//: Loading SubX programs from ELF binaries.
//: This will allow us to run them natively on a Linux kernel.
//: Based on https://github.com/kragen/stoneknifeforth/blob/702d2ebe1b/386.c
:(before "End Main")
assert(argc > 1);
if (is_equal(argv[1], "run")) {
// Outside of tests, traces must be explicitly requested.
if (Trace_file.is_open()) Trace_stream = new trace_stream;
trace(2, "run") << "=== Starting to run" << end();
assert(argc > 2);
reset();
cerr << std::hex;
load_elf(argv[2], argc, argv);
while (EIP < End_of_program) // weak final-gasp termination check
run_one_instruction();
raise << "executed past end of the world: " << EIP << " vs " << End_of_program << '\n' << end();
return 1;
}
:(code)
void load_elf(const string& filename, int argc, char* argv[]) {
int fd = open(filename.c_str(), O_RDONLY);
if (fd < 0) raise << filename.c_str() << ": open" << perr() << '\n' << die();
off_t size = lseek(fd, 0, SEEK_END);
lseek(fd, 0, SEEK_SET);
uint8_t* elf_contents = static_cast<uint8_t*>(malloc(size));
if (elf_contents == NULL) raise << "malloc(" << size << ')' << perr() << '\n' << die();
ssize_t read_size = read(fd, elf_contents, size);
if (size != read_size) raise << "read → " << size << " (!= " << read_size << ')' << perr() << '\n' << die();
load_elf_contents(elf_contents, size, argc, argv);
free(elf_contents);
}
void load_elf_contents(uint8_t* elf_contents, size_t size, int argc, char* argv[]) {
uint8_t magic[5] = {0};
memcpy(magic, elf_contents, 4);
if (memcmp(magic, "\177ELF", 4) != 0)
raise << "Invalid ELF file; starts with \"" << magic << '"' << die();
if (elf_contents[4] != 1)
raise << "Only 32-bit ELF files (4-byte words; virtual addresses up to 4GB) supported.\n" << die();
if (elf_contents[5] != 1)
raise << "Only little-endian ELF files supported.\n" << die();
// unused: remaining 10 bytes of e_ident
uint32_t e_machine_type = u32_in(&elf_contents[16]);
if (e_machine_type != 0x00030002)
raise << "ELF type/machine 0x" << HEXWORD << e_machine_type << " isn't i386 executable\n" << die();
// unused: e_version. We only support version 1, and later versions will be backwards compatible.
uint32_t e_entry = u32_in(&elf_contents[24]);
uint32_t e_phoff = u32_in(&elf_contents[28]);
// unused: e_shoff
// unused: e_flags
uint32_t e_ehsize = u16_in(&elf_contents[40]);
if (e_ehsize < 52) raise << "Invalid binary; ELF header too small\n" << die();
uint32_t e_phentsize = u16_in(&elf_contents[42]);
uint32_t e_phnum = u16_in(&elf_contents[44]);
trace(90, "load") << e_phnum << " entries in the program header, each " << e_phentsize << " bytes long" << end();
// unused: e_shentsize
// unused: e_shnum
// unused: e_shstrndx
set<uint32_t> overlap; // to detect overlapping segments
for (size_t i = 0; i < e_phnum; ++i)
load_segment_from_program_header(elf_contents, i, size, e_phoff + i*e_phentsize, e_ehsize, overlap);
// initialize code and stack
assert(overlap.find(STACK_SEGMENT) == overlap.end());
Mem.push_back(vma(STACK_SEGMENT));
assert(overlap.find(AFTER_STACK) == overlap.end());
// The stack grows downward.
Reg[ESP].u = AFTER_STACK;
Reg[EBP].u = 0;
EIP = e_entry;
// initialize args on stack
// no envp for now
// we wastefully use a separate page of memory for argv
Mem.push_back(vma(ARGV_DATA_SEGMENT));
uint32_t argv_data = ARGV_DATA_SEGMENT;
for (int i = argc-1; i >= /*skip 'subx_bin' and 'run'*/2; --i) {
push(argv_data);
for (size_t j = 0; j <= strlen(argv[i]); ++j) {
assert(overlap.find(argv_data) == overlap.end()); // don't bother comparing ARGV and STACK
write_mem_u8(argv_data, argv[i][j]);
argv_data += sizeof(char);
assert(argv_data < ARGV_DATA_SEGMENT + SEGMENT_ALIGNMENT);
}
}
push(argc-/*skip 'subx_bin' and 'run'*/2);
}
void push(uint32_t val) {
Reg[ESP].u -= 4;
if (Reg[ESP].u < STACK_SEGMENT) {
raise << "The stack overflowed its segment. "
<< "Maybe SPACE_FOR_SEGMENT should be larger? "
<< "Or you need to carve out an exception for the stack segment "
<< "to be larger.\n" << die();
}
trace(Callstack_depth+1, "run") << "decrementing ESP to 0x" << HEXWORD << Reg[ESP].u << end();
trace(Callstack_depth+1, "run") << "pushing value 0x" << HEXWORD << val << end();
write_mem_u32(Reg[ESP].u, val);
}
void load_segment_from_program_header(uint8_t* elf_contents, int segment_index, size_t size, uint32_t offset, uint32_t e_ehsize, set<uint32_t>& overlap) {
uint32_t p_type = u32_in(&elf_contents[offset]);
trace(90, "load") << "program header at offset " << offset << ": type " << p_type << end();
if (p_type != 1) {
trace(90, "load") << "ignoring segment at offset " << offset << " of non PT_LOAD type " << p_type << " (see http://refspecs.linuxbase.org/elf/elf.pdf)" << end();
return;
}
uint32_t p_offset = u32_in(&elf_contents[offset + 4]);
uint32_t p_vaddr = u32_in(&elf_contents[offset + 8]);
if (e_ehsize > p_vaddr) raise << "Invalid binary; program header overlaps ELF header\n" << die();
// unused: p_paddr
uint32_t p_filesz = u32_in(&elf_contents[offset + 16]);
uint32_t p_memsz = u32_in(&elf_contents[offset + 20]);
if (p_filesz != p_memsz)
raise << "Can't yet handle segments where p_filesz != p_memsz (see http://refspecs.linuxbase.org/elf/elf.pdf)\n" << die();
if (p_offset + p_filesz > size)
raise << "Invalid binary; segment at offset " << offset << " is too large: wants to end at " << p_offset+p_filesz << " but the file ends at " << size << '\n' << die();
if (p_memsz >= SEGMENT_ALIGNMENT) {
raise << "Code segment too small for SubX; for now please manually increase SEGMENT_ALIGNMENT.\n" << end();
return;
}
trace(90, "load") << "blitting file offsets (" << p_offset << ", " << (p_offset+p_filesz) << ") to addresses (" << p_vaddr << ", " << (p_vaddr+p_memsz) << ')' << end();
if (size > p_memsz) size = p_memsz;
Mem.push_back(vma(p_vaddr));
for (size_t i = 0; i < p_filesz; ++i) {
assert(overlap.find(p_vaddr+i) == overlap.end());
write_mem_u8(p_vaddr+i, elf_contents[p_offset+i]);
overlap.insert(p_vaddr+i);
}
if (segment_index == 0 && End_of_program < p_vaddr+p_memsz)
End_of_program = p_vaddr+p_memsz;
}
:(before "End Includes")
// Very primitive/fixed/insecure ELF segments for now.
// --- inaccessible: 0x00000000 -> 0x08047fff
// code: 0x09000000 -> 0x09ffffff (specified in ELF binary)
// data: 0x0a000000 -> 0x0affffff (specified in ELF binary)
// --- heap gets mmap'd somewhere here ---
// stack: 0xbdffffff -> 0xbd000000 (downward; not in ELF binary)
// argv hack: 0xbf000000 -> 0xbfffffff (not in ELF binary)
// --- reserved for kernel: 0xc0000000 -> ...
const uint32_t START_HEAP = 0x0b000000;
const uint32_t END_HEAP = 0xbd000000;
const uint32_t STACK_SEGMENT = 0xbd000000;
const uint32_t AFTER_STACK = 0xbe000000;
const uint32_t ARGV_DATA_SEGMENT = 0xbf000000;
// When updating the above memory map, don't forget to update `mmap`'s
// implementation in the 'syscalls' layer.
:(before "End Dump Info for Instruction")
//? dump_stack(); // slow
:(code)
void dump_stack() {
ostringstream out;
trace(Callstack_depth+1, "run") << "stack:" << end();
for (uint32_t a = AFTER_STACK-4; a > Reg[ESP].u; a -= 4)
trace(Callstack_depth+2, "run") << " 0x" << HEXWORD << a << " => 0x" << HEXWORD << read_mem_u32(a) << end();
trace(Callstack_depth+2, "run") << " 0x" << HEXWORD << Reg[ESP].u << " => 0x" << HEXWORD << read_mem_u32(Reg[ESP].u) << " <=== ESP" << end();
for (uint32_t a = Reg[ESP].u-4; a > Reg[ESP].u-40; a -= 4)
trace(Callstack_depth+2, "run") << " 0x" << HEXWORD << a << " => 0x" << HEXWORD << read_mem_u32(a) << end();
}
inline uint32_t u32_in(uint8_t* p) {
return p[0] | p[1] << 8 | p[2] << 16 | p[3] << 24;
}
inline uint16_t u16_in(uint8_t* p) {
return p[0] | p[1] << 8;
}
:(before "End Types")
struct perr {};
:(code)
ostream& operator<<(ostream& os, perr /*unused*/) {
if (errno)
os << ": " << strerror(errno);
return os;
}
:(before "End Includes")
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <stdarg.h>
#include <errno.h>
#include <unistd.h>