Opportunistic CoW for platforms that support it

litebox/src/fs/in_mem.rs

@@ -785,6 +785,21 @@ impl<Platform: sync::RawSyncPrimitivesProvider> super::FileSystem for FileSystem
             blksize: BLOCK_SIZE,
         })
     }
+
+    fn get_static_backing_data(&self, fd: &FileFd<Platform>) -> Option<&'static [u8]> {
+        let descriptor_table = self.litebox.descriptor_table();
+        let entry = descriptor_table.get_entry(fd)?;
+        match &entry.entry {
+            Descriptor::File { file, .. } => {
+                let file = file.read();
+                match &file.data {
+                    alloc::borrow::Cow::Borrowed(slice) => Some(*slice),
+                    alloc::borrow::Cow::Owned(_) => None,
+                }
+            }
+            Descriptor::Dir { .. } => None,
+        }
+    }
 }
 
 struct RootDir<Platform: sync::RawSyncPrimitivesProvider> {

litebox/src/fs/layered.rs

@@ -1336,6 +1336,21 @@ impl<
             blksize,
         })
     }
+
+    fn get_static_backing_data(
+        &self,
+        fd: &FileFd<Platform, Upper, Lower>,
+    ) -> Option<&'static [u8]> {
+        let entry = self
+            .litebox
+            .descriptor_table()
+            .with_entry(fd, |descriptor| Arc::clone(&descriptor.entry.entry))?;
+        match entry.as_ref() {
+            EntryX::Upper { fd } => self.upper.get_static_backing_data(fd),
+            EntryX::Lower { fd } => self.lower.get_static_backing_data(fd),
+            EntryX::Tombstone => unreachable!(),
+        }
+    }
 }
 
 struct Descriptor<Upper: super::FileSystem + 'static, Lower: super::FileSystem + 'static> {

litebox/src/fs/mod.rs

@@ -136,6 +136,17 @@ pub trait FileSystem: private::Sealed + FdEnabledSubsystem {
 
     /// Equivalent to [`Self::file_status`], but open an open `fd` instead.
     fn fd_file_status(&self, fd: &TypedFd<Self>) -> Result<FileStatus, FileStatusError>;
+
+    /// Get static backing data for a file, if available and supported.
+    ///
+    /// This method returns the (entire) underlying static byte slice if the file's contents are
+    /// backed by borrowed static data (e.g., loaded via `initialize_primarily_read_heavy_file`).
+    ///
+    /// Returns `None` if indicating no static backing data is available/supported.
+    #[expect(unused_variables, reason = "default body, non-underscored param names")]
+    fn get_static_backing_data(&self, fd: &TypedFd<Self>) -> Option<&'static [u8]> {
+        None
+    }
 }
 
 bitflags! {

litebox/src/mm/linux.rs

@@ -315,6 +315,19 @@ impl<Platform: PageManagementProvider<ALIGN> + 'static, const ALIGN: usize> Vmem
         self.vmas.iter()
     }
 
+    /// Insert an already-allocated region (e.g., via CoW) without calling the platform allocator.
+    ///
+    /// Any existing tracked mappings that overlap `range` are silently removed from tracking
+    /// (without calling the platform deallocator) before inserting. Use [`Self::overlapping`] to
+    /// check for overlap before running this if needed.
+    pub(super) fn register_existing_mapping_overwrite(
+        &mut self,
+        range: PageRange<ALIGN>,
+        vma: VmArea,
+    ) {
+        self.vmas.insert(range.into(), vma);
+    }
+
     /// Gets an iterator over all the stored ranges that are
     /// either partially or completely overlapped by the given range.
     pub(super) fn overlapping(

litebox/src/mm/mod.rs

@@ -14,8 +14,8 @@ use core::ops::Range;
 
 use alloc::vec::Vec;
 use linux::{
-    CreatePagesFlags, MappingError, PageFaultError, PageRange, VmFlags, Vmem, VmemPageFaultHandler,
-    VmemProtectError, VmemUnmapError,
+    CreatePagesFlags, MappingError, PageFaultError, PageRange, VmArea, VmFlags, Vmem,
+    VmemPageFaultHandler, VmemProtectError, VmemUnmapError,
 };
 
 use crate::{
@@ -591,6 +591,39 @@ where
         )
     }
 
+    /// Register an already-allocated memory region in the VMA tracker.
+    ///
+    /// This is used when memory has been allocated by some means other than the normal
+    /// `create_*_pages` path (e.g., CoW mappings created directly by the platform), so that the
+    /// page manager tracks the region for future `mprotect`, `munmap`, etc.
+    ///
+    /// If `replace` is `true`, any overlapping tracked mappings are evicted from the tracker
+    /// (without calling the platform deallocator) before inserting. Otherwise, returns `None`
+    /// without registering if the provided `range` overlaps with any existing mapping.
+    ///
+    /// # Safety
+    ///
+    /// The `range` must be an already-mapped region with the given `permissions`.
+    #[must_use]
+    pub unsafe fn register_existing_mapping(
+        &self,
+        range: PageRange<ALIGN>,
+        permissions: MemoryRegionPermissions,
+        is_file_backed: bool,
+        replace: bool,
+    ) -> Option<()> {
+        let vma = VmArea::new(
+            VmFlags::from(permissions) | VmFlags::VM_MAY_ACCESS_FLAGS,
+            is_file_backed,
+        );
+        let mut vmem = self.vmem.write();
+        if !replace && vmem.overlapping(range.into()).next().is_some() {
+            return None;
+        }
+        vmem.register_existing_mapping_overwrite(range, vma);
+        Some(())
+    }
+
     /// Returns all mappings in a vector.
     pub fn mappings(&self) -> Vec<(Range<usize>, VmFlags)> {
         self.vmem

litebox/src/platform/page_mgmt.rs

@@ -172,6 +172,25 @@ pub trait PageManagementProvider<const ALIGN: usize>: RawPointerProvider {
     ///
     /// Note that the returned ranges should be `ALIGN`-aligned.
     fn reserved_pages(&self) -> impl Iterator<Item = &Range<usize>>;
+
+    /// Attempt to allocate pages with copy-on-write semantics backed by static data.
+    ///
+    /// This method allows platforms that support it to create CoW mappings instead of performing
+    /// expensive page-by-page memory copies. This is particularly useful when mapping pre-loaded
+    /// file data that was mmap'd by the host.
+    ///
+    /// The default implementation returns unsupported CoW. Platforms that DO support COW should
+    /// override this method to unlock better performance.
+    #[expect(unused_variables, reason = "default body, non-underscored param names")]
+    fn try_allocate_cow_pages(
+        &self,
+        suggested_start: usize,
+        source_data: &'static [u8],
+        permissions: MemoryRegionPermissions,
+        fixed_address_behavior: FixedAddressBehavior,
+    ) -> Result<Self::RawMutPointer<u8>, CowAllocationError> {
+        Err(CowAllocationError::UnsupportedByPlatform)
+    }
 }
 
 /// Behavior when allocating pages at a fixed address.
@@ -243,3 +262,29 @@ pub enum PermissionUpdateError {
     #[error("provided range contains unallocated pages")]
     Unallocated,
 }
+
+/// Possible errors for [`PageManagementProvider::try_allocate_cow_pages`]
+///
+/// ```text
+///  ____________________
+/// ( Maybe the grass is )
+/// ( greener on the     )
+/// ( other side?        )
+///  --------------------
+///         o   ^__^
+///          o  (oo)\_______
+///             (__)\       )\/\
+///                 ||----w |
+///                 ||     ||
+/// ```
+#[derive(Error, Debug)]
+pub enum CowAllocationError {
+    #[error("copy-on-write page allocation is not supported for this particular platform")]
+    UnsupportedByPlatform,
+    #[error("source region is not copy-on-writable")]
+    UnsupportedSourceRegion,
+    #[error("unaligned request")]
+    Unaligned,
+    #[error("internal failure in creating CoW pages")]
+    InternalFailure,
+}

litebox_platform_linux_userland/src/lib.rs

@@ -9,12 +9,15 @@
 
 use std::cell::Cell;
 use std::os::fd::{AsRawFd as _, FromRawFd as _};
+use std::path::PathBuf;
 use std::sync::atomic::{AtomicI32, AtomicU32, Ordering};
 use std::time::Duration;
 
 use litebox::fs::OFlags;
 use litebox::platform::UnblockedOrTimedOut;
-use litebox::platform::page_mgmt::{FixedAddressBehavior, MemoryRegionPermissions};
+use litebox::platform::page_mgmt::{
+    CowAllocationError, FixedAddressBehavior, MemoryRegionPermissions,
+};
 use litebox::platform::{ImmediatelyWokenUp, RawConstPointer as _};
 use litebox::shim::ContinueOperation;
 use litebox::utils::{ReinterpretSignedExt, ReinterpretUnsignedExt as _, TruncateExt};
@@ -40,6 +43,9 @@ pub struct LinuxUserland {
     reserved_pages: Vec<core::ops::Range<usize>>,
     /// The base address of the VDSO.
     vdso_address: Option<usize>,
+    /// CoW-eligible memory regions. Maps start address of the static slice, to the info needed to
+    /// re-mmap the file.
+    cow_regions: std::sync::RwLock<std::collections::BTreeMap<usize, CowRegionInfo>>,
 }
 
 impl core::fmt::Debug for LinuxUserland {
@@ -48,6 +54,15 @@ impl core::fmt::Debug for LinuxUserland {
     }
 }
 
+/// Information about a CoW-eligible memory region backed by a file.
+#[derive(Debug, Clone)]
+struct CowRegionInfo {
+    /// The path to the backing file on the host filesystem.
+    file_path: PathBuf,
+    /// Length of the backing file.
+    file_length: usize,
+}
+
 const IF_NAMESIZE: usize = 16;
 /// Use TUN device
 const IFF_TUN: i32 = 0x0001;
@@ -163,10 +178,53 @@ impl LinuxUserland {
             seccomp_interception_enabled: std::sync::atomic::AtomicBool::new(false),
             reserved_pages,
             vdso_address,
+            cow_regions: std::sync::RwLock::new(std::collections::BTreeMap::new()),
         };
         Box::leak(Box::new(platform))
     }
 
+    /// Register a CoW-eligible memory region backed by a file.
+    ///
+    /// # Panics
+    ///
+    /// Panics if an overlapping region is already registered.
+    pub fn register_cow_region(&self, data: &'static [u8], file_path: impl Into<PathBuf>) {
+        let start = data.as_ptr() as usize;
+        let info = CowRegionInfo {
+            file_path: file_path.into(),
+            file_length: data.len(),
+        };
+
+        let mut regions = self.cow_regions.write().unwrap();
+        assert!(
+            regions.range(start..start + data.len()).next().is_none(),
+            "Attempting to register an overlapping region"
+        );
+        let old = regions.insert(start, info);
+        assert!(old.is_none());
+    }
+
+    /// Look up the file backing a static slice for CoW mapping.
+    ///
+    /// Returns `Some((file_path, offset_in_file))` if the slice is backed by a registered
+    /// CoW region, `None` otherwise.
+    fn lookup_cow_region(&self, source_data: &'static [u8]) -> Option<(PathBuf, usize)> {
+        let slice_start = source_data.as_ptr() as usize;
+        let slice_len = source_data.len();
+
+        let regions = self.cow_regions.read().unwrap();
+
+        if let Some((&region_start, info)) = regions.range(..=slice_start).next_back() {
+            let region_end = region_start.checked_add(info.file_length).unwrap();
+            let slice_end = slice_start.checked_add(slice_len).unwrap();
+
+            if slice_start >= region_start && slice_end <= region_end {
+                return Some((info.file_path.clone(), slice_start - region_start));
+            }
+        }
+        None
+    }
+
     /// Enable seccomp syscall interception on the platform.
     ///
     /// # Panics
@@ -1490,6 +1548,89 @@ impl<const ALIGN: usize> litebox::platform::PageManagementProvider<ALIGN> for Li
     fn reserved_pages(&self) -> impl Iterator<Item = &core::ops::Range<usize>> {
         self.reserved_pages.iter()
     }
+
+    fn try_allocate_cow_pages(
+        &self,
+        suggested_start: usize,
+        source_data: &'static [u8],
+        permissions: MemoryRegionPermissions,
+        fixed_address_behavior: FixedAddressBehavior,
+    ) -> Result<Self::RawMutPointer<u8>, CowAllocationError> {
+        let Some((file_path, file_offset)) = self.lookup_cow_region(source_data) else {
+            return Err(CowAllocationError::UnsupportedSourceRegion);
+        };
+        if !file_offset.is_multiple_of(ALIGN) {
+            return Err(CowAllocationError::Unaligned);
+        }
+
+        let file_path_cstr =
+            std::ffi::CString::new(file_path.as_os_str().as_encoded_bytes()).unwrap();
+        // TODO(jb): We should likely be storing pre-opened FDs, right?
+        let fd = unsafe {
+            syscalls::syscall4(
+                syscalls::Sysno::open,
+                file_path_cstr.as_ptr() as usize,
+                OFlags::RDONLY.bits() as usize,
+                0,
+                // Unused by the syscall but would be checked by Seccomp filter if enabled.
+                syscall_intercept::SYSCALL_ARG_MAGIC,
+            )
+        };
+        let fd = fd.expect("file should remain unchanged on host");
+
+        let mut flags = MapFlags::MAP_PRIVATE;
+        match fixed_address_behavior {
+            FixedAddressBehavior::Hint => {}
+            FixedAddressBehavior::Replace => flags |= MapFlags::MAP_FIXED,
+            FixedAddressBehavior::NoReplace => flags |= MapFlags::MAP_FIXED_NOREPLACE,
+        }
+
+        let result = unsafe {
+            syscalls::syscall6(
+                {
+                    #[cfg(target_arch = "x86_64")]
+                    {
+                        syscalls::Sysno::mmap
+                    }
+                    #[cfg(target_arch = "x86")]
+                    {
+                        syscalls::Sysno::mmap2
+                    }
+                },
+                suggested_start,
+                source_data.len(),
+                prot_flags(permissions).bits().reinterpret_as_unsigned() as usize,
+                (flags.bits().reinterpret_as_unsigned()
+                    // This is to ensure it won't be intercepted by Seccomp if enabled.
+                    | syscall_intercept::MMAP_FLAG_MAGIC) as usize,
+                fd,
+                {
+                    #[cfg(target_arch = "x86_64")]
+                    {
+                        file_offset
+                    }
+                    #[cfg(target_arch = "x86")]
+                    {
+                        // mmap2 takes offset in pages, not bytes
+                        file_offset / ALIGN
+                    }
+                },
+            )
+        };
+
+        let _ = unsafe {
+            syscalls::syscall2(
+                syscalls::Sysno::close,
+                fd, // This is to ensure it won't be intercepted by Seccomp if enabled.
+                syscall_intercept::SYSCALL_ARG_MAGIC,
+            )
+        };
+
+        match result {
+            Ok(ptr) => Ok(UserMutPtr::from_usize(ptr)),
+            Err(_) => Err(CowAllocationError::InternalFailure),
+        }
+    }
 }
 
 impl litebox::platform::StdioProvider for LinuxUserland {