vdso_join_timens() question

vdso_join_timens() question

[Christian Brauner]

Hey,

I'm about to finish a patch to add CLONE_NEWTIME support to setns().
Since setns() now allows to attach to a multiple namespaces at the same
time I've also reworked it to be atomic (already upstream). Either all
namespaces are switched or no namespace is switched. All namespaces
basically now have a commit mode after which installation should ideally
not fail anymore. That could work for CLONE_NEWTIME too, I think. The
only blocker to this is vdso_join_timens() which can fail due to
mmap_write_lock_killable().

Is it possible to change this to mmap_write_lock()? So sm like:

diff --git a/arch/x86/entry/vdso/vma.c b/arch/x86/entry/vdso/vma.c
index ea7c1f0b79df..5c5b4cc61fce 100644
--- a/arch/x86/entry/vdso/vma.c
+++ b/arch/x86/entry/vdso/vma.c
@@ -144,8 +144,7 @@ int vdso_join_timens(struct task_struct *task, struct time_namespace *ns)
        struct mm_struct *mm = task->mm;
        struct vm_area_struct *vma;

-       if (mmap_write_lock_killable(mm))
-               return -EINTR;
+       mmap_write_lock(mm);

        for (vma = mm->mmap; vma; vma = vma->vm_next) {
                unsigned long size = vma->vm_end - vma->vm_start;

vdso_join_timens() is called in two places. Once during fork() and once
during timens_install(). I would only need the mmap_write_lock() change
for the latter. So alternatively we could have:

__vdso_join_timens_unlocked()

and then have/expose:

vdso_join_timens_fork()
{
        if (mmap_write_lock_killable(mm))
                return -EINTR;
	__vdso_join_timens_unlocked()
	mmap_write_unlock(mm);
}

and 

vdso_join_timens_install()
{
        mmap_write_lock(mm);
	__vdso_join_timens_unlocked()
	mmap_write_unlock(mm);
}

Thanks!
Christian

Re: vdso_join_timens() question

[Dmitry Safonov]

Hi Christian,

On 6/11/20 12:02 PM, Christian Brauner wrote:
> Hey,
> 
> I'm about to finish a patch to add CLONE_NEWTIME support to setns().
> Since setns() now allows to attach to a multiple namespaces at the same
> time I've also reworked it to be atomic (already upstream). Either all
> namespaces are switched or no namespace is switched. All namespaces
> basically now have a commit mode after which installation should ideally
> not fail anymore. That could work for CLONE_NEWTIME too, I think. The
> only blocker to this is vdso_join_timens() which can fail due to
> mmap_write_lock_killable().
> 
> Is it possible to change this to mmap_write_lock()? So sm like:
> 
> diff --git a/arch/x86/entry/vdso/vma.c b/arch/x86/entry/vdso/vma.c
> index ea7c1f0b79df..5c5b4cc61fce 100644
> --- a/arch/x86/entry/vdso/vma.c
> +++ b/arch/x86/entry/vdso/vma.c
> @@ -144,8 +144,7 @@ int vdso_join_timens(struct task_struct *task, struct time_namespace *ns)
>         struct mm_struct *mm = task->mm;
>         struct vm_area_struct *vma;
> 
> -       if (mmap_write_lock_killable(mm))
> -               return -EINTR;
> +       mmap_write_lock(mm);
> 
>         for (vma = mm->mmap; vma; vma = vma->vm_next) {
>                 unsigned long size = vma->vm_end - vma->vm_start;

I think, it should be fine.

I'm thinking if it actually could be downgraded to mmap_read_lock()..
Probably, write lock was being over-cautious.

> vdso_join_timens() is called in two places. Once during fork() and once
> during timens_install(). I would only need the mmap_write_lock() change
> for the latter. So alternatively we could have:
> 
> __vdso_join_timens_unlocked()
> 
> and then have/expose:
> 
> vdso_join_timens_fork()
> {
>         if (mmap_write_lock_killable(mm))
>                 return -EINTR;
> 	__vdso_join_timens_unlocked()
> 	mmap_write_unlock(mm);
> }
> 
> and 
> 
> vdso_join_timens_install()
> {
>         mmap_write_lock(mm);
> 	__vdso_join_timens_unlocked()
> 	mmap_write_unlock(mm);
> }

I think it's not needed. On fork() it's called on creation of new timens:

:	if (nsproxy->time_ns == nsproxy->time_ns_for_children)
:		return 0;

So the vdso_join_timens() is called on setns() or on creation of new
namespace, which both are quite heavy operations themselves (in sense of
locking).

Thanks,
          Dmitry

Re: vdso_join_timens() question

[Christian Brauner]

On Thu, Jun 11, 2020 at 08:17:12PM +0100, Dmitry Safonov wrote:
> Hi Christian,
> 
> On 6/11/20 12:02 PM, Christian Brauner wrote:
> > Hey,
> > 
> > I'm about to finish a patch to add CLONE_NEWTIME support to setns().
> > Since setns() now allows to attach to a multiple namespaces at the same
> > time I've also reworked it to be atomic (already upstream). Either all
> > namespaces are switched or no namespace is switched. All namespaces
> > basically now have a commit mode after which installation should ideally
> > not fail anymore. That could work for CLONE_NEWTIME too, I think. The
> > only blocker to this is vdso_join_timens() which can fail due to
> > mmap_write_lock_killable().
> > 
> > Is it possible to change this to mmap_write_lock()? So sm like:
> > 
> > diff --git a/arch/x86/entry/vdso/vma.c b/arch/x86/entry/vdso/vma.c
> > index ea7c1f0b79df..5c5b4cc61fce 100644
> > --- a/arch/x86/entry/vdso/vma.c
> > +++ b/arch/x86/entry/vdso/vma.c
> > @@ -144,8 +144,7 @@ int vdso_join_timens(struct task_struct *task, struct time_namespace *ns)
> >         struct mm_struct *mm = task->mm;
> >         struct vm_area_struct *vma;
> > 
> > -       if (mmap_write_lock_killable(mm))
> > -               return -EINTR;
> > +       mmap_write_lock(mm);
> > 
> >         for (vma = mm->mmap; vma; vma = vma->vm_next) {
> >                 unsigned long size = vma->vm_end - vma->vm_start;
> 
> I think, it should be fine.

Great! :)

> 
> I'm thinking if it actually could be downgraded to mmap_read_lock()..
> Probably, write lock was being over-cautious.

Let's use mmap_write_lock() for now and revisit this later, I guess.

> 
> > vdso_join_timens() is called in two places. Once during fork() and once
> > during timens_install(). I would only need the mmap_write_lock() change
> > for the latter. So alternatively we could have:
> > 
> > __vdso_join_timens_unlocked()
> > 
> > and then have/expose:
> > 
> > vdso_join_timens_fork()
> > {
> >         if (mmap_write_lock_killable(mm))
> >                 return -EINTR;
> > 	__vdso_join_timens_unlocked()
> > 	mmap_write_unlock(mm);
> > }
> > 
> > and 
> > 
> > vdso_join_timens_install()
> > {
> >         mmap_write_lock(mm);
> > 	__vdso_join_timens_unlocked()
> > 	mmap_write_unlock(mm);
> > }
> 
> I think it's not needed. On fork() it's called on creation of new timens:
> 
> :	if (nsproxy->time_ns == nsproxy->time_ns_for_children)
> :		return 0;
> 
> So the vdso_join_timens() is called on setns() or on creation of new
> namespace, which both are quite heavy operations themselves (in sense of
> locking).

Ok, great. It'll like be the week after next until I can send the
patchset for setns(<nsfd>/<pidfd>, CLONE_NEWTIME) out.

Thanks!
Christian