Differences between revisions 6 and 7
Revision 6 as of 2008-07-24 01:23:01
Size: 5828
Editor: BeJonas
Comment: re mkdir on nfs
Revision 7 as of 2008-11-22 14:08:52
Size: 5828
Editor: localhost
Comment: converted to 1.6 markup
Deletions are marked like this. Additions are marked like this.
Line 1: Line 1:
[[Anchor(faq45)]] <<Anchor(faq45)>>
Line 121: Line 121:
 Are you sure mkdir has problems with being atomic on NFS? I thought that affected only open, but I'm not really sure. -- BeJonas [[DateTime(2008-07-24T01:22:59Z)]]  Are you sure mkdir has problems with being atomic on NFS? I thought that affected only open, but I'm not really sure. -- BeJonas <<DateTime(2008-07-24T01:22:59Z)>>

How can I ensure that only one instance of a script is running at a time (mutual exclusion)?

We need some means of mutual exclusion. One easy way is to use a "lock": any number of processes can try to acquire the lock simultaneously, but only one of them will succeed.

How can we implement this using shell scripts? Some people suggest creating a lock file, and checking for its presence:

  •  # locking example -- WRONG
     lockfile=/tmp/myscript.lock
     if [ -f "$lockfile" ]
     then                      # lock is already held
         echo >&2 "cannot acquire lock, giving up: $lockfile"
         exit 0
     else                      # nobody owns the lock
         > "$lockfile"         # create the file
         #...continue script
     fi

This example does not work, because there is a time window between checking and creating the file. Assume two processes are running the code at the same time. Both check if the lockfile exists, and both get the result that it does not exist. Now both processes assume they have acquired the lock -- a disaster waiting to happen. We need an atomic check-and-create operation, and fortunately there is one: mkdir, the command to create a directory:

  •  # locking example -- CORRECT
     # Bourne
     lockdir=/tmp/myscript.lock
     if mkdir "$lockdir"
     then    # directory did not exist, but was created successfully
         echo >&2 "successfully acquired lock: $lockdir"
         # continue script
     else
         echo >&2 "cannot acquire lock, giving up on $lockdir"
         exit 0
     fi

Here, even when two processes call mkdir at the same time, only one process can succeed at most. This atomicity of check-and-create is ensured at the operating system kernel level.

Instead of using mkdir we could also have used the program to create a symbolic link, ln -s.

Note that we cannot use mkdir -p to automatically create missing path components: mkdir -p does not return an error if the directory exists already, but that's the feature we rely upon to ensure mutual exclusion.

Now let's spice up this example by automatically removing the lock when the script finishes:

  •  # POSIX (maybe Bourne?)
     lockdir=/tmp/myscript.lock
     if mkdir "$lockdir"
     then
         echo >&2 "successfully acquired lock"
    
         # Remove lockdir when the script finishes, or when it receives a signal
         trap 'rm -rf "$lockdir"' 0    # remove directory when script finishes
         trap "exit 2" 1 2 3 15        # terminate script when receiving signal
    
         # Optionally create temporary files in this directory, because
         # they will be removed automatically:
         tmpfile=$lockdir/filelist
    
     else
         echo >&2 "cannot acquire lock, giving up on $lockdir"
         exit 0
     fi

This example is much better. There is still the problem that a stale lock could remain when the script is terminated with a signal not caught (or signal 9, SIGKILL), or could be created by a user (either accidentally or maliciously), but it's a good step towards reliable mutual exclusion. An example that remedies the "stale lock" problem (contributed by Charles Duffy) follows:

  • Are we sure this code's correct? There seems to be a discrepancy between the names LOCK_DEFAULT_NAME and DEFAULT_NAME; and it checks for processes in what looks to be a race condition; and it uses the Linux-specific /proc file system and the GNU-specific egrep -o to do so.... I don't trust it. It looks overly complex and fragile. And quite non-portable. -- GreyCat

     LOCK_DEFAULT_NAME=$0
     LOCK_HOSTNAME="$(hostname -f)"
    
     ## function to take the lock if free; will fail otherwise
     function grab-lock {
       local PROGRAMNAME="${1:-$DEFAULT_NAME}"
       local PID=${2:-$$}
       (
         umask 000;
         mkdir -p "/tmp/${PROGRAMNAME}-lock"
         mkdir "/tmp/${PROGRAMNAME}-lock/held" || return 1
         mkdir "/tmp/${PROGRAMNAME}-lock/held/${LOCK_HOSTNAME}--pid-${PID}" && return 0 || return 1
       ) 2>/dev/null
       return $?
     }
    
     ## function to nicely let go of the lock
     function release-lock {
       local PROGRAMNAME="${1:-$DEFAULT_NAME}"
       local PID=${2:-$$}
       (
         rmdir "/tmp/${PROGRAMNAME}-lock/held/${LOCK_HOSTNAME}--pid-${PID}" || true
         rmdir "/tmp/${PROGRAMNAME}-lock/held" && return 0 || return 1
       ) 2>/dev/null
       return $?
     }
    
     ## function to force anyone else off of the lock
     function break-lock {
       local PROGRAMNAME="${1:-$DEFAULT_NAME}"
       (
         [ -d "/tmp/${PROGRAMNAME}-lock/held" ] || return 0
         for DIR in "/tmp/${PROGRAMNAME}-lock/held/${LOCK_HOSTNAME}--pid-"* ; do
           OTHERPID="$(echo $DIR | egrep -o '[0-9]+$')"
           [ -d /proc/${OTHERPID} ] || rmdir $DIR
         done
         rmdir /tmp/${PROGRAMNAME}-lock/held && return 0 || return 1
       ) 2>/dev/null
       return $?
     }
    
     ## function to take the lock nicely, freeing it first if needed
     function get-lock {
       break-lock "$@" && grab-lock "$@"
     }

Discussion

I believe using if (set -C; >$lockfile); then ... is equally safe if not safer. The Bash source uses open(filename, flags|O_EXCL, mode); which should be atomic on almost all platforms (with the exception of some versions of NFS where mkdir may not be atomic either). I haven't traced the path of the flags variable, which must contain O_CREAT, nor have I looked at any other shells. I wouldn't suggest using this until someone else can backup my claims. --Andy753421

  • Using set -C does not work with ksh88. Ksh88 does not use O_EXCL, when you set noclobber (-C). --jrw32982

    Are you sure mkdir has problems with being atomic on NFS? I thought that affected only open, but I'm not really sure. -- BeJonas 2008-07-24 01:22:59

For more discussion on these issues, see ProcessManagement.

BashFAQ/045 (last edited 2016-07-29 15:44:32 by GreyCat)