Re: [PATCH v3 04/10] fscrypt: Drop d_revalidate once the key is added

[Gabriel Krisman Bertazi <krisman@suse.de>]

Eric Biggers <ebiggers@kernel.org> writes:

> On Thu, Jan 25, 2024 at 05:20:56PM -0300, Gabriel Krisman Bertazi wrote:
>> Eric Biggers <ebiggers@kernel.org> writes:
>> 
>> > On Fri, Jan 19, 2024 at 03:47:36PM -0300, Gabriel Krisman Bertazi wrote:
>> >> /*
>> >>  * When d_splice_alias() moves a directory's no-key alias to its plaintext alias
>> >>  * as a result of the encryption key being added, DCACHE_NOKEY_NAME must be
>> >>  * cleared.  Note that we don't have to support arbitrary moves of this flag
>> >>  * because fscrypt doesn't allow no-key names to be the source or target of a
>> >>  * rename().
>> >>  */
>> >>  static inline void fscrypt_handle_d_move(struct dentry *dentry)
>> >>  {
>> >>  	dentry->d_flags &= ~DCACHE_NOKEY_NAME;
>> >> +
>> >> +	/*
>> >> +	 * Save the d_revalidate call cost during VFS operations.  We
>> >> +	 * can do it because, when the key is available, the dentry
>> >> +	 * can't go stale and the key won't go away without eviction.
>> >> +	 */
>> >> +	if (dentry->d_op && dentry->d_op->d_revalidate == fscrypt_d_revalidate)
>> >> +		dentry->d_flags &= ~DCACHE_OP_REVALIDATE;
>> >>  }
>> >
>> > Is there any way to optimize this further for the case where fscrypt is not
>> > being used?  This is called unconditionally from d_move().  I've generally been
>> > trying to avoid things like this where the fscrypt support slows things down for
>> > everyone even when they're not using the feature.
>> 
>> The problem would be figuring out whether the filesystem has fscrypt
>> enabled.  I think we can rely on sb->s_cop always being set:
>> 
>> if (sb->s_cop)
>>    fscrypt_handle_d_move(dentry);
>> 
>> What do you think?
>
> That's better, I just wonder if there's an even better way.  Do you need to do
> this for dentries that don't have DCACHE_NOKEY_NAME set?  If not, it would be
> more efficient to test for DCACHE_NOKEY_NAME before clearing the flags.

I like that.  We don't need it for dentries without DCACHE_NOKEY_NAME,
because those dentries have the d_revalidate disabled at lookup-time.

I raced my v4 with your comment, but I'll spin a v5 folding in this
suggestion shortly.

-- 
Gabriel Krisman Bertazi