vlibc/tests/syscall_test.c

/*
 * vlibc — syscall layer test (todo 1).
 *
 * Exercises the raw __syscall<n>() entry points and the errno TCB-slot
 * mechanism end to end:
 *
 *   1. happy path: __syscall3(SYS_write, 1, "x", 1) emits "x" on stdout and
 *      returns 1 (one byte written);
 *   2. failure path: syscall_ret(SYS_openat on a nonexistent path) returns -1
 *      and leaves errno == ENOENT in the calling thread's TCB slot;
 *   3. TCB offset mechanics: a fake TCB (a local buffer whose first word is
 *      its own address, as the real TCB layout will have) is installed as
 *      the FS thread pointer via SYS_arch_prctl; __errno_location() must
 *      then point exactly at fake + VLIBC_TCB_ERRNO_OFF, a failing syscall
 *      must land its errno in that slot, and the real thread pointer is
 *      restored afterwards.
 *   4. register pinning: a 4-arg (SYS_rt_sigprocmask) and a 5-arg
 *      (SYS_ppoll) syscall whose later arguments are load-bearing must see
 *      them in r10/r8, where the x86_64 kernel ABI places args 4 and 5.
 *
 * All diagnostics go through raw SYS_write (no printf): between installing
 * and restoring the fake thread pointer the test must not call any libc
 * function, whose TLS reads would see the fake TCB.
 *
 * Not part of the library proper; compiled manually for this todo (the
 * tests/ + make check wiring is owned by a later todo).
 */

#include "../src/internal/errno.h"
#include "../src/internal/syscall.h"
#include "../src/internal/types.h"

/* Flag constants the public <fcntl.h> will own; test-local copies. */
#define TEST_AT_FDCWD (-100)
#define TEST_O_RDONLY 0
#define TEST_O_CLOEXEC 0x80000

/* Constants the public <signal.h>/<poll.h> will own; test-local copies. */
#define TEST_SIG_BLOCK 0
#define TEST_SIG_SETMASK 2
#define TEST_SIGUSR1 10

/* arch_prctl codes (kernel UAPI). */
#define TEST_ARCH_SET_FS 0x1002
#define TEST_ARCH_GET_FS 0x1003

static int failures;

/* Write a NUL-terminated string to fd via the raw syscall layer. */
static void
say(int fd, const char *s)
{
    ssize_t n = 0;

    while (s[n] != '\0')
    {
        n++;
    }
    syscall_ret(__syscall3(SYS_write, fd, (long)s, (long)n));
}

static void
check(int cond, const char *what)
{
    if (!cond)
    {
        say(2, "FAIL: ");
        say(2, what);
        say(2, "\n");
        failures++;
    }
}

/*
 * A stand-in thread control block. The real TCB layout (owned by the
 * startup todo) will store the TCB's own address at offset 0 — the self
 * pointer __builtin_thread_pointer() reads through %fs:0 — and errno at
 * VLIBC_TCB_ERRNO_OFF. This buffer mirrors exactly those two slots.
 */
static unsigned long fake_tcb[32];

/*
 * The fake TCB must be installed over a saved copy of the real FS base:
 * the kernel's arch_prctl only reads or writes the current base, so the
 * save must happen while the real thread pointer is still installed.
 */
static int
install_fake_tcb(unsigned long *real_fs) // NOLINT(readability-non-const-parameter)
{
    long r;

    fake_tcb[0] = (unsigned long)fake_tcb; /* TCB self pointer at slot 0 */
    r = syscall_ret(__syscall2(SYS_arch_prctl, TEST_ARCH_GET_FS, (long)real_fs));
    if (r != 0)
    {
        return 1;
    }
    r = syscall_ret(__syscall2(SYS_arch_prctl, TEST_ARCH_SET_FS, (long)fake_tcb));
    if (r != 0)
    {
        return 1;
    }
    return 0;
}

static int
restore_real_tcb(unsigned long real_fs)
{
    return syscall_ret(__syscall2(SYS_arch_prctl, TEST_ARCH_SET_FS, (long)real_fs)) != 0;
}

struct test_timespec
{
    long tv_sec;
    long tv_nsec;
};

/*
 * 4-arg regression probe: block SIGUSR1, read the mask back through a NULL
 * set, restore. The 4th argument (sigsetsize) must reach the kernel in r10;
 * placed anywhere else it fails with -EINVAL.
 */
static int
probe_rt_sigprocmask(void)
{
    unsigned long mask = 1UL << (TEST_SIGUSR1 - 1);
    unsigned long oldmask = 0;
    unsigned long readback = 0;

    if (__syscall4(SYS_rt_sigprocmask, TEST_SIG_BLOCK, (long)&mask, (long)&oldmask, 8) != 0)
    {
        return 0;
    }
    if (__syscall4(SYS_rt_sigprocmask, TEST_SIG_BLOCK, 0, (long)&readback, 8) != 0)
    {
        return 0;
    }
    if (__syscall4(SYS_rt_sigprocmask, TEST_SIG_SETMASK, (long)&oldmask, 0, 8) != 0)
    {
        return 0;
    }
    return (readback & (1UL << (TEST_SIGUSR1 - 1))) != 0;
}

/*
 * 5-arg regression probe: ppoll with no fds and a zero timeout must return
 * 0. The 4th argument (sigmask) must reach r10 and the 5th (sigsetsize) r8;
 * misplaced, the kernel reads a pointer value where the set size belongs and
 * answers -EINVAL.
 */
static int
probe_ppoll(void)
{
    struct test_timespec ts = {0, 0};
    unsigned long mask = 0;

    return __syscall5(SYS_ppoll, 0, 0, (long)&ts, (long)&mask, 8) == 0;
}

/*
 * Failure scenario (-f): SYS_openat on a nonexistent path must return -1
 * and leave errno == ENOENT in the TCB slot. Prints the observed result to
 * stdout; exits 0 only when the failure behaved exactly as specified.
 */
static int
failure_scenario(void)
{
    unsigned long real_fs = 0;
    long r;
    int saved_errno;

    if (install_fake_tcb(&real_fs))
    {
        return 1;
    }
    r = syscall_ret(__syscall4(SYS_openat, TEST_AT_FDCWD, (long)"/nonexistent/vlibc",
                               TEST_O_RDONLY | TEST_O_CLOEXEC, 0));
    saved_errno = errno;
    if (restore_real_tcb(real_fs))
    {
        return 1;
    }

    if (r == -1 && saved_errno == ENOENT)
    {
        say(1, "openat=-1 errno=ENOENT\n");
        return 0;
    }
    say(1, "openat FAIL\n");
    return 1;
}

int
main(int argc, char **argv)
{
    long r;
    unsigned long real_fs = 0;

    if (argc == 2 && argv[1][0] == '-' && argv[1][1] == 'f')
    {
        return failure_scenario();
    }

    /* 1. Happy path: one raw byte to stdout, byte count returned. */
    r = __syscall3(SYS_write, 1, (long)"x", 1);
    check(r == 1, "SYS_write returned byte count");

    /* 2. Point FS at the fake TCB and verify the errno slot mechanics. */
    check(install_fake_tcb(&real_fs) == 0, "arch_prctl fake TCB install");

    {
        unsigned long now_fs = 0;

        r = syscall_ret(__syscall2(SYS_arch_prctl, TEST_ARCH_GET_FS, (long)&now_fs));
        check(r == 0 && now_fs == (unsigned long)fake_tcb,
              "arch_prctl(ARCH_GET_FS) reads fake base");
    }

    check((char *)__errno_location() == (char *)fake_tcb + VLIBC_TCB_ERRNO_OFF,
          "__errno_location == fake TCB + VLIBC_TCB_ERRNO_OFF");

    r = syscall_ret(__syscall4(SYS_openat, TEST_AT_FDCWD, (long)"/nonexistent/vlibc",
                               TEST_O_RDONLY | TEST_O_CLOEXEC, 0));
    check(r == -1, "SYS_openat on nonexistent path returns -1");
    check(errno == ENOENT, "errno == ENOENT via TCB slot");

    /* 3. Restore the real thread pointer; the fake slot keeps its value. */
    check(restore_real_tcb(real_fs) == 0, "arch_prctl(ARCH_SET_FS, real TCB) restored");
    check(*(int *)((char *)fake_tcb + VLIBC_TCB_ERRNO_OFF) == ENOENT,
          "fake TCB errno slot retains ENOENT after restore");

    /* 4. Register pinning for 4- and 5-argument syscalls (r10/r8). */
    check(probe_rt_sigprocmask(), "rt_sigprocmask 4-arg r10 pinning");
    check(probe_ppoll(), "ppoll 5-arg r10/r8 pinning");

    /*
     * VLIBC_TCB_ERRNO_OFF addresses slot 1 of whichever TCB the FS thread
     * pointer selects; under the host libc that slot is its private TLS
     * state, so an errno round-trip through the real thread pointer would
     * corrupt the host libc. The fake-TCB path proves the offset mechanics;
     * the real TCB is exercised once the startup todo installs vlibc's own
     * thread pointer.
     */

    if (failures > 0)
    {
        say(2, "FAILED\n");
    }
    return failures == 0 ? 0 : 1;
}