Re: [edk2-devel] RFC: EXT4 filesystem driver


Andrew Fish <afish@...>
 

On Jul 22, 2021, at 7:07 PM, Desimone, Nathaniel L <nathaniel.l.desimone@intel.com> wrote:

Hi Pedro,

-----Original Message-----
From: devel@edk2.groups.io <mailto:devel@edk2.groups.io> <devel@edk2.groups.io <mailto:devel@edk2.groups.io>> On Behalf Of Pedro
Falcato
Sent: Thursday, July 22, 2021 9:54 AM
To: Andrew Fish <afish@apple.com <mailto:afish@apple.com>>
Cc: edk2-devel-groups-io <devel@edk2.groups.io <mailto:devel@edk2.groups.io>>; mhaeuser@posteo.de <mailto:mhaeuser@posteo.de>;
rfc@edk2.groups.io <mailto:rfc@edk2.groups.io>
Subject: Re: [edk2-devel] RFC: EXT4 filesystem driver

Hi Andrew, Marvin,

RE: The package name: It doesn't sound like a bad idea to have something
like a FileSystemPkg and have a bunch of different filesystems inside of it,
but I'll defer to you and my mentors' judgement; we could also drop that
issue for now and take care of it afterwards, since it may need further
changes that are not a part of GSoC and would just delay the process.

With respect to the write capabilities of the driver, I'm not entirely sure
whether or not it's useful. I've been thinking about it today, and it seems like
there's not much that could go wrong? The write path isn't excessively
complex. Except of course in the event of an untimely power cut, but those
/should/ be easily detected by the checksums. My initial idea was to have it
up to speed with FatPkg and other filesystems by implementing all of
EFI_FILE_PROTOCOL, including the write portions. If Apple's HFS+ and APFS
drivers don't have those, it may be a decent idea to reduce the scope of the
ext4 driver as well. I don't see a big need for write support; on the other
hand, I've only worked on UEFI bootloaders before, which may be an outlier
in that regard. Further feedback is appreciated.
The most commonly used reason to for writing to the filesystem in a production environment is capsule updates. Most capsule update implementations will stage the capsule on the EFI System Partition and then reset the system to unlock flash. The second most useful is the UEFI Shell and all the many applications that run within it will write to the filesystem for a large variety of reasons. I think it would be a useful feature to have as one could conceivably start using EFI System Partitions formatted as ext4.
The EFI System Partition is defined to be FAT32 by the UEFI Spec for interoperability. It defines the file system drivers required for the firmware and OS. So changing that is not really an option.

You can still install the UEFI Shell to a read only file system, you just need to do it from the OS :). We actually do this on Macs quite often. You just run the macOS bless command and reboot to the UEFI Shell.

Thanks,

Andrew Fish


As for the tests, UEFI SCTs already seem to have some tests on
EFI_FILE_PROTOCOL's. Further testing may require some sort of fuzzing,
which is what I want to, although in a simplified way. With fuzzing we could
hammer the filesystem code with all sorts of different calls in different
orders, we could also mutate the disk structures to test if the driver is secure
and can handle corruption in a nice, safe way. A future (GSoC or not) project
could also attempt to use compiler-generated coverage instrumentation (see
LLVM's LibFuzzer and SanitizerCoverage for an example).

I'm not sure about all OSes, but at least Linux ext2/3/4 drivers are very robust
and can handle and work around any corrupted FS I
(accidentally) throw at them. However, running fsck is the best way to detect
corruption; note that licensing may be an issue as, for example, ext4's fsck is
GPL2 licensed.

Best Regards,
Pedro

On Thu, 22 Jul 2021 at 16:58, Andrew Fish <afish@apple.com> wrote:



On Jul 22, 2021, at 3:24 AM, Marvin Häuser <mhaeuser@posteo.de> wrote:

On 22.07.21 01:12, Pedro Falcato wrote:

EXT4 (fourth extended filesystem) is a filesystem developed for Linux
that has been in wide use (desktops, servers, smartphones) since 2008.

The Ext4Pkg implements the Simple File System Protocol for a partition
that is formatted with the EXT4 file system. This allows UEFI Drivers,
UEFI Applications, UEFI OS Loaders, and the UEFI Shell to access files
on an EXT4 partition and supports booting a UEFI OS Loader from an
EXT4 partition.
This project is one of the TianoCore Google Summer of Code projects.

Right now, Ext4Pkg only contains a single member, Ext4Dxe, which is a
UEFI driver that consumes Block I/O, Disk I/O and (optionally) Disk
I/O 2 Protocols, and produces the Simple File System protocol. It
allows mounting ext4 filesystems exclusively.

Brief overhead of EXT4:
Layout of an EXT2/3/4 filesystem:
(note: this driver has been developed using
https://www.kernel.org/doc/html/latest/filesystems/ext4/index.html as
documentation).

An ext2/3/4 filesystem (here on out referred to as simply an ext4
filesystem, due to the similarities) is composed of various concepts:

1) Superblock
The superblock is the structure near (1024 bytes offset from the
start) the start of the partition, and describes the filesystem in general.
Here, we get to know the size of the filesystem's blocks, which
features it supports or not, whether it's been cleanly unmounted, how
many blocks we have, etc.

2) Block groups
EXT4 filesystems are divided into block groups, and each block group
covers
s_blocks_per_group(8 * Block Size) blocks. Each block group has an
associated block group descriptor; these are present directly after
the superblock. Each block group descriptor contains the location of
the inode table, and the inode and block bitmaps (note these bitmaps
are only a block long, which gets us the 8 * Block Size formula covered
previously).

3) Blocks
The ext4 filesystem is divided into blocks, of size s_log_block_size ^ 1024.
Blocks can be allocated using individual block groups's bitmaps. Note
that block 0 is invalid and its presence on extents/block tables means
it's part of a file hole, and that particular location must be read as
a block full of zeros.

4) Inodes
The ext4 filesystem divides files/directories into inodes (originally
index nodes). Each file/socket/symlink/directory/etc (here on out
referred to as a file, since there is no distinction under the ext4
filesystem) is stored as a /nameless/ inode, that is stored in some
block group's inode table. Each inode has s_inode_size size (or
GOOD_OLD_INODE_SIZE if it's an old filesystem), and holds various
metadata about the file. Since the largest inode structure right now
is ~160 bytes, the rest of the inode contains inline extended
attributes. Inodes' data is stored using either data blocks (under ext2/3) or
extents (under ext4).

5) Extents
Ext4 inodes store data in extents. These let N contiguous logical
blocks that are represented by N contiguous physical blocks be
represented by a single extent structure, which minimizes filesystem
metadata bloat and speeds up block mapping (particularly due to the
fact that high-quality
ext4 implementations like linux's try /really/ hard to make the file
contiguous, so it's common to have files with almost 0 fragmentation).
Inodes that use extents store them in a tree, and the top of the tree
is stored on i_data. The tree's leaves always start with an
EXT4_EXTENT_HEADER and contain EXT4_EXTENT_INDEX on eh_depth != 0
and
EXT4_EXTENT on eh_depth = 0; these entries are always sorted by
logical block.

6) Directories
Ext4 directories are files that store name -> inode mappings for the
logical directory; this is where files get their names, which means
ext4 inodes do not themselves have names, since they can be linked
(present) multiple times with different names. Directories can store
entries in two different ways:
1) Classical linear directories: They store entries as a mostly-linked
mostly-list of EXT4_DIR_ENTRY.
2) Hash tree directories: These are used for larger directories, with
hundreds of entries, and are designed in a backwards-compatible way.
These are not yet implemented in the Ext4Dxe driver.

7) Journal
Ext3/4 filesystems have a journal to help protect the filesystem
against system crashes. This is not yet implemented in Ext4Dxe but is
described in detail in the Linux kernel's documentation.

The EDK2 implementation of ext4 is based only on the public
documentation available at
https://www.kernel.org/doc/html/latest/filesystems/ext4/index.html
and
the FreeBSD ext2fs driver (available at
https://github.com/freebsd/freebsd-src/tree/main/sys/fs/ext2fs,
BSD-2-Clause-FreeBSD licensed). It is licensed as
SPDX-License-Identifier: BSD-2-Clause-Patent.

After a brief discussion with the community, the proposed package
location is edk2-platform/Features/Ext4Pkg (relevant discussion:
https://edk2.groups.io/g/devel/topic/83060185).

I was the main contributor and I would like to maintain the package in
the future, if possible.


While I personally don't like it's outside of the EDK II core, I kind of get it.
However I would strongly suggest to choose a more general package name,
like "LinuxFsPkg", or "NixFsPkg", or maybe even just "FileSystemPkg" (and
move FAT over some day?). Imagine someone wants to import BTRFS next
year, should it really be "BtrfsPkg"? I understand it follows the "FatPkg"
convention, but I feel like people forget FatPkg was special as to its awkward
license before Microsoft allowed a change a few years ago. Maintainers.txt
already has the concept of different Reviewers per subfolder, maybe it could
be extended a little to have a common EDK II contributor to officially maintain
the package, but have you be a Maintainer or something like a Reviewer+ to
your driver? Or you could maintain the entire package of course.


Marvin,

Good point that the FatPkg was more about license boundary than
anything else, so I’m not opposed to a more generic package name.

Current limitations:
1) The Ext4Dxe driver is, at the moment, read-only.
2) The Ext4Dxe driver at the moment cannot mount older (ext2/3)
filesystems. Ensuring compatibility with those may not be a bad idea.

I intend to test the package using the UEFI SCTs present in edk2-test,
and implement any other needed unit tests myself using the already
available unit test framework. I also intend to (privately) fuzz the
UEFI driver with bad/unusual disk images, to improve the security and
reliability of the driver.

In the future, ext4 write support should be added so edk2 has a
fully-featured RW ext4 driver. There could also be a focus on
supporting the older ext4-like filesystems, as I mentioned in the
limitations, but that is open for discussion.


I may be alone, but I strongly object. One of our projects (OpenCore) has a
disgusting way of writing files because the FAT32 driver in Aptio IV firmwares
may corrupt the filesystem when resizing files. To be honest, it may corrupt
with other usages too and we never noticed, because basically we wrote the
code to require the least amount of (complex) FS operations.

The issue with EDK II is, there is a lot of own code and a lot of users, but
little testing. By that I do not mean that developers do not test their code,
but that nobody sits down and performs all sorts of FS manipulations in all
sorts of orders and closely observes the result for regression-testing. Users
will not really test it either, as UEFI to them should just somehow boot to
Windows. If at least the code was shared with a codebase that is known-
trusted (e.g. the Linux kernel itself), that'd be much easier to trust, but
realistically this is not going to happen.
My point is, if a company like AMI cannot guarantee writing does not
damage the FS for a very simple FS, how do you plan to guarantee yours
doesn't for a much more complex FS? I'd rather have only one simple FS type
that supports writing for most use-cases (e.g. logging).

At the very least I would beg you to have a PCD to turn write support
off - if it will be off by default, that'd be great of course. :) Was there any
discussion yet as to why write support is needed in the first place you could
point me to?


I think having a default PCD option of read only is a good idea.

EFI on Mac carries HFS+ and APFS EFI file system drivers and both of those
are read only for safety, security, and to avoid the need to validate them. So I
think some products may want to have the option to ship read only versions
of the file system.

Seems like having EFI base file system tests would be useful. I’d imaging
with OVMF it would be possible to implement a very robust test
infrastructure. Seems like the hard bits would be generating the test cases
and figuring out how to validate the tests did the correct thing. I’m guess the
OS based file system drivers are robust and try to work around bugs
gracefully? Maybe there is a way to turn on OS logging, or even run an OS
based fsck on the volume after the tests complete. Regardless this seems
like an interesting project, maybe we can add it to next years GSoC?

Thanks,

Andrew Fish

Thanks for your work!

Best regards,
Marvin

The driver's handling of unclean unmounting through forced shutdown is
unclear.
Is there a position in edk2 on how to handle such cases? I don't think
FAT32 has a "this filesystem is/was dirty" and even though it seems to
me that stopping a system from booting/opening the partition because
"we may find some tiny irregularities" is not the best course of
action, I can't find a clear answer.

The driver also had to add implementations of CRC32C and CRC16, and
after talking with my mentor we quickly reached the conclusion that
these may be good candidates for inclusion in MdePkg. We also
discussed moving the Ucs2 <-> Utf8 conversion library in RedfishPkg
(BaseUcs2Utf8Lib) into MdePkg as well. Any comments?

Feel free to ask any questions you may find relevant.

Best Regards,

Pedro Falcato










--
Pedro Falcato


Join rfc@edk2.groups.io to automatically receive all group messages.