NSO is the main executable format.

It starts with the "NSO" header and mainly describes .text, .rodata, and .data segments (like a short-form of ELF program headers). If the segments are compressed, they are compressed using LZ4.

Header

Offset Size Description
0x0 4 Magic "NSO0"
0x4 4 Version (always 0)
0x8 4 Reserved
0xC 4 Flags
0x10 0xC SegmentHeader for .text
0x1C 0x4 ModuleNameOffset (calculated by sizeof(header))
0x20 0xC SegmentHeader for .rodata
0x2C 0x4 ModuleNameSize
0x30 0xC SegmentHeader for .data
0x3C 0x4 BssSize
0x40 0x20 ModuleId
0x60 0x4 TextFileSize (.text compressed size)
0x64 0x4 RoFileSize (.rodata compressed size)
0x68 0x4 DataFileSize (.data compressed size)
0x6C 0x1C Reserved
0x88 0x8 SegmentHeaderRelative for .api_info
0x90 0x8 SegmentHeaderRelative for .dynstr
0x98 0x8 SegmentHeaderRelative for .dynsym
0xA0 0x20 TextHash (SHA-256 hash over the decompressed .text section using the above size)
0xC0 0x20 RoHash (SHA-256 hash over the decompressed .rodata section using the above size)
0xE0 0x20 DataHash (SHA-256 hash over the decompressed .data section using the above size)
0x100 Variable Compressed sections

Most data in Switch binaries are standard ELF structures, however some are custom. For example, the MOD header is essentially a replacement for a PT_DYNAMIC program header.

Flags

Bits Description
0 TextCompress (.text section is compressed)
1 RoCompress (.rodata section is compressed)
2 DataCompress (.data section is compressed)
3 TextHash (.text hash must be checked when loading)
4 RoHash (.rodata hash must be checked when loading)
5 DataHash (.data hash must be checked when loading)

SegmentHeader

Offset Size Description
0x0 0x4 FileOffset
0x4 0x4 MemoryOffset
0x8 0x4 Size (decompressed)

ModuleId

Value of "build id" from ELF's GNU .note section. Contains variable sized digest, up to 32bytes.

SegmentHeaderRelative

Offset Size Description
0x0 4 Offset
0x4 4 Size

Offset is relative to the .rodata section.

MOD

Offset Size Description
0x00 4 Reserved
0x04 4 MagicOffset (always 8, so it works when MOD is at image_base + 0)
0x08 4 Magic "MOD0"
0x0C 4 .dynamic offset
0x10 4 .bss start offset
0x14 4 .bss end offset
0x18 4 .eh_frame_hdr start offset
0x1C 4 .eh_frame_hdr end offset
0x20 4 Offset to runtime-generated module object (typically equal to .bss base)

All offsets are signed 32bit values relative to the magic field. The 32bits at image base + 4 must point to the magic field. The MOD structure is designed such that it can be placed at image base and point to itself. The 2 fields preceding the magic field get copied around with the structure, even if it is relocated to somewhere besides the image base. If MOD is not located at image base, the value at offset 4 must still point to the MOD magic. In the case of .text being at image base, this implies that the first instruction can only be an unconditional branch over the offset literal.

Arguments

Loader maps memory and writes the arguments to {end of rwdata section specified by last SegmentHeader}. Official processes use argdata_addr = {page-aligned _end}. svcQueryMemory is used by official sw to verify that argdata_addr is mapped RW, since this memory is only mapped when arguments are specified via that command. Afterwards, official sw aligns the argdata_addr to 4-bytes.

The structure located at argdata_addr is as follows:

Offset Size Description
0x0 0x4 This is the total allocated space relative to argdata_addr, used for calculating the max size of the argv ptr array. Normally 0x9000?
0x4 0x4 This is the total_bytesize of the actual argdata string.
0x8 0x18 Unused by official sw.
0x20 See above Actual argdata string.
  • The copy of the args used with the argv array is written by official processes to actual_argdata_string+actual_argdata_size.
  • argv_ptrarray written by official processes is at (args_copy+actual_argdata_size) + 0x9 & ~0x7.