feat: kernel W^X (per-section perms, firmware unmapped, stack guard, boot check), halt on kernel faults, region W^X across processes

crates/adapter-riscv/linker.ld

@@ -1,8 +1,11 @@
 OUTPUT_ARCH(riscv)
 ENTRY(_start)
 
+/* Every boundary below is page aligned: the kernel maps each range with
+   its own permissions (W^X), see mmu::init_kernel_page_table. */
 SECTIONS {
     . = 0x80200000;
+    _text_start = .;
 
     .text : {
         *(.text.entry)
@@ -10,13 +13,22 @@ SECTIONS {
     }
 
     . = ALIGN(4096);
+    _text_end = .;
+
     .rodata : {
         *(.rodata .rodata.*)
+        *(.srodata .srodata.*)
+    }
+    .eh_frame : {
+        *(.eh_frame)
     }
 
     . = ALIGN(4096);
+    _rodata_end = .;
+
     .data : {
         *(.data .data.*)
+        *(.sdata .sdata.*)
     }
 
     . = ALIGN(4096);
@@ -27,7 +39,12 @@ SECTIONS {
         _bss_end = .;
     }
 
+    /* Left unmapped, so overflowing the kernel stack faults instead of
+       silently overwriting .bss. */
     . = ALIGN(4096);
+    _stack_guard = .;
+    . += 4096;
+
     _stack_bottom = .;
     . += 0x80000;
     _stack_top = .;

crates/adapter-riscv/src/boot/hardware.rs

@@ -16,6 +16,26 @@ pub fn init(uart: &Uart) -> &'static mut mmu::PageTable {
     let kernel_pt = mmu::init_kernel_page_table();
     mmu::activate(kernel_pt);
     uart.write_str("k-os: MMU enabled (Sv39)\n");
+    match mmu::check_wx(kernel_pt) {
+        Ok(code_pages) => {
+            uart.write_str("k-os: W^X verified, ");
+            uart.write_dec(code_pages);
+            uart.write_str(" code pages\n");
+        }
+        Err(v) => {
+            let (what, va) = match v {
+                mmu::page_table::WxViolation::WritableCode(va) => ("writable code", va),
+                mmu::page_table::WxViolation::ExecutableData(va) => ("executable data", va),
+                mmu::page_table::WxViolation::MappedHole(va) => ("mapped hole", va),
+            };
+            uart.write_str("k-os: W^X violation: ");
+            uart.write_str(what);
+            uart.write_str(" at ");
+            uart.write_hex(va);
+            uart.put_char(b'\n');
+            crate::halt(uart, "kernel page table violates W^X");
+        }
+    }
 
     trap::init();
     crate::hw::timer::init();

crates/adapter-riscv/src/mmu/mod.rs

@@ -3,6 +3,6 @@ pub mod page_table;
 
 pub use allocator::init_allocator;
 pub use page_table::{
-    PAGE_SIZE, PageTable, Sv39Flags, USER_VA_END, USER_VA_START, activate, active_root,
+    PAGE_SIZE, PageTable, Sv39Flags, USER_VA_END, USER_VA_START, activate, active_root, check_wx,
     create_user_root, init_kernel_page_table, map_user_page, unmap_user_page,
 };

crates/adapter-riscv/src/mmu/page_table.rs

@@ -22,9 +22,15 @@ pub struct Sv39Flags {
 }
 
 impl Sv39Flags {
-    pub fn kernel_rwx() -> Self {
+    pub fn kernel_rx() -> Self {
         Self {
-            bits: PTE_R | PTE_W | PTE_X | PTE_G,
+            bits: PTE_R | PTE_X | PTE_G,
+        }
+    }
+
+    pub fn kernel_ro() -> Self {
+        Self {
+            bits: PTE_R | PTE_G,
         }
     }
 
@@ -140,16 +146,33 @@ pub fn init_kernel_page_table() -> &'static mut PageTable {
         )
     };
 
-    let (va, pa) = identity(board::RAM_BASE);
+    // The kernel image, then everything after it, each with only the
+    // access it needs: code is never writable, data never executable.
+    // RAM below the image belongs to the firmware and stays unmapped, as
+    // does the guard page under the kernel stack.
+    let layout = KernelLayout::from_linker();
+    let ram_end = board::RAM_BASE + board::RAM_SIZE;
+    for (start, end, flags) in [
+        (layout.text_start, layout.text_end, Sv39Flags::kernel_rx()),
+        (layout.text_end, layout.rodata_end, Sv39Flags::kernel_ro()),
+        (
+            layout.rodata_end,
+            layout.stack_guard,
+            Sv39Flags::kernel_rw(),
+        ),
+        (layout.stack_bottom, ram_end, Sv39Flags::kernel_rw()),
+    ] {
+        let (va, pa) = identity(start);
         map(
             root,
             Mapping {
                 va,
                 pa,
-            pages: board::RAM_SIZE / PAGE_SIZE,
+                pages: (end - start) / PAGE_SIZE,
-            flags: Sv39Flags::kernel_rwx(),
+                flags,
             },
         );
+    }
 
     let (va, pa) = identity(board::UART_BASE);
     map(
@@ -198,6 +221,102 @@ pub fn init_kernel_page_table() -> &'static mut PageTable {
     root
 }
 
+/// Page-aligned boundaries of the kernel image, from linker.ld.
+pub struct KernelLayout {
+    pub text_start: usize,
+    pub text_end: usize,
+    pub rodata_end: usize,
+    pub stack_guard: usize,
+    pub stack_bottom: usize,
+}
+
+impl KernelLayout {
+    pub fn from_linker() -> Self {
+        unsafe extern "C" {
+            static _text_start: u8;
+            static _text_end: u8;
+            static _rodata_end: u8;
+            static _stack_guard: u8;
+            static _stack_bottom: u8;
+        }
+        // SAFETY: only the symbols' addresses are taken, never their values.
+        unsafe {
+            Self {
+                text_start: &_text_start as *const u8 as usize,
+                text_end: &_text_end as *const u8 as usize,
+                rodata_end: &_rodata_end as *const u8 as usize,
+                stack_guard: &_stack_guard as *const u8 as usize,
+                stack_bottom: &_stack_bottom as *const u8 as usize,
+            }
+        }
+    }
+}
+
+/// What `check_wx` found wrong with a kernel page table, if anything.
+pub enum WxViolation {
+    /// A page that is both writable and executable.
+    WritableCode(usize),
+    /// An executable page outside the kernel's text.
+    ExecutableData(usize),
+    /// A mapped page under the firmware, or the stack guard.
+    MappedHole(usize),
+}
+
+/// Walks every leaf of `root` and checks the W^X layout: nothing is both
+/// writable and executable, only kernel text is executable, and the
+/// firmware and the stack guard are unmapped. Returns the number of
+/// executable pages.
+pub fn check_wx(root: &PageTable) -> Result<usize, WxViolation> {
+    let layout = KernelLayout::from_linker();
+    let text = layout.text_start..layout.text_end;
+    let guard = layout.stack_guard..layout.stack_bottom;
+    let firmware = board::RAM_BASE..layout.text_start;
+    let mut executable = 0;
+    for (va, bits) in leaves(root) {
+        if bits & PTE_W != 0 && bits & PTE_X != 0 {
+            return Err(WxViolation::WritableCode(va));
+        }
+        if bits & PTE_X != 0 {
+            if !text.contains(&va) {
+                return Err(WxViolation::ExecutableData(va));
+            }
+            executable += 1;
+        }
+        if guard.contains(&va) || firmware.contains(&va) {
+            return Err(WxViolation::MappedHole(va));
+        }
+    }
+    Ok(executable)
+}
+
+/// Every 4 KiB leaf in `root` as (virtual address, PTE flag bits).
+/// The kernel maps nothing larger.
+fn leaves(root: &PageTable) -> impl Iterator<Item = (usize, u64)> + '_ {
+    let next = |e: &PageTableEntry| table_at((e.ppn() << 12) as usize) as &PageTable;
+    root.entries
+        .iter()
+        .enumerate()
+        .filter(|(_, e)| e.is_valid())
+        .flat_map(move |(i2, e2)| {
+            next(e2)
+                .entries
+                .iter()
+                .enumerate()
+                .filter(|(_, e)| e.is_valid())
+                .flat_map(move |(i1, e1)| {
+                    next(e1)
+                        .entries
+                        .iter()
+                        .enumerate()
+                        .filter(|(_, e)| e.is_valid())
+                        .map(move |(i0, e0)| {
+                            let va = (i2 << 30) | (i1 << 21) | (i0 << 12);
+                            (va, e0.0 & 0x3ff)
+                        })
+                })
+        })
+}
+
 pub fn activate(root: &PageTable) {
     let ppn = (root as *const PageTable as u64) >> 12;
     let satp = (8u64 << 60) | ppn;

crates/adapter-riscv/src/trap/mod.rs

@@ -123,6 +123,8 @@ const CAUSE_INTERRUPT_BIT: usize = 1 << 63;
 const CAUSE_S_TIMER: usize = 5;
 const CAUSE_S_EXTERNAL: usize = 9;
 const CAUSE_ECALL_U: usize = 8;
+/// Set in `sstatus` when the trap came from supervisor mode.
+const SSTATUS_SPP: usize = 1 << 8;
 
 pub fn init() {
     unsafe extern "C" {
@@ -347,6 +349,22 @@ fn handle_exception(frame: &mut TrapFrame, scause: usize) {
                 }
             }
         }
+        _ if frame.sstatus & SSTATUS_SPP != 0 => {
+            // The kernel itself faulted (a W^X violation, the stack guard,
+            // a bad pointer). Its state can no longer be trusted, and no
+            // user thread is to blame.
+            let uart = crate::hw::uart::Uart::new();
+            uart.write_str("KERNEL FAULT: scause=");
+            uart.write_hex(scause);
+            let stval: usize;
+            unsafe { core::arch::asm!("csrr {0}, stval", out(reg) stval) };
+            uart.write_str(" stval=");
+            uart.write_hex(stval);
+            uart.write_str(" sepc=");
+            uart.write_hex(frame.sepc);
+            uart.put_char(b'\n');
+            crate::halt(&uart, "fault in kernel mode");
+        }
         _ => {
             let uart = crate::hw::uart::Uart::new();
             uart.write_str("FAULT: scause=");

crates/domain/src/error.rs

@@ -31,6 +31,9 @@ pub enum KernelError {
     /// The target virtual range already holds a mapping.
     AddressInUse,
     InvalidRegion,
+    /// The region is mapped executable somewhere and the request would map
+    /// it writable, or the other way round (W^X across processes).
+    WritableAndExecutable,
     /// The object a capability would refer to no longer exists.
     NoSuchObject,
     /// The capability names the wrong kind of object for this operation.

crates/kernel/src/memory.rs

@@ -385,6 +385,18 @@ impl<P: Platform> Kernel<P> {
         if region.mappings().any(|m| m.process == pid) {
             return Err(KernelError::AddressInUse);
         }
+        // W^X holds across processes too: nobody may write what someone
+        // else can execute.
+        let clashes = |other: PagePerms| {
+            matches!(
+                (perms, other),
+                (PagePerms::ReadWrite, PagePerms::ReadExecute)
+                    | (PagePerms::ReadExecute, PagePerms::ReadWrite)
+            )
+        };
+        if region.mappings().any(|m| clashes(m.perms)) {
+            return Err(KernelError::WritableAndExecutable);
+        }
         if !Self::in_user_window(va, pages) {
             return Err(KernelError::InvalidAddress);
         }

crates/kernel/src/syscall/mod.rs

@@ -38,7 +38,9 @@ pub fn error_code(error: &KernelError) -> ErrorCode {
         InvalidSlot(_) | RevokedCapability(_) | NoSuchObject | InvalidEndpoint(_) => {
             ErrorCode::InvalidCap
         }
-        PermissionDenied { .. } | ProcessCreateDenied => ErrorCode::PermissionDenied,
+        PermissionDenied { .. } | ProcessCreateDenied | WritableAndExecutable => {
+            ErrorCode::PermissionDenied
+        }
         WrongObject => ErrorCode::WrongObject,
         NoFreeSlot => ErrorCode::NoFreeSlot,
         OutOfMemory => ErrorCode::OutOfMemory,

crates/kernel/src/tests/revocation.rs

@@ -355,6 +355,43 @@ fn a_region_can_be_remapped_after_its_mapping_was_revoked() {
     assert!(map(&mut t, tb, again, 0x50000, PagePerms::ReadWrite));
 }
 
+#[test]
+fn a_region_is_never_writable_in_one_process_and_executable_in_another() {
+    let mut t = Tree::new();
+    let (b, ta, tb, a_to_b) = (t.b, t.ta, t.tb, t.a_to_b);
+    let r = region(&mut t, ta, 1);
+    let in_b = t.give(ta, r, a_to_b, b);
+    assert!(map(&mut t, ta, r, 0x30000, PagePerms::ReadWrite));
+    assert!(!map(&mut t, tb, in_b, 0x50000, PagePerms::ReadExecute));
+    assert!(t.k.platform_mut().mapping(b, 0x50000).is_none());
+    // Read-only alongside either is fine.
+    assert!(map(&mut t, tb, in_b, 0x50000, PagePerms::ReadOnly));
+}
+
+#[test]
+fn code_can_be_written_then_made_executable_once_the_writer_unmaps() {
+    let mut t = Tree::new();
+    let (b, ta, tb, a_to_b) = (t.b, t.ta, t.tb, t.a_to_b);
+    let r = region(&mut t, ta, 1);
+    let writer = slot_of(t.run(
+        ta,
+        Syscall::Mint {
+            slot: r,
+            badge: Badge::NONE,
+            permissions: all(),
+        },
+    ));
+    assert!(map(&mut t, ta, writer, 0x30000, PagePerms::ReadWrite));
+    let in_b = t.give(ta, r, a_to_b, b);
+    assert!(!map(&mut t, tb, in_b, 0x50000, PagePerms::ReadExecute));
+    assert!(ok(&t.run(ta, Syscall::DropCap { slot: writer })));
+    assert!(map(&mut t, tb, in_b, 0x50000, PagePerms::ReadExecute));
+    assert!(
+        !map(&mut t, ta, r, 0x30000, PagePerms::ReadWrite),
+        "and nobody may write it while it is executable"
+    );
+}
+
 #[test]
 fn create_region_is_charged_to_the_caller() {
     let mut t = Tree::new();

crates/kernel/src/tests/storm.rs

@@ -15,6 +15,7 @@ use platform::Platform;
 use types::address::DataRef;
 use types::bounded::CapSlot;
 use types::id::ThreadId;
+use types::mapping::PagePerms;
 use types::syscall_abi::*;
 
 impl Kernel<MockPlatform> {
@@ -96,6 +97,10 @@ impl Kernel<MockPlatform> {
             }
         }
         for (_, region) in self.regions.iter() {
+            let has = |want| region.mappings().any(|m| m.perms == want);
+            if has(PagePerms::ReadWrite) && has(PagePerms::ReadExecute) {
+                return Err(String::from("region mapped both writable and executable"));
+            }
             for m in region.mappings() {
                 if let Some(via) = m.via
                     && (!self.resolves(via) || via.process != m.process)
@@ -308,6 +313,19 @@ fn new_round(k: &mut Kernel<MockPlatform>, init: ProcessId, init_thread: ThreadI
     }
 }
 
+/// The slots of every region capability `pid` holds.
+fn region_slots(k: &Kernel<MockPlatform>, pid: Option<ProcessId>) -> std::vec::Vec<CapSlot> {
+    let Some(process) = pid.and_then(|p| k.processes.get(p).ok()) else {
+        return std::vec::Vec::new();
+    };
+    process
+        .capabilities()
+        .iter()
+        .filter(|(_, c)| matches!(c.resource(), domain::resource::Resource::Memory(_)))
+        .map(|(s, _)| s)
+        .collect()
+}
+
 #[test]
 fn the_kernel_survives_a_syscall_storm() {
     let mut total = [0usize; 36 + 1];
@@ -357,7 +375,20 @@ fn the_kernel_survives_a_syscall_storm() {
             if is_init && (nr == SYS_EXIT || nr == SYS_EXIT_THREAD) {
                 continue;
             }
-            let args: [usize; 6] = core::array::from_fn(|_| rng.arg());
+            let mut args: [usize; 6] = core::array::from_fn(|_| rng.arg());
+            // Random arguments almost never name a region cap, a free
+            // address and valid perms at once; half the time, pick one of
+            // the caller's region caps (often the hub's shared region) so
+            // regions really are mapped into several processes.
+            if nr == SYS_MAP_REGION && rng.below(2) == 0 {
+                let regions = region_slots(&k, k.thread_process(tid));
+                if !regions.is_empty() {
+                    args[0] = regions[rng.below(regions.len())].value() as usize;
+                    args[1] = 0x10000 + rng.below(32) * 4096;
+                    // Read-write or read-execute: the pairs W^X forbids.
+                    args[2] = [PERMS_READ_WRITE, PERMS_READ_EXECUTE][rng.below(2)];
+                }
+            }
             // init keeps the capability it creates processes with.
             if is_init && nr == SYS_DROP_CAP && args[0] == 0 {
                 continue;

docs/ADAPTERS.md

@@ -13,7 +13,12 @@ This generates `crates/adapter-<arch>/` with the standard structure. Then:
 1. Add `"crates/adapter-<arch>"` to workspace `members` in `Cargo.toml`
 2. Fill in `board.rs` with device addresses for the target board
 3. Add boot assembly (`global_asm!` in `main.rs` or a `.S` file)
-4. Update `linker.ld` for the target's entry address and memory layout
+4. Update `linker.ld` for the target's entry address and memory layout.
+   Keep text, read-only data and writable data on separate pages and
+   export their page-aligned boundaries (`_text_start`, `_text_end`,
+   `_rodata_end`, `_stack_guard`, `_stack_bottom`): the kernel maps each
+   range with only the access it needs (W^X) and leaves a guard page under
+   the stack.
 5. Implement the TODOs in `hw/`, `mmu/`, `trap/`, and `platform.rs`
 
 ## Structure

docs/ipc.md

@@ -37,7 +37,11 @@ from a revocation further up.
 
 A region mapping made with `MapRegion(cap, va, perms)` lasts only as long
 as `cap`: deleting or revoking `cap` unmaps it. `perms` needs `Read`, plus
-`Write` for read-write or `Execute` for read-execute.
+`Write` for read-write or `Execute` for read-execute. W^X holds across
+processes: while a region is mapped read-write anywhere it cannot be mapped
+read-execute anywhere, and the other way round (`PermissionDenied`). To
+load code, map it writable, fill it, drop that mapping, then map it
+executable.
 
 | Object       | Permission | Allows                                     |
 |--------------|------------|--------------------------------------------|