[CRIU] [PATCH 4/5] restore: restore pocesses which share one fdtable (v3)

[Andrew Vagin]

On Wed, Dec 26, 2012 at 02:21:40PM +0400, Pavel Emelyanov wrote:
> > @@ -585,13 +585,34 @@ static int pstree_wait_helpers()
> >  
> >  static int restore_one_alive_task(int pid, CoreEntry *core)
> >  {
> > +	struct pstree_item *child;
> > +
> >  	pr_info("Restoring resources\n");
> >  
> >  	if (pstree_wait_helpers())
> >  		return -1;
> >  
> > -	if (prepare_fds(current))
> > -		return -1;
> > +	/*
> > +	 * Wait all children, who share a current fd table.
> > +	 * We should be sure, that children don't use any file
> > +	 * descriptor while fdtable is being restored.
> > +	 */
> 
> This leads to deadlock. Currently we have shmem restoring tasks wait
> for the one with the least pid to restore it. If for any reason child
> waits for its parent to restore shmem, and at the same time parent
> waits for this child to finish with this fdt mutex we're stuck.

Summary of a personal discussion:

open_vmas() waits a lock, which released in restorer.
open_vmas is calledd after preprare_fds(), so this deadlock is
impossiable.

But we decided, that fdtable should be restored by the process
with the minimal pid. It should be true for all kind of resources.
> 
> > +	list_for_each_entry(child, &current->children, sibling) {
> > +		if (!shared_fdtable(child))
> > +			continue;
> > +		futex_wait_until(child->rst->fdt_lock, FDT_LOCK_SYNC);
> > +	}
> > +
> > +	if (!shared_fdtable(current)) {
> > +		if (prepare_fds(current))
> > +			return -1;
> > +	} else {
> > +		/* Notify a parent, that a current is ready for restoring fdtable */
> > +		futex_set_and_wake(current->rst->fdt_lock, FDT_LOCK_SYNC);
> > +		futex_wait_until(current->parent->rst->fdt_lock, FDT_LOCK_DONE);
> > +	}
> > +
> > +	futex_set_and_wake(current->rst->fdt_lock, FDT_LOCK_DONE);
> >  
> >  	if (prepare_fs(pid))
> >  		return -1;