Re: [PATCH] ext4: fix race conditions in ->d_compare() and ->d_hash()

[Eric Biggers <ebiggers@kernel.org>]

On Fri, Jan 24, 2020 at 02:15:31PM +0800, Gao Xiang wrote:
> On Thu, Jan 23, 2020 at 09:42:56PM -0800, Eric Biggers wrote:
> > On Fri, Jan 24, 2020 at 01:34:23PM +0800, Gao Xiang wrote:
> > > On Thu, Jan 23, 2020 at 09:16:01PM -0800, Eric Biggers wrote:
> > > 
> > > []
> > > 
> > > > So we need READ_ONCE() to ensure that a consistent value is used.
> > > 
> > > By the way, my understanding is all pointer could be accessed
> > > atomicly guaranteed by compiler. In my opinion, we generally
> > > use READ_ONCE() on pointers for other uses (such as, avoid
> > > accessing a variable twice due to compiler optimization and
> > > it will break some logic potentially or need some data
> > > dependency barrier...)
> > > 
> > > Thanks,
> > > Gao Xiang
> > 
> > But that *is* why we need READ_ONCE() here.  Without it, there's no guarantee
> > that the compiler doesn't load the variable twice.  Please read:
> > https://github.com/google/ktsan/wiki/READ_ONCE-and-WRITE_ONCE
> 
> After scanning the patch, it seems the parent variable (dentry->d_parent)
> only referenced once as below:
> 
> -	struct inode *inode = dentry->d_parent->d_inode;
> +	const struct dentry *parent = READ_ONCE(dentry->d_parent);
> +	const struct inode *inode = READ_ONCE(parent->d_inode);
> 
> So I think it is enough as
> 
> 	const struct inode *inode = READ_ONCE(dentry->d_parent->d_inode);
> 
> to access parent inode once to avoid parent inode being accessed
> for more time (and all pointers dereference should be in atomic
> by compilers) as one reason on
> 
> 	if (!inode || !IS_CASEFOLDED(inode) || ...
> 
> or etc.
> 
> Thanks for your web reference, I will look into it. I think there
> is no worry about dentry->d_parent here because of this only one
> dereference on dentry->d_parent.
> 
> You could ignore my words anyway, just my little thought though.
> Other part of the patch is ok.
> 

While that does make it really unlikely to cause a real-world problem, it's
still undefined behavior to not properly annotate a data race, it would make the
code harder to understand as there would be no indication that there's a data
race, and it would confuse tools that try to automatically detect data races.
So let's keep the READ_ONCE() on d_parent.

- Eric