Skip to contentSkip to conversation

ipc, mqueue: lazy call kern_mount_data in new namespaces

RFC PATCH Superseded Giuseppe Scrivano 3 participants 6 messages Nov 27, 2017 – Nov 30, 2017

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 ipc, mqueue: lazy call kern_mount_data in new namespaces

kern_mount_data is a relatively expensive operation when creating a new IPC namespace, so delay the mount until its first usage when not creating the the global namespace.

On my machine, the time for creating 1000 new IPC namespaces dropped from ~8s to ~2s.

---
 include/linux/ipc_namespace.h |  2 +-
 ipc/mqueue.c                  | 47 ++++++++++++++++++++++++++++++++++---------
 ipc/namespace.c               |  2 +-
 3 files changed, 39 insertions(+), 12 deletions(-)
include/linux/ipc_namespace.h +1 -1
index b5630c8eb2f3..a06617c113f1 100644
@@ -81,7 +81,7 @@ static inline void shm_destroy_orphaned(struct ipc_namespace *ns) {}
 #endif /* CONFIG_SYSVIPC */
 
 #ifdef CONFIG_POSIX_MQUEUE
-extern int mq_init_ns(struct ipc_namespace *ns);
+extern int mq_init_ns(struct ipc_namespace *ns, bool mount);
 /*
  * POSIX Message Queue default values:
  *
ipc/mqueue.c +37 -10

96 lines not on this page · raw message

ipc/namespace.c +1 -1
index f59a89966f92..9d3689577f66 100644
@@ -65,7 +65,7 @@ static struct ipc_namespace *create_ipc_ns(struct user_namespace *user_ns,
 	if (err)
 		goto fail_destroy_msg;
 
-	err = mq_init_ns(ns);
+	err = mq_init_ns(ns, false);
 	if (err)
 		goto fail_destroy_shm;
 
Signature · 1 line
2.14.3
#1 lore raw

This is akpm domain.

Quoted 6 lines · Giuseppe Scrivano

kern_mount_data is a relatively expensive operation when creating a new IPC namespace, so delay the mount until its first usage when not creating the the global namespace.

On my machine, the time for creating 1000 new IPC namespaces dropped from ~8s to ~2s.

Nice.

Quoted 143 lines

---

include/linux/ipc_namespace.h |  2 +-
ipc/mqueue.c                  | 47 ++++++++++++++++++++++++++++++++++---------
ipc/namespace.c               |  2 +-
3 files changed, 39 insertions(+), 12 deletions(-)
include/linux/ipc_namespace.h +1 -1
index b5630c8eb2f3..a06617c113f1 100644
@@ -81,7 +81,7 @@ static inline void shm_destroy_orphaned(struct ipc_namespace *ns) {}
 #endif /* CONFIG_SYSVIPC */
 
 #ifdef CONFIG_POSIX_MQUEUE
-extern int mq_init_ns(struct ipc_namespace *ns);
+extern int mq_init_ns(struct ipc_namespace *ns, bool mount);
 /*
 * POSIX Message Queue default values:
 *
ipc/mqueue.c +37 -10
index d24025626310..46d2795bbf93 100644
@@ -87,6 +87,7 @@ struct mqueue_inode_info {
 	unsigned long qsize; /* size of queue in memory (sum of all msgs) */
 };
 
+static struct file_system_type mqueue_fs_type;
 static const struct inode_operations mqueue_dir_inode_operations;
 static const struct file_operations mqueue_file_operations;
 static const struct super_operations mqueue_super_ops;
@@ -776,10 +777,24 @@ static int do_mq_open(const char __user *u_name, int oflag, umode_t mode,
 	struct filename *name;
 	int fd, error;
 	struct ipc_namespace *ipc_ns = current->nsproxy->ipc_ns;
-	struct vfsmount *mnt = ipc_ns->mq_mnt;
-	struct dentry *root = mnt->mnt_root;
+	struct vfsmount *mnt;
+	struct dentry *root;
 	int ro;
 
+	spin_lock(&mq_lock);
+	mnt = ipc_ns->mq_mnt;
+	if (unlikely(!mnt)) {
+		mnt = kern_mount_data(&mqueue_fs_type, ipc_ns);
+		if (IS_ERR(mnt)) {
+			spin_unlock(&mq_lock);
+			return PTR_ERR(mnt);
+		}
+		ipc_ns->mq_mnt = mnt;
+	}
+	spin_unlock(&mq_lock);
+
+	root = mnt->mnt_root;
+
 	audit_mq_open(oflag, mode, attr);
 
 	if (IS_ERR(name = getname(u_name)))
@@ -863,6 +878,9 @@ SYSCALL_DEFINE1(mq_unlink, const char __user *, u_name)
 	struct ipc_namespace *ipc_ns = current->nsproxy->ipc_ns;
 	struct vfsmount *mnt = ipc_ns->mq_mnt;
 
+	if (!mnt)
+		return -ENOENT;
+
 	name = getname(u_name);
 	if (IS_ERR(name))
 		return PTR_ERR(name);
@@ -1581,7 +1599,8 @@ static struct file_system_type mqueue_fs_type = {
 	.fs_flags = FS_USERNS_MOUNT,
 };
 
-int mq_init_ns(struct ipc_namespace *ns)
+
+int mq_init_ns(struct ipc_namespace *ns, bool mount)
 {
 	ns->mq_queues_count  = 0;
 	ns->mq_queues_max    = DFLT_QUEUESMAX;
@@ -1590,23 +1609,31 @@ int mq_init_ns(struct ipc_namespace *ns)
 	ns->mq_msg_default   = DFLT_MSG;
 	ns->mq_msgsize_default  = DFLT_MSGSIZE;
 
-	ns->mq_mnt = kern_mount_data(&mqueue_fs_type, ns);
-	if (IS_ERR(ns->mq_mnt)) {
-		int err = PTR_ERR(ns->mq_mnt);
+	if (!mount)
 		ns->mq_mnt = NULL;
-		return err;
+	else {
+		int err;
+
+		ns->mq_mnt = kern_mount_data(&mqueue_fs_type, ns);
+		if (IS_ERR(ns->mq_mnt)) {
+			err = PTR_ERR(ns->mq_mnt);
+			ns->mq_mnt = NULL;
+			return err;
+		}
 	}
 	return 0;
 }
 
 void mq_clear_sbinfo(struct ipc_namespace *ns)
 {
-	ns->mq_mnt->mnt_sb->s_fs_info = NULL;
+	if (ns->mq_mnt)
+		ns->mq_mnt->mnt_sb->s_fs_info = NULL;
 }
 
 void mq_put_mnt(struct ipc_namespace *ns)
 {
-	kern_unmount(ns->mq_mnt);
+	if (ns->mq_mnt)
+		kern_unmount(ns->mq_mnt);
 }
 
 static int __init init_mqueue_fs(void)
@@ -1628,7 +1655,7 @@ static int __init init_mqueue_fs(void)
 
 	spin_lock_init(&mq_lock);
 
-	error = mq_init_ns(&init_ipc_ns);
+	error = mq_init_ns(&init_ipc_ns, true);
 	if (error)
 		goto out_filesystem;
 
ipc/namespace.c +1 -1
index f59a89966f92..9d3689577f66 100644
@@ -65,7 +65,7 @@ static struct ipc_namespace *create_ipc_ns(struct user_namespace *user_ns,
 	if (err)
 		goto fail_destroy_msg;
 
-	err = mq_init_ns(ns);
+	err = mq_init_ns(ns, false);
 	if (err)
 		goto fail_destroy_shm;
 
Signature · 1 line
2.14.3
#2 lore raw
In reply to Giuseppe Scrivano
Quoted 6 lines · On Mon, 27 Nov 2017 13:55:50 +0100 Giuseppe Scrivano <[email protected]>…

kern_mount_data is a relatively expensive operation when creating a new IPC namespace, so delay the mount until its first usage when not creating the the global namespace.

On my machine, the time for creating 1000 new IPC namespaces dropped from ~8s to ~2s.

OK, but this simply moves the expense so it happens later on. Why is that better?

#3 lore raw
Quoted 2 lines · Andrew Morton

OK, but this simply moves the expense so it happens later on. Why is that better?

the optimization is for new IPC namespaces that don't use mq_open. In this case there won't be any kern_mount_data cost at all.

Regards,
Giuseppe

#4 lore raw
Quoted 8 lines · On Wed, 29 Nov 2017 11:33:28 +0100 Giuseppe Scrivano <[email protected]>…
Quoted 2 lines · Andrew Morton

OK, but this simply moves the expense so it happens later on. Why is that better?

the optimization is for new IPC namespaces that don't use mq_open. In this case there won't be any kern_mount_data cost at all.

Fair enough. Please add this paragraph (or similar) to the changelog:

: This is a net saving for new IPC namespaces that don't use mq_open(). In : this case there won't be any kern_mount_data() cost at all

And.. the patch calls kern_mount_data()->vfs_kern_mount()->...->kmem_cache_zalloc(GFP_KERNEL) under spin_lock(). This should have created a might_sleep() warning in your testing, but obviously did not.

Could you please find out why? Do you have CONFIG_DEBUG_ATOMIC_SLEEP=n, I hope? Please peruse Documentation/process/submit-checklist.rst, section 12...

I assume a suitable fix would be to create a new mutex (static to do_mq_open()) to prevent concurrent mounting.

#5 lore raw
Quoted 21 lines · Andrew Morton
Quoted 8 lines · On Wed, 29 Nov 2017 11:33:28 +0100 Giuseppe Scrivano <[email protected]>…
Quoted 2 lines · Andrew Morton

OK, but this simply moves the expense so it happens later on. Why is that better?

the optimization is for new IPC namespaces that don't use mq_open. In this case there won't be any kern_mount_data cost at all.

Fair enough. Please add this paragraph (or similar) to the changelog:

: This is a net saving for new IPC namespaces that don't use mq_open(). In : this case there won't be any kern_mount_data() cost at all

And.. the patch calls kern_mount_data()->vfs_kern_mount()->...->kmem_cache_zalloc(GFP_KERNEL) under spin_lock(). This should have created a might_sleep() warning in your testing, but obviously did not.

Quoted 6 lines

Could you please find out why? Do you have CONFIG_DEBUG_ATOMIC_SLEEP=n, I hope? Please peruse Documentation/process/submit-checklist.rst, section 12...

I assume a suitable fix would be to create a new mutex (static to do_mq_open()) to prevent concurrent mounting.

thanks for the hints.

Indeed, that was a mistake on my side as I didn't use CONFIG_DEBUG_ATOMIC_SLEEP=y. The might_sleep() warning is correctly raised once I enable CONFIG_DEBUG_ATOMIC_SLEEP (and the other options suggested in submit-checklist.rst).

Giuseppe