Convert namespace_sem to a mutex
From: Bharata B Rao <[email protected]>
namespace_sem is a rwsem. It is acquired as read sem at only one place(used by /proc/mounts, /proc/<pid>/mounts and /proc/<pid>/mountstats). In all other cases it is acquired as a write sem. So, as there is not more than one reader for this sem, this can be a generic sem (and not rwsem) and if so it can be a mutex.
Patch is for 2.6.22-rc1-mm1.
--- fs/namespace.c | 48 ++++++++++++++++++++++++------------------------ 1 files changed, 24 insertions(+), 24 deletions(-)
fs/namespace.c +24 -24
@@ -37,7 +37,7 @@ static int event;static struct list_head *mount_hashtable__read_mostly ;static int hash_mask__read_mostly , hash_bits__read_mostly ;static struct kmem_cache *mnt_cache__read_mostly ; -static struct rw_semaphore namespace_sem; +static struct mutex namespace_mutex;int nr_user_mounts;int max_user_mounts =1024 ;
@@ -396,7 +396,7 @@ static void *m_start(struct seq_file *m,struct list_head *p;loff_t l = *pos; -down_read (&namespace_sem); +mutex_lock (&namespace_mutex);list_for_each (p, &n->list)if (!l--) return list_entry(p, struct vfsmount, mnt_list);
@@ -413,7 +413,7 @@ static void *m_next(struct seq_file *m,static void m_stop (struct seq_file *m,void *v) { -up_read (&namespace_sem); +mutex_unlock (&namespace_mutex); }static inline void mangle (struct seq_file *m,const char *s)
@@ -683,7 +683,7 @@ static int do_umount(struct vfsmount *mnreturn retval; } -down_write (&namespace_sem); +mutex_lock (&namespace_mutex);spin_lock (&vfsmount_lock); event++;
@@ -696,7 +696,7 @@ static int do_umount(struct vfsmount *mnspin_unlock (&vfsmount_lock);if (retval)security_sb_umount_busy (mnt); -up_write (&namespace_sem); +mutex_unlock (&namespace_mutex);release_mounts (&umount_list);return retval; }
@@ -1002,12 +1002,12 @@ static int do_change_type(struct nameidaif (nd->dentry != nd->mnt->mnt_root)return -EINVAL ; -down_write (&namespace_sem); +mutex_lock (&namespace_mutex);spin_lock (&vfsmount_lock);for (m = mnt; m; m = (recurse ?next_mnt (m, mnt) :NULL ))change_mnt_propagation (m, type);spin_unlock (&vfsmount_lock); -up_write (&namespace_sem); +mutex_unlock (&namespace_mutex);return 0 ; }
@@ -1030,7 +1030,7 @@ static int do_loopback(struct nameidataif (err)return err; -down_write (&namespace_sem); +mutex_lock (&namespace_mutex); err = -EINVAL ;if (IS_MNT_UNBINDABLE (old_nd.mnt)) goto out;
@@ -1062,7 +1062,7 @@ static int do_loopback(struct nameidata } out: -up_write (&namespace_sem); +mutex_unlock (&namespace_mutex);path_release (&old_nd);return err; }
@@ -1124,7 +1124,7 @@ static int do_move_mount(struct nameidatif (err)return err; -down_write (&namespace_sem); +mutex_lock (&namespace_mutex);while (d_mountpoint (nd->dentry) &&follow_down (&nd->mnt, &nd->dentry)) ; err = -EINVAL ;
@@ -1176,7 +1176,7 @@ static int do_move_mount(struct nameidat out1:mutex_unlock (&nd->dentry->d_inode->i_mutex); out: -up_write (&namespace_sem); +mutex_unlock (&namespace_mutex);if (!err)path_release (&parent_nd);path_release (&old_nd);
@@ -1238,7 +1238,7 @@ int do_add_mount(struct vfsmount *newmnt {int err; -down_write (&namespace_sem); +mutex_lock (&namespace_mutex);/* Something was mounted here while we slept */ while (d_mountpoint (nd->dentry) &&follow_down (&nd->mnt, &nd->dentry)) ;
@@ -1267,11 +1267,11 @@ int do_add_mount(struct vfsmount *newmntlist_add_tail (&newmnt->mnt_expire, fslist);spin_unlock (&vfsmount_lock); } -up_write (&namespace_sem); +mutex_unlock (&namespace_mutex);return 0 ; unlock: -up_write (&namespace_sem); +mutex_unlock (&namespace_mutex);mntput (newmnt);return err; }
@@ -1337,9 +1337,9 @@ static void expire_mount_list(struct lisget_mnt_ns (ns);spin_unlock (&vfsmount_lock); -down_write (&namespace_sem); +mutex_lock (&namespace_mutex);expire_mount (mnt, mounts, &umounts); -up_write (&namespace_sem); +mutex_unlock (&namespace_mutex);release_mounts (&umounts);mntput (mnt);put_mnt_ns (ns);
@@ -1612,12 +1612,12 @@ static struct mnt_namespace *dup_mnt_ns(init_waitqueue_head (&new_ns->poll); new_ns->event =0 ; -down_write (&namespace_sem); +mutex_lock (&namespace_mutex);/* First pass: copy the tree topology */ new_ns->root =copy_tree (mnt_ns->root, mnt_ns->root->mnt_root,CL_COPY_ALL |CL_EXPIRE ,0 );if (IS_ERR (new_ns->root)) { -up_write (&namespace_sem); +mutex_unlock (&namespace_mutex);kfree (new_ns);return NULL ; }
@@ -1651,7 +1651,7 @@ static struct mnt_namespace *dup_mnt_ns( p =next_mnt (p, mnt_ns->root); q =next_mnt (q, new_ns->root); } -up_write (&namespace_sem); +mutex_unlock (&namespace_mutex);if (rootmnt)mntput (rootmnt);
@@ -1850,7 +1850,7 @@ asmlinkage long sys_pivot_root(const cha user_nd.mnt =mntget (current->fs->rootmnt); user_nd.dentry =dget (current->fs->root);read_unlock (¤t->fs->lock); -down_write (&namespace_sem); +mutex_lock (&namespace_mutex);mutex_lock (&old_nd.dentry->d_inode->i_mutex); error = -EINVAL ; if (IS_MNT_SHARED(old_nd.mnt) ||
@@ -1905,7 +1905,7 @@ asmlinkage long sys_pivot_root(const chapath_release (&parent_nd); out2:mutex_unlock (&old_nd.dentry->d_inode->i_mutex); -up_write (&namespace_sem); +mutex_unlock (&namespace_mutex);path_release (&user_nd);path_release (&old_nd); out1:
@@ -1951,7 +1951,7 @@ void __init mnt_init(unsigned long mempaint i;int err; -init_rwsem (&namespace_sem); +mutex_init (&namespace_mutex); mnt_cache =kmem_cache_create ("mnt_cache" ,sizeof (struct vfsmount ),0 ,SLAB_HWCACHE_ALIGN |SLAB_PANIC ,NULL ,NULL );
@@ -2008,11 +2008,11 @@ void __put_mnt_ns(struct mnt_namespace *LIST_HEAD (umount_list); ns->root =NULL ;spin_unlock (&vfsmount_lock); -down_write (&namespace_sem); +mutex_lock (&namespace_mutex);spin_lock (&vfsmount_lock);umount_tree (root,0 , &umount_list);spin_unlock (&vfsmount_lock); -up_write (&namespace_sem); +mutex_unlock (&namespace_mutex);release_mounts (&umount_list);kfree (ns); }
Quoted 3 lines · Bharata B Rao
From: Bharata B Rao <[email protected]>
namespace_semis a rwsem. It is acquired as read sem at only one place(used
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Actually, this ...
Quoted 2 lines
by
/proc/mounts, /proc/<pid>/mounts and /proc/<pid>/mountstats). In all other cases it is acquired as a write sem. So, as there is not more than one reader
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
... does not mean this. Multiple threads could be reading mounts or mountstats, and otoh mount(2) and umount(2) (acquire it as write sem) could be less frequent?
Quoted 2 lines
for this sem, this can be a generic sem (and not rwsem) and if so it can be a mutex.
Quoted 7 lines · Bharata B Rao
From: Bharata B Rao <[email protected]>
namespace_semis a rwsem. It is acquired as read sem at only oneplace(used by /proc/mounts, /proc/<pid>/mounts and /proc/<pid>/mountstats). In all other cases it is acquired as a write sem. So, as there is not more than one reader for this sem, this can be a generic sem (and not rwsem) and if so it can be a mutex.
Except that read accesses outnumber write ones by far and we have no reason for serializing them against each other. NAK.
Quoted 11 lines · Al Viro
Quoted 7 lines · Bharata B Rao
From: Bharata B Rao <[email protected]>
namespace_semis a rwsem. It is acquired as read sem at only oneplace(used by /proc/mounts, /proc/<pid>/mounts and /proc/<pid>/mountstats). In all other cases it is acquired as a write sem. So, as there is not more than one reader for this sem, this can be a generic sem (and not rwsem) and if so it can be a mutex.Except that read accesses outnumber write ones by far and we have no reason for serializing them against each other. NAK.
Ok, I wasn't aware that this rwsem had many concurrent readers. From whatever little I could see (during system boot and during kernel compilation with make -j), the number of concurrent readers very rarely touched 3 and was 1 mostly. May be there are other concurrent readers of this sem. Any idea from where they could be coming from ?
Regards,
Bharata.
Quoted 5 lines · Bharata B Rao
Ok, I wasn't aware that this rwsem had many concurrent readers. From whatever little I could see (during system boot and during kernel compilation with make -j), the number of concurrent readers very rarely touched 3 and was 1 mostly. May be there are other concurrent readers of this sem. Any idea from where they could be coming from ?
Anybody reading from /proc/mounts? You do realize that on systems where /etc/mtab is a symlink to /proc/mounts every df(1), etc. - anything that wants to find out all mountpoints - will read through that file?
The whole point is that we need exclusion between those who modify mount tree *and* between them and anyone who looks at the mount tree. We don't need any exclusion between several processes that happen to examine the current mount tree at the same time. And since those are far more frequent than those who modify the tree...