Re: [PATCH] generic/556: add a check to make sure ext4 supports encrypted casefolding

[Eric Biggers <ebiggers@kernel.org>]

On Mon, Jun 21, 2021 at 10:37:58AM -0700, Eric Biggers wrote:
> On Mon, Jun 21, 2021 at 12:49:01PM -0400, Theodore Ts'o wrote:
> > diff --git a/tests/generic/556 b/tests/generic/556
> > index 3145188c..7916a08e 100755
> > --- a/tests/generic/556
> > +++ b/tests/generic/556
> > @@ -16,6 +16,7 @@ status=1 # failure is thea default
> >  . ./common/attr
> >  
> >  _supported_fs generic
> > +_require_encrypted_casefold
> 
> This isn't an encrypt+casefold test though, but rather just a casefold test.  It
> only becomes an encrypt+casefold test when $MOUNT_OPTIONS includes
> test_dummy_encryption.
> 
> I think we shouldn't update the test itself, but rather make
> _require_scratch_casefold() call _require_encrypted_casefold() if
> test_dummy_encryption is in $MOUNT_OPTIONS.
> 

Hi Ted, just to follow up on this patch which never got merged:

I tried to reproduce the test failure with the 5.4 and 5.10 kernels, whose ext4
supported casefold but not encrypt+casefold.  However, it does *not* reproduce
with the "ext4/encrypt" configuration of kvm-xfstests, since
_require_scratch_casefold() already detects that the filesystem cannot mounted
with both the encrypt and casefold features:

        if ! _try_scratch_mount &>>seqres.full; then
                _notrun "kernel can't mount filesystem with the encoding set by userspace"
        fi

I think the problem is that you were running xfstests with test_dummy_encryption
but without adding "-O encrypt" to EXT_MKFS_OPTIONS.  That sort of works because
of the following code in __ext4_fill_super() in the kernel which force-enables
the encrypt feature if the test_dummy_encryption mount option was given:

	if (DUMMY_ENCRYPTION_ENABLED(sbi) && !sb_rdonly(sb) &&
	    !ext4_has_feature_encrypt(sb)) {
		ext4_set_feature_encrypt(sb);
		ext4_commit_super(sb);
	}

However, I don't think ext4 should be doing this.  The kernel should generally
not be messing with the feature flags, and this code can force-enable 'encrypt'
in cases where e2fsprogs would *not* allow it (e.g. the encrypt+casefold
situation being encountered here).  Also, f2fs doesn't allow this.

So I am planning to send a kernel patch to remove the above code from ext4.

- Eric