* [RFC PATCH 0/1] fs/namespace: defer free_mount from namespace_unlock
@ 2023-01-19 20:55 Eric Chanudet
2023-01-19 20:55 ` [RFC PATCH 1/1] " Eric Chanudet
2023-01-19 21:14 ` [RFC PATCH 0/1] " Eric Chanudet
0 siblings, 2 replies; 4+ messages in thread
From: Eric Chanudet @ 2023-01-19 20:55 UTC (permalink / raw)
To: Alexander Viro
Cc: linux-fsdevel, linux-kernel, Alexander Larsson, Andrew Halaney,
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
^ permalink raw reply [flat|nested] 4+ messages in thread
* [RFC PATCH 1/1] fs/namespace: defer free_mount from namespace_unlock
2023-01-19 20:55 [RFC PATCH 0/1] fs/namespace: defer free_mount from namespace_unlock Eric Chanudet
@ 2023-01-19 20:55 ` Eric Chanudet
2023-01-19 22:03 ` Al Viro
2023-01-19 21:14 ` [RFC PATCH 0/1] " Eric Chanudet
1 sibling, 1 reply; 4+ messages in thread
From: Eric Chanudet @ 2023-01-19 20:55 UTC (permalink / raw)
To: Alexander Viro
Cc: linux-fsdevel, linux-kernel, Alexander Larsson, Andrew Halaney,
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
^ permalink raw reply related [flat|nested] 4+ messages in thread
* Re: [RFC PATCH 0/1] fs/namespace: defer free_mount from namespace_unlock
2023-01-19 20:55 [RFC PATCH 0/1] fs/namespace: defer free_mount from namespace_unlock Eric Chanudet
2023-01-19 20:55 ` [RFC PATCH 1/1] " Eric Chanudet
@ 2023-01-19 21:14 ` Eric Chanudet
1 sibling, 0 replies; 4+ messages in thread
From: Eric Chanudet @ 2023-01-19 21:14 UTC (permalink / raw)
To: Alexander Viro
Cc: linux-fsdevel, linux-kernel, Alexander Larsson, Andrew Halaney
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
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [RFC PATCH 1/1] fs/namespace: defer free_mount from namespace_unlock
2023-01-19 20:55 ` [RFC PATCH 1/1] " Eric Chanudet
@ 2023-01-19 22:03 ` Al Viro
0 siblings, 0 replies; 4+ messages in thread
From: Al Viro @ 2023-01-19 22:03 UTC (permalink / raw)
To: Eric Chanudet
Cc: linux-fsdevel, linux-kernel, Alexander Larsson, Andrew Halaney
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.
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2023-01-19 22:21 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-01-19 20:55 [RFC PATCH 0/1] fs/namespace: defer free_mount from namespace_unlock Eric Chanudet
2023-01-19 20:55 ` [RFC PATCH 1/1] " Eric Chanudet
2023-01-19 22:03 ` Al Viro
2023-01-19 21:14 ` [RFC PATCH 0/1] " Eric Chanudet
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).