Mandatory file locking semantics (was: sqlite3: bug with monotone)

[Warren Young]

On 6/4/2013 02:48, Corinna Vinschen wrote:
> On Jun  3 11:36, Warren Young wrote:
>> Could you add an O_MAND open(2) option as well to turn on the same
>> feature?  This will avoid a race condition.
>
> If you call F_LCK_MANDATORY right after open in the opening routine,
> there won't be (much of) a race.  I know what you mean (it's the
> O_CLOEXEC dilemma)

As long as everyone plays by the rules, there actually isn't a race.

For this style of locking to work at all, every program involved has to 
enable the mandatory locking flag before doing any I/O on the file.  In 
that sense, locking is already atomic.

Adding O_MAND changes the call pattern from:

    fd = open(..., O_RDWR);
    fcntl(fd, F_LCK_MANDATORY, 1);
    fcntl(fd, F_SETLK, &lock);

to:

    fd = open(..., O_RDWR | O_MAND);
    fcntl(fd, F_SETLK, &lock);

Since there are two required calls due to the POSIX design, there's 
still a hole here, where an uncooperative program could call open(2) for 
writing on the same file without acquiring the lock.  To fix *that* race 
condition, you'd need an extension to open(2) that lets it acquire the 
lock on open.  That just isn't in the cards.

Or am I missing something?  Does mandatory locking on Linux or SysV 
somehow avoid the "uncooperative program" race?




That having been said, I still think O_MAND works better.

Cygwin's mandatory locking is effectively a flag on the file that, once 
set, stays with the file descriptor as long as it is open.

You can later unset the flag, but under what condition would that ever 
be the right thing?  That effectively tries to mix locking semantics on 
the same file descriptor!

The F_LCK_MANDATORY design also allows you to do arbitrary I/O between 
open() and setting the flag, another thing which can never be correct. 
It violates the rule laid out above that allows two Cygwin programs to 
cooperatively enable mandatory file locking "atomically".

If a Cygwin program wants mandatory locking, it should be enabled 
continuously from open() to close().



If those weren't enough reasons, it now feels weird that you use fcntl() 
to set the locking flag even when you're going to use flock() or lockf() 
to do the actual locking.

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple