代码

// duck.ac 1001 探针：验证裸 clone() 多线程在评测机上是否真并行 + mmap 400MB 是否可用。
// 判定：故意不排序（预期 WA），看报告的用时：
//   ~600ms  => 4 核并行 OK（3 个 worker + main 各自 spin 0.5s @ 3.6GHz）
//   ~2.1s   => clone 可用但被串行化
//   RE      => clone 崩溃（TLS/栈保护问题）
// 附加：mmap 400MB + 4KB 步长预触页 + 读校验，验证内存路径（约 +100~200ms）。
#include <cstdint>
#include <sched.h>
#include <sys/mman.h>
#include <immintrin.h>

extern "C" int clone(int (*)(void*), void*, int, void*, ...);

// 最小 TCB：glibc 栈保护读 fs:0x28，pointer guard 在 fs:0x30
struct TCB {
  void* tcb;
  void* dtv;
  void* self;
  long multiple_threads;
  long gscope_flag;
  void* sysinfo;
  void* pad1;
  unsigned long guard;  // offset 0x28
  unsigned long ptr_guard;
};
static char stacks[3][1 << 18] __attribute__((aligned(4096)));  // 256KB/线程
static TCB tcbs[3] __attribute__((aligned(64)));
struct Arg {
  volatile unsigned long long done;
  unsigned long long spin_cycles;
} __attribute__((aligned(64)));
static Arg args[3];

static int worker(void* p) {
  Arg* a = (Arg*)p;
  unsigned long long end = __rdtsc() + a->spin_cycles;
  while (__rdtsc() < end) {
  }
  __atomic_store_n(&a->done, 1, __ATOMIC_RELEASE);
  return 0;
}

void sort(unsigned* a, int n) {
  (void)a;
  (void)n;
  const unsigned long long SPIN = 1800000000ULL;  // ~0.5s @ 3.6GHz
  for (int t = 0; t < 3; ++t) {
    args[t].spin_cycles = SPIN;
    tcbs[t].tcb = &tcbs[t];
    tcbs[t].self = &tcbs[t];
    tcbs[t].multiple_threads = 1;
    tcbs[t].guard = 0x1234567890ABCDEFUL;
    tcbs[t].ptr_guard = 0xFEDCBA0987654321UL;
  }
  for (int t = 0; t < 3; ++t) {
    void* top = stacks[t] + sizeof(stacks[t]);
    clone(worker, top,
          CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND | CLONE_THREAD |
              CLONE_SYSVSEM | CLONE_SETTLS,
          &args[t], nullptr, &tcbs[t], &tcbs[t]);
  }
  unsigned long long end = __rdtsc() + SPIN;
  while (__rdtsc() < end) {
  }
  for (int t = 0; t < 3; ++t)
    while (!__atomic_load_n(&args[t].done, __ATOMIC_ACQUIRE)) {
    }
  // 内存路径探测：mmap 400MB + 预触 + 读
  volatile unsigned* v = (volatile unsigned*)mmap(
      nullptr, 400000000UL, PROT_READ | PROT_WRITE,
      MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
  if (v != MAP_FAILED) {
    for (size_t i = 0; i < 400000000UL; i += 4096) *(volatile char*)((char*)v + i) = 0;
    unsigned s = 0;
    for (size_t i = 0; i < 100000000UL; i += 1024) s += v[i];
    if (s == 0xdeadbeef) *const_cast<unsigned*>(v) = s;  // 防优化
  }
}

评测结果