER_SUBPROG_ITERS: col #%d sum=%d%s",
      col, arr2d_col_sums[col],
      col == ARRAY_SIZE(arr2d[0]) - 1 ? "\n" : "");
 }

 return 0;
}

struct {
 __uint(type, BPF_MAP_TYPE_HASH);
 __type(key, int);
 __type(value, int);
 __uint(max_entries, 1000);
} hash_map SEC(".maps");

SEC("?raw_tp")
__failure __msg("invalid mem access 'scalar'")
int iter_err_too_permissive1(const void *ctx)
{
 int *map_val = NULL;
 int key = 0;

 MY_PID_GUARD();

 map_val = bpf_map_lookup_elem(&hash_map, &key);
 if (!map_val)
  return 0;

 bpf_repeat(1000000) {
  map_val = NULL;
 }

 *map_val = 123;

 return 0;
}

SEC("?raw_tp")
__failure __msg("invalid mem access 'map_value_or_null'")
int iter_err_too_permissive2(const void *ctx)
{
 int *map_val = NULL;
 int key = 0;

 MY_PID_GUARD();

 map_val = bpf_map_lookup_elem(&hash_map, &key);
 if (!map_val)
  return 0;

 bpf_repeat(1000000) {
  map_val = bpf_map_lookup_elem(&hash_map, &key);
 }

 *map_val = 123;

 return 0;
}

SEC("?raw_tp")
__failure __msg("invalid mem access 'map_value_or_null'")
int iter_err_too_permissive3(const void *ctx)
{
 int *map_val = NULL;
 int key = 0;
 bool found = false;

 MY_PID_GUARD();

 bpf_repeat(1000000) {
  map_val = bpf_map_lookup_elem(&hash_map, &key);
  found = true;
 }

 if (found)
  *map_val = 123;

 return 0;
}

SEC("raw_tp")
__success
int iter_tricky_but_fine(const void *ctx)
{
 int *map_val = NULL;
 int key = 0;
 bool found = false;

 MY_PID_GUARD();

 bpf_repeat(1000000) {
  map_val = bpf_map_lookup_elem(&hash_map, &key);
  if (map_val) {
   found = true;
   break;
  }
 }

 if (found)
  *map_val = 123;

 return 0;
}

#define __bpf_memzero(p, sz) bpf_probe_read_kernel((p), (sz), 0)

SEC("raw_tp")
__success
int iter_stack_array_loop(const void *ctx)
{
 long arr1[16], arr2[16], sum = 0;
 int i;

 MY_PID_GUARD();

 /* zero-init arr1 and arr2 in such a way that verifier doesn't know
 * it's all zeros; if we don't do that, we'll make BPF verifier track
 * all combination of zero/non-zero stack slots for arr1/arr2, which
 * will lead to O(2^(ARRAY_SIZE(arr1)+ARRAY_SIZE(arr2))) different
 * states
 */
 __bpf_memzero(arr1, sizeof(arr1));
 __bpf_memzero(arr2, sizeof(arr1));

 /* validate that we can break and continue when using bpf_for() */
 bpf_for(i, 0, ARRAY_SIZE(arr1)) {
  if (i & 1) {
   arr1[i] = i;
   continue;
  } else {
   arr2[i] = i;
   break;
  }
 }

 bpf_for(i, 0, ARRAY_SIZE(arr1)) {
  sum += arr1[i] + arr2[i];
 }

 return sum;
}

static __noinline void fill(struct bpf_iter_num *it, int *arr, __u32 n, int mul)
{
 int *t, i;

 while ((t = bpf_iter_num_next(it))) {
  i = *t;
  if (i >= n)
   break;
  arr[i] =  i * mul;
 }
}

static __noinline int sum(struct bpf_iter_num *it, int *arr, __u32 n)
{
 int *t, i, sum = 0;

 while ((t = bpf_iter_num_next(it))) {
  i = *t;
  if ((__u32)i >= n)
   break;
  sum += arr[i];
 }

 return sum;
}

SEC("raw_tp")
__success
int iter_pass_iter_ptr_to_subprog(const void *ctx)
{
 int arr1[16], arr2[32];
 struct bpf_iter_num it;
 int n, sum1, sum2;

 MY_PID_GUARD();

 /* fill arr1 */
 n = ARRAY_SIZE(arr1);
 bpf_iter_num_new(&it, 0, n);
 fill(&it, arr1, n, 2);
 bpf_iter_num_destroy(&it);

 /* fill arr2 */
 n = ARRAY_SIZE(arr2);
 bpf_iter_num_new(&it, 0, n);
 fill(&it, arr2, n, 10);
 bpf_iter_num_destroy(&it);

 /* sum arr1 */
 n = ARRAY_SIZE(arr1);
 bpf_iter_num_new(&it, 0, n);
 sum1 = sum(&it, arr1, n);
 bpf_iter_num_destroy(&it);

 /* sum arr2 */
 n = ARRAY_SIZE(arr2);
 bpf_iter_num_new(&it, 0, n);
 sum2 = sum(&it, arr2, n);
 bpf_iter_num_destroy(&it);

 bpf_printk("sum1=%d, sum2=%d", sum1, sum2);

 return 0;
}

SEC("?raw_tp")
__failure
__msg("R1 type=scalar expected=fp")
__naked int delayed_read_mark(void)
{
 /* This is equivalent to C program below.
 * The call to bpf_iter_num_next() is reachable with r7 values &fp[-16] and 0xdead.
 * State with r7=&fp[-16] is visited first and follows r6 != 42 ... continue branch.
 * At this point iterator next() call is reached with r7 that has no read mark.
 * Loop body with r7=0xdead would only be visited if verifier would decide to continue
 * with second loop iteration. Absence of read mark on r7 might affect state
 * equivalent logic used for iterator convergence tracking.
 *
 * r7 = &fp[-16]
 * fp[-16] = 0
 * r6 = bpf_get_prandom_u32()
 * bpf_iter_num_new(&fp[-8], 0, 10)
 * while (bpf_iter_num_next(&fp[-8])) {
 *   r6++
 *   if (r6 != 42) {
 *     r7 = 0xdead
 *     continue;
 *   }
 *   bpf_probe_read_user(r7, 8, 0xdeadbeef); // this is not safe
 * }
 * bpf_iter_num_destroy(&fp[-8])
 * return 0
 */
 asm volatile (
  "r7 = r10;"
  "r7 += -16;"
  "r0 = 0;"
  "*(u64 *)(r7 + 0) = r0;"
  "call %[bpf_get_prandom_u32];"
  "r6 = r0;"
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
 "1:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto 2f;"
  "r6 += 1;"
  "if r6 != 42 goto 3f;"
  "r7 = 0xdead;"
  "goto 1b;"
 "3:"
  "r1 = r7;"
  "r2 = 8;"
  "r3 = 0xdeadbeef;"
  "call %[bpf_probe_read_user];"
  "goto 1b;"
 "2:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
  :
  : __imm(bpf_get_prandom_u32),
    __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_next),
    __imm(bpf_iter_num_destroy),
    __imm(bpf_probe_read_user)
  : __clobber_all
 );
}

SEC("?raw_tp")
__failure
__msg("math between fp pointer and register with unbounded")
__naked int delayed_precision_mark(void)
{
 /* This is equivalent to C program below.
 * The test is similar to delayed_iter_mark but verifies that incomplete
 * precision don't fool verifier.
 * The call to bpf_iter_num_next() is reachable with r7 values -16 and -32.
 * State with r7=-16 is visited first and follows r6 != 42 ... continue branch.
 * At this point iterator next() call is reached with r7 that has no read
 * and precision marks.
 * Loop body with r7=-32 would only be visited if verifier would decide to continue
 * with second loop iteration. Absence of precision mark on r7 might affect state
 * equivalent logic used for iterator convergence tracking.
 *
 * r8 = 0
 * fp[-16] = 0
 * r7 = -16
 * r6 = bpf_get_prandom_u32()
 * bpf_iter_num_new(&fp[-8], 0, 10)
 * while (bpf_iter_num_next(&fp[-8])) {
 *   if (r6 != 42) {
 *     r7 = -32
 *     r6 = bpf_get_prandom_u32()
 *     continue;
 *   }
 *   r0 = r10
 *   r0 += r7
 *   r8 = *(u64 *)(r0 + 0)           // this is not safe
 *   r6 = bpf_get_prandom_u32()
 * }
 * bpf_iter_num_destroy(&fp[-8])
 * return r8
 */
 asm volatile (
  "r8 = 0;"
  "*(u64 *)(r10 - 16) = r8;"
  "r7 = -16;"
  "call %[bpf_get_prandom_u32];"
  "r6 = r0;"
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
 "1:"
  "r1 = r10;"
  "r1 += -8;\n"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto 2f;"
  "if r6 != 42 goto 3f;"
  "r7 = -33;"
  "call %[bpf_get_prandom_u32];"
  "r6 = r0;"
  "goto 1b;\n"
 "3:"
  "r0 = r10;"
  "r0 += r7;"
  "r8 = *(u64 *)(r0 + 0);"
  "call %[bpf_get_prandom_u32];"
  "r6 = r0;"
  "goto 1b;\n"
 "2:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r0 = r8;"
  "exit;"
  :
  : __imm(bpf_get_prandom_u32),
    __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_next),
    __imm(bpf_iter_num_destroy),
    __imm(bpf_probe_read_user)
  : __clobber_all
 );
}

SEC("?raw_tp")
__failure
__msg("math between fp pointer and register with unbounded")
__flag(BPF_F_TEST_STATE_FREQ)
__naked int loop_state_deps1(void)
{
 /* This is equivalent to C program below.
 *
 * The case turns out to be tricky in a sense that:
 * - states with c=-25 are explored only on a second iteration
 *   of the outer loop;
 * - states with read+precise mark on c are explored only on
 *   second iteration of the inner loop and in a state which
 *   is pushed to states stack first.
 *
 * Depending on the details of iterator convergence logic
 * verifier might stop states traversal too early and miss
 * unsafe c=-25 memory access.
 *
 *   j = iter_new();  // fp[-16]
 *   a = 0;  // r6
 *   b = 0;  // r7
 *   c = -24;  // r8
 *   while (iter_next(j)) {
 *     i = iter_new();  // fp[-8]
 *     a = 0;  // r6
 *     b = 0;  // r7
 *     while (iter_next(i)) {
 *  if (a == 1) {
 *    a = 0;
 *    b = 1;
 *  } else if (a == 0) {
 *    a = 1;
 *    if (random() == 42)
 *      continue;
 *    if (b == 1) {
 *      *(r10 + c) = 7;  // this is not safe
 *      iter_destroy(i);
 *      iter_destroy(j);
 *      return;
 *    }
 *  }
 *     }
 *     iter_destroy(i);
 *     a = 0;
 *     b = 0;
 *     c = -25;
 *   }
 *   iter_destroy(j);
 *   return;
 */
 asm volatile (
  "r1 = r10;"
  "r1 += -16;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
  "r6 = 0;"
  "r7 = 0;"
  "r8 = -24;"
 "j_loop_%=:"
  "r1 = r10;"
  "r1 += -16;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto j_loop_end_%=;"
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
  "r6 = 0;"
  "r7 = 0;"
 "i_loop_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto i_loop_end_%=;"
 "check_one_r6_%=:"
  "if r6 != 1 goto check_zero_r6_%=;"
  "r6 = 0;"
  "r7 = 1;"
  "goto i_loop_%=;"
 "check_zero_r6_%=:"
  "if r6 != 0 goto i_loop_%=;"
  "r6 = 1;"
  "call %[bpf_get_prandom_u32];"
  "if r0 != 42 goto check_one_r7_%=;"
  "goto i_loop_%=;"
 "check_one_r7_%=:"
  "if r7 != 1 goto i_loop_%=;"
  "r0 = r10;"
  "r0 += r8;"
  "r1 = 7;"
  "*(u64 *)(r0 + 0) = r1;"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r1 = r10;"
  "r1 += -16;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
 "i_loop_end_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r6 = 0;"
  "r7 = 0;"
  "r8 = -25;"
  "goto j_loop_%=;"
 "j_loop_end_%=:"
  "r1 = r10;"
  "r1 += -16;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
  :
  : __imm(bpf_get_prandom_u32),
    __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_next),
    __imm(bpf_iter_num_destroy)
  : __clobber_all
 );
}

SEC("?raw_tp")
__failure
__msg("math between fp pointer and register with unbounded")
__flag(BPF_F_TEST_STATE_FREQ)
__naked int loop_state_deps2(void)
{
 /* This is equivalent to C program below.
 *
 * The case turns out to be tricky in a sense that:
 * - states with read+precise mark on c are explored only on a second
 *   iteration of the first inner loop and in a state which is pushed to
 *   states stack first.
 * - states with c=-25 are explored only on a second iteration of the
 *   second inner loop and in a state which is pushed to states stack
 *   first.
 *
 * Depending on the details of iterator convergence logic
 * verifier might stop states traversal too early and miss
 * unsafe c=-25 memory access.
 *
 *   j = iter_new();             // fp[-16]
 *   a = 0;                      // r6
 *   b = 0;                      // r7
 *   c = -24;                    // r8
 *   while (iter_next(j)) {
 *     i = iter_new();           // fp[-8]
 *     a = 0;                    // r6
 *     b = 0;                    // r7
 *     while (iter_next(i)) {
 *       if (a == 1) {
 *         a = 0;
 *         b = 1;
 *       } else if (a == 0) {
 *         a = 1;
 *         if (random() == 42)
 *           continue;
 *         if (b == 1) {
 *           *(r10 + c) = 7;     // this is not safe
 *           iter_destroy(i);
 *           iter_destroy(j);
 *           return;
 *         }
 *       }
 *     }
 *     iter_destroy(i);
 *     i = iter_new();           // fp[-8]
 *     a = 0;                    // r6
 *     b = 0;                    // r7
 *     while (iter_next(i)) {
 *       if (a == 1) {
 *         a = 0;
 *         b = 1;
 *       } else if (a == 0) {
 *         a = 1;
 *         if (random() == 42)
 *           continue;
 *         if (b == 1) {
 *           a = 0;
 *           c = -25;
 *         }
 *       }
 *     }
 *     iter_destroy(i);
 *   }
 *   iter_destroy(j);
 *   return;
 */
 asm volatile (
  "r1 = r10;"
  "r1 += -16;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
  "r6 = 0;"
  "r7 = 0;"
  "r8 = -24;"
 "j_loop_%=:"
  "r1 = r10;"
  "r1 += -16;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto j_loop_end_%=;"

  /* first inner loop */
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
  "r6 = 0;"
  "r7 = 0;"
 "i_loop_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto i_loop_end_%=;"
 "check_one_r6_%=:"
  "if r6 != 1 goto check_zero_r6_%=;"
  "r6 = 0;"
  "r7 = 1;"
  "goto i_loop_%=;"
 "check_zero_r6_%=:"
  "if r6 != 0 goto i_loop_%=;"
  "r6 = 1;"
  "call %[bpf_get_prandom_u32];"
  "if r0 != 42 goto check_one_r7_%=;"
  "goto i_loop_%=;"
 "check_one_r7_%=:"
  "if r7 != 1 goto i_loop_%=;"
  "r0 = r10;"
  "r0 += r8;"
  "r1 = 7;"
  "*(u64 *)(r0 + 0) = r1;"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r1 = r10;"
  "r1 += -16;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
 "i_loop_end_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"

  /* second inner loop */
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
  "r6 = 0;"
  "r7 = 0;"
 "i2_loop_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto i2_loop_end_%=;"
 "check2_one_r6_%=:"
  "if r6 != 1 goto check2_zero_r6_%=;"
  "r6 = 0;"
  "r7 = 1;"
  "goto i2_loop_%=;"
 "check2_zero_r6_%=:"
  "if r6 != 0 goto i2_loop_%=;"
  "r6 = 1;"
  "call %[bpf_get_prandom_u32];"
  "if r0 != 42 goto check2_one_r7_%=;"
  "goto i2_loop_%=;"
 "check2_one_r7_%=:"
  "if r7 != 1 goto i2_loop_%=;"
  "r6 = 0;"
  "r8 = -25;"
  "goto i2_loop_%=;"
 "i2_loop_end_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"

  "r6 = 0;"
  "r7 = 0;"
  "goto j_loop_%=;"
 "j_loop_end_%=:"
  "r1 = r10;"
  "r1 += -16;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
  :
  : __imm(bpf_get_prandom_u32),
    __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_next),
    __imm(bpf_iter_num_destroy)
  : __clobber_all
 );
}

SEC("?raw_tp")
__failure
__msg("math between fp pointer and register with unbounded")
__flag(BPF_F_TEST_STATE_FREQ)
__naked int loop_state_deps3(void)
{
 /* This is equivalent to a C program below.
 *
 *   if (random() != 24) {       // assume false branch is placed first
 *     i = iter_new();           // fp[-8]
 *     while (iter_next(i));
 *     iter_destroy(i);
 *     return;
 *   }
 *
 *   for (i = 10; i > 0; i--);   // increase dfs_depth for child states
 *
 *   i = iter_new();             // fp[-8]
 *   b = -24;                    // r8
 *   for (;;) {                  // checkpoint (L)
 *     if (iter_next(i))         // checkpoint (N)
 *       break;
 *     if (random() == 77) {     // assume false branch is placed first
 *       *(u64 *)(r10 + b) = 7;  // this is not safe when b == -25
 *       iter_destroy(i);
 *       return;
 *     }
 *     if (random() == 42) {     // assume false branch is placed first
 *       b = -25;
 *     }
 *   }
 *   iter_destroy(i);
 *
 * In case of a buggy verifier first loop might poison
 * env->cur_state->loop_entry with a state having 0 branches
 * and small dfs_depth. This would trigger NOT_EXACT states
 * comparison for some states within second loop.
 * Specifically, checkpoint (L) might be problematic if:
 * - branch with '*(u64 *)(r10 + b) = 7' is not explored yet;
 * - checkpoint (L) is first reached in state {b=-24};
 * - traversal is pruned at checkpoint (N) setting checkpoint's (L)
 *   branch count to 0, thus making it eligible for use in pruning;
 * - checkpoint (L) is next reached in state {b=-25},
 *   this would cause NOT_EXACT comparison with a state {b=-24}
 *   while 'b' is not marked precise yet.
 */
 asm volatile (
  "call %[bpf_get_prandom_u32];"
  "if r0 == 24 goto 2f;"
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 5;"
  "call %[bpf_iter_num_new];"
 "1:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_next];"
  "if r0 != 0 goto 1b;"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
 "2:"
  /* loop to increase dfs_depth */
  "r0 = 10;"
 "3:"
  "r0 -= 1;"
  "if r0 != 0 goto 3b;"
  /* end of loop */
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
  "r8 = -24;"
 "main_loop_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto main_loop_end_%=;"
  /* first if */
  "call %[bpf_get_prandom_u32];"
  "if r0 == 77 goto unsafe_write_%=;"
  /* second if */
  "call %[bpf_get_prandom_u32];"
  "if r0 == 42 goto poison_r8_%=;"
  /* iterate */
  "goto main_loop_%=;"
 "main_loop_end_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"

 "unsafe_write_%=:"
  "r0 = r10;"
  "r0 += r8;"
  "r1 = 7;"
  "*(u64 *)(r0 + 0) = r1;"
  "goto main_loop_end_%=;"

 "poison_r8_%=:"
  "r8 = -25;"
  "goto main_loop_%=;"
  :
  : __imm(bpf_get_prandom_u32),
    __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_next),
    __imm(bpf_iter_num_destroy)
  : __clobber_all
 );
}

SEC("?raw_tp")
__success
__naked int triple_continue(void)
{
 /* This is equivalent to C program below.
 * High branching factor of the loop body turned out to be
 * problematic for one of the iterator convergence tracking
 * algorithms explored.
 *
 * r6 = bpf_get_prandom_u32()
 * bpf_iter_num_new(&fp[-8], 0, 10)
 * while (bpf_iter_num_next(&fp[-8])) {
 *   if (bpf_get_prandom_u32() != 42)
 *     continue;
 *   if (bpf_get_prandom_u32() != 42)
 *     continue;
 *   if (bpf_get_prandom_u32() != 42)
 *     continue;
 *   r0 += 0;
 * }
 * bpf_iter_num_destroy(&fp[-8])
 * return 0
 */
 asm volatile (
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
 "loop_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto loop_end_%=;"
  "call %[bpf_get_prandom_u32];"
  "if r0 != 42 goto loop_%=;"
  "call %[bpf_get_prandom_u32];"
  "if r0 != 42 goto loop_%=;"
  "call %[bpf_get_prandom_u32];"
  "if r0 != 42 goto loop_%=;"
  "r0 += 0;"
  "goto loop_%=;"
 "loop_end_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
  :
  : __imm(bpf_get_prandom_u32),
    __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_next),
    __imm(bpf_iter_num_destroy)
  : __clobber_all
 );
}

SEC("?raw_tp")
__success
__naked int widen_spill(void)
{
 /* This is equivalent to C program below.
 * The counter is stored in fp[-16], if this counter is not widened
 * verifier states representing loop iterations would never converge.
 *
 * fp[-16] = 0
 * bpf_iter_num_new(&fp[-8], 0, 10)
 * while (bpf_iter_num_next(&fp[-8])) {
 *   r0 = fp[-16];
 *   r0 += 1;
 *   fp[-16] = r0;
 * }
 * bpf_iter_num_destroy(&fp[-8])
 * return 0
 */
 asm volatile (
  "r0 = 0;"
  "*(u64 *)(r10 - 16) = r0;"
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
 "loop_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto loop_end_%=;"
  "r0 = *(u64 *)(r10 - 16);"
  "r0 += 1;"
  "*(u64 *)(r10 - 16) = r0;"
  "goto loop_%=;"
 "loop_end_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
  :
  : __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_next),
    __imm(bpf_iter_num_destroy)
  : __clobber_all
 );
}

SEC("raw_tp")
__success
__naked int checkpoint_states_deletion(void)
{
 /* This is equivalent to C program below.
 *
 *   int *a, *b, *c, *d, *e, *f;
 *   int i, sum = 0;
 *   bpf_for(i, 0, 10) {
 *     a = bpf_map_lookup_elem(&amap, &i);
 *     b = bpf_map_lookup_elem(&amap, &i);
 *     c = bpf_map_lookup_elem(&amap, &i);
 *     d = bpf_map_lookup_elem(&amap, &i);
 *     e = bpf_map_lookup_elem(&amap, &i);
 *     f = bpf_map_lookup_elem(&amap, &i);
 *     if (a) sum += 1;
 *     if (b) sum += 1;
 *     if (c) sum += 1;
 *     if (d) sum += 1;
 *     if (e) sum += 1;
 *     if (f) sum += 1;
 *   }
 *   return 0;
 *
 * The body of the loop spawns multiple simulation paths
 * with different combination of NULL/non-NULL information for a/b/c/d/e/f.
 * Each combination is unique from states_equal() point of view.
 * Explored states checkpoint is created after each iterator next call.
 * Iterator convergence logic expects that eventually current state
 * would get equal to one of the explored states and thus loop
 * exploration would be finished (at-least for a specific path).
 * Verifier evicts explored states with high miss to hit ratio
 * to to avoid comparing current state with too many explored
 * states per instruction.
 * This test is designed to "stress test" eviction policy defined using formula:
 *
 *    sl->miss_cnt > sl->hit_cnt * N + N // if true sl->state is evicted
 *
 * Currently N is set to 64, which allows for 6 variables in this test.
 */
 asm volatile (
  "r6 = 0;"                  /* a */
  "r7 = 0;"                  /* b */
  "r8 = 0;"                  /* c */
  "*(u64 *)(r10 - 24) = r6;" /* d */
  "*(u64 *)(r10 - 32) = r6;" /* e */
  "*(u64 *)(r10 - 40) = r6;" /* f */
  "r9 = 0;"                  /* sum */
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
 "loop_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto loop_end_%=;"

  "*(u64 *)(r10 - 16) = r0;"

  "r1 = %[amap] ll;"
  "r2 = r10;"
  "r2 += -16;"
  "call %[bpf_map_lookup_elem];"
  "r6 = r0;"

  "r1 = %[amap] ll;"
  "r2 = r10;"
  "r2 += -16;"
  "call %[bpf_map_lookup_elem];"
  "r7 = r0;"

  "r1 = %[amap] ll;"
  "r2 = r10;"
  "r2 += -16;"
  "call %[bpf_map_lookup_elem];"
  "r8 = r0;"

  "r1 = %[amap] ll;"
  "r2 = r10;"
  "r2 += -16;"
  "call %[bpf_map_lookup_elem];"
  "*(u64 *)(r10 - 24) = r0;"

  "r1 = %[amap] ll;"
  "r2 = r10;"
  "r2 += -16;"
  "call %[bpf_map_lookup_elem];"
  "*(u64 *)(r10 - 32) = r0;"

  "r1 = %[amap] ll;"
  "r2 = r10;"
  "r2 += -16;"
  "call %[bpf_map_lookup_elem];"
  "*(u64 *)(r10 - 40) = r0;"

  "if r6 == 0 goto +1;"
  "r9 += 1;"
  "if r7 == 0 goto +1;"
  "r9 += 1;"
  "if r8 == 0 goto +1;"
  "r9 += 1;"
  "r0 = *(u64 *)(r10 - 24);"
  "if r0 == 0 goto +1;"
  "r9 += 1;"
  "r0 = *(u64 *)(r10 - 32);"
  "if r0 == 0 goto +1;"
  "r9 += 1;"
  "r0 = *(u64 *)(r10 - 40);"
  "if r0 == 0 goto +1;"
  "r9 += 1;"

  "goto loop_%=;"
 "loop_end_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
  :
  : __imm(bpf_map_lookup_elem),
    __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_next),
    __imm(bpf_iter_num_destroy),
    __imm_addr(amap)
  : __clobber_all
 );
}

struct {
 int data[32];
 int n;
} loop_data;

SEC("raw_tp")
__success
int iter_arr_with_actual_elem_count(const void *ctx)
{
 int i, n = loop_data.n, sum = 0;

 if (n > ARRAY_SIZE(loop_data.data))
  return 0;

 bpf_for(i, 0, n) {
  /* no rechecking of i against ARRAY_SIZE(loop_data.n) */
  sum += loop_data.data[i];
 }

 return sum;
}

__u32 upper, select_n, result;
__u64 global;

static __noinline bool nest_2(char *str)
{
 /* some insns (including branch insns) to ensure stacksafe() is triggered
 * in nest_2(). This way, stacksafe() can compare frame associated with nest_1().
 */
 if (str[0] == 't')
  return true;
 if (str[1] == 'e')
  return true;
 if (str[2] == 's')
  return true;
 if (str[3] == 't')
  return true;
 return false;
}

static __noinline bool nest_1(int n)
{
 /* case 0: allocate stack, case 1: no allocate stack */
 switch (n) {
 case 0: {
  char comm[16];

  if (bpf_get_current_comm(comm, 16))
   return false;
  return nest_2(comm);
 }
 case 1:
  return nest_2((char *)&global);
 default:
  return false;
 }
}

SEC("raw_tp")
__success
int iter_subprog_check_stacksafe(const void *ctx)
{
 long i;

 bpf_for(i, 0, upper) {
  if (!nest_1(select_n)) {
   result = 1;
   return 0;
  }
 }

 result = 2;
 return 0;
}

struct bpf_iter_num global_it;

SEC("raw_tp")
__failure __msg("arg#0 expected pointer to an iterator on stack")
int iter_new_bad_arg(const void *ctx)
{
 bpf_iter_num_new(&global_it, 0, 1);
 return 0;
}

SEC("raw_tp")
__failure __msg("arg#0 expected pointer to an iterator on stack")
int iter_next_bad_arg(const void *ctx)
{
 bpf_iter_num_next(&global_it);
 return 0;
}

SEC("raw_tp")
__failure __msg("arg#0 expected pointer to an iterator on stack")
int iter_destroy_bad_arg(const void *ctx)
{
 bpf_iter_num_destroy(&global_it);
 return 0;
}

SEC("raw_tp")
__success
int clean_live_states(const void *ctx)
{
 char buf[1];
 int i, j, k, l, m, n, o;

 bpf_for(i, 0, 10)
 bpf_for(j, 0, 10)
 bpf_for(k, 0, 10)
 bpf_for(l, 0, 10)
 bpf_for(m, 0, 10)
 bpf_for(n, 0, 10)
 bpf_for(o, 0, 10) {
  if (unlikely(bpf_get_prandom_u32()))
   buf[0] = 42;
  bpf_printk("%s", buf);
 }
 return 0;
}

SEC("?raw_tp")
__flag(BPF_F_TEST_STATE_FREQ)
__failure __msg("misaligned stack access off 0+-31+0 size 8")
__naked int absent_mark_in_the_middle_state(void)
{
 /* This is equivalent to C program below.
 *
 * r8 = bpf_get_prandom_u32();
 * r6 = -32;
 * bpf_iter_num_new(&fp[-8], 0, 10);
 * if (unlikely(bpf_get_prandom_u32()))
 *   r6 = -31;
 * while (bpf_iter_num_next(&fp[-8])) {
 *   if (unlikely(bpf_get_prandom_u32()))
 *     *(fp + r6) = 7;
 * }
 * bpf_iter_num_destroy(&fp[-8])
 * return 0
 */
 asm volatile (
  "call %[bpf_get_prandom_u32];"
  "r8 = r0;"
  "r7 = 0;"
  "r6 = -32;"
  "r0 = 0;"
  "*(u64 *)(r10 - 16) = r0;"
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
  "call %[bpf_get_prandom_u32];"
  "if r0 == r8 goto change_r6_%=;"
 "loop_%=:"
  "call noop;"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto loop_end_%=;"
  "call %[bpf_get_prandom_u32];"
  "if r0 == r8 goto use_r6_%=;"
  "goto loop_%=;"
 "loop_end_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
 "use_r6_%=:"
  "r0 = r10;"
  "r0 += r6;"
  "r1 = 7;"
  "*(u64 *)(r0 + 0) = r1;"
  "goto loop_%=;"
 "change_r6_%=:"
  "r6 = -31;"
  "goto loop_%=;"
  :
  : __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_next),
    __imm(bpf_iter_num_destroy),
    __imm(bpf_get_prandom_u32)
  : __clobber_all
 );
}

__used __naked
static int noop(void)
{
 asm volatile (
  "r0 = 0;"
  "exit;"
 );
}

SEC("?raw_tp")
__flag(BPF_F_TEST_STATE_FREQ)
__failure __msg("misaligned stack access off 0+-31+0 size 8")
__naked int absent_mark_in_the_middle_state2(void)
{
 /* This is equivalent to C program below.
 *
 *     r8 = bpf_get_prandom_u32();
 *     r6 = -32;
 *     bpf_iter_num_new(&fp[-8], 0, 10);
 *     if (unlikely(bpf_get_prandom_u32())) {
 *       r6 = -31;
 * jump_into_loop:
 *       goto +0;
 *       goto loop;
 *     }
 *     if (unlikely(bpf_get_prandom_u32()))
 *       goto jump_into_loop;
 * loop:
 *     while (bpf_iter_num_next(&fp[-8])) {
 *       if (unlikely(bpf_get_prandom_u32()))
 *         *(fp + r6) = 7;
 *     }
 *     bpf_iter_num_destroy(&fp[-8])
 *     return 0
 */
 asm volatile (
  "call %[bpf_get_prandom_u32];"
  "r8 = r0;"
  "r7 = 0;"
  "r6 = -32;"
  "r0 = 0;"
  "*(u64 *)(r10 - 16) = r0;"
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
  "call %[bpf_get_prandom_u32];"
  "if r0 == r8 goto change_r6_%=;"
  "call %[bpf_get_prandom_u32];"
  "if r0 == r8 goto jump_into_loop_%=;"
 "loop_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto loop_end_%=;"
  "call %[bpf_get_prandom_u32];"
  "if r0 == r8 goto use_r6_%=;"
  "goto loop_%=;"
 "loop_end_%=:"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
 "use_r6_%=:"
  "r0 = r10;"
  "r0 += r6;"
  "r1 = 7;"
  "*(u64 *)(r0 + 0) = r1;"
  "goto loop_%=;"
 "change_r6_%=:"
  "r6 = -31;"
 "jump_into_loop_%=: "
  "goto +0;"
  "goto loop_%=;"
  :
  : __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_next),
    __imm(bpf_iter_num_destroy),
    __imm(bpf_get_prandom_u32)
  : __clobber_all
 );
}

SEC("?raw_tp")
__flag(BPF_F_TEST_STATE_FREQ)
__failure __msg("misaligned stack access off 0+-31+0 size 8")
__naked int absent_mark_in_the_middle_state3(void)
{
 /*
 * bpf_iter_num_new(&fp[-8], 0, 10)
 * loop1(-32, &fp[-8])
 * loop1_wrapper(&fp[-8])
 * bpf_iter_num_destroy(&fp[-8])
 */
 asm volatile (
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
  /* call #1 */
  "r1 = -32;"
  "r2 = r10;"
  "r2 += -8;"
  "call loop1;"
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  /* call #2 */
  "r1 = r10;"
  "r1 += -8;"
  "r2 = 0;"
  "r3 = 10;"
  "call %[bpf_iter_num_new];"
  "r1 = r10;"
  "r1 += -8;"
  "call loop1_wrapper;"
  /* return */
  "r1 = r10;"
  "r1 += -8;"
  "call %[bpf_iter_num_destroy];"
  "r0 = 0;"
  "exit;"
  :
  : __imm(bpf_iter_num_new),
    __imm(bpf_iter_num_destroy),
    __imm(bpf_get_prandom_u32)
  : __clobber_all
 );
}

__used __naked
static int loop1(void)
{
 /*
 *  int loop1(num, iter) {
 *     r6 = num;
 *     r7 = iter;
 *     while (bpf_iter_num_next(r7)) {
 *       if (unlikely(bpf_get_prandom_u32()))
 *         *(fp + r6) = 7;
 *     }
 *     return 0
 *  }
 */
 asm volatile (
  "r6 = r1;"
  "r7 = r2;"
  "call %[bpf_get_prandom_u32];"
  "r8 = r0;"
 "loop_%=:"
  "r1 = r7;"
  "call %[bpf_iter_num_next];"
  "if r0 == 0 goto loop_end_%=;"
  "call %[bpf_get_prandom_u32];"
  "if r0 == r8 goto use_r6_%=;"
  "goto loop_%=;"
 "loop_end_%=:"
  "r0 = 0;"
  "exit;"
 "use_r6_%=:"
  "r0 = r10;"
  "r0 += r6;"
  "r1 = 7;"
  "*(u64 *)(r0 + 0) = r1;"
  "goto loop_%=;"
  :
  : __imm(bpf_iter_num_next),
    __imm(bpf_get_prandom_u32)
  : __clobber_all
 );
}

__used __naked
static int loop1_wrapper(void)
{
 /*
 *  int loop1_wrapper(iter) {
 *    r6 = -32;
 *    r7 = iter;
 *    if (unlikely(bpf_get_prandom_u32()))
 *      r6 = -31;
 *    loop1(r6, r7);
 *    return 0;
 *  }
 */
 asm volatile (
  "r6 = -32;"
  "r7 = r1;"
  "call %[bpf_get_prandom_u32];"
  "r8 = r0;"
  "call %[bpf_get_prandom_u32];"
  "if r0 == r8 goto change_r6_%=;"
 "loop_%=:"
  "r1 = r6;"
  "r2 = r7;"
  "call loop1;"
  "r0 = 0;"
  "exit;"
 "change_r6_%=:"
  "r6 = -31;"
  "goto loop_%=;"
  :
  : __imm(bpf_iter_num_next),
    __imm(bpf_get_prandom_u32)
  : __clobber_all
 );
}

char _license[] SEC("license") = "GPL";