Filesystems and Cryptography

Niels de Vos pointed out that a lot of filesystems were adding support for fscrypt, which encrypts user files, and that it might be useful for kernel users to be able to enable that kernel feature for some filesystems but not others. He proposed a simple kernel config option that would allow per-filesystem encryption. He remarked, "This RFC is mostly for checking the acceptance of this solution, or if an other direction is preferred."

Theodore Ts'o didn't like this idea at all. He remarked pointedly, "I'm not sure what's the motivation behind adding this configuration option. If memory serves, early in the fscrypt development we did have per-file system CONFIG's for fscrypt, but we consciously removed it, just as we no longer have per-file system CONFIG's to enable or disable Posix ACL's or extended attributes, in the name of simplifying the kernel config."

Niels confirmed that the Linux distributions were a big motivator for his initial post. He said, "This is mostly why I sent this RFC. We are interested in enabling fscrypt for CephFS (soonish) as a network filesystem, but not for local filesystems (we recommend dm-crypt for those). The idea is that functionality that isn't available, can also not (easily) cause breakage."

He pointed out that the kernel actually did have special per-filesystem config options already, for ACLs and other security features. He said, "Because these exist already, I did not expect too much concerns with proposing a CONFIG_EXT4_FS_ENCRYPTION."

But Ted replied, "Actually, I was thinking of getting rid of them, as we've already gotten rid of [EXT4_FS_XATTR]." He reiterated, "there are tons of file system features that have not and/or still are not supported for distros, but for which we don't have kernel config knobs. This includes ext4's bigalloc and inline data, btrfs's dedup and reflink support, xfs online fsck, etc., etc., etc. Heck, ext4 is only supported up to a certain size by Red Hat, and we don't have a Kernel config so that the kernel will absolutely refuse to mount an ext4 file system larger than The Officially Supported RHEL Capacity Limit for Ext4."

Eric Biggers remarked, regarding Niels' proposal, "as others have pointed out, it doesn't seem worth the complexity to do this." And he went on to offer some historical background:

"Before Linux v5.1, we did have per-filesystem options for this: CONFIG_EXT4_ENCRYPTION, CONFIG_F2FS_FS_ENCRYPTION, and CONFIG_UBIFS_FS_ENCRYPTION. If you enabled one of these, it selected CONFIG_FS_ENCRYPTION to get the code in fs/crypto/. CONFIG_FS_ENCRYPTION was a tristate, so the code in fs/crypto/ could be built as a loadable module if it was only needed by filesystems that were loadable modules themselves.

"Having fs/crypto/ possibly be a loadable module was problematic, though, because it made it impossible to call into fs/crypto/ from built-in code such as fs/buffer.c, fs/ioctl.c, fs/libfs.c, fs/super.c, fs/iomap/direct-io.c, etc. So that's why we made CONFIG_FS_ENCRYPTION into a bool. At the same time, we decided to simplify the kconfig options by removing the per-filesystem options so that it worked like CONFIG_QUOTA, CONFIG_FS_DAX, CONFIG_FS_POSIX_ACL, etc."

At around this point in the conversation, Niels remarked, "What you are explaining makes sense, and I am not sure if there is another good reason why splitting out fscrypt support per filesystem would be required. I'm checking with the folks that suggested doing this, and see where we go from there." He added, "I understand that there is a preference for reducing the number of Kconfig options for filesystems. That indeed would make it a little easier for users, so I am supportive of that as well."

And that was the end of the thread.

It's a tough balancing act! Obviously, it would be great if every feature had its own config option, as well as maybe more options to simplify having multiple features together, perhaps with exceptions specified by yet more config options. In some cases, I'm sure the kernel developers feel that a given set of features really does need such fine-grained control by system administrators, while in others they would prefer to keep the config system as simple as possible, even if it means sacrificing some behaviors system administrators might actually want.