lkml.org 
[lkml]   [2013]   [Feb]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v2 2/2] cgroup: fix cgroup_path() vs rename() race, take 2
Hello, Li.

On Mon, Feb 18, 2013 at 09:16:48AM +0800, Li Zefan wrote:
> @@ -171,6 +171,7 @@ struct cgroup {
>
> struct cgroup *parent; /* my parent */
> struct dentry *dentry; /* cgroup fs entry, RCU protected */
> + char __rcu *name; /* a copy of dentry->d_name */

A brief explanation of why this is necessary and how rcu is used would
be nice.

> +static char *cgroup_alloc_name(struct dentry *dentry)
> +{
> + char *name;
> +
> + name = kmalloc(dentry->d_name.len + 1, GFP_KERNEL);
> + if (!name)
> + return NULL;
> + memcpy(name, dentry->d_name.name, dentry->d_name.len);
> + name[dentry->d_name.len] = '\0';
> + return name;
> +}

While d_name has length field, it's always properly NULL terminated,
so kstrdup() should suffice here. Right, Al?

> @@ -1613,13 +1626,19 @@ static struct dentry *cgroup_mount(struct file_system_type *fs_type,
...
> - inode = sb->s_root->d_inode;
> + dentry = sb->s_root;
> + inode = dentry->d_inode;
> +
> + root_cgrp->name = cgroup_alloc_name(dentry);
> + if (!root_cgrp->name)
> + goto drop_new_super;

Don't we need an RCU assignment? Is it safe because it isn't online
yet? But wouldn't this still trigger sparse warning?

> @@ -1751,6 +1770,8 @@ static void cgroup_kill_sb(struct super_block *sb) {
> mutex_unlock(&cgroup_root_mutex);
> mutex_unlock(&cgroup_mutex);
>
> + synchronize_rcu();

An explanation on what we're synchronizing would be nice. Barriers
without explanation sucks because there's nothing directly linking the
barriers to the things which are being protected.

> @@ -2539,13 +2558,41 @@ static int cgroup_file_release(struct inode *inode, struct file *file)
> static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
> struct inode *new_dir, struct dentry *new_dentry)
> {
...
> + old_name = cgrp->name;
> + rcu_assign_pointer(cgrp->name, name);
> +
> + synchronize_rcu();

Please don't call synchronize_rcu() from interface which is directly
visible to userland. It leads to sporadic difficult-to-reproduce
latencies which hurt enough in corner cases and this is kmalloc
memory. It's not like kfree_rcu() is difficult to use or anything.

> + kfree(old_name);
> + return 0;
> }
>
> static struct simple_xattrs *__d_xattrs(struct dentry *dentry)
> @@ -4144,9 +4191,13 @@ static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
> if (!cgrp)
> return -ENOMEM;
>
> + cgrp->name = cgroup_alloc_name(dentry);
> + if (!cgrp->name)
> + goto err_free_cgrp;

Ditto with assignment.

Thanks.

--
tejun


\
 
 \ /
  Last update: 2013-02-18 19:21    [W:0.062 / U:0.016 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site