ipc, mqueue: lazy call kern_mount_data in new namespaces
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
@@ -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
@@ -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
This is akpm domain.
Quoted 6 lines · Giuseppe Scrivano
kern_mount_datais 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
Signed-off-by Giuseppe Scrivano <[email protected]>---
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@@ -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@@ -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@@ -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
Quoted 6 lines · On Mon, 27 Nov 2017 13:55:50 +0100 Giuseppe Scrivano <[email protected]>…
kern_mount_datais 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?
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
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 anykern_mount_datacost 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.
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 anykern_mount_datacost 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 anykern_mount_data()cost at allAnd.. the patch calls kern_mount_data()->vfs_kern_mount()->...->kmem_cache_zalloc(GFP_KERNEL) under
spin_lock(). This should have created amight_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 peruseDocumentation/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