Difference between revisions of "ARM binaries"

From WiiBrew
Jump to navigation Jump to search
m (Hallowizer moved page ARM Binaries to ARM binaries: sentence case)
(→‎ELF format: updated descriptions to not be as outdated)
 
(4 intermediate revisions by 3 users not shown)
Line 2: Line 2:
  
 
== ELF format ==
 
== ELF format ==
IOS modules, at least, use bare [[ELF]] files. The files seem to be compiled with GCC 3.4.3, and they are EABI compliant. The file is in big endian format. All ELF files are stripped and don't include function names or symbols.
+
Individual IOS modules use bare [[ELF]] files. The files seem to be compiled with GCC 3.4.3, and they are EABI compliant. The file is in big endian format. All ELF files are stripped and don't include function names or symbols.
  
 
=== ELF Note Structure ===
 
=== ELF Note Structure ===
IOS modules are loaded by the kernel (syscall load_module) according to an ELF note. There are exactly as many IOSELFNoteData as threads created by IOS.
+
IOS modules in the main binary are loaded by the main kernel thread according to an ELF note. There are exactly as many IOSELFNoteData as process main threads in IOS.
  
 
<source lang="c">
 
<source lang="c">
Line 75: Line 75:
 
|}
 
|}
  
=== BOOT2 elf loader ===
+
=== Official ELF loaders ===
 +
The [[boot2]] and [[IOS]] elf stub loaders seem to be position-independent: they can be loaded at any address and will still work, as long as it doesn't overlap with the destination of the ELF load.
  
The BOOT2 elf stub loader sets up a stack, calculates its own address, and switches to THUMB mode. Then it does the following:
+
The official ELF loaders set up a stack and calculate their own address, switches to THUMB mode, and then they do the following:
  
 
<source lang="c">
 
<source lang="c">
Line 85: Line 86:
 
</source>
 
</source>
  
0xD800000 seems to be the start of the (a?) hardware register space.
+
This sets the [[Hollywood/Registers#HW_SRNPROT|HW_SRNPROT]] register to enable the [[Starlet/Main Memory|SRAM]] mirror at 0xFFFE0000.
  
After this, it loads the ELF file, and then '''zeroes out the memory area where the ELF file resides'''. Then it goes back to ARM mode and vectors to 0xFFFF0000 (the entrypoint of the ARM / vector table). The entire BOOT2 code seems to be position-independent: it can be loaded at any address and will still work, as long as it doesn't overlap with the destination of the ELF load. The entire BOOT2 file cleartext is loaded and then the loader is called, so the loader can calculate the offset of the header simply by subtracting 0x10 from the PC at its entrypoint.
+
After this, it loads the ELF file, and then '''zeroes out the memory area where most of the ELF loader resides'''. Then it goes back to ARM mode and vectors to 0xFFFF0000 (the entrypoint of the ARM / vector table). The entire BOOT2 file cleartext is loaded and then the loader is called, so the loader can calculate the offset of the header simply by subtracting 0x10 from the PC at its entrypoint.
  
 
=== Embedded Broadway Code ===
 
=== Embedded Broadway Code ===
Line 93: Line 94:
  
 
=== Dynamic Linker ===
 
=== Dynamic Linker ===
In later IOS versions ([[IOS28]] and later) the single IOS ARM binary was devided into several modules/libraries. The modules are loaded dynamically. The code is statically linked to a fixed address. Each module can register driver entry points at the operating systems. Functions of other modules are not directly called. Syscalls are used to communicate with other modules. The calls are forwarded to the approperiate module function.
+
In later IOS versions ([[IOS28]] and later) the single monolithic IOS ARM binary was split into individual dynamically loaded modules. The main kernel is the only file to include the ELFLOADER header and stub; each individual module is a regular ELF file, loaded dynamically by the kernel to a fixed address specified in the ELF header. No functions are called directly between modules - [[IOS/Syscalls#Syscalls_.28via_undefined_instructions.29|Syscalls]] are used for individual modules to communicate with the main kernel, and resource managers are used for modules to communicate between each other (such as [[:/dev/di|DI]] interfacing with [[:/dev/es]] for ticket verification)
  
 
=== Extract ELF file ===
 
=== Extract ELF file ===

Latest revision as of 08:02, 23 September 2022

ARM binaries are contained inside .wad files in update partitions. The .wad format and how to decrypt it is described in WAD Files.

ELF format

Individual IOS modules use bare ELF files. The files seem to be compiled with GCC 3.4.3, and they are EABI compliant. The file is in big endian format. All ELF files are stripped and don't include function names or symbols.

ELF Note Structure

IOS modules in the main binary are loaded by the main kernel thread according to an ELF note. There are exactly as many IOSELFNoteData as process main threads in IOS.

struct IOSELFNoteData
{
  u32 version_maybe; // always 0x0b?
  u32 pid;
  u32 unk2; // unused
  u32 entry_point;
  u32 unk4; // unused
  u32 priority;
  u32 unk6; // unused
  u32 stack_size;
  u32 unk8; // unused
  u32 stack_top;
};

struct IOSELFNote
{
  // Standard ELF note header
  u32 name_size; // always 0x0
  u32 descriptor_size;
  u32 note_type; // always 0x6

  // IOS specific data
  IOSELFNoteData data[];
};

ELFLOADER format

The ELFLOADER ARM binary format is used for the "bootup" files, including the IOS kernel (or the entirety of the IOS in earlier versions which are monolithic) and BOOT2. Once decrypted, the data has the following format:

Start End Length Description
0x000 0x004 0x004 Header size = 0x0010
0x004 0x008 0x004 Offset to ELF file after header
0x008 0x00C 0x004 Size of ELF file
0x00C 0x010 0x004 Unknown "ddrInit", set to 0 on boot
0x010 variable variable ELF file stub loader binary
variable variable variable ELF file

Official ELF loaders

The boot2 and IOS elf stub loaders seem to be position-independent: they can be loaded at any address and will still work, as long as it doesn't overlap with the destination of the ELF load.

The official ELF loaders set up a stack and calculate their own address, switches to THUMB mode, and then they do the following:

 if( ! (*((u32 *)0xD800060) & 0x20) ) {
   *((u32 *)0xD800060) |= 0x20;
 }

This sets the HW_SRNPROT register to enable the SRAM mirror at 0xFFFE0000.

After this, it loads the ELF file, and then zeroes out the memory area where most of the ELF loader resides. Then it goes back to ARM mode and vectors to 0xFFFF0000 (the entrypoint of the ARM / vector table). The entire BOOT2 file cleartext is loaded and then the loader is called, so the loader can calculate the offset of the header simply by subtracting 0x10 from the PC at its entrypoint.

Embedded Broadway Code

Some ARM binaries include PowerPC code for the Broadway. For example MIOS include code at address 0x00003400 and 0x01300000. The first code sets the BATs up, initializes the powerpc and vectors to the second code (which can be addressed at 0x81300000 then) then. This code is executed by writing a small stub to the EXI boot buffer and setting a bit in another register then.

Dynamic Linker

In later IOS versions (IOS28 and later) the single monolithic IOS ARM binary was split into individual dynamically loaded modules. The main kernel is the only file to include the ELFLOADER header and stub; each individual module is a regular ELF file, loaded dynamically by the kernel to a fixed address specified in the ELF header. No functions are called directly between modules - Syscalls are used for individual modules to communicate with the main kernel, and resource managers are used for modules to communicate between each other (such as DI interfacing with /dev/es for ticket verification)

Extract ELF file

The following program extracts the ELF from the ARM binary. Normal ELF tools can handle the generated output.

#include <stdio.h>
#include <stdint.h>
#include <stdlib.h>
#include <netinet/in.h>

/** Header for Wii ARM binaries. */
typedef struct {
	/** Size of this header. */
	uint32_t headerSize;
	/** Offset to ELF file. */
	uint32_t offset;
	/** Size of ELF file. */
	uint32_t size;
	/** Padded with zeroes. */
	uint32_t resevered;
} arm_binary_header_t;

int main(int argc, char *argv[])
{
	const char *inFilename;
	const char *outFilename;
	FILE *fin = NULL;
	FILE *fout = NULL;
	arm_binary_header_t header;
	char *buffer = NULL;

	if (argc != 3) {
		fprintf(stderr, "%s: [Wii ARM Binary] [ELF output file]\n\n", argv[0]);
		fprintf(stderr, "Extract ELF from Wii ARM binary.\n");
		fprintf(stderr, "Error: Parameter wrong.\n");
		return -1;
	}
	inFilename = argv[1];
	outFilename = argv[2];

	fin = fopen(inFilename, "rb");
	if (fin == NULL) {
		fprintf(stderr, "Error: Failed to open input file \"%s\".\n", inFilename);
		return -2;
	}
	if (fread(&header, sizeof(header), 1, fin) != 1) {
		fclose(fin);
		fprintf(stderr, "Error: Input file \"%s\" is too small.\n", inFilename);
		return -3;
	}
	header.headerSize = ntohl(header.headerSize);
	header.offset = ntohl(header.offset);
	header.size = ntohl(header.size);

	if (header.headerSize != sizeof(arm_binary_header_t)) {
		fclose(fin);
		fprintf(stderr, "Error: Input file \"%s\" is not a Wii ARM binary.\n", inFilename);
		return -4;
	}
	if (fseek(fin, header.offset, SEEK_CUR) != 0) {
		fclose(fin);
		fprintf(stderr, "Error: Input file \"%s\" is too small (seek to %d failed).\n", inFilename, header.offset);
		return -5;
	}
	buffer = malloc(header.size);
	if (buffer == NULL) {
		fclose(fin);
		fprintf(stderr, "Error: Out of memory.\n");
		return -6;
	}
	if (fread(buffer, header.size, 1, fin) != 1) {
		fclose(fin);
		fprintf(stderr, "Error: Input file \"%s\" is too small.\n", inFilename);
		return -7;
	}
	fclose(fin);

	fout = fopen(outFilename, "wb");
	if (fout == NULL) {
		fclose(fin);
		fprintf(stderr, "Error: Failed to open output file \"%s\".\n", outFilename);
		return -8;
	}
	if (fwrite(buffer, header.size, 1, fout) != 1) {
		fclose(fin);
		fprintf(stderr, "Error: Output file \"%s\" write error (disc full?).\n", outFilename);
		return -9;
	}
	fclose(fout);
	free(buffer);
	return 0;
}