* vdso_join_timens() question
@ 2020-06-11 11:02 Christian Brauner
2020-06-11 19:17 ` Dmitry Safonov
0 siblings, 1 reply; 3+ messages in thread
From: Christian Brauner @ 2020-06-11 11:02 UTC (permalink / raw)
To: Andrei Vagin, Dmitry Safonov, Thomas Gleixner; +Cc: linux-kernel
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
^ permalink raw reply related [flat|nested] 3+ messages in thread
* Re: vdso_join_timens() question
2020-06-11 11:02 vdso_join_timens() question Christian Brauner
@ 2020-06-11 19:17 ` Dmitry Safonov
2020-06-12 14:47 ` Christian Brauner
0 siblings, 1 reply; 3+ messages in thread
From: Dmitry Safonov @ 2020-06-11 19:17 UTC (permalink / raw)
To: Christian Brauner, Andrei Vagin, Dmitry Safonov, Thomas Gleixner
Cc: linux-kernel
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
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: vdso_join_timens() question
2020-06-11 19:17 ` Dmitry Safonov
@ 2020-06-12 14:47 ` Christian Brauner
0 siblings, 0 replies; 3+ messages in thread
From: Christian Brauner @ 2020-06-12 14:47 UTC (permalink / raw)
To: Dmitry Safonov; +Cc: Andrei Vagin, Thomas Gleixner, linux-kernel
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
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2020-06-12 14:47 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-06-11 11:02 vdso_join_timens() question Christian Brauner
2020-06-11 19:17 ` Dmitry Safonov
2020-06-12 14:47 ` Christian Brauner
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).