/*
 * vdsotest.c -- stage 19: the vDSO, checked from a real C library (musl).
 *
 * Built with `musl-gcc -static` as a classic ET_EXEC program (linked at
 * Rustbelt's user base, see build.rs). musl finds the vDSO through
 * AT_SYSINFO_EHDR and then calls __vdso_clock_gettime instead of making a
 * system call. We compare what the vDSO says with what the real system
 * call says, check the clocks never go backwards, and that it all still
 * works in a forked child and after execve.
 *
 *   vdsotest            all checks (exit code = number of failed checks)
 *   vdsotest calls      just call the clock functions a lot, print nothing
 *                       (tests/vdso.rs looks at the --trace output)
 *   vdsotest exec       (used internally: the program after execve)
 */
#define _GNU_SOURCE
#include <elf.h>
#include <errno.h>
#include <sched.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/auxv.h>
#include <sys/mman.h>
#include <sys/syscall.h>
#include <sys/time.h>
#include <sys/wait.h>
#include <time.h>
#include <unistd.h>

static int fails;
#define CHECK(cond, ...)                                              \
	do {                                                          \
		if (!(cond)) {                                        \
			fails++;                                      \
			printf("vdsotest: FAIL line %d: ", __LINE__); \
			printf(__VA_ARGS__);                          \
			printf("\n");                                 \
		}                                                     \
	} while (0)

extern char _start[]; /* musl's entry point (crt1.o) */

static long long ns(struct timespec t) { return t.tv_sec * 1000000000LL + t.tv_nsec; }
static long long us(struct timeval t) { return t.tv_sec * 1000000LL + t.tv_usec; }

/* The real system call, bypassing the C library (and so the vDSO). */
static long long sys_ns(int clock)
{
	struct timespec t;
	if (syscall(SYS_clock_gettime, clock, &t) != 0)
		return -1;
	return ns(t);
}

static long long vdso_ns(int clock)
{
	struct timespec t;
	if (clock_gettime(clock, &t) != 0)
		return -1;
	return ns(t);
}

/* The image behind AT_SYSINFO_EHDR is an ELF shared object, and the page
 * in front of it (the vvar page) is readable but can't be made writable. */
static void image(void)
{
	unsigned long base = getauxval(AT_SYSINFO_EHDR);
	CHECK(base != 0, "no AT_SYSINFO_EHDR");
	if (!base)
		return;
	const Elf64_Ehdr *eh = (const Elf64_Ehdr *)base;
	CHECK(memcmp(eh->e_ident, ELFMAG, SELFMAG) == 0, "vDSO has no ELF magic");
	CHECK(eh->e_type == ET_DYN, "vDSO e_type %d", eh->e_type);
	void *vvar = (void *)(base - 4096);
	unsigned seq = *(volatile unsigned *)vvar;
	CHECK(seq % 2 == 0, "vvar seq %u is odd", seq);
	errno = 0;
	CHECK(mprotect(vvar, 4096, PROT_READ | PROT_WRITE) == -1 && errno == EACCES, "mprotect(vvar, RW): errno %d", errno);
	/* Other auxv entries a C library looks at. */
	CHECK(getauxval(AT_PAGESZ) == 4096, "AT_PAGESZ");
	CHECK(getauxval(AT_CLKTCK) == 100, "AT_CLKTCK");
	CHECK(getauxval(AT_HWCAP) & (1 << 4), "AT_HWCAP has no TSC bit");
	CHECK(getauxval(AT_RANDOM) != 0, "AT_RANDOM");
	CHECK(getauxval(AT_SECURE) == 0, "AT_SECURE");
	CHECK(getauxval(AT_ENTRY) == (unsigned long)_start, "AT_ENTRY %lx", getauxval(AT_ENTRY));
	CHECK(getauxval(AT_PHNUM) > 0 && getauxval(AT_PHENT) == sizeof(Elf64_Phdr), "AT_PHNUM/AT_PHENT");
	CHECK(getauxval(AT_BASE) == 0, "AT_BASE (no dynamic loader)");
	const char *platform = (const char *)getauxval(AT_PLATFORM);
	CHECK(platform && strcmp(platform, "x86_64") == 0, "AT_PLATFORM");
	const char *execfn = (const char *)getauxval(AT_EXECFN);
	CHECK(execfn && strstr(execfn, "vdsotest"), "AT_EXECFN");
}

/* Each fast clock: vDSO, then syscall, then vDSO again must be in order
 * (they compute the same formula from the same numbers); and the clocks
 * never go backwards. */
static void clocks(void)
{
	static const int fast[] = { CLOCK_REALTIME, CLOCK_MONOTONIC, CLOCK_MONOTONIC_RAW, CLOCK_REALTIME_COARSE,
				    CLOCK_MONOTONIC_COARSE };
	for (unsigned i = 0; i < sizeof fast / sizeof fast[0]; i++) {
		int c = fast[i];
		for (int round = 0; round < 100; round++) {
			long long a = vdso_ns(c), b = sys_ns(c), d = vdso_ns(c);
			if (!(a > 0 && a <= b && b <= d)) {
				CHECK(0, "clock %d out of order: vdso %lld, syscall %lld, vdso %lld", c, a, b, d);
				break;
			}
		}
	}
	long long prev = vdso_ns(CLOCK_MONOTONIC);
	for (int i = 0; i < 100000; i++) {
		long long now = vdso_ns(CLOCK_MONOTONIC);
		if (now < prev) {
			CHECK(0, "CLOCK_MONOTONIC went back from %lld to %lld", prev, now);
			break;
		}
		prev = now;
	}
	/* Nanoseconds really: two readings in a row usually differ by less
	 * than a microsecond, but they must not be stuck on whole µs. */
	int sub_us = 0;
	for (int i = 0; i < 100 && !sub_us; i++)
		sub_us = vdso_ns(CLOCK_MONOTONIC) % 1000 != 0;
	CHECK(sub_us, "CLOCK_MONOTONIC only has microseconds");
	/* The wall clock is about now (after 2020), and REALTIME = MONOTONIC + boot. */
	CHECK(vdso_ns(CLOCK_REALTIME) / 1000000000 > 1577836800LL, "CLOCK_REALTIME is before 2020");
	/* Other clocks take the syscall path inside the vDSO, and still work. */
	struct timespec t;
	CHECK(clock_gettime(CLOCK_BOOTTIME, &t) == 0 && ns(t) > 0, "CLOCK_BOOTTIME");
	CHECK(clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &t) == 0, "CLOCK_PROCESS_CPUTIME_ID");
	CHECK(clock_gettime(CLOCK_THREAD_CPUTIME_ID, &t) == 0, "CLOCK_THREAD_CPUTIME_ID");
	errno = 0;
	CHECK(clock_gettime(12345, &t) == -1 && errno == EINVAL, "bad clock: errno %d", errno);

	/* gettimeofday and time: also from the vDSO, also agreeing. */
	struct timeval tv1, tv2, tvs;
	struct timezone tz = { 99, 99 };
	gettimeofday(&tv1, NULL);
	CHECK(syscall(SYS_gettimeofday, &tvs, &tz) == 0, "gettimeofday syscall");
	gettimeofday(&tv2, NULL);
	CHECK(us(tv1) <= us(tvs) && us(tvs) <= us(tv2), "gettimeofday out of order: %lld %lld %lld", us(tv1), us(tvs),
	      us(tv2));
	CHECK(tz.tz_minuteswest == 0 && tz.tz_dsttime == 0, "timezone not UTC");
	CHECK(tv1.tv_usec >= 0 && tv1.tv_usec < 1000000, "tv_usec %ld", (long)tv1.tv_usec);
	time_t t1 = time(NULL), ts = syscall(SYS_time, NULL), t2;
	time(&t2);
	CHECK(t1 <= ts && ts <= t2 && t2 - t1 < 2, "time out of order: %ld %ld %ld", (long)t1, (long)ts, (long)t2);
	CHECK(t1 == tv1.tv_sec || t1 == tv1.tv_sec + 1, "time %ld vs gettimeofday %ld", (long)t1, (long)tv1.tv_sec);

	/* sched_getcpu (getcpu in the vDSO): one of our 4 CPUs. */
	int cpu = sched_getcpu();
	CHECK(cpu >= 0 && cpu < 64, "sched_getcpu %d", cpu);
	unsigned scpu = 999, node = 999;
	CHECK(syscall(SYS_getcpu, &scpu, &node, NULL) == 0 && scpu < 64 && node == 0, "getcpu syscall %u %u", scpu, node);
	/* Pin ourselves? No sched_setaffinity here: just check both agree
	 * at least once in a few tries (the scheduler may move us between). */
	int same = 0;
	for (int i = 0; i < 50 && !same; i++) {
		cpu = sched_getcpu();
		syscall(SYS_getcpu, &scpu, NULL, NULL);
		same = (unsigned)cpu == scpu;
	}
	CHECK(same, "sched_getcpu %d never matched getcpu %u", cpu, scpu);
}

/* Writing to the vvar page is a SIGSEGV (it's mapped read-only). */
static void vvar_is_read_only(void)
{
	unsigned long base = getauxval(AT_SYSINFO_EHDR);
	if (!base)
		return;
	pid_t pid = fork();
	if (pid == 0) {
		*(volatile unsigned *)(base - 4096) = 1;
		_exit(0);
	}
	int status = 0;
	waitpid(pid, &status, 0);
	CHECK(WIFSIGNALED(status) && WTERMSIG(status) == SIGSEGV, "write to vvar: status %#x", status);
}

/* A forked child has the vDSO too (same pages, same answers). */
static void after_fork(void)
{
	long long before = vdso_ns(CLOCK_MONOTONIC);
	pid_t pid = fork();
	if (pid == 0) {
		long long a = vdso_ns(CLOCK_MONOTONIC), b = sys_ns(CLOCK_MONOTONIC), c = vdso_ns(CLOCK_MONOTONIC);
		_exit(getauxval(AT_SYSINFO_EHDR) && a >= before && a <= b && b <= c ? 0 : 1);
	}
	int status = 0;
	waitpid(pid, &status, 0);
	CHECK(WIFEXITED(status) && WEXITSTATUS(status) == 0, "vDSO after fork: status %#x", status);
}

/* And a program started by execve gets a fresh one. */
static void after_exec(const char *self)
{
	pid_t pid = fork();
	if (pid == 0) {
		char *argv[] = { (char *)self, "exec", NULL };
		execv(self, argv);
		_exit(99);
	}
	int status = 0;
	waitpid(pid, &status, 0);
	CHECK(WIFEXITED(status) && WEXITSTATUS(status) == 0, "vdsotest after execve: status %#x", status);
}

/* A dynamically linked program (PT_INTERP) can't run here: ENOEXEC. */
static void dynamic_is_enoexec(void)
{
	char *argv[] = { "dynhello", NULL };
	errno = 0;
	execv("/mnt/disk0/bin/dynhello", argv);
	CHECK(errno == ENOEXEC, "execve of a dynamic program: errno %d", errno);
}

int main(int argc, char **argv)
{
	if (argc > 1 && strcmp(argv[1], "calls") == 0) {
		/* Lots of clock reads; tests/vdso.rs checks none is a syscall. */
		struct timespec t;
		struct timeval tv;
		long long sum = 0;
		for (int i = 0; i < 1000; i++) {
			clock_gettime(CLOCK_MONOTONIC, &t);
			sum += t.tv_nsec;
			clock_gettime(CLOCK_REALTIME, &t);
			gettimeofday(&tv, NULL);
			sum += time(NULL) + sched_getcpu();
		}
		return sum == 42; /* use the result; 0 */
	}
	if (argc > 1 && strcmp(argv[1], "exec") == 0) {
		image();
		clocks();
		return fails;
	}
	image();
	clocks();
	vvar_is_read_only();
	after_fork();
	after_exec((const char *)getauxval(AT_EXECFN));
	dynamic_is_enoexec();
	printf("vdsotest: %d failed\n", fails);
	return fails;
}