On Thu, Aug 20, 2020 at 01:29:32PM +0200, Michal Hocko wrote:
> On Thu 20-08-20 13:13:55, Michal Hocko wrote:
> > On Thu 20-08-20 12:55:56, Oleg Nesterov wrote:
> > > On 08/19, Suren Baghdasaryan wrote:
> > > >
> > > > Since the combination of CLONE_VM and !CLONE_SIGHAND is rarely
> > > > used the additional mutex lock in that path of the clone() syscall should
> > > > not affect its overall performance. Clearing the MMF_PROC_SHARED flag
> > > > (when the last process sharing the mm exits) is left out of this patch to
> > > > keep it simple and because it is believed that this threading model is
> > > > rare.
> > >
> > > vfork() ?
> >
> > Could you be more specific?
> >
> > > > --- a/kernel/fork.c
> > > > +++ b/kernel/fork.c
> > > > @@ -1403,6 +1403,15 @@ static int copy_mm(unsigned long clone_flags, struct task_struct *tsk)
> > > > if (clone_flags & CLONE_VM) {
> > > > mmget(oldmm);
> > > > mm = oldmm;
> > > > + if (!(clone_flags & CLONE_SIGHAND)) {
> > >
> > > I agree with Christian, you need CLONE_THREAD
> >
> > This was my suggestion to Suren, likely because I've misrememberd which
> > clone flag is responsible for the signal delivery. But now, after double
> > checking we do explicitly disallow CLONE_SIGHAND && !CLONE_VM. So
> > CLONE_THREAD is the right thing to check.
>
> I have tried to remember but I have to say that after reading man page I
> am still confused. So what is the actual difference between CLONE_THREAD
> and CLONE_SIGHAND? Essentially all we care about from the OOM (and
CLONE_THREAD implies CLONE_SIGHAND
CLONE_SIGHAND implies CLONE_VM but CLONE_SIGHAND doesn't imply CLONE_THREAD.
> oom_score_adj) POV is that signals are delivered to all entities and
> that thay share signal struct. copy_signal is checking for CLONE_THREAD
If a thread has a separate sighand struct it can have separate handlers
(Oleg will correct me if wrong.). But fatal signals will take the whole
thread-group down and can't be ignored which is the only thing you care
about with OOM afair.
What you care about is that the oom_score_adj{_min} settings are shared
and they live in struct signal_struct and whether that's shared or not
is basically guided by CLONE_THREAD.
> but CLONE_THREAD requires CLONE_SIGHAND AFAIU. So is there any cae where
> checking for CLONE_SIGHAND would wrong for our purpose?
Without having spent a long time thinking deeply about this it likely
wouldn't. But using CLONE_SIGHAND is very irritating since it doesn't
clearly express what you want this for. Especially since there's now a
difference between the check in copy_signal() and copy_mm() and a
disconnect to what is expressed in the commit message too, imho.
Christian
