On Thu, Sep 5, 2013 at 1:18 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> On Thu, Sep 05, 2013 at 11:44:37AM +0200, Miklos Szeredi wrote:
>> +static bool __has_unlinked_ancestor(struct dentry *dentry)
>> +{
>> + struct dentry *this;
>> +
>> + for (this = dentry; !IS_ROOT(this); this = this->d_parent) {
>> + int is_unhashed;
>> +
>> + /* Need exclusion wrt. check_submounts_and_drop() */
>> + spin_lock(&this->d_lock);
>> + is_unhashed = d_unhashed(this);
>> + spin_unlock(&this->d_lock);
>> +
>> + if (is_unhashed)
>> + return true;
>> + }
>> + return false;
>> +}
>> +
>> +/*
>> + * Called by mount code to check if the mountpoint is reachable (e.g. NFS can
>> + * unhash a directory dentry and then the complete subtree can become
>> + * unreachable).
>> + */
>> +bool has_unlinked_ancestor(struct dentry *dentry)
>> +{
>> + bool found;
>> +
>> + /* Need exclusion wrt. check_submounts_and_drop() */
>> + write_seqlock(&rename_lock);
>> + found = __has_unlinked_ancestor(dentry);
>> + write_sequnlock(&rename_lock);
>> +
>> + return found;
>> +}
>> +
>> /*
>> * Search the dentry child list of the specified parent,
>> * and move any unused dentries to the end of the unused
>> diff --git a/fs/internal.h b/fs/internal.h
>> index 7c5f01c..d232355 100644
>> --- a/fs/internal.h
>> +++ b/fs/internal.h
>> @@ -126,6 +126,7 @@ extern int invalidate_inodes(struct super_block *, bool);
>> * dcache.c
>> */
>> extern struct dentry *__d_alloc(struct super_block *, const struct qstr *);
>> +extern bool has_unlinked_ancestor(struct dentry *dentry);
>>
>> /*
>> * read_write.c
>> diff --git a/fs/namespace.c b/fs/namespace.c
>> index a45ba4f..91b1c39 100644
>> --- a/fs/namespace.c
>> +++ b/fs/namespace.c
>> @@ -634,6 +634,15 @@ static struct mountpoint *new_mountpoint(struct dentry *dentry)
>> }
>> dentry->d_flags |= DCACHE_MOUNTED;
>> spin_unlock(&dentry->d_lock);
>> +
>> + if (has_unlinked_ancestor(dentry)) {
>> + spin_lock(&dentry->d_lock);
>> + dentry->d_flags &= ~DCACHE_MOUNTED;
>> + spin_unlock(&dentry->d_lock);
>> + kfree(mp);
>> + return ERR_PTR(-ENOENT);
>> + }
>
> Something's really odd with locking here. You are take d_lock, do one
> check, set flag, drop d_lock, grab rename_lock, do another check (taking
> and dropping d_lock in process), and, in case that check fails, grab
> d_lock again to clear the flag.
>
> At the very least it's a massive overkill. Just grab rename_lock, then
> d_lock, then do the damn check and set the flag only on success. Moreover,
> with rename_lock held, do you need d_lock on ancestors to mess with in
> has_unlinked_ancestor()?
Yes, we need hard exclusion for the __d_drop() part. rename_lock can
provide one if we always take it for write in
check_submounts_and_drop(). But if we only take it for read then
that's not enough.
And we do in fact also need DCACHE_MOUNTED set *before* checking
ancestors. Otherwise check_submounts_and_drop() could succeed and
has_unlinked_ancestor() return false, resulting in a dropped dentry
and a mount below it. Though this is mostly theoretical at this
point.
Thanks,
Miklos
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
