Skip to contentSkip to conversation

eccryptfs: select CONFIG_BUFFER_HEAD

PATCH In review Arnd Bergmann 9 participants 15 messages Oct 28, 2024 – Feb 16 Ecrypt file system

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 eccryptfs: select CONFIG_BUFFER_HEAD

From: Arnd Bergmann <[email protected]>

The ecryptfs file system uses functions from fs/buffer.c that are only available when CONFIG_BUFFER_HEAD is enabled:

ld.lld-20: error: undefined symbol: block_dirty_folio

Quoted 1 line
Quoted 1 line
Quoted 1 line

vmlinux.o:(ecryptfs_aops)

ld.lld-20: error: undefined symbol: block_invalidate_folio

Quoted 1 line
Quoted 1 line
Quoted 1 line

vmlinux.o:(ecryptfs_aops)

When CONFIG_BLOCK is turned off completely, this is not needed, so add a conditional 'select BUFFER_HEAD'.

Fixes 7ba13abbd31e ("fs: Turn block_invalidatepage into block_invalidate_folio")

---
Ideally we would not depend on buffer heads and instead remove the dependency here, but I could not immediately figure out how to do that.

---
 fs/ecryptfs/Kconfig | 1 +
 1 file changed, 1 insertion(+)
fs/ecryptfs/Kconfig +1 -0
index 1bdeaa6d5790..b3c603c4f808 100644
@@ -2,6 +2,7 @@
 config ECRYPT_FS
 	tristate "eCrypt filesystem layer support"
 	depends on KEYS && CRYPTO && (ENCRYPTED_KEYS || ENCRYPTED_KEYS=n)
+	select BUFFER_HEAD if BLOCK
 	select CRYPTO_ECB
 	select CRYPTO_CBC
 	select CRYPTO_MD5
Signature · 1 line
2.39.5
#1 lore raw

ecryptfs is unmaintained and untested

Quoted 10 lines · Arnd Bergmann

The ecryptfs file system uses functions from fs/buffer.c that are only available when CONFIG_BUFFER_HEAD is enabled:

ld.lld-20: error: undefined symbol: block_dirty_folio

Quoted 1 line
Quoted 1 line
Quoted 1 line

vmlinux.o:(ecryptfs_aops)

ld.lld-20: error: undefined symbol: block_invalidate_folio

Quoted 1 line
Quoted 1 line
Quoted 1 line

vmlinux.o:(ecryptfs_aops)

When CONFIG_BLOCK is turned off completely, this is not needed, so add a conditional 'select BUFFER_HEAD'.

The comment says it doesn't work without CONFIG_BLOCK:

/*
* XXX: This is pretty broken for multiple reasons: ecryptfs does not
* actually use buffer_heads, and ecryptfs will crash without
* CONFIG_BLOCK. But it matches the behavior before the default for
* address_space_operations without the ->dirty_folio method was
* cleaned up, so this is the best we can do without maintainer
* feedback.

This comment has been there since June 2021, so I think we can just delete ecryptfs now?

If we can't delete it for some reason, I think we can use filemap_dirty_folio() and remove the setting of invalidate_folio() as block_invalidate_folio() is a no-op if there are no folio_buffers. ie this in lieu of your patch:

+++ b/fs/ecryptfs/mmap.c
@@ -514,17 +514,9 @@ static sector_t ecryptfs_bmap(struct address_space *mapping, sector_t block)

 const struct address_space_operations ecryptfs_aops = {
        /*
-        * XXX: This is pretty broken for multiple reasons: ecryptfs does not
-        * actually use buffer_heads, and ecryptfs will crash without
-        * CONFIG_BLOCK.  But it matches the behavior before the default for
-        * address_space_operations without the ->dirty_folio method was
-        * cleaned up, so this is the best we can do without maintainer
-        * feedback.
+        * XXX: This entire filesystem is unmaintained and untested.
         */

-#ifdef CONFIG_BLOCK
- .dirty_folio = block_dirty_folio,
- .invalidate_folio = block_invalidate_folio,
-#endif
+ .dirty_folio = filemap_dirty_folio,
.writepages = ecryptfs_writepages,
.read_folio = ecryptfs_read_folio,
.write_begin = ecryptfs_write_begin,

#2 lore raw

Re: ecryptfs is unmaintained and untested

Quoted 12 lines · Matthew Wilcox

On Mon, Oct 28, 2024 at 02:18:45PM +0000, Arnd Bergmann wrote:

The comment says it doesn't work without CONFIG_BLOCK:

/*
* XXX: This is pretty broken for multiple reasons: ecryptfs does not
* actually use buffer_heads, and ecryptfs will crash without
* CONFIG_BLOCK. But it matches the behavior before the default for
* address_space_operations without the ->dirty_folio method was
* cleaned up, so this is the best we can do without maintainer
* feedback.

Quoted 2 lines

This comment has been there since June 2021, so I think we can just delete ecryptfs now?

I have no opinion on removing ecryptfs, but I don't how possibly removing it is related to the patch I sent, as far as I can tell it just means it relies on both CONFIG_BLOCK and CONFIG_BUFFER_HEAD then.

Is there any indication that the last users that had files on ecryptfs are unable to update their kernels?

Review

Quoted 11 lines

If we can't delete it for some reason, I think we can use filemap_dirty_folio() and remove the setting of invalidate_folio() as block_invalidate_folio() is a no-op if there are no folio_buffers. ie this in lieu of your patch:

(quoted diff) +1 -4
-#ifdef CONFIG_BLOCK
-       .dirty_folio    = block_dirty_folio,
-       .invalidate_folio = block_invalidate_folio,
-#endif
+       .dirty_folio = filemap_dirty_folio,
        .writepages = ecryptfs_writepages,

This clearly addresses the build failure as well, so no objections from me, but I don't understand what the functional difference is here and would rely on you to write a changelog text for that change.

Arnd

#3 lore raw

Re: ecryptfs is unmaintained and untested

Quoted 5 lines · Arnd Bergmann
Quoted 3 lines · Matthew Wilcox

This comment has been there since June 2021, so I think we can just delete ecryptfs now?

Quoted 7 lines

I have no opinion on removing ecryptfs, but I don't how possibly removing it is related to the patch I sent, as far as I can tell it just means it relies on both CONFIG_BLOCK and CONFIG_BUFFER_HEAD then.

Is there any indication that the last users that had files on ecryptfs are unable to update their kernels?

Debian is still shipping ecryptfs-utils and is building and including the ecryptfs kernel module in their distro kernel.`

So it seems likely that there are probably a non-zero (although probably relatively small) number of ecryptfs users out there.

- Ted

maintainer
#4 lore raw

Re: ecryptfs is unmaintained and untested

Quoted 11 lines · Theodore Ts'o
Quoted 10 lines · Arnd Bergmann
Quoted 3 lines · Matthew Wilcox

This comment has been there since June 2021, so I think we can just delete ecryptfs now?

I have no opinion on removing ecryptfs, but I don't how possibly removing it is related to the patch I sent, as far as I can tell it just means it relies on both CONFIG_BLOCK and CONFIG_BUFFER_HEAD then.

Quoted 8 lines
Quoted 2 lines

Is there any indication that the last users that had files on ecryptfs are unable to update their kernels?

Debian is still shipping ecryptfs-utils and is building and including the ecryptfs kernel module in their distro kernel.`

So it seems likely that there are probably a non-zero (although probably relatively small) number of ecryptfs users out there.

It would be good to discuss how we can get the message out to users to migrate off of eCryptfs so that functionality can be reduced and eventually it can be removed.

What do folks think about the following?

1. Print loud warnings at mount time that eCryptfs is deprecated and give a specific date when write support will be removed.
2. Remove write support at that date, while retaining read-only support to allow any lagging users to move their data to fscrypt or other alternatives.
3. Print loud warnings at mount that eCryptfs will be removed and give a specific date.
4. Remove it.

Suggestions on lead times for #2 and #4 would be appreciated.

Tyler

#5 lore raw

Re: ecryptfs is unmaintained and untested

Quoted 16 lines

It would be good to discuss how we can get the message out to users to migrate off of eCryptfs so that functionality can be reduced and eventually it can be removed.

What do folks think about the following?

1. Print loud warnings at mount time that eCryptfs is deprecated and give a specific date when write support will be removed.
2. Remove write support at that date, while retaining read-only support to allow any lagging users to move their data to fscrypt or other alternatives.
3. Print loud warnings at mount that eCryptfs will be removed and give a specific date.
4. Remove it.

Quoted 1 line

Suggestions on lead times for #2 and #4 would be appreciated.

FWIW. Ever since Ubuntu dropped eCryptfs home-directory encryption already in I believe 18.04 from now 8 years ago, Linux Mint has probably been the single biggest consumer of eCryptfs: while based on Ubuntu, they've added back eCryptfs home-directory encryption in the installer ever since Ubuntu dropped it.

In the Mint project lead's latest blog he/they moreover proudly announced that their upcoming new & shiny GUI user-management tool will now support enabling home-directory also for at runtime added accounts (which up to now is done with adduser --encrypt-home from the command line only) rather than just the first, installer- created user account, and I am pretty much certain that the idea there still is *eCryptfs* home-directory-encryption. Mint is not the kind of project that routinely goes low-level.

https://blog.linuxmint.com/?p=4991

In any case then this is to say that they seem to not be planning on moving off of eCryptfs -- and Linux Mint is (probably?) the most popular Linux Distribution for/among new Linux users.

Have been commenting on this a bit as well on their forum, advocating for fully doing away with eCryptfs, but quite unsure that'll have any meaningful impact. Added the Mint project lead to the CC on this. If there's any progress to report on slashing eCryptfs he/they may want to be aware.

Regards,
Rene

#6 lore raw

Re: ecryptfs is unmaintained and untested

In reply to Theodore Ts'o
Quoted 11 lines · Theodore Ts'o
Quoted 10 lines · Arnd Bergmann
Quoted 3 lines · Matthew Wilcox

This comment has been there since June 2021, so I think we can just delete ecryptfs now?

I have no opinion on removing ecryptfs, but I don't how possibly removing it is related to the patch I sent, as far as I can tell it just means it relies on both CONFIG_BLOCK and CONFIG_BUFFER_HEAD then.

Quoted 8 lines
Quoted 2 lines

Is there any indication that the last users that had files on ecryptfs are unable to update their kernels?

Debian is still shipping ecryptfs-utils and is building and including the ecryptfs kernel module in their distro kernel.`

So it seems likely that there are probably a non-zero (although probably relatively small) number of ecryptfs users out there.

Yeah. Sadly I'm one, as I needed something to migrate off of when encfs was deprecated.

Is there another soon-to-be-deprecated filesystem to encrypt directories I should move to? :)

I definitely think we need some loud warnings and Tylers' suggestion for a read-only grace period would be helpful.

thanks
-john

#7 lore raw

Re: ecryptfs is unmaintained and untested

Quoted 6 lines · John Stultz

Yeah. Sadly I'm one, as I needed something to migrate off of when encfs was deprecated.

Is there another soon-to-be-deprecated filesystem to encrypt directories I should move to? :)

Well, the closest way of encrypting directories is fscrypt. The good news is that it works on top of btrfs, ext4, f2fs, and ubifs, and it's not likely to be deprecated given that it is used by chromeos and android. The bad news is that the integration with traditional Linux desktop setups (e.g., login, etc.) was never completed.

This is probably because for many desktop and server configurations, using dm-crypt is actually better suited and more secure. It certainly doesn't solve the "just encrypt a directory hierarchy in a file system" and the "support multiple users' who might have different encryption keys and which are mutually suspicious" use cases. But this appears to not be sufficiently interesting for distributions to do that integration work.

- Ted

#8 lore raw

Re: ecryptfs is unmaintained and untested

Quoted 5 lines · Theodore Ts'o
Quoted 4 lines · John Stultz

Yeah. Sadly I'm one, as I needed something to migrate off of when encfs was deprecated.

Quoted 8 lines
Quoted 2 lines

Is there another soon-to-be-deprecated filesystem to encrypt directories I should move to? :)

Well, the closest way of encrypting directories is fscrypt. The good news is that it works on top of btrfs, ext4, f2fs, and ubifs, and it's not likely to be deprecated given that it is used by chromeos and android. The bad news is that the integration with traditional Linux desktop setups (e.g., login, etc.) was never completed.

Yeah, though to my understanding fscrypt complicates backing up the data in its encrypted form.
Having the fuse/overlay encryption approaches is nice because you can just backup the underlying filesystem layer and ignore the overlay mount.

Quoted 7 lines

This is probably because for many desktop and server configurations, using dm-crypt is actually better suited and more secure. It certainly doesn't solve the "just encrypt a directory hierarchy in a file system" and the "support multiple users' who might have different encryption keys and which are mutually suspicious" use cases. But this appears to not be sufficiently interesting for distributions to do that integration work.

Mostly I avoid dm-crypt for personal files as I want the majority of things (family pictures, etc) to be as simply recoverable as possible. It's only for a small amount of things like email archives and tax/financial documents that I'd like to have it be non-trivial to access if my backup drive or desktop was stolen.

I've wondered if maybe something as simple as fuse mounting a password protected zip file would do, but I'm guessing something a little more modern like a fuse + age approach would be better. Unfortunately I'm not finding anything so far.

thanks
-john

#9 lore raw

Re: ecryptfs is unmaintained and untested

John Stultz - 14.10.25, 18:38:52 CEST:

Quoted 5 lines

Mostly I avoid dm-crypt for personal files as I want the majority of things (family pictures, etc) to be as simply recoverable as possible. It's only for a small amount of things like email archives and tax/financial documents that I'd like to have it be non-trivial to access if my backup drive or desktop was stolen.

See my hints about CryFS or gocryptfs in my other mail as used by Plasma Vault. I believe it might suit your use case quite well.

Signature · 1 line
Martin
#10 lore raw

Re: ecryptfs is unmaintained and untested

In reply to John Stultz
Quoted 2 lines · John Stultz

Yeah, though to my understanding fscrypt complicates backing up the data in its encrypted form.

Unfortunately, yes, that's correct. Michael and I did throw around a rough design for doing encrypted backups and saving the encrypted per-file encryption key. Actually doing the _backup_ wasn't that difficult; but doing the *restore* was very tricky/painful. Ultimately, we never implemented it because it wasn't necessarily for the Android/ChromeOS use case, and because we weren't getting a lot of interest for the desktop, without which having a better general-purpose backup is lower priority.

Quoted 4 lines

I've wondered if maybe something as simple as fuse mounting a password protected zip file would do, but I'm guessing something a little more modern like a fuse + age approach would be better. Unfortunately I'm not finding anything so far.

Darrick is doing a lot of work to significantly improve the performance of fuse2fs. So perhaps fuse mounting a dm-crypt device backed by a loop device might be a possibility?

- Ted

#11 lore raw

Re: ecryptfs is unmaintained and untested

In reply to Theodore Ts'o

Hi.

Theodore Ts'o - 14.10.25, 16:39:16 CEST:

Quoted 7 lines

This is probably because for many desktop and server configurations, using dm-crypt is actually better suited and more secure. It certainly doesn't solve the "just encrypt a directory hierarchy in a file system" and the "support multiple users' who might have different encryption keys and which are mutually suspicious" use cases. But this appears to not be sufficiently interesting for distributions to do that integration work.

If it is just about encrypting a sub directory of the home directory there has been work to support that on Plasma desktop via Plasma Vault. It supports CryFS as default, EncFS (with a security warning about it) and gocryptfs. CryFS is interesting as it also obfuscates the directory hierarchy as well as object names. All of them are FUSE filesystems.

Maybe one of these – excluding EncFS – could be used for encrypting the complete home directory of a user. Preferably CryFS maybe. But I bet it will be quite a bit slower than ecryptfs¹. And I am not aware of any other desktop or distribution integration work regarding CryFS, gocryptfs or another alternative.

[1] "The increase in security when compared to other file systems comes at a performance cost. CryFS is fast enough to be used in practice. I'm getting a read speed to 170MB/s and a write speed of 80MB/s on my SSD machine, but other file systems are even faster."

https://www.cryfs.org/comparison

Best,

Signature · 1 line
Martin
#12 lore raw

Re: ecryptfs is unmaintained and untested

In reply to Theodore Ts'o
Quoted 5 lines · Theodore Ts'o
Quoted 4 lines · John Stultz

Yeah. Sadly I'm one, as I needed something to migrate off of when encfs was deprecated.

Quoted 8 lines
Quoted 2 lines

Is there another soon-to-be-deprecated filesystem to encrypt directories I should move to? :)

Well, the closest way of encrypting directories is fscrypt. The good news is that it works on top of btrfs, ext4, f2fs, and ubifs, and it's not likely to be deprecated given that it is used by chromeos and android. The bad news is that the integration with traditional Linux desktop setups (e.g., login, etc.) was never completed.

The current set of filesystems that support fscrypt is ext4, f2fs, ubifs, cephfs, and (out-of-tree) Lustre. btrfs's support for fscrypt is still under development, I'm afraid. I'm told it's starting to be worked on again.

While the main user of the fscrypt kernel feature is Android which has its own userspace, there's also a userspace tool for general-purpose Linux distros, also called fscrypt. See
https://github.com/google/fscrypt and
https://wiki.archlinux.org/title/Fscrypt

I've been maintaining the 'fscrypt' userspace tool, and in the past I've done quite a bit of work to improve it. I also use it to encrypt the home directory on my personal desktop.

It's true that it really could use some love, though. It's not something that I've been prioritizing recently, and no one else has stepped up either. (Compare to eCryptfs where Ubuntu adopted it, and Canonical stepped up to develop and maintain ecryptfs-tools. That's just not something that's happened for fscrypt.)

There are other userspace programs that use the fscrypt kernel feature too, such as systemd-homed and a new one called dirlock:
https://lwn.net/Articles/1038859/

- Eric

#13 lore raw

Re: ecryptfs is unmaintained and untested

Quoted 3 lines · Eric Biggers

I've been maintaining the 'fscrypt' userspace tool, and in the past I've done quite a bit of work to improve it. I also use it to encrypt the home directory on my personal desktop.

Do you have integration with PAM? I assume you must if you are encrypting your home directory, since there would need to have some way to insert the key to unlock your home directory as part of single-sign on. Or are you only encrypting something like the Private directory on your home directory, and entering the password separately fromo the login password.

One of the really nice things of the ecryptfs integration, especialy on Ubuntu (and I don't remember if Tyler was responsible for that work; if so, kudos!) wasthat it was particularly eeamless.

- Ted

#14 lore raw

Re: ecryptfs is unmaintained and untested

Quoted 12 lines · Theodore Ts'o
Quoted 3 lines · Eric Biggers

I've been maintaining the 'fscrypt' userspace tool, and in the past I've done quite a bit of work to improve it. I also use it to encrypt the home directory on my personal desktop.

Do you have integration with PAM? I assume you must if you are encrypting your home directory, since there would need to have some way to insert the key to unlock your home directory as part of single-sign on. Or are you only encrypting something like the Private directory on your home directory, and entering the password separately fromo the login password.

Quoted 3 lines

One of the really nice things of the ecryptfs integration, especialy on Ubuntu (and I don't remember if Tyler was responsible for that work; if so, kudos!) wasthat it was particularly eeamless.

Yes, there's a PAM module called pam_fscrypt.

- Eric