Filesystems and Ramfs #459
Conversation
Co-authored-by: Benedikt Weber <[email protected]>
This commit is the merged version with upstream, fix coming later
Co-authored-by: SqYtCO <[email protected]>
This mainly includes: - `sed -i 's/KernelResult/Result/g' rust/kernel/**/*.rs` - fixing some `CStr`s - adding small things I missed when rebasing I tested with the VM and everything seems to work. Namely, ```shell $ ./run.sh $ echo a > bs3/a $ cat bs3/a a ``` worked, so it should all be good. I think there are some new compiler warnings, we'll have to see through them some other time :^)
I converted some names to their intended casing and put some underscores before unused variable names. I also added some Rust style where I saw something odd.
This also replaces some uses of bindings::mode_t with the Mode type from the kernel crate. There are still some things that require another indirection to be ported to Rust types, but most of the code we have now uses the new type. I think that wrapping the `i_mapping` field could be difficult to get right.
This also includes the first wrapper, `generic_file_llseek`.
This method should be preferred over accessing the member variable directly, because it enforces an implementation of `file_operations::FileOperations`. For now, we added a `FileOpenAdapter` that doesn't return anything, because we currently don't need anything more complicated.
This commit adds the call to `Inode::set_file_operations`, the `FileOperations` implementation is not done yet though.
Previously, the kernel crate implemented `read_iter` in terms of `read`. Now, the default implementation of the trait method does just that, but it can also be overridden.
This moves the shared behavior of the `build` method from the file
operations vtable into its own trait, so that other types can also
implement it. Namely, we added the unit struct `SimpleDirOperations`
that returns a ref to the C global.
This could be expanded, I can see a general `trait BuildVtable<Table>`
and a macro like
declare_c_vtable!(
SimpleDirOperations,
bindings::file_operations,
bindings::simple_dir_operations,
);
that expands to
struct SimpleDirOperations;
impl BuildVtable<bindings::file_operations> for SimpleDirOperations {
fn build() -> &'static bindings::file_operations {
&bindings::simple_dir_operations
}
}
so that we can easily make these constants accessible. The downside is,
that the interface for the caller is more abstract / complicated, since
the vtable struct has to be constructed on the calling site
.set_file_operations::<FileOperationsVtable<NopFileOpenAdapter, Bs2RamfsFileOps>>();
We could however aid with some helper methods
.set_file_ops_from_trait::<Bs2RamfsFileOps>();
// ...
fn set_file_ops_from_trait<Ops: FileOperations<A::Arg>, A: FileOpenAdapter = NopFileOpenAdapter>() { ... }
Implementing the `BuildVtable` trait for all implementors of
`FileOperations` is not possible, because there might be multiple types
`A` that satisfy the requirements, thus making a definite choice
impossible.
This macro mirrors the `std::dbg` macro. To view the output, you might have to set the log level to "info" using `dmesg -n 6`.
…was just pasted from FileOperations
Partly abridged from the fat branch Co-authored-by: Tamino Bauknecht <[email protected]>
Recent C implementations of ramfs don't implement it either
Co-authored-by: Niklas Mohrin <[email protected]>
I really tried to write a good FileSystemFlags abstraction. But that damn const static stuff... argh... So, I guess we have to live with this :/
We already have a macro for flag types in ec61599 We could also try to make this as one trait that gives you free impls of BitAnd and BitOr etc., and maybe add a derive(BitFlag) or so (we would need one method that returns self.0 on these types). We could then also have an associated type with the trait, so that we can also get rid of ModeInt. I don't know how much we loose on the "const" side of things with a trait though.
When you do it, please follow There seems to be also a few files that shouldn't be in the PR, e.g. the
Definitely :) |
| #[macro_export] | ||
| macro_rules! declare_fs_type { | ||
| ($T:ty, $S:ident) => { | ||
| static mut $S: $crate::bindings::file_system_type = $crate::bindings::file_system_type { |
There was a problem hiding this comment.
Where is this static mutated? Is it implicitly synchronized, or does it need a mutex?
There was a problem hiding this comment.
We need a *mut bindings::file_system_type for mount_nodev and I think we faced problems when this was just a static (are normal statics in read-only memory?). In the C impl, this was just a (non-const) global variable and we figured this would be the closest match. The vtable approach from FileOperations doesn't work either because the fields are mutated in C code
Maybe we are missing something though
There was a problem hiding this comment.
IIUC, This is only used for fs registration. It will not be mutated later from fs side.
|
Great work! I'm more than happy to work together. I was moving really slow on my own and I'm gonna stop that work and come here to work on this. What is your workflow?? Is there any work that I can immediately start working on it?? |
|
@foxhlchen We would start with the rebase in the next couple of days. We usually work in pairs and would meet up every week in a set spot, since the lecture is over (and exams have started), we didn't stick to the schedule the last two weeks or so. I think that we would be fine with going back to that workflow and working in groups of 1-3. I was thinking we could move conversations to a zulip thread and maybe discuss progress in a jitsi or similar if that is okay for you (we are not on zulip yet, gonna request an invitation later today) I am not really sure which existing parts need work right now, but afaik no one has started working on a |
Yes, In my opinion, this should be done first. Some have been changed since you've forked the code (e.g. Error). We need to use what
Sounds great! We can have regular meetups or meetups as needed. Apart from that, the email also works for me [email protected] as well. Is jitsi a zoom like product? I think that will work fine for me.
Ok, I will look into it! |
|
Sadly, we never got around to properly finishing this, closing :) |
Closes #369
WIP
Hey everyone, we are now starting to shift our focus away from coding for the lecture and can start working on your requirements. :)
There are some TODOs already
FileSystemRegistration(+ macro for module?)static mutwe have forstruct filesystem_type(maybe rather staticUnsafeCell?)*Operationsand considervtablemacro from Implement a#[vtable]macro #322FileSystemBase -> FileSystemOperationsdeclare_*_operationsmacro to allow for something likedeclare_file_operations(stat=simple_statfs)(seems very hard, but maybe we can come up with a nice way somehow); or other suggestions for how to get rid of the huge boilerplate for things where the C implementation just writes a function into the vtable, whereas we need to enumerate every argumentprivate/ in animpl FileOperations, but failed to come up with an elegant solution (elegant meaning nounsafe)). I saw thatPinalready implementsPointerWrapper, but we will have to see how it all plays out - maybe we will even ignore pinning for now@foxhlchen We gladly invite you to work this out together :)
Reviewers: There is probably no point in reviewing until we have rebased at least :^)