Re: spinlock contention of files->file_lock

[Ingo Molnar <mingo@kernel.org>]

* Al Viro <viro@ZenIV.linux.org.uk> wrote:

> On Wed, Oct 02, 2013 at 07:13:19AM +0200, Ingo Molnar wrote:
> > 
> > * Al Viro <viro@ZenIV.linux.org.uk> wrote:
> > 
> > > On Tue, Oct 01, 2013 at 02:41:58PM -0700, Eric Dumazet wrote:
> > > > Maybe I am missing something obvious ?
> > > 
> > > Yes.  do_execve_common() starts with unshare_files(); there can be
> > > no other thread capable of modifying that descriptor table.
> > 
> > Btw., might the Android Binder:
> > 
> >   drivers/staging/android/binder.c:       struct files_struct *files = proc->files;
> > ...
> >   drivers/staging/android/binder.c:               __fd_install(proc->files, fd, file);
> > ...
> >   drivers/staging/android/binder.c:       retval = __close_fd(proc->files, fd);
> > 
> > violate that assumption?
> 
> Not unless your thread has managed to call an ioctl between entering
> do_execve_common() and calling do_close_on_exec() ;-)

Indeed - while the binder interface appears to allow the insertion of fds 
into other task's file tables, it refcounts its task->files access and 
only ever receives it via get_files_struct(current), so it cannot possibly 
interfere with a private file table resulting from unshare_files().

Thanks,

	Ingo