This updates the vm_area_struct support to use the approach we discussed at LPC where there are several different Rust wrappers for vm_area_struct depending on the kind of access you have to the vma. Each case allows a different set of operations on the vma. This series depends on vfs.rust.file and char-misc-next which are scheduled to land upstream in the current merge window. --- Changes in v9: - Be more explicit about VmAreaNew being used with f_ops->mmap(). - Point out that clearing VM_MAYWRITE is irreversible. - Use __vm_flags to set the flags. - Use as_ and into_ prefixes for conversions. - Update lock_vma_under_rcu docs and commit msg - Mention that VmAreaRef::end is exclusive. - Reword docs for zap_page_range_single. - Minor fixes to flag docs. - Add way to access current->mm without a refcount increment. - Link to v8: https://lore.kernel.org/r/20241120-vma-v8-0-eb31425da66b@google.com Changes in v8: - Split series into more commits to ease review. - Improve read locks based on Lorenzo's doc: either the mmap or vma lock can be used. - Get rid of mmap write lock because it's possible to avoid the need for it. - Do not allow invalid flag combinations on VmAreaNew. - Link to v7: https://lore.kernel.org/r/20241014-vma-v7-0-01e32f861195@google.com Changes in v7: - Make the mmap read/write lock guards respect strict owner semantics. - Link to v6: https://lore.kernel.org/r/20241010-vma-v6-0-d89039b6f573@google.com Changes in v6: - Introduce VmArea{Ref,Mut,New} distinction. - Add a second patchset for miscdevice. - Rebase on char-misc-next (currently on v6.12-rc2). - Link to v5: https://lore.kernel.org/r/20240806-vma-v5-1-04018f05de2b@google.com Changes in v5: - Rename VmArea::from_raw_vma to from_raw. - Use Pin for mutable VmArea references. - Go through `ARef::from` in `mmgrab_current`. - Link to v4: https://lore.kernel.org/r/20240802-vma-v4-1-091a87058a43@google.com Changes in v4: - Pull out ARef::into_raw into a separate patch. - Update invariants and struct documentation. - Rename from_raw_mm to from_raw. - Link to v3: https://lore.kernel.org/r/20240801-vma-v3-1-db6c1c0afda9@google.com Changes in v3: - Reorder entries in mm.rs. - Use ARef for mmput_async helper. - Clarify that VmArea requires you to hold the mmap read or write lock. - Link to v2: https://lore.kernel.org/r/20240727-vma-v2-1-ab3e5927dc3a@google.com Changes in v2: - mm.rs is redesigned from scratch making use of AsRef - Add notes about whether destructors may sleep - Rename Area to VmArea - Link to v1: https://lore.kernel.org/r/20240723-vma-v1-1-32ad5a0118ee@google.com --- Alice Ryhl (8): mm: rust: add abstraction for struct mm_struct mm: rust: add vm_area_struct methods that require read access mm: rust: add vm_insert_page mm: rust: add lock_vma_under_rcu mm: rust: add mmput_async support mm: rust: add VmAreaNew for f_ops->mmap() rust: miscdevice: add mmap support task: rust: rework how current is accessed rust/helpers/helpers.c | 1 + rust/helpers/mm.c | 50 ++++++ rust/kernel/lib.rs | 1 + rust/kernel/miscdevice.rs | 28 +++ rust/kernel/mm.rs | 323 +++++++++++++++++++++++++++++++++++ rust/kernel/mm/virt.rs | 421 ++++++++++++++++++++++++++++++++++++++++++++++ rust/kernel/task.rs | 64 +++++-- 7 files changed, 875 insertions(+), 13 deletions(-) --- base-commit: 8b8379ae7c70734686e98cf7455e6b25e614d593 change-id: 20240723-vma-f80119f9fb35 Best regards, -- Alice Ryhl <aliceryhl@google.com> These abstractions allow you to reference a `struct mm_struct` using both mmgrab and mmget refcounts. This is done using two Rust types: * Mm - represents an mm_struct where you don't know anything about the value of mm_users. * MmWithUser - represents an mm_struct where you know at compile time that mm_users is non-zero. This allows us to encode in the type system whether a method requires that mm_users is non-zero or not. For instance, you can always call `mmget_not_zero` but you can only call `mmap_read_lock` when mm_users is non-zero. It's possible to access current->mm without a refcount increment, but that is added in a later patch of this series. Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/helpers/helpers.c | 1 + rust/helpers/mm.c | 39 +++++++++ rust/kernel/lib.rs | 1 + rust/kernel/mm.rs | 219 +++++++++++++++++++++++++++++++++++++++++++++++++ 4 files changed, 260 insertions(+) diff --git a/rust/helpers/helpers.c b/rust/helpers/helpers.c index XXXXXXX..XXXXXXX 100644 --- a/rust/helpers/helpers.c +++ b/rust/helpers/helpers.c @@ -XXX,XX +XXX,XX @@ #include "err.c" #include "fs.c" #include "kunit.c" +#include "mm.c" #include "mutex.c" #include "page.c" #include "rbtree.c" diff --git a/rust/helpers/mm.c b/rust/helpers/mm.c new file mode 100644 index XXXXXXX..XXXXXXX --- /dev/null +++ b/rust/helpers/mm.c @@ -XXX,XX +XXX,XX @@ +// SPDX-License-Identifier: GPL-2.0 + +#include <linux/mm.h> +#include <linux/sched/mm.h> + +void rust_helper_mmgrab(struct mm_struct *mm) +{ + mmgrab(mm); +} + +void rust_helper_mmdrop(struct mm_struct *mm) +{ + mmdrop(mm); +} + +void rust_helper_mmget(struct mm_struct *mm) +{ + mmget(mm); +} + +bool rust_helper_mmget_not_zero(struct mm_struct *mm) +{ + return mmget_not_zero(mm); +} + +void rust_helper_mmap_read_lock(struct mm_struct *mm) +{ + mmap_read_lock(mm); +} + +bool rust_helper_mmap_read_trylock(struct mm_struct *mm) +{ + return mmap_read_trylock(mm); +} + +void rust_helper_mmap_read_unlock(struct mm_struct *mm) +{ + mmap_read_unlock(mm); +} diff --git a/rust/kernel/lib.rs b/rust/kernel/lib.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/lib.rs +++ b/rust/kernel/lib.rs @@ -XXX,XX +XXX,XX @@ pub mod kunit; pub mod list; pub mod miscdevice; +pub mod mm; #[cfg(CONFIG_NET)] pub mod net; pub mod page; diff --git a/rust/kernel/mm.rs b/rust/kernel/mm.rs new file mode 100644 index XXXXXXX..XXXXXXX --- /dev/null +++ b/rust/kernel/mm.rs @@ -XXX,XX +XXX,XX @@ +// SPDX-License-Identifier: GPL-2.0 + +// Copyright (C) 2024 Google LLC. + +//! Memory management. +//! +//! C header: [`include/linux/mm.h`](srctree/include/linux/mm.h) + +use crate::{ + bindings, + types::{ARef, AlwaysRefCounted, NotThreadSafe, Opaque}, +}; +use core::{ops::Deref, ptr::NonNull}; + +/// A wrapper for the kernel's `struct mm_struct`. +/// +/// Since `mm_users` may be zero, the associated address space may not exist anymore. You can use +/// [`mmget_not_zero`] to be able to access the address space. +/// +/// The `ARef<Mm>` smart pointer holds an `mmgrab` refcount. Its destructor may sleep. +/// +/// # Invariants +/// +/// Values of this type are always refcounted using `mmgrab`. +/// +/// [`mmget_not_zero`]: Mm::mmget_not_zero +#[repr(transparent)] +pub struct Mm { + mm: Opaque<bindings::mm_struct>, +} + +// SAFETY: It is safe to call `mmdrop` on another thread than where `mmgrab` was called. +unsafe impl Send for Mm {} +// SAFETY: All methods on `Mm` can be called in parallel from several threads. +unsafe impl Sync for Mm {} + +// SAFETY: By the type invariants, this type is always refcounted. +unsafe impl AlwaysRefCounted for Mm { + #[inline] + fn inc_ref(&self) { + // SAFETY: The pointer is valid since self is a reference. + unsafe { bindings::mmgrab(self.as_raw()) }; + } + + #[inline] + unsafe fn dec_ref(obj: NonNull<Self>) { + // SAFETY: The caller is giving up their refcount. + unsafe { bindings::mmdrop(obj.cast().as_ptr()) }; + } +} + +/// A wrapper for the kernel's `struct mm_struct`. +/// +/// This type is like [`Mm`], but with non-zero `mm_users`. It can only be used when `mm_users` can +/// be proven to be non-zero at compile-time, usually because the relevant code holds an `mmget` +/// refcount. It can be used to access the associated address space. +/// +/// The `ARef<MmWithUser>` smart pointer holds an `mmget` refcount. Its destructor may sleep. +/// +/// # Invariants +/// +/// Values of this type are always refcounted using `mmget`. The value of `mm_users` is non-zero. +#[repr(transparent)] +pub struct MmWithUser { + mm: Mm, +} + +// SAFETY: It is safe to call `mmput` on another thread than where `mmget` was called. +unsafe impl Send for MmWithUser {} +// SAFETY: All methods on `MmWithUser` can be called in parallel from several threads. +unsafe impl Sync for MmWithUser {} + +// SAFETY: By the type invariants, this type is always refcounted. +unsafe impl AlwaysRefCounted for MmWithUser { + #[inline] + fn inc_ref(&self) { + // SAFETY: The pointer is valid since self is a reference. + unsafe { bindings::mmget(self.as_raw()) }; + } + + #[inline] + unsafe fn dec_ref(obj: NonNull<Self>) { + // SAFETY: The caller is giving up their refcount. + unsafe { bindings::mmput(obj.cast().as_ptr()) }; + } +} + +// Make all `Mm` methods available on `MmWithUser`. +impl Deref for MmWithUser { + type Target = Mm; + + #[inline] + fn deref(&self) -> &Mm { + &self.mm + } +} + +// These methods are safe to call even if `mm_users` is zero. +impl Mm { + /// Call `mmgrab` on `current.mm`. + #[inline] + pub fn mmgrab_current() -> Option<ARef<Mm>> { + // SAFETY: It's safe to get the `mm` field from current. + let mm = unsafe { + let current = bindings::get_current(); + (*current).mm + }; + + if mm.is_null() { + return None; + } + + // SAFETY: The value of `current->mm` is guaranteed to be null or a valid `mm_struct`. We + // just checked that it's not null. Furthermore, the returned `&Mm` is valid only for the + // duration of this function, and `current->mm` will stay valid for that long. + let mm = unsafe { Mm::from_raw(mm) }; + + // This increments the refcount using `mmgrab`. + Some(ARef::from(mm)) + } + + /// Returns a raw pointer to the inner `mm_struct`. + #[inline] + pub fn as_raw(&self) -> *mut bindings::mm_struct { + self.mm.get() + } + + /// Obtain a reference from a raw pointer. + /// + /// # Safety + /// + /// The caller must ensure that `ptr` points at an `mm_struct`, and that it is not deallocated + /// during the lifetime 'a. + #[inline] + pub unsafe fn from_raw<'a>(ptr: *const bindings::mm_struct) -> &'a Mm { + // SAFETY: Caller promises that the pointer is valid for 'a. Layouts are compatible due to + // repr(transparent). + unsafe { &*ptr.cast() } + } + + /// Calls `mmget_not_zero` and returns a handle if it succeeds. + #[inline] + pub fn mmget_not_zero(&self) -> Option<ARef<MmWithUser>> { + // SAFETY: The pointer is valid since self is a reference. + let success = unsafe { bindings::mmget_not_zero(self.as_raw()) }; + + if success { + // SAFETY: We just created an `mmget` refcount. + Some(unsafe { ARef::from_raw(NonNull::new_unchecked(self.as_raw().cast())) }) + } else { + None + } + } +} + +// These methods require `mm_users` to be non-zero. +impl MmWithUser { + /// Obtain a reference from a raw pointer. + /// + /// # Safety + /// + /// The caller must ensure that `ptr` points at an `mm_struct`, and that `mm_users` remains + /// non-zero for the duration of the lifetime 'a. + #[inline] + pub unsafe fn from_raw<'a>(ptr: *const bindings::mm_struct) -> &'a MmWithUser { + // SAFETY: Caller promises that the pointer is valid for 'a. The layout is compatible due + // to repr(transparent). + unsafe { &*ptr.cast() } + } + + /// Lock the mmap read lock. + #[inline] + pub fn mmap_read_lock(&self) -> MmapReadGuard<'_> { + // SAFETY: The pointer is valid since self is a reference. + unsafe { bindings::mmap_read_lock(self.as_raw()) }; + + // INVARIANT: We just acquired the read lock. + MmapReadGuard { + mm: self, + _nts: NotThreadSafe, + } + } + + /// Try to lock the mmap read lock. + #[inline] + pub fn mmap_read_trylock(&self) -> Option<MmapReadGuard<'_>> { + // SAFETY: The pointer is valid since self is a reference. + let success = unsafe { bindings::mmap_read_trylock(self.as_raw()) }; + + if success { + // INVARIANT: We just acquired the read lock. + Some(MmapReadGuard { + mm: self, + _nts: NotThreadSafe, + }) + } else { + None + } + } +} + +/// A guard for the mmap read lock. +/// +/// # Invariants +/// +/// This `MmapReadGuard` guard owns the mmap read lock. +pub struct MmapReadGuard<'a> { + mm: &'a MmWithUser, + // `mmap_read_lock` and `mmap_read_unlock` must be called on the same thread + _nts: NotThreadSafe, +} + +impl Drop for MmapReadGuard<'_> { + #[inline] + fn drop(&mut self) { + // SAFETY: We hold the read lock by the type invariants. + unsafe { bindings::mmap_read_unlock(self.mm.as_raw()) }; + } +} -- 2.47.0.371.ga323438b13-goog This adds a type called VmAreaRef which is used when referencing a vma that you have read access to. Here, read access means that you hold either the mmap read lock or the vma read lock (or stronger). Additionally, a vma_lookup method is added to the mmap read guard, which enables you to obtain a &VmAreaRef in safe Rust code. This patch only provides a way to lock the mmap read lock, but a follow-up patch also provides a way to just lock the vma read lock. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/helpers/mm.c | 6 ++ rust/kernel/mm.rs | 21 ++++++ rust/kernel/mm/virt.rs | 176 +++++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 203 insertions(+) diff --git a/rust/helpers/mm.c b/rust/helpers/mm.c index XXXXXXX..XXXXXXX 100644 --- a/rust/helpers/mm.c +++ b/rust/helpers/mm.c @@ -XXX,XX +XXX,XX @@ void rust_helper_mmap_read_unlock(struct mm_struct *mm) { mmap_read_unlock(mm); } + +struct vm_area_struct *rust_helper_vma_lookup(struct mm_struct *mm, + unsigned long addr) +{ + return vma_lookup(mm, addr); +} diff --git a/rust/kernel/mm.rs b/rust/kernel/mm.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm.rs +++ b/rust/kernel/mm.rs @@ -XXX,XX +XXX,XX @@ }; use core::{ops::Deref, ptr::NonNull}; +pub mod virt; + /// A wrapper for the kernel's `struct mm_struct`. /// /// Since `mm_users` may be zero, the associated address space may not exist anymore. You can use @@ -XXX,XX +XXX,XX @@ pub struct MmapReadGuard<'a> { _nts: NotThreadSafe, } +impl<'a> MmapReadGuard<'a> { + /// Look up a vma at the given address. + #[inline] + pub fn vma_lookup(&self, vma_addr: usize) -> Option<&virt::VmAreaRef> { + // SAFETY: We hold a reference to the mm, so the pointer must be valid. Any value is okay + // for `vma_addr`. + let vma = unsafe { bindings::vma_lookup(self.mm.as_raw(), vma_addr as _) }; + + if vma.is_null() { + None + } else { + // SAFETY: We just checked that a vma was found, so the pointer is valid. Furthermore, + // the returned area will borrow from this read lock guard, so it can only be used + // while the mmap read lock is still held. + unsafe { Some(virt::VmAreaRef::from_raw(vma)) } + } + } +} + impl Drop for MmapReadGuard<'_> { #[inline] fn drop(&mut self) { diff --git a/rust/kernel/mm/virt.rs b/rust/kernel/mm/virt.rs new file mode 100644 index XXXXXXX..XXXXXXX --- /dev/null +++ b/rust/kernel/mm/virt.rs @@ -XXX,XX +XXX,XX @@ +// SPDX-License-Identifier: GPL-2.0 + +// Copyright (C) 2024 Google LLC. + +//! Virtual memory. + +use crate::{bindings, types::Opaque}; + +/// A wrapper for the kernel's `struct vm_area_struct` with read access. +/// +/// It represents an area of virtual memory. +/// +/// # Invariants +/// +/// The caller must hold the mmap read lock or the vma read lock. +#[repr(transparent)] +pub struct VmAreaRef { + vma: Opaque<bindings::vm_area_struct>, +} + +// Methods you can call when holding the mmap or vma read lock (or strong). They must be usable no +// matter what the vma flags are. +impl VmAreaRef { + /// Access a virtual memory area given a raw pointer. + /// + /// # Safety + /// + /// Callers must ensure that `vma` is valid for the duration of 'a, and that the mmap or vma + /// read lock (or stronger) is held for at least the duration of 'a. + #[inline] + pub unsafe fn from_raw<'a>(vma: *const bindings::vm_area_struct) -> &'a Self { + // SAFETY: The caller ensures that the invariants are satisfied for the duration of 'a. + unsafe { &*vma.cast() } + } + + /// Returns a raw pointer to this area. + #[inline] + pub fn as_ptr(&self) -> *mut bindings::vm_area_struct { + self.vma.get() + } + + /// Returns the flags associated with the virtual memory area. + /// + /// The possible flags are a combination of the constants in [`flags`]. + #[inline] + pub fn flags(&self) -> vm_flags_t { + // SAFETY: By the type invariants, the caller holds at least the mmap read lock, so this + // access is not a data race. + unsafe { (*self.as_ptr()).__bindgen_anon_2.vm_flags as _ } + } + + /// Returns the (inclusive) start address of the virtual memory area. + #[inline] + pub fn start(&self) -> usize { + // SAFETY: By the type invariants, the caller holds at least the mmap read lock, so this + // access is not a data race. + unsafe { (*self.as_ptr()).__bindgen_anon_1.__bindgen_anon_1.vm_start as _ } + } + + /// Returns the (exclusive) end address of the virtual memory area. + #[inline] + pub fn end(&self) -> usize { + // SAFETY: By the type invariants, the caller holds at least the mmap read lock, so this + // access is not a data race. + unsafe { (*self.as_ptr()).__bindgen_anon_1.__bindgen_anon_1.vm_end as _ } + } + + /// Zap pages in the given page range. + /// + /// This clears page table mappings for the range at the leaf level, leaving all other page + /// tables intact, and freeing any memory referenced by the VMA in this range. That is, + /// anonymous memory is completely freed, file-backed memory has its reference count on page + /// cache folio's dropped, any dirty data will still be written back to disk as usual. + #[inline] + pub fn zap_page_range_single(&self, address: usize, size: usize) { + // SAFETY: By the type invariants, the caller has read access to this VMA, which is + // sufficient for this method call. This method has no requirements on the vma flags. Any + // value of `address` and `size` is allowed. + unsafe { + bindings::zap_page_range_single( + self.as_ptr(), + address as _, + size as _, + core::ptr::null_mut(), + ) + }; + } +} + +/// The integer type used for vma flags. +#[doc(inline)] +pub use bindings::vm_flags_t; + +/// All possible flags for [`VmAreaRef`]. +pub mod flags { + use super::vm_flags_t; + use crate::bindings; + + /// No flags are set. + pub const NONE: vm_flags_t = bindings::VM_NONE as _; + + /// Mapping allows reads. + pub const READ: vm_flags_t = bindings::VM_READ as _; + + /// Mapping allows writes. + pub const WRITE: vm_flags_t = bindings::VM_WRITE as _; + + /// Mapping allows execution. + pub const EXEC: vm_flags_t = bindings::VM_EXEC as _; + + /// Mapping is shared. + pub const SHARED: vm_flags_t = bindings::VM_SHARED as _; + + /// Mapping may be updated to allow reads. + pub const MAYREAD: vm_flags_t = bindings::VM_MAYREAD as _; + + /// Mapping may be updated to allow writes. + pub const MAYWRITE: vm_flags_t = bindings::VM_MAYWRITE as _; + + /// Mapping may be updated to allow execution. + pub const MAYEXEC: vm_flags_t = bindings::VM_MAYEXEC as _; + + /// Mapping may be updated to be shared. + pub const MAYSHARE: vm_flags_t = bindings::VM_MAYSHARE as _; + + /// Page-ranges managed without `struct page`, just pure PFN. + pub const PFNMAP: vm_flags_t = bindings::VM_PFNMAP as _; + + /// Memory mapped I/O or similar. + pub const IO: vm_flags_t = bindings::VM_IO as _; + + /// Do not copy this vma on fork. + pub const DONTCOPY: vm_flags_t = bindings::VM_DONTCOPY as _; + + /// Cannot expand with mremap(). + pub const DONTEXPAND: vm_flags_t = bindings::VM_DONTEXPAND as _; + + /// Lock the pages covered when they are faulted in. + pub const LOCKONFAULT: vm_flags_t = bindings::VM_LOCKONFAULT as _; + + /// Is a VM accounted object. + pub const ACCOUNT: vm_flags_t = bindings::VM_ACCOUNT as _; + + /// Should the VM suppress accounting. + pub const NORESERVE: vm_flags_t = bindings::VM_NORESERVE as _; + + /// Huge TLB Page VM. + pub const HUGETLB: vm_flags_t = bindings::VM_HUGETLB as _; + + /// Synchronous page faults. (DAX-specific) + pub const SYNC: vm_flags_t = bindings::VM_SYNC as _; + + /// Architecture-specific flag. + pub const ARCH_1: vm_flags_t = bindings::VM_ARCH_1 as _; + + /// Wipe VMA contents in child on fork. + pub const WIPEONFORK: vm_flags_t = bindings::VM_WIPEONFORK as _; + + /// Do not include in the core dump. + pub const DONTDUMP: vm_flags_t = bindings::VM_DONTDUMP as _; + + /// Not soft dirty clean area. + pub const SOFTDIRTY: vm_flags_t = bindings::VM_SOFTDIRTY as _; + + /// Can contain `struct page` and pure PFN pages. + pub const MIXEDMAP: vm_flags_t = bindings::VM_MIXEDMAP as _; + + /// MADV_HUGEPAGE marked this vma. + pub const HUGEPAGE: vm_flags_t = bindings::VM_HUGEPAGE as _; + + /// MADV_NOHUGEPAGE marked this vma. + pub const NOHUGEPAGE: vm_flags_t = bindings::VM_NOHUGEPAGE as _; + + /// KSM may merge identical pages. + pub const MERGEABLE: vm_flags_t = bindings::VM_MERGEABLE as _; +} -- 2.47.0.371.ga323438b13-goog The vm_insert_page method is only usable on vmas with the VM_MIXEDMAP flag, so we introduce a new type to keep track of such vmas. The approach used in this patch assumes that we will not need to encode many flag combinations in the type. I don't think we need to encode more than VM_MIXEDMAP and VM_PFNMAP as things are now. However, if that becomes necessary, using generic parameters in a single type would scale better as the number of flags increases. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/kernel/mm/virt.rs | 70 +++++++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 69 insertions(+), 1 deletion(-) diff --git a/rust/kernel/mm/virt.rs b/rust/kernel/mm/virt.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm/virt.rs +++ b/rust/kernel/mm/virt.rs @@ -XXX,XX +XXX,XX @@ //! Virtual memory. -use crate::{bindings, types::Opaque}; +use crate::{ + bindings, + error::{to_result, Result}, + page::Page, + types::Opaque, +}; + +use core::ops::Deref; /// A wrapper for the kernel's `struct vm_area_struct` with read access. /// @@ -XXX,XX +XXX,XX @@ pub fn zap_page_range_single(&self, address: usize, size: usize) { ) }; } + + /// Check whether the `VM_MIXEDMAP` flag is set. + /// + /// This can be used to access methods that require `VM_MIXEDMAP` to be set. + #[inline] + pub fn as_mixedmap_vma(&self) -> Option<&VmAreaMixedMap> { + if self.flags() & flags::MIXEDMAP != 0 { + // SAFETY: We just checked that `VM_MIXEDMAP` is set. All other requirements are + // satisfied by the type invariants of `VmAreaRef`. + Some(unsafe { VmAreaMixedMap::from_raw(self.as_ptr()) }) + } else { + None + } + } +} + +/// A wrapper for the kernel's `struct vm_area_struct` with read access and `VM_MIXEDMAP` set. +/// +/// It represents an area of virtual memory. +/// +/// # Invariants +/// +/// The caller must hold the mmap read lock or the vma read lock. The `VM_MIXEDMAP` flag must be +/// set. +#[repr(transparent)] +pub struct VmAreaMixedMap { + vma: VmAreaRef, +} + +// Make all `VmAreaRef` methods available on `VmAreaMixedMap`. +impl Deref for VmAreaMixedMap { + type Target = VmAreaRef; + + #[inline] + fn deref(&self) -> &VmAreaRef { + &self.vma + } +} + +impl VmAreaMixedMap { + /// Access a virtual memory area given a raw pointer. + /// + /// # Safety + /// + /// Callers must ensure that `vma` is valid for the duration of 'a, and that the mmap read lock + /// (or stronger) is held for at least the duration of 'a. The `VM_MIXEDMAP` flag must be set. + #[inline] + pub unsafe fn from_raw<'a>(vma: *const bindings::vm_area_struct) -> &'a Self { + // SAFETY: The caller ensures that the invariants are satisfied for the duration of 'a. + unsafe { &*vma.cast() } + } + + /// Maps a single page at the given address within the virtual memory area. + /// + /// This operation does not take ownership of the page. + #[inline] + pub fn vm_insert_page(&self, address: usize, page: &Page) -> Result { + // SAFETY: The caller has read access and has verified that `VM_MIXEDMAP` is set. The page + // is order 0. The address is checked on the C side so it can take any value. + to_result(unsafe { bindings::vm_insert_page(self.as_ptr(), address as _, page.as_ptr()) }) + } } /// The integer type used for vma flags. -- 2.47.0.371.ga323438b13-goog Currently, the binder driver always uses the mmap lock to make changes to its vma. Because the mmap lock is global to the process, this can involve significant contention. However, the kernel has a feature called per-vma locks, which can significantly reduce contention. For example, you can take a vma lock in parallel with an mmap write lock. This is important because contention on the mmap lock has been a long-term recurring challenge for the Binder driver. This patch introduces support for using `lock_vma_under_rcu` from Rust. The Rust Binder driver will be able to use this to reduce contention on the mmap lock. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/helpers/mm.c | 5 +++++ rust/kernel/mm.rs | 56 +++++++++++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 61 insertions(+) diff --git a/rust/helpers/mm.c b/rust/helpers/mm.c index XXXXXXX..XXXXXXX 100644 --- a/rust/helpers/mm.c +++ b/rust/helpers/mm.c @@ -XXX,XX +XXX,XX @@ struct vm_area_struct *rust_helper_vma_lookup(struct mm_struct *mm, { return vma_lookup(mm, addr); } + +void rust_helper_vma_end_read(struct vm_area_struct *vma) +{ + vma_end_read(vma); +} diff --git a/rust/kernel/mm.rs b/rust/kernel/mm.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm.rs +++ b/rust/kernel/mm.rs @@ -XXX,XX +XXX,XX @@ use core::{ops::Deref, ptr::NonNull}; pub mod virt; +use virt::VmAreaRef; /// A wrapper for the kernel's `struct mm_struct`. /// @@ -XXX,XX +XXX,XX @@ pub unsafe fn from_raw<'a>(ptr: *const bindings::mm_struct) -> &'a MmWithUser { unsafe { &*ptr.cast() } } + /// Attempt to access a vma using the vma read lock. + /// + /// This is an optimistic trylock operation, so it may fail if there is contention. In that + /// case, you should fall back to taking the mmap read lock. + /// + /// When per-vma locks are disabled, this always returns `None`. + #[inline] + pub fn lock_vma_under_rcu(&self, vma_addr: usize) -> Option<VmaReadGuard<'_>> { + #[cfg(CONFIG_PER_VMA_LOCK)] + { + // SAFETY: Calling `bindings::lock_vma_under_rcu` is always okay given an mm where + // `mm_users` is non-zero. + let vma = unsafe { bindings::lock_vma_under_rcu(self.as_raw(), vma_addr as _) }; + if !vma.is_null() { + return Some(VmaReadGuard { + // SAFETY: If `lock_vma_under_rcu` returns a non-null ptr, then it points at a + // valid vma. The vma is stable for as long as the vma read lock is held. + vma: unsafe { VmAreaRef::from_raw(vma) }, + _nts: NotThreadSafe, + }); + } + } + + None + } + /// Lock the mmap read lock. #[inline] pub fn mmap_read_lock(&self) -> MmapReadGuard<'_> { @@ -XXX,XX +XXX,XX @@ fn drop(&mut self) { unsafe { bindings::mmap_read_unlock(self.mm.as_raw()) }; } } + +/// A guard for the vma read lock. +/// +/// # Invariants +/// +/// This `VmaReadGuard` guard owns the vma read lock. +pub struct VmaReadGuard<'a> { + vma: &'a VmAreaRef, + // `vma_end_read` must be called on the same thread as where the lock was taken + _nts: NotThreadSafe, +} + +// Make all `VmAreaRef` methods available on `VmaReadGuard`. +impl Deref for VmaReadGuard<'_> { + type Target = VmAreaRef; + + #[inline] + fn deref(&self) -> &VmAreaRef { + self.vma + } +} + +impl Drop for VmaReadGuard<'_> { + #[inline] + fn drop(&mut self) { + // SAFETY: We hold the read lock by the type invariants. + unsafe { bindings::vma_end_read(self.vma.as_ptr()) }; + } +} -- 2.47.0.371.ga323438b13-goog Adds an MmWithUserAsync type that uses mmput_async when dropped but is otherwise identical to MmWithUser. This has to be done using a separate type because the thing we are changing is the destructor. Rust Binder needs this to avoid a certain deadlock. See commit 9a9ab0d96362 ("binder: fix race between mmput() and do_exit()") for details. It's also needed in the shrinker to avoid cleaning up the mm in the shrinker's context. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/kernel/mm.rs | 49 +++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 49 insertions(+) diff --git a/rust/kernel/mm.rs b/rust/kernel/mm.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm.rs +++ b/rust/kernel/mm.rs @@ -XXX,XX +XXX,XX @@ fn deref(&self) -> &Mm { } } +/// A wrapper for the kernel's `struct mm_struct`. +/// +/// This type is identical to `MmWithUser` except that it uses `mmput_async` when dropping a +/// refcount. This means that the destructor of `ARef<MmWithUserAsync>` is safe to call in atomic +/// context. +/// +/// # Invariants +/// +/// Values of this type are always refcounted using `mmget`. The value of `mm_users` is non-zero. +#[repr(transparent)] +pub struct MmWithUserAsync { + mm: MmWithUser, +} + +// SAFETY: It is safe to call `mmput_async` on another thread than where `mmget` was called. +unsafe impl Send for MmWithUserAsync {} +// SAFETY: All methods on `MmWithUserAsync` can be called in parallel from several threads. +unsafe impl Sync for MmWithUserAsync {} + +// SAFETY: By the type invariants, this type is always refcounted. +unsafe impl AlwaysRefCounted for MmWithUserAsync { + fn inc_ref(&self) { + // SAFETY: The pointer is valid since self is a reference. + unsafe { bindings::mmget(self.as_raw()) }; + } + + unsafe fn dec_ref(obj: NonNull<Self>) { + // SAFETY: The caller is giving up their refcount. + unsafe { bindings::mmput_async(obj.cast().as_ptr()) }; + } +} + +// Make all `MmWithUser` methods available on `MmWithUserAsync`. +impl Deref for MmWithUserAsync { + type Target = MmWithUser; + + #[inline] + fn deref(&self) -> &MmWithUser { + &self.mm + } +} + // These methods are safe to call even if `mm_users` is zero. impl Mm { /// Call `mmgrab` on `current.mm`. @@ -XXX,XX +XXX,XX @@ pub unsafe fn from_raw<'a>(ptr: *const bindings::mm_struct) -> &'a MmWithUser { unsafe { &*ptr.cast() } } + /// Use `mmput_async` when dropping this refcount. + #[inline] + pub fn into_mmput_async(me: ARef<MmWithUser>) -> ARef<MmWithUserAsync> { + // SAFETY: The layouts and invariants are compatible. + unsafe { ARef::from_raw(ARef::into_raw(me).cast()) } + } + /// Attempt to access a vma using the vma read lock. /// /// This is an optimistic trylock operation, so it may fail if there is contention. In that -- 2.47.0.371.ga323438b13-goog This type will be used when setting up a new vma in an f_ops->mmap() hook. Using a separate type from VmAreaRef allows us to have a separate set of operations that you are only able to use during the mmap() hook. For example, the VM_MIXEDMAP flag must not be changed after the initial setup that happens during the f_ops->mmap() hook. To avoid setting invalid flag values, the methods for clearing VM_MAYWRITE and similar involve a check of VM_WRITE, and return an error if VM_WRITE is set. Trying to use `try_clear_maywrite` without checking the return value results in a compilation error because the `Result` type is marked #[must_use]. For now, there's only a method for VM_MIXEDMAP and not VM_PFNMAP. When we add a VM_PFNMAP method, we will need some way to prevent you from setting both VM_MIXEDMAP and VM_PFNMAP on the same vma. Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/kernel/mm/virt.rs | 179 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 178 insertions(+), 1 deletion(-) diff --git a/rust/kernel/mm/virt.rs b/rust/kernel/mm/virt.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm/virt.rs +++ b/rust/kernel/mm/virt.rs @@ -XXX,XX +XXX,XX @@ use crate::{ bindings, - error::{to_result, Result}, + error::{code::EINVAL, to_result, Result}, page::Page, types::Opaque, }; @@ -XXX,XX +XXX,XX @@ pub fn vm_insert_page(&self, address: usize, page: &Page) -> Result { } } +/// A builder for setting up a vma in an `f_ops->mmap()` hook. +/// +/// # Invariants +/// +/// For the duration of 'a, the referenced vma must be undergoing initialization in an +/// `f_ops->mmap()` hook. +pub struct VmAreaNew { + vma: VmAreaRef, +} + +// Make all `VmAreaRef` methods available on `VmAreaNew`. +impl Deref for VmAreaNew { + type Target = VmAreaRef; + + #[inline] + fn deref(&self) -> &VmAreaRef { + &self.vma + } +} + +impl VmAreaNew { + /// Access a virtual memory area given a raw pointer. + /// + /// # Safety + /// + /// Callers must ensure that `vma` is undergoing initial vma setup for the duration of 'a. + #[inline] + pub unsafe fn from_raw<'a>(vma: *const bindings::vm_area_struct) -> &'a Self { + // SAFETY: The caller ensures that the invariants are satisfied for the duration of 'a. + unsafe { &*vma.cast() } + } + + /// Internal method for updating the vma flags. + /// + /// # Safety + /// + /// This must not be used to set the flags to an invalid value. + #[inline] + unsafe fn update_flags(&self, set: vm_flags_t, unset: vm_flags_t) { + let mut flags = self.flags(); + flags |= set; + flags &= !unset; + + // SAFETY: This is not a data race: the vma is undergoing initial setup, so it's not yet + // shared. Additionally, `VmAreaNew` is `!Sync`, so it cannot be used to write in parallel. + // The caller promises that this does not set the flags to an invalid value. + unsafe { (*self.as_ptr()).__bindgen_anon_2.__vm_flags = flags }; + } + + /// Set the `VM_MIXEDMAP` flag on this vma. + /// + /// This enables the vma to contain both `struct page` and pure PFN pages. Returns a reference + /// that can be used to call `vm_insert_page` on the vma. + #[inline] + pub fn set_mixedmap(&self) -> &VmAreaMixedMap { + // SAFETY: We don't yet provide a way to set VM_PFNMAP, so this cannot put the flags in an + // invalid state. + unsafe { self.update_flags(flags::MIXEDMAP, 0) }; + + // SAFETY: We just set `VM_MIXEDMAP` on the vma. + unsafe { VmAreaMixedMap::from_raw(self.vma.as_ptr()) } + } + + /// Set the `VM_IO` flag on this vma. + /// + /// This marks the vma as being a memory-mapped I/O region. + #[inline] + pub fn set_io(&self) { + // SAFETY: Setting the VM_IO flag is always okay. + unsafe { self.update_flags(flags::IO, 0) }; + } + + /// Set the `VM_DONTEXPAND` flag on this vma. + /// + /// This prevents the vma from being expanded with `mremap()`. + #[inline] + pub fn set_dontexpand(&self) { + // SAFETY: Setting the VM_DONTEXPAND flag is always okay. + unsafe { self.update_flags(flags::DONTEXPAND, 0) }; + } + + /// Set the `VM_DONTCOPY` flag on this vma. + /// + /// This prevents the vma from being copied on fork. This option is only permanent if `VM_IO` + /// is set. + #[inline] + pub fn set_dontcopy(&self) { + // SAFETY: Setting the VM_DONTCOPY flag is always okay. + unsafe { self.update_flags(flags::DONTCOPY, 0) }; + } + + /// Set the `VM_DONTDUMP` flag on this vma. + /// + /// This prevents the vma from being included in core dumps. This option is only permanent if + /// `VM_IO` is set. + #[inline] + pub fn set_dontdump(&self) { + // SAFETY: Setting the VM_DONTDUMP flag is always okay. + unsafe { self.update_flags(flags::DONTDUMP, 0) }; + } + + /// Returns whether `VM_READ` is set. + /// + /// This flag indicates whether userspace is mapping this vma as readable. + #[inline] + pub fn get_read(&self) -> bool { + (self.flags() & flags::READ) != 0 + } + + /// Try to clear the `VM_MAYREAD` flag, failing if `VM_READ` is set. + /// + /// This flag indicates whether userspace is allowed to make this vma readable with + /// `mprotect()`. + /// + /// Note that this operation is irreversible. Once `VM_MAYREAD` has been cleared, it can never + /// be set again. + #[inline] + pub fn try_clear_mayread(&self) -> Result { + if self.get_read() { + return Err(EINVAL); + } + // SAFETY: Clearing `VM_MAYREAD` is okay when `VM_READ` is not set. + unsafe { self.update_flags(0, flags::MAYREAD) }; + Ok(()) + } + + /// Returns whether `VM_WRITE` is set. + /// + /// This flag indicates whether userspace is mapping this vma as writable. + #[inline] + pub fn get_write(&self) -> bool { + (self.flags() & flags::WRITE) != 0 + } + + /// Try to clear the `VM_MAYWRITE` flag, failing if `VM_WRITE` is set. + /// + /// This flag indicates whether userspace is allowed to make this vma writable with + /// `mprotect()`. + /// + /// Note that this operation is irreversible. Once `VM_MAYWRITE` has been cleared, it can never + /// be set again. + #[inline] + pub fn try_clear_maywrite(&self) -> Result { + if self.get_write() { + return Err(EINVAL); + } + // SAFETY: Clearing `VM_MAYWRITE` is okay when `VM_WRITE` is not set. + unsafe { self.update_flags(0, flags::MAYWRITE) }; + Ok(()) + } + + /// Returns whether `VM_EXEC` is set. + /// + /// This flag indicates whether userspace is mapping this vma as executable. + #[inline] + pub fn get_exec(&self) -> bool { + (self.flags() & flags::EXEC) != 0 + } + + /// Try to clear the `VM_MAYEXEC` flag, failing if `VM_EXEC` is set. + /// + /// This flag indicates whether userspace is allowed to make this vma executable with + /// `mprotect()`. + /// + /// Note that this operation is irreversible. Once `VM_MAYEXEC` has been cleared, it can never + /// be set again. + #[inline] + pub fn try_clear_mayexec(&self) -> Result { + if self.get_exec() { + return Err(EINVAL); + } + // SAFETY: Clearing `VM_MAYEXEC` is okay when `VM_EXEC` is not set. + unsafe { self.update_flags(0, flags::MAYEXEC) }; + Ok(()) + } +} + /// The integer type used for vma flags. #[doc(inline)] pub use bindings::vm_flags_t; -- 2.47.0.371.ga323438b13-goog Add the ability to write a file_operations->mmap hook in Rust when using the miscdevice abstraction. The `vma` argument to the `mmap` hook uses the `VmAreaNew` type from the previous commit; this type provides the correct set of operations for a file_operations->mmap hook. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/kernel/miscdevice.rs | 28 ++++++++++++++++++++++++++++ 1 file changed, 28 insertions(+) diff --git a/rust/kernel/miscdevice.rs b/rust/kernel/miscdevice.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/miscdevice.rs +++ b/rust/kernel/miscdevice.rs @@ -XXX,XX +XXX,XX @@ use crate::{ bindings, error::{to_result, Error, Result, VTABLE_DEFAULT_ERROR}, + mm::virt::VmAreaNew, prelude::*, str::CStr, types::{ForeignOwnable, Opaque}, @@ -XXX,XX +XXX,XX @@ fn release(device: Self::Ptr) { drop(device); } + /// Handle for mmap. + fn mmap(_device: <Self::Ptr as ForeignOwnable>::Borrowed<'_>, _vma: &VmAreaNew) -> Result { + kernel::build_error(VTABLE_DEFAULT_ERROR) + } + /// Handler for ioctls. /// /// The `cmd` argument is usually manipulated using the utilties in [`kernel::ioctl`]. @@ -XXX,XX +XXX,XX @@ impl<T: MiscDevice> VtableHelper<T> { const VTABLE: bindings::file_operations = bindings::file_operations { open: Some(fops_open::<T>), release: Some(fops_release::<T>), + mmap: maybe_fn(T::HAS_MMAP, fops_mmap::<T>), unlocked_ioctl: maybe_fn(T::HAS_IOCTL, fops_ioctl::<T>), #[cfg(CONFIG_COMPAT)] compat_ioctl: if T::HAS_COMPAT_IOCTL { @@ -XXX,XX +XXX,XX @@ impl<T: MiscDevice> VtableHelper<T> { 0 } +/// # Safety +/// +/// `file` must be a valid file that is associated with a `MiscDeviceRegistration<T>`. +/// `vma` must be a vma that is currently being mmap'ed with this file. +unsafe extern "C" fn fops_mmap<T: MiscDevice>( + file: *mut bindings::file, + vma: *mut bindings::vm_area_struct, +) -> c_int { + // SAFETY: The mmap call of a file can access the private data. + let private = unsafe { (*file).private_data }; + // SAFETY: Mmap calls can borrow the private data of the file. + let device = unsafe { <T::Ptr as ForeignOwnable>::borrow(private) }; + // SAFETY: The caller provides a vma that is undergoing initial VMA setup. + let area = unsafe { VmAreaNew::from_raw(vma) }; + + match T::mmap(device, area) { + Ok(()) => 0, + Err(err) => err.to_errno() as c_int, + } +} + /// # Safety /// /// `file` must be a valid file that is associated with a `MiscDeviceRegistration<T>`. -- 2.47.0.371.ga323438b13-goog Introduce a new type called `CurrentTask` that lets you perform various operations that are only safe on the `current` task. Use the new type to provide a way to access the current mm without incrementing its refcount. With this change, you can write stuff such as let vma = current!().mm().lock_vma_under_rcu(addr); without incrementing any refcounts. Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- Reviewers: Does accessing task->mm on a non-current task require rcu protection? Christian: If you submit the PidNamespace abstractions this cycle, I'll update this to also apply to PidNamespace. --- rust/kernel/mm.rs | 22 ------------------ rust/kernel/task.rs | 64 ++++++++++++++++++++++++++++++++++++++++++----------- 2 files changed, 51 insertions(+), 35 deletions(-) diff --git a/rust/kernel/mm.rs b/rust/kernel/mm.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm.rs +++ b/rust/kernel/mm.rs @@ -XXX,XX +XXX,XX @@ fn deref(&self) -> &MmWithUser { // These methods are safe to call even if `mm_users` is zero. impl Mm { - /// Call `mmgrab` on `current.mm`. - #[inline] - pub fn mmgrab_current() -> Option<ARef<Mm>> { - // SAFETY: It's safe to get the `mm` field from current. - let mm = unsafe { - let current = bindings::get_current(); - (*current).mm - }; - - if mm.is_null() { - return None; - } - - // SAFETY: The value of `current->mm` is guaranteed to be null or a valid `mm_struct`. We - // just checked that it's not null. Furthermore, the returned `&Mm` is valid only for the - // duration of this function, and `current->mm` will stay valid for that long. - let mm = unsafe { Mm::from_raw(mm) }; - - // This increments the refcount using `mmgrab`. - Some(ARef::from(mm)) - } - /// Returns a raw pointer to the inner `mm_struct`. #[inline] pub fn as_raw(&self) -> *mut bindings::mm_struct { diff --git a/rust/kernel/task.rs b/rust/kernel/task.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/task.rs +++ b/rust/kernel/task.rs @@ -XXX,XX +XXX,XX @@ unsafe impl Send for Task {} // synchronised by C code (e.g., `signal_pending`). unsafe impl Sync for Task {} +/// Represents a [`Task`] obtained from the `current` global. +/// +/// This type exists to provide more efficient operations that are only valid on the current task. +/// For example, to retrieve the pid-namespace of a task, you must use rcu protection unless it is +/// the current task. +/// +/// # Invariants +/// +/// Must be equal to `current` of some thread that is currently running somewhere. +pub struct CurrentTask(Task); + +// Make all `Task` methods available on `CurrentTask`. +impl Deref for CurrentTask { + type Target = Task; + #[inline] + fn deref(&self) -> &Task { + &self.0 + } +} + /// The type of process identifiers (PIDs). type Pid = bindings::pid_t; @@ -XXX,XX +XXX,XX @@ pub fn current_raw() -> *mut bindings::task_struct { /// # Safety /// /// Callers must ensure that the returned object doesn't outlive the current task/thread. - pub unsafe fn current() -> impl Deref<Target = Task> { - struct TaskRef<'a> { - task: &'a Task, - _not_send: NotThreadSafe, + pub unsafe fn current() -> impl Deref<Target = CurrentTask> { + struct TaskRef { + task: *const CurrentTask, } - impl Deref for TaskRef<'_> { - type Target = Task; + impl Deref for TaskRef { + type Target = CurrentTask; fn deref(&self) -> &Self::Target { - self.task + // SAFETY: The returned reference borrows from this `TaskRef`, so it cannot outlive + // the `TaskRef`, which the caller of `Task::current()` has promised will not + // outlive the task/thread for which `self.task` is the `current` pointer. Thus, it + // is okay to return a `CurrentTask` reference here. + unsafe { &*self.task } } } - let current = Task::current_raw(); TaskRef { - // SAFETY: If the current thread is still running, the current task is valid. Given - // that `TaskRef` is not `Send`, we know it cannot be transferred to another thread - // (where it could potentially outlive the caller). - task: unsafe { &*current.cast() }, - _not_send: NotThreadSafe, + task: Task::current_raw().cast(), } } @@ -XXX,XX +XXX,XX @@ pub fn wake_up(&self) { } } +impl CurrentTask { + /// Access the address space of this task. + /// + /// To increment the refcount of the referenced `mm`, you can use `ARef::from`. + #[inline] + pub fn mm(&self) -> Option<&MmWithUser> { + let mm = unsafe { (*self.as_ptr()).mm }; + + if mm.is_null() { + None + } else { + // SAFETY: If `current->mm` is non-null, then it references a valid mm with a non-zero + // value of `mm_users`. The returned `&MmWithUser` borrows from `CurrentTask`, so the + // `&MmWithUser` cannot escape the current task, meaning `mm_users` can't reach zero + // while the reference is still live. + Some(unsafe { MmWithUser::from_raw(mm) }) + } + } +} + // SAFETY: The type invariants guarantee that `Task` is always refcounted. unsafe impl crate::types::AlwaysRefCounted for Task { fn inc_ref(&self) { -- 2.47.0.371.ga323438b13-goog This updates the vm_area_struct support to use the approach we discussed at LPC where there are several different Rust wrappers for vm_area_struct depending on the kind of access you have to the vma. Each case allows a different set of operations on the vma. This depends on several PULLs that have landed / are landing during the currently ongoing merge window. To: Miguel Ojeda <ojeda@kernel.org> To: Matthew Wilcox <willy@infradead.org> To: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> To: Vlastimil Babka <vbabka@suse.cz> To: John Hubbard <jhubbard@nvidia.com> To: Liam R. Howlett <Liam.Howlett@oracle.com> To: Andrew Morton <akpm@linux-foundation.org> To: Greg Kroah-Hartman <gregkh@linuxfoundation.org> To: Arnd Bergmann <arnd@arndb.de> To: Christian Brauner <brauner@kernel.org> To: Jann Horn <jannh@google.com> To: Suren Baghdasaryan <surenb@google.com> Cc: Alex Gaynor <alex.gaynor@gmail.com> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Gary Guo <gary@garyguo.net> Cc: Björn Roy Baron <bjorn3_gh@protonmail.com> Cc: Benno Lossin <benno.lossin@proton.me> Cc: Andreas Hindborg <a.hindborg@kernel.org> Cc: Trevor Gross <tmgross@umich.edu> Cc: linux-kernel@vger.kernel.org Cc: linux-mm@kvack.org Cc: rust-for-linux@vger.kernel.org Cc: Alice Ryhl <aliceryhl@google.com> Changes in v10: - Update docs for `set_io`. - Check address in `zap_page_range_single`. - Completely redo the last patch. - Link to v9: https://lore.kernel.org/r/20241122-vma-v9-0-7127bfcdd54e@google.com Changes in v9: - Be more explicit about VmAreaNew being used with f_ops->mmap(). - Point out that clearing VM_MAYWRITE is irreversible. - Use __vm_flags to set the flags. - Use as_ and into_ prefixes for conversions. - Update lock_vma_under_rcu docs and commit msg - Mention that VmAreaRef::end is exclusive. - Reword docs for zap_page_range_single. - Minor fixes to flag docs. - Add way to access current->mm without a refcount increment. - Link to v8: https://lore.kernel.org/r/20241120-vma-v8-0-eb31425da66b@google.com Changes in v8: - Split series into more commits to ease review. - Improve read locks based on Lorenzo's doc: either the mmap or vma lock can be used. - Get rid of mmap write lock because it's possible to avoid the need for it. - Do not allow invalid flag combinations on VmAreaNew. - Link to v7: https://lore.kernel.org/r/20241014-vma-v7-0-01e32f861195@google.com Changes in v7: - Make the mmap read/write lock guards respect strict owner semantics. - Link to v6: https://lore.kernel.org/r/20241010-vma-v6-0-d89039b6f573@google.com Changes in v6: - Introduce VmArea{Ref,Mut,New} distinction. - Add a second patchset for miscdevice. - Rebase on char-misc-next (currently on v6.12-rc2). - Link to v5: https://lore.kernel.org/r/20240806-vma-v5-1-04018f05de2b@google.com Changes in v5: - Rename VmArea::from_raw_vma to from_raw. - Use Pin for mutable VmArea references. - Go through `ARef::from` in `mmgrab_current`. - Link to v4: https://lore.kernel.org/r/20240802-vma-v4-1-091a87058a43@google.com Changes in v4: - Pull out ARef::into_raw into a separate patch. - Update invariants and struct documentation. - Rename from_raw_mm to from_raw. - Link to v3: https://lore.kernel.org/r/20240801-vma-v3-1-db6c1c0afda9@google.com Changes in v3: - Reorder entries in mm.rs. - Use ARef for mmput_async helper. - Clarify that VmArea requires you to hold the mmap read or write lock. - Link to v2: https://lore.kernel.org/r/20240727-vma-v2-1-ab3e5927dc3a@google.com Changes in v2: - mm.rs is redesigned from scratch making use of AsRef - Add notes about whether destructors may sleep - Rename Area to VmArea - Link to v1: https://lore.kernel.org/r/20240723-vma-v1-1-32ad5a0118ee@google.com --- Alice Ryhl (8): mm: rust: add abstraction for struct mm_struct mm: rust: add vm_area_struct methods that require read access mm: rust: add vm_insert_page mm: rust: add lock_vma_under_rcu mm: rust: add mmput_async support mm: rust: add VmAreaNew for f_ops->mmap() rust: miscdevice: add mmap support task: rust: rework how current is accessed rust/helpers/helpers.c | 1 + rust/helpers/mm.c | 50 ++++++ rust/kernel/lib.rs | 1 + rust/kernel/miscdevice.rs | 28 +++ rust/kernel/mm.rs | 323 ++++++++++++++++++++++++++++++++++ rust/kernel/mm/virt.rs | 429 ++++++++++++++++++++++++++++++++++++++++++++++ rust/kernel/task.rs | 284 +++++++++++++++++------------- 7 files changed, 999 insertions(+), 117 deletions(-) --- base-commit: a62eb7bdef0b0009d1e141198e790fbc588ad0a1 change-id: 20240723-vma-f80119f9fb35 Best regards, -- Alice Ryhl <aliceryhl@google.com> These abstractions allow you to reference a `struct mm_struct` using both mmgrab and mmget refcounts. This is done using two Rust types: * Mm - represents an mm_struct where you don't know anything about the value of mm_users. * MmWithUser - represents an mm_struct where you know at compile time that mm_users is non-zero. This allows us to encode in the type system whether a method requires that mm_users is non-zero or not. For instance, you can always call `mmget_not_zero` but you can only call `mmap_read_lock` when mm_users is non-zero. It's possible to access current->mm without a refcount increment, but that is added in a later patch of this series. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/helpers/helpers.c | 1 + rust/helpers/mm.c | 39 +++++++++ rust/kernel/lib.rs | 1 + rust/kernel/mm.rs | 219 +++++++++++++++++++++++++++++++++++++++++++++++++ 4 files changed, 260 insertions(+) diff --git a/rust/helpers/helpers.c b/rust/helpers/helpers.c index XXXXXXX..XXXXXXX 100644 --- a/rust/helpers/helpers.c +++ b/rust/helpers/helpers.c @@ -XXX,XX +XXX,XX @@ #include "fs.c" #include "jump_label.c" #include "kunit.c" +#include "mm.c" #include "mutex.c" #include "page.c" #include "pid_namespace.c" diff --git a/rust/helpers/mm.c b/rust/helpers/mm.c new file mode 100644 index XXXXXXX..XXXXXXX --- /dev/null +++ b/rust/helpers/mm.c @@ -XXX,XX +XXX,XX @@ +// SPDX-License-Identifier: GPL-2.0 + +#include <linux/mm.h> +#include <linux/sched/mm.h> + +void rust_helper_mmgrab(struct mm_struct *mm) +{ + mmgrab(mm); +} + +void rust_helper_mmdrop(struct mm_struct *mm) +{ + mmdrop(mm); +} + +void rust_helper_mmget(struct mm_struct *mm) +{ + mmget(mm); +} + +bool rust_helper_mmget_not_zero(struct mm_struct *mm) +{ + return mmget_not_zero(mm); +} + +void rust_helper_mmap_read_lock(struct mm_struct *mm) +{ + mmap_read_lock(mm); +} + +bool rust_helper_mmap_read_trylock(struct mm_struct *mm) +{ + return mmap_read_trylock(mm); +} + +void rust_helper_mmap_read_unlock(struct mm_struct *mm) +{ + mmap_read_unlock(mm); +} diff --git a/rust/kernel/lib.rs b/rust/kernel/lib.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/lib.rs +++ b/rust/kernel/lib.rs @@ -XXX,XX +XXX,XX @@ pub mod kunit; pub mod list; pub mod miscdevice; +pub mod mm; #[cfg(CONFIG_NET)] pub mod net; pub mod page; diff --git a/rust/kernel/mm.rs b/rust/kernel/mm.rs new file mode 100644 index XXXXXXX..XXXXXXX --- /dev/null +++ b/rust/kernel/mm.rs @@ -XXX,XX +XXX,XX @@ +// SPDX-License-Identifier: GPL-2.0 + +// Copyright (C) 2024 Google LLC. + +//! Memory management. +//! +//! C header: [`include/linux/mm.h`](srctree/include/linux/mm.h) + +use crate::{ + bindings, + types::{ARef, AlwaysRefCounted, NotThreadSafe, Opaque}, +}; +use core::{ops::Deref, ptr::NonNull}; + +/// A wrapper for the kernel's `struct mm_struct`. +/// +/// Since `mm_users` may be zero, the associated address space may not exist anymore. You can use +/// [`mmget_not_zero`] to be able to access the address space. +/// +/// The `ARef<Mm>` smart pointer holds an `mmgrab` refcount. Its destructor may sleep. +/// +/// # Invariants +/// +/// Values of this type are always refcounted using `mmgrab`. +/// +/// [`mmget_not_zero`]: Mm::mmget_not_zero +#[repr(transparent)] +pub struct Mm { + mm: Opaque<bindings::mm_struct>, +} + +// SAFETY: It is safe to call `mmdrop` on another thread than where `mmgrab` was called. +unsafe impl Send for Mm {} +// SAFETY: All methods on `Mm` can be called in parallel from several threads. +unsafe impl Sync for Mm {} + +// SAFETY: By the type invariants, this type is always refcounted. +unsafe impl AlwaysRefCounted for Mm { + #[inline] + fn inc_ref(&self) { + // SAFETY: The pointer is valid since self is a reference. + unsafe { bindings::mmgrab(self.as_raw()) }; + } + + #[inline] + unsafe fn dec_ref(obj: NonNull<Self>) { + // SAFETY: The caller is giving up their refcount. + unsafe { bindings::mmdrop(obj.cast().as_ptr()) }; + } +} + +/// A wrapper for the kernel's `struct mm_struct`. +/// +/// This type is like [`Mm`], but with non-zero `mm_users`. It can only be used when `mm_users` can +/// be proven to be non-zero at compile-time, usually because the relevant code holds an `mmget` +/// refcount. It can be used to access the associated address space. +/// +/// The `ARef<MmWithUser>` smart pointer holds an `mmget` refcount. Its destructor may sleep. +/// +/// # Invariants +/// +/// Values of this type are always refcounted using `mmget`. The value of `mm_users` is non-zero. +#[repr(transparent)] +pub struct MmWithUser { + mm: Mm, +} + +// SAFETY: It is safe to call `mmput` on another thread than where `mmget` was called. +unsafe impl Send for MmWithUser {} +// SAFETY: All methods on `MmWithUser` can be called in parallel from several threads. +unsafe impl Sync for MmWithUser {} + +// SAFETY: By the type invariants, this type is always refcounted. +unsafe impl AlwaysRefCounted for MmWithUser { + #[inline] + fn inc_ref(&self) { + // SAFETY: The pointer is valid since self is a reference. + unsafe { bindings::mmget(self.as_raw()) }; + } + + #[inline] + unsafe fn dec_ref(obj: NonNull<Self>) { + // SAFETY: The caller is giving up their refcount. + unsafe { bindings::mmput(obj.cast().as_ptr()) }; + } +} + +// Make all `Mm` methods available on `MmWithUser`. +impl Deref for MmWithUser { + type Target = Mm; + + #[inline] + fn deref(&self) -> &Mm { + &self.mm + } +} + +// These methods are safe to call even if `mm_users` is zero. +impl Mm { + /// Call `mmgrab` on `current.mm`. + #[inline] + pub fn mmgrab_current() -> Option<ARef<Mm>> { + // SAFETY: It's safe to get the `mm` field from current. + let mm = unsafe { + let current = bindings::get_current(); + (*current).mm + }; + + if mm.is_null() { + return None; + } + + // SAFETY: The value of `current->mm` is guaranteed to be null or a valid `mm_struct`. We + // just checked that it's not null. Furthermore, the returned `&Mm` is valid only for the + // duration of this function, and `current->mm` will stay valid for that long. + let mm = unsafe { Mm::from_raw(mm) }; + + // This increments the refcount using `mmgrab`. + Some(ARef::from(mm)) + } + + /// Returns a raw pointer to the inner `mm_struct`. + #[inline] + pub fn as_raw(&self) -> *mut bindings::mm_struct { + self.mm.get() + } + + /// Obtain a reference from a raw pointer. + /// + /// # Safety + /// + /// The caller must ensure that `ptr` points at an `mm_struct`, and that it is not deallocated + /// during the lifetime 'a. + #[inline] + pub unsafe fn from_raw<'a>(ptr: *const bindings::mm_struct) -> &'a Mm { + // SAFETY: Caller promises that the pointer is valid for 'a. Layouts are compatible due to + // repr(transparent). + unsafe { &*ptr.cast() } + } + + /// Calls `mmget_not_zero` and returns a handle if it succeeds. + #[inline] + pub fn mmget_not_zero(&self) -> Option<ARef<MmWithUser>> { + // SAFETY: The pointer is valid since self is a reference. + let success = unsafe { bindings::mmget_not_zero(self.as_raw()) }; + + if success { + // SAFETY: We just created an `mmget` refcount. + Some(unsafe { ARef::from_raw(NonNull::new_unchecked(self.as_raw().cast())) }) + } else { + None + } + } +} + +// These methods require `mm_users` to be non-zero. +impl MmWithUser { + /// Obtain a reference from a raw pointer. + /// + /// # Safety + /// + /// The caller must ensure that `ptr` points at an `mm_struct`, and that `mm_users` remains + /// non-zero for the duration of the lifetime 'a. + #[inline] + pub unsafe fn from_raw<'a>(ptr: *const bindings::mm_struct) -> &'a MmWithUser { + // SAFETY: Caller promises that the pointer is valid for 'a. The layout is compatible due + // to repr(transparent). + unsafe { &*ptr.cast() } + } + + /// Lock the mmap read lock. + #[inline] + pub fn mmap_read_lock(&self) -> MmapReadGuard<'_> { + // SAFETY: The pointer is valid since self is a reference. + unsafe { bindings::mmap_read_lock(self.as_raw()) }; + + // INVARIANT: We just acquired the read lock. + MmapReadGuard { + mm: self, + _nts: NotThreadSafe, + } + } + + /// Try to lock the mmap read lock. + #[inline] + pub fn mmap_read_trylock(&self) -> Option<MmapReadGuard<'_>> { + // SAFETY: The pointer is valid since self is a reference. + let success = unsafe { bindings::mmap_read_trylock(self.as_raw()) }; + + if success { + // INVARIANT: We just acquired the read lock. + Some(MmapReadGuard { + mm: self, + _nts: NotThreadSafe, + }) + } else { + None + } + } +} + +/// A guard for the mmap read lock. +/// +/// # Invariants +/// +/// This `MmapReadGuard` guard owns the mmap read lock. +pub struct MmapReadGuard<'a> { + mm: &'a MmWithUser, + // `mmap_read_lock` and `mmap_read_unlock` must be called on the same thread + _nts: NotThreadSafe, +} + +impl Drop for MmapReadGuard<'_> { + #[inline] + fn drop(&mut self) { + // SAFETY: We hold the read lock by the type invariants. + unsafe { bindings::mmap_read_unlock(self.mm.as_raw()) }; + } +} -- 2.47.0.338.g60cca15819-goog This adds a type called VmAreaRef which is used when referencing a vma that you have read access to. Here, read access means that you hold either the mmap read lock or the vma read lock (or stronger). Additionally, a vma_lookup method is added to the mmap read guard, which enables you to obtain a &VmAreaRef in safe Rust code. This patch only provides a way to lock the mmap read lock, but a follow-up patch also provides a way to just lock the vma read lock. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Reviewed-by: Jann Horn <jannh@google.com> Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/helpers/mm.c | 6 ++ rust/kernel/mm.rs | 21 ++++++ rust/kernel/mm/virt.rs | 182 +++++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 209 insertions(+) diff --git a/rust/helpers/mm.c b/rust/helpers/mm.c index XXXXXXX..XXXXXXX 100644 --- a/rust/helpers/mm.c +++ b/rust/helpers/mm.c @@ -XXX,XX +XXX,XX @@ void rust_helper_mmap_read_unlock(struct mm_struct *mm) { mmap_read_unlock(mm); } + +struct vm_area_struct *rust_helper_vma_lookup(struct mm_struct *mm, + unsigned long addr) +{ + return vma_lookup(mm, addr); +} diff --git a/rust/kernel/mm.rs b/rust/kernel/mm.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm.rs +++ b/rust/kernel/mm.rs @@ -XXX,XX +XXX,XX @@ }; use core::{ops::Deref, ptr::NonNull}; +pub mod virt; + /// A wrapper for the kernel's `struct mm_struct`. /// /// Since `mm_users` may be zero, the associated address space may not exist anymore. You can use @@ -XXX,XX +XXX,XX @@ pub struct MmapReadGuard<'a> { _nts: NotThreadSafe, } +impl<'a> MmapReadGuard<'a> { + /// Look up a vma at the given address. + #[inline] + pub fn vma_lookup(&self, vma_addr: usize) -> Option<&virt::VmAreaRef> { + // SAFETY: We hold a reference to the mm, so the pointer must be valid. Any value is okay + // for `vma_addr`. + let vma = unsafe { bindings::vma_lookup(self.mm.as_raw(), vma_addr as _) }; + + if vma.is_null() { + None + } else { + // SAFETY: We just checked that a vma was found, so the pointer is valid. Furthermore, + // the returned area will borrow from this read lock guard, so it can only be used + // while the mmap read lock is still held. + unsafe { Some(virt::VmAreaRef::from_raw(vma)) } + } + } +} + impl Drop for MmapReadGuard<'_> { #[inline] fn drop(&mut self) { diff --git a/rust/kernel/mm/virt.rs b/rust/kernel/mm/virt.rs new file mode 100644 index XXXXXXX..XXXXXXX --- /dev/null +++ b/rust/kernel/mm/virt.rs @@ -XXX,XX +XXX,XX @@ +// SPDX-License-Identifier: GPL-2.0 + +// Copyright (C) 2024 Google LLC. + +//! Virtual memory. + +use crate::{bindings, types::Opaque}; + +/// A wrapper for the kernel's `struct vm_area_struct` with read access. +/// +/// It represents an area of virtual memory. +/// +/// # Invariants +/// +/// The caller must hold the mmap read lock or the vma read lock. +#[repr(transparent)] +pub struct VmAreaRef { + vma: Opaque<bindings::vm_area_struct>, +} + +// Methods you can call when holding the mmap or vma read lock (or strong). They must be usable no +// matter what the vma flags are. +impl VmAreaRef { + /// Access a virtual memory area given a raw pointer. + /// + /// # Safety + /// + /// Callers must ensure that `vma` is valid for the duration of 'a, and that the mmap or vma + /// read lock (or stronger) is held for at least the duration of 'a. + #[inline] + pub unsafe fn from_raw<'a>(vma: *const bindings::vm_area_struct) -> &'a Self { + // SAFETY: The caller ensures that the invariants are satisfied for the duration of 'a. + unsafe { &*vma.cast() } + } + + /// Returns a raw pointer to this area. + #[inline] + pub fn as_ptr(&self) -> *mut bindings::vm_area_struct { + self.vma.get() + } + + /// Returns the flags associated with the virtual memory area. + /// + /// The possible flags are a combination of the constants in [`flags`]. + #[inline] + pub fn flags(&self) -> vm_flags_t { + // SAFETY: By the type invariants, the caller holds at least the mmap read lock, so this + // access is not a data race. + unsafe { (*self.as_ptr()).__bindgen_anon_2.vm_flags as _ } + } + + /// Returns the (inclusive) start address of the virtual memory area. + #[inline] + pub fn start(&self) -> usize { + // SAFETY: By the type invariants, the caller holds at least the mmap read lock, so this + // access is not a data race. + unsafe { (*self.as_ptr()).__bindgen_anon_1.__bindgen_anon_1.vm_start as _ } + } + + /// Returns the (exclusive) end address of the virtual memory area. + #[inline] + pub fn end(&self) -> usize { + // SAFETY: By the type invariants, the caller holds at least the mmap read lock, so this + // access is not a data race. + unsafe { (*self.as_ptr()).__bindgen_anon_1.__bindgen_anon_1.vm_end as _ } + } + + /// Zap pages in the given page range. + /// + /// This clears page table mappings for the range at the leaf level, leaving all other page + /// tables intact, and freeing any memory referenced by the VMA in this range. That is, + /// anonymous memory is completely freed, file-backed memory has its reference count on page + /// cache folio's dropped, any dirty data will still be written back to disk as usual. + #[inline] + pub fn zap_page_range_single(&self, address: usize, size: usize) { + let (end, did_overflow) = address.overflowing_add(size); + if did_overflow || address < self.start() || self.end() < end { + // TODO: call WARN_ONCE once Rust version of it is added + return; + } + + // SAFETY: By the type invariants, the caller has read access to this VMA, which is + // sufficient for this method call. This method has no requirements on the vma flags. The + // address range is checked to be within the vma. + unsafe { + bindings::zap_page_range_single( + self.as_ptr(), + address as _, + size as _, + core::ptr::null_mut(), + ) + }; + } +} + +/// The integer type used for vma flags. +#[doc(inline)] +pub use bindings::vm_flags_t; + +/// All possible flags for [`VmAreaRef`]. +pub mod flags { + use super::vm_flags_t; + use crate::bindings; + + /// No flags are set. + pub const NONE: vm_flags_t = bindings::VM_NONE as _; + + /// Mapping allows reads. + pub const READ: vm_flags_t = bindings::VM_READ as _; + + /// Mapping allows writes. + pub const WRITE: vm_flags_t = bindings::VM_WRITE as _; + + /// Mapping allows execution. + pub const EXEC: vm_flags_t = bindings::VM_EXEC as _; + + /// Mapping is shared. + pub const SHARED: vm_flags_t = bindings::VM_SHARED as _; + + /// Mapping may be updated to allow reads. + pub const MAYREAD: vm_flags_t = bindings::VM_MAYREAD as _; + + /// Mapping may be updated to allow writes. + pub const MAYWRITE: vm_flags_t = bindings::VM_MAYWRITE as _; + + /// Mapping may be updated to allow execution. + pub const MAYEXEC: vm_flags_t = bindings::VM_MAYEXEC as _; + + /// Mapping may be updated to be shared. + pub const MAYSHARE: vm_flags_t = bindings::VM_MAYSHARE as _; + + /// Page-ranges managed without `struct page`, just pure PFN. + pub const PFNMAP: vm_flags_t = bindings::VM_PFNMAP as _; + + /// Memory mapped I/O or similar. + pub const IO: vm_flags_t = bindings::VM_IO as _; + + /// Do not copy this vma on fork. + pub const DONTCOPY: vm_flags_t = bindings::VM_DONTCOPY as _; + + /// Cannot expand with mremap(). + pub const DONTEXPAND: vm_flags_t = bindings::VM_DONTEXPAND as _; + + /// Lock the pages covered when they are faulted in. + pub const LOCKONFAULT: vm_flags_t = bindings::VM_LOCKONFAULT as _; + + /// Is a VM accounted object. + pub const ACCOUNT: vm_flags_t = bindings::VM_ACCOUNT as _; + + /// Should the VM suppress accounting. + pub const NORESERVE: vm_flags_t = bindings::VM_NORESERVE as _; + + /// Huge TLB Page VM. + pub const HUGETLB: vm_flags_t = bindings::VM_HUGETLB as _; + + /// Synchronous page faults. (DAX-specific) + pub const SYNC: vm_flags_t = bindings::VM_SYNC as _; + + /// Architecture-specific flag. + pub const ARCH_1: vm_flags_t = bindings::VM_ARCH_1 as _; + + /// Wipe VMA contents in child on fork. + pub const WIPEONFORK: vm_flags_t = bindings::VM_WIPEONFORK as _; + + /// Do not include in the core dump. + pub const DONTDUMP: vm_flags_t = bindings::VM_DONTDUMP as _; + + /// Not soft dirty clean area. + pub const SOFTDIRTY: vm_flags_t = bindings::VM_SOFTDIRTY as _; + + /// Can contain `struct page` and pure PFN pages. + pub const MIXEDMAP: vm_flags_t = bindings::VM_MIXEDMAP as _; + + /// MADV_HUGEPAGE marked this vma. + pub const HUGEPAGE: vm_flags_t = bindings::VM_HUGEPAGE as _; + + /// MADV_NOHUGEPAGE marked this vma. + pub const NOHUGEPAGE: vm_flags_t = bindings::VM_NOHUGEPAGE as _; + + /// KSM may merge identical pages. + pub const MERGEABLE: vm_flags_t = bindings::VM_MERGEABLE as _; +} -- 2.47.0.338.g60cca15819-goog The vm_insert_page method is only usable on vmas with the VM_MIXEDMAP flag, so we introduce a new type to keep track of such vmas. The approach used in this patch assumes that we will not need to encode many flag combinations in the type. I don't think we need to encode more than VM_MIXEDMAP and VM_PFNMAP as things are now. However, if that becomes necessary, using generic parameters in a single type would scale better as the number of flags increases. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/kernel/mm/virt.rs | 70 +++++++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 69 insertions(+), 1 deletion(-) diff --git a/rust/kernel/mm/virt.rs b/rust/kernel/mm/virt.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm/virt.rs +++ b/rust/kernel/mm/virt.rs @@ -XXX,XX +XXX,XX @@ //! Virtual memory. -use crate::{bindings, types::Opaque}; +use crate::{ + bindings, + error::{to_result, Result}, + page::Page, + types::Opaque, +}; + +use core::ops::Deref; /// A wrapper for the kernel's `struct vm_area_struct` with read access. /// @@ -XXX,XX +XXX,XX @@ pub fn zap_page_range_single(&self, address: usize, size: usize) { ) }; } + + /// Check whether the `VM_MIXEDMAP` flag is set. + /// + /// This can be used to access methods that require `VM_MIXEDMAP` to be set. + #[inline] + pub fn as_mixedmap_vma(&self) -> Option<&VmAreaMixedMap> { + if self.flags() & flags::MIXEDMAP != 0 { + // SAFETY: We just checked that `VM_MIXEDMAP` is set. All other requirements are + // satisfied by the type invariants of `VmAreaRef`. + Some(unsafe { VmAreaMixedMap::from_raw(self.as_ptr()) }) + } else { + None + } + } +} + +/// A wrapper for the kernel's `struct vm_area_struct` with read access and `VM_MIXEDMAP` set. +/// +/// It represents an area of virtual memory. +/// +/// # Invariants +/// +/// The caller must hold the mmap read lock or the vma read lock. The `VM_MIXEDMAP` flag must be +/// set. +#[repr(transparent)] +pub struct VmAreaMixedMap { + vma: VmAreaRef, +} + +// Make all `VmAreaRef` methods available on `VmAreaMixedMap`. +impl Deref for VmAreaMixedMap { + type Target = VmAreaRef; + + #[inline] + fn deref(&self) -> &VmAreaRef { + &self.vma + } +} + +impl VmAreaMixedMap { + /// Access a virtual memory area given a raw pointer. + /// + /// # Safety + /// + /// Callers must ensure that `vma` is valid for the duration of 'a, and that the mmap read lock + /// (or stronger) is held for at least the duration of 'a. The `VM_MIXEDMAP` flag must be set. + #[inline] + pub unsafe fn from_raw<'a>(vma: *const bindings::vm_area_struct) -> &'a Self { + // SAFETY: The caller ensures that the invariants are satisfied for the duration of 'a. + unsafe { &*vma.cast() } + } + + /// Maps a single page at the given address within the virtual memory area. + /// + /// This operation does not take ownership of the page. + #[inline] + pub fn vm_insert_page(&self, address: usize, page: &Page) -> Result { + // SAFETY: The caller has read access and has verified that `VM_MIXEDMAP` is set. The page + // is order 0. The address is checked on the C side so it can take any value. + to_result(unsafe { bindings::vm_insert_page(self.as_ptr(), address as _, page.as_ptr()) }) + } } /// The integer type used for vma flags. -- 2.47.0.338.g60cca15819-goog Currently, the binder driver always uses the mmap lock to make changes to its vma. Because the mmap lock is global to the process, this can involve significant contention. However, the kernel has a feature called per-vma locks, which can significantly reduce contention. For example, you can take a vma lock in parallel with an mmap write lock. This is important because contention on the mmap lock has been a long-term recurring challenge for the Binder driver. This patch introduces support for using `lock_vma_under_rcu` from Rust. The Rust Binder driver will be able to use this to reduce contention on the mmap lock. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Reviewed-by: Jann Horn <jannh@google.com> Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/helpers/mm.c | 5 +++++ rust/kernel/mm.rs | 56 +++++++++++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 61 insertions(+) diff --git a/rust/helpers/mm.c b/rust/helpers/mm.c index XXXXXXX..XXXXXXX 100644 --- a/rust/helpers/mm.c +++ b/rust/helpers/mm.c @@ -XXX,XX +XXX,XX @@ struct vm_area_struct *rust_helper_vma_lookup(struct mm_struct *mm, { return vma_lookup(mm, addr); } + +void rust_helper_vma_end_read(struct vm_area_struct *vma) +{ + vma_end_read(vma); +} diff --git a/rust/kernel/mm.rs b/rust/kernel/mm.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm.rs +++ b/rust/kernel/mm.rs @@ -XXX,XX +XXX,XX @@ use core::{ops::Deref, ptr::NonNull}; pub mod virt; +use virt::VmAreaRef; /// A wrapper for the kernel's `struct mm_struct`. /// @@ -XXX,XX +XXX,XX @@ pub unsafe fn from_raw<'a>(ptr: *const bindings::mm_struct) -> &'a MmWithUser { unsafe { &*ptr.cast() } } + /// Attempt to access a vma using the vma read lock. + /// + /// This is an optimistic trylock operation, so it may fail if there is contention. In that + /// case, you should fall back to taking the mmap read lock. + /// + /// When per-vma locks are disabled, this always returns `None`. + #[inline] + pub fn lock_vma_under_rcu(&self, vma_addr: usize) -> Option<VmaReadGuard<'_>> { + #[cfg(CONFIG_PER_VMA_LOCK)] + { + // SAFETY: Calling `bindings::lock_vma_under_rcu` is always okay given an mm where + // `mm_users` is non-zero. + let vma = unsafe { bindings::lock_vma_under_rcu(self.as_raw(), vma_addr as _) }; + if !vma.is_null() { + return Some(VmaReadGuard { + // SAFETY: If `lock_vma_under_rcu` returns a non-null ptr, then it points at a + // valid vma. The vma is stable for as long as the vma read lock is held. + vma: unsafe { VmAreaRef::from_raw(vma) }, + _nts: NotThreadSafe, + }); + } + } + + None + } + /// Lock the mmap read lock. #[inline] pub fn mmap_read_lock(&self) -> MmapReadGuard<'_> { @@ -XXX,XX +XXX,XX @@ fn drop(&mut self) { unsafe { bindings::mmap_read_unlock(self.mm.as_raw()) }; } } + +/// A guard for the vma read lock. +/// +/// # Invariants +/// +/// This `VmaReadGuard` guard owns the vma read lock. +pub struct VmaReadGuard<'a> { + vma: &'a VmAreaRef, + // `vma_end_read` must be called on the same thread as where the lock was taken + _nts: NotThreadSafe, +} + +// Make all `VmAreaRef` methods available on `VmaReadGuard`. +impl Deref for VmaReadGuard<'_> { + type Target = VmAreaRef; + + #[inline] + fn deref(&self) -> &VmAreaRef { + self.vma + } +} + +impl Drop for VmaReadGuard<'_> { + #[inline] + fn drop(&mut self) { + // SAFETY: We hold the read lock by the type invariants. + unsafe { bindings::vma_end_read(self.vma.as_ptr()) }; + } +} -- 2.47.0.338.g60cca15819-goog Adds an MmWithUserAsync type that uses mmput_async when dropped but is otherwise identical to MmWithUser. This has to be done using a separate type because the thing we are changing is the destructor. Rust Binder needs this to avoid a certain deadlock. See commit 9a9ab0d96362 ("binder: fix race between mmput() and do_exit()") for details. It's also needed in the shrinker to avoid cleaning up the mm in the shrinker's context. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/kernel/mm.rs | 49 +++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 49 insertions(+) diff --git a/rust/kernel/mm.rs b/rust/kernel/mm.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm.rs +++ b/rust/kernel/mm.rs @@ -XXX,XX +XXX,XX @@ fn deref(&self) -> &Mm { } } +/// A wrapper for the kernel's `struct mm_struct`. +/// +/// This type is identical to `MmWithUser` except that it uses `mmput_async` when dropping a +/// refcount. This means that the destructor of `ARef<MmWithUserAsync>` is safe to call in atomic +/// context. +/// +/// # Invariants +/// +/// Values of this type are always refcounted using `mmget`. The value of `mm_users` is non-zero. +#[repr(transparent)] +pub struct MmWithUserAsync { + mm: MmWithUser, +} + +// SAFETY: It is safe to call `mmput_async` on another thread than where `mmget` was called. +unsafe impl Send for MmWithUserAsync {} +// SAFETY: All methods on `MmWithUserAsync` can be called in parallel from several threads. +unsafe impl Sync for MmWithUserAsync {} + +// SAFETY: By the type invariants, this type is always refcounted. +unsafe impl AlwaysRefCounted for MmWithUserAsync { + fn inc_ref(&self) { + // SAFETY: The pointer is valid since self is a reference. + unsafe { bindings::mmget(self.as_raw()) }; + } + + unsafe fn dec_ref(obj: NonNull<Self>) { + // SAFETY: The caller is giving up their refcount. + unsafe { bindings::mmput_async(obj.cast().as_ptr()) }; + } +} + +// Make all `MmWithUser` methods available on `MmWithUserAsync`. +impl Deref for MmWithUserAsync { + type Target = MmWithUser; + + #[inline] + fn deref(&self) -> &MmWithUser { + &self.mm + } +} + // These methods are safe to call even if `mm_users` is zero. impl Mm { /// Call `mmgrab` on `current.mm`. @@ -XXX,XX +XXX,XX @@ pub unsafe fn from_raw<'a>(ptr: *const bindings::mm_struct) -> &'a MmWithUser { unsafe { &*ptr.cast() } } + /// Use `mmput_async` when dropping this refcount. + #[inline] + pub fn into_mmput_async(me: ARef<MmWithUser>) -> ARef<MmWithUserAsync> { + // SAFETY: The layouts and invariants are compatible. + unsafe { ARef::from_raw(ARef::into_raw(me).cast()) } + } + /// Attempt to access a vma using the vma read lock. /// /// This is an optimistic trylock operation, so it may fail if there is contention. In that -- 2.47.0.338.g60cca15819-goog This type will be used when setting up a new vma in an f_ops->mmap() hook. Using a separate type from VmAreaRef allows us to have a separate set of operations that you are only able to use during the mmap() hook. For example, the VM_MIXEDMAP flag must not be changed after the initial setup that happens during the f_ops->mmap() hook. To avoid setting invalid flag values, the methods for clearing VM_MAYWRITE and similar involve a check of VM_WRITE, and return an error if VM_WRITE is set. Trying to use `try_clear_maywrite` without checking the return value results in a compilation error because the `Result` type is marked #[must_use]. For now, there's only a method for VM_MIXEDMAP and not VM_PFNMAP. When we add a VM_PFNMAP method, we will need some way to prevent you from setting both VM_MIXEDMAP and VM_PFNMAP on the same vma. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Reviewed-by: Jann Horn <jannh@google.com> Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/kernel/mm/virt.rs | 181 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 180 insertions(+), 1 deletion(-) diff --git a/rust/kernel/mm/virt.rs b/rust/kernel/mm/virt.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm/virt.rs +++ b/rust/kernel/mm/virt.rs @@ -XXX,XX +XXX,XX @@ use crate::{ bindings, - error::{to_result, Result}, + error::{code::EINVAL, to_result, Result}, page::Page, types::Opaque, }; @@ -XXX,XX +XXX,XX @@ pub fn vm_insert_page(&self, address: usize, page: &Page) -> Result { } } +/// A builder for setting up a vma in an `f_ops->mmap()` hook. +/// +/// # Invariants +/// +/// For the duration of 'a, the referenced vma must be undergoing initialization in an +/// `f_ops->mmap()` hook. +pub struct VmAreaNew { + vma: VmAreaRef, +} + +// Make all `VmAreaRef` methods available on `VmAreaNew`. +impl Deref for VmAreaNew { + type Target = VmAreaRef; + + #[inline] + fn deref(&self) -> &VmAreaRef { + &self.vma + } +} + +impl VmAreaNew { + /// Access a virtual memory area given a raw pointer. + /// + /// # Safety + /// + /// Callers must ensure that `vma` is undergoing initial vma setup for the duration of 'a. + #[inline] + pub unsafe fn from_raw<'a>(vma: *const bindings::vm_area_struct) -> &'a Self { + // SAFETY: The caller ensures that the invariants are satisfied for the duration of 'a. + unsafe { &*vma.cast() } + } + + /// Internal method for updating the vma flags. + /// + /// # Safety + /// + /// This must not be used to set the flags to an invalid value. + #[inline] + unsafe fn update_flags(&self, set: vm_flags_t, unset: vm_flags_t) { + let mut flags = self.flags(); + flags |= set; + flags &= !unset; + + // SAFETY: This is not a data race: the vma is undergoing initial setup, so it's not yet + // shared. Additionally, `VmAreaNew` is `!Sync`, so it cannot be used to write in parallel. + // The caller promises that this does not set the flags to an invalid value. + unsafe { (*self.as_ptr()).__bindgen_anon_2.__vm_flags = flags }; + } + + /// Set the `VM_MIXEDMAP` flag on this vma. + /// + /// This enables the vma to contain both `struct page` and pure PFN pages. Returns a reference + /// that can be used to call `vm_insert_page` on the vma. + #[inline] + pub fn set_mixedmap(&self) -> &VmAreaMixedMap { + // SAFETY: We don't yet provide a way to set VM_PFNMAP, so this cannot put the flags in an + // invalid state. + unsafe { self.update_flags(flags::MIXEDMAP, 0) }; + + // SAFETY: We just set `VM_MIXEDMAP` on the vma. + unsafe { VmAreaMixedMap::from_raw(self.vma.as_ptr()) } + } + + /// Set the `VM_IO` flag on this vma. + /// + /// This is used for memory mapped IO and similar. The flag tells other parts of the kernel to + /// avoid looking at the pages. For memory mapped IO this is useful as accesses to the pages + /// could have side effects. + #[inline] + pub fn set_io(&self) { + // SAFETY: Setting the VM_IO flag is always okay. + unsafe { self.update_flags(flags::IO, 0) }; + } + + /// Set the `VM_DONTEXPAND` flag on this vma. + /// + /// This prevents the vma from being expanded with `mremap()`. + #[inline] + pub fn set_dontexpand(&self) { + // SAFETY: Setting the VM_DONTEXPAND flag is always okay. + unsafe { self.update_flags(flags::DONTEXPAND, 0) }; + } + + /// Set the `VM_DONTCOPY` flag on this vma. + /// + /// This prevents the vma from being copied on fork. This option is only permanent if `VM_IO` + /// is set. + #[inline] + pub fn set_dontcopy(&self) { + // SAFETY: Setting the VM_DONTCOPY flag is always okay. + unsafe { self.update_flags(flags::DONTCOPY, 0) }; + } + + /// Set the `VM_DONTDUMP` flag on this vma. + /// + /// This prevents the vma from being included in core dumps. This option is only permanent if + /// `VM_IO` is set. + #[inline] + pub fn set_dontdump(&self) { + // SAFETY: Setting the VM_DONTDUMP flag is always okay. + unsafe { self.update_flags(flags::DONTDUMP, 0) }; + } + + /// Returns whether `VM_READ` is set. + /// + /// This flag indicates whether userspace is mapping this vma as readable. + #[inline] + pub fn get_read(&self) -> bool { + (self.flags() & flags::READ) != 0 + } + + /// Try to clear the `VM_MAYREAD` flag, failing if `VM_READ` is set. + /// + /// This flag indicates whether userspace is allowed to make this vma readable with + /// `mprotect()`. + /// + /// Note that this operation is irreversible. Once `VM_MAYREAD` has been cleared, it can never + /// be set again. + #[inline] + pub fn try_clear_mayread(&self) -> Result { + if self.get_read() { + return Err(EINVAL); + } + // SAFETY: Clearing `VM_MAYREAD` is okay when `VM_READ` is not set. + unsafe { self.update_flags(0, flags::MAYREAD) }; + Ok(()) + } + + /// Returns whether `VM_WRITE` is set. + /// + /// This flag indicates whether userspace is mapping this vma as writable. + #[inline] + pub fn get_write(&self) -> bool { + (self.flags() & flags::WRITE) != 0 + } + + /// Try to clear the `VM_MAYWRITE` flag, failing if `VM_WRITE` is set. + /// + /// This flag indicates whether userspace is allowed to make this vma writable with + /// `mprotect()`. + /// + /// Note that this operation is irreversible. Once `VM_MAYWRITE` has been cleared, it can never + /// be set again. + #[inline] + pub fn try_clear_maywrite(&self) -> Result { + if self.get_write() { + return Err(EINVAL); + } + // SAFETY: Clearing `VM_MAYWRITE` is okay when `VM_WRITE` is not set. + unsafe { self.update_flags(0, flags::MAYWRITE) }; + Ok(()) + } + + /// Returns whether `VM_EXEC` is set. + /// + /// This flag indicates whether userspace is mapping this vma as executable. + #[inline] + pub fn get_exec(&self) -> bool { + (self.flags() & flags::EXEC) != 0 + } + + /// Try to clear the `VM_MAYEXEC` flag, failing if `VM_EXEC` is set. + /// + /// This flag indicates whether userspace is allowed to make this vma executable with + /// `mprotect()`. + /// + /// Note that this operation is irreversible. Once `VM_MAYEXEC` has been cleared, it can never + /// be set again. + #[inline] + pub fn try_clear_mayexec(&self) -> Result { + if self.get_exec() { + return Err(EINVAL); + } + // SAFETY: Clearing `VM_MAYEXEC` is okay when `VM_EXEC` is not set. + unsafe { self.update_flags(0, flags::MAYEXEC) }; + Ok(()) + } +} + /// The integer type used for vma flags. #[doc(inline)] pub use bindings::vm_flags_t; -- 2.47.0.338.g60cca15819-goog Add the ability to write a file_operations->mmap hook in Rust when using the miscdevice abstraction. The `vma` argument to the `mmap` hook uses the `VmAreaNew` type from the previous commit; this type provides the correct set of operations for a file_operations->mmap hook. Acked-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com> (for mm bits) Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/kernel/miscdevice.rs | 28 ++++++++++++++++++++++++++++ 1 file changed, 28 insertions(+) diff --git a/rust/kernel/miscdevice.rs b/rust/kernel/miscdevice.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/miscdevice.rs +++ b/rust/kernel/miscdevice.rs @@ -XXX,XX +XXX,XX @@ use crate::{ bindings, error::{to_result, Error, Result, VTABLE_DEFAULT_ERROR}, + mm::virt::VmAreaNew, prelude::*, str::CStr, types::{ForeignOwnable, Opaque}, @@ -XXX,XX +XXX,XX @@ fn release(device: Self::Ptr) { drop(device); } + /// Handle for mmap. + fn mmap(_device: <Self::Ptr as ForeignOwnable>::Borrowed<'_>, _vma: &VmAreaNew) -> Result { + kernel::build_error!(VTABLE_DEFAULT_ERROR) + } + /// Handler for ioctls. /// /// The `cmd` argument is usually manipulated using the utilties in [`kernel::ioctl`]. @@ -XXX,XX +XXX,XX @@ impl<T: MiscDevice> VtableHelper<T> { const VTABLE: bindings::file_operations = bindings::file_operations { open: Some(fops_open::<T>), release: Some(fops_release::<T>), + mmap: maybe_fn(T::HAS_MMAP, fops_mmap::<T>), unlocked_ioctl: maybe_fn(T::HAS_IOCTL, fops_ioctl::<T>), #[cfg(CONFIG_COMPAT)] compat_ioctl: if T::HAS_COMPAT_IOCTL { @@ -XXX,XX +XXX,XX @@ impl<T: MiscDevice> VtableHelper<T> { 0 } +/// # Safety +/// +/// `file` must be a valid file that is associated with a `MiscDeviceRegistration<T>`. +/// `vma` must be a vma that is currently being mmap'ed with this file. +unsafe extern "C" fn fops_mmap<T: MiscDevice>( + file: *mut bindings::file, + vma: *mut bindings::vm_area_struct, +) -> c_int { + // SAFETY: The mmap call of a file can access the private data. + let private = unsafe { (*file).private_data }; + // SAFETY: Mmap calls can borrow the private data of the file. + let device = unsafe { <T::Ptr as ForeignOwnable>::borrow(private) }; + // SAFETY: The caller provides a vma that is undergoing initial VMA setup. + let area = unsafe { VmAreaNew::from_raw(vma) }; + + match T::mmap(device, area) { + Ok(()) => 0, + Err(err) => err.to_errno() as c_int, + } +} + /// # Safety /// /// `file` must be a valid file that is associated with a `MiscDeviceRegistration<T>`. -- 2.47.0.338.g60cca15819-goog Introduce a new type called `CurrentTask` that lets you perform various operations that are only safe on the `current` task. Use the new type to provide a way to access the current mm without incrementing its refcount. With this change, you can write stuff such as let vma = current!().mm().lock_vma_under_rcu(addr); without incrementing any refcounts. This replaces the existing abstractions for accessing the current pid namespace. With the old approach, every field access to current involves both a macro and a unsafe helper function. The new approach simplifies that to a single safe function on the `CurrentTask` type. This makes it less heavy-weight to add additional current accessors in the future. That said, creating a `CurrentTask` type like the one in this patch requires that we are careful to ensure that it cannot escape the current task or otherwise access things after they are freed. To do this, I declared that it cannot escape the current "task context" where I defined a "task context" as essentially the region in which `current` remains unchanged. So e.g., release_task() or begin_new_exec() would leave the task context. If a userspace thread returns to userspace and later makes another syscall, then I consider the two syscalls to be different task contexts. This allows values stored in that task to be modified between syscalls, even if they're guaranteed to be immutable during a syscall. Ensuring correctness of `CurrentTask` is slightly tricky if we also want the ability to have a safe `kthread_use_mm()` implementation in Rust. To support that safely, there are two patterns we need to ensure are safe: // Case 1: current!() called inside the scope. let mm; kthread_use_mm(some_mm, || { mm = current!().mm(); }); drop(some_mm); mm.do_something(); // UAF and: // Case 2: current!() called before the scope. let mm; let task = current!(); kthread_use_mm(some_mm, || { mm = task.mm(); }); drop(some_mm); mm.do_something(); // UAF The existing `current!()` abstraction already natively prevents the first case: The `&CurrentTask` would be tied to the inner scope, so the borrow-checker ensures that no reference derived from it can escape the scope. Fixing the second case is a bit more tricky. The solution is to essentially pretend that the contents of the scope execute on an different thread, which means that only thread-safe types can cross the boundary. Since `CurrentTask` is marked `NotThreadSafe`, attempts to move it to another thread will fail, and this includes our fake pretend thread boundary. This has the disadvantage that other types that aren't thread-safe for reasons unrelated to `current` also cannot be moved across the `kthread_use_mm()` boundary. I consider this an acceptable tradeoff. Cc: Christian Brauner <brauner@kernel.org> Signed-off-by: Alice Ryhl <aliceryhl@google.com> --- rust/kernel/mm.rs | 22 ---- rust/kernel/task.rs | 284 ++++++++++++++++++++++++++++++---------------------- 2 files changed, 167 insertions(+), 139 deletions(-) diff --git a/rust/kernel/mm.rs b/rust/kernel/mm.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/mm.rs +++ b/rust/kernel/mm.rs @@ -XXX,XX +XXX,XX @@ fn deref(&self) -> &MmWithUser { // These methods are safe to call even if `mm_users` is zero. impl Mm { - /// Call `mmgrab` on `current.mm`. - #[inline] - pub fn mmgrab_current() -> Option<ARef<Mm>> { - // SAFETY: It's safe to get the `mm` field from current. - let mm = unsafe { - let current = bindings::get_current(); - (*current).mm - }; - - if mm.is_null() { - return None; - } - - // SAFETY: The value of `current->mm` is guaranteed to be null or a valid `mm_struct`. We - // just checked that it's not null. Furthermore, the returned `&Mm` is valid only for the - // duration of this function, and `current->mm` will stay valid for that long. - let mm = unsafe { Mm::from_raw(mm) }; - - // This increments the refcount using `mmgrab`. - Some(ARef::from(mm)) - } - /// Returns a raw pointer to the inner `mm_struct`. #[inline] pub fn as_raw(&self) -> *mut bindings::mm_struct { diff --git a/rust/kernel/task.rs b/rust/kernel/task.rs index XXXXXXX..XXXXXXX 100644 --- a/rust/kernel/task.rs +++ b/rust/kernel/task.rs @@ -XXX,XX +XXX,XX @@ use crate::{ bindings, ffi::{c_int, c_long, c_uint}, + mm::MmWithUser, pid_namespace::PidNamespace, types::{ARef, NotThreadSafe, Opaque}, }; @@ -XXX,XX +XXX,XX @@ #[macro_export] macro_rules! current { () => { - // SAFETY: Deref + addr-of below create a temporary `TaskRef` that cannot outlive the - // caller. + // SAFETY: This expression creates a temporary value that is dropped at the end of the + // caller's scope. The following mechanisms ensure that the resulting `&CurrentTask` cannot + // leave current task context: + // + // * To return to userspace, the caller must leave the current scope. + // * Operations such as `begin_new_exec()` are necessarily unsafe and the caller of + // `begin_new_exec()` is responsible for safety. + // * Rust abstractions for things such as a `kthread_use_mm()` scope must require the + // closure to be `Send`, so the `NotThreadSafe` field of `CurrentTask` ensures that the + // `&CurrentTask` cannot cross the scope in either direction. unsafe { &*$crate::task::Task::current() } }; } -/// Returns the currently running task's pid namespace. -#[macro_export] -macro_rules! current_pid_ns { - () => { - // SAFETY: Deref + addr-of below create a temporary `PidNamespaceRef` that cannot outlive - // the caller. - unsafe { &*$crate::task::Task::current_pid_ns() } - }; -} - /// Wraps the kernel's `struct task_struct`. /// /// # Invariants @@ -XXX,XX +XXX,XX @@ unsafe impl Send for Task {} // synchronised by C code (e.g., `signal_pending`). unsafe impl Sync for Task {} +/// Represents the [`Task`] in the `current` global. +/// +/// This type exists to provide more efficient operations that are only valid on the current task. +/// For example, to retrieve the pid-namespace of a task, you must use rcu protection unless it is +/// the current task. +/// +/// # Invariants +/// +/// Each value of this type must only be accessed from the task context it was created within. +/// +/// Of course, every thread is in a different task context, but for the purposes of this invariant, +/// these operations also permanently leave the task context: +/// +/// * Returning to userspace from system call context. +/// * Calling `release_task()`. +/// * Calling `begin_new_exec()` in a binary format loader. +/// +/// Other operations temporarily create a new sub-context: +/// +/// * Calling `kthread_use_mm()` creates a new context, and `kthread_unuse_mm()` returns to the +/// old context. +/// +/// This means that a `CurrentTask` obtained before a `kthread_use_mm()` call may be used again +/// once `kthread_unuse_mm()` is called, but it must not be used between these two calls. +/// Conversely, a `CurrentTask` obtained between a `kthread_use_mm()`/`kthread_unuse_mm()` pair +/// must not be used after `kthread_unuse_mm()`. +#[repr(transparent)] +pub struct CurrentTask(Task, NotThreadSafe); + +// Make all `Task` methods available on `CurrentTask`. +impl Deref for CurrentTask { + type Target = Task; + #[inline] + fn deref(&self) -> &Task { + &self.0 + } +} + /// The type of process identifiers (PIDs). type Pid = bindings::pid_t; @@ -XXX,XX +XXX,XX @@ pub fn current_raw() -> *mut bindings::task_struct { /// /// # Safety /// - /// Callers must ensure that the returned object doesn't outlive the current task/thread. - pub unsafe fn current() -> impl Deref<Target = Task> { - struct TaskRef<'a> { - task: &'a Task, - _not_send: NotThreadSafe, + /// Callers must ensure that the returned object is only used to access a [`CurrentTask`] + /// within the task context that was active when this function was called. For more details, + /// see the invariants section for [`CurrentTask`]. + pub unsafe fn current() -> impl Deref<Target = CurrentTask> { + struct TaskRef { + task: *const CurrentTask, } - impl Deref for TaskRef<'_> { - type Target = Task; + impl Deref for TaskRef { + type Target = CurrentTask; fn deref(&self) -> &Self::Target { - self.task + // SAFETY: The returned reference borrows from this `TaskRef`, so it cannot outlive + // the `TaskRef`, which the caller of `Task::current()` has promised will not + // outlive the task/thread for which `self.task` is the `current` pointer. Thus, it + // is okay to return a `CurrentTask` reference here. + unsafe { &*self.task } } } - let current = Task::current_raw(); TaskRef { - // SAFETY: If the current thread is still running, the current task is valid. Given - // that `TaskRef` is not `Send`, we know it cannot be transferred to another thread - // (where it could potentially outlive the caller). - task: unsafe { &*current.cast() }, - _not_send: NotThreadSafe, - } - } - - /// Returns a PidNamespace reference for the currently executing task's/thread's pid namespace. - /// - /// This function can be used to create an unbounded lifetime by e.g., storing the returned - /// PidNamespace in a global variable which would be a bug. So the recommended way to get the - /// current task's/thread's pid namespace is to use the [`current_pid_ns`] macro because it is - /// safe. - /// - /// # Safety - /// - /// Callers must ensure that the returned object doesn't outlive the current task/thread. - pub unsafe fn current_pid_ns() -> impl Deref<Target = PidNamespace> { - struct PidNamespaceRef<'a> { - task: &'a PidNamespace, - _not_send: NotThreadSafe, - } - - impl Deref for PidNamespaceRef<'_> { - type Target = PidNamespace; - - fn deref(&self) -> &Self::Target { - self.task - } - } - - // The lifetime of `PidNamespace` is bound to `Task` and `struct pid`. - // - // The `PidNamespace` of a `Task` doesn't ever change once the `Task` is alive. A - // `unshare(CLONE_NEWPID)` or `setns(fd_pidns/pidfd, CLONE_NEWPID)` will not have an effect - // on the calling `Task`'s pid namespace. It will only effect the pid namespace of children - // created by the calling `Task`. This invariant guarantees that after having acquired a - // reference to a `Task`'s pid namespace it will remain unchanged. - // - // When a task has exited and been reaped `release_task()` will be called. This will set - // the `PidNamespace` of the task to `NULL`. So retrieving the `PidNamespace` of a task - // that is dead will return `NULL`. Note, that neither holding the RCU lock nor holding a - // referencing count to - // the `Task` will prevent `release_task()` being called. - // - // In order to retrieve the `PidNamespace` of a `Task` the `task_active_pid_ns()` function - // can be used. There are two cases to consider: - // - // (1) retrieving the `PidNamespace` of the `current` task - // (2) retrieving the `PidNamespace` of a non-`current` task - // - // From system call context retrieving the `PidNamespace` for case (1) is always safe and - // requires neither RCU locking nor a reference count to be held. Retrieving the - // `PidNamespace` after `release_task()` for current will return `NULL` but no codepath - // like that is exposed to Rust. - // - // Retrieving the `PidNamespace` from system call context for (2) requires RCU protection. - // Accessing `PidNamespace` outside of RCU protection requires a reference count that - // must've been acquired while holding the RCU lock. Note that accessing a non-`current` - // task means `NULL` can be returned as the non-`current` task could have already passed - // through `release_task()`. - // - // To retrieve (1) the `current_pid_ns!()` macro should be used which ensure that the - // returned `PidNamespace` cannot outlive the calling scope. The associated - // `current_pid_ns()` function should not be called directly as it could be abused to - // created an unbounded lifetime for `PidNamespace`. The `current_pid_ns!()` macro allows - // Rust to handle the common case of accessing `current`'s `PidNamespace` without RCU - // protection and without having to acquire a reference count. - // - // For (2) the `task_get_pid_ns()` method must be used. This will always acquire a - // reference on `PidNamespace` and will return an `Option` to force the caller to - // explicitly handle the case where `PidNamespace` is `None`, something that tends to be - // forgotten when doing the equivalent operation in `C`. Missing RCU primitives make it - // difficult to perform operations that are otherwise safe without holding a reference - // count as long as RCU protection is guaranteed. But it is not important currently. But we - // do want it in the future. - // - // Note for (2) the required RCU protection around calling `task_active_pid_ns()` - // synchronizes against putting the last reference of the associated `struct pid` of - // `task->thread_pid`. The `struct pid` stored in that field is used to retrieve the - // `PidNamespace` of the caller. When `release_task()` is called `task->thread_pid` will be - // `NULL`ed and `put_pid()` on said `struct pid` will be delayed in `free_pid()` via - // `call_rcu()` allowing everyone with an RCU protected access to the `struct pid` acquired - // from `task->thread_pid` to finish. - // - // SAFETY: The current task's pid namespace is valid as long as the current task is running. - let pidns = unsafe { bindings::task_active_pid_ns(Task::current_raw()) }; - PidNamespaceRef { - // SAFETY: If the current thread is still running, the current task and its associated - // pid namespace are valid. `PidNamespaceRef` is not `Send`, so we know it cannot be - // transferred to another thread (where it could potentially outlive the current - // `Task`). The caller needs to ensure that the PidNamespaceRef doesn't outlive the - // current task/thread. - task: unsafe { PidNamespace::from_ptr(pidns) }, - _not_send: NotThreadSafe, + // CAST: The layout of `struct task_struct` and `CurrentTask` is identical. + task: Task::current_raw().cast(), } } @@ -XXX,XX +XXX,XX @@ pub fn wake_up(&self) { } } +impl CurrentTask { + /// Access the address space of the current task. + /// + /// This function does not touch the refcount of the mm. + #[inline] + pub fn mm(&self) -> Option<&MmWithUser> { + // SAFETY: The `mm` field of `current` is not modified from other threads, so reading it is + // not a data race. + let mm = unsafe { (*self.as_ptr()).mm }; + + if mm.is_null() { + return None; + } + + // SAFETY: If `current->mm` is non-null, then it references a valid mm with a non-zero + // value of `mm_users`. Furthermore, the returned `&MmWithUser` borrows from this + // `CurrentTask`, so it cannot escape the scope in which the current pointer was obtained. + // + // This is safe even if `kthread_use_mm()`/`kthread_unuse_mm()` are used. There are two + // relevant cases: + // * If the `&CurrentTask` was created before `kthread_use_mm()`, then it cannot be + // accessed during the `kthread_use_mm()`/`kthread_unuse_mm()` scope due to the + // `NotThreadSafe` field of `CurrentTask`. + // * If the `&CurrentTask` was created within a `kthread_use_mm()`/`kthread_unuse_mm()` + // scope, then the `&CurrentTask` cannot escape that scope, so the returned `&MmWithUser` + // also cannot escape that scope. + // In either case, it's not possible to read `current->mm` and keep using it after the + // scope is ended with `kthread_unuse_mm()`. + Some(unsafe { MmWithUser::from_raw(mm) }) + } + + /// Access the pid namespace of the current task. + /// + /// This function does not touch the refcount of the namespace or use RCU protection. + #[doc(alias = "task_active_pid_ns")] + #[inline] + pub fn active_pid_ns(&self) -> Option<&PidNamespace> { + // SAFETY: It is safe to call `task_active_pid_ns` without RCU protection when calling it + // on the current task. + let active_ns = unsafe { bindings::task_active_pid_ns(self.as_ptr()) }; + + if active_ns.is_null() { + return None; + } + + // The lifetime of `PidNamespace` is bound to `Task` and `struct pid`. + // + // The `PidNamespace` of a `Task` doesn't ever change once the `Task` is alive. A + // `unshare(CLONE_NEWPID)` or `setns(fd_pidns/pidfd, CLONE_NEWPID)` will not have an effect + // on the calling `Task`'s pid namespace. It will only effect the pid namespace of children + // created by the calling `Task`. This invariant guarantees that after having acquired a + // reference to a `Task`'s pid namespace it will remain unchanged. + // + // When a task has exited and been reaped `release_task()` will be called. This will set + // the `PidNamespace` of the task to `NULL`. So retrieving the `PidNamespace` of a task + // that is dead will return `NULL`. Note, that neither holding the RCU lock nor holding a + // referencing count to the `Task` will prevent `release_task()` being called. + // + // In order to retrieve the `PidNamespace` of a `Task` the `task_active_pid_ns()` function + // can be used. There are two cases to consider: + // + // (1) retrieving the `PidNamespace` of the `current` task + // (2) retrieving the `PidNamespace` of a non-`current` task + // + // From system call context retrieving the `PidNamespace` for case (1) is always safe and + // requires neither RCU locking nor a reference count to be held. Retrieving the + // `PidNamespace` after `release_task()` for current will return `NULL` but no codepath + // like that is exposed to Rust. + // + // Retrieving the `PidNamespace` from system call context for (2) requires RCU protection. + // Accessing `PidNamespace` outside of RCU protection requires a reference count that + // must've been acquired while holding the RCU lock. Note that accessing a non-`current` + // task means `NULL` can be returned as the non-`current` task could have already passed + // through `release_task()`. + // + // To retrieve (1) the `&CurrentTask` type should be used which ensures that the returned + // `PidNamespace` cannot outlive the current task context. The `CurrentTask::active_pid_ns` + // function allows Rust to handle the common case of accessing `current`'s `PidNamespace` + // without RCU protection and without having to acquire a reference count. + // + // For (2) the `task_get_pid_ns()` method must be used. This will always acquire a + // reference on `PidNamespace` and will return an `Option` to force the caller to + // explicitly handle the case where `PidNamespace` is `None`, something that tends to be + // forgotten when doing the equivalent operation in `C`. Missing RCU primitives make it + // difficult to perform operations that are otherwise safe without holding a reference + // count as long as RCU protection is guaranteed. But it is not important currently. But we + // do want it in the future. + // + // Note for (2) the required RCU protection around calling `task_active_pid_ns()` + // synchronizes against putting the last reference of the associated `struct pid` of + // `task->thread_pid`. The `struct pid` stored in that field is used to retrieve the + // `PidNamespace` of the caller. When `release_task()` is called `task->thread_pid` will be + // `NULL`ed and `put_pid()` on said `struct pid` will be delayed in `free_pid()` via + // `call_rcu()` allowing everyone with an RCU protected access to the `struct pid` acquired + // from `task->thread_pid` to finish. + // + // SAFETY: If `current`'s pid ns is non-null, then it references a valid pid ns. + // Furthermore, the returned `&PidNamespace` borrows from this `CurrentTask`, so it cannot + // escape the scope in which the current pointer was obtained. + Some(unsafe { PidNamespace::from_ptr(active_ns) }) + } +} + // SAFETY: The type invariants guarantee that `Task` is always refcounted. unsafe impl crate::types::AlwaysRefCounted for Task { fn inc_ref(&self) { -- 2.47.0.338.g60cca15819-goog