Received-SPF: pass (google.com: domain of linux-ext4+bounces-506-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Date: Tue, 19 Dec 2023 16:12:22 -0700
From: Eric Biggers <ebiggers@kernel.org>
To: Gabriel Krisman Bertazi <krisman@suse.de>
Cc: viro@zeniv.linux.org.uk, 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 0/8] Revert setting casefolding dentry operations
 through s_d_op
Message-ID: <20231219231222.GI38652@quark.localdomain>
References: <20231215211608.6449-1-krisman@suse.de>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20231215211608.6449-1-krisman@suse.de>

On Fri, Dec 15, 2023 at 04:16:00PM -0500, Gabriel Krisman Bertazi wrote:
> [Apologies for the quick spin of a v2.  The only difference are a couple
> fixes to the build when CONFIG_UNICODE=n caught by LKP and detailed in
> each patch changelog.]
> 
> When case-insensitive and fscrypt were adapted to work together, we moved the
> code that sets the dentry operations for case-insensitive dentries(d_hash and
> d_compare) to happen from a helper inside ->lookup.  This is because fscrypt
> wants to set d_revalidate only on some dentries, so it does it only for them in
> d_revalidate.
> 
> But, case-insensitive hooks are actually set on all dentries in the filesystem,
> so the natural place to do it is through s_d_op and let d_alloc handle it [1].
> In addition, doing it inside the ->lookup is a problem for case-insensitive
> dentries that are not created through ->lookup, like those coming
> open-by-fhandle[2], which will not see the required d_ops.
> 
> This patchset therefore reverts to using sb->s_d_op to set the dentry operations
> for case-insensitive filesystems.  In order to set case-insensitive hooks early
> and not require every dentry to have d_revalidate in case-insensitive
> filesystems, it introduces a patch suggested by Al Viro to disable d_revalidate
> on some dentries on the fly.
> 
> It survives fstests encrypt and quick groups without regressions.  Based on
> v6.7-rc1.
> 
> [1] https://lore.kernel.org/linux-fsdevel/20231123195327.GP38156@ZenIV/
> [2] https://lore.kernel.org/linux-fsdevel/20231123171255.GN38156@ZenIV/
> 
> Gabriel Krisman Bertazi (8):
>   dcache: Add helper to disable d_revalidate for a specific dentry
>   fscrypt: Drop d_revalidate if key is available
>   libfs: Merge encrypted_ci_dentry_ops and ci_dentry_ops
>   libfs: Expose generic_ci_dentry_ops outside of libfs
>   ext4: Set the case-insensitive dentry operations through ->s_d_op
>   f2fs: Set the case-insensitive dentry operations through ->s_d_op
>   libfs: Don't support setting casefold operations during lookup
>   fscrypt: Move d_revalidate configuration back into fscrypt

Thanks Gabriel, this series looks good.  Sorry that we missed this when adding
the support for encrypt+casefold.

It's slightly awkward that some lines of code added by patches 5-6 are removed
in patch 8.  These changes look very hard to split up, though, so you've
probably done about the best that can be done.

One question/request: besides performance, the other reason we're so careful
about minimizing when ->d_revalidate is set for fscrypt is so that overlayfs
works on encrypted directories.  This is because overlayfs is not compatible
with ->d_revalidate.  I think your solution still works for that, since
DCACHE_OP_REVALIDATE will be cleared after the first call to
fscrypt_d_revalidate(), and when checking for usupported dentries overlayfs does
indeed check for DCACHE_OP_REVALIDATE instead of ->d_revalidate directly.
However, that does rely on that very first call to ->d_revalidate actually
happening before the check is done.  It would be nice to verify that
overlayfs+fscrypt indeed continues to work, and explicitly mention this
somewhere (I don't see any mention of overlayfs+fscrypt in the series).

- Eric