nintendoswitch: Add dynamic loader for runtime loading PIE sections

builder/objcopy.go

@@ -9,6 +9,10 @@ import (
 	"github.com/marcinbor85/gohex"
 )
 
+// maxPadBytes is the maximum allowed bytes to be padded in a rom extraction
+// this value is currently defined by Nintendo Switch Page Alignment (4096 bytes)
+const maxPadBytes = 4095
+
 // objcopyError is an error returned by functions that act like objcopy.
 type objcopyError struct {
 	Op  string
@@ -70,7 +74,12 @@ func extractROM(path string) (uint64, []byte, error) {
 	var rom []byte
 	for _, prog := range progs {
 		if prog.Paddr != progs[0].Paddr+uint64(len(rom)) {
-			return 0, nil, objcopyError{"ROM segments are non-contiguous: " + path, nil}
+			diff := prog.Paddr - (progs[0].Paddr + uint64(len(rom)))
+			if diff > maxPadBytes {
+				return 0, nil, objcopyError{"ROM segments are non-contiguous: " + path, nil}
+			}
+			// Pad the difference
+			rom = append(rom, make([]byte, diff)...)
 		}
 		data, err := ioutil.ReadAll(prog.Open())
 		if err != nil {

src/runtime/dynamic_arm64.go

@@ -0,0 +1,55 @@
+package runtime
+
+import (
+	"debug/elf"
+	"unsafe"
+)
+
+const debugLoader = false
+
+//export __dynamic_loader
+func dynamicLoader(base uintptr, dyn *elf.Dyn64) {
+	var rela *elf.Rela64
+	relasz := uint64(0)
+
+	if debugLoader {
+		println("ASLR Base: ", base)
+	}
+
+	for dyn.Tag != int64(elf.DT_NULL) {
+		switch elf.DynTag(dyn.Tag) {
+		case elf.DT_RELA:
+			rela = (*elf.Rela64)(unsafe.Pointer(base + uintptr(dyn.Val)))
+		case elf.DT_RELASZ:
+			relasz = uint64(dyn.Val) / uint64(unsafe.Sizeof(elf.Rela64{}))
+		}
+
+		ptr := uintptr(unsafe.Pointer(dyn))
+		ptr += unsafe.Sizeof(elf.Dyn64{})
+		dyn = (*elf.Dyn64)(unsafe.Pointer(ptr))
+	}
+
+	if rela == nil {
+		runtimePanic("bad reloc")
+	}
+	if rela == nil {
+		runtimePanic("bad reloc")
+	}
+
+	if debugLoader {
+		println("Sections to load: ", relasz)
+	}
+
+	for relasz > 0 && rela != nil {
+		switch elf.R_AARCH64(rela.Info) {
+		case elf.R_AARCH64_RELATIVE:
+			ptr := (*uint64)(unsafe.Pointer(base + uintptr(rela.Off)))
+			*ptr = uint64(base + uintptr(rela.Addend))
+		}
+
+		rptr := uintptr(unsafe.Pointer(rela))
+		rptr += unsafe.Sizeof(elf.Rela64{})
+		rela = (*elf.Rela64)(unsafe.Pointer(rptr))
+		relasz--
+	}
+}

src/runtime/runtime_nintendoswitch.go

@@ -2,6 +2,8 @@
 
 package runtime
 
+import "unsafe"
+
 type timeUnit int64
 
 const asyncScheduler = false
@@ -60,6 +62,16 @@ func abort() {
 	}
 }
 
+//export write
+func write(fd int32, buf *byte, count int) int {
+	// TODO: Proper handling write
+	for i := 0; i < count; i++ {
+		putchar(*buf)
+		buf = (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(buf)) + 1))
+	}
+	return count
+}
+
 //export sleepThread
 func sleepThread(nanos uint64)
 

targets/nintendoswitch.json

@@ -20,6 +20,10 @@
     "-fno-exceptions", "-fno-unwind-tables",
     "-ffunction-sections", "-fdata-sections"
   ],
+  "ldflags": [
+    "-pie",
+    "-z", "notext"
+  ],
   "linkerscript": "targets/nintendoswitch.ld",
   "extra-files": [
     "targets/nintendoswitch.s",

targets/nintendoswitch.ld

@@ -1,69 +1,84 @@
+PHDRS
+{
+  text    PT_LOAD FLAGS(5) /* Read | Execute */;
+  rodata  PT_LOAD FLAGS(4) /* Read */;
+  data    PT_LOAD FLAGS(6) /* Read | Write */;
+  bss     PT_LOAD FLAGS(6) /* Read | Write */;
+  dynamic PT_DYNAMIC;
+}
+
 SECTIONS
 {
+  /* Code and file header */
   . = 0;
 
-  /* Code and file header */
-
-  .text : {
+  .text : ALIGN(0x1000) {
     HIDDEN(__text_start = .);
     KEEP(*(.text.jmp))
 
     . = 0x80;
 
     *(.text .text.*)
+    *(.plt .plt.*)
 
-    . = ALIGN(0x1000);
     HIDDEN(__text_end = .);
     HIDDEN(__text_size = . - __text_start);
   }
 
   /* Read-only sections */
+  . = ALIGN(0x1000);
 
-  .rodata : {
-    HIDDEN(__rodata_start = .);
-
-    *(.rodata .rodata.*)
-
-    *(.got)
+  HIDDEN(__rodata_start = .);
+  .rodata : { *(.rodata .rodata.*) }  :rodata
 
+  .mod0 : {
     KEEP(crt0.nso.o(.data.mod0))
     KEEP(crt0.nro.o(.data.mod0))
     KEEP(crt0.lib.nro.o(.data.mod0))
-    KEEP(*(.data.mod0))
-
-    HIDDEN(__dynamic_start = .);
-    *(.dynamic)
-
-    . = ALIGN(0x1000);
-    HIDDEN(__rodata_end = .);
-    HIDDEN(__rodata_size = . - __rodata_start);
   }
 
+  .dynsym            : { *(.dynsym) } :rodata
+  .dynstr            : { *(.dynstr) } :rodata
+  .rela.dyn          : { *(.rela.*) } :rodata
+
+  HIDDEN(__rodata_end = .);
+  HIDDEN(__rodata_size = . - __rodata_start);
+
   /* Read-write sections */
+  . = ALIGN(0x1000);
 
   .data : {
     HIDDEN(__data_start = .);
 
     *(.data .data.*)
+    *(.got .got.*)
+    *(.got.plt .got.plt.*)
 
     HIDDEN(__data_end = .);
     HIDDEN(__data_size = . - __data_start);
+  } :data
+
+  .dynamic : {
+    HIDDEN(__dynamic_start = .);
+    *(.dynamic)
   }
 
   /* BSS section */
-
+  . = ALIGN(0x1000);
   .bss : {
     HIDDEN(__bss_start = .);
 
     *(.bss .bss.*)
     *(COMMON)
+    . = ALIGN(8);
 
     HIDDEN(__bss_end = .);
     HIDDEN(__bss_size = . - __bss_start);
-  }
+  } : bss
 
   /DISCARD/ :
   {
     *(.eh_frame) /* This is probably unnecessary and bloats the binary. */
+    *(.eh_frame_hdr)
   }
 }

targets/nintendoswitch.s

@@ -6,8 +6,7 @@
 _start:
     b start
     .word _mod_header - _start
-    .word 0
-    .word 0
+    .ascii "HOMEBREW"
 
     .ascii "NRO0"              // magic
     .word 0                    // version (always 0)
@@ -15,11 +14,11 @@ _start:
     .word 0                    // flags (unused)
 
     // segment headers
-    .word __text_start
+    .word 0 // __text_start
     .word __text_size
-    .word __rodata_start
+    .word 0 //__rodata_start
     .word __rodata_size
-    .word __data_start
+    .word 0 //__data_start
     .word __data_size
     .word __bss_size
     .word 0
@@ -45,18 +44,17 @@ _mod_header:
 .section .text, "x"
 .global start
 start:
-
-    // save lr
-    mov  x7, x30
-
-    // get aslr base
-    bl   +4
-    sub  x6, x30, #0x88
+    // Get ASLR Base
+    adrp x6, _start
 
     // context ptr and main thread handle
     mov  x5, x0
     mov  x4, x1
 
+    // Save lr, context pointer, main thread handler
+    adrp x0, _aslr_base
+    str  x6, [x0, #:lo12:_aslr_base]
+
     // clear .bss
     adrp x5, __bss_start
     add x5, x5, #:lo12:__bss_start
@@ -71,7 +69,27 @@ bssloop:
     b bssloop
 
 run:
+    // process .dynamic section
+    adrp x0, _aslr_base
+    ldr  x0, [x0, #:lo12:_aslr_base]
+    adrp x1, _DYNAMIC
+    add  x1, x1, #:lo12:_DYNAMIC
+    bl   __dynamic_loader
+
+    // set LR to svcExitProcess if it's null
+    adrp x3, exit
+    add x3, x3, #:lo12:exit
+    cmp x30, xzr
+    csel x30, x3, x30, eq
+
     // call entrypoint
-    adrp x30, exit
-    add  x30, x30, #:lo12:exit
+    mov x3, sp
+    sub sp, sp, 0x10
+    stp x29, x30, [sp]
     b    main
+
+.section .data.horizon
+.align 8
+.global _aslr_base // Placeholder for ASLR Base Address
+_aslr_base:
+    .dword 0