Push BKL into do_mount()
[linux-2.6.git] / fs / namespace.c
index a40685d..4740f7b 100644 (file)
@@ -27,6 +27,7 @@
 #include <linux/ramfs.h>
 #include <linux/log2.h>
 #include <linux/idr.h>
+#include <linux/fs_struct.h>
 #include <asm/uaccess.h>
 #include <asm/unistd.h>
 #include "pnode.h"
@@ -130,10 +131,20 @@ struct vfsmount *alloc_vfsmnt(const char *name)
                INIT_LIST_HEAD(&mnt->mnt_share);
                INIT_LIST_HEAD(&mnt->mnt_slave_list);
                INIT_LIST_HEAD(&mnt->mnt_slave);
-               atomic_set(&mnt->__mnt_writers, 0);
+#ifdef CONFIG_SMP
+               mnt->mnt_writers = alloc_percpu(int);
+               if (!mnt->mnt_writers)
+                       goto out_free_devname;
+#else
+               mnt->mnt_writers = 0;
+#endif
        }
        return mnt;
 
+#ifdef CONFIG_SMP
+out_free_devname:
+       kfree(mnt->mnt_devname);
+#endif
 out_free_id:
        mnt_free_id(mnt);
 out_free_cache:
@@ -170,65 +181,38 @@ int __mnt_is_readonly(struct vfsmount *mnt)
 }
 EXPORT_SYMBOL_GPL(__mnt_is_readonly);
 
-struct mnt_writer {
-       /*
-        * If holding multiple instances of this lock, they
-        * must be ordered by cpu number.
-        */
-       spinlock_t lock;
-       struct lock_class_key lock_class; /* compiles out with !lockdep */
-       unsigned long count;
-       struct vfsmount *mnt;
-} ____cacheline_aligned_in_smp;
-static DEFINE_PER_CPU(struct mnt_writer, mnt_writers);
+static inline void inc_mnt_writers(struct vfsmount *mnt)
+{
+#ifdef CONFIG_SMP
+       (*per_cpu_ptr(mnt->mnt_writers, smp_processor_id()))++;
+#else
+       mnt->mnt_writers++;
+#endif
+}
 
-static int __init init_mnt_writers(void)
+static inline void dec_mnt_writers(struct vfsmount *mnt)
 {
-       int cpu;
-       for_each_possible_cpu(cpu) {
-               struct mnt_writer *writer = &per_cpu(mnt_writers, cpu);
-               spin_lock_init(&writer->lock);
-               lockdep_set_class(&writer->lock, &writer->lock_class);
-               writer->count = 0;
-       }
-       return 0;
+#ifdef CONFIG_SMP
+       (*per_cpu_ptr(mnt->mnt_writers, smp_processor_id()))--;
+#else
+       mnt->mnt_writers--;
+#endif
 }
-fs_initcall(init_mnt_writers);
 
-static void unlock_mnt_writers(void)
+static unsigned int count_mnt_writers(struct vfsmount *mnt)
 {
+#ifdef CONFIG_SMP
+       unsigned int count = 0;
        int cpu;
-       struct mnt_writer *cpu_writer;
 
        for_each_possible_cpu(cpu) {
-               cpu_writer = &per_cpu(mnt_writers, cpu);
-               spin_unlock(&cpu_writer->lock);
+               count += *per_cpu_ptr(mnt->mnt_writers, cpu);
        }
-}
 
-static inline void __clear_mnt_count(struct mnt_writer *cpu_writer)
-{
-       if (!cpu_writer->mnt)
-               return;
-       /*
-        * This is in case anyone ever leaves an invalid,
-        * old ->mnt and a count of 0.
-        */
-       if (!cpu_writer->count)
-               return;
-       atomic_add(cpu_writer->count, &cpu_writer->mnt->__mnt_writers);
-       cpu_writer->count = 0;
-}
- /*
- * must hold cpu_writer->lock
- */
-static inline void use_cpu_writer_for_mount(struct mnt_writer *cpu_writer,
-                                         struct vfsmount *mnt)
-{
-       if (cpu_writer->mnt == mnt)
-               return;
-       __clear_mnt_count(cpu_writer);
-       cpu_writer->mnt = mnt;
+       return count;
+#else
+       return mnt->mnt_writers;
+#endif
 }
 
 /*
@@ -252,74 +236,73 @@ static inline void use_cpu_writer_for_mount(struct mnt_writer *cpu_writer,
 int mnt_want_write(struct vfsmount *mnt)
 {
        int ret = 0;
-       struct mnt_writer *cpu_writer;
 
-       cpu_writer = &get_cpu_var(mnt_writers);
-       spin_lock(&cpu_writer->lock);
+       preempt_disable();
+       inc_mnt_writers(mnt);
+       /*
+        * The store to inc_mnt_writers must be visible before we pass
+        * MNT_WRITE_HOLD loop below, so that the slowpath can see our
+        * incremented count after it has set MNT_WRITE_HOLD.
+        */
+       smp_mb();
+       while (mnt->mnt_flags & MNT_WRITE_HOLD)
+               cpu_relax();
+       /*
+        * After the slowpath clears MNT_WRITE_HOLD, mnt_is_readonly will
+        * be set to match its requirements. So we must not load that until
+        * MNT_WRITE_HOLD is cleared.
+        */
+       smp_rmb();
        if (__mnt_is_readonly(mnt)) {
+               dec_mnt_writers(mnt);
                ret = -EROFS;
                goto out;
        }
-       use_cpu_writer_for_mount(cpu_writer, mnt);
-       cpu_writer->count++;
 out:
-       spin_unlock(&cpu_writer->lock);
-       put_cpu_var(mnt_writers);
+       preempt_enable();
        return ret;
 }
 EXPORT_SYMBOL_GPL(mnt_want_write);
 
-static void lock_mnt_writers(void)
-{
-       int cpu;
-       struct mnt_writer *cpu_writer;
-
-       for_each_possible_cpu(cpu) {
-               cpu_writer = &per_cpu(mnt_writers, cpu);
-               spin_lock(&cpu_writer->lock);
-               __clear_mnt_count(cpu_writer);
-               cpu_writer->mnt = NULL;
-       }
+/**
+ * mnt_clone_write - get write access to a mount
+ * @mnt: the mount on which to take a write
+ *
+ * This is effectively like mnt_want_write, except
+ * it must only be used to take an extra write reference
+ * on a mountpoint that we already know has a write reference
+ * on it. This allows some optimisation.
+ *
+ * After finished, mnt_drop_write must be called as usual to
+ * drop the reference.
+ */
+int mnt_clone_write(struct vfsmount *mnt)
+{
+       /* superblock may be r/o */
+       if (__mnt_is_readonly(mnt))
+               return -EROFS;
+       preempt_disable();
+       inc_mnt_writers(mnt);
+       preempt_enable();
+       return 0;
 }
+EXPORT_SYMBOL_GPL(mnt_clone_write);
 
-/*
- * These per-cpu write counts are not guaranteed to have
- * matched increments and decrements on any given cpu.
- * A file open()ed for write on one cpu and close()d on
- * another cpu will imbalance this count.  Make sure it
- * does not get too far out of whack.
+/**
+ * mnt_want_write_file - get write access to a file's mount
+ * @file: the file who's mount on which to take a write
+ *
+ * This is like mnt_want_write, but it takes a file and can
+ * do some optimisations if the file is open for write already
  */
-static void handle_write_count_underflow(struct vfsmount *mnt)
+int mnt_want_write_file(struct file *file)
 {
-       if (atomic_read(&mnt->__mnt_writers) >=
-           MNT_WRITER_UNDERFLOW_LIMIT)
-               return;
-       /*
-        * It isn't necessary to hold all of the locks
-        * at the same time, but doing it this way makes
-        * us share a lot more code.
-        */
-       lock_mnt_writers();
-       /*
-        * vfsmount_lock is for mnt_flags.
-        */
-       spin_lock(&vfsmount_lock);
-       /*
-        * If coalescing the per-cpu writer counts did not
-        * get us back to a positive writer count, we have
-        * a bug.
-        */
-       if ((atomic_read(&mnt->__mnt_writers) < 0) &&
-           !(mnt->mnt_flags & MNT_IMBALANCED_WRITE_COUNT)) {
-               WARN(1, KERN_DEBUG "leak detected on mount(%p) writers "
-                               "count: %d\n",
-                       mnt, atomic_read(&mnt->__mnt_writers));
-               /* use the flag to keep the dmesg spam down */
-               mnt->mnt_flags |= MNT_IMBALANCED_WRITE_COUNT;
-       }
-       spin_unlock(&vfsmount_lock);
-       unlock_mnt_writers();
+       if (!(file->f_mode & FMODE_WRITE))
+               return mnt_want_write(file->f_path.mnt);
+       else
+               return mnt_clone_write(file->f_path.mnt);
 }
+EXPORT_SYMBOL_GPL(mnt_want_write_file);
 
 /**
  * mnt_drop_write - give up write access to a mount
@@ -331,37 +314,9 @@ static void handle_write_count_underflow(struct vfsmount *mnt)
  */
 void mnt_drop_write(struct vfsmount *mnt)
 {
-       int must_check_underflow = 0;
-       struct mnt_writer *cpu_writer;
-
-       cpu_writer = &get_cpu_var(mnt_writers);
-       spin_lock(&cpu_writer->lock);
-
-       use_cpu_writer_for_mount(cpu_writer, mnt);
-       if (cpu_writer->count > 0) {
-               cpu_writer->count--;
-       } else {
-               must_check_underflow = 1;
-               atomic_dec(&mnt->__mnt_writers);
-       }
-
-       spin_unlock(&cpu_writer->lock);
-       /*
-        * Logically, we could call this each time,
-        * but the __mnt_writers cacheline tends to
-        * be cold, and makes this expensive.
-        */
-       if (must_check_underflow)
-               handle_write_count_underflow(mnt);
-       /*
-        * This could be done right after the spinlock
-        * is taken because the spinlock keeps us on
-        * the cpu, and disables preemption.  However,
-        * putting it here bounds the amount that
-        * __mnt_writers can underflow.  Without it,
-        * we could theoretically wrap __mnt_writers.
-        */
-       put_cpu_var(mnt_writers);
+       preempt_disable();
+       dec_mnt_writers(mnt);
+       preempt_enable();
 }
 EXPORT_SYMBOL_GPL(mnt_drop_write);
 
@@ -369,24 +324,41 @@ static int mnt_make_readonly(struct vfsmount *mnt)
 {
        int ret = 0;
 
-       lock_mnt_writers();
+       spin_lock(&vfsmount_lock);
+       mnt->mnt_flags |= MNT_WRITE_HOLD;
        /*
-        * With all the locks held, this value is stable
+        * After storing MNT_WRITE_HOLD, we'll read the counters. This store
+        * should be visible before we do.
         */
-       if (atomic_read(&mnt->__mnt_writers) > 0) {
-               ret = -EBUSY;
-               goto out;
-       }
+       smp_mb();
+
        /*
-        * nobody can do a successful mnt_want_write() with all
-        * of the counts in MNT_DENIED_WRITE and the locks held.
+        * With writers on hold, if this value is zero, then there are
+        * definitely no active writers (although held writers may subsequently
+        * increment the count, they'll have to wait, and decrement it after
+        * seeing MNT_READONLY).
+        *
+        * It is OK to have counter incremented on one CPU and decremented on
+        * another: the sum will add up correctly. The danger would be when we
+        * sum up each counter, if we read a counter before it is incremented,
+        * but then read another CPU's count which it has been subsequently
+        * decremented from -- we would see more decrements than we should.
+        * MNT_WRITE_HOLD protects against this scenario, because
+        * mnt_want_write first increments count, then smp_mb, then spins on
+        * MNT_WRITE_HOLD, so it can't be decremented by another CPU while
+        * we're counting up here.
         */
-       spin_lock(&vfsmount_lock);
-       if (!ret)
+       if (count_mnt_writers(mnt) > 0)
+               ret = -EBUSY;
+       else
                mnt->mnt_flags |= MNT_READONLY;
+       /*
+        * MNT_READONLY must become visible before ~MNT_WRITE_HOLD, so writers
+        * that become unheld will see MNT_READONLY.
+        */
+       smp_wmb();
+       mnt->mnt_flags &= ~MNT_WRITE_HOLD;
        spin_unlock(&vfsmount_lock);
-out:
-       unlock_mnt_writers();
        return ret;
 }
 
@@ -397,11 +369,10 @@ static void __mnt_unmake_readonly(struct vfsmount *mnt)
        spin_unlock(&vfsmount_lock);
 }
 
-int simple_set_mnt(struct vfsmount *mnt, struct super_block *sb)
+void simple_set_mnt(struct vfsmount *mnt, struct super_block *sb)
 {
        mnt->mnt_sb = sb;
        mnt->mnt_root = dget(sb->s_root);
-       return 0;
 }
 
 EXPORT_SYMBOL(simple_set_mnt);
@@ -410,6 +381,9 @@ void free_vfsmnt(struct vfsmount *mnt)
 {
        kfree(mnt->mnt_devname);
        mnt_free_id(mnt);
+#ifdef CONFIG_SMP
+       free_percpu(mnt->mnt_writers);
+#endif
        kmem_cache_free(mnt_cache, mnt);
 }
 
@@ -442,11 +416,11 @@ struct vfsmount *__lookup_mnt(struct vfsmount *mnt, struct dentry *dentry,
  * lookup_mnt increments the ref count before returning
  * the vfsmount struct.
  */
-struct vfsmount *lookup_mnt(struct vfsmount *mnt, struct dentry *dentry)
+struct vfsmount *lookup_mnt(struct path *path)
 {
        struct vfsmount *child_mnt;
        spin_lock(&vfsmount_lock);
-       if ((child_mnt = __lookup_mnt(mnt, dentry, 1)))
+       if ((child_mnt = __lookup_mnt(path->mnt, path->dentry, 1)))
                mntget(child_mnt);
        spin_unlock(&vfsmount_lock);
        return child_mnt;
@@ -604,36 +578,18 @@ static struct vfsmount *clone_mnt(struct vfsmount *old, struct dentry *root,
 
 static inline void __mntput(struct vfsmount *mnt)
 {
-       int cpu;
        struct super_block *sb = mnt->mnt_sb;
        /*
-        * We don't have to hold all of the locks at the
-        * same time here because we know that we're the
-        * last reference to mnt and that no new writers
-        * can come in.
-        */
-       for_each_possible_cpu(cpu) {
-               struct mnt_writer *cpu_writer = &per_cpu(mnt_writers, cpu);
-               if (cpu_writer->mnt != mnt)
-                       continue;
-               spin_lock(&cpu_writer->lock);
-               atomic_add(cpu_writer->count, &mnt->__mnt_writers);
-               cpu_writer->count = 0;
-               /*
-                * Might as well do this so that no one
-                * ever sees the pointer and expects
-                * it to be valid.
-                */
-               cpu_writer->mnt = NULL;
-               spin_unlock(&cpu_writer->lock);
-       }
-       /*
         * This probably indicates that somebody messed
         * up a mnt_want/drop_write() pair.  If this
         * happens, the filesystem was probably unable
         * to make r/w->r/o transitions.
         */
-       WARN_ON(atomic_read(&mnt->__mnt_writers));
+       /*
+        * atomic_dec_and_lock() used to deal with ->mnt_count decrements
+        * provides barriers, so count_mnt_writers() below is safe.  AV
+        */
+       WARN_ON(count_mnt_writers(mnt));
        dput(mnt->mnt_root);
        free_vfsmnt(mnt);
        deactivate_super(sb);
@@ -693,12 +649,16 @@ static inline void mangle(struct seq_file *m, const char *s)
  */
 int generic_show_options(struct seq_file *m, struct vfsmount *mnt)
 {
-       const char *options = mnt->mnt_sb->s_options;
+       const char *options;
+
+       rcu_read_lock();
+       options = rcu_dereference(mnt->mnt_sb->s_options);
 
        if (options != NULL && options[0]) {
                seq_putc(m, ',');
                mangle(m, options);
        }
+       rcu_read_unlock();
 
        return 0;
 }
@@ -719,11 +679,22 @@ EXPORT_SYMBOL(generic_show_options);
  */
 void save_mount_options(struct super_block *sb, char *options)
 {
-       kfree(sb->s_options);
-       sb->s_options = kstrdup(options, GFP_KERNEL);
+       BUG_ON(sb->s_options);
+       rcu_assign_pointer(sb->s_options, kstrdup(options, GFP_KERNEL));
 }
 EXPORT_SYMBOL(save_mount_options);
 
+void replace_mount_options(struct super_block *sb, char *options)
+{
+       char *old = sb->s_options;
+       rcu_assign_pointer(sb->s_options, options);
+       if (old) {
+               synchronize_rcu();
+               kfree(old);
+       }
+}
+EXPORT_SYMBOL(replace_mount_options);
+
 #ifdef CONFIG_PROC_FS
 /* iterator */
 static void *m_start(struct seq_file *m, loff_t *pos)
@@ -778,6 +749,7 @@ static void show_mnt_opts(struct seq_file *m, struct vfsmount *mnt)
                { MNT_NOATIME, ",noatime" },
                { MNT_NODIRATIME, ",nodiratime" },
                { MNT_RELATIME, ",relatime" },
+               { MNT_STRICTATIME, ",strictatime" },
                { 0, NULL }
        };
        const struct proc_fs_info *fs_infop;
@@ -1070,9 +1042,7 @@ static int do_umount(struct vfsmount *mnt, int flags)
         */
 
        if (flags & MNT_FORCE && sb->s_op->umount_begin) {
-               lock_kernel();
                sb->s_op->umount_begin(sb);
-               unlock_kernel();
        }
 
        /*
@@ -1128,7 +1098,7 @@ static int do_umount(struct vfsmount *mnt, int flags)
  * unixes. Our API is identical to OSF/1 to avoid making a mess of AMD
  */
 
-asmlinkage long sys_umount(char __user * name, int flags)
+SYSCALL_DEFINE2(umount, char __user *, name, int, flags)
 {
        struct path path;
        int retval;
@@ -1160,7 +1130,7 @@ out:
 /*
  *     The 2.0 compatible umount. No flags.
  */
-asmlinkage long sys_oldumount(char __user * name)
+SYSCALL_DEFINE1(oldumount, char __user *, name)
 {
        return sys_umount(name, 0);
 }
@@ -1237,11 +1207,11 @@ Enomem:
        return NULL;
 }
 
-struct vfsmount *collect_mounts(struct vfsmount *mnt, struct dentry *dentry)
+struct vfsmount *collect_mounts(struct path *path)
 {
        struct vfsmount *tree;
        down_write(&namespace_sem);
-       tree = copy_tree(mnt, dentry, CL_COPY_ALL | CL_PRIVATE);
+       tree = copy_tree(path->mnt, path->dentry, CL_COPY_ALL | CL_PRIVATE);
        up_write(&namespace_sem);
        return tree;
 }
@@ -1374,7 +1344,7 @@ static int attach_recursive_mnt(struct vfsmount *source_mnt,
        if (parent_path) {
                detach_mnt(source_mnt, parent_path);
                attach_mnt(source_mnt, path);
-               touch_mnt_namespace(current->nsproxy->mnt_ns);
+               touch_mnt_namespace(parent_path->mnt->mnt_ns);
        } else {
                mnt_set_mountpoint(dest_mnt, dest_dentry, source_mnt);
                commit_tree(source_mnt);
@@ -1414,7 +1384,7 @@ static int graft_tree(struct vfsmount *mnt, struct path *path)
                goto out_unlock;
 
        err = -ENOENT;
-       if (IS_ROOT(path->dentry) || !d_unhashed(path->dentry))
+       if (!d_unlinked(path->dentry))
                err = attach_recursive_mnt(mnt, path, NULL);
 out_unlock:
        mutex_unlock(&path->dentry->d_inode->i_mutex);
@@ -1585,7 +1555,7 @@ static int do_move_mount(struct path *path, char *old_name)
 
        down_write(&namespace_sem);
        while (d_mountpoint(path->dentry) &&
-              follow_down(&path->mnt, &path->dentry))
+              follow_down(path))
                ;
        err = -EINVAL;
        if (!check_mnt(path->mnt) || !check_mnt(old_path.mnt))
@@ -1596,7 +1566,7 @@ static int do_move_mount(struct path *path, char *old_name)
        if (IS_DEADDIR(path->dentry->d_inode))
                goto out1;
 
-       if (!IS_ROOT(path->dentry) && d_unhashed(path->dentry))
+       if (d_unlinked(path->dentry))
                goto out1;
 
        err = -EINVAL;
@@ -1679,10 +1649,10 @@ int do_add_mount(struct vfsmount *newmnt, struct path *path,
        down_write(&namespace_sem);
        /* Something was mounted here while we slept */
        while (d_mountpoint(path->dentry) &&
-              follow_down(&path->mnt, &path->dentry))
+              follow_down(path))
                ;
        err = -EINVAL;
-       if (!check_mnt(path->mnt))
+       if (!(mnt_flags & MNT_SHRINKABLE) && !check_mnt(path->mnt))
                goto unlock;
 
        /* Refuse the same filesystem on the same mount point */
@@ -1917,6 +1887,10 @@ long do_mount(char *dev_name, char *dir_name, char *type_page,
        if (data_page)
                ((char *)data_page)[PAGE_SIZE - 1] = 0;
 
+       /* Default to relatime unless overriden */
+       if (!(flags & MS_NOATIME))
+               mnt_flags |= MNT_RELATIME;
+
        /* Separate the per-mountpoint flags */
        if (flags & MS_NOSUID)
                mnt_flags |= MNT_NOSUID;
@@ -1928,13 +1902,14 @@ long do_mount(char *dev_name, char *dir_name, char *type_page,
                mnt_flags |= MNT_NOATIME;
        if (flags & MS_NODIRATIME)
                mnt_flags |= MNT_NODIRATIME;
-       if (flags & MS_RELATIME)
-               mnt_flags |= MNT_RELATIME;
+       if (flags & MS_STRICTATIME)
+               mnt_flags &= ~(MNT_RELATIME | MNT_NOATIME);
        if (flags & MS_RDONLY)
                mnt_flags |= MNT_READONLY;
 
        flags &= ~(MS_NOSUID | MS_NOEXEC | MS_NODEV | MS_ACTIVE |
-                  MS_NOATIME | MS_NODIRATIME | MS_RELATIME| MS_KERNMOUNT);
+                  MS_NOATIME | MS_NODIRATIME | MS_RELATIME| MS_KERNMOUNT |
+                  MS_STRICTATIME);
 
        /* ... and get the mountpoint */
        retval = kern_path(dir_name, LOOKUP_FOLLOW, &path);
@@ -1946,6 +1921,7 @@ long do_mount(char *dev_name, char *dir_name, char *type_page,
        if (retval)
                goto dput_out;
 
+       lock_kernel();
        if (flags & MS_REMOUNT)
                retval = do_remount(&path, flags & ~MS_REMOUNT, mnt_flags,
                                    data_page);
@@ -1958,6 +1934,7 @@ long do_mount(char *dev_name, char *dir_name, char *type_page,
        else
                retval = do_new_mount(&path, type_page, flags, mnt_flags,
                                      dev_name, data_page);
+       unlock_kernel();
 dput_out:
        path_put(&path);
        return retval;
@@ -2045,9 +2022,8 @@ struct mnt_namespace *copy_mnt_ns(unsigned long flags, struct mnt_namespace *ns,
        return new_ns;
 }
 
-asmlinkage long sys_mount(char __user * dev_name, char __user * dir_name,
-                         char __user * type, unsigned long flags,
-                         void __user * data)
+SYSCALL_DEFINE5(mount, char __user *, dev_name, char __user *, dir_name,
+               char __user *, type, unsigned long, flags, void __user *, data)
 {
        int retval;
        unsigned long data_page;
@@ -2072,10 +2048,8 @@ asmlinkage long sys_mount(char __user * dev_name, char __user * dir_name,
        if (retval < 0)
                goto out3;
 
-       lock_kernel();
        retval = do_mount((char *)dev_page, dir_page, (char *)type_page,
                          flags, (void *)data_page);
-       unlock_kernel();
        free_page(data_page);
 
 out3:
@@ -2088,66 +2062,6 @@ out1:
 }
 
 /*
- * Replace the fs->{rootmnt,root} with {mnt,dentry}. Put the old values.
- * It can block. Requires the big lock held.
- */
-void set_fs_root(struct fs_struct *fs, struct path *path)
-{
-       struct path old_root;
-
-       write_lock(&fs->lock);
-       old_root = fs->root;
-       fs->root = *path;
-       path_get(path);
-       write_unlock(&fs->lock);
-       if (old_root.dentry)
-               path_put(&old_root);
-}
-
-/*
- * Replace the fs->{pwdmnt,pwd} with {mnt,dentry}. Put the old values.
- * It can block. Requires the big lock held.
- */
-void set_fs_pwd(struct fs_struct *fs, struct path *path)
-{
-       struct path old_pwd;
-
-       write_lock(&fs->lock);
-       old_pwd = fs->pwd;
-       fs->pwd = *path;
-       path_get(path);
-       write_unlock(&fs->lock);
-
-       if (old_pwd.dentry)
-               path_put(&old_pwd);
-}
-
-static void chroot_fs_refs(struct path *old_root, struct path *new_root)
-{
-       struct task_struct *g, *p;
-       struct fs_struct *fs;
-
-       read_lock(&tasklist_lock);
-       do_each_thread(g, p) {
-               task_lock(p);
-               fs = p->fs;
-               if (fs) {
-                       atomic_inc(&fs->count);
-                       task_unlock(p);
-                       if (fs->root.dentry == old_root->dentry
-                           && fs->root.mnt == old_root->mnt)
-                               set_fs_root(fs, new_root);
-                       if (fs->pwd.dentry == old_root->dentry
-                           && fs->pwd.mnt == old_root->mnt)
-                               set_fs_pwd(fs, new_root);
-                       put_fs_struct(fs);
-               } else
-                       task_unlock(p);
-       } while_each_thread(g, p);
-       read_unlock(&tasklist_lock);
-}
-
-/*
  * pivot_root Semantics:
  * Moves the root file system of the current process to the directory put_old,
  * makes new_root as the new root file system of the current process, and sets
@@ -2172,8 +2086,8 @@ static void chroot_fs_refs(struct path *old_root, struct path *new_root)
  *    though, so you may need to say mount --bind /nfs/my_root /nfs/my_root
  *    first.
  */
-asmlinkage long sys_pivot_root(const char __user * new_root,
-                              const char __user * put_old)
+SYSCALL_DEFINE2(pivot_root, const char __user *, new_root,
+               const char __user *, put_old)
 {
        struct vfsmount *tmp;
        struct path new, old, parent_path, root_parent, root;
@@ -2215,9 +2129,9 @@ asmlinkage long sys_pivot_root(const char __user * new_root,
        error = -ENOENT;
        if (IS_DEADDIR(new.dentry->d_inode))
                goto out2;
-       if (d_unhashed(new.dentry) && !IS_ROOT(new.dentry))
+       if (d_unlinked(new.dentry))
                goto out2;
-       if (d_unhashed(old.dentry) && !IS_ROOT(old.dentry))
+       if (d_unlinked(old.dentry))
                goto out2;
        error = -EBUSY;
        if (new.mnt == root.mnt ||