On Fri, Apr 18, 2025 at 09:59:23PM +0200, Christian Brauner wrote:
> On Fri, Apr 18, 2025 at 10:47:10AM +0200, Christian Brauner wrote:
> > On Fri, Apr 18, 2025 at 09:20:52AM +0800, Ian Kent wrote:
> > > > On 18/4/25 00:28, Christian Brauner wrote:
> > > > > I'm not sure how much that would buy us. If it doesn't then it should be
> > > > > possible to play with the following possibly strange idea:
> > > > >
> > > > > diff --git a/fs/mount.h b/fs/mount.h
> > > > > index 7aecf2a60472..51b86300dc50 100644
> > > > > --- a/fs/mount.h
> > > > > +++ b/fs/mount.h
> > > > > @@ -61,6 +61,7 @@ struct mount {
> > > > > struct rb_node mnt_node; /* node in the ns->mounts
> > > > > rbtree */
> > > > > struct rcu_head mnt_rcu;
> > > > > struct llist_node mnt_llist;
> > > > > + unsigned long mnt_rcu_unmount_seq;
> > > > > };
> > > > > #ifdef CONFIG_SMP
> > > > > struct mnt_pcp __percpu *mnt_pcp;
> > > > > diff --git a/fs/namespace.c b/fs/namespace.c
> > > > > index d9ca80dcc544..aae9df75beed 100644
> > > > > --- a/fs/namespace.c
> > > > > +++ b/fs/namespace.c
> > > > > @@ -1794,6 +1794,7 @@ static void namespace_unlock(void)
> > > > > struct hlist_head head;
> > > > > struct hlist_node *p;
> > > > > struct mount *m;
> > > > > + bool needs_synchronize_rcu = false;
> > > > > LIST_HEAD(list);
> > > > >
> > > > > hlist_move_list(&unmounted, &head);
> > > > > @@ -1817,7 +1818,16 @@ static void namespace_unlock(void)
> > > > > if (likely(hlist_empty(&head)))
> > > > > return;
> > > > >
> > > > > - synchronize_rcu_expedited();
> > > > > + hlist_for_each_entry_safe(m, p, &head, mnt_umount) {
> > > > > + if (!poll_state_synchronize_rcu(m->mnt_rcu_unmount_seq))
> > > > > + continue;
> >
> > This has a bug. This needs to be:
> >
> > /* A grace period has already elapsed. */
> > if (poll_state_synchronize_rcu(m->mnt_rcu_unmount_seq))
> > continue;
> >
> > /* Oh oh, we have to pay up. */
> > needs_synchronize_rcu = true;
> > break;
> >
> > which I'm pretty sure will eradicate most of the performance gain you've
> > seen because fundamentally the two version shouldn't be different (Note,
> > I drafted this while on my way out the door. r.
I failed to notice... Once corrected the numbers are back to that of
mainline, same as with the other version.
> > I would test the following version where we pay the cost of the
> > smb_mb() from poll_state_synchronize_rcu() exactly one time:
> >
> > diff --git a/fs/mount.h b/fs/mount.h
> > index 7aecf2a60472..51b86300dc50 100644
> > --- a/fs/mount.h
> > +++ b/fs/mount.h
> > @@ -61,6 +61,7 @@ struct mount {
> > struct rb_node mnt_node; /* node in the ns->mounts rbtree */
> > struct rcu_head mnt_rcu;
> > struct llist_node mnt_llist;
> > + unsigned long mnt_rcu_unmount_seq;
> > };
> > #ifdef CONFIG_SMP
> > struct mnt_pcp __percpu *mnt_pcp;
> > diff --git a/fs/namespace.c b/fs/namespace.c
> > index d9ca80dcc544..dd367c54bc29 100644
> > --- a/fs/namespace.c
> > +++ b/fs/namespace.c
> > @@ -1794,6 +1794,7 @@ static void namespace_unlock(void)
> > struct hlist_head head;
> > struct hlist_node *p;
> > struct mount *m;
> > + unsigned long mnt_rcu_unmount_seq = 0;
> > LIST_HEAD(list);
> >
> > hlist_move_list(&unmounted, &head);
> > @@ -1817,7 +1818,10 @@ static void namespace_unlock(void)
> > if (likely(hlist_empty(&head)))
> > return;
> >
> > - synchronize_rcu_expedited();
> > + hlist_for_each_entry_safe(m, p, &head, mnt_umount)
> > + mnt_rcu_unmount_seq = max(m->mnt_rcu_unmount_seq, mnt_rcu_unmount_seq);
> > +
> > + cond_synchronize_rcu_expedited(mnt_rcu_unmount_seq);
> >
> > hlist_for_each_entry_safe(m, p, &head, mnt_umount) {
> > hlist_del(&m->mnt_umount);
> > @@ -1923,8 +1927,10 @@ static void umount_tree(struct mount *mnt, enum umount_tree_flags how)
> > }
> > }
> > change_mnt_propagation(p, MS_PRIVATE);
> > - if (disconnect)
> > + if (disconnect) {
> > + p->mnt_rcu_unmount_seq = get_state_synchronize_rcu();
> > hlist_add_head(&p->mnt_umount, &unmounted);
> > + }
> >
> > /*
> > * At this point p->mnt_ns is NULL, notification will be queued
That did not show improved performances.
ftrace confirms time is spent in synchronize_rcu as it used to, e.g:
5) | namespace_unlock() {
5) | cond_synchronize_rcu_expedited() {
5) | synchronize_rcu_expedited() {
5) 0.088 us | rcu_gp_is_normal();
5) | synchronize_rcu_normal() {
5) * 15797.30 us | }
5) * 15797.70 us | }
5) * 15797.94 us | }
5) * 15856.13 us | }
(rcu_expedited=1 was mentioned upthread iirc, PREEMPT_RT discards it,
the trace above was with it:
# cat /sys/kernel/rcu_expedited
1)
0001-UNTESTED-fs-namespace-defer-RCU-sync-for-MNT_DETACH-.patch yields
the expected improved performances, although I'm still seeing the
corruption reported earlier that doesn't happen on mainline.
> I'm appending a patch that improves on the first version of this patch.
> Instead of simply sampling the current rcu state and hoping that the rcu
> grace period has elapsed by the time we get to put the mounts we sample
> the rcu state and kick off a new grace period at the end of
> umount_tree(). That could even get us some performance improvement by on
> non-RT kernels. I have no clue how well this will fare on RT though.
0001-UNTESTED-mount-sample-and-kick-of-grace-period-durin.patch does not
show the expected performance improvements:
# perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs /mnt' -- umount /mnt
0.022602 +- 0.000757 seconds time elapsed ( +- 3.35% )
# perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs /mnt' -- umount -l /mnt
0.02145 +- 0.00130 seconds time elapsed ( +- 6.05% )
Best,
--
Eric Chanudet
