crt0/dynamic: refactoring + add support for reprotecting relro segment

nx/source/runtime/dynamic.c

@@ -1,34 +1,42 @@
 #include "result.h"
+#include "kernel/svc.h"
 #include "runtime/diag.h"
 #include <elf.h>
+#include <string.h>
 
-void __nx_dynamic(uintptr_t base, const Elf64_Dyn* dyn)
+typedef struct Mod0Header {
+	u32 magic_mod0;
+	s32 dyn_offset;
+	s32 bss_start_offset;
+	s32 bss_end_offset;
+	s32 eh_frame_hdr_start_offset;
+	s32 eh_frame_hdr_end_offset;
+	s32 unused;
+
+	u32 magic_lny0;
+	s32 got_start_offset;
+	s32 got_end_offset;
+
+	u32 magic_lny1;
+	s32 relro_start_offset;
+	s32 relro_end_offset;
+} Mod0Header;
+
+NX_INLINE void* _dynResolveOffset(const Mod0Header* mod0, s32 offset)
 {
-	const Elf64_Rela* rela = NULL;
-	u64 relasz = 0;
+	return (void*)((uintptr_t)mod0 + offset);
+}
 
-	for (; dyn->d_tag != DT_NULL; dyn++)
-	{
-		switch (dyn->d_tag)
-		{
-			case DT_RELA:
-				rela = (const Elf64_Rela*)(base + dyn->d_un.d_ptr);
+static void _dynProcessRela(uintptr_t base, const Elf64_Rela* rela, size_t relasz)
+{
+	for (; relasz--; rela++) {
+		switch (ELF64_R_TYPE(rela->r_info)) {
+			default: {
+				diagAbortWithResult(MAKERESULT(Module_Libnx, LibnxError_BadReloc));
 				break;
-			case DT_RELASZ:
-				relasz = dyn->d_un.d_val / sizeof(Elf64_Rela);
-				break;
-		}
-	}
+			}
 
-	if (rela == NULL)
-		diagAbortWithResult(MAKERESULT(Module_Libnx, LibnxError_BadReloc));
-
-	for (; relasz--; rela++)
-	{
-		switch (ELF64_R_TYPE(rela->r_info))
-		{
-			case R_AARCH64_RELATIVE:
-			{
+			case R_AARCH64_RELATIVE: {
 				u64* ptr = (u64*)(base + rela->r_offset);
 				*ptr = base + rela->r_addend;
 				break;
@@ -36,3 +44,57 @@ void __nx_dynamic(uintptr_t base, const Elf64_Dyn* dyn)
 		}
 	}
 }
+
+void __nx_dynamic(uintptr_t base, const Mod0Header* mod0)
+{
+	// Return early if MOD0 header has been invalidated
+	if (mod0->magic_mod0 != 0x30444f4d) { // MOD0
+		return;
+	}
+
+	// Clear the BSS area
+	u8* bss_start = _dynResolveOffset(mod0, mod0->bss_start_offset);
+	u8* bss_end = _dynResolveOffset(mod0, mod0->bss_end_offset);
+	if (bss_start != bss_end) {
+		memset(bss_start, 0, bss_end - bss_start);
+	}
+
+	// Retrieve pointer to the ELF dynamic section
+	const Elf64_Dyn* dyn = _dynResolveOffset(mod0, mod0->dyn_offset);
+
+	// Extract relevant information from the ELF dynamic section
+	const Elf64_Rela* rela = NULL;
+	size_t relasz = 0;
+	for (; dyn->d_tag != DT_NULL; dyn++) {
+		switch (dyn->d_tag) {
+			case DT_RELA:
+				rela = (const Elf64_Rela*)(base + dyn->d_un.d_ptr);
+				break;
+
+			case DT_RELASZ:
+				relasz = dyn->d_un.d_val / sizeof(Elf64_Rela);
+				break;
+		}
+	}
+
+	// Apply RELA relocations if present
+	if (rela && relasz) {
+		_dynProcessRela(base, rela, relasz);
+	}
+
+	// Return early if LNY0/LNY1 extensions are not present
+	if (mod0->magic_lny0 != 0x30594e4c || mod0->magic_lny1 != 0x31594e4c) { // LNY0, LNY1
+		return;
+	}
+
+	// Reprotect relro segment as read-only now that we're done processing relocations
+	u8* relro_start = _dynResolveOffset(mod0, mod0->relro_start_offset);
+	size_t relro_sz = (u8*)_dynResolveOffset(mod0, mod0->relro_end_offset) - relro_start;
+	Result rc = svcSetMemoryPermission(relro_start, relro_sz, Perm_R);
+	if (R_FAILED(rc)) {
+		diagAbortWithResult(rc);
+	}
+
+	// Lock the relro segment's permissions
+	svcSetMemoryAttribute(relro_start, relro_sz, MemAttr_IsPermissionLocked, MemAttr_IsPermissionLocked);
+}

nx/source/runtime/switch_crt0.s

@@ -23,36 +23,20 @@ _start:
     b    __libnx_exception_entry
 
 .Lcrt0_main_entry:
-    // Get pointer to MOD0 struct (contains offsets to important places)
-    adr x28, __nx_mod0
-
-    // Calculate BSS address/size
-    ldp  w8, w9, [x28, #8] // load BSS start/end offset from MOD0
-    sub  w9, w9, w8        // calculate BSS size
-    add  w9, w9, #7        // round up to 8
-    bic  w9, w9, #7        // ^
-    add  x8, x28, x8       // fixup the start pointer
-
-    // Clear the BSS in 8-byte units
-1:  subs w9, w9, #8
-    str  xzr, [x8], #8
-    bne  1b
-
     // Preserve registers across function calls
     mov x25, x0  // entrypoint argument 0
     mov x26, x1  // entrypoint argument 1
     mov x27, x30 // loader return address
+    mov x28, sp  // initial stack pointer
+
+    // Perform runtime linking on ourselves (including relocations)
+    adr  x0, _start    // get aslr base
+    adr  x1, __nx_mod0 // get pointer to MOD0 struct
+    bl   __nx_dynamic
 
     // Save initial stack pointer
-    mov  x8, sp
     adrp x9, __stack_top
-    str  x8, [x9, #:lo12:__stack_top]
-
-    // Parse ELF .dynamic section (which applies relocations to our module)
-    adr  x0, _start    // get aslr base
-    ldr  w1, [x28, #4] // pointer to .dynamic section
-    add  x1, x28, x1
-    bl   __nx_dynamic
+    str  x28, [x9, #:lo12:__stack_top]
 
     // Perform system initialization
     mov  x0, x25
@@ -65,8 +49,8 @@ _start:
     ldr  w0, [x0, #:lo12:__system_argc]
     adrp x1, __system_argv // argv
     ldr  x1, [x1, #:lo12:__system_argv]
-    adrp x30, exit
-    add  x30, x30, #:lo12:exit
+    adrp x30, :got:exit
+    ldr  x30, [x30, #:got_lo12:exit]
     b    main
 
 .global __nx_exit
@@ -95,6 +79,10 @@ __nx_mod0:
     .word  __got_start__        - __nx_mod0
     .word  __got_end__          - __nx_mod0
 
+    .ascii "LNY1"
+    .word  __relro_start        - __nx_mod0
+    .word  __data_start         - __nx_mod0
+
 .section .bss.__stack_top, "aw", %nobits
 .global __stack_top
 .align 3