Re: flock still buggy

[Corinna Vinschen <corinna-cygwin at cygwin dot com>]

On Jul 21 23:26, Eric Blake wrote:
> I finally figured out why autoconf is still failing its flock-related tests, 
> and why perl was reliably failing even though my simple attempts in C were 
> always passing.  It turns out that if you do:
> 
> open
> flock(LOCK_EX)
> if (!fork)
>   execlp("sleep","sleep","10",NULL);
> sleep(10);
> 
> then ProcessExplorer shows that the Event in the global namespace of flock-dev-
> ino\20-2-* exists in both parent and child, and with a notification level of 
> false, blocking any outside influence until both the parent and forkee exit.  
> But if you do:
> 
> open
> fcntl (fd, F_SETFD, FD_CLOEXEC | fcntl (fd, F_GETFD))
> flock(LOCK_EX)
> if (!fork)
>   execlp("sleep","sleep","10",NULL);
> sleep(10);
> 
> then only the parent holds a handle to the Event, but with a notification level 
> of true, allowing any outside party to do whatever they want.

Do you have a working C testcase to demonstrate this?

> I'm still trying to figure out why the close-on-exec cleanup appears to be 
> spuriously triggering the flock Event to unlock.  But my understanding is that 
> F_FLOCK locks should survive over exec, so the close-on-exec cleanup should 
> only trigger lock release on F_POSIX locks.

I have a hunch it's a thinko in fhandler_base::del_my_locks.  In case of
close_on_exec, the underlying file handle is already invalid.

Here's a question:  If you strace this, do you get a debug message
from get_obj_handle_count with a status code from NtQueryObject?

Does the below patch fix the problem?  Does the reasoning sound...
reasonable?

Index: flock.cc
===================================================================
RCS file: /cvs/src/src/winsup/cygwin/flock.cc,v
retrieving revision 1.23
diff -u -p -r1.23 flock.cc
--- flock.cc	14 Jul 2009 17:37:42 -0000	1.23
+++ flock.cc	22 Jul 2009 10:07:17 -0000
@@ -350,8 +350,18 @@ fhandler_base::del_my_locks (bool after_
   inode_t *node = inode_t::get (get_dev (), get_ino (), false);
   if (node)
     {
+      /* In the close-on-exec case, our io handle is already invalid.
+         We can't use it to test for the object reference count.
+	 However, that shouldn't be necessary for the following reason.
+	 After exec, there are no threads in the current process
+	 waiting for the lock.  So, either we're the only process
+	 accessing the file table entry and there are no threads
+	 which require signalling, or we have a parent process still
+	 accessing the file object and signalling the lock event would
+	 be premature. */
       bool no_locks_left =
-	node->del_my_locks (after_fork ? 0 : get_unique_id (), get_handle ());
+	node->del_my_locks (after_fork ? 0 : get_unique_id (),
+			    close_on_exec () ? NULL : get_handle ());
       if (no_locks_left)
 	{
 	  LIST_REMOVE (node, i_next);


Corinna

-- 
Corinna Vinschen                  Please, send mails regarding Cygwin to
Cygwin Project Co-Leader          cygwin AT cygwin DOT com
Red Hat

--
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