// SPDX-License-Identifier: GPL-2.0 // Copyright (C) 2025 Google LLC. //! This module has utilities for managing a page range where unused pages may be reclaimed by a //! vma shrinker. // To avoid deadlocks, locks are taken in the order: // // 1. mmap lock // 2. spinlock // 3. lru spinlock // // The shrinker will use trylock methods because it locks them in a different order. use core::{ marker::PhantomPinned, mem::{size_of, size_of_val, MaybeUninit}, ptr, }; use kernel::{ bindings, error::Result, ffi::{c_ulong, c_void}, mm::{virt, Mm, MmWithUser}, new_mutex, new_spinlock, page::{Page, PAGE_SHIFT, PAGE_SIZE}, prelude::*, str::CStr, sync::{aref::ARef, Mutex, SpinLock}, task::Pid, transmute::FromBytes, types::Opaque, uaccess::UserSliceReader, }; /// Represents a shrinker that can be registered with the kernel. /// /// Each shrinker can be used by many `ShrinkablePageRange` objects. #[repr(C)] pub(crate) struct Shrinker { inner: Opaque<*mut bindings::shrinker>, list_lru: Opaque, } // SAFETY: The shrinker and list_lru are thread safe. unsafe impl Send for Shrinker {} // SAFETY: The shrinker and list_lru are thread safe. unsafe impl Sync for Shrinker {} impl Shrinker { /// Create a new shrinker. /// /// # Safety /// /// Before using this shrinker with a `ShrinkablePageRange`, the `register` method must have /// been called exactly once, and it must not have returned an error. pub(crate) const unsafe fn new() -> Self { Self { inner: Opaque::uninit(), list_lru: Opaque::uninit(), } } /// Register this shrinker with the kernel. pub(crate) fn register(&'static self, name: &CStr) -> Result<()> { // SAFETY: These fields are not yet used, so it's okay to zero them. unsafe { self.inner.get().write(ptr::null_mut()); self.list_lru.get().write_bytes(0, 1); } // SAFETY: The field is not yet used, so we can initialize it. let ret = unsafe { bindings::__list_lru_init(self.list_lru.get(), false, ptr::null_mut()) }; if ret != 0 { return Err(Error::from_errno(ret)); } // SAFETY: The `name` points at a valid c string. let shrinker = unsafe { bindings::shrinker_alloc(0, name.as_char_ptr()) }; if shrinker.is_null() { // SAFETY: We initialized it, so its okay to destroy it. unsafe { bindings::list_lru_destroy(self.list_lru.get()) }; return Err(Error::from_errno(ret)); } // SAFETY: We're about to register the shrinker, and these are the fields we need to // initialize. (All other fields are already zeroed.) unsafe { (&raw mut (*shrinker).count_objects).write(Some(rust_shrink_count)); (&raw mut (*shrinker).scan_objects).write(Some(rust_shrink_scan)); (&raw mut (*shrinker).private_data).write(self.list_lru.get().cast()); } // SAFETY: The new shrinker has been fully initialized, so we can register it. unsafe { bindings::shrinker_register(shrinker) }; // SAFETY: This initializes the pointer to the shrinker so that we can use it. unsafe { self.inner.get().write(shrinker) }; Ok(()) } } /// A container that manages a page range in a vma. /// /// The pages can be thought of as an array of booleans of whether the pages are usable. The /// methods `use_range` and `stop_using_range` set all booleans in a range to true or false /// respectively. Initially, no pages are allocated. When a page is not used, it is not freed /// immediately. Instead, it is made available to the memory shrinker to free it if the device is /// under memory pressure. /// /// It's okay for `use_range` and `stop_using_range` to race with each other, although there's no /// way to know whether an index ends up with true or false if a call to `use_range` races with /// another call to `stop_using_range` on a given index. /// /// It's also okay for the two methods to race with themselves, e.g. if two threads call /// `use_range` on the same index, then that's fine and neither call will return until the page is /// allocated and mapped. /// /// The methods that read or write to a range require that the page is marked as in use. So it is /// _not_ okay to call `stop_using_range` on a page that is in use by the methods that read or /// write to the page. #[pin_data(PinnedDrop)] pub(crate) struct ShrinkablePageRange { /// Shrinker object registered with the kernel. shrinker: &'static Shrinker, /// Pid using this page range. Only used as debugging information. pid: Pid, /// The mm for the relevant process. mm: ARef, /// Used to synchronize calls to `vm_insert_page` and `zap_page_range_single`. #[pin] mm_lock: Mutex<()>, /// Spinlock protecting changes to pages. #[pin] lock: SpinLock, /// Must not move, since page info has pointers back. #[pin] _pin: PhantomPinned, } struct Inner { /// Array of pages. /// /// Since this is also accessed by the shrinker, we can't use a `Box`, which asserts exclusive /// ownership. To deal with that, we manage it using raw pointers. pages: *mut PageInfo, /// Length of the `pages` array. size: usize, /// The address of the vma to insert the pages into. vma_addr: usize, } // SAFETY: proper locking is in place for `Inner` unsafe impl Send for Inner {} type StableMmGuard = kernel::sync::lock::Guard<'static, (), kernel::sync::lock::mutex::MutexBackend>; /// An array element that describes the current state of a page. /// /// There are three states: /// /// * Free. The page is None. The `lru` element is not queued. /// * Available. The page is Some. The `lru` element is queued to the shrinker's lru. /// * Used. The page is Some. The `lru` element is not queued. /// /// When an element is available, the shrinker is able to free the page. #[repr(C)] struct PageInfo { lru: bindings::list_head, page: Option, range: *const ShrinkablePageRange, } impl PageInfo { /// # Safety /// /// The caller ensures that writing to `me.page` is ok, and that the page is not currently set. unsafe fn set_page(me: *mut PageInfo, page: Page) { // SAFETY: This pointer offset is in bounds. let ptr = unsafe { &raw mut (*me).page }; // SAFETY: The pointer is valid for writing, so also valid for reading. if unsafe { (*ptr).is_some() } { pr_err!("set_page called when there is already a page"); // SAFETY: We will initialize the page again below. unsafe { ptr::drop_in_place(ptr) }; } // SAFETY: The pointer is valid for writing. unsafe { ptr::write(ptr, Some(page)) }; } /// # Safety /// /// The caller ensures that reading from `me.page` is ok for the duration of 'a. unsafe fn get_page<'a>(me: *const PageInfo) -> Option<&'a Page> { // SAFETY: This pointer offset is in bounds. let ptr = unsafe { &raw const (*me).page }; // SAFETY: The pointer is valid for reading. unsafe { (*ptr).as_ref() } } /// # Safety /// /// The caller ensures that writing to `me.page` is ok for the duration of 'a. unsafe fn take_page(me: *mut PageInfo) -> Option { // SAFETY: This pointer offset is in bounds. let ptr = unsafe { &raw mut (*me).page }; // SAFETY: The pointer is valid for reading. unsafe { (*ptr).take() } } /// Add this page to the lru list, if not already in the list. /// /// # Safety /// /// The pointer must be valid, and it must be the right shrinker and nid. unsafe fn list_lru_add(me: *mut PageInfo, nid: i32, shrinker: &'static Shrinker) { // SAFETY: This pointer offset is in bounds. let lru_ptr = unsafe { &raw mut (*me).lru }; // SAFETY: The lru pointer is valid, and we're not using it with any other lru list. unsafe { bindings::list_lru_add(shrinker.list_lru.get(), lru_ptr, nid, ptr::null_mut()) }; } /// Remove this page from the lru list, if it is in the list. /// /// # Safety /// /// The pointer must be valid, and it must be the right shrinker and nid. unsafe fn list_lru_del(me: *mut PageInfo, nid: i32, shrinker: &'static Shrinker) { // SAFETY: This pointer offset is in bounds. let lru_ptr = unsafe { &raw mut (*me).lru }; // SAFETY: The lru pointer is valid, and we're not using it with any other lru list. unsafe { bindings::list_lru_del(shrinker.list_lru.get(), lru_ptr, nid, ptr::null_mut()) }; } } impl ShrinkablePageRange { /// Create a new `ShrinkablePageRange` using the given shrinker. pub(crate) fn new(shrinker: &'static Shrinker) -> impl PinInit { try_pin_init!(Self { shrinker, pid: kernel::current!().pid(), mm: ARef::from(&**kernel::current!().mm().ok_or(ESRCH)?), mm_lock <- new_mutex!((), "ShrinkablePageRange::mm"), lock <- new_spinlock!(Inner { pages: ptr::null_mut(), size: 0, vma_addr: 0, }, "ShrinkablePageRange"), _pin: PhantomPinned, }) } pub(crate) fn stable_trylock_mm(&self) -> Option { // SAFETY: This extends the duration of the reference. Since this call happens before // `mm_lock` is taken in the destructor of `ShrinkablePageRange`, the destructor will block // until the returned guard is dropped. This ensures that the guard is valid until dropped. let mm_lock = unsafe { &*ptr::from_ref(&self.mm_lock) }; mm_lock.try_lock() } /// Register a vma with this page range. Returns the size of the region. pub(crate) fn register_with_vma(&self, vma: &virt::VmaNew) -> Result { let num_bytes = usize::min(vma.end() - vma.start(), bindings::SZ_4M as usize); let num_pages = num_bytes >> PAGE_SHIFT; if !ptr::eq::(&*self.mm, &**vma.mm()) { pr_debug!("Failed to register with vma: invalid vma->vm_mm"); return Err(EINVAL); } if num_pages == 0 { pr_debug!("Failed to register with vma: size zero"); return Err(EINVAL); } let mut pages = KVVec::::with_capacity(num_pages, GFP_KERNEL)?; // SAFETY: This just initializes the pages array. unsafe { let self_ptr = self as *const ShrinkablePageRange; for i in 0..num_pages { let info = pages.as_mut_ptr().add(i); (&raw mut (*info).range).write(self_ptr); (&raw mut (*info).page).write(None); let lru = &raw mut (*info).lru; (&raw mut (*lru).next).write(lru); (&raw mut (*lru).prev).write(lru); } } let mut inner = self.lock.lock(); if inner.size > 0 { pr_debug!("Failed to register with vma: already registered"); drop(inner); return Err(EBUSY); } inner.pages = pages.into_raw_parts().0; inner.size = num_pages; inner.vma_addr = vma.start(); Ok(num_pages) } /// Make sure that the given pages are allocated and mapped. /// /// Must not be called from an atomic context. pub(crate) fn use_range(&self, start: usize, end: usize) -> Result<()> { if start >= end { return Ok(()); } let mut inner = self.lock.lock(); assert!(end <= inner.size); for i in start..end { // SAFETY: This pointer offset is in bounds. let page_info = unsafe { inner.pages.add(i) }; // SAFETY: The pointer is valid, and we hold the lock so reading from the page is okay. if let Some(page) = unsafe { PageInfo::get_page(page_info) } { // Since we're going to use the page, we should remove it from the lru list so that // the shrinker will not free it. // // SAFETY: The pointer is valid, and this is the right shrinker. // // The shrinker can't free the page between the check and this call to // `list_lru_del` because we hold the lock. unsafe { PageInfo::list_lru_del(page_info, page.nid(), self.shrinker) }; } else { // We have to allocate a new page. Use the slow path. drop(inner); // SAFETY: `i < end <= inner.size` so `i` is in bounds. match unsafe { self.use_page_slow(i) } { Ok(()) => {} Err(err) => { pr_warn!("Error in use_page_slow: {:?}", err); return Err(err); } } inner = self.lock.lock(); } } Ok(()) } /// Mark the given page as in use, slow path. /// /// Must not be called from an atomic context. /// /// # Safety /// /// Assumes that `i` is in bounds. #[cold] unsafe fn use_page_slow(&self, i: usize) -> Result<()> { let new_page = Page::alloc_page(GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO)?; let mm_mutex = self.mm_lock.lock(); let inner = self.lock.lock(); // SAFETY: This pointer offset is in bounds. let page_info = unsafe { inner.pages.add(i) }; // SAFETY: The pointer is valid, and we hold the lock so reading from the page is okay. if let Some(page) = unsafe { PageInfo::get_page(page_info) } { // The page was already there, or someone else added the page while we didn't hold the // spinlock. // // SAFETY: The pointer is valid, and this is the right shrinker. // // The shrinker can't free the page between the check and this call to // `list_lru_del` because we hold the lock. unsafe { PageInfo::list_lru_del(page_info, page.nid(), self.shrinker) }; return Ok(()); } let vma_addr = inner.vma_addr; // Release the spinlock while we insert the page into the vma. drop(inner); // No overflow since we stay in bounds of the vma. let user_page_addr = vma_addr + (i << PAGE_SHIFT); // We use `mmput_async` when dropping the `mm` because `use_page_slow` is usually used from // a remote process. If the call to `mmput` races with the process shutting down, then the // caller of `use_page_slow` becomes responsible for cleaning up the `mm`, which doesn't // happen until it returns to userspace. However, the caller might instead go to sleep and // wait for the owner of the `mm` to wake it up, which doesn't happen because it's in the // middle of a shutdown process that won't complete until the `mm` is dropped. This can // amount to a deadlock. // // Using `mmput_async` avoids this, because then the `mm` cleanup is instead queued to a // workqueue. MmWithUser::into_mmput_async(self.mm.mmget_not_zero().ok_or(ESRCH)?) .mmap_read_lock() .vma_lookup(vma_addr) .ok_or(ESRCH)? .as_mixedmap_vma() .ok_or(ESRCH)? .vm_insert_page(user_page_addr, &new_page) .inspect_err(|err| { pr_warn!( "Failed to vm_insert_page({}): vma_addr:{} i:{} err:{:?}", user_page_addr, vma_addr, i, err ) })?; let inner = self.lock.lock(); // SAFETY: The `page_info` pointer is valid and currently does not have a page. The page // can be written to since we hold the lock. // // We released and reacquired the spinlock since we checked that the page is null, but we // always hold the mm_lock mutex when setting the page to a non-null value, so it's not // possible for someone else to have changed it since our check. unsafe { PageInfo::set_page(page_info, new_page) }; drop(inner); drop(mm_mutex); Ok(()) } /// If the given page is in use, then mark it as available so that the shrinker can free it. /// /// May be called from an atomic context. pub(crate) fn stop_using_range(&self, start: usize, end: usize) { if start >= end { return; } let inner = self.lock.lock(); assert!(end <= inner.size); for i in (start..end).rev() { // SAFETY: The pointer is in bounds. let page_info = unsafe { inner.pages.add(i) }; // SAFETY: Okay for reading since we have the lock. if let Some(page) = unsafe { PageInfo::get_page(page_info) } { // SAFETY: The pointer is valid, and it's the right shrinker. unsafe { PageInfo::list_lru_add(page_info, page.nid(), self.shrinker) }; } } } /// Helper for reading or writing to a range of bytes that may overlap with several pages. /// /// # Safety /// /// All pages touched by this operation must be in use for the duration of this call. unsafe fn iterate(&self, mut offset: usize, mut size: usize, mut cb: T) -> Result where T: FnMut(&Page, usize, usize) -> Result, { if size == 0 { return Ok(()); } let (pages, num_pages) = { let inner = self.lock.lock(); (inner.pages, inner.size) }; let num_bytes = num_pages << PAGE_SHIFT; // Check that the request is within the buffer. if offset.checked_add(size).ok_or(EFAULT)? > num_bytes { return Err(EFAULT); } let mut page_index = offset >> PAGE_SHIFT; offset &= PAGE_SIZE - 1; while size > 0 { let available = usize::min(size, PAGE_SIZE - offset); // SAFETY: The pointer is in bounds. let page_info = unsafe { pages.add(page_index) }; // SAFETY: The caller guarantees that this page is in the "in use" state for the // duration of this call to `iterate`, so nobody will change the page. let page = unsafe { PageInfo::get_page(page_info) }; if page.is_none() { pr_warn!("Page is null!"); } let page = page.ok_or(EFAULT)?; cb(page, offset, available)?; size -= available; page_index += 1; offset = 0; } Ok(()) } /// Copy from userspace into this page range. /// /// # Safety /// /// All pages touched by this operation must be in use for the duration of this call. pub(crate) unsafe fn copy_from_user_slice( &self, reader: &mut UserSliceReader, offset: usize, size: usize, ) -> Result { // SAFETY: `self.iterate` has the same safety requirements as `copy_from_user_slice`. unsafe { self.iterate(offset, size, |page, offset, to_copy| { page.copy_from_user_slice_raw(reader, offset, to_copy) }) } } /// Copy from this page range into kernel space. /// /// # Safety /// /// All pages touched by this operation must be in use for the duration of this call. pub(crate) unsafe fn read(&self, offset: usize) -> Result { let mut out = MaybeUninit::::uninit(); let mut out_offset = 0; // SAFETY: `self.iterate` has the same safety requirements as `read`. unsafe { self.iterate(offset, size_of::(), |page, offset, to_copy| { // SAFETY: The sum of `offset` and `to_copy` is bounded by the size of T. let obj_ptr = (out.as_mut_ptr() as *mut u8).add(out_offset); // SAFETY: The pointer points is in-bounds of the `out` variable, so it is valid. page.read_raw(obj_ptr, offset, to_copy)?; out_offset += to_copy; Ok(()) })?; } // SAFETY: We just initialised the data. Ok(unsafe { out.assume_init() }) } /// Copy from kernel space into this page range. /// /// # Safety /// /// All pages touched by this operation must be in use for the duration of this call. pub(crate) unsafe fn write(&self, offset: usize, obj: &T) -> Result { let mut obj_offset = 0; // SAFETY: `self.iterate` has the same safety requirements as `write`. unsafe { self.iterate(offset, size_of_val(obj), |page, offset, to_copy| { // SAFETY: The sum of `offset` and `to_copy` is bounded by the size of T. let obj_ptr = (obj as *const T as *const u8).add(obj_offset); // SAFETY: We have a reference to the object, so the pointer is valid. page.write_raw(obj_ptr, offset, to_copy)?; obj_offset += to_copy; Ok(()) }) } } /// Write zeroes to the given range. /// /// # Safety /// /// All pages touched by this operation must be in use for the duration of this call. pub(crate) unsafe fn fill_zero(&self, offset: usize, size: usize) -> Result { // SAFETY: `self.iterate` has the same safety requirements as `copy_into`. unsafe { self.iterate(offset, size, |page, offset, len| { page.fill_zero_raw(offset, len) }) } } } #[pinned_drop] impl PinnedDrop for ShrinkablePageRange { fn drop(self: Pin<&mut Self>) { let (pages, size) = { let lock = self.lock.lock(); (lock.pages, lock.size) }; if size == 0 { return; } // Note: This call is also necessary for the safety of `stable_trylock_mm`. let mm_lock = self.mm_lock.lock(); // This is the destructor, so unlike the other methods, we only need to worry about races // with the shrinker here. Since we hold the `mm_lock`, we also can't race with the // shrinker, and after this loop, the shrinker will not access any of our pages since we // removed them from the lru list. for i in 0..size { // SAFETY: Loop is in-bounds of the size. let p_ptr = unsafe { pages.add(i) }; // SAFETY: No other readers, so we can read. if let Some(p) = unsafe { PageInfo::get_page(p_ptr) } { // SAFETY: The pointer is valid and it's the right shrinker. unsafe { PageInfo::list_lru_del(p_ptr, p.nid(), self.shrinker) }; } } drop(mm_lock); // SAFETY: `pages` was allocated as an `KVVec` with capacity `size`. Furthermore, // all `size` elements are initialized. Also, the array is no longer shared with the // shrinker due to the above loop. drop(unsafe { KVVec::from_raw_parts(pages, size, size) }); } } /// # Safety /// Called by the shrinker. #[no_mangle] unsafe extern "C" fn rust_shrink_count( shrink: *mut bindings::shrinker, _sc: *mut bindings::shrink_control, ) -> c_ulong { // SAFETY: We can access our own private data. let list_lru = unsafe { (*shrink).private_data.cast::() }; // SAFETY: Accessing the lru list is okay. Just an FFI call. unsafe { bindings::list_lru_count(list_lru) } } /// # Safety /// Called by the shrinker. #[no_mangle] unsafe extern "C" fn rust_shrink_scan( shrink: *mut bindings::shrinker, sc: *mut bindings::shrink_control, ) -> c_ulong { // SAFETY: We can access our own private data. let list_lru = unsafe { (*shrink).private_data.cast::() }; // SAFETY: Caller guarantees that it is safe to read this field. let nr_to_scan = unsafe { (*sc).nr_to_scan }; // SAFETY: Accessing the lru list is okay. Just an FFI call. unsafe { bindings::list_lru_walk( list_lru, Some(bindings::rust_shrink_free_page_wrap), ptr::null_mut(), nr_to_scan, ) } } const LRU_SKIP: bindings::lru_status = bindings::lru_status_LRU_SKIP; const LRU_REMOVED_ENTRY: bindings::lru_status = bindings::lru_status_LRU_REMOVED_RETRY; /// # Safety /// Called by the shrinker. #[no_mangle] unsafe extern "C" fn rust_shrink_free_page( item: *mut bindings::list_head, lru: *mut bindings::list_lru_one, _cb_arg: *mut c_void, ) -> bindings::lru_status { // Fields that should survive after unlocking the lru lock. let page; let page_index; let mm; let mmap_read; let mm_mutex; let vma_addr; { // CAST: The `list_head` field is first in `PageInfo`. let info = item as *mut PageInfo; // SAFETY: The `range` field of `PageInfo` is immutable. let range = unsafe { &*((*info).range) }; mm = match range.mm.mmget_not_zero() { Some(mm) => MmWithUser::into_mmput_async(mm), None => return LRU_SKIP, }; mm_mutex = match range.stable_trylock_mm() { Some(guard) => guard, None => return LRU_SKIP, }; mmap_read = match mm.mmap_read_trylock() { Some(guard) => guard, None => return LRU_SKIP, }; // We can't lock it normally here, since we hold the lru lock. let inner = match range.lock.try_lock() { Some(inner) => inner, None => return LRU_SKIP, }; // SAFETY: The item is in this lru list, so it's okay to remove it. unsafe { bindings::list_lru_isolate(lru, item) }; // SAFETY: Both pointers are in bounds of the same allocation. page_index = unsafe { info.offset_from(inner.pages) } as usize; // SAFETY: We hold the spinlock, so we can take the page. // // This sets the page pointer to zero before we unmap it from the vma. However, we call // `zap_page_range` before we release the mmap lock, so `use_page_slow` will not be able to // insert a new page until after our call to `zap_page_range`. page = unsafe { PageInfo::take_page(info) }; vma_addr = inner.vma_addr; // From this point on, we don't access this PageInfo or ShrinkablePageRange again, because // they can be freed at any point after we unlock `lru_lock`. This is with the exception of // `mm_mutex` which is kept alive by holding the lock. } // SAFETY: The lru lock is locked when this method is called. unsafe { bindings::spin_unlock(&raw mut (*lru).lock) }; if let Some(vma) = mmap_read.vma_lookup(vma_addr) { let user_page_addr = vma_addr + (page_index << PAGE_SHIFT); vma.zap_page_range_single(user_page_addr, PAGE_SIZE); } drop(mmap_read); drop(mm_mutex); drop(mm); drop(page); // SAFETY: We just unlocked the lru lock, but it should be locked when we return. unsafe { bindings::spin_lock(&raw mut (*lru).lock) }; LRU_REMOVED_ENTRY }