[edk2-devel] RFC: EXT4 filesystem driver

Fri Jul 23 03:59:57 UTC 2021

> On Jul 22, 2021, at 7:07 PM, Desimone, Nathaniel L <nathaniel.l.desimone at intel.com> wrote:
> 
> Hi Pedro,
> 
>> -----Original Message-----
>> From: devel at edk2.groups.io <mailto:devel at edk2.groups.io> <devel at edk2.groups.io <mailto:devel at edk2.groups.io>> On Behalf Of Pedro
>> Falcato
>> Sent: Thursday, July 22, 2021 9:54 AM
>> To: Andrew Fish <afish at apple.com <mailto:afish at apple.com>>
>> Cc: edk2-devel-groups-io <devel at edk2.groups.io <mailto:devel at edk2.groups.io>>; mhaeuser at posteo.de <mailto:mhaeuser at posteo.de>;
>> rfc at edk2.groups.io <mailto:rfc at 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 at apple.com> wrote:
>>> 
>>> 
>>> 
>>> On Jul 22, 2021, at 3:24 AM, Marvin Häuser <mhaeuser at 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
>> 
>> 
>> 

-=-=-=-=-=-=-=-=-=-=-=-
Groups.io Links: You receive all messages sent to this group.
View/Reply Online (#78117): https://edk2.groups.io/g/devel/message/78117
Mute This Topic: https://groups.io/mt/84368561/1813853
Group Owner: devel+owner at edk2.groups.io
Unsubscribe: https://edk2.groups.io/g/devel/unsub [edk2-devel-archive at redhat.com]
-=-=-=-=-=-=-=-=-=-=-=-

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/edk2-devel-archive/attachments/20210722/ae2062c5/attachment.htm>