Received-SPF: pass (google.com: domain of linux-ext4+bounces-512-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Date: Thu, 21 Dec 2023 07:39:40 +0000
From: Al Viro <viro@zeniv.linux.org.uk>
To: Gabriel Krisman Bertazi <krisman@suse.de>
Cc: ebiggers@kernel.org, jaegeuk@kernel.org, tytso@mit.edu,
	linux-f2fs-devel@lists.sourceforge.net, linux-ext4@vger.kernel.org,
	linux-fsdevel@vger.kernel.org
Subject: Re: [PATCH v2 8/8] fscrypt: Move d_revalidate configuration back
 into fscrypt
Message-ID: <20231221073940.GC1674809@ZenIV>
References: <20231215211608.6449-1-krisman@suse.de>
 <20231215211608.6449-9-krisman@suse.de>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20231215211608.6449-9-krisman@suse.de>
Sender: Al Viro <viro@ftp.linux.org.uk>

On Fri, Dec 15, 2023 at 04:16:08PM -0500, Gabriel Krisman Bertazi wrote:

> +static const struct dentry_operations fscrypt_dentry_ops = {
> +	.d_revalidate = fscrypt_d_revalidate,
> +};
> +
>  int __fscrypt_prepare_lookup(struct inode *dir, struct dentry *dentry,
>  			     struct fscrypt_name *fname)
>  {
> @@ -106,6 +110,10 @@ int __fscrypt_prepare_lookup(struct inode *dir, struct dentry *dentry,
>  		spin_lock(&dentry->d_lock);
>  		dentry->d_flags |= DCACHE_NOKEY_NAME;
>  		spin_unlock(&dentry->d_lock);
> +
> +		/* Give preference to the filesystem hooks, if any. */
> +		if (!dentry->d_op)
> +			d_set_d_op(dentry, &fscrypt_dentry_ops);
>  	}
>  	return err;

Hmm...  Could we simply set ->s_d_op to &fscrypt_dentry_ops in non-ci case
*AND* have __fscrypt_prepare_lookup() clear DCACHE_OP_REVALIDATE in case
when it's not setting DCACHE_NOKEY_NAME and ->d_op->d_revalidate is
equal to fscrypt_d_revalidate?  I mean,

	spin_lock(&dentry->d_lock);
        if (fname->is_nokey_name)
                dentry->d_flags |= DCACHE_NOKEY_NAME;
        else if (dentry->d_flags & DCACHE_OP_REVALIDATE &&
		 dentry->d_op->d_revalidate == fscrypt_d_revalidate)
		dentry->d_flags &= ~DCACHE_OP_REVALIDATE;
	spin_unlock(&dentry->d_lock);

here + always set ->s_d_op for ext4 and friends (conditional upon
the CONFIG_UNICODE).

No encryption - fine, you get ->is_nokey_name false from the very
beginning, DCACHE_OP_REVALIDATE is cleared and VFS won't ever call
->d_revalidate(); not even the first time.  

Yes, you pay minimal price in dentry_unlink_inode() when we hit
        if (dentry->d_op && dentry->d_op->d_iput)
and bugger off after the second fetch instead of the first one.
I would be quite surprised if it turns out to be measurable,
but if it is, we can always add DCACHE_OP_IPUT to flags.
Similar for ->d_op->d_release (called in the end of
__dentry_kill()).  Again, that only makes sense if we get
a measurable overhead from that.