flint_lock.cc

Go to the documentation of this file.

/* Copyright (C) 2005,2006,2007,2008,2009,2010,2011,2012,2013,2014,2015,2016,2017 Olly Betts
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public License as
 * published by the Free Software Foundation; either version 2 of the
 * License, or (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301
 * USA
 */

#include <config.h>

#include "flint_lock.h"

#ifndef __WIN32__
#include <cerrno>

#include "safefcntl.h"
#include <unistd.h>
#include <cstdlib>
#include <sys/types.h>
#include "safesyssocket.h"
#include <sys/wait.h>
#include <signal.h>
#include <cstring>
#endif

#include "closefrom.h"
#include "errno_to_string.h"
#include "omassert.h"

#ifdef __CYGWIN__
# include <cygwin/version.h>
# include <sys/cygwin.h>
#endif

#ifdef FLINTLOCK_USE_FLOCK
# include <sys/file.h>
#endif

#include "xapian/error.h"

using namespace std;

#ifndef F_OFD_SETLK
# ifdef __linux__
// Apparently defining _GNU_SOURCE should get us F_OFD_SETLK, etc, but that
// doesn't actually seem to work, so hard-code the known values.
#  define F_OFD_GETLK   36
#  define F_OFD_SETLK   37
#  define F_OFD_SETLKW  38
# endif
#endif

XAPIAN_NORETURN(static void throw_cannot_test_lock());
static void
throw_cannot_test_lock()
{
    throw Xapian::FeatureUnavailableError("Can't test lock without trying to "
                                          "take it");
}

bool
FlintLock::test() const
{
    // A database which doesn't support update can't be locked for update.
    if (filename.empty()) return false;

#if defined __CYGWIN__ || defined __WIN32__
    if (hFile != INVALID_HANDLE_VALUE) return true;
    // Doesn't seem to be possible to check if the lock is held without briefly
    // taking the lock.
    throw_cannot_test_lock();
#elif defined FLINTLOCK_USE_FLOCK
    if (fd != -1) return true;
    // Doesn't seem to be possible to check if the lock is held without briefly
    // taking the lock.
    throw_cannot_test_lock();
#else
    if (fd != -1) return true;
    int lockfd = open(filename.c_str(), O_WRONLY | O_CREAT | O_TRUNC | O_CLOEXEC, 0666);
    if (lockfd < 0) {
        // Couldn't open lockfile.
        reason why = ((errno == EMFILE || errno == ENFILE) ? FDLIMIT : UNKNOWN);
        throw_databaselockerror(why, filename, "Testing lock");
    }

    struct flock fl;
    fl.l_type = F_WRLCK;
    fl.l_whence = SEEK_SET;
    fl.l_start = 0;
    fl.l_len = 1;
    fl.l_pid = 0;
    while (fcntl(lockfd, F_GETLK, &fl) == -1) {
        if (errno != EINTR) {
            // Translate known errno values into a reason code.
            int e = errno;
            close(lockfd);
            if (e == ENOSYS) {
                // F_GETLK always failed with ENOSYS on older GNU Hurd libc
                // versions: https://bugs.debian.org/190367
                throw_cannot_test_lock();
            }
            reason why = (e == ENOLCK ? UNSUPPORTED : UNKNOWN);
            throw_databaselockerror(why, filename, "Testing lock");
        }
    }
    close(lockfd);
    return fl.l_type != F_UNLCK;
#endif
}

FlintLock::reason
FlintLock::lock(bool exclusive, bool wait, string & explanation) {
    // Currently we only support exclusive locks.
    (void)exclusive;
    Assert(exclusive);
#if defined __CYGWIN__ || defined __WIN32__
    Assert(hFile == INVALID_HANDLE_VALUE);
#ifdef __CYGWIN__
    char fnm[MAX_PATH];
#if CYGWIN_VERSION_API_MAJOR == 0 && CYGWIN_VERSION_API_MINOR < 181
    cygwin_conv_to_win32_path(filename.c_str(), fnm);
#else
    if (cygwin_conv_path(CCP_POSIX_TO_WIN_A|CCP_RELATIVE, filename.c_str(),
                         fnm, MAX_PATH) < 0) {
        explanation.assign("cygwin_conv_path failed: ");
        errno_to_string(errno, explanation);
        return UNKNOWN;
    }
#endif
#else
    const char *fnm = filename.c_str();
#endif
retry:
    // FIXME: Use LockFileEx() for locking, which would allow proper blocking
    // and also byte-range locking for when we implement MVCC.  But is there a
    // way to interwork with the CreateFile()-based locking while doing so?
    hFile = CreateFile(fnm, GENERIC_WRITE, FILE_SHARE_READ,
                       NULL, OPEN_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
    if (hFile != INVALID_HANDLE_VALUE) return SUCCESS;
    if (GetLastError() == ERROR_ALREADY_EXISTS) {
        if (wait) {
            Sleep(1000);
            goto retry;
        }
        return INUSE;
    }
    explanation = string();
    return UNKNOWN;
#elif defined FLINTLOCK_USE_FLOCK
    // This is much simpler than using fcntl() due to saner semantics around
    // releasing locks when closing other descriptors on the same file (at
    // least on platforms where flock() isn't just a compatibility wrapper
    // around fcntl()).  We can't simply switch to this without breaking
    // locking compatibility with previous releases, though it might be useful
    // for porting to platforms without fcntl() locking.
    //
    // Also, flock() is problematic over NFS at least on Linux - it's been
    // supported since Linux 2.6.12 but it's actually emulated by taking out an
    // fcntl() byte-range lock on the entire file, which means that a process
    // on the NFS server can get a (genuine) flock() lock on the same file a
    // process on an NFS client has locked by flock() emulated as an fcntl()
    // lock.
    Assert(fd == -1);
    int lockfd = open(filename.c_str(), O_WRONLY | O_CREAT | O_TRUNC | O_CLOEXEC, 0666);
    if (lockfd < 0) {
        // Couldn't open lockfile.
        explanation.assign("Couldn't open lockfile: ");
        errno_to_string(errno, explanation);
        return ((errno == EMFILE || errno == ENFILE) ? FDLIMIT : UNKNOWN);
    }

    int op = LOCK_EX;
    if (!wait) op |= LOCK_NB;
    while (flock(lockfd, op) == -1) {
        if (errno != EINTR) {
            // Lock failed - translate known errno values into a reason code.
            close(lockfd);
            switch (errno) {
                case EWOULDBLOCK:
                    return INUSE;
                case ENOLCK:
                    return UNSUPPORTED; // FIXME: what do we get for NFS?
                default:
                    return UNKNOWN;
            }
        }
    }

    fd = lockfd;
    return SUCCESS;
#else
    Assert(fd == -1);
#if defined F_SETFD && defined FD_CLOEXEC
    int lockfd = open(filename.c_str(), O_WRONLY | O_CREAT | O_TRUNC | O_CLOEXEC, 0666);
#else
    int lockfd = open(filename.c_str(), O_WRONLY | O_CREAT | O_TRUNC, 0666);
#endif
    if (lockfd < 0) {
        // Couldn't open lockfile.
        explanation.assign("Couldn't open lockfile: ");
        errno_to_string(errno, explanation);
        return ((errno == EMFILE || errno == ENFILE) ? FDLIMIT : UNKNOWN);
    }

#ifdef F_OFD_SETLK
    // F_OFD_SETLK has exactly the semantics we want, so use it if it's
    // available.  Support was added in Linux 3.15, and it was accepted
    // for POSIX issue 8 on 2022-12-15:
    // https://austingroupbugs.net/view.php?id=768

    // Use a static flag so we don't repeatedly try F_OFD_SETLK when
    // the kernel in use doesn't support it.  This should be safe in a
    // threaded context - at worst multiple threads might end up trying
    // F_OFD_SETLK and then setting f_ofd_setlk_fails to true.
    static bool f_ofd_setlk_fails = false;
    if (!f_ofd_setlk_fails) {
        struct flock fl;
        fl.l_type = F_WRLCK;
        fl.l_whence = SEEK_SET;
        fl.l_start = 0;
        fl.l_len = 1;
        fl.l_pid = 0;
        while (fcntl(lockfd, wait ? F_OFD_SETLKW : F_OFD_SETLK, &fl) == -1) {
            if (errno != EINTR) {
                if (errno == EINVAL) {
                    // F_OFD_SETLK not supported by this kernel.
                    goto no_ofd_support;
                }
                // Lock failed - translate known errno values into a reason
                // code.
                int e = errno;
                close(lockfd);
                switch (e) {
                    case EACCES: case EAGAIN:
                        return INUSE;
                    case ENOLCK:
                        return UNSUPPORTED;
                    default:
                        return UNKNOWN;
                }
            }
        }
        fd = lockfd;
        pid = 0;
        return SUCCESS;
no_ofd_support:
        f_ofd_setlk_fails = true;
    }
#endif

    // If stdin and/or stdout have been closed, it is possible that lockfd could
    // be 0 or 1.  We need fds 0 and 1 to be available in the child process to
    // be stdin and stdout, and we can't use dup() on lockfd after locking it,
    // as the lock won't be transferred, so we handle this corner case here by
    // using F_DUPFD or by calling dup() once or twice so that lockfd >= 2.
    if (rare(lockfd < 2)) {
        // Note this temporarily requires one or two spare fds to work, but
        // then we need two spare for socketpair() to succeed below anyway.
#ifdef F_DUPFD
        // Where available, F_DUPFD allows us to directly get the first unused
        // fd which is at least 2.
        int lockfd_dup = fcntl(lockfd, F_DUPFD, 2);
        int eno = errno;
        close(lockfd);
        if (lockfd_dup < 0) {
            return ((eno == EMFILE || eno == ENFILE) ? FDLIMIT : UNKNOWN);
        }
        lockfd = lockfd_dup;
#else
        // Otherwise we have to call dup() until we get one, though at least
        // that's only at most twice.
        int lockfd_dup = dup(lockfd);
        if (rare(lockfd_dup < 2)) {
            int eno = 0;
            if (lockfd_dup < 0) {
                eno = errno;
                close(lockfd);
            } else {
                int lockfd_dup2 = dup(lockfd);
                if (lockfd_dup2 < 0) {
                    eno = errno;
                }
                close(lockfd);
                close(lockfd_dup);
                lockfd = lockfd_dup2;
            }
            if (eno) {
                return ((eno == EMFILE || eno == ENFILE) ? FDLIMIT : UNKNOWN);
            }
        } else {
            close(lockfd);
            lockfd = lockfd_dup;
        }
#endif
    }

    int fds[2];
    if (socketpair(AF_UNIX, SOCK_STREAM|SOCK_CLOEXEC, PF_UNSPEC, fds) < 0) {
        // Couldn't create socketpair.
        explanation.assign("Couldn't create socketpair: ");
        errno_to_string(errno, explanation);
        reason why = ((errno == EMFILE || errno == ENFILE) ? FDLIMIT : UNKNOWN);
        (void)close(lockfd);
        return why;
    }

    pid_t child = fork();

    if (child == 0) {
        // Child process.
        close(fds[0]);

#if defined F_SETFD && defined FD_CLOEXEC
        // Clear close-on-exec flag, if we set it when we called socketpair().
        // Clearing it here means there's no window where another thread in the
        // parent process could fork()+exec() and end up with this fd still
        // open (assuming close-on-exec is supported).
        //
        // We can't use a preprocessor check on the *value* of SOCK_CLOEXEC as
        // on Linux SOCK_CLOEXEC is an enum, with '#define SOCK_CLOEXEC
        // SOCK_CLOEXEC' to allow '#ifdef SOCK_CLOEXEC' to work.
        if (SOCK_CLOEXEC != 0)
            (void)fcntl(fds[1], F_SETFD, 0);
        if (O_CLOEXEC != 0)
            (void)fcntl(lockfd, F_SETFD, 0);
#endif
        // Connect pipe to stdin and stdout.
        dup2(fds[1], 0);
        dup2(fds[1], 1);

        // Make sure we don't hang on to open files which may get deleted but
        // not have their disk space released until we exit.  Close these
        // before we try to get the lock because if one of them is open on
        // the lock file then closing it after obtaining the lock would release
        // the lock, which would be really bad.
        for (int i = 2; i < lockfd; ++i) {
            // Retry on EINTR; just ignore other errors (we'll get
            // EBADF if the fd isn't open so that's OK).
            while (close(i) < 0 && errno == EINTR) { }
        }
        closefrom(lockfd + 1);

        reason why = SUCCESS;
        {
            struct flock fl;
            fl.l_type = F_WRLCK;
            fl.l_whence = SEEK_SET;
            fl.l_start = 0;
            fl.l_len = 1;
            while (fcntl(lockfd, wait ? F_SETLKW : F_SETLK, &fl) == -1) {
                if (errno != EINTR) {
                    // Lock failed - translate known errno values into a reason
                    // code.
                    if (errno == EACCES || errno == EAGAIN) {
                        why = INUSE;
                    } else if (errno == ENOLCK) {
                        why = UNSUPPORTED;
                    } else {
                        _exit(0);
                    }
                    break;
                }
            }
        }

        {
            // Tell the parent if we got the lock, and if not, why not.
            char ch = static_cast<char>(why);
            while (write(1, &ch, 1) < 0) {
                // EINTR means a signal interrupted us, so retry.
                // Otherwise we're DOOMED!  The best we can do is just exit
                // and the parent process should get EOF and know the lock
                // failed.
                if (errno != EINTR) _exit(1);
            }
            if (why != SUCCESS) _exit(0);
        }

        // Make sure we don't block unmount of partition holding the current
        // directory.
        if (chdir("/") < 0) {
            // We can't usefully do anything in response to an error, so just
            // ignore it - the worst harm it can do is make it impossible to
            // unmount a partition.
            //
            // We need the if statement because glibc's _FORTIFY_SOURCE mode
            // gives a warning even if we cast the result to void.
        }

        // FIXME: use special statically linked helper instead of cat.
        execl("/bin/cat", "/bin/cat", static_cast<void*>(NULL));
        // Emulate cat ourselves (we try to avoid this to reduce VM overhead).
        char ch;
        while (read(0, &ch, 1) != 0) {
            /* Do nothing */
        }
        _exit(0);
    }

    close(lockfd);
    close(fds[1]);

    if (child == -1) {
        // Couldn't fork.
        explanation.assign("Couldn't fork: ");
        errno_to_string(errno, explanation);
        close(fds[0]);
        return UNKNOWN;
    }

    reason why = UNKNOWN;

    // Parent process.
    while (true) {
        char ch;
        ssize_t n = read(fds[0], &ch, 1);
        if (n == 1) {
            why = static_cast<reason>(ch);
            if (why != SUCCESS) break;
            // Got the lock.
            fd = fds[0];
            pid = child;
            return SUCCESS;
        }
        if (n == 0) {
            // EOF means the lock failed.
            explanation.assign("Got EOF reading from child process");
            break;
        }
        if (errno != EINTR) {
            // Treat unexpected errors from read() as failure to get the lock.
            explanation.assign("Error reading from child process: ");
            errno_to_string(errno, explanation);
            break;
        }
    }

    close(fds[0]);

    int status;
    while (waitpid(child, &status, 0) < 0) {
        if (errno != EINTR) break;
    }

    return why;
#endif
}

void
FlintLock::release() {
#if defined __CYGWIN__ || defined __WIN32__
    if (hFile == INVALID_HANDLE_VALUE) return;
    CloseHandle(hFile);
    hFile = INVALID_HANDLE_VALUE;
#elif defined FLINTLOCK_USE_FLOCK
    if (fd < 0) return;
    close(fd);
    fd = -1;
#else
    if (fd < 0) return;
    close(fd);
    fd = -1;
#ifdef F_OFD_SETLK
    if (pid == 0) return;
#endif
    // Kill the child process which is holding the lock.  Use SIGKILL since
    // that can't be caught or ignored (we used to use SIGHUP, but if the
    // application has set that to SIG_IGN, the child process inherits that
    // setting, which sometimes results in the child process not exiting -
    // noted on Linux).
    //
    // The only likely error from kill is ESRCH (pid doesn't exist).  The other
    // possibilities (according to the Linux man page) are EINVAL (invalid
    // signal) and EPERM (don't have permission to SIGKILL the process) but in
    // none of the cases does calling waitpid do us any good!
    if (kill(pid, SIGKILL) == 0) {
        int status;
        while (waitpid(pid, &status, 0) < 0) {
            if (errno != EINTR) break;
        }
    }
#endif
}

void
FlintLock::throw_databaselockerror(FlintLock::reason why,
                                   const string & db_dir,
                                   const string & explanation) const
{
    string msg("Unable to get write lock on ");
    msg += db_dir;
    if (why == FlintLock::INUSE) {
        msg += ": already locked";
    } else if (why == FlintLock::UNSUPPORTED) {
        msg += ": locking probably not supported by this FS";
    } else if (why == FlintLock::FDLIMIT) {
        msg += ": too many open files";
    } else if (why == FlintLock::UNKNOWN) {
        if (!explanation.empty())
            msg += ": " + explanation;
    }
    throw Xapian::DatabaseLockError(msg);
}