[PATCH 06/14] sysfs: Rewrite sysfs_get_dentry

[Eric W. Biederman]

Tejun Heo <teheo at suse.de> writes:

> Eric W. Biederman wrote:
>> Currently sysfs_get_dentry is very hairy and requires all kinds
>> of locking magic.  This patch rewrites sysfs_get_dentry to
>> not use the cached sd->s_dentry, and instead simply lookup
>> and create dcache entries.
>
> I wanted to kill sd->s_dentry from the beginning.  The obstacle was
> actually the shadow directory support, ironic.
>
>> +	struct qstr name;
>> +	struct inode *inode;
>>  
>> +	parent_dentry = NULL;
>> +	dentry = dget(sysfs_sb->s_root);
>>  
>> +	do {
>> +		/* Find the first ancestor I have not looked up */
>> +		cur = sd;
>> +		while (cur->s_parent != dentry->d_fsdata)
>>  			cur = cur->s_parent;
>>  
>>  		/* look it up */
>>  		dput(parent_dentry);
>
> Shouldn't this be done after looking up the child?
Yes and that is what this is.  Delaying it a little longer
made the conditionals easier to write and verify for correctness.

>> +		parent_dentry = dentry;
>> +		name.name = cur->s_name;
>> +		name.len = strlen(cur->s_name);
>> +		dentry = d_hash_and_lookup(parent_dentry, &name);
>> +		if (dentry)
>> +			continue;
>> +		if (!create)
>> +			goto out;
>
> Probably missing dentry = NULL?
d_hash_and_lookup sets dentry, and we can't get here if (dentry != NULL)

>> +		dentry = d_alloc(parent_dentry, &name);
>> +		if (!dentry) {
>> +			dentry = ERR_PTR(-ENOMEM);
>> +			goto out;
>>  		}
>> +		inode = sysfs_get_inode(cur);
>> +		if (!inode) {
>>  			dput(dentry);
>> +			dentry = ERR_PTR(-ENOMEM);
>> +			goto out;
>>  		}
>> +		d_instantiate(dentry, inode);
>> +		sysfs_attach_dentry(cur, dentry);
>> +	} while (cur != sd);
>>  
>> +out:
>> +	dput(parent_dentry);
>> +	return dentry;
>> +}
>>  
>> @@ -750,6 +725,12 @@ static struct dentry * sysfs_lookup(struct inode *dir,
> struct dentry *dentry,
>>  	struct inode *inode;
>>  
>>  	mutex_lock(&sysfs_mutex);
>> +
>> +	/* Guard against races with sysfs_get_dentry */
>> +	result = d_hash_and_lookup(dentry->d_parent, &dentry->d_name);
>> +	if (result)
>> +		goto out;
>> +
>
> Hmmm... This is tricky but probably better than the previous hairy
> sysfs_get_dentry().  I think it would be worthwhile to comment about
> creating dentry/inode behind vfs's back in __sysfs_get_dentry().

Yes. Good point.  That is sufficiently non-intuitive and non-obvious
to deserve a comment.

> One thing I'm worried about is that dentry can now be looked up without
> holding i_mutex.  In sysfs proper, there's no synchronization problem
> but I'm not sure whether we're willing to cross that line set by vfs.
> It might come back and bite our asses hard later.

It's a reasonable concern.  I'm wondering if there are any precedents
set by distributed filesystems.  Because in a sense that is what
we are.

As for crossing that line I don't know what to say it makes the
code a lot cleaner, and if we are merged into the kernel at
least it will be visible if someone rewrites the vfs.

If sysfs_mutex nested the other way things would be easier,
and we could grab all of the i_mutexes we wanted.  I wonder if we can
be annoying in sysfs_lookup and treat that as the lock inversion
case using mutex_trylock etc.  And have sysfs_mutex be on the
outside for the rest of the cases?

Anyway back to bed with me.  I've been dreaming up to many silly
ways to abuse the dcache...

Eric