Re: fs_struct refcounting: spinlock vs atomic

[Al Viro]

On Thu, Feb 15, 2018 at 02:46:19PM +0100, Enrico Weigelt wrote:
> On 15.02.2018 10:14, Richard Weinberger wrote:
> > On Wed, Feb 14, 2018 at 10:13 PM, Enrico Weigelt <lkml@xxxxxxxxx> wrote:
> > > Hi folks,
> > > 
> > > 
> > > in fork.c, a spinlock is held for fs_struct refcounting, while other
> > > places - eg. switch_task_namespaces uses atomic_dec_and_test() on
> > > the nsproxy.
> > > 
> > > What's the exact difference here ? Could the atomic counting also used
> > > for fs_struct ?
> > 
> > Well, the spinlock protects more than just the counter. So atomic won't do it.
> 
> Okay. Is that needed in that case ?
> 
> See unshare() syscall:
> 
> if (new_fs) {
> 	fs = current->fs;
> 	spin_lock(&fs->lock);
> 	current->fs = new_fs;
> 	if (--fs->users)
> 		new_fs = NULL;
> 	else
> 		new_fs = fs;
> 	spin_unlock(&fs->lock);
> }
> 
> Seems to me, that we're just refcounting here, and once it went dont to
> zero, nobody else can access it anymore.

Not true.  We also assume that once fs_struct has been locked, the number of
tasks with reference to it won't change.  See fs/exec.c:check_unsafe_exec(),
for example.