[RFC PATCH 0/1] fs/namespace: defer free_mount from namespace_unlock

[RFC PATCH 0/1] fs/namespace: defer free_mount from namespace_unlock

[Eric Chanudet]

We noticed a significant slow down when running a containers on an
Aarch64 system with the RT patch set using the following test:
# mkdir -p "rootfs/bin"
# printf "int main(){return 0;}" | gcc  -x c - -static -o rootfs/bin/sh
# crun spec
# perf stat -r 10 --table --null -- crun run test

 Performance counter stats for 'crun run test' (10 runs):

            # Table of individual measurements:
            0.3902 (-0.1941) ##########
            0.5791 (-0.0053) #
            0.5785 (-0.0058) #
            0.5891 (+0.0047) #
            0.6682 (+0.0839) ###
            0.5507 (-0.0337) ##
            0.5888 (+0.0044) #
            0.5797 (-0.0047) #
            0.5977 (+0.0133) #
            0.7217 (+0.1374) ####

            # Final result:
            0.5844 +- 0.0269 seconds time elapsed  ( +-  4.60% )

A 6.2 non-RT kernel results on the same hardware for comparison:

 Performance counter stats for 'crun run test' (10 runs):

            # Table of individual measurements:
            0.1680 (+0.1375) #################
            0.0074 (-0.0231) ###############################################################
            0.0073 (-0.0232) ################################################################
            0.0070 (-0.0235) ####################################################################
            0.0072 (-0.0233) #################################################################
            0.0794 (+0.0489) #############
            0.0078 (-0.0227) ##########################################################
            0.0070 (-0.0235) ###################################################################
            0.0070 (-0.0235) ####################################################################
            0.0068 (-0.0237) ######################################################################

            # Final result:
            0.0305 +- 0.0169 seconds time elapsed  ( +- 55.33% )

It looks like there is one bottleneck in:
-> do_umount 
 -> namespace_unlock
  -> synchronize_rcu_expedited

With the following patch, namespace_unlock will queue up the resources
that needs to be released and defer the operation through call_rcu to
return without waiting for the grace period.

Alexander Larsson (1):
  fs/namespace: defer free_mount from namespace_unlock

 fs/namespace.c | 42 +++++++++++++++++++++++++++++++++++-------
 1 file changed, 35 insertions(+), 7 deletions(-)

-- 
2.39.0

[RFC PATCH 1/1] fs/namespace: defer free_mount from namespace_unlock

[Eric Chanudet]

From: Alexander Larsson <alexl@redhat.com>

Use call_rcu to defer releasing the umount'ed or detached filesystem
when calling namepsace_unlock().

Calling synchronize_rcu_expedited() has a significant cost on RT kernel
that default to rcupdate.rcu_normal_after_boot=1.

For example, on a 6.2-rt1 kernel:
perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs mnt' -- umount mnt
           0.07464 +- 0.00396 seconds time elapsed  ( +-  5.31% )

With this change applied:
perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs mnt' -- umount mnt
        0.00162604 +- 0.00000637 seconds time elapsed  ( +-  0.39% )

Waiting for the grace period before completing the syscall does not seem
mandatory. The struct mount umount'ed are queued up for release in a
separate list and no longer accessible to following syscalls.

Signed-off-by: Alexander Larsson <alexl@redhat.com>
Signed-off-by: Eric Chanudet <echanude@redhat.com>
---
 fs/namespace.c | 42 +++++++++++++++++++++++++++++++++++-------
 1 file changed, 35 insertions(+), 7 deletions(-)

diff --git a/fs/namespace.c b/fs/namespace.c
index ab467ee58341..11d219a6e83c 100644
--- a/fs/namespace.c
+++ b/fs/namespace.c
@@ -44,6 +44,11 @@ static unsigned int m_hash_shift __read_mostly;
 static unsigned int mp_hash_mask __read_mostly;
 static unsigned int mp_hash_shift __read_mostly;
 
+struct mount_delayed_release {
+	struct rcu_head rcu;
+	struct hlist_head release_list;
+};
+
 static __initdata unsigned long mhash_entries;
 static int __init set_mhash_entries(char *str)
 {
@@ -1582,11 +1587,31 @@ int may_umount(struct vfsmount *mnt)
 
 EXPORT_SYMBOL(may_umount);
 
-static void namespace_unlock(void)
+static void free_mounts(struct hlist_head *mount_list)
 {
-	struct hlist_head head;
 	struct hlist_node *p;
 	struct mount *m;
+
+	hlist_for_each_entry_safe(m, p, mount_list, mnt_umount) {
+		hlist_del(&m->mnt_umount);
+		mntput(&m->mnt);
+	}
+}
+
+static void delayed_mount_release(struct rcu_head *head)
+{
+	struct mount_delayed_release *drelease =
+		container_of(head, struct mount_delayed_release, rcu);
+
+	free_mounts(&drelease->release_list);
+	kfree(drelease);
+}
+
+static void namespace_unlock(void)
+{
+	struct hlist_head head;
+	struct mount_delayed_release *drelease;
+
 	LIST_HEAD(list);
 
 	hlist_move_list(&unmounted, &head);
@@ -1599,12 +1624,15 @@ static void namespace_unlock(void)
 	if (likely(hlist_empty(&head)))
 		return;
 
-	synchronize_rcu_expedited();
-
-	hlist_for_each_entry_safe(m, p, &head, mnt_umount) {
-		hlist_del(&m->mnt_umount);
-		mntput(&m->mnt);
+	drelease = kmalloc(sizeof(*drelease), GFP_KERNEL);
+	if (unlikely(!drelease)) {
+		synchronize_rcu_expedited();
+		free_mounts(&head);
+		return;
 	}
+
+	hlist_move_list(&head, &drelease->release_list);
+	call_rcu(&drelease->rcu, delayed_mount_release);
 }
 
 static inline void namespace_lock(void)
-- 
2.39.0

Re: [RFC PATCH 1/1] fs/namespace: defer free_mount from namespace_unlock

[Al Viro]

On Thu, Jan 19, 2023 at 03:55:21PM -0500, Eric Chanudet wrote:
> From: Alexander Larsson <alexl@redhat.com>
> 
> Use call_rcu to defer releasing the umount'ed or detached filesystem
> when calling namepsace_unlock().
> 
> Calling synchronize_rcu_expedited() has a significant cost on RT kernel
> that default to rcupdate.rcu_normal_after_boot=1.
> 
> For example, on a 6.2-rt1 kernel:
> perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs mnt' -- umount mnt
>            0.07464 +- 0.00396 seconds time elapsed  ( +-  5.31% )
> 
> With this change applied:
> perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs mnt' -- umount mnt
>         0.00162604 +- 0.00000637 seconds time elapsed  ( +-  0.39% )
> 
> Waiting for the grace period before completing the syscall does not seem
> mandatory. The struct mount umount'ed are queued up for release in a
> separate list and no longer accessible to following syscalls.

You *really* do not want to have umount(2) return without having
the filesystems shut down.

Re: [RFC PATCH 0/1] fs/namespace: defer free_mount from namespace_unlock

[Eric Chanudet]

On Thu, Jan 19, 2023 at 03:55:20PM -0500, Eric Chanudet wrote:
> We noticed a significant slow down when running a containers on an
> Aarch64 system with the RT patch set using the following test:
> [...]
> With the following patch, namespace_unlock will queue up the resources
> that needs to be released and defer the operation through call_rcu to
> return without waiting for the grace period.

I did not CC:linux-rt-users. Resending with them included, my apologies.

-- 
Eric Chanudet