[sheepdog] [PATCH] sheep: implement a correct mutex of base directory

[Hitoshi Mitake]

At Tue, 02 Apr 2013 12:13:23 +0900,
MORITA Kazutaka wrote:
> 
> At Tue,  2 Apr 2013 11:40:48 +0900,
> Hitoshi Mitake wrote:
> > 
> > sheep employes lockf() for mutex of lock_base_dir() now. But the
> > lockf() is not suitable for sheep bacause sheep calls daemon(3) after
> > the lockf(). daemon(3) forks internally and a parent process exits
> > immediately. In a case of sheep, daemon() must be called after locking
> > base dir so the lock owner, parent process, exits and the lock will be
> > released even though the child process is running. This is the reason
> > current lock_base_dir() doesn't work well.
> > 
> > This patch makes lock_base_dir() correct with a new mechanism based on
> > O_EXCL of open(2).
> 
> The lock will not be released if the program exits unexpectedly, no?

Sorry, unlock_base_dir() should also be called in
crash_handler(). I'll fix it in v2.

> 
> init_base_path() was called after daemon() in the old code, and the
> order was changed in f092748b.  I think we should fix the commit.

I think calling daemon() before init_base_path() is not a good
idea. Because after calling daemon(), error messages can be written to
sheep.log only. And this means that multiple sheeps share single
sheep.log file.

Thanks,
Hitoshi