src/lib/file-dotlock.c

	file-dotlock.c revision 021665c0147e5c655efb2c3f0eff2549b31f6b5a
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina/* Copyright (c) 2003-2008 Dovecot authors, see the included COPYING file */
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include "lib.h"
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include "ioloop.h"
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include "str.h"
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include "hex-binary.h"
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include "hostpid.h"
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include "randgen.h"
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include "write-full.h"
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include "safe-mkstemp.h"
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include "nfs-workarounds.h"
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include "file-dotlock.h"
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include <stdio.h>
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include <stdlib.h>
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include <signal.h>
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include <time.h>
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include <utime.h>
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#include <sys/stat.h>
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#define DEFAULT_LOCK_SUFFIX ".lock"
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina/* 0.1 .. 0.2msec */
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#define LOCK_RANDOM_USLEEP_TIME (100000 + (unsigned int)rand() % 100000)
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina/* If the dotlock is newer than this, don't verify that the PID it contains
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina   is valid (since it most likely is). */
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#define STALE_PID_CHECK_SECS 2
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina/* Maximum difference between current time and create file's ctime before
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina   logging a warning. Should be less than a second in normal operation. */
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina#define MAX_TIME_DIFF 30
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březinastruct dotlock {
968e2bc3bb28458b25c3b19868ea50f0325d8b0aPavel Březina    struct dotlock_settings settings;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    dev_t dev;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    ino_t ino;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    time_t mtime;
968e2bc3bb28458b25c3b19868ea50f0325d8b0aPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    char *path;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    char *lock_path;
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina    int fd;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    time_t lock_time;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina};
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březinastruct file_change_info {
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    dev_t dev;
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina    ino_t ino;
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina    off_t size;
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina    time_t ctime, mtime;
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina};
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březinastruct lock_info {
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    const struct dotlock_settings *set;
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina    const char *path, *lock_path, *temp_path;
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina    int fd;
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina    struct file_change_info lock_info;
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina    struct file_change_info file_info;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    time_t last_pid_check;
e6092759dc4268b105b79478aa9b1398f21ed7a9Pavel Březina    time_t last_change;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    unsigned int have_pid:1;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    unsigned int pid_read:1;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    unsigned int use_io_notify:1;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    unsigned int lock_stated:1;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina};
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březinastatic struct dotlock *
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březinafile_dotlock_alloc(const struct dotlock_settings *settings)
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina{
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    struct dotlock *dotlock;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    dotlock = i_new(struct dotlock, 1);
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    dotlock->settings = *settings;
968e2bc3bb28458b25c3b19868ea50f0325d8b0aPavel Březina    if (dotlock->settings.lock_suffix == NULL)
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina        dotlock->settings.lock_suffix = DEFAULT_LOCK_SUFFIX;
968e2bc3bb28458b25c3b19868ea50f0325d8b0aPavel Březina    dotlock->fd = -1;
968e2bc3bb28458b25c3b19868ea50f0325d8b0aPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    return dotlock;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina}
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březinastatic pid_t read_local_pid(const char *lock_path)
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina{
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    char buf[512], *host;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    int fd;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    ssize_t ret;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    pid_t pid;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    fd = open(lock_path, O_RDONLY);
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    if (fd == -1)
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina        return -1; /* ignore the actual error */
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    /* read line */
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    ret = read(fd, buf, sizeof(buf)-1);
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    (void)close(fd);
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina    if (ret <= 0)
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina        return -1;
f8c829e72968b574e1c9bda96f4d5f206622358fPavel Březina
    /* fix the string */
    if (buf[ret-1] == '\n')
        ret--;
    buf[ret] = '\0';

    /* it should contain pid:host */
    host = strchr(buf, ':');
    if (host == NULL)
        return -1;
    *host++ = '\0';

    /* host must be ours */
    if (strcmp(host, my_hostname) != 0)
        return -1;

    if (!is_numeric(buf, '\0'))
        return -1;
    pid = (pid_t)strtoul(buf, NULL, 0);
    if (pid <= 0)
        return -1;
    return pid;
}

static bool
update_change_info(const struct stat *st, struct file_change_info *change,
           time_t *last_change_r, time_t now, bool check_ctime)
{
    /* ctime is checked only if we're not doing NFS attribute cache
       flushes. it changes them. */
    if (change->ino != st->st_ino || !CMP_DEV_T(change->dev, st->st_dev) ||
        (change->ctime != st->st_ctime && check_ctime) ||
        change->mtime != st->st_mtime || change->size != st->st_size) {
        time_t change_time = now;

        if (change->ctime == 0) {
            /* First check, set last_change to file's change time.
               Use mtime instead if it's higher, but only if it's
               not higher than current time, because the mtime
               can also be used for keeping metadata. */
            change_time = st->st_mtime <= now &&
                (st->st_mtime > st->st_ctime || !check_ctime) ?
                st->st_mtime : st->st_ctime;
        }
        if (*last_change_r < change_time)
            *last_change_r = change_time;
        change->ino = st->st_ino;
        change->dev = st->st_dev;
        change->ctime = st->st_ctime;
        change->mtime = st->st_mtime;
        change->size = st->st_size;
        return TRUE;
    }
    return FALSE;
}

static int update_lock_info(time_t now, struct lock_info *lock_info,
                bool *changed_r)
{
    struct stat st;

    /* don't waste time flushing attribute cache the first time we're here.
       if it's stale we'll get back here soon. */
    if (lock_info->set->nfs_flush && lock_info->lock_stated) {
        nfs_flush_file_handle_cache(lock_info->lock_path);
        nfs_flush_attr_cache_unlocked(lock_info->lock_path);
    }

    lock_info->lock_stated = TRUE;
    if (nfs_safe_lstat(lock_info->lock_path, &st) < 0) {
        if (errno != ENOENT) {
            i_error("lstat(%s) failed: %m", lock_info->lock_path);
            return -1;
        }
        return 1;
    }

    *changed_r = update_change_info(&st, &lock_info->lock_info,
                    &lock_info->last_change, now,
                    !lock_info->set->nfs_flush);
    return 0;
}

static int dotlock_override(struct lock_info *lock_info)
{
    if (unlink(lock_info->lock_path) < 0 && errno != ENOENT) {
        i_error("unlink(%s) failed: %m",
            lock_info->lock_path);
        return -1;
    }

    /* make sure we sleep for a while after overriding the lock file.
       otherwise another process might try to override it at the same time
       and unlink our newly created dotlock. */
    if (lock_info->use_io_notify)
        usleep(LOCK_RANDOM_USLEEP_TIME);
    return 0;
}

static int check_lock(time_t now, struct lock_info *lock_info)
{
    time_t stale_timeout = lock_info->set->stale_timeout;
    pid_t pid = -1;
    bool changed;
    int ret;

    if ((ret = update_lock_info(now, lock_info, &changed)) != 0)
        return ret;
    if (changed || !lock_info->pid_read) {
        /* either our first check or someone else got the lock file.
           if the dotlock was created only a couple of seconds ago,
           don't bother to read its PID. */
        if (lock_info->lock_info.mtime >= now - STALE_PID_CHECK_SECS)
            lock_info->pid_read = FALSE;
        else {
            pid = read_local_pid(lock_info->lock_path);
            lock_info->pid_read = TRUE;
        }
        lock_info->have_pid = pid != -1;
    } else if (!lock_info->have_pid) {
        /* no pid checking */
    } else {
        if (lock_info->last_pid_check == now) {
            /* we just checked the pid */
            return 0;
        }

        /* re-read the pid. even if all times and inodes are the same,
           the PID in the file might have changed if lock files were
           rapidly being recreated. */
        pid = read_local_pid(lock_info->lock_path);
        lock_info->have_pid = pid != -1;
    }

    if (lock_info->have_pid) {
        /* we've local PID. Check if it exists. */
        if (kill(pid, 0) == 0 || errno != ESRCH) {
            if (pid != getpid())
                return 0;
            /* it's us. either we're locking it again, or it's a
               stale lock file with same pid than us. either way,
               recreate it.. */
        }

        /* doesn't exist - now check again if the dotlock was just
           deleted or replaced */
        if ((ret = update_lock_info(now, lock_info, &changed)) != 0)
            return ret;

        if (!changed) {
            /* still there, go ahead and override it */
            return dotlock_override(lock_info);
        }
        return 1;
    }

    if (stale_timeout == 0) {
        /* no change checking */
        return 0;
    }

    if (now > lock_info->last_change + stale_timeout) {
        struct stat st;

        /* possibly stale lock file. check also the timestamp of the
           file we're protecting. */
        if (lock_info->set->nfs_flush) {
            nfs_flush_file_handle_cache(lock_info->path);
            nfs_flush_attr_cache_maybe_locked(lock_info->path);
        }
        if (nfs_safe_stat(lock_info->path, &st) < 0) {
            if (errno == ENOENT) {
                /* file doesn't exist. treat it as if
                   it hasn't changed */
            } else {
                i_error("stat(%s) failed: %m", lock_info->path);
                return -1;
            }
        } else {
            (void)update_change_info(&st, &lock_info->file_info,
                         &lock_info->last_change, now,
                         !lock_info->set->nfs_flush);
        }
    }

    if (now > lock_info->last_change + stale_timeout) {
        /* no changes for a while, assume stale lock */
        return dotlock_override(lock_info);
    }

    return 0;
}

static int file_write_pid(int fd, const char *path)
{
    const char *str;

    /* write our pid and host, if possible */
    str = t_strdup_printf("%s:%s", my_pid, my_hostname);
    if (write_full(fd, str, strlen(str)) < 0) {
        /* failed, leave it empty then */
        if (ftruncate(fd, 0) < 0) {
            i_error("ftruncate(%s) failed: %m", path);
            return -1;
        }
    }
    return 0;
}

static int try_create_lock_hardlink(struct lock_info *lock_info, bool write_pid,
                    string_t *tmp_path)
{
    const char *temp_prefix = lock_info->set->temp_prefix;
    const char *p;

    if (lock_info->temp_path == NULL) {
        /* we'll need our temp file first. */
        i_assert(lock_info->fd == -1);

        p = strrchr(lock_info->lock_path, '/');

        str_truncate(tmp_path, 0);
        if (temp_prefix != NULL) {
            if (*temp_prefix != '/' && p != NULL) {
                /* add directory */
                str_append_n(tmp_path, lock_info->lock_path,
                         p - lock_info->lock_path);
                str_append_c(tmp_path, '/');
            }
            str_append(tmp_path, temp_prefix);
        } else {
            if (p != NULL) {
                /* add directory */
                str_append_n(tmp_path, lock_info->lock_path,
                         p - lock_info->lock_path);
                str_append_c(tmp_path, '/');
            }
            str_printfa(tmp_path, ".temp.%s.%s.",
                    my_hostname, my_pid);
        }

        lock_info->fd = safe_mkstemp(tmp_path, 0666,
                         (uid_t)-1, (gid_t)-1);
        if (lock_info->fd == -1)
            return -1;

        if (write_pid) {
            if (file_write_pid(lock_info->fd,
                       str_c(tmp_path)) < 0) {
                (void)close(lock_info->fd);
                lock_info->fd = -1;
                return -1;
            }
        }

                lock_info->temp_path = str_c(tmp_path);
    }

    if (nfs_safe_link(lock_info->temp_path,
              lock_info->lock_path, TRUE) < 0) {
        if (errno == EEXIST)
            return 0;

        i_error("link(%s, %s) failed: %m",
            lock_info->temp_path, lock_info->lock_path);
        return -1;
    }

    if (unlink(lock_info->temp_path) < 0) {
        i_error("unlink(%s) failed: %m", lock_info->temp_path);
        /* non-fatal, continue */
    }
    lock_info->temp_path = NULL;
    return 1;
}

static int try_create_lock_excl(struct lock_info *lock_info, bool write_pid)
{
    int fd;

    fd = open(lock_info->lock_path, O_RDWR | O_EXCL | O_CREAT, 0666);
    if (fd == -1) {
        if (errno == EEXIST)
            return 0;

        if (errno != ENOENT)
            i_error("open(%s) failed: %m", lock_info->lock_path);
        return -1;
    }

    if (write_pid) {
        if (file_write_pid(fd, lock_info->lock_path) < 0) {
            (void)close(fd);
            return -1;
        }
    }

    lock_info->fd = fd;
    return 1;
}

static void dotlock_wait_end(struct ioloop *ioloop)
{
    io_loop_stop(ioloop);
}

static void dotlock_wait(struct lock_info *lock_info)
{
    struct ioloop *ioloop;
    struct io *io;
    struct timeout *to;

    if (!lock_info->use_io_notify) {
        usleep(LOCK_RANDOM_USLEEP_TIME);
        return;
    }

    ioloop = io_loop_create();
    switch (io_add_notify(lock_info->lock_path, dotlock_wait_end,
                  ioloop, &io)) {
    case IO_NOTIFY_ADDED:
        break;
    case IO_NOTIFY_NOTFOUND:
        /* the lock file doesn't exist anymore, don't sleep */
        io_loop_destroy(&ioloop);
        return;
    case IO_NOTIFY_DISABLED:
        /* listening for files not supported */
        io_loop_destroy(&ioloop);
        lock_info->use_io_notify = FALSE;
        usleep(LOCK_RANDOM_USLEEP_TIME);
        return;
    }
    to = timeout_add(LOCK_RANDOM_USLEEP_TIME/1000,
             dotlock_wait_end, ioloop);
    io_loop_run(ioloop);
    io_remove(&io);
    timeout_remove(&to);
    io_loop_destroy(&ioloop);
}

static int dotlock_create(const char *path, struct dotlock *dotlock,
              enum dotlock_create_flags flags, bool write_pid,
              const char **lock_path_r)
{
    const struct dotlock_settings *set = &dotlock->settings;
    const char *lock_path;
    struct lock_info lock_info;
    struct stat st;
    unsigned int stale_notify_threshold;
    unsigned int change_secs, wait_left;
    time_t now, max_wait_time, last_notify;
    string_t *tmp_path;
    int ret;
    bool do_wait;

    now = time(NULL);

    lock_path = *lock_path_r = t_strconcat(path, set->lock_suffix, NULL);
    stale_notify_threshold = set->stale_timeout / 2;
    max_wait_time = (flags & DOTLOCK_CREATE_FLAG_NONBLOCK) != 0 ? 0 :
        now + set->timeout;
    tmp_path = t_str_new(256);

    memset(&lock_info, 0, sizeof(lock_info));
    lock_info.path = path;
    lock_info.set = set;
    lock_info.lock_path = lock_path;
    lock_info.fd = -1;
    lock_info.use_io_notify = set->use_io_notify;

    last_notify = 0; do_wait = FALSE;

    do {
        if (do_wait) {
            dotlock_wait(&lock_info);
            do_wait = FALSE;
        }

        ret = check_lock(now, &lock_info);
        if (ret < 0)
            break;

        if (ret == 1) {
            if ((flags & DOTLOCK_CREATE_FLAG_CHECKONLY) != 0)
                break;

            ret = set->use_excl_lock ?
                try_create_lock_excl(&lock_info, write_pid) :
                try_create_lock_hardlink(&lock_info, write_pid,
                             tmp_path);
            if (ret != 0)
                break;
        }

        do_wait = TRUE;
        if (last_notify != now && set->callback != NULL) {
            last_notify = now;
            change_secs = now - lock_info.last_change;
            wait_left = max_wait_time - now;

            if (change_secs >= stale_notify_threshold &&
                change_secs <= wait_left) {
                unsigned int secs_left =
                    set->stale_timeout < change_secs ?
                    0 : set->stale_timeout - change_secs;
                if (!set->callback(secs_left, TRUE,
                           set->context)) {
                    /* we don't want to override */
                    lock_info.last_change = now;
                }
            } else {
                (void)set->callback(wait_left, FALSE,
                            set->context);
            }
        }

        now = time(NULL);
    } while (now < max_wait_time);

    if (ret > 0) {
        if (fstat(lock_info.fd, &st) < 0) {
            i_error("fstat(%s) failed: %m", lock_path);
            ret = -1;
        } else {
            /* successful dotlock creation */
            dotlock->dev = st.st_dev;
            dotlock->ino = st.st_ino;

            dotlock->path = i_strdup(path);
            dotlock->fd = lock_info.fd;
                        dotlock->lock_time = now;
            lock_info.fd = -1;

            if (st.st_ctime + MAX_TIME_DIFF < now ||
                st.st_ctime - MAX_TIME_DIFF > now) {
                i_warning("Created dotlock file's timestamp is "
                      "different than current time "
                      "(%s vs %s): %s", dec2str(st.st_ctime),
                      dec2str(now), path);
            }
        }
    }

    if (lock_info.fd != -1) {
        int old_errno = errno;

        if (close(lock_info.fd) < 0)
            i_error("close(%s) failed: %m", lock_path);
        errno = old_errno;
    }
    if (lock_info.temp_path != NULL) {
        if (unlink(lock_info.temp_path) < 0)
            i_error("unlink(%s) failed: %m", lock_info.temp_path);
    }

    if (ret == 0)
        errno = EAGAIN;
    return ret;
}

static void file_dotlock_free(struct dotlock **_dotlock)
{
    struct dotlock *dotlock = *_dotlock;
    int old_errno;

    *_dotlock = NULL;

    if (dotlock->fd != -1) {
        old_errno = errno;
        if (close(dotlock->fd) < 0)
            i_error("close(%s) failed: %m", dotlock->path);
        dotlock->fd = -1;
        errno = old_errno;
    }

    i_free(dotlock->path);
    i_free(dotlock->lock_path);
    i_free(dotlock);
}

static int file_dotlock_create_real(struct dotlock *dotlock, const char *path,
                    enum dotlock_create_flags flags)
{
    const char *lock_path;
    struct stat st;
    int fd, ret;

    ret = dotlock_create(path, dotlock, flags, TRUE, &lock_path);
    if (ret <= 0 || (flags & DOTLOCK_CREATE_FLAG_CHECKONLY) != 0)
        return ret;

    fd = dotlock->fd;
    dotlock->fd = -1;

    if (close(fd) < 0) {
        i_error("close(%s) failed: %m", lock_path);
        return -1;
    }

    /* With NFS the writes may have been flushed only when closing the
       file. Get the mtime again after that to avoid "dotlock was modified"
       errors. */
    if (lstat(lock_path, &st) < 0) {
        if (errno != ENOENT)
            i_error("stat(%s) failed: %m", lock_path);
        else {
            i_error("dotlock %s was immediately deleted under us",
                lock_path);
        }
        return -1;
    }
    /* extra sanity check won't hurt.. */
    if (st.st_dev != dotlock->dev || st.st_ino != dotlock->ino) {
        errno = ENOENT;
        i_error("dotlock %s was immediately recreated under us",
            lock_path);
        return -1;
    }
    dotlock->mtime = st.st_mtime;
    return 1;
}

int file_dotlock_create(const struct dotlock_settings *set, const char *path,
            enum dotlock_create_flags flags,
            struct dotlock **dotlock_r)
{
    struct dotlock *dotlock;
    int ret;

    dotlock = file_dotlock_alloc(set);
    T_FRAME(
        ret = file_dotlock_create_real(dotlock, path, flags);
    );
    if (ret <= 0 || (flags & DOTLOCK_CREATE_FLAG_CHECKONLY) != 0)
        file_dotlock_free(&dotlock);

    *dotlock_r = dotlock;
    return ret;
}

int file_dotlock_delete(struct dotlock **dotlock_p)
{
    struct dotlock *dotlock;
    const char *lock_path;
        struct stat st;

    dotlock = *dotlock_p;
    *dotlock_p = NULL;

    lock_path = file_dotlock_get_lock_path(dotlock);
    if (nfs_safe_lstat(lock_path, &st) < 0) {
        if (errno == ENOENT) {
            i_warning("Our dotlock file %s was deleted "
                  "(kept it %d secs)", lock_path,
                  (int)(time(NULL) - dotlock->lock_time));
            file_dotlock_free(&dotlock);
            return 0;
        }

        i_error("lstat(%s) failed: %m", lock_path);
        file_dotlock_free(&dotlock);
        return -1;
    }

    if (dotlock->ino != st.st_ino ||
        !CMP_DEV_T(dotlock->dev, st.st_dev)) {
        i_warning("Our dotlock file %s was overridden "
              "(kept it %d secs)", lock_path,
              (int)(dotlock->lock_time - time(NULL)));
        errno = EEXIST;
        file_dotlock_free(&dotlock);
        return 0;
    }

    if (dotlock->mtime != st.st_mtime && dotlock->fd == -1) {
        i_warning("Our dotlock file %s was modified (%s vs %s), "
              "assuming it wasn't overridden (kept it %d secs)",
              lock_path,
              dec2str(dotlock->mtime), dec2str(st.st_mtime),
              (int)(time(NULL) - dotlock->lock_time));
    }

    if (unlink(lock_path) < 0) {
        if (errno == ENOENT) {
            i_warning("Our dotlock file %s was deleted "
                  "(kept it %d secs)", lock_path,
                  (int)(time(NULL) - dotlock->lock_time));
            file_dotlock_free(&dotlock);
            return 0;
        }

        i_error("unlink(%s) failed: %m", lock_path);
        file_dotlock_free(&dotlock);
        return -1;
    }

    file_dotlock_free(&dotlock);
    return 1;
}

int file_dotlock_open(const struct dotlock_settings *set, const char *path,
              enum dotlock_create_flags flags,
              struct dotlock **dotlock_r)
{
    struct dotlock *dotlock;
    int ret;

    dotlock = file_dotlock_alloc(set);
    T_FRAME(
        const char *lock_path;

        ret = dotlock_create(path, dotlock, flags, FALSE, &lock_path);
    );

    if (ret <= 0) {
        file_dotlock_free(&dotlock);
        *dotlock_r = NULL;
        return -1;
    }

    *dotlock_r = dotlock;
    return dotlock->fd;
}

int file_dotlock_replace(struct dotlock **dotlock_p,
             enum dotlock_replace_flags flags)
{
    struct dotlock *dotlock;
    const char *lock_path;
    int fd;

    dotlock = *dotlock_p;
    *dotlock_p = NULL;

    fd = dotlock->fd;
    if ((flags & DOTLOCK_REPLACE_FLAG_DONT_CLOSE_FD) != 0)
        dotlock->fd = -1;

    lock_path = file_dotlock_get_lock_path(dotlock);
    if ((flags & DOTLOCK_REPLACE_FLAG_VERIFY_OWNER) != 0 &&
        !file_dotlock_is_locked(dotlock)) {
        i_warning("Our dotlock file %s was overridden "
              "(kept it %d secs)", lock_path,
              (int)(time(NULL) - dotlock->lock_time));
        errno = EEXIST;
        file_dotlock_free(&dotlock);
        return 0;
    }

    if (rename(lock_path, dotlock->path) < 0) {
        i_error("rename(%s, %s) failed: %m", lock_path, dotlock->path);
        file_dotlock_free(&dotlock);
        return -1;
    }
    file_dotlock_free(&dotlock);
    return 1;
}

int file_dotlock_touch(struct dotlock *dotlock)
{
    time_t now = time(NULL);
    struct utimbuf buf;
    int ret = 0;

    if (dotlock->mtime == now)
        return 0;

    dotlock->mtime = now;
    buf.actime = buf.modtime = now;

    T_FRAME(
        const char *lock_path = file_dotlock_get_lock_path(dotlock);
        if (utime(lock_path, &buf) < 0) {
            i_error("utime(%s) failed: %m", lock_path);
            ret = -1;
        }
    );
    return ret;
}

bool file_dotlock_is_locked(struct dotlock *dotlock)
{
    struct stat st, st2;
    const char *lock_path;

    lock_path = file_dotlock_get_lock_path(dotlock);
    if (fstat(dotlock->fd, &st) < 0) {
        i_error("fstat(%s) failed: %m", lock_path);
        return FALSE;
    }

    if (nfs_safe_lstat(lock_path, &st2) < 0) {
        i_error("lstat(%s) failed: %m", lock_path);
        return FALSE;
    }
    return st.st_ino == st2.st_ino && CMP_DEV_T(st.st_dev, st2.st_dev);
}

const char *file_dotlock_get_lock_path(struct dotlock *dotlock)
{
    if (dotlock->lock_path == NULL) {
        dotlock->lock_path =
            i_strconcat(dotlock->path,
                    dotlock->settings.lock_suffix, NULL);
    }
    return dotlock->lock_path;
}