Motivation

Libvirt metadata locking
- chown(2), setfilecon(3)
- XATTRs to store original owner/context
- Not atomic, need to mutually exclude with other daemons

DB locking files on update
- Lock only those parts of file that need updating

PID files

Current state

Lock types

Advisory
- BSD locks
- POSIX lock
- Open File Description locks

Mandatory

BSD locks

int fd = open("/some/path", O_RDWR);
flock(fd, LOCK_EX);

Associated with file description
Shared or exclusive lock

Except:

Always lock the entire file
Not POSIX (even though widely available)
No atomic lock promotion
On older Linux NOP returning success
On newer Linux silently converted to POSIX lock for NFS

POSIX locks

int fd = open("/some/path", O_RDWR);
struct flock fl = {
    .l_type = F_WRLCK,
    .l_whence = SEEK_SET,
    .l_start = 1,
    .l_len = 2,
};

fcntl(fd, F_SETLKW, &fl);

Associated with [i-node, pid]
Ranges
Kernel detects & prevents deadlocks
Read or write lock
Works across NFS

POSIX locks

Except:

Not shared across fork(2)
Doesn't work on SMB
close(2) releases lock
- The lock acquired through one FD might be released by another thread through different FD
- Workaround: reimplement open(2) & close(2)
- Not possible if linking with some other library (/etc/passwd, getpwent(3))
- Hardlinks
struct flock

POSIX locks

int fd = open("/some/path", O_RDWR);
lockf(fd, F_LOCK, len);

Just a fancy wrapper over fcntl(fd, F_SETLK, ...)

Open File Description locks

int fd = open("/some/path", O_RDWR);
struct flock fl = {
    .l_type = F_WRLCK,
    .l_whence = SEEK_SET,
    .l_start = 1,
    .l_len = 2,
};

fcntl(fd, F_OFD_SETLK, &fl);

Associated with file description
Ranges
Kernel detects & prevents deadlocks
Read or write lock
Works across NFS
Safe around close(2)

Mandatory locking

Do not use it!

Really, don't.

Solution?

Libvirt
- POSIX locks
- fork(2) & auditing

sqlite
- POSIX locks
- Reimplemented open, close & pthread locks

PID files
- BSD locks

Thank You!

mprivozn@redhat.com