https://github.com/akkartik/mu/blob/master/subx/012elf.cc
  1 //: Generating ELF binaries for SubX programs.
  2 //: This will allow us to run them natively on a Linux kernel.
  3 //: Based on https://github.com/kragen/stoneknifeforth/blob/702d2ebe1b/386.c
  4 
  5 :(before "End Main")
  6 assert(argc > 1);
  7 if (is_equal(argv[1], "run")) {
  8   START_TRACING_UNTIL_END_OF_SCOPE;
  9   assert(argc > 2);
 10   reset();
 11   cerr << std::hex;
 12   load_elf(argv[2], argc, argv);
 13   while (EIP < End_of_program)  // weak final-gasp termination check
 14     run_one_instruction();
 15   raise << "executed past end of the world: " << EIP << " vs " << End_of_program << '\n' << end();
 16   return 1;
 17 }
 18 
 19 :(code)
 20 void load_elf(const string& filename, int argc, char* argv[]) {
 21   int fd = open(filename.c_str(), O_RDONLY);
 22   if (fd < 0) raise << filename.c_str() << ": open" << perr() << '\n' << die();
 23   off_t size = lseek(fd, 0, SEEK_END);
 24   lseek(fd, 0, SEEK_SET);
 25   uint8_t* elf_contents = static_cast<uint8_t*>(malloc(size));
 26   if (elf_contents == NULL) raise << "malloc(" << size << ')' << perr() << '\n' << die();
 27   ssize_t read_size = read(fd, elf_contents, size);
 28   if (size != read_size) raise << "read → " << size << " (!= " << read_size << ')' << perr() << '\n' << die();
 29   load_elf_contents(elf_contents, size, argc, argv);
 30   free(elf_contents);
 31 }
 32 
 33 void load_elf_contents(uint8_t* elf_contents, size_t size, int argc, char* argv[]) {
 34   uint8_t magic[5] = {0};
 35   memcpy(magic, elf_contents, 4);
 36   if (memcmp(magic, "\177ELF", 4) != 0)
 37     raise << "Invalid ELF file; starts with \"" << magic << '"' << die();
 38   if (elf_contents[4] != 1)
 39     raise << "Only 32-bit ELF files (4-byte words; virtual addresses up to 4GB) supported.\n" << die();
 40   if (elf_contents[5] != 1)
 41     raise << "Only little-endian ELF files supported.\n" << die();
 42   // unused: remaining 10 bytes of e_ident
 43   uint32_t e_machine_type = u32_in(&elf_contents[16]);
 44   if (e_machine_type != 0x00030002)
 45     raise << "ELF type/machine 0x" << HEXWORD << e_machine_type << " isn't i386 executable\n" << die();
 46   // unused: e_version. We only support version 1, and later versions will be backwards compatible.
 47   uint32_t e_entry = u32_in(&elf_contents[24]);
 48   uint32_t e_phoff = u32_in(&elf_contents[28]);
 49   // unused: e_shoff
 50   // unused: e_flags
 51   uint32_t e_ehsize = u16_in(&elf_contents[40]);
 52   if (e_ehsize < 52) raise << "Invalid binary; ELF header too small\n" << die();
 53   uint32_t e_phentsize = u16_in(&elf_contents[42]);
 54   uint32_t e_phnum = u16_in(&elf_contents[44]);
 55   trace(90, "load") << e_phnum << " entries in the program header, each " << e_phentsize << " bytes long" << end();
 56   // unused: e_shentsize
 57   // unused: e_shnum
 58   // unused: e_shstrndx
 59 
 60   set<uint32_t> overlap;  // to detect overlapping segments
 61   for (size_t i = 0;  i < e_phnum;  ++i)
 62     load_segment_from_program_header(elf_contents, i, size, e_phoff + i*e_phentsize, e_ehsize, overlap);
 63 
 64   // initialize code and stack
 65   assert(overlap.find(STACK_SEGMENT) == overlap.end());
 66   Mem.push_back(vma(STACK_SEGMENT));
 67   assert(overlap.find(AFTER_STACK) == overlap.end());
 68   Reg[ESP].u = AFTER_STACK;
 69   Reg[EBP].u = 0;
 70   EIP = e_entry;
 71 
 72   // initialize args on stack
 73   // no envp for now
 74   // we wastefully use a separate page of memory for argv
 75   Mem.push_back(vma(ARGV_DATA_SEGMENT));
 76   uint32_t argv_data = ARGV_DATA_SEGMENT;
 77   for (int i = argc-1;  i >= /*skip 'subx_bin' and 'run'*/2;  --i) {
 78     push(argv_data);
 79     for (size_t j = 0;  j <= strlen(argv[i]);  ++j) {
 80       assert(overlap.find(argv_data) == overlap.end());  // don't bother comparing ARGV and STACK
 81       write_mem_u8(argv_data, argv[i][j]);
 82       argv_data += sizeof(char);
 83       assert(argv_data < ARGV_DATA_SEGMENT + SEGMENT_ALIGNMENT);
 84     }
 85   }
 86   push(argc-/*skip 'subx_bin' and 'run'*/2);
 87 }
 88 
 89 void push(uint32_t val) {
 90   Reg[ESP].u -= 4;
 91   trace(90, "run") << "decrementing ESP to 0x" << HEXWORD << Reg[ESP].u << end();
 92   trace(90, "run") << "pushing value 0x" << HEXWORD << val << end();
 93   write_mem_u32(Reg[ESP].u, val);
 94 }
 95 
 96 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) {
 97   uint32_t p_type = u32_in(&elf_contents[offset]);
 98   trace(90, "load") << "program header at offset " << offset << ": type " << p_type << end();
 99   if (p_type != 1) {
100     trace(90, "load") << "ignoring segment at offset " << offset << " of non PT_LOAD type " << p_type << " (see http://refspecs.linuxbase.org/elf/elf.pdf)" << end();
101     return;
102   }
103   uint32_t p_offset = u32_in(&elf_contents[offset + 4]);
104   uint32_t p_vaddr = u32_in(&elf_contents[offset + 8]);
105   if (e_ehsize > p_vaddr) raise << "Invalid binary; program header overlaps ELF header\n" << die();
106   // unused: p_paddr
107   uint32_t p_filesz = u32_in(&elf_contents[offset + 16]);
108   uint32_t p_memsz = u32_in(&elf_contents[offset + 20]);
109   if (p_filesz != p_memsz)
110     raise << "Can't yet handle segments where p_filesz != p_memsz (see http://refspecs.linuxbase.org/elf/elf.pdf)\n" << die();
111 
112   if (p_offset + p_filesz > size)
113     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();
114   if (p_memsz >= SEGMENT_ALIGNMENT) {
115     raise << "Code segment too small for SubX; for now please manually increase SEGMENT_ALIGNMENT.\n" << end();
116     return;
117   }
118   trace(90, "load") << "blitting file offsets (" << p_offset << ", " << (p_offset+p_filesz) << ") to addresses (" << p_vaddr << ", " << (p_vaddr+p_memsz) << ')' << end();
119   if (size > p_memsz) size = p_memsz;
120   Mem.push_back(vma(p_vaddr));
121   for (size_t i = 0;  i < p_filesz;  ++i) {
122     assert(overlap.find(p_vaddr+i) == overlap.end());
123     write_mem_u8(p_vaddr+i, elf_contents[p_offset+i]);
124     overlap.insert(p_vaddr+i);
125   }
126   if (segment_index == 0 && End_of_program < p_vaddr+p_memsz)
127     End_of_program = p_vaddr+p_memsz;
128 }
129 
130 :(before "End Includes")
131 // Very primitive/fixed/insecure ELF segments for now.
132 //   code: 0x09000000 -> 0x09ffffff
133 //   data/heap: 0x0a000000 -> 0x0affffff
134 //   stack: 0x0b000ffc -> 0x0b000000 (downward)
135 const int CODE_SEGMENT = 0x09000000;
136 const int DATA_SEGMENT = 0x0a000000;
137 const int STACK_SEGMENT = 0x0b000000;
138 const int AFTER_STACK = 0x0b000ffc;  // forget final word because of the off-by-one with SEGMENT_ALIGNMENT;
139 const int ARGV_DATA_SEGMENT = 0x0c000000;
140 :(code)
141 void dump_stack() {
142   cerr << "stack:\n";
143   for (uint32_t a = AFTER_STACK-4;  a > Reg[ESP].u;  a -= 4)
144     cerr << "  0x" << HEXWORD << a << " => 0x" << HEXWORD << read_mem_u32(a) << '\n';
145   cerr << "  0x" << HEXWORD << Reg[ESP].u << " => 0x" << HEXWORD << read_mem_u32(Reg[ESP].u) << "  <=== ESP\n";
146   for (uint32_t a = Reg[ESP].u-4;  a > Reg[ESP].u-40;  a -= 4)
147     cerr << "  0x" << HEXWORD << a << " => 0x" << HEXWORD << read_mem_u32(a) << '\n';
148 }
149 
150 inline uint32_t u32_in(uint8_t* p) {
151   return p[0] | p[1] << 8 | p[2] << 16 | p[3] << 24;
152 }
153 
154 inline uint16_t u16_in(uint8_t* p) {
155   return p[0] | p[1] << 8;
156 }
157 
158 :(before "End Types")
159 struct perr {};
160 :(code)
161 ostream& operator<<(ostream& os, perr /*unused*/) {
162   if (errno)
163     os << ": " << strerror(errno);
164   return os;
165 }
166 
167 :(before "End Types")
168 struct die {};
169 :(code)
170 ostream& operator<<(ostream& /*unused*/, die /*unused*/) {
171   if (Trace_stream) Trace_stream->newline();
172   exit(1);
173 }
174 
175 :(before "End Includes")
176 #include <sys/types.h>
177 #include <sys/stat.h>
178 #include <fcntl.h>
179 #include <stdarg.h>
180 #include <errno.h>
181 #include <unistd.h>