runtime: Validate ksupport ELF against hard-coded address ranges

This would have caught the reduction in header padding with LLD 14.
In theory, we could just get rid of the hard-coded kernel CPU address
ranges altogether and use ksupport.elf as the one source of truth; the
code already exists in dyld. The actual base address of the file would
still need to be forwarded to the kernel-side libunwind glue, though,
as there doesn't seem to be a clean way to get the equivalent of
KSUPPORT_HEADER_SIZE through the linker script. I have left this as-is
with the hard-coded KERNELCPU_… constants for now.

artiq/firmware/Cargo.lock

@@ -320,6 +320,7 @@ dependencies = [
  "build_misoc",
  "byteorder",
  "cslice",
+ "dyld",
  "eh",
  "failure",
  "failure_derive",

artiq/firmware/libdyld/lib.rs

@@ -5,7 +5,7 @@ use elf::*;
 
 pub mod elf;
 
-fn read_unaligned<T: Copy>(data: &[u8], offset: usize) -> Result<T, ()> {
+pub fn read_unaligned<T: Copy>(data: &[u8], offset: usize) -> Result<T, ()> {
     if data.len() < offset + mem::size_of::<T>() {
         Err(())
     } else {
@@ -75,6 +75,25 @@ impl<'a> fmt::Display for Error<'a> {
     }
 }
 
+pub fn is_elf_for_current_arch(ehdr: &Elf32_Ehdr, e_type: u16) -> bool {
+    const IDENT: [u8; EI_NIDENT] = [
+        ELFMAG0,    ELFMAG1,     ELFMAG2,    ELFMAG3,
+        ELFCLASS32, ELFDATA2LSB, EV_CURRENT, ELFOSABI_NONE,
+        /* ABI version */ 0, /* padding */ 0, 0, 0, 0, 0, 0, 0
+    ];
+    if ehdr.e_ident != IDENT { return false; }
+
+    if ehdr.e_type != e_type { return false; }
+
+    #[cfg(target_arch = "riscv32")]
+    const ARCH: u16 = EM_RISCV;
+    #[cfg(not(target_arch = "riscv32"))]
+    const ARCH: u16 = EM_NONE;
+    if ehdr.e_machine != ARCH { return false; }
+
+    true
+}
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum Arch {
     RiscV,
@@ -268,25 +287,16 @@ impl<'a> Library<'a> {
         let ehdr = read_unaligned::<Elf32_Ehdr>(data, 0)
                                   .map_err(|()| "cannot read ELF header")?;
 
-        const IDENT: [u8; EI_NIDENT] = [
-            ELFMAG0,    ELFMAG1,     ELFMAG2,    ELFMAG3,
-            ELFCLASS32, ELFDATA2LSB, EV_CURRENT, ELFOSABI_NONE,
-            /* ABI version */ 0, /* padding */ 0, 0, 0, 0, 0, 0, 0
-        ];
-
-        #[cfg(target_arch = "riscv32")]
-        const ARCH: u16 = EM_RISCV;
-        #[cfg(not(target_arch = "riscv32"))]
-        const ARCH: u16 = EM_NONE;
+        if !is_elf_for_current_arch(&ehdr, ET_DYN) {
+            return Err("not a shared library for current architecture")?
+        }
 
         #[cfg(all(target_feature = "f", target_feature = "d"))]
         const FLAGS: u32 = EF_RISCV_FLOAT_ABI_DOUBLE;
-
         #[cfg(not(all(target_feature = "f", target_feature = "d")))]
         const FLAGS: u32 = EF_RISCV_FLOAT_ABI_SOFT;
-
-        if ehdr.e_ident != IDENT || ehdr.e_type != ET_DYN || ehdr.e_machine != ARCH || ehdr.e_flags != FLAGS {
-            return Err("not a shared library for current architecture")?
+        if ehdr.e_flags != FLAGS {
+            return Err("unexpected flags for shared library (wrong floating point ABI?)")?
         }
 
         let mut dyn_off = None;

artiq/firmware/runtime/Cargo.toml

@@ -19,6 +19,7 @@ byteorder = { version = "1.0", default-features = false }
 cslice = { version = "0.3" }
 log = { version = "0.4", default-features = false }
 managed = { version = "^0.7.1", default-features = false, features = ["alloc", "map"] }
+dyld = { path = "../libdyld" }
 eh = { path = "../libeh" }
 unwind_backtrace = { path = "../libunwind_backtrace" }
 io = { path = "../libio", features = ["byteorder"] }

artiq/firmware/runtime/kernel.rs

@@ -1,5 +1,5 @@
-use core::ptr;
 use board_misoc::csr;
+use core::{ptr, slice};
 use mailbox;
 use rpc_queue;
 
@@ -13,15 +13,20 @@ pub unsafe fn start() {
 
     stop();
 
-    extern {
+    extern "C" {
         static _binary____ksupport_ksupport_elf_start: u8;
         static _binary____ksupport_ksupport_elf_end: u8;
     }
-    let ksupport_start = &_binary____ksupport_ksupport_elf_start as *const _;
-    let ksupport_end   = &_binary____ksupport_ksupport_elf_end as *const _;
-    ptr::copy_nonoverlapping(ksupport_start,
-                             (KERNELCPU_EXEC_ADDRESS - KSUPPORT_HEADER_SIZE) as *mut u8,
-                             ksupport_end as usize - ksupport_start as usize);
+    let ksupport_elf_start = &_binary____ksupport_ksupport_elf_start as *const u8;
+    let ksupport_elf_end = &_binary____ksupport_ksupport_elf_end as *const u8;
+    let ksupport_elf = slice::from_raw_parts(
+        ksupport_elf_start,
+        ksupport_elf_end as usize - ksupport_elf_start as usize,
+    );
+
+    if let Err(msg) = load_image(&ksupport_elf) {
+        panic!("failed to load kernel CPU image (ksupport.elf): {}", msg);
+    }
 
     csr::kernel_cpu::reset_write(0);
 
@@ -41,6 +46,44 @@ pub unsafe fn stop() {
     rpc_queue::init();
 }
 
+/// Loads the given image for execution on the kernel CPU.
+///
+/// The entire image including the headers is copied into memory for later use by libunwind, but
+/// placed such that the text section ends up at the right location in memory. Currently, we just
+/// hard-code the address range, but at least verify that this matches the ELF program header given
+/// in the image (avoids loading the – non-relocatable – code at the wrong address on toolchain/…
+/// changes).
+unsafe fn load_image(image: &[u8]) -> Result<(), &'static str> {
+    use dyld::elf::*;
+    use dyld::{is_elf_for_current_arch, read_unaligned};
+
+    let ehdr = read_unaligned::<Elf32_Ehdr>(image, 0).map_err(|()| "could not read ELF header")?;
+
+    // The check assumes the two CPUs share the same architecture. This is just to avoid inscrutable
+    // errors; we do not functionally rely on this.
+    if !is_elf_for_current_arch(&ehdr, ET_EXEC) {
+        return Err("not an executable for kernel CPU architecture");
+    }
+
+    // First program header should be the main text/… LOAD (see ksupport.ld).
+    let phdr = read_unaligned::<Elf32_Phdr>(image, ehdr.e_phoff as usize)
+        .map_err(|()| "could not read program header")?;
+    if phdr.p_type != PT_LOAD {
+        return Err("unexpected program header type");
+    }
+    if phdr.p_vaddr + phdr.p_memsz > KERNELCPU_LAST_ADDRESS as u32 {
+        // This is a weak sanity check only; we also need to fit in the stack, etc.
+        return Err("too large for kernel CPU address range");
+    }
+    const TARGET_ADDRESS: u32 = (KERNELCPU_EXEC_ADDRESS - KSUPPORT_HEADER_SIZE) as _;
+    if phdr.p_vaddr - phdr.p_offset != TARGET_ADDRESS {
+        return Err("unexpected load address/offset");
+    }
+
+    ptr::copy_nonoverlapping(image.as_ptr(), TARGET_ADDRESS as *mut u8, image.len());
+    Ok(())
+}
+
 pub fn validate(ptr: usize) -> bool {
     ptr >= KERNELCPU_EXEC_ADDRESS && ptr <= KERNELCPU_LAST_ADDRESS
 }

artiq/firmware/runtime/main.rs

@@ -1,6 +1,7 @@
 #![feature(lang_items, panic_info_message, const_btree_new, iter_advance_by)]
 #![no_std]
 
+extern crate dyld;
 extern crate eh;
 #[macro_use]
 extern crate alloc;