https://github.com/akkartik/mu/blob/main/012elf.cc
  1 //: Loading SubX programs from ELF binaries.
  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   // Outside of tests, traces must be explicitly requested.
  9   if (Trace_file.is_open()) Trace_stream = new trace_stream;
 10   trace(2, "run") << "=== Starting to run" << end();
 11   if (argc <= 2) {
 12     raise << "Not enough arguments provided.\n" << die();
 13   }
 14   reset();
 15   cerr << std::hex;
 16   load_elf(argv[2], argc, argv);
 17   while (EIP < End_of_program)  // weak final-gasp termination check
 18     run_one_instruction();
 19   raise << "executed past end of the world: " << EIP << " vs " << End_of_program << '\n' << end();
 20   return 1;
 21 }
 22 
 23 :(code)
 24 void load_elf(const string& filename, int argc, char* argv[]) {
 25   int fd = open(filename.c_str(), O_RDONLY);
 26   if (fd < 0) raise << filename.c_str() << ": open" << perr() << '\n' << die();
 27   off_t size = lseek(fd, 0, SEEK_END);
 28   lseek(fd, 0, SEEK_SET);
 29   uint8_t* elf_contents = static_cast<uint8_t*>(malloc(size));
 30   if (elf_contents == NULL) raise << "malloc(" << size << ')' << perr() << '\n' << die();
 31   ssize_t read_size = read(fd, elf_contents, size);
 32   if (size != read_size) raise << "read → " << size << " (!= " << read_size << ')' << perr() << '\n' << die();
 33   load_elf_contents(elf_contents, size, argc, argv);
 34   free(elf_contents);
 35 }
 36 
 37 void load_elf_contents(uint8_t* elf_contents, size_t size, int argc, char* argv[]) {
 38   uint8_t magic[5] = {0};
 39   memcpy(magic, elf_contents, 4);
 40   if (memcmp(magic, "\177ELF", 4) != 0)
 41     raise << "Invalid ELF file; starts with \"" << magic << '"' << die();
 42   if (elf_contents[4] != 1)
 43     raise << "Only 32-bit ELF files (4-byte words; virtual addresses up to 4GB) supported.\n" << die();
 44   if (elf_contents[5] != 1)
 45     raise << "Only little-endian ELF files supported.\n" << die();
 46   // unused: remaining 10 bytes of e_ident
 47   uint32_t e_machine_type = u32_in(&elf_contents[16]);
 48   if (e_machine_type != 0x00030002)
 49     raise << "ELF type/machine 0x" << HEXWORD << e_machine_type << " isn't i386 executable\n" << die();
 50   // unused: e_version. We only support version 1, and later versions will be backwards compatible.
 51   uint32_t e_entry = u32_in(&elf_contents[24]);
 52   uint32_t e_phoff = u32_in(&elf_contents[28]);
 53   // unused: e_shoff
 54   // unused: e_flags
 55   uint32_t e_ehsize = u16_in(&elf_contents[40]);
 56   if (e_ehsize < 52) raise << "Invalid binary; ELF header too small\n" << die();
 57   uint32_t e_phentsize = u16_in(&elf_contents[42]);
 58   uint32_t e_phnum = u16_in(&elf_contents[44]);
 59   trace(90, "load") << e_phnum << " entries in the program header, each " << e_phentsize << " bytes long" << end();
 60   // unused: e_shentsize
 61   // unused: e_shnum
 62   // unused: e_shstrndx
 63 
 64   set<uint32_t> overlap;  // to detect overlapping segments
 65   for (size_t i = 0;  i < e_phnum;  ++i)
 66     load_segment_from_program_header(elf_contents, i, size, e_phoff + i*e_phentsize, e_ehsize, overlap);
 67 
 68   // initialize code and stack
 69   assert(overlap.find(STACK_SEGMENT) == overlap.end());
 70   Mem.push_back(vma(STACK_SEGMENT));
 71   assert(overlap.find(AFTER_STACK) == overlap.end());
 72   // The stack grows downward.
 73   Reg[ESP].u = AFTER_STACK;
 74   Reg[EBP].u = 0;
 75   EIP = e_entry;
 76 
 77   // initialize args on stack
 78   // no envp for now
 79   // we wastefully use a separate page of memory for argv
 80   Mem.push_back(vma(ARGV_DATA_SEGMENT));
 81   uint32_t argv_data = ARGV_DATA_SEGMENT;
 82   for (int i = argc-1;  i >= /*skip 'subx_bin' and 'run'*/2;  --i) {
 83     push(argv_data);
 84     for (size_t j = 0;  j <= strlen(argv[i]);  ++j) {
 85       assert(overlap.find(argv_data) == overlap.end());  // don't bother comparing ARGV and STACK
 86       write_mem_u8(argv_data, argv[i][j]);
 87       argv_data += sizeof(char);
 88       assert(argv_data < ARGV_DATA_SEGMENT + SEGMENT_ALIGNMENT);
 89     }
 90   }
 91   push(argc-/*skip 'subx_bin' and 'run'*/2);
 92 }
 93 
 94 void push(uint32_t val) {
 95   Reg[ESP].u -= 4;
 96   if (Reg[ESP].u < STACK_SEGMENT) {
 97     raise << "The stack overflowed its segment. "
 98           << "Maybe SPACE_FOR_SEGMENT should be larger? "
 99           << "Or you need to carve out an exception for the stack segment "
100           << "to be larger.\n" << die();
101   }
102   trace(Callstack_depth+1, "run") << "decrementing ESP to 0x" << HEXWORD << Reg[ESP].u << end();
103   trace(Callstack_depth+1, "run") << "pushing value 0x" << HEXWORD << val << end();
104   write_mem_u32(Reg[ESP].u, val);
105 }
106 
107 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) {
108   uint32_t p_type = u32_in(&elf_contents[offset]);
109   trace(90, "load") << "program header at offset " << offset << ": type " << p_type << end();
110   if (p_type != 1) {
111     trace(90, "load") << "ignoring segment at offset " << offset << " of non PT_LOAD type " << p_type << " (see http://refspecs.linuxbase.org/elf/elf.pdf)" << end();
112     return;
113   }
114   uint32_t p_offset = u32_in(&elf_contents[offset + 4]);
115   uint32_t p_vaddr = u32_in(&elf_contents[offset + 8]);
116   if (e_ehsize > p_vaddr) raise << "Invalid binary; program header overlaps ELF header\n" << die();
117   // unused: p_paddr
118   uint32_t p_filesz = u32_in(&elf_contents[offset + 16]);
119   uint32_t p_memsz = u32_in(&elf_contents[offset + 20]);
120   if (p_filesz != p_memsz)
121     raise << "Can't yet handle segments where p_filesz != p_memsz (see http://refspecs.linuxbase.org/elf/elf.pdf)\n" << die();
122 
123   if (p_offset + p_filesz > size)
124     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();
125   if (p_memsz >= SEGMENT_ALIGNMENT) {
126     raise << "Code segment too small for SubX; for now please manually increase SEGMENT_ALIGNMENT.\n" << end();
127     return;
128   }
129   trace(90, "load") << "blitting file offsets (" << p_offset << ", " << (p_offset+p_filesz) << ") to addresses (" << p_vaddr << ", " << (p_vaddr+p_memsz) << ')' << end();
130   if (size > p_memsz) size = p_memsz;
131   Mem.push_back(vma(p_vaddr));
132   for (size_t i = 0;  i < p_filesz;  ++i) {
133     assert(overlap.find(p_vaddr+i) == overlap.end());
134     write_mem_u8(p_vaddr+i, elf_contents[p_offset+i]);
135     overlap.insert(p_vaddr+i);
136   }
137   if (segment_index == 0 && End_of_program < p_vaddr+p_memsz)
138     End_of_program = p_vaddr+p_memsz;
139 }
140 
141 :(before "End Includes")
142 // Very primitive/fixed/insecure ELF segments for now.
143 //   --- inaccessible:        0x00000000 -> 0x08047fff
144 //   code:                    0x09000000 -> 0x09ffffff (specified in ELF binary)
145 //   data:                    0x0a000000 -> 0x0affffff (specified in ELF binary)
146 //                      --- heap gets mmap'd somewhere here ---
147 //   stack:                   0xbdffffff -> 0xbd000000 (downward; not in ELF binary)
148 //   argv hack:               0xbf000000 -> 0xbfffffff (not in ELF binary)
149 //   --- reserved for kernel: 0xc0000000 -> ...
150 const uint32_t START_HEAP        = 0x0b000000;
151 const uint32_t END_HEAP          = 0xbd000000;
152 const uint32_t STACK_SEGMENT     = 0xbd000000;
153 const uint32_t AFTER_STACK       = 0xbe000000;
154 const uint32_t ARGV_DATA_SEGMENT = 0xbf000000;
155 // When updating the above memory map, don't forget to update `mmap`'s
156 // implementation in the 'syscalls' layer.
157 :(before "End Dump Info for Instruction")
158 //? dump_stack();  // slow
159 :(code)
160 void dump_stack() {
161   ostringstream out;
162   trace(Callstack_depth+1, "run") << "stack:" << end();
163   for (uint32_t a = AFTER_STACK-4;  a > Reg[ESP].u;  a -= 4)
164     trace(Callstack_depth+2, "run") << "  0x" << HEXWORD << a << " => 0x" << HEXWORD << read_mem_u32(a) << end();
165   trace(Callstack_depth+2, "run") << "  0x" << HEXWORD << Reg[ESP].u << " => 0x" << HEXWORD << read_mem_u32(Reg[ESP].u) << "  <=== ESP" << end();
166   for (uint32_t a = Reg[ESP].u-4;  a > Reg[ESP].u-40;  a -= 4)
167     trace(Callstack_depth+2, "run") << "  0x" << HEXWORD << a << " => 0x" << HEXWORD << read_mem_u32(a) << end();
168 }
169 
170 inline uint32_t u32_in(uint8_t* p) {
171   return p[0] | p[1] << 8 | p[2] << 16 | p[3] << 24;
172 }
173 
174 inline uint16_t u16_in(uint8_t* p) {
175   return p[0] | p[1] << 8;
176 }
177 
178 :(before "End Types")
179 struct perr {};
180 :(code)
181 ostream& operator<<(ostream& os, perr /*unused*/) {
182   if (errno)
183     os << ": " << strerror(errno);
184   return os;
185 }
186 
187 :(before "End Includes")
188 #include <sys/types.h>
189 #include <sys/stat.h>
190 #include <fcntl.h>
191 #include <stdarg.h>
192 #include <errno.h>
193 #include <unistd.h>