| Message ID | 20240209-alice-file-v5-3-a37886783025@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | File abstractions needed by Rust Binder | expand |
On Fri, Feb 09, 2024 at 11:18:16AM +0000, Alice Ryhl wrote: > From: Wedson Almeida Filho <wedsonaf@gmail.com> > > This abstraction makes it possible to manipulate the open files for a > process. The new `File` struct wraps the C `struct file`. When accessing > it using the smart pointer `ARef<File>`, the pointer will own a > reference count to the file. When accessing it as `&File`, then the > reference does not own a refcount, but the borrow checker will ensure > that the reference count does not hit zero while the `&File` is live. > > Since this is intended to manipulate the open files of a process, we > introduce an `fget` constructor that corresponds to the C `fget` > method. In future patches, it will become possible to create a new fd in > a process and bind it to a `File`. Rust Binder will use these to send > fds from one process to another. > > We also provide a method for accessing the file's flags. Rust Binder > will use this to access the flags of the Binder fd to check whether the > non-blocking flag is set, which affects what the Binder ioctl does. > > This introduces a struct for the EBADF error type, rather than just > using the Error type directly. This has two advantages: > * `File::from_fd` returns a `Result<ARef<File>, BadFdError>`, which the Sorry, where's that method? > compiler will represent as a single pointer, with null being an error. > This is possible because the compiler understands that `BadFdError` > has only one possible value, and it also understands that the > `ARef<File>` smart pointer is guaranteed non-null. > * Additionally, we promise to users of the method that the method can > only fail with EBADF, which means that they can rely on this promise > without having to inspect its implementation. > That said, there are also two disadvantages: > * Defining additional error types involves boilerplate. > * The question mark operator will only utilize the `From` trait once, > which prevents you from using the question mark operator on > `BadFdError` in methods that return some third error type that the > kernel `Error` is convertible into. (However, it works fine in methods > that return `Error`.) > > Signed-off-by: Wedson Almeida Filho <wedsonaf@gmail.com> > Co-developed-by: Daniel Xu <dxu@dxuuu.xyz> > Signed-off-by: Daniel Xu <dxu@dxuuu.xyz> > Co-developed-by: Alice Ryhl <aliceryhl@google.com> > Reviewed-by: Martin Rodriguez Reboredo <yakoyoku@gmail.com> > Reviewed-by: Trevor Gross <tmgross@umich.edu> > Reviewed-by: Benno Lossin <benno.lossin@proton.me> > Signed-off-by: Alice Ryhl <aliceryhl@google.com> > --- > fs/file.c | 7 ++ > rust/bindings/bindings_helper.h | 2 + > rust/helpers.c | 7 ++ > rust/kernel/file.rs | 254 ++++++++++++++++++++++++++++++++++++++++ > rust/kernel/lib.rs | 1 + > 5 files changed, 271 insertions(+) > > diff --git a/fs/file.c b/fs/file.c > index 3b683b9101d8..f2eab5fcb87f 100644 > --- a/fs/file.c > +++ b/fs/file.c > @@ -1127,6 +1127,13 @@ EXPORT_SYMBOL(task_lookup_next_fdget_rcu); > * > * The fput_needed flag returned by fget_light should be passed to the > * corresponding fput_light. > + * > + * (As an exception to rule 2, you can call filp_close between fget_light and > + * fput_light provided that you capture a real refcount with get_file before > + * the call to filp_close, and ensure that this real refcount is fput *after* > + * the fput_light call.) > + * > + * See also the documentation in rust/kernel/file.rs. > */ > static unsigned long __fget_light(unsigned int fd, fmode_t mask) > { > diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h > index 936651110c39..41fcd2905ed4 100644 > --- a/rust/bindings/bindings_helper.h > +++ b/rust/bindings/bindings_helper.h > @@ -9,6 +9,8 @@ > #include <kunit/test.h> > #include <linux/errname.h> > #include <linux/ethtool.h> > +#include <linux/file.h> > +#include <linux/fs.h> > #include <linux/jiffies.h> > #include <linux/mdio.h> > #include <linux/phy.h> > diff --git a/rust/helpers.c b/rust/helpers.c > index 70e59efd92bc..03141a3608a4 100644 > --- a/rust/helpers.c > +++ b/rust/helpers.c > @@ -25,6 +25,7 @@ > #include <linux/build_bug.h> > #include <linux/err.h> > #include <linux/errname.h> > +#include <linux/fs.h> > #include <linux/mutex.h> > #include <linux/refcount.h> > #include <linux/sched/signal.h> > @@ -157,6 +158,12 @@ void rust_helper_init_work_with_key(struct work_struct *work, work_func_t func, > } > EXPORT_SYMBOL_GPL(rust_helper_init_work_with_key); > > +struct file *rust_helper_get_file(struct file *f) > +{ > + return get_file(f); > +} > +EXPORT_SYMBOL_GPL(rust_helper_get_file); > + > /* > * `bindgen` binds the C `size_t` type as the Rust `usize` type, so we can > * use it in contexts where Rust expects a `usize` like slice (array) indices. > diff --git a/rust/kernel/file.rs b/rust/kernel/file.rs > new file mode 100644 > index 000000000000..cf8ebf619379 > --- /dev/null > +++ b/rust/kernel/file.rs > @@ -0,0 +1,254 @@ > +// SPDX-License-Identifier: GPL-2.0 > + > +//! Files and file descriptors. > +//! > +//! C headers: [`include/linux/fs.h`](srctree/include/linux/fs.h) and > +//! [`include/linux/file.h`](srctree/include/linux/file.h) > + > +use crate::{ > + bindings, > + error::{code::*, Error, Result}, > + types::{ARef, AlwaysRefCounted, Opaque}, > +}; > +use core::ptr; > + > +/// Flags associated with a [`File`]. > +pub mod flags { > + /// File is opened in append mode. > + pub const O_APPEND: u32 = bindings::O_APPEND; > + > + /// Signal-driven I/O is enabled. > + pub const O_ASYNC: u32 = bindings::FASYNC; > + > + /// Close-on-exec flag is set. > + pub const O_CLOEXEC: u32 = bindings::O_CLOEXEC; > + > + /// File was created if it didn't already exist. > + pub const O_CREAT: u32 = bindings::O_CREAT; > + > + /// Direct I/O is enabled for this file. > + pub const O_DIRECT: u32 = bindings::O_DIRECT; > + > + /// File must be a directory. > + pub const O_DIRECTORY: u32 = bindings::O_DIRECTORY; > + > + /// Like [`O_SYNC`] except metadata is not synced. > + pub const O_DSYNC: u32 = bindings::O_DSYNC; > + > + /// Ensure that this file is created with the `open(2)` call. > + pub const O_EXCL: u32 = bindings::O_EXCL; > + > + /// Large file size enabled (`off64_t` over `off_t`). > + pub const O_LARGEFILE: u32 = bindings::O_LARGEFILE; > + > + /// Do not update the file last access time. > + pub const O_NOATIME: u32 = bindings::O_NOATIME; > + > + /// File should not be used as process's controlling terminal. > + pub const O_NOCTTY: u32 = bindings::O_NOCTTY; > + > + /// If basename of path is a symbolic link, fail open. > + pub const O_NOFOLLOW: u32 = bindings::O_NOFOLLOW; > + > + /// File is using nonblocking I/O. > + pub const O_NONBLOCK: u32 = bindings::O_NONBLOCK; > + > + /// File is using nonblocking I/O. > + /// > + /// This is effectively the same flag as [`O_NONBLOCK`] on all architectures > + /// except SPARC64. > + pub const O_NDELAY: u32 = bindings::O_NDELAY; > + > + /// Used to obtain a path file descriptor. > + pub const O_PATH: u32 = bindings::O_PATH; > + > + /// Write operations on this file will flush data and metadata. > + pub const O_SYNC: u32 = bindings::O_SYNC; > + > + /// This file is an unnamed temporary regular file. > + pub const O_TMPFILE: u32 = bindings::O_TMPFILE; > + > + /// File should be truncated to length 0. > + pub const O_TRUNC: u32 = bindings::O_TRUNC; > + > + /// Bitmask for access mode flags. > + /// > + /// # Examples > + /// > + /// ``` > + /// use kernel::file; > + /// # fn do_something() {} > + /// # let flags = 0; > + /// if (flags & file::flags::O_ACCMODE) == file::flags::O_RDONLY { > + /// do_something(); > + /// } > + /// ``` > + pub const O_ACCMODE: u32 = bindings::O_ACCMODE; > + > + /// File is read only. > + pub const O_RDONLY: u32 = bindings::O_RDONLY; > + > + /// File is write only. > + pub const O_WRONLY: u32 = bindings::O_WRONLY; > + > + /// File can be both read and written. > + pub const O_RDWR: u32 = bindings::O_RDWR; > +} > + > +/// Wraps the kernel's `struct file`. > +/// > +/// This represents an open file rather than a file on a filesystem. Processes generally reference > +/// open files using file descriptors. However, file descriptors are not the same as files. A file > +/// descriptor is just an integer that corresponds to a file, and a single file may be referenced > +/// by multiple file descriptors. > +/// > +/// # Refcounting > +/// > +/// Instances of this type are reference-counted. The reference count is incremented by the > +/// `fget`/`get_file` functions and decremented by `fput`. The Rust type `ARef<File>` represents a > +/// pointer that owns a reference count on the file. > +/// > +/// Whenever a process opens a file descriptor (fd), it stores a pointer to the file in its `struct > +/// files_struct`. This pointer owns a reference count to the file, ensuring the file isn't > +/// prematurely deleted while the file descriptor is open. In Rust terminology, the pointers in > +/// `struct files_struct` are `ARef<File>` pointers. > +/// > +/// ## Light refcounts > +/// > +/// Whenever a process has an fd to a file, it may use something called a "light refcount" as a > +/// performance optimization. Light refcounts are acquired by calling `fdget` and released with > +/// `fdput`. The idea behind light refcounts is that if the fd is not closed between the calls to > +/// `fdget` and `fdput`, then the refcount cannot hit zero during that time, as the `struct > +/// files_struct` holds a reference until the fd is closed. This means that it's safe to access the > +/// file even if `fdget` does not increment the refcount. > +/// > +/// The requirement that the fd is not closed during a light refcount applies globally across all > +/// threads - not just on the thread using the light refcount. For this reason, light refcounts are > +/// only used when the `struct files_struct` is not shared with other threads, since this ensures > +/// that other unrelated threads cannot suddenly start using the fd and close it. Therefore, > +/// calling `fdget` on a shared `struct files_struct` creates a normal refcount instead of a light > +/// refcount. When the fdget() calling task doesn't have a shared file descriptor table fdget() will not increment the reference count, yes. This also implies that you cannot have task A use fdget() and then pass f.file to task B that holds on to it while A returns to userspace. It's irrelevant that task A won't drop the reference count or that task B won't drop the reference count. Because task A could return back to userspace and immediately close the fd via a regular close() system call at which point task B has a UAF. In other words a file that has been gotten via fdget() can't be Send to another task without the Send implying taking a reference to it. > +/// > +/// Light reference counts must be released with `fdput` before the system call returns to > +/// userspace. This means that if you wait until the current system call returns to userspace, then > +/// all light refcounts that existed at the time have gone away. > +/// > +/// ## Rust references > +/// > +/// The reference type `&File` is similar to light refcounts: > +/// > +/// * `&File` references don't own a reference count. They can only exist as long as the reference > +/// count stays positive, and can only be created when there is some mechanism in place to ensure > +/// this. > +/// > +/// * The Rust borrow-checker normally ensures this by enforcing that the `ARef<File>` from which > +/// a `&File` is created outlives the `&File`. The section confuses me a little: Does the borrow-checker always ensure that a &File stays valid or are there circumstances where it doesn't or are you saying it doesn't enforce it? > +/// > +/// * Using the unsafe [`File::from_ptr`] means that it is up to the caller to ensure that the > +/// `&File` only exists while the reference count is positive. What is this used for in binder? If it's not used don't add it. > +/// > +/// * You can think of `fdget` as using an fd to look up an `ARef<File>` in the `struct Could you explain why there isn't an explicit fdget() then and you have that from_ptr() method? > +/// files_struct` and create an `&File` from it. The "fd cannot be closed" rule is like the Rust > +/// rule "the `ARef<File>` must outlive the `&File`". > +/// > +/// # Invariants > +/// > +/// * Instances of this type are refcounted using the `f_count` field. > +/// * If an fd with active light refcounts is closed, then it must be the case that the file > +/// refcount is positive until all light refcounts of the fd have been dropped. > +/// * A light refcount must be dropped before returning to userspace. > +#[repr(transparent)] > +pub struct File(Opaque<bindings::file>); > + > +// SAFETY: > +// - `File::dec_ref` can be called from any thread. > +// - It is okay to send ownership of `struct file` across thread boundaries. > +unsafe impl Send for File {} > + > +// SAFETY: All methods defined on `File` that take `&self` are safe to call even if other threads > +// are concurrently accessing the same `struct file`, because those methods either access immutable > +// properties or have proper synchronization to ensure that such accesses are safe. > +unsafe impl Sync for File {} > + > +impl File { > + /// Constructs a new `struct file` wrapper from a file descriptor. > + /// > + /// The file descriptor belongs to the current process. > + pub fn fget(fd: u32) -> Result<ARef<Self>, BadFdError> { > + // SAFETY: FFI call, there are no requirements on `fd`. > + let ptr = ptr::NonNull::new(unsafe { bindings::fget(fd) }).ok_or(BadFdError)?; > + > + // SAFETY: `bindings::fget` either returns null or a valid pointer to a file, and we > + // checked for null above. > + // > + // INVARIANT: `bindings::fget` creates a refcount, and we pass ownership of the refcount to > + // the new `ARef<File>`. > + Ok(unsafe { ARef::from_raw(ptr.cast()) }) > + } > + > + /// Creates a reference to a [`File`] from a valid pointer. > + /// > + /// # Safety > + /// > + /// The caller must ensure that `ptr` points at a valid file and that the file's refcount is > + /// positive for the duration of 'a. > + pub unsafe fn from_ptr<'a>(ptr: *const bindings::file) -> &'a File { > + // SAFETY: The caller guarantees that the pointer is not dangling and stays valid for the > + // duration of 'a. The cast is okay because `File` is `repr(transparent)`. > + // > + // INVARIANT: The safety requirements guarantee that the refcount does not hit zero during > + // 'a. > + unsafe { &*ptr.cast() } > + } > + > + /// Returns a raw pointer to the inner C struct. > + #[inline] > + pub fn as_ptr(&self) -> *mut bindings::file { > + self.0.get() > + } > + > + /// Returns the flags associated with the file. > + /// > + /// The flags are a combination of the constants in [`flags`]. > + pub fn flags(&self) -> u32 { > + // This `read_volatile` is intended to correspond to a READ_ONCE call. > + // > + // SAFETY: The file is valid because the shared reference guarantees a nonzero refcount. > + // > + // FIXME(read_once): Replace with `read_once` when available on the Rust side. > + unsafe { core::ptr::addr_of!((*self.as_ptr()).f_flags).read_volatile() } > + } > +} > + > +// SAFETY: The type invariants guarantee that `File` is always ref-counted. This implementation > +// makes `ARef<File>` own a normal refcount. > +unsafe impl AlwaysRefCounted for File { > + fn inc_ref(&self) { > + // SAFETY: The existence of a shared reference means that the refcount is nonzero. > + unsafe { bindings::get_file(self.as_ptr()) }; > + } > + > + unsafe fn dec_ref(obj: ptr::NonNull<File>) { > + // SAFETY: To call this method, the caller passes us ownership of a normal refcount, so we > + // may drop it. The cast is okay since `File` has the same representation as `struct file`. > + unsafe { bindings::fput(obj.cast().as_ptr()) } > + } > +} > + > +/// Represents the `EBADF` error code. > +/// > +/// Used for methods that can only fail with `EBADF`. > +#[derive(Copy, Clone, Eq, PartialEq)] > +pub struct BadFdError; > + > +impl From<BadFdError> for Error { > + fn from(_: BadFdError) -> Error { > + EBADF > + } > +} > + > +impl core::fmt::Debug for BadFdError { > + fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result { > + f.pad("EBADF") > + } > +} > diff --git a/rust/kernel/lib.rs b/rust/kernel/lib.rs > index b89ecf4e97a0..9353dd713a20 100644 > --- a/rust/kernel/lib.rs > +++ b/rust/kernel/lib.rs > @@ -34,6 +34,7 @@ > mod allocator; > mod build_assert; > pub mod error; > +pub mod file; > pub mod init; > pub mod ioctl; > #[cfg(CONFIG_KUNIT)] > > -- > 2.43.0.687.g38aa6559b0-goog >
Christian Brauner <brauner@kernel.org> wrote: > On Fri, Feb 09, 2024 at 11:18:16AM +0000, Alice Ryhl wrote: >> From: Wedson Almeida Filho <wedsonaf@gmail.com> >> >> This abstraction makes it possible to manipulate the open files for a >> process. The new `File` struct wraps the C `struct file`. When accessing >> it using the smart pointer `ARef<File>`, the pointer will own a >> reference count to the file. When accessing it as `&File`, then the >> reference does not own a refcount, but the borrow checker will ensure >> that the reference count does not hit zero while the `&File` is live. >> >> Since this is intended to manipulate the open files of a process, we >> introduce an `fget` constructor that corresponds to the C `fget` >> method. In future patches, it will become possible to create a new fd in >> a process and bind it to a `File`. Rust Binder will use these to send >> fds from one process to another. >> >> We also provide a method for accessing the file's flags. Rust Binder >> will use this to access the flags of the Binder fd to check whether the >> non-blocking flag is set, which affects what the Binder ioctl does. >> >> This introduces a struct for the EBADF error type, rather than just >> using the Error type directly. This has two advantages: >> * `File::from_fd` returns a `Result<ARef<File>, BadFdError>`, which the > > Sorry, where's that method? Sorry, this is supposed to say `File::fget`. >> +/// Wraps the kernel's `struct file`. >> +/// >> +/// This represents an open file rather than a file on a filesystem. Processes generally reference >> +/// open files using file descriptors. However, file descriptors are not the same as files. A file >> +/// descriptor is just an integer that corresponds to a file, and a single file may be referenced >> +/// by multiple file descriptors. >> +/// >> +/// # Refcounting >> +/// >> +/// Instances of this type are reference-counted. The reference count is incremented by the >> +/// `fget`/`get_file` functions and decremented by `fput`. The Rust type `ARef<File>` represents a >> +/// pointer that owns a reference count on the file. >> +/// >> +/// Whenever a process opens a file descriptor (fd), it stores a pointer to the file in its `struct >> +/// files_struct`. This pointer owns a reference count to the file, ensuring the file isn't >> +/// prematurely deleted while the file descriptor is open. In Rust terminology, the pointers in >> +/// `struct files_struct` are `ARef<File>` pointers. >> +/// >> +/// ## Light refcounts >> +/// >> +/// Whenever a process has an fd to a file, it may use something called a "light refcount" as a >> +/// performance optimization. Light refcounts are acquired by calling `fdget` and released with >> +/// `fdput`. The idea behind light refcounts is that if the fd is not closed between the calls to >> +/// `fdget` and `fdput`, then the refcount cannot hit zero during that time, as the `struct >> +/// files_struct` holds a reference until the fd is closed. This means that it's safe to access the >> +/// file even if `fdget` does not increment the refcount. >> +/// >> +/// The requirement that the fd is not closed during a light refcount applies globally across all >> +/// threads - not just on the thread using the light refcount. For this reason, light refcounts are >> +/// only used when the `struct files_struct` is not shared with other threads, since this ensures >> +/// that other unrelated threads cannot suddenly start using the fd and close it. Therefore, >> +/// calling `fdget` on a shared `struct files_struct` creates a normal refcount instead of a light >> +/// refcount. > > When the fdget() calling task doesn't have a shared file descriptor > table fdget() will not increment the reference count, yes. This > also implies that you cannot have task A use fdget() and then pass > f.file to task B that holds on to it while A returns to userspace. It's > irrelevant that task A won't drop the reference count or that task B > won't drop the reference count. Because task A could return back to > userspace and immediately close the fd via a regular close() system call > at which point task B has a UAF. In other words a file that has been > gotten via fdget() can't be Send to another task without the Send > implying taking a reference to it. That matches my understanding. I suppose that technically you can still send it to another thread *if* you ensure that the current thread waits until that other thread stops using the file before returning to userspace. >> +/// >> +/// Light reference counts must be released with `fdput` before the system call returns to >> +/// userspace. This means that if you wait until the current system call returns to userspace, then >> +/// all light refcounts that existed at the time have gone away. >> +/// >> +/// ## Rust references >> +/// >> +/// The reference type `&File` is similar to light refcounts: >> +/// >> +/// * `&File` references don't own a reference count. They can only exist as long as the reference >> +/// count stays positive, and can only be created when there is some mechanism in place to ensure >> +/// this. >> +/// >> +/// * The Rust borrow-checker normally ensures this by enforcing that the `ARef<File>` from which >> +/// a `&File` is created outlives the `&File`. > > The section confuses me a little: Does the borrow-checker always ensure > that a &File stays valid or are there circumstances where it doesn't or > are you saying it doesn't enforce it? The borrow-checker always ensures it. A &File is actually short-hand for &'a File, where 'a is some unspecified lifetime. We say that &'a File is annotated with 'a. The borrow-checker rejects any code that tries to use a reference after the end of the lifetime annotated on it. So as long as you annotate the reference with a sufficiently short lifetime, the borrow checker will prevent UAF. And outside of cases like from_ptr, the borrow-checker also takes care of ensuring that the lifetimes are sufficiently short. (Technically &'a File and &'b File are two completely different types, so &File is technically a class of types and not a single type. Rust will automatically convert &'long File to &'short File.) >> +/// >> +/// * Using the unsafe [`File::from_ptr`] means that it is up to the caller to ensure that the >> +/// `&File` only exists while the reference count is positive. > > What is this used for in binder? If it's not used don't add it. This is used on the boundary between the Rust part of Binder and the binderfs component that is implemented in C. For example: unsafe extern "C" fn rust_binder_open( inode: *mut bindings::inode, file_ptr: *mut bindings::file, ) -> core::ffi::c_int { // SAFETY: The `rust_binderfs.c` file ensures that `i_private` is set to the return value of a // successful call to `rust_binder_new_device`. let ctx = unsafe { Arc::<Context>::borrow((*inode).i_private) }; // SAFETY: The caller provides a valid file pointer to a new `struct file`. let file = unsafe { File::from_ptr(file_ptr) }; let process = match Process::open(ctx, file) { Ok(process) => process, Err(err) => return err.to_errno(), }; // SAFETY: This file is associated with Rust binder, so we own the `private_data` field. unsafe { (*file_ptr).private_data = process.into_foreign().cast_mut(); } 0 } Here, rust_binder_open is the open function in a struct file_operations vtable. In this case, file_ptr is guaranteed by the caller to be valid for the duration of the call to rust_binder_open. Binder uses from_ptr to get a &File from the raw pointer. As far as I understand, the caller of rust_binder_open uses fdget to ensure that file_ptr is valid for the duration of the call, but the Rust code doesn't assume that it does this with fdget. As long as file_ptr is valid for the duration of the rust_binder_open call, this use of from_ptr is okay. It will continue to work even if the caller is changed to use fget. As for how this code ensures that `file` ends up annotated with a sufficiently short lifetime, well, that has to do with the signature of Process::open. Here it is: impl Process { pub(crate) fn open(ctx: ArcBorrow<'_, Context>, file: &File) -> Result<Arc<Process>> { Self::new(ctx.into(), ARef::from(file.cred())) } } In this case, &File is used without specifying a lifetime. It's a function argument, so this means that the lifetime annotated on the `file` argument will be exactly the duration in which Process::open is called. So any attempt to use `file` after the end of the call to Process::open will be rejected by the borrow-checker. (E.g., if Process::open tried to schedule something on the workqueue using `file`, then that would not compile. Storing it in a global variable would not compile either.) This means that the borrow-checker will not catch mistakes in rust_binder_open, but it *will* catch mistakes in Process::open, and anything called by Process::open. These examples come from patch 2 of the Binder RFC: https://lore.kernel.org/rust-for-linux/20231101-rust-binder-v1-2-08ba9197f637@google.com/ >> +/// >> +/// * You can think of `fdget` as using an fd to look up an `ARef<File>` in the `struct > > Could you explain why there isn't an explicit fdget() then and you have > that from_ptr() method? I don't provide an fdget implementation because Binder never calls it. However, if you would like, I would be happy to add one to the patchset. As for from_ptr, see above. Alice
On Wed, Mar 20, 2024 at 06:09:05PM +0000, Alice Ryhl wrote: > Christian Brauner <brauner@kernel.org> wrote: > > On Fri, Feb 09, 2024 at 11:18:16AM +0000, Alice Ryhl wrote: > >> From: Wedson Almeida Filho <wedsonaf@gmail.com> > >> > >> This abstraction makes it possible to manipulate the open files for a > >> process. The new `File` struct wraps the C `struct file`. When accessing > >> it using the smart pointer `ARef<File>`, the pointer will own a > >> reference count to the file. When accessing it as `&File`, then the > >> reference does not own a refcount, but the borrow checker will ensure > >> that the reference count does not hit zero while the `&File` is live. > >> > >> Since this is intended to manipulate the open files of a process, we > >> introduce an `fget` constructor that corresponds to the C `fget` > >> method. In future patches, it will become possible to create a new fd in > >> a process and bind it to a `File`. Rust Binder will use these to send > >> fds from one process to another. > >> > >> We also provide a method for accessing the file's flags. Rust Binder > >> will use this to access the flags of the Binder fd to check whether the > >> non-blocking flag is set, which affects what the Binder ioctl does. > >> > >> This introduces a struct for the EBADF error type, rather than just > >> using the Error type directly. This has two advantages: > >> * `File::from_fd` returns a `Result<ARef<File>, BadFdError>`, which the > > > > Sorry, where's that method? > > Sorry, this is supposed to say `File::fget`. > > >> +/// Wraps the kernel's `struct file`. > >> +/// > >> +/// This represents an open file rather than a file on a filesystem. Processes generally reference > >> +/// open files using file descriptors. However, file descriptors are not the same as files. A file > >> +/// descriptor is just an integer that corresponds to a file, and a single file may be referenced > >> +/// by multiple file descriptors. > >> +/// > >> +/// # Refcounting > >> +/// > >> +/// Instances of this type are reference-counted. The reference count is incremented by the > >> +/// `fget`/`get_file` functions and decremented by `fput`. The Rust type `ARef<File>` represents a > >> +/// pointer that owns a reference count on the file. > >> +/// > >> +/// Whenever a process opens a file descriptor (fd), it stores a pointer to the file in its `struct > >> +/// files_struct`. This pointer owns a reference count to the file, ensuring the file isn't > >> +/// prematurely deleted while the file descriptor is open. In Rust terminology, the pointers in > >> +/// `struct files_struct` are `ARef<File>` pointers. > >> +/// > >> +/// ## Light refcounts > >> +/// > >> +/// Whenever a process has an fd to a file, it may use something called a "light refcount" as a > >> +/// performance optimization. Light refcounts are acquired by calling `fdget` and released with > >> +/// `fdput`. The idea behind light refcounts is that if the fd is not closed between the calls to > >> +/// `fdget` and `fdput`, then the refcount cannot hit zero during that time, as the `struct > >> +/// files_struct` holds a reference until the fd is closed. This means that it's safe to access the > >> +/// file even if `fdget` does not increment the refcount. > >> +/// > >> +/// The requirement that the fd is not closed during a light refcount applies globally across all > >> +/// threads - not just on the thread using the light refcount. For this reason, light refcounts are > >> +/// only used when the `struct files_struct` is not shared with other threads, since this ensures > >> +/// that other unrelated threads cannot suddenly start using the fd and close it. Therefore, > >> +/// calling `fdget` on a shared `struct files_struct` creates a normal refcount instead of a light > >> +/// refcount. > > > > When the fdget() calling task doesn't have a shared file descriptor > > table fdget() will not increment the reference count, yes. This > > also implies that you cannot have task A use fdget() and then pass > > f.file to task B that holds on to it while A returns to userspace. It's > > irrelevant that task A won't drop the reference count or that task B > > won't drop the reference count. Because task A could return back to > > userspace and immediately close the fd via a regular close() system call > > at which point task B has a UAF. In other words a file that has been > > gotten via fdget() can't be Send to another task without the Send > > implying taking a reference to it. > > That matches my understanding. > > I suppose that technically you can still send it to another thread *if* you > ensure that the current thread waits until that other thread stops using the > file before returning to userspace. _Technically_ yes, but it would be brittle as hell. The problem is that fdget() _relies_ on being single-threaded for the time that fd is used until fdput(). There's locking assumptions that build on that e.g., for concurrent read/write. So no, that shouldn't be allowed. Look at how this broke our back when we introduced pidfd_getfd() where we steal an fd from another task. I have a lengthy explanation how that can be used to violate our elided-locking which is based on assuming that we're always single-threaded and the file can't be suddenly shared with another task. So maybe doable but it would make the semantics even more intricate. > >> +/// > >> +/// Light reference counts must be released with `fdput` before the system call returns to > >> +/// userspace. This means that if you wait until the current system call returns to userspace, then > >> +/// all light refcounts that existed at the time have gone away. > >> +/// > >> +/// ## Rust references > >> +/// > >> +/// The reference type `&File` is similar to light refcounts: > >> +/// > >> +/// * `&File` references don't own a reference count. They can only exist as long as the reference > >> +/// count stays positive, and can only be created when there is some mechanism in place to ensure > >> +/// this. > >> +/// > >> +/// * The Rust borrow-checker normally ensures this by enforcing that the `ARef<File>` from which > >> +/// a `&File` is created outlives the `&File`. > > > > The section confuses me a little: Does the borrow-checker always ensure > > that a &File stays valid or are there circumstances where it doesn't or > > are you saying it doesn't enforce it? > > The borrow-checker always ensures it. Ok, thanks. > > A &File is actually short-hand for &'a File, where 'a is some > unspecified lifetime. We say that &'a File is annotated with 'a. The > borrow-checker rejects any code that tries to use a reference after the > end of the lifetime annotated on it. Thanks for the explanation. > > So as long as you annotate the reference with a sufficiently short > lifetime, the borrow checker will prevent UAF. And outside of cases like Sorry, but can you explain "sufficiently short lifetime"? > from_ptr, the borrow-checker also takes care of ensuring that the > lifetimes are sufficiently short. > > (Technically &'a File and &'b File are two completely different types, > so &File is technically a class of types and not a single type. Rust > will automatically convert &'long File to &'short File.) > > >> +/// > >> +/// * Using the unsafe [`File::from_ptr`] means that it is up to the caller to ensure that the > >> +/// `&File` only exists while the reference count is positive. > > > > What is this used for in binder? If it's not used don't add it. > > This is used on the boundary between the Rust part of Binder and the > binderfs component that is implemented in C. For example: I see, I'm being foiled by my own code... > > unsafe extern "C" fn rust_binder_open( > inode: *mut bindings::inode, > file_ptr: *mut bindings::file, > ) -> core::ffi::c_int { > // SAFETY: The `rust_binderfs.c` file ensures that `i_private` is set to the return value of a > // successful call to `rust_binder_new_device`. > let ctx = unsafe { Arc::<Context>::borrow((*inode).i_private) }; > > // SAFETY: The caller provides a valid file pointer to a new `struct file`. > let file = unsafe { File::from_ptr(file_ptr) }; We need a better name for this helper than from_ptr() imho. I think from_ptr() and as_ptr() is odd for C. How weird would it be to call that from_raw_file() and as_raw_file()? But bigger picture I somewhat struggle with the semantics of this because this is not an interface that we have in C and this is really about a low-level contract between C and Rust. Specifically this states that this pointer is _somehow_ guaranteed valid. And really, this seems a bit of a hack. Naively, I think this should probably not be necessary if file_operations are properly wrapped. Or it should at least be demotable to a purely internal method that can't be called directly or something. So what I mean is. fdget() may or may not take a reference. The C interface internally knows whether a reference is owned or not by abusing the lower two bits in a pointer to keep track of that. Naively, I would expect the same information to be available to rust so that it's clear to Rust wheter it's dealing with an explicitly referenced file or an elided-reference file. Maybe that's not possible and I'm not well-versed enough to see that yet. > let process = match Process::open(ctx, file) { > Ok(process) => process, > Err(err) => return err.to_errno(), > }; > // SAFETY: This file is associated with Rust binder, so we own the `private_data` field. > unsafe { > (*file_ptr).private_data = process.into_foreign().cast_mut(); > } > 0 > } > > Here, rust_binder_open is the open function in a struct file_operations > vtable. In this case, file_ptr is guaranteed by the caller to be valid Where's the code that wraps struct file_operations? > for the duration of the call to rust_binder_open. Binder uses from_ptr > to get a &File from the raw pointer. > > As far as I understand, the caller of rust_binder_open uses fdget to > ensure that file_ptr is valid for the duration of the call, but the Rust > code doesn't assume that it does this with fdget. As long as file_ptr is > valid for the duration of the rust_binder_open call, this use of > from_ptr is okay. It will continue to work even if the caller is changed > to use fget. Ok. > > As for how this code ensures that `file` ends up annotated with a > sufficiently short lifetime, well, that has to do with the signature of > Process::open. Here it is: > > impl Process { > pub(crate) fn open(ctx: ArcBorrow<'_, Context>, file: &File) -> Result<Arc<Process>> { > Self::new(ctx.into(), ARef::from(file.cred())) > } > } > > In this case, &File is used without specifying a lifetime. It's a > function argument, so this means that the lifetime annotated on the > `file` argument will be exactly the duration in which Process::open is > called. So any attempt to use `file` after the end of the call to > Process::open will be rejected by the borrow-checker. (E.g., if > Process::open tried to schedule something on the workqueue using `file`, > then that would not compile. Storing it in a global variable would not > compile either.) > > This means that the borrow-checker will not catch mistakes in > rust_binder_open, but it *will* catch mistakes in Process::open, and > anything called by Process::open. > > These examples come from patch 2 of the Binder RFC: > https://lore.kernel.org/rust-for-linux/20231101-rust-binder-v1-2-08ba9197f637@google.com/ > > >> +/// > >> +/// * You can think of `fdget` as using an fd to look up an `ARef<File>` in the `struct > > > > Could you explain why there isn't an explicit fdget() then and you have > > that from_ptr() method? > > I don't provide an fdget implementation because Binder never calls it. > However, if you would like, I would be happy to add one to the patchset. > > As for from_ptr, see above. > > Alice
Christian Brauner <brauner@kernel.org> wrote: > On Wed, Mar 20, 2024 at 06:09:05PM +0000, Alice Ryhl wrote: >> Christian Brauner <brauner@kernel.org> wrote: >>> On Fri, Feb 09, 2024 at 11:18:16AM +0000, Alice Ryhl wrote: >>>> +/// Wraps the kernel's `struct file`. >>>> +/// >>>> +/// This represents an open file rather than a file on a filesystem. Processes generally reference >>>> +/// open files using file descriptors. However, file descriptors are not the same as files. A file >>>> +/// descriptor is just an integer that corresponds to a file, and a single file may be referenced >>>> +/// by multiple file descriptors. >>>> +/// >>>> +/// # Refcounting >>>> +/// >>>> +/// Instances of this type are reference-counted. The reference count is incremented by the >>>> +/// `fget`/`get_file` functions and decremented by `fput`. The Rust type `ARef<File>` represents a >>>> +/// pointer that owns a reference count on the file. >>>> +/// >>>> +/// Whenever a process opens a file descriptor (fd), it stores a pointer to the file in its `struct >>>> +/// files_struct`. This pointer owns a reference count to the file, ensuring the file isn't >>>> +/// prematurely deleted while the file descriptor is open. In Rust terminology, the pointers in >>>> +/// `struct files_struct` are `ARef<File>` pointers. >>>> +/// >>>> +/// ## Light refcounts >>>> +/// >>>> +/// Whenever a process has an fd to a file, it may use something called a "light refcount" as a >>>> +/// performance optimization. Light refcounts are acquired by calling `fdget` and released with >>>> +/// `fdput`. The idea behind light refcounts is that if the fd is not closed between the calls to >>>> +/// `fdget` and `fdput`, then the refcount cannot hit zero during that time, as the `struct >>>> +/// files_struct` holds a reference until the fd is closed. This means that it's safe to access the >>>> +/// file even if `fdget` does not increment the refcount. >>>> +/// >>>> +/// The requirement that the fd is not closed during a light refcount applies globally across all >>>> +/// threads - not just on the thread using the light refcount. For this reason, light refcounts are >>>> +/// only used when the `struct files_struct` is not shared with other threads, since this ensures >>>> +/// that other unrelated threads cannot suddenly start using the fd and close it. Therefore, >>>> +/// calling `fdget` on a shared `struct files_struct` creates a normal refcount instead of a light >>>> +/// refcount. >>> >>> When the fdget() calling task doesn't have a shared file descriptor >>> table fdget() will not increment the reference count, yes. This >>> also implies that you cannot have task A use fdget() and then pass >>> f.file to task B that holds on to it while A returns to userspace. It's >>> irrelevant that task A won't drop the reference count or that task B >>> won't drop the reference count. Because task A could return back to >>> userspace and immediately close the fd via a regular close() system call >>> at which point task B has a UAF. In other words a file that has been >>> gotten via fdget() can't be Send to another task without the Send >>> implying taking a reference to it. >> >> That matches my understanding. >> >> I suppose that technically you can still send it to another thread *if* you >> ensure that the current thread waits until that other thread stops using the >> file before returning to userspace. > > _Technically_ yes, but it would be brittle as hell. The problem is that > fdget() _relies_ on being single-threaded for the time that fd is used > until fdput(). There's locking assumptions that build on that e.g., for > concurrent read/write. So no, that shouldn't be allowed. > > Look at how this broke our back when we introduced pidfd_getfd() where > we steal an fd from another task. I have a lengthy explanation how that > can be used to violate our elided-locking which is based on assuming > that we're always single-threaded and the file can't be suddenly shared > with another task. So maybe doable but it would make the semantics even > more intricate. Hmm, the part about elided locking is surprising to me, and may be an issue. Can you give more details on that? Because the current abstractions here *do* actually allow what I described, since we implement Sync for File. I'm not familiar with the pidfd_getfd discussion you are referring to. Do you have a link? I'm thinking that we may have to provide two different `struct file` wrappers to accurately model this API in Rust. Perhaps they could be called File and LocalFile, where one is marked as thread safe and the other isn't. I can make all LocalFile methods available on File to avoid having to duplicate methods that are available on both. But it's not clear to me that this is even enough. Even if we give you a &LocalFile to prevent you from moving it across threads, you can just call File::fget to get an ARef<File> to the same file and then move *that* across threads. This kind of global requirement is not so easy to model. Maybe klint [1] could do it ... atomic context violations are a similar kind of global check. But having klint do it would be far out. Or maybe File::fget should also return a LocalFile? But this raises a different question to me. Let's say process A uses Binder to send its own fd to process B, and the following things happen: 1. Process A enters the ioctl and takes fdget on the fd. 2. Process A calls fget on the same fd to send it to another process. 3. Process A goes to sleep, waiting for process B to respond. 4. Process B receives the message, installs the fd, and returns to userspace. 5. Process B responds to the transaction, but does not close the fd. 6a. Process A finishes sleeping, and returns to userspace from the ioctl. 6b. Process B tries to do an operation (e.g. read) on the fd. Let's say that 6a and 6b run in parallel. Could this potentially result in a data race between step 6a and 6b? I'm guessing that step 6a probably doesn't touch any of the code that has elided locking assumptions, so in practice I guess there's not a problem ... but if you make any sort of elided locking assumption as you exit from the ioctl (before reaching the fdput), then it seems to me that you have a problem. >>>> +/// >>>> +/// Light reference counts must be released with `fdput` before the system call returns to >>>> +/// userspace. This means that if you wait until the current system call returns to userspace, then >>>> +/// all light refcounts that existed at the time have gone away. >>>> +/// >>>> +/// ## Rust references >>>> +/// >>>> +/// The reference type `&File` is similar to light refcounts: >>>> +/// >>>> +/// * `&File` references don't own a reference count. They can only exist as long as the reference >>>> +/// count stays positive, and can only be created when there is some mechanism in place to ensure >>>> +/// this. >>>> +/// >>>> +/// * The Rust borrow-checker normally ensures this by enforcing that the `ARef<File>` from which >>>> +/// a `&File` is created outlives the `&File`. >>> >>> The section confuses me a little: Does the borrow-checker always ensure >>> that a &File stays valid or are there circumstances where it doesn't or >>> are you saying it doesn't enforce it? >> >> The borrow-checker always ensures it. > > Ok, thanks. > >> >> A &File is actually short-hand for &'a File, where 'a is some >> unspecified lifetime. We say that &'a File is annotated with 'a. The >> borrow-checker rejects any code that tries to use a reference after the >> end of the lifetime annotated on it. > > Thanks for the explanation. > >> >> So as long as you annotate the reference with a sufficiently short >> lifetime, the borrow checker will prevent UAF. And outside of cases like > > Sorry, but can you explain "sufficiently short lifetime"? By "sufficiently short lifetime" I mean "lifetime that ends before the object is destroyed". Idea being that if the lifetime ends before the object is freed, and the borrow-checker rejects attempts to use it after the lifetime ends, then it follows that the borrow-checker prevents use-after-frees. >> from_ptr, the borrow-checker also takes care of ensuring that the >> lifetimes are sufficiently short. >> >> (Technically &'a File and &'b File are two completely different types, >> so &File is technically a class of types and not a single type. Rust >> will automatically convert &'long File to &'short File.) >> >>>> +/// >>>> +/// * Using the unsafe [`File::from_ptr`] means that it is up to the caller to ensure that the >>>> +/// `&File` only exists while the reference count is positive. >>> >>> What is this used for in binder? If it's not used don't add it. >> >> This is used on the boundary between the Rust part of Binder and the >> binderfs component that is implemented in C. For example: > > I see, I'm being foiled by my own code... > >> >> unsafe extern "C" fn rust_binder_open( >> inode: *mut bindings::inode, >> file_ptr: *mut bindings::file, >> ) -> core::ffi::c_int { >> // SAFETY: The `rust_binderfs.c` file ensures that `i_private` is set to the return value of a >> // successful call to `rust_binder_new_device`. >> let ctx = unsafe { Arc::<Context>::borrow((*inode).i_private) }; >> >> // SAFETY: The caller provides a valid file pointer to a new `struct file`. >> let file = unsafe { File::from_ptr(file_ptr) }; > > We need a better name for this helper than from_ptr() imho. I think > from_ptr() and as_ptr() is odd for C. How weird would it be to call > that from_raw_file() and as_raw_file()? That's a reasonable name. I would be happy to rename to that, and I don't think it is unidiomatic. > But bigger picture I somewhat struggle with the semantics of this > because this is not an interface that we have in C and this is really > about a low-level contract between C and Rust. Specifically this states > that this pointer is _somehow_ guaranteed valid. And really, this seems > a bit of a hack. Indeed ... but I think it's a quite common hack. After all, any time you dereference a raw pointer in Rust, you are making the same assumption. > Naively, I think this should probably not be necessary if > file_operations are properly wrapped. Or it should at least be demotable > to a purely internal method that can't be called directly or something. Yes, the usage here of File::from_ptr could probably be hidden inside a suitably designed file_operations wrapper. The thing is, Rust Binder doesn't currently use such a wrapper at all. It just exports a global of type file_operations and the C code in binderfs then references that global. Rust Binder used to use such an abstraction, but I ripped it out before sending the Rust Binder RFC because it didn't actually help. It was designed for cases where the file system is also implemented in Rust, so to get it to expose a file_operations global to the C code in binderfs, I had to reach inside its internal implementation. It did not save me from doing stuff such as using File::from_ptr from Binder. Now, you could most likely modify those file_operations abstractions to support my use-case better. But calling into C is already unsafe, so unless we get multiple drivers that have a similar C/Rust split, it's not clear that it's useful to extract the logic from Binder. I would prefer to wait for the file_operations abstraction to get upstreamed by the people working on VFS bindings, and then we can decide whether we should rewrite binderfs into Rust and get rid of the logic, or whether it's worth to expand the file_operations abstraction to support split C/Rust drivers like the current binderfs. > So what I mean is. fdget() may or may not take a reference. The C > interface internally knows whether a reference is owned or not by > abusing the lower two bits in a pointer to keep track of that. Naively, > I would expect the same information to be available to rust so that it's > clear to Rust wheter it's dealing with an explicitly referenced file or > an elided-reference file. Maybe that's not possible and I'm not > well-versed enough to see that yet. I'm sure Rust can access the same information, but I don't think I'm currently doing anything that cares about the distinction? >> let process = match Process::open(ctx, file) { >> Ok(process) => process, >> Err(err) => return err.to_errno(), >> }; >> // SAFETY: This file is associated with Rust binder, so we own the `private_data` field. >> unsafe { >> (*file_ptr).private_data = process.into_foreign().cast_mut(); >> } >> 0 >> } >> >> Here, rust_binder_open is the open function in a struct file_operations >> vtable. In this case, file_ptr is guaranteed by the caller to be valid > > Where's the code that wraps struct file_operations? Please see drivers/android/rust_binder.rs in the binderfs patch [2] in the Rust Binder RFC. >> for the duration of the call to rust_binder_open. Binder uses from_ptr >> to get a &File from the raw pointer. >> >> As far as I understand, the caller of rust_binder_open uses fdget to >> ensure that file_ptr is valid for the duration of the call, but the Rust >> code doesn't assume that it does this with fdget. As long as file_ptr is >> valid for the duration of the rust_binder_open call, this use of >> from_ptr is okay. It will continue to work even if the caller is changed >> to use fget. > > Ok. > [1]: https://rust-for-linux.com/klint [2]: https://lore.kernel.org/rust-for-linux/20231101-rust-binder-v1-2-08ba9197f637@google.com/ Alice
On Mon, Apr 01, 2024 at 12:09:08PM +0000, Alice Ryhl wrote: > Christian Brauner <brauner@kernel.org> wrote: > > On Wed, Mar 20, 2024 at 06:09:05PM +0000, Alice Ryhl wrote: > >> Christian Brauner <brauner@kernel.org> wrote: > >>> On Fri, Feb 09, 2024 at 11:18:16AM +0000, Alice Ryhl wrote: > >>>> +/// Wraps the kernel's `struct file`. > >>>> +/// > >>>> +/// This represents an open file rather than a file on a filesystem. Processes generally reference > >>>> +/// open files using file descriptors. However, file descriptors are not the same as files. A file > >>>> +/// descriptor is just an integer that corresponds to a file, and a single file may be referenced > >>>> +/// by multiple file descriptors. > >>>> +/// > >>>> +/// # Refcounting > >>>> +/// > >>>> +/// Instances of this type are reference-counted. The reference count is incremented by the > >>>> +/// `fget`/`get_file` functions and decremented by `fput`. The Rust type `ARef<File>` represents a > >>>> +/// pointer that owns a reference count on the file. > >>>> +/// > >>>> +/// Whenever a process opens a file descriptor (fd), it stores a pointer to the file in its `struct > >>>> +/// files_struct`. This pointer owns a reference count to the file, ensuring the file isn't > >>>> +/// prematurely deleted while the file descriptor is open. In Rust terminology, the pointers in > >>>> +/// `struct files_struct` are `ARef<File>` pointers. > >>>> +/// > >>>> +/// ## Light refcounts > >>>> +/// > >>>> +/// Whenever a process has an fd to a file, it may use something called a "light refcount" as a > >>>> +/// performance optimization. Light refcounts are acquired by calling `fdget` and released with > >>>> +/// `fdput`. The idea behind light refcounts is that if the fd is not closed between the calls to > >>>> +/// `fdget` and `fdput`, then the refcount cannot hit zero during that time, as the `struct > >>>> +/// files_struct` holds a reference until the fd is closed. This means that it's safe to access the > >>>> +/// file even if `fdget` does not increment the refcount. > >>>> +/// > >>>> +/// The requirement that the fd is not closed during a light refcount applies globally across all > >>>> +/// threads - not just on the thread using the light refcount. For this reason, light refcounts are > >>>> +/// only used when the `struct files_struct` is not shared with other threads, since this ensures > >>>> +/// that other unrelated threads cannot suddenly start using the fd and close it. Therefore, > >>>> +/// calling `fdget` on a shared `struct files_struct` creates a normal refcount instead of a light > >>>> +/// refcount. > >>> > >>> When the fdget() calling task doesn't have a shared file descriptor > >>> table fdget() will not increment the reference count, yes. This > >>> also implies that you cannot have task A use fdget() and then pass > >>> f.file to task B that holds on to it while A returns to userspace. It's > >>> irrelevant that task A won't drop the reference count or that task B > >>> won't drop the reference count. Because task A could return back to > >>> userspace and immediately close the fd via a regular close() system call > >>> at which point task B has a UAF. In other words a file that has been > >>> gotten via fdget() can't be Send to another task without the Send > >>> implying taking a reference to it. > >> > >> That matches my understanding. > >> > >> I suppose that technically you can still send it to another thread *if* you > >> ensure that the current thread waits until that other thread stops using the > >> file before returning to userspace. > > > > _Technically_ yes, but it would be brittle as hell. The problem is that > > fdget() _relies_ on being single-threaded for the time that fd is used > > until fdput(). There's locking assumptions that build on that e.g., for > > concurrent read/write. So no, that shouldn't be allowed. > > > > Look at how this broke our back when we introduced pidfd_getfd() where > > we steal an fd from another task. I have a lengthy explanation how that > > can be used to violate our elided-locking which is based on assuming > > that we're always single-threaded and the file can't be suddenly shared > > with another task. So maybe doable but it would make the semantics even > > more intricate. > > Hmm, the part about elided locking is surprising to me, and may be an > issue. Can you give more details on that? Because the current So what I referred to was that we do have fdget_pos(). Roughly, if there's more than one reference on the file then we need to acquire a mutex but if it's only a single reference then we can avoid taking the mutex because we know that we're the only one that has a reference to that file and no one else can acquire one. Whether or not that mutex was taken is taken track of in struct fd. So you can't share a file after fdget_pos() has been called on it and you haven't taken the position mutex. So let's say you had: * Tread A that calls fdget_pos() on file1 and the reference count is one. So Thread A doesn't acquire the file position mutex for file1. * Now somehow that file1 becomes shared, e.g., Thread B calls fget() on it and now Thread B does some operation that requires the file position mutex. => Thread A and Thread B race on the file position. So just because you have a reference to a file from somewhere it doesn't mean you can just share it with another thread. So if yo have an arbitrary reference to a file in Rust and that somehow can be shared with another thread you risk races here. > abstractions here *do* actually allow what I described, since we > implement Sync for File. > > I'm not familiar with the pidfd_getfd discussion you are referring to. > Do you have a link? https://lore.kernel.org/linux-fsdevel/20230724-vfs-fdget_pos-v1-1-a4abfd7103f3@kernel.org pidfd_getfd() can be used to steal a file descriptor from another task. It's like a non-cooperative SCM_RIGHTS. That means you can have exactly the scenario described above where a file assumed to be non-shared is suddenly shared and you have racing reads/writes. For readdir we nowadays always take the file position mutex because of the pidfd_getfd() business because that might corrupt internal state. > > I'm thinking that we may have to provide two different `struct file` > wrappers to accurately model this API in Rust. Perhaps they could be > called File and LocalFile, where one is marked as thread safe and the > other isn't. I can make all LocalFile methods available on File to avoid > having to duplicate methods that are available on both. But isn't that just struct file and struct fd? Ideally we'd stay close to something like this. > > But it's not clear to me that this is even enough. Even if we give you a > &LocalFile to prevent you from moving it across threads, you can just > call File::fget to get an ARef<File> to the same file and then move > *that* across threads. Yes, absolutely. > > This kind of global requirement is not so easy to model. Maybe klint [1] > could do it ... atomic context violations are a similar kind of global > check. But having klint do it would be far out. > > Or maybe File::fget should also return a LocalFile? > > But this raises a different question to me. Let's say process A uses > Binder to send its own fd to process B, and the following things happen: > > 1. Process A enters the ioctl and takes fdget on the fd. > 2. Process A calls fget on the same fd to send it to another process. > 3. Process A goes to sleep, waiting for process B to respond. > 4. Process B receives the message, installs the fd, and returns to userspace. > 5. Process B responds to the transaction, but does not close the fd. The fd just installed in 4. and the fd you're referring to in 5. are identical, right? IOW, we're not talking about two different fd (dup) for the same file, right? > 6a. Process A finishes sleeping, and returns to userspace from the ioctl. > 6b. Process B tries to do an operation (e.g. read) on the fd. > > Let's say that 6a and 6b run in parallel. > > Could this potentially result in a data race between step 6a and 6b? I'm > guessing that step 6a probably doesn't touch any of the code that has > elided locking assumptions, so in practice I guess there's not a problem > ... but if you make any sort of elided locking assumption as you exit > from the ioctl (before reaching the fdput), then it seems to me that you > have a problem. Yes, 6a doesn't touch any code that has elided locking assumptions. 1'. Process A enters the ioctl and takes fdget() on the fd. Process A holds the only reference to that file and the file descriptor table isn't shared. Therefore, f_count is left untouched and remains at 1. 2'. Process A calls fget() which unconditionally bumps f_count bringing it to 2 and sending it another process (Presumably you intend to imply that this reference is now owned by the second process.). 3'. [as 3.] 4'. Process B installs the file into it's file descriptor table consuming that reference from 2'. The f_count remains at 2 with the reference from 2' now being owned by Process B. 5'. Since Process B isn't closing the fd and has just called fd_install() it returns to userspace with f_count untouched and still at 2. 6'a. Process A finishes sleeping and returns to userspace calling fdput(). Since the original fdget() was done without bumping the reference count the fdput() of Process A will not decrement the reference count. So f_count remains at 2. 6'b. Process B performs a read/write syscall and calls fdget_pos(). fdget_pos() sees that this file has f_count > 1 and takes the file position mutex. So this isn't a problem. The problem is when a file becomes shared implicitly without the original owner of the file knowing. > > >>>> +/// > >>>> +/// Light reference counts must be released with `fdput` before the system call returns to > >>>> +/// userspace. This means that if you wait until the current system call returns to userspace, then > >>>> +/// all light refcounts that existed at the time have gone away. > >>>> +/// > >>>> +/// ## Rust references > >>>> +/// > >>>> +/// The reference type `&File` is similar to light refcounts: > >>>> +/// > >>>> +/// * `&File` references don't own a reference count. They can only exist as long as the reference > >>>> +/// count stays positive, and can only be created when there is some mechanism in place to ensure > >>>> +/// this. > >>>> +/// > >>>> +/// * The Rust borrow-checker normally ensures this by enforcing that the `ARef<File>` from which > >>>> +/// a `&File` is created outlives the `&File`. > >>> > >>> The section confuses me a little: Does the borrow-checker always ensure > >>> that a &File stays valid or are there circumstances where it doesn't or > >>> are you saying it doesn't enforce it? > >> > >> The borrow-checker always ensures it. > > > > Ok, thanks. > > > >> > >> A &File is actually short-hand for &'a File, where 'a is some > >> unspecified lifetime. We say that &'a File is annotated with 'a. The > >> borrow-checker rejects any code that tries to use a reference after the > >> end of the lifetime annotated on it. > > > > Thanks for the explanation. > > > >> > >> So as long as you annotate the reference with a sufficiently short > >> lifetime, the borrow checker will prevent UAF. And outside of cases like > > > > Sorry, but can you explain "sufficiently short lifetime"? > > By "sufficiently short lifetime" I mean "lifetime that ends before the > object is destroyed". Ah, ok. It sounded like it was a specific concept that Rust is implementing in contrast to long-term lifetime or sm. Thanks! > > Idea being that if the lifetime ends before the object is freed, and the > borrow-checker rejects attempts to use it after the lifetime ends, then > it follows that the borrow-checker prevents use-after-frees. > > >> from_ptr, the borrow-checker also takes care of ensuring that the > >> lifetimes are sufficiently short. > >> > >> (Technically &'a File and &'b File are two completely different types, > >> so &File is technically a class of types and not a single type. Rust > >> will automatically convert &'long File to &'short File.) > >> > >>>> +/// > >>>> +/// * Using the unsafe [`File::from_ptr`] means that it is up to the caller to ensure that the > >>>> +/// `&File` only exists while the reference count is positive. > >>> > >>> What is this used for in binder? If it's not used don't add it. > >> > >> This is used on the boundary between the Rust part of Binder and the > >> binderfs component that is implemented in C. For example: > > > > I see, I'm being foiled by my own code... > > > >> > >> unsafe extern "C" fn rust_binder_open( > >> inode: *mut bindings::inode, > >> file_ptr: *mut bindings::file, > >> ) -> core::ffi::c_int { > >> // SAFETY: The `rust_binderfs.c` file ensures that `i_private` is set to the return value of a > >> // successful call to `rust_binder_new_device`. > >> let ctx = unsafe { Arc::<Context>::borrow((*inode).i_private) }; > >> > >> // SAFETY: The caller provides a valid file pointer to a new `struct file`. > >> let file = unsafe { File::from_ptr(file_ptr) }; > > > > We need a better name for this helper than from_ptr() imho. I think > > from_ptr() and as_ptr() is odd for C. How weird would it be to call > > that from_raw_file() and as_raw_file()? > > That's a reasonable name. I would be happy to rename to that, and I > don't think it is unidiomatic. Thanks! > > > But bigger picture I somewhat struggle with the semantics of this > > because this is not an interface that we have in C and this is really > > about a low-level contract between C and Rust. Specifically this states > > that this pointer is _somehow_ guaranteed valid. And really, this seems > > a bit of a hack. > > Indeed ... but I think it's a quite common hack. After all, any time you > dereference a raw pointer in Rust, you are making the same assumption. > > > Naively, I think this should probably not be necessary if > > file_operations are properly wrapped. Or it should at least be demotable > > to a purely internal method that can't be called directly or something. > > Yes, the usage here of File::from_ptr could probably be hidden inside a > suitably designed file_operations wrapper. The thing is, Rust Binder > doesn't currently use such a wrapper at all. It just exports a global of > type file_operations and the C code in binderfs then references that > global. Yeah. > > Rust Binder used to use such an abstraction, but I ripped it out before > sending the Rust Binder RFC because it didn't actually help. It was > designed for cases where the file system is also implemented in Rust, so > to get it to expose a file_operations global to the C code in binderfs, > I had to reach inside its internal implementation. It did not save me > from doing stuff such as using File::from_ptr from Binder. > > Now, you could most likely modify those file_operations abstractions to > support my use-case better. But calling into C is already unsafe, so > unless we get multiple drivers that have a similar C/Rust split, it's > not clear that it's useful to extract the logic from Binder. I would > prefer to wait for the file_operations abstraction to get upstreamed by > the people working on VFS bindings, and then we can decide whether we > should rewrite binderfs into Rust and get rid of the logic, or whether > it's worth to expand the file_operations abstraction to support split > C/Rust drivers like the current binderfs. > > > So what I mean is. fdget() may or may not take a reference. The C > > interface internally knows whether a reference is owned or not by > > abusing the lower two bits in a pointer to keep track of that. Naively, > > I would expect the same information to be available to rust so that it's > > clear to Rust wheter it's dealing with an explicitly referenced file or > > an elided-reference file. Maybe that's not possible and I'm not > > well-versed enough to see that yet. > > I'm sure Rust can access the same information, but I don't think I'm > currently doing anything that cares about the distinction? Ok. My main goal is that we end up with an almost 1:1 correspondence between the Rust and C interface so it's easy for current maintainers and developers that don't want to care about Rust to continue to do so and also just somewhat verify that changes they do are sane. > > >> let process = match Process::open(ctx, file) { > >> Ok(process) => process, > >> Err(err) => return err.to_errno(), > >> }; > >> // SAFETY: This file is associated with Rust binder, so we own the `private_data` field. > >> unsafe { > >> (*file_ptr).private_data = process.into_foreign().cast_mut(); > >> } > >> 0 > >> } > >> > >> Here, rust_binder_open is the open function in a struct file_operations > >> vtable. In this case, file_ptr is guaranteed by the caller to be valid > > > > Where's the code that wraps struct file_operations? > > Please see drivers/android/rust_binder.rs in the binderfs patch [2] in > the Rust Binder RFC. Ok, I need to just find time to do that...:) > > >> for the duration of the call to rust_binder_open. Binder uses from_ptr > >> to get a &File from the raw pointer. > >> > >> As far as I understand, the caller of rust_binder_open uses fdget to > >> ensure that file_ptr is valid for the duration of the call, but the Rust > >> code doesn't assume that it does this with fdget. As long as file_ptr is > >> valid for the duration of the rust_binder_open call, this use of > >> from_ptr is okay. It will continue to work even if the caller is changed > >> to use fget. > > > > Ok. > > > > [1]: https://rust-for-linux.com/klint > [2]: https://lore.kernel.org/rust-for-linux/20231101-rust-binder-v1-2-08ba9197f637@google.com/ > > Alice
Christian Brauner <brauner@kernel.org> wrote: > On Mon, Apr 01, 2024 at 12:09:08PM +0000, Alice Ryhl wrote: >> Christian Brauner <brauner@kernel.org> wrote: >>> On Wed, Mar 20, 2024 at 06:09:05PM +0000, Alice Ryhl wrote: >>>> Christian Brauner <brauner@kernel.org> wrote: >>>>> On Fri, Feb 09, 2024 at 11:18:16AM +0000, Alice Ryhl wrote: >>>>>> +/// Wraps the kernel's `struct file`. >>>>>> +/// >>>>>> +/// This represents an open file rather than a file on a filesystem. Processes generally reference >>>>>> +/// open files using file descriptors. However, file descriptors are not the same as files. A file >>>>>> +/// descriptor is just an integer that corresponds to a file, and a single file may be referenced >>>>>> +/// by multiple file descriptors. >>>>>> +/// >>>>>> +/// # Refcounting >>>>>> +/// >>>>>> +/// Instances of this type are reference-counted. The reference count is incremented by the >>>>>> +/// `fget`/`get_file` functions and decremented by `fput`. The Rust type `ARef<File>` represents a >>>>>> +/// pointer that owns a reference count on the file. >>>>>> +/// >>>>>> +/// Whenever a process opens a file descriptor (fd), it stores a pointer to the file in its `struct >>>>>> +/// files_struct`. This pointer owns a reference count to the file, ensuring the file isn't >>>>>> +/// prematurely deleted while the file descriptor is open. In Rust terminology, the pointers in >>>>>> +/// `struct files_struct` are `ARef<File>` pointers. >>>>>> +/// >>>>>> +/// ## Light refcounts >>>>>> +/// >>>>>> +/// Whenever a process has an fd to a file, it may use something called a "light refcount" as a >>>>>> +/// performance optimization. Light refcounts are acquired by calling `fdget` and released with >>>>>> +/// `fdput`. The idea behind light refcounts is that if the fd is not closed between the calls to >>>>>> +/// `fdget` and `fdput`, then the refcount cannot hit zero during that time, as the `struct >>>>>> +/// files_struct` holds a reference until the fd is closed. This means that it's safe to access the >>>>>> +/// file even if `fdget` does not increment the refcount. >>>>>> +/// >>>>>> +/// The requirement that the fd is not closed during a light refcount applies globally across all >>>>>> +/// threads - not just on the thread using the light refcount. For this reason, light refcounts are >>>>>> +/// only used when the `struct files_struct` is not shared with other threads, since this ensures >>>>>> +/// that other unrelated threads cannot suddenly start using the fd and close it. Therefore, >>>>>> +/// calling `fdget` on a shared `struct files_struct` creates a normal refcount instead of a light >>>>>> +/// refcount. >>>>> >>>>> When the fdget() calling task doesn't have a shared file descriptor >>>>> table fdget() will not increment the reference count, yes. This >>>>> also implies that you cannot have task A use fdget() and then pass >>>>> f.file to task B that holds on to it while A returns to userspace. It's >>>>> irrelevant that task A won't drop the reference count or that task B >>>>> won't drop the reference count. Because task A could return back to >>>>> userspace and immediately close the fd via a regular close() system call >>>>> at which point task B has a UAF. In other words a file that has been >>>>> gotten via fdget() can't be Send to another task without the Send >>>>> implying taking a reference to it. >>>> >>>> That matches my understanding. >>>> >>>> I suppose that technically you can still send it to another thread *if* you >>>> ensure that the current thread waits until that other thread stops using the >>>> file before returning to userspace. >>> >>> _Technically_ yes, but it would be brittle as hell. The problem is that >>> fdget() _relies_ on being single-threaded for the time that fd is used >>> until fdput(). There's locking assumptions that build on that e.g., for >>> concurrent read/write. So no, that shouldn't be allowed. >>> >>> Look at how this broke our back when we introduced pidfd_getfd() where >>> we steal an fd from another task. I have a lengthy explanation how that >>> can be used to violate our elided-locking which is based on assuming >>> that we're always single-threaded and the file can't be suddenly shared >>> with another task. So maybe doable but it would make the semantics even >>> more intricate. >> >> Hmm, the part about elided locking is surprising to me, and may be an >> issue. Can you give more details on that? Because the current > > So what I referred to was that we do have fdget_pos(). Roughly, if > there's more than one reference on the file then we need to acquire a > mutex but if it's only a single reference then we can avoid taking the > mutex because we know that we're the only one that has a reference to > that file and no one else can acquire one. Whether or not that mutex was > taken is taken track of in struct fd. > > So you can't share a file after fdget_pos() has been called on it and > you haven't taken the position mutex. So let's say you had: > > * Tread A that calls fdget_pos() on file1 and the reference count is > one. So Thread A doesn't acquire the file position mutex for file1. > * Now somehow that file1 becomes shared, e.g., Thread B calls fget() on > it and now Thread B does some operation that requires the file > position mutex. > => Thread A and Thread B race on the file position. > > So just because you have a reference to a file from somewhere it doesn't > mean you can just share it with another thread. > > So if yo have an arbitrary reference to a file in Rust and that somehow > can be shared with another thread you risk races here. > >> abstractions here *do* actually allow what I described, since we >> implement Sync for File. >> >> I'm not familiar with the pidfd_getfd discussion you are referring to. >> Do you have a link? > > https://lore.kernel.org/linux-fsdevel/20230724-vfs-fdget_pos-v1-1-a4abfd7103f3@kernel.org > > pidfd_getfd() can be used to steal a file descriptor from another task. > It's like a non-cooperative SCM_RIGHTS. That means you can have exactly > the scenario described above where a file assumed to be non-shared is > suddenly shared and you have racing reads/writes. > > For readdir we nowadays always take the file position mutex because of > the pidfd_getfd() business because that might corrupt internal state. > >> >> I'm thinking that we may have to provide two different `struct file` >> wrappers to accurately model this API in Rust. Perhaps they could be >> called File and LocalFile, where one is marked as thread safe and the >> other isn't. I can make all LocalFile methods available on File to avoid >> having to duplicate methods that are available on both. > > But isn't that just struct file and struct fd? Ideally we'd stay close > to something like this. Right, that kind of naming seems sensible. But I still need to understand the details a bit better. See below on fdget_pos. >> But it's not clear to me that this is even enough. Even if we give you a >> &LocalFile to prevent you from moving it across threads, you can just >> call File::fget to get an ARef<File> to the same file and then move >> *that* across threads. > > Yes, absolutely. One of my challenges is that Binder wants to call File::fget, immediately move it to another thread, and then call fd_install. And it would be pretty unfortunate if that requires unsafe. But like I argue below, it seems hard to design a safe API for this in the face of fdget_pos. >> This kind of global requirement is not so easy to model. Maybe klint [1] >> could do it ... atomic context violations are a similar kind of global >> check. But having klint do it would be far out. >> >> Or maybe File::fget should also return a LocalFile? >> >> But this raises a different question to me. Let's say process A uses >> Binder to send its own fd to process B, and the following things happen: >> >> 1. Process A enters the ioctl and takes fdget on the fd. >> 2. Process A calls fget on the same fd to send it to another process. >> 3. Process A goes to sleep, waiting for process B to respond. >> 4. Process B receives the message, installs the fd, and returns to userspace. >> 5. Process B responds to the transaction, but does not close the fd. > > The fd just installed in 4. and the fd you're referring to in 5. are > identical, right? IOW, we're not talking about two different fd (dup) > for the same file, right? I'm referring to whatever fd_install does given the `struct file` I got from fget in step 2. >> 6a. Process A finishes sleeping, and returns to userspace from the ioctl. >> 6b. Process B tries to do an operation (e.g. read) on the fd. >> >> Let's say that 6a and 6b run in parallel. >> >> Could this potentially result in a data race between step 6a and 6b? I'm >> guessing that step 6a probably doesn't touch any of the code that has >> elided locking assumptions, so in practice I guess there's not a problem >> ... but if you make any sort of elided locking assumption as you exit >> from the ioctl (before reaching the fdput), then it seems to me that you >> have a problem. > > Yes, 6a doesn't touch any code that has elided locking assumptions. > > 1'. Process A enters the ioctl and takes fdget() on the fd. Process A > holds the only reference to that file and the file descriptor table > isn't shared. Therefore, f_count is left untouched and remains at 1. > 2'. Process A calls fget() which unconditionally bumps f_count bringing > it to 2 and sending it another process (Presumably you intend to > imply that this reference is now owned by the second process.). > 3'. [as 3.] > 4'. Process B installs the file into it's file descriptor table > consuming that reference from 2'. The f_count remains at 2 with the > reference from 2' now being owned by Process B. > 5'. Since Process B isn't closing the fd and has just called > fd_install() it returns to userspace with f_count untouched and > still at 2. > 6'a. Process A finishes sleeping and returns to userspace calling > fdput(). Since the original fdget() was done without bumping the > reference count the fdput() of Process A will not decrement the > reference count. So f_count remains at 2. > 6'b. Process B performs a read/write syscall and calls fdget_pos(). > fdget_pos() sees that this file has f_count > 1 and takes the > file position mutex. > > So this isn't a problem. The problem is when a file becomes shared > implicitly without the original owner of the file knowing. Hmm. Yes, but the ioctl code that called fdget doesn't really know that the ioctl shared the file? So why is it okay? It really seems like there are two different things going on here. When it comes to fdget, we only really care about operations that could remove it from the local file descriptor table, since fdget relies on the refcount in that table remaining valid until fdput. On the other hand, for fdget_pos it also matters whether it gets installed in other file descriptor tables. Threads that reference it through a different fd table will still access the same position. And so this means that between fdget/fdput, there's never any problem with installing the `struct file` into another file descriptor table. Nothing you can do from that other fd table could cause the local fd table to drop its refcount on the file. Whereas such an install can be a problem between fdget_pos/fdput_pos, since that could introduce a race on the position. Is this correct? I was thinking that if we have some sort of File/LocalFile distinction (or File/Fd), then we may be able to get it to work by limiting what a File can do. For example, let's say that the only thing you can do with a File is install it into fd tables, then by the previous logic, there's no problem with it being safe to move across threads even if there's an active fdget. But the fdget_pos kind of throws a wrench into that, because now I can no longer say "it's always safe to do File::fget, move it to another thread, and install it into the remote fd table", since that could cause races on the position if there's an active fdget_pos when we call File::fget. >>>>>> +/// >>>>>> +/// Light reference counts must be released with `fdput` before the system call returns to >>>>>> +/// userspace. This means that if you wait until the current system call returns to userspace, then >>>>>> +/// all light refcounts that existed at the time have gone away. >>>>>> +/// >>>>>> +/// ## Rust references >>>>>> +/// >>>>>> +/// The reference type `&File` is similar to light refcounts: >>>>>> +/// >>>>>> +/// * `&File` references don't own a reference count. They can only exist as long as the reference >>>>>> +/// count stays positive, and can only be created when there is some mechanism in place to ensure >>>>>> +/// this. >>>>>> +/// >>>>>> +/// * The Rust borrow-checker normally ensures this by enforcing that the `ARef<File>` from which >>>>>> +/// a `&File` is created outlives the `&File`. >>>>> >>>>> The section confuses me a little: Does the borrow-checker always ensure >>>>> that a &File stays valid or are there circumstances where it doesn't or >>>>> are you saying it doesn't enforce it? >>>> >>>> The borrow-checker always ensures it. >>> >>> Ok, thanks. >>> >>>> >>>> A &File is actually short-hand for &'a File, where 'a is some >>>> unspecified lifetime. We say that &'a File is annotated with 'a. The >>>> borrow-checker rejects any code that tries to use a reference after the >>>> end of the lifetime annotated on it. >>> >>> Thanks for the explanation. >>> >>>> >>>> So as long as you annotate the reference with a sufficiently short >>>> lifetime, the borrow checker will prevent UAF. And outside of cases like >>> >>> Sorry, but can you explain "sufficiently short lifetime"? >> >> By "sufficiently short lifetime" I mean "lifetime that ends before the >> object is destroyed". > > Ah, ok. It sounded like it was a specific concept that Rust is > implementing in contrast to long-term lifetime or sm. Thanks! > >> >> Idea being that if the lifetime ends before the object is freed, and the >> borrow-checker rejects attempts to use it after the lifetime ends, then >> it follows that the borrow-checker prevents use-after-frees. >> >>>> from_ptr, the borrow-checker also takes care of ensuring that the >>>> lifetimes are sufficiently short. >>>> >>>> (Technically &'a File and &'b File are two completely different types, >>>> so &File is technically a class of types and not a single type. Rust >>>> will automatically convert &'long File to &'short File.) >>>> >>>>>> +/// >>>>>> +/// * Using the unsafe [`File::from_ptr`] means that it is up to the caller to ensure that the >>>>>> +/// `&File` only exists while the reference count is positive. >>>>> >>>>> What is this used for in binder? If it's not used don't add it. >>>> >>>> This is used on the boundary between the Rust part of Binder and the >>>> binderfs component that is implemented in C. For example: >>> >>> I see, I'm being foiled by my own code... >>> >>>> >>>> unsafe extern "C" fn rust_binder_open( >>>> inode: *mut bindings::inode, >>>> file_ptr: *mut bindings::file, >>>> ) -> core::ffi::c_int { >>>> // SAFETY: The `rust_binderfs.c` file ensures that `i_private` is set to the return value of a >>>> // successful call to `rust_binder_new_device`. >>>> let ctx = unsafe { Arc::<Context>::borrow((*inode).i_private) }; >>>> >>>> // SAFETY: The caller provides a valid file pointer to a new `struct file`. >>>> let file = unsafe { File::from_ptr(file_ptr) }; >>> >>> We need a better name for this helper than from_ptr() imho. I think >>> from_ptr() and as_ptr() is odd for C. How weird would it be to call >>> that from_raw_file() and as_raw_file()? >> >> That's a reasonable name. I would be happy to rename to that, and I >> don't think it is unidiomatic. > > Thanks! > >> >>> But bigger picture I somewhat struggle with the semantics of this >>> because this is not an interface that we have in C and this is really >>> about a low-level contract between C and Rust. Specifically this states >>> that this pointer is _somehow_ guaranteed valid. And really, this seems >>> a bit of a hack. >> >> Indeed ... but I think it's a quite common hack. After all, any time you >> dereference a raw pointer in Rust, you are making the same assumption. >> >>> Naively, I think this should probably not be necessary if >>> file_operations are properly wrapped. Or it should at least be demotable >>> to a purely internal method that can't be called directly or something. >> >> Yes, the usage here of File::from_ptr could probably be hidden inside a >> suitably designed file_operations wrapper. The thing is, Rust Binder >> doesn't currently use such a wrapper at all. It just exports a global of >> type file_operations and the C code in binderfs then references that >> global. > > Yeah. > >> >> Rust Binder used to use such an abstraction, but I ripped it out before >> sending the Rust Binder RFC because it didn't actually help. It was >> designed for cases where the file system is also implemented in Rust, so >> to get it to expose a file_operations global to the C code in binderfs, >> I had to reach inside its internal implementation. It did not save me >> from doing stuff such as using File::from_ptr from Binder. >> >> Now, you could most likely modify those file_operations abstractions to >> support my use-case better. But calling into C is already unsafe, so >> unless we get multiple drivers that have a similar C/Rust split, it's >> not clear that it's useful to extract the logic from Binder. I would >> prefer to wait for the file_operations abstraction to get upstreamed by >> the people working on VFS bindings, and then we can decide whether we >> should rewrite binderfs into Rust and get rid of the logic, or whether >> it's worth to expand the file_operations abstraction to support split >> C/Rust drivers like the current binderfs. >> >>> So what I mean is. fdget() may or may not take a reference. The C >>> interface internally knows whether a reference is owned or not by >>> abusing the lower two bits in a pointer to keep track of that. Naively, >>> I would expect the same information to be available to rust so that it's >>> clear to Rust wheter it's dealing with an explicitly referenced file or >>> an elided-reference file. Maybe that's not possible and I'm not >>> well-versed enough to see that yet. >> >> I'm sure Rust can access the same information, but I don't think I'm >> currently doing anything that cares about the distinction? > > Ok. My main goal is that we end up with an almost 1:1 correspondence > between the Rust and C interface so it's easy for current maintainers > and developers that don't want to care about Rust to continue to do so > and also just somewhat verify that changes they do are sane. Sure, that goal makes total sense to me. Alice
On Tue, Apr 02, 2024 at 09:39:57AM +0000, Alice Ryhl wrote: > Christian Brauner <brauner@kernel.org> wrote: > > On Mon, Apr 01, 2024 at 12:09:08PM +0000, Alice Ryhl wrote: > >> Christian Brauner <brauner@kernel.org> wrote: > >>> On Wed, Mar 20, 2024 at 06:09:05PM +0000, Alice Ryhl wrote: > >>>> Christian Brauner <brauner@kernel.org> wrote: > >>>>> On Fri, Feb 09, 2024 at 11:18:16AM +0000, Alice Ryhl wrote: > >>>>>> +/// Wraps the kernel's `struct file`. > >>>>>> +/// > >>>>>> +/// This represents an open file rather than a file on a filesystem. Processes generally reference > >>>>>> +/// open files using file descriptors. However, file descriptors are not the same as files. A file > >>>>>> +/// descriptor is just an integer that corresponds to a file, and a single file may be referenced > >>>>>> +/// by multiple file descriptors. > >>>>>> +/// > >>>>>> +/// # Refcounting > >>>>>> +/// > >>>>>> +/// Instances of this type are reference-counted. The reference count is incremented by the > >>>>>> +/// `fget`/`get_file` functions and decremented by `fput`. The Rust type `ARef<File>` represents a > >>>>>> +/// pointer that owns a reference count on the file. > >>>>>> +/// > >>>>>> +/// Whenever a process opens a file descriptor (fd), it stores a pointer to the file in its `struct > >>>>>> +/// files_struct`. This pointer owns a reference count to the file, ensuring the file isn't > >>>>>> +/// prematurely deleted while the file descriptor is open. In Rust terminology, the pointers in > >>>>>> +/// `struct files_struct` are `ARef<File>` pointers. > >>>>>> +/// > >>>>>> +/// ## Light refcounts > >>>>>> +/// > >>>>>> +/// Whenever a process has an fd to a file, it may use something called a "light refcount" as a > >>>>>> +/// performance optimization. Light refcounts are acquired by calling `fdget` and released with > >>>>>> +/// `fdput`. The idea behind light refcounts is that if the fd is not closed between the calls to > >>>>>> +/// `fdget` and `fdput`, then the refcount cannot hit zero during that time, as the `struct > >>>>>> +/// files_struct` holds a reference until the fd is closed. This means that it's safe to access the > >>>>>> +/// file even if `fdget` does not increment the refcount. > >>>>>> +/// > >>>>>> +/// The requirement that the fd is not closed during a light refcount applies globally across all > >>>>>> +/// threads - not just on the thread using the light refcount. For this reason, light refcounts are > >>>>>> +/// only used when the `struct files_struct` is not shared with other threads, since this ensures > >>>>>> +/// that other unrelated threads cannot suddenly start using the fd and close it. Therefore, > >>>>>> +/// calling `fdget` on a shared `struct files_struct` creates a normal refcount instead of a light > >>>>>> +/// refcount. > >>>>> > >>>>> When the fdget() calling task doesn't have a shared file descriptor > >>>>> table fdget() will not increment the reference count, yes. This > >>>>> also implies that you cannot have task A use fdget() and then pass > >>>>> f.file to task B that holds on to it while A returns to userspace. It's > >>>>> irrelevant that task A won't drop the reference count or that task B > >>>>> won't drop the reference count. Because task A could return back to > >>>>> userspace and immediately close the fd via a regular close() system call > >>>>> at which point task B has a UAF. In other words a file that has been > >>>>> gotten via fdget() can't be Send to another task without the Send > >>>>> implying taking a reference to it. > >>>> > >>>> That matches my understanding. > >>>> > >>>> I suppose that technically you can still send it to another thread *if* you > >>>> ensure that the current thread waits until that other thread stops using the > >>>> file before returning to userspace. > >>> > >>> _Technically_ yes, but it would be brittle as hell. The problem is that > >>> fdget() _relies_ on being single-threaded for the time that fd is used > >>> until fdput(). There's locking assumptions that build on that e.g., for > >>> concurrent read/write. So no, that shouldn't be allowed. > >>> > >>> Look at how this broke our back when we introduced pidfd_getfd() where > >>> we steal an fd from another task. I have a lengthy explanation how that > >>> can be used to violate our elided-locking which is based on assuming > >>> that we're always single-threaded and the file can't be suddenly shared > >>> with another task. So maybe doable but it would make the semantics even > >>> more intricate. > >> > >> Hmm, the part about elided locking is surprising to me, and may be an > >> issue. Can you give more details on that? Because the current > > > > So what I referred to was that we do have fdget_pos(). Roughly, if > > there's more than one reference on the file then we need to acquire a > > mutex but if it's only a single reference then we can avoid taking the > > mutex because we know that we're the only one that has a reference to > > that file and no one else can acquire one. Whether or not that mutex was > > taken is taken track of in struct fd. > > > > So you can't share a file after fdget_pos() has been called on it and > > you haven't taken the position mutex. So let's say you had: > > > > * Tread A that calls fdget_pos() on file1 and the reference count is > > one. So Thread A doesn't acquire the file position mutex for file1. > > * Now somehow that file1 becomes shared, e.g., Thread B calls fget() on > > it and now Thread B does some operation that requires the file > > position mutex. > > => Thread A and Thread B race on the file position. > > > > So just because you have a reference to a file from somewhere it doesn't > > mean you can just share it with another thread. > > > > So if yo have an arbitrary reference to a file in Rust and that somehow > > can be shared with another thread you risk races here. > > > >> abstractions here *do* actually allow what I described, since we > >> implement Sync for File. > >> > >> I'm not familiar with the pidfd_getfd discussion you are referring to. > >> Do you have a link? > > > > https://lore.kernel.org/linux-fsdevel/20230724-vfs-fdget_pos-v1-1-a4abfd7103f3@kernel.org > > > > pidfd_getfd() can be used to steal a file descriptor from another task. > > It's like a non-cooperative SCM_RIGHTS. That means you can have exactly > > the scenario described above where a file assumed to be non-shared is > > suddenly shared and you have racing reads/writes. > > > > For readdir we nowadays always take the file position mutex because of > > the pidfd_getfd() business because that might corrupt internal state. > > > >> > >> I'm thinking that we may have to provide two different `struct file` > >> wrappers to accurately model this API in Rust. Perhaps they could be > >> called File and LocalFile, where one is marked as thread safe and the > >> other isn't. I can make all LocalFile methods available on File to avoid > >> having to duplicate methods that are available on both. > > > > But isn't that just struct file and struct fd? Ideally we'd stay close > > to something like this. > > Right, that kind of naming seems sensible. But I still need to > understand the details a bit better. See below on fdget_pos. > > >> But it's not clear to me that this is even enough. Even if we give you a > >> &LocalFile to prevent you from moving it across threads, you can just > >> call File::fget to get an ARef<File> to the same file and then move > >> *that* across threads. > > > > Yes, absolutely. > > One of my challenges is that Binder wants to call File::fget, > immediately move it to another thread, and then call fd_install. And > it would be pretty unfortunate if that requires unsafe. But like I argue > below, it seems hard to design a safe API for this in the face of > fdget_pos. > > >> This kind of global requirement is not so easy to model. Maybe klint [1] > >> could do it ... atomic context violations are a similar kind of global > >> check. But having klint do it would be far out. > >> > >> Or maybe File::fget should also return a LocalFile? > >> > >> But this raises a different question to me. Let's say process A uses > >> Binder to send its own fd to process B, and the following things happen: > >> > >> 1. Process A enters the ioctl and takes fdget on the fd. > >> 2. Process A calls fget on the same fd to send it to another process. > >> 3. Process A goes to sleep, waiting for process B to respond. > >> 4. Process B receives the message, installs the fd, and returns to userspace. > >> 5. Process B responds to the transaction, but does not close the fd. > > > > The fd just installed in 4. and the fd you're referring to in 5. are > > identical, right? IOW, we're not talking about two different fd (dup) > > for the same file, right? > > I'm referring to whatever fd_install does given the `struct file` I got > from fget in step 2. > > >> 6a. Process A finishes sleeping, and returns to userspace from the ioctl. > >> 6b. Process B tries to do an operation (e.g. read) on the fd. > >> > >> Let's say that 6a and 6b run in parallel. > >> > >> Could this potentially result in a data race between step 6a and 6b? I'm > >> guessing that step 6a probably doesn't touch any of the code that has > >> elided locking assumptions, so in practice I guess there's not a problem > >> ... but if you make any sort of elided locking assumption as you exit > >> from the ioctl (before reaching the fdput), then it seems to me that you > >> have a problem. > > > > Yes, 6a doesn't touch any code that has elided locking assumptions. > > > > 1'. Process A enters the ioctl and takes fdget() on the fd. Process A > > holds the only reference to that file and the file descriptor table > > isn't shared. Therefore, f_count is left untouched and remains at 1. > > 2'. Process A calls fget() which unconditionally bumps f_count bringing > > it to 2 and sending it another process (Presumably you intend to > > imply that this reference is now owned by the second process.). > > 3'. [as 3.] > > 4'. Process B installs the file into it's file descriptor table > > consuming that reference from 2'. The f_count remains at 2 with the > > reference from 2' now being owned by Process B. > > 5'. Since Process B isn't closing the fd and has just called > > fd_install() it returns to userspace with f_count untouched and > > still at 2. > > 6'a. Process A finishes sleeping and returns to userspace calling > > fdput(). Since the original fdget() was done without bumping the > > reference count the fdput() of Process A will not decrement the > > reference count. So f_count remains at 2. > > 6'b. Process B performs a read/write syscall and calls fdget_pos(). > > fdget_pos() sees that this file has f_count > 1 and takes the > > file position mutex. > > > > So this isn't a problem. The problem is when a file becomes shared > > implicitly without the original owner of the file knowing. > > Hmm. Yes, but the ioctl code that called fdget doesn't really know that > the ioctl shared the file? So why is it okay? Why does it matter to the ioctl() code itself? The ioctl() code never calls fdget_pos(). > > It really seems like there are two different things going on here. When > it comes to fdget, we only really care about operations that could > remove it from the local file descriptor table, since fdget relies on > the refcount in that table remaining valid until fdput. Yes. > > On the other hand, for fdget_pos it also matters whether it gets > installed in other file descriptor tables. Threads that reference it > through a different fd table will still access the same position. Yes, they operate on the same f_pos. > > And so this means that between fdget/fdput, there's never any problem > with installing the `struct file` into another file descriptor table. > Nothing you can do from that other fd table could cause the local fd > table to drop its refcount on the file. Whereas such an install can be > a problem between fdget_pos/fdput_pos, since that could introduce a race > on the position. > > Is this correct? Yes, but that would imply you're sharing and installing a file into a file descriptor table from a read/write/seek codepath. I don't see how this can happen without something like e.g., pidfd_getfd(). And the fd_install()ing task would then have to go back to userspace and issue a concurrent read/write/seek system call while the other thread is still reading/writing. Overall, we really only care about f_pos consistency because posix requires atomicity between reads/writes/seeks. For pidfd_getfd() where such sharing can happen non-cooperatively we just don't care as we've just declared this to be an instance where we're outside of posix guarantees. And for readdir() we unconditionally acquire the mutex. I think io_uring is racing on f_pos as well under certain circumstances (REQ_F_CUR_POS?) as they don't use fdget_pos() at all. But iirc Jens dislikes that they ever allowed that. > > I was thinking that if we have some sort of File/LocalFile distinction > (or File/Fd), then we may be able to get it to work by limiting what a > File can do. For example, let's say that the only thing you can do with > a File is install it into fd tables, then by the previous logic, there's > no problem with it being safe to move across threads even if there's an > active fdget. > > But the fdget_pos kind of throws a wrench into that, because now I can > no longer say "it's always safe to do File::fget, move it to another > thread, and install it into the remote fd table", since that could cause > races on the position if there's an active fdget_pos when we call > File::fget. I think I understand why that's a problem for you but let me try and spell it out so I can understand where you're coming from: You want the Rust compiler to reject a file becoming shared implicitly from any codepath that is beneath fdget_pos() even though no such code yet exists (ignoring pidfd_getfd()). So it's a correctness issue to you. > > >>>>>> +/// > >>>>>> +/// Light reference counts must be released with `fdput` before the system call returns to > >>>>>> +/// userspace. This means that if you wait until the current system call returns to userspace, then > >>>>>> +/// all light refcounts that existed at the time have gone away. > >>>>>> +/// > >>>>>> +/// ## Rust references > >>>>>> +/// > >>>>>> +/// The reference type `&File` is similar to light refcounts: > >>>>>> +/// > >>>>>> +/// * `&File` references don't own a reference count. They can only exist as long as the reference > >>>>>> +/// count stays positive, and can only be created when there is some mechanism in place to ensure > >>>>>> +/// this. > >>>>>> +/// > >>>>>> +/// * The Rust borrow-checker normally ensures this by enforcing that the `ARef<File>` from which > >>>>>> +/// a `&File` is created outlives the `&File`. > >>>>> > >>>>> The section confuses me a little: Does the borrow-checker always ensure > >>>>> that a &File stays valid or are there circumstances where it doesn't or > >>>>> are you saying it doesn't enforce it? > >>>> > >>>> The borrow-checker always ensures it. > >>> > >>> Ok, thanks. > >>> > >>>> > >>>> A &File is actually short-hand for &'a File, where 'a is some > >>>> unspecified lifetime. We say that &'a File is annotated with 'a. The > >>>> borrow-checker rejects any code that tries to use a reference after the > >>>> end of the lifetime annotated on it. > >>> > >>> Thanks for the explanation. > >>> > >>>> > >>>> So as long as you annotate the reference with a sufficiently short > >>>> lifetime, the borrow checker will prevent UAF. And outside of cases like > >>> > >>> Sorry, but can you explain "sufficiently short lifetime"? > >> > >> By "sufficiently short lifetime" I mean "lifetime that ends before the > >> object is destroyed". > > > > Ah, ok. It sounded like it was a specific concept that Rust is > > implementing in contrast to long-term lifetime or sm. Thanks! > > > >> > >> Idea being that if the lifetime ends before the object is freed, and the > >> borrow-checker rejects attempts to use it after the lifetime ends, then > >> it follows that the borrow-checker prevents use-after-frees. > >> > >>>> from_ptr, the borrow-checker also takes care of ensuring that the > >>>> lifetimes are sufficiently short. > >>>> > >>>> (Technically &'a File and &'b File are two completely different types, > >>>> so &File is technically a class of types and not a single type. Rust > >>>> will automatically convert &'long File to &'short File.) > >>>> > >>>>>> +/// > >>>>>> +/// * Using the unsafe [`File::from_ptr`] means that it is up to the caller to ensure that the > >>>>>> +/// `&File` only exists while the reference count is positive. > >>>>> > >>>>> What is this used for in binder? If it's not used don't add it. > >>>> > >>>> This is used on the boundary between the Rust part of Binder and the > >>>> binderfs component that is implemented in C. For example: > >>> > >>> I see, I'm being foiled by my own code... > >>> > >>>> > >>>> unsafe extern "C" fn rust_binder_open( > >>>> inode: *mut bindings::inode, > >>>> file_ptr: *mut bindings::file, > >>>> ) -> core::ffi::c_int { > >>>> // SAFETY: The `rust_binderfs.c` file ensures that `i_private` is set to the return value of a > >>>> // successful call to `rust_binder_new_device`. > >>>> let ctx = unsafe { Arc::<Context>::borrow((*inode).i_private) }; > >>>> > >>>> // SAFETY: The caller provides a valid file pointer to a new `struct file`. > >>>> let file = unsafe { File::from_ptr(file_ptr) }; > >>> > >>> We need a better name for this helper than from_ptr() imho. I think > >>> from_ptr() and as_ptr() is odd for C. How weird would it be to call > >>> that from_raw_file() and as_raw_file()? > >> > >> That's a reasonable name. I would be happy to rename to that, and I > >> don't think it is unidiomatic. > > > > Thanks! > > > >> > >>> But bigger picture I somewhat struggle with the semantics of this > >>> because this is not an interface that we have in C and this is really > >>> about a low-level contract between C and Rust. Specifically this states > >>> that this pointer is _somehow_ guaranteed valid. And really, this seems > >>> a bit of a hack. > >> > >> Indeed ... but I think it's a quite common hack. After all, any time you > >> dereference a raw pointer in Rust, you are making the same assumption. > >> > >>> Naively, I think this should probably not be necessary if > >>> file_operations are properly wrapped. Or it should at least be demotable > >>> to a purely internal method that can't be called directly or something. > >> > >> Yes, the usage here of File::from_ptr could probably be hidden inside a > >> suitably designed file_operations wrapper. The thing is, Rust Binder > >> doesn't currently use such a wrapper at all. It just exports a global of > >> type file_operations and the C code in binderfs then references that > >> global. > > > > Yeah. > > > >> > >> Rust Binder used to use such an abstraction, but I ripped it out before > >> sending the Rust Binder RFC because it didn't actually help. It was > >> designed for cases where the file system is also implemented in Rust, so > >> to get it to expose a file_operations global to the C code in binderfs, > >> I had to reach inside its internal implementation. It did not save me > >> from doing stuff such as using File::from_ptr from Binder. > >> > >> Now, you could most likely modify those file_operations abstractions to > >> support my use-case better. But calling into C is already unsafe, so > >> unless we get multiple drivers that have a similar C/Rust split, it's > >> not clear that it's useful to extract the logic from Binder. I would > >> prefer to wait for the file_operations abstraction to get upstreamed by > >> the people working on VFS bindings, and then we can decide whether we > >> should rewrite binderfs into Rust and get rid of the logic, or whether > >> it's worth to expand the file_operations abstraction to support split > >> C/Rust drivers like the current binderfs. > >> > >>> So what I mean is. fdget() may or may not take a reference. The C > >>> interface internally knows whether a reference is owned or not by > >>> abusing the lower two bits in a pointer to keep track of that. Naively, > >>> I would expect the same information to be available to rust so that it's > >>> clear to Rust wheter it's dealing with an explicitly referenced file or > >>> an elided-reference file. Maybe that's not possible and I'm not > >>> well-versed enough to see that yet. > >> > >> I'm sure Rust can access the same information, but I don't think I'm > >> currently doing anything that cares about the distinction? > > > > Ok. My main goal is that we end up with an almost 1:1 correspondence > > between the Rust and C interface so it's easy for current maintainers > > and developers that don't want to care about Rust to continue to do so > > and also just somewhat verify that changes they do are sane. > > Sure, that goal makes total sense to me. > > Alice
On Wed, Apr 3, 2024 at 8:02 AM Christian Brauner <brauner@kernel.org> wrote: > > On Tue, Apr 02, 2024 at 09:39:57AM +0000, Alice Ryhl wrote: > > Christian Brauner <brauner@kernel.org> wrote: > > > On Mon, Apr 01, 2024 at 12:09:08PM +0000, Alice Ryhl wrote: > > >> Christian Brauner <brauner@kernel.org> wrote: > > >>> On Wed, Mar 20, 2024 at 06:09:05PM +0000, Alice Ryhl wrote: > > >>>> Christian Brauner <brauner@kernel.org> wrote: > > >>>>> On Fri, Feb 09, 2024 at 11:18:16AM +0000, Alice Ryhl wrote: > > >>>>>> +/// Wraps the kernel's `struct file`. > > >>>>>> +/// > > >>>>>> +/// This represents an open file rather than a file on a filesystem. Processes generally reference > > >>>>>> +/// open files using file descriptors. However, file descriptors are not the same as files. A file > > >>>>>> +/// descriptor is just an integer that corresponds to a file, and a single file may be referenced > > >>>>>> +/// by multiple file descriptors. > > >>>>>> +/// > > >>>>>> +/// # Refcounting > > >>>>>> +/// > > >>>>>> +/// Instances of this type are reference-counted. The reference count is incremented by the > > >>>>>> +/// `fget`/`get_file` functions and decremented by `fput`. The Rust type `ARef<File>` represents a > > >>>>>> +/// pointer that owns a reference count on the file. > > >>>>>> +/// > > >>>>>> +/// Whenever a process opens a file descriptor (fd), it stores a pointer to the file in its `struct > > >>>>>> +/// files_struct`. This pointer owns a reference count to the file, ensuring the file isn't > > >>>>>> +/// prematurely deleted while the file descriptor is open. In Rust terminology, the pointers in > > >>>>>> +/// `struct files_struct` are `ARef<File>` pointers. > > >>>>>> +/// > > >>>>>> +/// ## Light refcounts > > >>>>>> +/// > > >>>>>> +/// Whenever a process has an fd to a file, it may use something called a "light refcount" as a > > >>>>>> +/// performance optimization. Light refcounts are acquired by calling `fdget` and released with > > >>>>>> +/// `fdput`. The idea behind light refcounts is that if the fd is not closed between the calls to > > >>>>>> +/// `fdget` and `fdput`, then the refcount cannot hit zero during that time, as the `struct > > >>>>>> +/// files_struct` holds a reference until the fd is closed. This means that it's safe to access the > > >>>>>> +/// file even if `fdget` does not increment the refcount. > > >>>>>> +/// > > >>>>>> +/// The requirement that the fd is not closed during a light refcount applies globally across all > > >>>>>> +/// threads - not just on the thread using the light refcount. For this reason, light refcounts are > > >>>>>> +/// only used when the `struct files_struct` is not shared with other threads, since this ensures > > >>>>>> +/// that other unrelated threads cannot suddenly start using the fd and close it. Therefore, > > >>>>>> +/// calling `fdget` on a shared `struct files_struct` creates a normal refcount instead of a light > > >>>>>> +/// refcount. > > >>>>> > > >>>>> When the fdget() calling task doesn't have a shared file descriptor > > >>>>> table fdget() will not increment the reference count, yes. This > > >>>>> also implies that you cannot have task A use fdget() and then pass > > >>>>> f.file to task B that holds on to it while A returns to userspace. It's > > >>>>> irrelevant that task A won't drop the reference count or that task B > > >>>>> won't drop the reference count. Because task A could return back to > > >>>>> userspace and immediately close the fd via a regular close() system call > > >>>>> at which point task B has a UAF. In other words a file that has been > > >>>>> gotten via fdget() can't be Send to another task without the Send > > >>>>> implying taking a reference to it. > > >>>> > > >>>> That matches my understanding. > > >>>> > > >>>> I suppose that technically you can still send it to another thread *if* you > > >>>> ensure that the current thread waits until that other thread stops using the > > >>>> file before returning to userspace. > > >>> > > >>> _Technically_ yes, but it would be brittle as hell. The problem is that > > >>> fdget() _relies_ on being single-threaded for the time that fd is used > > >>> until fdput(). There's locking assumptions that build on that e.g., for > > >>> concurrent read/write. So no, that shouldn't be allowed. > > >>> > > >>> Look at how this broke our back when we introduced pidfd_getfd() where > > >>> we steal an fd from another task. I have a lengthy explanation how that > > >>> can be used to violate our elided-locking which is based on assuming > > >>> that we're always single-threaded and the file can't be suddenly shared > > >>> with another task. So maybe doable but it would make the semantics even > > >>> more intricate. > > >> > > >> Hmm, the part about elided locking is surprising to me, and may be an > > >> issue. Can you give more details on that? Because the current > > > > > > So what I referred to was that we do have fdget_pos(). Roughly, if > > > there's more than one reference on the file then we need to acquire a > > > mutex but if it's only a single reference then we can avoid taking the > > > mutex because we know that we're the only one that has a reference to > > > that file and no one else can acquire one. Whether or not that mutex was > > > taken is taken track of in struct fd. > > > > > > So you can't share a file after fdget_pos() has been called on it and > > > you haven't taken the position mutex. So let's say you had: > > > > > > * Tread A that calls fdget_pos() on file1 and the reference count is > > > one. So Thread A doesn't acquire the file position mutex for file1. > > > * Now somehow that file1 becomes shared, e.g., Thread B calls fget() on > > > it and now Thread B does some operation that requires the file > > > position mutex. > > > => Thread A and Thread B race on the file position. > > > > > > So just because you have a reference to a file from somewhere it doesn't > > > mean you can just share it with another thread. > > > > > > So if yo have an arbitrary reference to a file in Rust and that somehow > > > can be shared with another thread you risk races here. > > > > > >> abstractions here *do* actually allow what I described, since we > > >> implement Sync for File. > > >> > > >> I'm not familiar with the pidfd_getfd discussion you are referring to. > > >> Do you have a link? > > > > > > https://lore.kernel.org/linux-fsdevel/20230724-vfs-fdget_pos-v1-1-a4abfd7103f3@kernel.org > > > > > > pidfd_getfd() can be used to steal a file descriptor from another task. > > > It's like a non-cooperative SCM_RIGHTS. That means you can have exactly > > > the scenario described above where a file assumed to be non-shared is > > > suddenly shared and you have racing reads/writes. > > > > > > For readdir we nowadays always take the file position mutex because of > > > the pidfd_getfd() business because that might corrupt internal state. > > > > > >> > > >> I'm thinking that we may have to provide two different `struct file` > > >> wrappers to accurately model this API in Rust. Perhaps they could be > > >> called File and LocalFile, where one is marked as thread safe and the > > >> other isn't. I can make all LocalFile methods available on File to avoid > > >> having to duplicate methods that are available on both. > > > > > > But isn't that just struct file and struct fd? Ideally we'd stay close > > > to something like this. > > > > Right, that kind of naming seems sensible. But I still need to > > understand the details a bit better. See below on fdget_pos. > > > > >> But it's not clear to me that this is even enough. Even if we give you a > > >> &LocalFile to prevent you from moving it across threads, you can just > > >> call File::fget to get an ARef<File> to the same file and then move > > >> *that* across threads. > > > > > > Yes, absolutely. > > > > One of my challenges is that Binder wants to call File::fget, > > immediately move it to another thread, and then call fd_install. And > > it would be pretty unfortunate if that requires unsafe. But like I argue > > below, it seems hard to design a safe API for this in the face of > > fdget_pos. > > > > >> This kind of global requirement is not so easy to model. Maybe klint [1] > > >> could do it ... atomic context violations are a similar kind of global > > >> check. But having klint do it would be far out. > > >> > > >> Or maybe File::fget should also return a LocalFile? > > >> > > >> But this raises a different question to me. Let's say process A uses > > >> Binder to send its own fd to process B, and the following things happen: > > >> > > >> 1. Process A enters the ioctl and takes fdget on the fd. > > >> 2. Process A calls fget on the same fd to send it to another process. > > >> 3. Process A goes to sleep, waiting for process B to respond. > > >> 4. Process B receives the message, installs the fd, and returns to userspace. > > >> 5. Process B responds to the transaction, but does not close the fd. > > > > > > The fd just installed in 4. and the fd you're referring to in 5. are > > > identical, right? IOW, we're not talking about two different fd (dup) > > > for the same file, right? > > > > I'm referring to whatever fd_install does given the `struct file` I got > > from fget in step 2. > > > > >> 6a. Process A finishes sleeping, and returns to userspace from the ioctl. > > >> 6b. Process B tries to do an operation (e.g. read) on the fd. > > >> > > >> Let's say that 6a and 6b run in parallel. > > >> > > >> Could this potentially result in a data race between step 6a and 6b? I'm > > >> guessing that step 6a probably doesn't touch any of the code that has > > >> elided locking assumptions, so in practice I guess there's not a problem > > >> ... but if you make any sort of elided locking assumption as you exit > > >> from the ioctl (before reaching the fdput), then it seems to me that you > > >> have a problem. > > > > > > Yes, 6a doesn't touch any code that has elided locking assumptions. > > > > > > 1'. Process A enters the ioctl and takes fdget() on the fd. Process A > > > holds the only reference to that file and the file descriptor table > > > isn't shared. Therefore, f_count is left untouched and remains at 1. > > > 2'. Process A calls fget() which unconditionally bumps f_count bringing > > > it to 2 and sending it another process (Presumably you intend to > > > imply that this reference is now owned by the second process.). > > > 3'. [as 3.] > > > 4'. Process B installs the file into it's file descriptor table > > > consuming that reference from 2'. The f_count remains at 2 with the > > > reference from 2' now being owned by Process B. > > > 5'. Since Process B isn't closing the fd and has just called > > > fd_install() it returns to userspace with f_count untouched and > > > still at 2. > > > 6'a. Process A finishes sleeping and returns to userspace calling > > > fdput(). Since the original fdget() was done without bumping the > > > reference count the fdput() of Process A will not decrement the > > > reference count. So f_count remains at 2. > > > 6'b. Process B performs a read/write syscall and calls fdget_pos(). > > > fdget_pos() sees that this file has f_count > 1 and takes the > > > file position mutex. > > > > > > So this isn't a problem. The problem is when a file becomes shared > > > implicitly without the original owner of the file knowing. > > > > Hmm. Yes, but the ioctl code that called fdget doesn't really know that > > the ioctl shared the file? So why is it okay? > > Why does it matter to the ioctl() code itself? The ioctl() code never > calls fdget_pos(). > > > > > It really seems like there are two different things going on here. When > > it comes to fdget, we only really care about operations that could > > remove it from the local file descriptor table, since fdget relies on > > the refcount in that table remaining valid until fdput. > > Yes. > > > > > On the other hand, for fdget_pos it also matters whether it gets > > installed in other file descriptor tables. Threads that reference it > > through a different fd table will still access the same position. > > Yes, they operate on the same f_pos. > > > > > And so this means that between fdget/fdput, there's never any problem > > with installing the `struct file` into another file descriptor table. > > Nothing you can do from that other fd table could cause the local fd > > table to drop its refcount on the file. Whereas such an install can be > > a problem between fdget_pos/fdput_pos, since that could introduce a race > > on the position. > > > > Is this correct? > > Yes, but that would imply you're sharing and installing a file into a > file descriptor table from a read/write/seek codepath. I don't see how > this can happen without something like e.g., pidfd_getfd(). And the > fd_install()ing task would then have to go back to userspace and issue a > concurrent read/write/seek system call while the other thread is still > reading/writing. > > Overall, we really only care about f_pos consistency because posix > requires atomicity between reads/writes/seeks. For pidfd_getfd() where > such sharing can happen non-cooperatively we just don't care as we've > just declared this to be an instance where we're outside of posix > guarantees. And for readdir() we unconditionally acquire the mutex. > > I think io_uring is racing on f_pos as well under certain circumstances > (REQ_F_CUR_POS?) as they don't use fdget_pos() at all. But iirc Jens > dislikes that they ever allowed that. > > > > > I was thinking that if we have some sort of File/LocalFile distinction > > (or File/Fd), then we may be able to get it to work by limiting what a > > File can do. For example, let's say that the only thing you can do with > > a File is install it into fd tables, then by the previous logic, there's > > no problem with it being safe to move across threads even if there's an > > active fdget. > > > > But the fdget_pos kind of throws a wrench into that, because now I can > > no longer say "it's always safe to do File::fget, move it to another > > thread, and install it into the remote fd table", since that could cause > > races on the position if there's an active fdget_pos when we call > > File::fget. > > I think I understand why that's a problem for you but let me try and > spell it out so I can understand where you're coming from: > > You want the Rust compiler to reject a file becoming shared implicitly > from any codepath that is beneath fdget_pos() even though no such code > yet exists (ignoring pidfd_getfd()). So it's a correctness issue to you. Yes, exactly. One of the design principles behind Rust is that if an API allows you to trigger memory unsafety, then it *must* be the case that you called an unsafe function somewhere, and that this call violated the documented safety requirements for that unsafe function. So if the File API provides a safe interface that you can use to trigger memory unsafety, then that's a correctness issue for me. Thanks for asking this. I should have clarified this previously. > > >>>>>> +/// > > >>>>>> +/// Light reference counts must be released with `fdput` before the system call returns to > > >>>>>> +/// userspace. This means that if you wait until the current system call returns to userspace, then > > >>>>>> +/// all light refcounts that existed at the time have gone away. > > >>>>>> +/// > > >>>>>> +/// ## Rust references > > >>>>>> +/// > > >>>>>> +/// The reference type `&File` is similar to light refcounts: > > >>>>>> +/// > > >>>>>> +/// * `&File` references don't own a reference count. They can only exist as long as the reference > > >>>>>> +/// count stays positive, and can only be created when there is some mechanism in place to ensure > > >>>>>> +/// this. > > >>>>>> +/// > > >>>>>> +/// * The Rust borrow-checker normally ensures this by enforcing that the `ARef<File>` from which > > >>>>>> +/// a `&File` is created outlives the `&File`. > > >>>>> > > >>>>> The section confuses me a little: Does the borrow-checker always ensure > > >>>>> that a &File stays valid or are there circumstances where it doesn't or > > >>>>> are you saying it doesn't enforce it? > > >>>> > > >>>> The borrow-checker always ensures it. > > >>> > > >>> Ok, thanks. > > >>> > > >>>> > > >>>> A &File is actually short-hand for &'a File, where 'a is some > > >>>> unspecified lifetime. We say that &'a File is annotated with 'a. The > > >>>> borrow-checker rejects any code that tries to use a reference after the > > >>>> end of the lifetime annotated on it. > > >>> > > >>> Thanks for the explanation. > > >>> > > >>>> > > >>>> So as long as you annotate the reference with a sufficiently short > > >>>> lifetime, the borrow checker will prevent UAF. And outside of cases like > > >>> > > >>> Sorry, but can you explain "sufficiently short lifetime"? > > >> > > >> By "sufficiently short lifetime" I mean "lifetime that ends before the > > >> object is destroyed". > > > > > > Ah, ok. It sounded like it was a specific concept that Rust is > > > implementing in contrast to long-term lifetime or sm. Thanks! > > > > > >> > > >> Idea being that if the lifetime ends before the object is freed, and the > > >> borrow-checker rejects attempts to use it after the lifetime ends, then > > >> it follows that the borrow-checker prevents use-after-frees. > > >> > > >>>> from_ptr, the borrow-checker also takes care of ensuring that the > > >>>> lifetimes are sufficiently short. > > >>>> > > >>>> (Technically &'a File and &'b File are two completely different types, > > >>>> so &File is technically a class of types and not a single type. Rust > > >>>> will automatically convert &'long File to &'short File.) > > >>>> > > >>>>>> +/// > > >>>>>> +/// * Using the unsafe [`File::from_ptr`] means that it is up to the caller to ensure that the > > >>>>>> +/// `&File` only exists while the reference count is positive. > > >>>>> > > >>>>> What is this used for in binder? If it's not used don't add it. > > >>>> > > >>>> This is used on the boundary between the Rust part of Binder and the > > >>>> binderfs component that is implemented in C. For example: > > >>> > > >>> I see, I'm being foiled by my own code... > > >>> > > >>>> > > >>>> unsafe extern "C" fn rust_binder_open( > > >>>> inode: *mut bindings::inode, > > >>>> file_ptr: *mut bindings::file, > > >>>> ) -> core::ffi::c_int { > > >>>> // SAFETY: The `rust_binderfs.c` file ensures that `i_private` is set to the return value of a > > >>>> // successful call to `rust_binder_new_device`. > > >>>> let ctx = unsafe { Arc::<Context>::borrow((*inode).i_private) }; > > >>>> > > >>>> // SAFETY: The caller provides a valid file pointer to a new `struct file`. > > >>>> let file = unsafe { File::from_ptr(file_ptr) }; > > >>> > > >>> We need a better name for this helper than from_ptr() imho. I think > > >>> from_ptr() and as_ptr() is odd for C. How weird would it be to call > > >>> that from_raw_file() and as_raw_file()? > > >> > > >> That's a reasonable name. I would be happy to rename to that, and I > > >> don't think it is unidiomatic. > > > > > > Thanks! > > > > > >> > > >>> But bigger picture I somewhat struggle with the semantics of this > > >>> because this is not an interface that we have in C and this is really > > >>> about a low-level contract between C and Rust. Specifically this states > > >>> that this pointer is _somehow_ guaranteed valid. And really, this seems > > >>> a bit of a hack. > > >> > > >> Indeed ... but I think it's a quite common hack. After all, any time you > > >> dereference a raw pointer in Rust, you are making the same assumption. > > >> > > >>> Naively, I think this should probably not be necessary if > > >>> file_operations are properly wrapped. Or it should at least be demotable > > >>> to a purely internal method that can't be called directly or something. > > >> > > >> Yes, the usage here of File::from_ptr could probably be hidden inside a > > >> suitably designed file_operations wrapper. The thing is, Rust Binder > > >> doesn't currently use such a wrapper at all. It just exports a global of > > >> type file_operations and the C code in binderfs then references that > > >> global. > > > > > > Yeah. > > > > > >> > > >> Rust Binder used to use such an abstraction, but I ripped it out before > > >> sending the Rust Binder RFC because it didn't actually help. It was > > >> designed for cases where the file system is also implemented in Rust, so > > >> to get it to expose a file_operations global to the C code in binderfs, > > >> I had to reach inside its internal implementation. It did not save me > > >> from doing stuff such as using File::from_ptr from Binder. > > >> > > >> Now, you could most likely modify those file_operations abstractions to > > >> support my use-case better. But calling into C is already unsafe, so > > >> unless we get multiple drivers that have a similar C/Rust split, it's > > >> not clear that it's useful to extract the logic from Binder. I would > > >> prefer to wait for the file_operations abstraction to get upstreamed by > > >> the people working on VFS bindings, and then we can decide whether we > > >> should rewrite binderfs into Rust and get rid of the logic, or whether > > >> it's worth to expand the file_operations abstraction to support split > > >> C/Rust drivers like the current binderfs. > > >> > > >>> So what I mean is. fdget() may or may not take a reference. The C > > >>> interface internally knows whether a reference is owned or not by > > >>> abusing the lower two bits in a pointer to keep track of that. Naively, > > >>> I would expect the same information to be available to rust so that it's > > >>> clear to Rust wheter it's dealing with an explicitly referenced file or > > >>> an elided-reference file. Maybe that's not possible and I'm not > > >>> well-versed enough to see that yet. > > >> > > >> I'm sure Rust can access the same information, but I don't think I'm > > >> currently doing anything that cares about the distinction? > > > > > > Ok. My main goal is that we end up with an almost 1:1 correspondence > > > between the Rust and C interface so it's easy for current maintainers > > > and developers that don't want to care about Rust to continue to do so > > > and also just somewhat verify that changes they do are sane. > > > > Sure, that goal makes total sense to me. > > > > Alice
Christian Brauner <brauner@kernel.org> wrote: > On Tue, Apr 02, 2024 at 09:39:57AM +0000, Alice Ryhl wrote: >> Christian Brauner <brauner@kernel.org> wrote: >>> On Mon, Apr 01, 2024 at 12:09:08PM +0000, Alice Ryhl wrote: >>>> Christian Brauner <brauner@kernel.org> wrote: >>>>> On Wed, Mar 20, 2024 at 06:09:05PM +0000, Alice Ryhl wrote: >>>>>> Christian Brauner <brauner@kernel.org> wrote: >>>>>>> On Fri, Feb 09, 2024 at 11:18:16AM +0000, Alice Ryhl wrote: >>>>>>>> +/// Wraps the kernel's `struct file`. >>>>>>>> +/// >>>>>>>> +/// This represents an open file rather than a file on a filesystem. Processes generally reference >>>>>>>> +/// open files using file descriptors. However, file descriptors are not the same as files. A file >>>>>>>> +/// descriptor is just an integer that corresponds to a file, and a single file may be referenced >>>>>>>> +/// by multiple file descriptors. >>>>>>>> +/// >>>>>>>> +/// # Refcounting >>>>>>>> +/// >>>>>>>> +/// Instances of this type are reference-counted. The reference count is incremented by the >>>>>>>> +/// `fget`/`get_file` functions and decremented by `fput`. The Rust type `ARef<File>` represents a >>>>>>>> +/// pointer that owns a reference count on the file. >>>>>>>> +/// >>>>>>>> +/// Whenever a process opens a file descriptor (fd), it stores a pointer to the file in its `struct >>>>>>>> +/// files_struct`. This pointer owns a reference count to the file, ensuring the file isn't >>>>>>>> +/// prematurely deleted while the file descriptor is open. In Rust terminology, the pointers in >>>>>>>> +/// `struct files_struct` are `ARef<File>` pointers. >>>>>>>> +/// >>>>>>>> +/// ## Light refcounts >>>>>>>> +/// >>>>>>>> +/// Whenever a process has an fd to a file, it may use something called a "light refcount" as a >>>>>>>> +/// performance optimization. Light refcounts are acquired by calling `fdget` and released with >>>>>>>> +/// `fdput`. The idea behind light refcounts is that if the fd is not closed between the calls to >>>>>>>> +/// `fdget` and `fdput`, then the refcount cannot hit zero during that time, as the `struct >>>>>>>> +/// files_struct` holds a reference until the fd is closed. This means that it's safe to access the >>>>>>>> +/// file even if `fdget` does not increment the refcount. >>>>>>>> +/// >>>>>>>> +/// The requirement that the fd is not closed during a light refcount applies globally across all >>>>>>>> +/// threads - not just on the thread using the light refcount. For this reason, light refcounts are >>>>>>>> +/// only used when the `struct files_struct` is not shared with other threads, since this ensures >>>>>>>> +/// that other unrelated threads cannot suddenly start using the fd and close it. Therefore, >>>>>>>> +/// calling `fdget` on a shared `struct files_struct` creates a normal refcount instead of a light >>>>>>>> +/// refcount. >>>>>>> >>>>>>> When the fdget() calling task doesn't have a shared file descriptor >>>>>>> table fdget() will not increment the reference count, yes. This >>>>>>> also implies that you cannot have task A use fdget() and then pass >>>>>>> f.file to task B that holds on to it while A returns to userspace. It's >>>>>>> irrelevant that task A won't drop the reference count or that task B >>>>>>> won't drop the reference count. Because task A could return back to >>>>>>> userspace and immediately close the fd via a regular close() system call >>>>>>> at which point task B has a UAF. In other words a file that has been >>>>>>> gotten via fdget() can't be Send to another task without the Send >>>>>>> implying taking a reference to it. >>>>>> >>>>>> That matches my understanding. >>>>>> >>>>>> I suppose that technically you can still send it to another thread *if* you >>>>>> ensure that the current thread waits until that other thread stops using the >>>>>> file before returning to userspace. >>>>> >>>>> _Technically_ yes, but it would be brittle as hell. The problem is that >>>>> fdget() _relies_ on being single-threaded for the time that fd is used >>>>> until fdput(). There's locking assumptions that build on that e.g., for >>>>> concurrent read/write. So no, that shouldn't be allowed. >>>>> >>>>> Look at how this broke our back when we introduced pidfd_getfd() where >>>>> we steal an fd from another task. I have a lengthy explanation how that >>>>> can be used to violate our elided-locking which is based on assuming >>>>> that we're always single-threaded and the file can't be suddenly shared >>>>> with another task. So maybe doable but it would make the semantics even >>>>> more intricate. >>>> >>>> Hmm, the part about elided locking is surprising to me, and may be an >>>> issue. Can you give more details on that? Because the current >>> >>> So what I referred to was that we do have fdget_pos(). Roughly, if >>> there's more than one reference on the file then we need to acquire a >>> mutex but if it's only a single reference then we can avoid taking the >>> mutex because we know that we're the only one that has a reference to >>> that file and no one else can acquire one. Whether or not that mutex was >>> taken is taken track of in struct fd. >>> >>> So you can't share a file after fdget_pos() has been called on it and >>> you haven't taken the position mutex. So let's say you had: >>> >>> * Tread A that calls fdget_pos() on file1 and the reference count is >>> one. So Thread A doesn't acquire the file position mutex for file1. >>> * Now somehow that file1 becomes shared, e.g., Thread B calls fget() on >>> it and now Thread B does some operation that requires the file >>> position mutex. >>> => Thread A and Thread B race on the file position. >>> >>> So just because you have a reference to a file from somewhere it doesn't >>> mean you can just share it with another thread. >>> >>> So if yo have an arbitrary reference to a file in Rust and that somehow >>> can be shared with another thread you risk races here. >>> >>>> abstractions here *do* actually allow what I described, since we >>>> implement Sync for File. >>>> >>>> I'm not familiar with the pidfd_getfd discussion you are referring to. >>>> Do you have a link? >>> >>> https://lore.kernel.org/linux-fsdevel/20230724-vfs-fdget_pos-v1-1-a4abfd7103f3@kernel.org >>> >>> pidfd_getfd() can be used to steal a file descriptor from another task. >>> It's like a non-cooperative SCM_RIGHTS. That means you can have exactly >>> the scenario described above where a file assumed to be non-shared is >>> suddenly shared and you have racing reads/writes. >>> >>> For readdir we nowadays always take the file position mutex because of >>> the pidfd_getfd() business because that might corrupt internal state. >>> >>>> >>>> I'm thinking that we may have to provide two different `struct file` >>>> wrappers to accurately model this API in Rust. Perhaps they could be >>>> called File and LocalFile, where one is marked as thread safe and the >>>> other isn't. I can make all LocalFile methods available on File to avoid >>>> having to duplicate methods that are available on both. >>> >>> But isn't that just struct file and struct fd? Ideally we'd stay close >>> to something like this. >> >> Right, that kind of naming seems sensible. But I still need to >> understand the details a bit better. See below on fdget_pos. >> >>>> But it's not clear to me that this is even enough. Even if we give you a >>>> &LocalFile to prevent you from moving it across threads, you can just >>>> call File::fget to get an ARef<File> to the same file and then move >>>> *that* across threads. >>> >>> Yes, absolutely. >> >> One of my challenges is that Binder wants to call File::fget, >> immediately move it to another thread, and then call fd_install. And >> it would be pretty unfortunate if that requires unsafe. But like I argue >> below, it seems hard to design a safe API for this in the face of >> fdget_pos. >> >>>> This kind of global requirement is not so easy to model. Maybe klint [1] >>>> could do it ... atomic context violations are a similar kind of global >>>> check. But having klint do it would be far out. >>>> >>>> Or maybe File::fget should also return a LocalFile? >>>> >>>> But this raises a different question to me. Let's say process A uses >>>> Binder to send its own fd to process B, and the following things happen: >>>> >>>> 1. Process A enters the ioctl and takes fdget on the fd. >>>> 2. Process A calls fget on the same fd to send it to another process. >>>> 3. Process A goes to sleep, waiting for process B to respond. >>>> 4. Process B receives the message, installs the fd, and returns to userspace. >>>> 5. Process B responds to the transaction, but does not close the fd. >>> >>> The fd just installed in 4. and the fd you're referring to in 5. are >>> identical, right? IOW, we're not talking about two different fd (dup) >>> for the same file, right? >> >> I'm referring to whatever fd_install does given the `struct file` I got >> from fget in step 2. >> >>>> 6a. Process A finishes sleeping, and returns to userspace from the ioctl. >>>> 6b. Process B tries to do an operation (e.g. read) on the fd. >>>> >>>> Let's say that 6a and 6b run in parallel. >>>> >>>> Could this potentially result in a data race between step 6a and 6b? I'm >>>> guessing that step 6a probably doesn't touch any of the code that has >>>> elided locking assumptions, so in practice I guess there's not a problem >>>> ... but if you make any sort of elided locking assumption as you exit >>>> from the ioctl (before reaching the fdput), then it seems to me that you >>>> have a problem. >>> >>> Yes, 6a doesn't touch any code that has elided locking assumptions. >>> >>> 1'. Process A enters the ioctl and takes fdget() on the fd. Process A >>> holds the only reference to that file and the file descriptor table >>> isn't shared. Therefore, f_count is left untouched and remains at 1. >>> 2'. Process A calls fget() which unconditionally bumps f_count bringing >>> it to 2 and sending it another process (Presumably you intend to >>> imply that this reference is now owned by the second process.). >>> 3'. [as 3.] >>> 4'. Process B installs the file into it's file descriptor table >>> consuming that reference from 2'. The f_count remains at 2 with the >>> reference from 2' now being owned by Process B. >>> 5'. Since Process B isn't closing the fd and has just called >>> fd_install() it returns to userspace with f_count untouched and >>> still at 2. >>> 6'a. Process A finishes sleeping and returns to userspace calling >>> fdput(). Since the original fdget() was done without bumping the >>> reference count the fdput() of Process A will not decrement the >>> reference count. So f_count remains at 2. >>> 6'b. Process B performs a read/write syscall and calls fdget_pos(). >>> fdget_pos() sees that this file has f_count > 1 and takes the >>> file position mutex. >>> >>> So this isn't a problem. The problem is when a file becomes shared >>> implicitly without the original owner of the file knowing. >> >> Hmm. Yes, but the ioctl code that called fdget doesn't really know that >> the ioctl shared the file? So why is it okay? > > Why does it matter to the ioctl() code itself? The ioctl() code never > calls fdget_pos(). > >> >> It really seems like there are two different things going on here. When >> it comes to fdget, we only really care about operations that could >> remove it from the local file descriptor table, since fdget relies on >> the refcount in that table remaining valid until fdput. > > Yes. > >> >> On the other hand, for fdget_pos it also matters whether it gets >> installed in other file descriptor tables. Threads that reference it >> through a different fd table will still access the same position. > > Yes, they operate on the same f_pos. > >> >> And so this means that between fdget/fdput, there's never any problem >> with installing the `struct file` into another file descriptor table. >> Nothing you can do from that other fd table could cause the local fd >> table to drop its refcount on the file. Whereas such an install can be >> a problem between fdget_pos/fdput_pos, since that could introduce a race >> on the position. >> >> Is this correct? > > Yes, but that would imply you're sharing and installing a file into a > file descriptor table from a read/write/seek codepath. I don't see how > this can happen without something like e.g., pidfd_getfd(). And the > fd_install()ing task would then have to go back to userspace and issue a > concurrent read/write/seek system call while the other thread is still > reading/writing. > > Overall, we really only care about f_pos consistency because posix > requires atomicity between reads/writes/seeks. For pidfd_getfd() where > such sharing can happen non-cooperatively we just don't care as we've > just declared this to be an instance where we're outside of posix > guarantees. And for readdir() we unconditionally acquire the mutex. > > I think io_uring is racing on f_pos as well under certain circumstances > (REQ_F_CUR_POS?) as they don't use fdget_pos() at all. But iirc Jens > dislikes that they ever allowed that. > >> >> I was thinking that if we have some sort of File/LocalFile distinction >> (or File/Fd), then we may be able to get it to work by limiting what a >> File can do. For example, let's say that the only thing you can do with >> a File is install it into fd tables, then by the previous logic, there's >> no problem with it being safe to move across threads even if there's an >> active fdget. >> >> But the fdget_pos kind of throws a wrench into that, because now I can >> no longer say "it's always safe to do File::fget, move it to another >> thread, and install it into the remote fd table", since that could cause >> races on the position if there's an active fdget_pos when we call >> File::fget. > > I think I understand why that's a problem for you but let me try and > spell it out so I can understand where you're coming from: > > You want the Rust compiler to reject a file becoming shared implicitly > from any codepath that is beneath fdget_pos() even though no such code > yet exists (ignoring pidfd_getfd()). So it's a correctness issue to you. Hi Christian, I thought some more about this. Here's an idea of how we can encode this knowledge in the API. It seems like there are a few different things we might know about a pointer: 1. The pointer is fully shared, that is, all active fdget / fdget_pos references (if any) actually incremented the refcount / took the lock. 2. The pointer may have active fdget references that didn't increment the refcount, but all fdget_pos references (if any) incremented the refcount and took the lock. 3. The pointer may have active fdget or fdget_pos references that didn't increment the refcount / didn't take the lock. And orthogonal from that, all of the fdget / fdget_pos references are on a single thread, so we either know that we are on the same thread as them, or we don't know that. (Though we don't care about this distinction in the fully shared case.) I might encode this information with five different "states": * FullyShared - fully shared * MaybeFdget - might have fdget, not necessarily on this thread * MaybeLocalFdget - might have fdget, but if so, it's on this thread * MaybeFdgetPos - might have fdget_pos, not necessarily on this thread * MaybeLocalFdgetPos - might have fdget_pos, but if so, it's on this thread And you can make this a parameter on File, so you get pointer types like these: * ARef<File<MaybeFdget>> - A pointer to a file that owns a refcount and might have an active fdget, possibly on a different thread. * &File<FullyShared> - A reference to a file that doesn't own a refcount, and which is fully shared. * ARef<File<MaybeLocalFdgetPos>> - A pointer to a file that owns a refcount, and may have both active fdget and fdget_pos calls. If it does, they're on the same thread as this reference. This is the return type of fget. And if we introduce an Fdget smart pointer too, then it would probably be typed Fdget<MaybeLocalFdget> or Fdget<MaybeLocalFdgetPos>. You could do different things with these. For example, a MaybeFdget reference can be used with fd_install, but a MaybeFdgetPos reference can't. Similarly, you can close an fd using a FullyShared reference, but not using any of the weaker kinds. As for how you get each kind, well, the fget method would return an MaybeLocalFdgetPos reference, since we know that we're on the right thread if any fdget/fdget_pos references exist, but we don't otherwise know anything. And you could use something similar to the DeferredFdCloser to take a MaybeLocalFdgetPos or MaybeLocalFdget file reference any upgrade it to a FullyShared one by waiting until we return to userspace. (But you can't do this with the non-local variants.) When we write a Rust wrapper around file_operations that internally uses File::from_raw_file, then we can have it pass a &File<MaybeLocalFdget> to the ioctl handler, and a &File<MaybeLocalFdgetPos> to the read/write/seek handlers. That way, the ioctl handler can do things with the provided file that don't work when there's an fdget_pos, but the read/write/seek handlers can't. Or we can have unsafe methods that do upgrades. For example, the codepath in Binder that sends an fd from one process to another will most likely just call fget and then use unsafe to say "I know there are no active fdget_pos references since I'm in an ioctl" to upgrade the MaybeLocalFdgetPos it got from fget into an MaybeFdget (not FullyShared since the ioctl uses fdget), that it sends to the other process and then fd_installs. Does this sound sane to you? Alice
On Mon, Apr 8, 2024 at 9:45 AM Alice Ryhl <aliceryhl@google.com> wrote: > > Christian Brauner <brauner@kernel.org> wrote: > > On Tue, Apr 02, 2024 at 09:39:57AM +0000, Alice Ryhl wrote: > >> Christian Brauner <brauner@kernel.org> wrote: > >>> On Mon, Apr 01, 2024 at 12:09:08PM +0000, Alice Ryhl wrote: > >>>> Christian Brauner <brauner@kernel.org> wrote: > >>>>> On Wed, Mar 20, 2024 at 06:09:05PM +0000, Alice Ryhl wrote: > >>>>>> Christian Brauner <brauner@kernel.org> wrote: > >>>>>>> On Fri, Feb 09, 2024 at 11:18:16AM +0000, Alice Ryhl wrote: > >>>>>>>> +/// Wraps the kernel's `struct file`. > >>>>>>>> +/// > >>>>>>>> +/// This represents an open file rather than a file on a filesystem. Processes generally reference > >>>>>>>> +/// open files using file descriptors. However, file descriptors are not the same as files. A file > >>>>>>>> +/// descriptor is just an integer that corresponds to a file, and a single file may be referenced > >>>>>>>> +/// by multiple file descriptors. > >>>>>>>> +/// > >>>>>>>> +/// # Refcounting > >>>>>>>> +/// > >>>>>>>> +/// Instances of this type are reference-counted. The reference count is incremented by the > >>>>>>>> +/// `fget`/`get_file` functions and decremented by `fput`. The Rust type `ARef<File>` represents a > >>>>>>>> +/// pointer that owns a reference count on the file. > >>>>>>>> +/// > >>>>>>>> +/// Whenever a process opens a file descriptor (fd), it stores a pointer to the file in its `struct > >>>>>>>> +/// files_struct`. This pointer owns a reference count to the file, ensuring the file isn't > >>>>>>>> +/// prematurely deleted while the file descriptor is open. In Rust terminology, the pointers in > >>>>>>>> +/// `struct files_struct` are `ARef<File>` pointers. > >>>>>>>> +/// > >>>>>>>> +/// ## Light refcounts > >>>>>>>> +/// > >>>>>>>> +/// Whenever a process has an fd to a file, it may use something called a "light refcount" as a > >>>>>>>> +/// performance optimization. Light refcounts are acquired by calling `fdget` and released with > >>>>>>>> +/// `fdput`. The idea behind light refcounts is that if the fd is not closed between the calls to > >>>>>>>> +/// `fdget` and `fdput`, then the refcount cannot hit zero during that time, as the `struct > >>>>>>>> +/// files_struct` holds a reference until the fd is closed. This means that it's safe to access the > >>>>>>>> +/// file even if `fdget` does not increment the refcount. > >>>>>>>> +/// > >>>>>>>> +/// The requirement that the fd is not closed during a light refcount applies globally across all > >>>>>>>> +/// threads - not just on the thread using the light refcount. For this reason, light refcounts are > >>>>>>>> +/// only used when the `struct files_struct` is not shared with other threads, since this ensures > >>>>>>>> +/// that other unrelated threads cannot suddenly start using the fd and close it. Therefore, > >>>>>>>> +/// calling `fdget` on a shared `struct files_struct` creates a normal refcount instead of a light > >>>>>>>> +/// refcount. > >>>>>>> > >>>>>>> When the fdget() calling task doesn't have a shared file descriptor > >>>>>>> table fdget() will not increment the reference count, yes. This > >>>>>>> also implies that you cannot have task A use fdget() and then pass > >>>>>>> f.file to task B that holds on to it while A returns to userspace. It's > >>>>>>> irrelevant that task A won't drop the reference count or that task B > >>>>>>> won't drop the reference count. Because task A could return back to > >>>>>>> userspace and immediately close the fd via a regular close() system call > >>>>>>> at which point task B has a UAF. In other words a file that has been > >>>>>>> gotten via fdget() can't be Send to another task without the Send > >>>>>>> implying taking a reference to it. > >>>>>> > >>>>>> That matches my understanding. > >>>>>> > >>>>>> I suppose that technically you can still send it to another thread *if* you > >>>>>> ensure that the current thread waits until that other thread stops using the > >>>>>> file before returning to userspace. > >>>>> > >>>>> _Technically_ yes, but it would be brittle as hell. The problem is that > >>>>> fdget() _relies_ on being single-threaded for the time that fd is used > >>>>> until fdput(). There's locking assumptions that build on that e.g., for > >>>>> concurrent read/write. So no, that shouldn't be allowed. > >>>>> > >>>>> Look at how this broke our back when we introduced pidfd_getfd() where > >>>>> we steal an fd from another task. I have a lengthy explanation how that > >>>>> can be used to violate our elided-locking which is based on assuming > >>>>> that we're always single-threaded and the file can't be suddenly shared > >>>>> with another task. So maybe doable but it would make the semantics even > >>>>> more intricate. > >>>> > >>>> Hmm, the part about elided locking is surprising to me, and may be an > >>>> issue. Can you give more details on that? Because the current > >>> > >>> So what I referred to was that we do have fdget_pos(). Roughly, if > >>> there's more than one reference on the file then we need to acquire a > >>> mutex but if it's only a single reference then we can avoid taking the > >>> mutex because we know that we're the only one that has a reference to > >>> that file and no one else can acquire one. Whether or not that mutex was > >>> taken is taken track of in struct fd. > >>> > >>> So you can't share a file after fdget_pos() has been called on it and > >>> you haven't taken the position mutex. So let's say you had: > >>> > >>> * Tread A that calls fdget_pos() on file1 and the reference count is > >>> one. So Thread A doesn't acquire the file position mutex for file1. > >>> * Now somehow that file1 becomes shared, e.g., Thread B calls fget() on > >>> it and now Thread B does some operation that requires the file > >>> position mutex. > >>> => Thread A and Thread B race on the file position. > >>> > >>> So just because you have a reference to a file from somewhere it doesn't > >>> mean you can just share it with another thread. > >>> > >>> So if yo have an arbitrary reference to a file in Rust and that somehow > >>> can be shared with another thread you risk races here. > >>> > >>>> abstractions here *do* actually allow what I described, since we > >>>> implement Sync for File. > >>>> > >>>> I'm not familiar with the pidfd_getfd discussion you are referring to. > >>>> Do you have a link? > >>> > >>> https://lore.kernel.org/linux-fsdevel/20230724-vfs-fdget_pos-v1-1-a4abfd7103f3@kernel.org > >>> > >>> pidfd_getfd() can be used to steal a file descriptor from another task. > >>> It's like a non-cooperative SCM_RIGHTS. That means you can have exactly > >>> the scenario described above where a file assumed to be non-shared is > >>> suddenly shared and you have racing reads/writes. > >>> > >>> For readdir we nowadays always take the file position mutex because of > >>> the pidfd_getfd() business because that might corrupt internal state. > >>> > >>>> > >>>> I'm thinking that we may have to provide two different `struct file` > >>>> wrappers to accurately model this API in Rust. Perhaps they could be > >>>> called File and LocalFile, where one is marked as thread safe and the > >>>> other isn't. I can make all LocalFile methods available on File to avoid > >>>> having to duplicate methods that are available on both. > >>> > >>> But isn't that just struct file and struct fd? Ideally we'd stay close > >>> to something like this. > >> > >> Right, that kind of naming seems sensible. But I still need to > >> understand the details a bit better. See below on fdget_pos. > >> > >>>> But it's not clear to me that this is even enough. Even if we give you a > >>>> &LocalFile to prevent you from moving it across threads, you can just > >>>> call File::fget to get an ARef<File> to the same file and then move > >>>> *that* across threads. > >>> > >>> Yes, absolutely. > >> > >> One of my challenges is that Binder wants to call File::fget, > >> immediately move it to another thread, and then call fd_install. And > >> it would be pretty unfortunate if that requires unsafe. But like I argue > >> below, it seems hard to design a safe API for this in the face of > >> fdget_pos. > >> > >>>> This kind of global requirement is not so easy to model. Maybe klint [1] > >>>> could do it ... atomic context violations are a similar kind of global > >>>> check. But having klint do it would be far out. > >>>> > >>>> Or maybe File::fget should also return a LocalFile? > >>>> > >>>> But this raises a different question to me. Let's say process A uses > >>>> Binder to send its own fd to process B, and the following things happen: > >>>> > >>>> 1. Process A enters the ioctl and takes fdget on the fd. > >>>> 2. Process A calls fget on the same fd to send it to another process. > >>>> 3. Process A goes to sleep, waiting for process B to respond. > >>>> 4. Process B receives the message, installs the fd, and returns to userspace. > >>>> 5. Process B responds to the transaction, but does not close the fd. > >>> > >>> The fd just installed in 4. and the fd you're referring to in 5. are > >>> identical, right? IOW, we're not talking about two different fd (dup) > >>> for the same file, right? > >> > >> I'm referring to whatever fd_install does given the `struct file` I got > >> from fget in step 2. > >> > >>>> 6a. Process A finishes sleeping, and returns to userspace from the ioctl. > >>>> 6b. Process B tries to do an operation (e.g. read) on the fd. > >>>> > >>>> Let's say that 6a and 6b run in parallel. > >>>> > >>>> Could this potentially result in a data race between step 6a and 6b? I'm > >>>> guessing that step 6a probably doesn't touch any of the code that has > >>>> elided locking assumptions, so in practice I guess there's not a problem > >>>> ... but if you make any sort of elided locking assumption as you exit > >>>> from the ioctl (before reaching the fdput), then it seems to me that you > >>>> have a problem. > >>> > >>> Yes, 6a doesn't touch any code that has elided locking assumptions. > >>> > >>> 1'. Process A enters the ioctl and takes fdget() on the fd. Process A > >>> holds the only reference to that file and the file descriptor table > >>> isn't shared. Therefore, f_count is left untouched and remains at 1. > >>> 2'. Process A calls fget() which unconditionally bumps f_count bringing > >>> it to 2 and sending it another process (Presumably you intend to > >>> imply that this reference is now owned by the second process.). > >>> 3'. [as 3.] > >>> 4'. Process B installs the file into it's file descriptor table > >>> consuming that reference from 2'. The f_count remains at 2 with the > >>> reference from 2' now being owned by Process B. > >>> 5'. Since Process B isn't closing the fd and has just called > >>> fd_install() it returns to userspace with f_count untouched and > >>> still at 2. > >>> 6'a. Process A finishes sleeping and returns to userspace calling > >>> fdput(). Since the original fdget() was done without bumping the > >>> reference count the fdput() of Process A will not decrement the > >>> reference count. So f_count remains at 2. > >>> 6'b. Process B performs a read/write syscall and calls fdget_pos(). > >>> fdget_pos() sees that this file has f_count > 1 and takes the > >>> file position mutex. > >>> > >>> So this isn't a problem. The problem is when a file becomes shared > >>> implicitly without the original owner of the file knowing. > >> > >> Hmm. Yes, but the ioctl code that called fdget doesn't really know that > >> the ioctl shared the file? So why is it okay? > > > > Why does it matter to the ioctl() code itself? The ioctl() code never > > calls fdget_pos(). > > > >> > >> It really seems like there are two different things going on here. When > >> it comes to fdget, we only really care about operations that could > >> remove it from the local file descriptor table, since fdget relies on > >> the refcount in that table remaining valid until fdput. > > > > Yes. > > > >> > >> On the other hand, for fdget_pos it also matters whether it gets > >> installed in other file descriptor tables. Threads that reference it > >> through a different fd table will still access the same position. > > > > Yes, they operate on the same f_pos. > > > >> > >> And so this means that between fdget/fdput, there's never any problem > >> with installing the `struct file` into another file descriptor table. > >> Nothing you can do from that other fd table could cause the local fd > >> table to drop its refcount on the file. Whereas such an install can be > >> a problem between fdget_pos/fdput_pos, since that could introduce a race > >> on the position. > >> > >> Is this correct? > > > > Yes, but that would imply you're sharing and installing a file into a > > file descriptor table from a read/write/seek codepath. I don't see how > > this can happen without something like e.g., pidfd_getfd(). And the > > fd_install()ing task would then have to go back to userspace and issue a > > concurrent read/write/seek system call while the other thread is still > > reading/writing. > > > > Overall, we really only care about f_pos consistency because posix > > requires atomicity between reads/writes/seeks. For pidfd_getfd() where > > such sharing can happen non-cooperatively we just don't care as we've > > just declared this to be an instance where we're outside of posix > > guarantees. And for readdir() we unconditionally acquire the mutex. > > > > I think io_uring is racing on f_pos as well under certain circumstances > > (REQ_F_CUR_POS?) as they don't use fdget_pos() at all. But iirc Jens > > dislikes that they ever allowed that. > > > >> > >> I was thinking that if we have some sort of File/LocalFile distinction > >> (or File/Fd), then we may be able to get it to work by limiting what a > >> File can do. For example, let's say that the only thing you can do with > >> a File is install it into fd tables, then by the previous logic, there's > >> no problem with it being safe to move across threads even if there's an > >> active fdget. > >> > >> But the fdget_pos kind of throws a wrench into that, because now I can > >> no longer say "it's always safe to do File::fget, move it to another > >> thread, and install it into the remote fd table", since that could cause > >> races on the position if there's an active fdget_pos when we call > >> File::fget. > > > > I think I understand why that's a problem for you but let me try and > > spell it out so I can understand where you're coming from: > > > > You want the Rust compiler to reject a file becoming shared implicitly > > from any codepath that is beneath fdget_pos() even though no such code > > yet exists (ignoring pidfd_getfd()). So it's a correctness issue to you. > > Hi Christian, > > I thought some more about this. Here's an idea of how we can encode this > knowledge in the API. It seems like there are a few different things we > might know about a pointer: > > 1. The pointer is fully shared, that is, all active fdget / fdget_pos > references (if any) actually incremented the refcount / took the > lock. > > 2. The pointer may have active fdget references that didn't increment > the refcount, but all fdget_pos references (if any) incremented the > refcount and took the lock. > > 3. The pointer may have active fdget or fdget_pos references that didn't > increment the refcount / didn't take the lock. > > And orthogonal from that, all of the fdget / fdget_pos references are on > a single thread, so we either know that we are on the same thread as > them, or we don't know that. (Though we don't care about this > distinction in the fully shared case.) > > I might encode this information with five different "states": > > * FullyShared - fully shared > * MaybeFdget - might have fdget, not necessarily on this thread > * MaybeLocalFdget - might have fdget, but if so, it's on this thread > * MaybeFdgetPos - might have fdget_pos, not necessarily on this thread > * MaybeLocalFdgetPos - might have fdget_pos, but if so, it's on this thread > > And you can make this a parameter on File, so you get pointer types like > these: > > * ARef<File<MaybeFdget>> - A pointer to a file that owns a refcount and > might have an active fdget, possibly on a different thread. > > * &File<FullyShared> - A reference to a file that doesn't own a > refcount, and which is fully shared. > > * ARef<File<MaybeLocalFdgetPos>> - A pointer to a file that owns a > refcount, and may have both active fdget and fdget_pos calls. If it > does, they're on the same thread as this reference. This is the return > type of fget. > > And if we introduce an Fdget smart pointer too, then it would probably > be typed Fdget<MaybeLocalFdget> or Fdget<MaybeLocalFdgetPos>. > > You could do different things with these. For example, a MaybeFdget > reference can be used with fd_install, but a MaybeFdgetPos reference > can't. Similarly, you can close an fd using a FullyShared reference, but > not using any of the weaker kinds. > > As for how you get each kind, well, the fget method would return an > MaybeLocalFdgetPos reference, since we know that we're on the right > thread if any fdget/fdget_pos references exist, but we don't otherwise > know anything. > > And you could use something similar to the DeferredFdCloser to take a > MaybeLocalFdgetPos or MaybeLocalFdget file reference any upgrade it to a > FullyShared one by waiting until we return to userspace. (But you can't > do this with the non-local variants.) > > When we write a Rust wrapper around file_operations that internally uses > File::from_raw_file, then we can have it pass a &File<MaybeLocalFdget> > to the ioctl handler, and a &File<MaybeLocalFdgetPos> to the > read/write/seek handlers. That way, the ioctl handler can do things with > the provided file that don't work when there's an fdget_pos, but the > read/write/seek handlers can't. > > Or we can have unsafe methods that do upgrades. For example, the > codepath in Binder that sends an fd from one process to another will > most likely just call fget and then use unsafe to say "I know there are > no active fdget_pos references since I'm in an ioctl" to upgrade the > MaybeLocalFdgetPos it got from fget into an MaybeFdget (not FullyShared > since the ioctl uses fdget), that it sends to the other process and then > fd_installs. > > Does this sound sane to you? Forget what I said above, it's wrong. I forgot that the fdget regions are tied to the fdtable's refcount, not the file's refcount. Doing an fget and waiting until the current thread returns to userspace is not enough to guarantee that no non-shared fdget regions exist. But the logic still applies to fdget_pos regions. Alice
diff --git a/fs/file.c b/fs/file.c index 3b683b9101d8..f2eab5fcb87f 100644 --- a/fs/file.c +++ b/fs/file.c @@ -1127,6 +1127,13 @@ EXPORT_SYMBOL(task_lookup_next_fdget_rcu); * * The fput_needed flag returned by fget_light should be passed to the * corresponding fput_light. + * + * (As an exception to rule 2, you can call filp_close between fget_light and + * fput_light provided that you capture a real refcount with get_file before + * the call to filp_close, and ensure that this real refcount is fput *after* + * the fput_light call.) + * + * See also the documentation in rust/kernel/file.rs. */ static unsigned long __fget_light(unsigned int fd, fmode_t mask) { diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h index 936651110c39..41fcd2905ed4 100644 --- a/rust/bindings/bindings_helper.h +++ b/rust/bindings/bindings_helper.h @@ -9,6 +9,8 @@ #include <kunit/test.h> #include <linux/errname.h> #include <linux/ethtool.h> +#include <linux/file.h> +#include <linux/fs.h> #include <linux/jiffies.h> #include <linux/mdio.h> #include <linux/phy.h> diff --git a/rust/helpers.c b/rust/helpers.c index 70e59efd92bc..03141a3608a4 100644 --- a/rust/helpers.c +++ b/rust/helpers.c @@ -25,6 +25,7 @@ #include <linux/build_bug.h> #include <linux/err.h> #include <linux/errname.h> +#include <linux/fs.h> #include <linux/mutex.h> #include <linux/refcount.h> #include <linux/sched/signal.h> @@ -157,6 +158,12 @@ void rust_helper_init_work_with_key(struct work_struct *work, work_func_t func, } EXPORT_SYMBOL_GPL(rust_helper_init_work_with_key); +struct file *rust_helper_get_file(struct file *f) +{ + return get_file(f); +} +EXPORT_SYMBOL_GPL(rust_helper_get_file); + /* * `bindgen` binds the C `size_t` type as the Rust `usize` type, so we can * use it in contexts where Rust expects a `usize` like slice (array) indices. diff --git a/rust/kernel/file.rs b/rust/kernel/file.rs new file mode 100644 index 000000000000..cf8ebf619379 --- /dev/null +++ b/rust/kernel/file.rs @@ -0,0 +1,254 @@ +// SPDX-License-Identifier: GPL-2.0 + +//! Files and file descriptors. +//! +//! C headers: [`include/linux/fs.h`](srctree/include/linux/fs.h) and +//! [`include/linux/file.h`](srctree/include/linux/file.h) + +use crate::{ + bindings, + error::{code::*, Error, Result}, + types::{ARef, AlwaysRefCounted, Opaque}, +}; +use core::ptr; + +/// Flags associated with a [`File`]. +pub mod flags { + /// File is opened in append mode. + pub const O_APPEND: u32 = bindings::O_APPEND; + + /// Signal-driven I/O is enabled. + pub const O_ASYNC: u32 = bindings::FASYNC; + + /// Close-on-exec flag is set. + pub const O_CLOEXEC: u32 = bindings::O_CLOEXEC; + + /// File was created if it didn't already exist. + pub const O_CREAT: u32 = bindings::O_CREAT; + + /// Direct I/O is enabled for this file. + pub const O_DIRECT: u32 = bindings::O_DIRECT; + + /// File must be a directory. + pub const O_DIRECTORY: u32 = bindings::O_DIRECTORY; + + /// Like [`O_SYNC`] except metadata is not synced. + pub const O_DSYNC: u32 = bindings::O_DSYNC; + + /// Ensure that this file is created with the `open(2)` call. + pub const O_EXCL: u32 = bindings::O_EXCL; + + /// Large file size enabled (`off64_t` over `off_t`). + pub const O_LARGEFILE: u32 = bindings::O_LARGEFILE; + + /// Do not update the file last access time. + pub const O_NOATIME: u32 = bindings::O_NOATIME; + + /// File should not be used as process's controlling terminal. + pub const O_NOCTTY: u32 = bindings::O_NOCTTY; + + /// If basename of path is a symbolic link, fail open. + pub const O_NOFOLLOW: u32 = bindings::O_NOFOLLOW; + + /// File is using nonblocking I/O. + pub const O_NONBLOCK: u32 = bindings::O_NONBLOCK; + + /// File is using nonblocking I/O. + /// + /// This is effectively the same flag as [`O_NONBLOCK`] on all architectures + /// except SPARC64. + pub const O_NDELAY: u32 = bindings::O_NDELAY; + + /// Used to obtain a path file descriptor. + pub const O_PATH: u32 = bindings::O_PATH; + + /// Write operations on this file will flush data and metadata. + pub const O_SYNC: u32 = bindings::O_SYNC; + + /// This file is an unnamed temporary regular file. + pub const O_TMPFILE: u32 = bindings::O_TMPFILE; + + /// File should be truncated to length 0. + pub const O_TRUNC: u32 = bindings::O_TRUNC; + + /// Bitmask for access mode flags. + /// + /// # Examples + /// + /// ``` + /// use kernel::file; + /// # fn do_something() {} + /// # let flags = 0; + /// if (flags & file::flags::O_ACCMODE) == file::flags::O_RDONLY { + /// do_something(); + /// } + /// ``` + pub const O_ACCMODE: u32 = bindings::O_ACCMODE; + + /// File is read only. + pub const O_RDONLY: u32 = bindings::O_RDONLY; + + /// File is write only. + pub const O_WRONLY: u32 = bindings::O_WRONLY; + + /// File can be both read and written. + pub const O_RDWR: u32 = bindings::O_RDWR; +} + +/// Wraps the kernel's `struct file`. +/// +/// This represents an open file rather than a file on a filesystem. Processes generally reference +/// open files using file descriptors. However, file descriptors are not the same as files. A file +/// descriptor is just an integer that corresponds to a file, and a single file may be referenced +/// by multiple file descriptors. +/// +/// # Refcounting +/// +/// Instances of this type are reference-counted. The reference count is incremented by the +/// `fget`/`get_file` functions and decremented by `fput`. The Rust type `ARef<File>` represents a +/// pointer that owns a reference count on the file. +/// +/// Whenever a process opens a file descriptor (fd), it stores a pointer to the file in its `struct +/// files_struct`. This pointer owns a reference count to the file, ensuring the file isn't +/// prematurely deleted while the file descriptor is open. In Rust terminology, the pointers in +/// `struct files_struct` are `ARef<File>` pointers. +/// +/// ## Light refcounts +/// +/// Whenever a process has an fd to a file, it may use something called a "light refcount" as a +/// performance optimization. Light refcounts are acquired by calling `fdget` and released with +/// `fdput`. The idea behind light refcounts is that if the fd is not closed between the calls to +/// `fdget` and `fdput`, then the refcount cannot hit zero during that time, as the `struct +/// files_struct` holds a reference until the fd is closed. This means that it's safe to access the +/// file even if `fdget` does not increment the refcount. +/// +/// The requirement that the fd is not closed during a light refcount applies globally across all +/// threads - not just on the thread using the light refcount. For this reason, light refcounts are +/// only used when the `struct files_struct` is not shared with other threads, since this ensures +/// that other unrelated threads cannot suddenly start using the fd and close it. Therefore, +/// calling `fdget` on a shared `struct files_struct` creates a normal refcount instead of a light +/// refcount. +/// +/// Light reference counts must be released with `fdput` before the system call returns to +/// userspace. This means that if you wait until the current system call returns to userspace, then +/// all light refcounts that existed at the time have gone away. +/// +/// ## Rust references +/// +/// The reference type `&File` is similar to light refcounts: +/// +/// * `&File` references don't own a reference count. They can only exist as long as the reference +/// count stays positive, and can only be created when there is some mechanism in place to ensure +/// this. +/// +/// * The Rust borrow-checker normally ensures this by enforcing that the `ARef<File>` from which +/// a `&File` is created outlives the `&File`. +/// +/// * Using the unsafe [`File::from_ptr`] means that it is up to the caller to ensure that the +/// `&File` only exists while the reference count is positive. +/// +/// * You can think of `fdget` as using an fd to look up an `ARef<File>` in the `struct +/// files_struct` and create an `&File` from it. The "fd cannot be closed" rule is like the Rust +/// rule "the `ARef<File>` must outlive the `&File`". +/// +/// # Invariants +/// +/// * Instances of this type are refcounted using the `f_count` field. +/// * If an fd with active light refcounts is closed, then it must be the case that the file +/// refcount is positive until all light refcounts of the fd have been dropped. +/// * A light refcount must be dropped before returning to userspace. +#[repr(transparent)] +pub struct File(Opaque<bindings::file>); + +// SAFETY: +// - `File::dec_ref` can be called from any thread. +// - It is okay to send ownership of `struct file` across thread boundaries. +unsafe impl Send for File {} + +// SAFETY: All methods defined on `File` that take `&self` are safe to call even if other threads +// are concurrently accessing the same `struct file`, because those methods either access immutable +// properties or have proper synchronization to ensure that such accesses are safe. +unsafe impl Sync for File {} + +impl File { + /// Constructs a new `struct file` wrapper from a file descriptor. + /// + /// The file descriptor belongs to the current process. + pub fn fget(fd: u32) -> Result<ARef<Self>, BadFdError> { + // SAFETY: FFI call, there are no requirements on `fd`. + let ptr = ptr::NonNull::new(unsafe { bindings::fget(fd) }).ok_or(BadFdError)?; + + // SAFETY: `bindings::fget` either returns null or a valid pointer to a file, and we + // checked for null above. + // + // INVARIANT: `bindings::fget` creates a refcount, and we pass ownership of the refcount to + // the new `ARef<File>`. + Ok(unsafe { ARef::from_raw(ptr.cast()) }) + } + + /// Creates a reference to a [`File`] from a valid pointer. + /// + /// # Safety + /// + /// The caller must ensure that `ptr` points at a valid file and that the file's refcount is + /// positive for the duration of 'a. + pub unsafe fn from_ptr<'a>(ptr: *const bindings::file) -> &'a File { + // SAFETY: The caller guarantees that the pointer is not dangling and stays valid for the + // duration of 'a. The cast is okay because `File` is `repr(transparent)`. + // + // INVARIANT: The safety requirements guarantee that the refcount does not hit zero during + // 'a. + unsafe { &*ptr.cast() } + } + + /// Returns a raw pointer to the inner C struct. + #[inline] + pub fn as_ptr(&self) -> *mut bindings::file { + self.0.get() + } + + /// Returns the flags associated with the file. + /// + /// The flags are a combination of the constants in [`flags`]. + pub fn flags(&self) -> u32 { + // This `read_volatile` is intended to correspond to a READ_ONCE call. + // + // SAFETY: The file is valid because the shared reference guarantees a nonzero refcount. + // + // FIXME(read_once): Replace with `read_once` when available on the Rust side. + unsafe { core::ptr::addr_of!((*self.as_ptr()).f_flags).read_volatile() } + } +} + +// SAFETY: The type invariants guarantee that `File` is always ref-counted. This implementation +// makes `ARef<File>` own a normal refcount. +unsafe impl AlwaysRefCounted for File { + fn inc_ref(&self) { + // SAFETY: The existence of a shared reference means that the refcount is nonzero. + unsafe { bindings::get_file(self.as_ptr()) }; + } + + unsafe fn dec_ref(obj: ptr::NonNull<File>) { + // SAFETY: To call this method, the caller passes us ownership of a normal refcount, so we + // may drop it. The cast is okay since `File` has the same representation as `struct file`. + unsafe { bindings::fput(obj.cast().as_ptr()) } + } +} + +/// Represents the `EBADF` error code. +/// +/// Used for methods that can only fail with `EBADF`. +#[derive(Copy, Clone, Eq, PartialEq)] +pub struct BadFdError; + +impl From<BadFdError> for Error { + fn from(_: BadFdError) -> Error { + EBADF + } +} + +impl core::fmt::Debug for BadFdError { + fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result { + f.pad("EBADF") + } +} diff --git a/rust/kernel/lib.rs b/rust/kernel/lib.rs index b89ecf4e97a0..9353dd713a20 100644 --- a/rust/kernel/lib.rs +++ b/rust/kernel/lib.rs @@ -34,6 +34,7 @@ mod allocator; mod build_assert; pub mod error; +pub mod file; pub mod init; pub mod ioctl; #[cfg(CONFIG_KUNIT)]