| OLD | NEW |
| (Empty) |
| 1 /* | |
| 2 * Copyright (c) 2012 The Native Client Authors. All rights reserved. | |
| 3 * Use of this source code is governed by a BSD-style license that can be | |
| 4 * found in the LICENSE file. | |
| 5 */ | |
| 6 | |
| 7 /* | |
| 8 * This is the core of ia32-mode validator. Please note that this file | |
| 9 * combines ragel machine description and C language actions. Please read | |
| 10 * validator_internals.html first to understand how the whole thing is built: | |
| 11 * it explains how the byte sequences are constructed, what constructs like | |
| 12 * "@{}" or "REX_WRX?" mean, etc. | |
| 13 */ | |
| 14 | |
| 15 #include <assert.h> | |
| 16 #include <errno.h> | |
| 17 #include <stddef.h> | |
| 18 #include <stdio.h> | |
| 19 #include <stdlib.h> | |
| 20 #include <string.h> | |
| 21 | |
| 22 #include "native_client/src/trusted/validator_ragel/bitmap.h" | |
| 23 #include "native_client/src/trusted/validator_ragel/unreviewed/validator_interna
l.h" | |
| 24 | |
| 25 /* Ignore this information: it's not used by security model in IA32 mode. */ | |
| 26 #undef GET_VEX_PREFIX3 | |
| 27 #define GET_VEX_PREFIX3 0 | |
| 28 #undef SET_VEX_PREFIX3 | |
| 29 #define SET_VEX_PREFIX3(P) | |
| 30 | |
| 31 %%{ | |
| 32 machine x86_32_validator; | |
| 33 alphtype unsigned char; | |
| 34 variable p current_position; | |
| 35 variable pe end_of_bundle; | |
| 36 variable eof end_of_bundle; | |
| 37 variable cs current_state; | |
| 38 | |
| 39 include byte_machine "byte_machines.rl"; | |
| 40 | |
| 41 include prefixes_parsing_validator | |
| 42 "native_client/src/trusted/validator_ragel/unreviewed/parse_instruction.rl"; | |
| 43 include vex_actions_ia32 | |
| 44 "native_client/src/trusted/validator_ragel/unreviewed/parse_instruction.rl"; | |
| 45 include vex_parsing_ia32 | |
| 46 "native_client/src/trusted/validator_ragel/unreviewed/parse_instruction.rl"; | |
| 47 include displacement_fields_actions | |
| 48 "native_client/src/trusted/validator_ragel/unreviewed/parse_instruction.rl"; | |
| 49 include displacement_fields_parsing | |
| 50 "native_client/src/trusted/validator_ragel/unreviewed/parse_instruction.rl"; | |
| 51 include modrm_parsing_ia32_validator | |
| 52 "native_client/src/trusted/validator_ragel/unreviewed/parse_instruction.rl"; | |
| 53 include immediate_fields_actions | |
| 54 "native_client/src/trusted/validator_ragel/unreviewed/parse_instruction.rl"; | |
| 55 include immediate_fields_parsing_ia32 | |
| 56 "native_client/src/trusted/validator_ragel/unreviewed/parse_instruction.rl"; | |
| 57 include relative_fields_validator_actions | |
| 58 "native_client/src/trusted/validator_ragel/unreviewed/parse_instruction.rl"; | |
| 59 include relative_fields_parsing | |
| 60 "native_client/src/trusted/validator_ragel/unreviewed/parse_instruction.rl"; | |
| 61 include cpuid_actions | |
| 62 "native_client/src/trusted/validator_ragel/unreviewed/parse_instruction.rl"; | |
| 63 | |
| 64 # Action which marks last byte as not immediate. Most 3DNow! instructions, | |
| 65 # some AVX and XOP instructions have this property. It's referenced by | |
| 66 # decode_x86_32 machine in [autogenerated] "validator_x86_32_instruction.rl" | |
| 67 # file. | |
| 68 action last_byte_is_not_immediate { | |
| 69 instruction_info_collected |= LAST_BYTE_IS_NOT_IMMEDIATE; | |
| 70 } | |
| 71 | |
| 72 include decode_x86_32 "validator_x86_32_instruction.rl"; | |
| 73 | |
| 74 special_instruction = | |
| 75 # and $~0x1f, %eXX call %eXX | |
| 76 # vvvvvvvvvv | |
| 77 (0x83 0xe0 0xe0 0xff (0xd0|0xe0) | # naclcall/jmp %eax | |
| 78 0x83 0xe1 0xe0 0xff (0xd1|0xe1) | # naclcall/jmp %ecx | |
| 79 0x83 0xe2 0xe0 0xff (0xd2|0xe2) | # naclcall/jmp %edx | |
| 80 0x83 0xe3 0xe0 0xff (0xd3|0xe3) | # naclcall/jmp %ebx | |
| 81 0x83 0xe4 0xe0 0xff (0xd4|0xe4) | # naclcall/jmp %esp | |
| 82 0x83 0xe5 0xe0 0xff (0xd5|0xe5) | # naclcall/jmp %ebp | |
| 83 0x83 0xe6 0xe0 0xff (0xd6|0xe6) | # naclcall/jmp %esi | |
| 84 0x83 0xe7 0xe0 0xff (0xd7|0xe7)) # naclcall/jmp %edi | |
| 85 # ^^^^ ^^^^ | |
| 86 # and $~0x1f, %eXX jmp %eXX | |
| 87 @{ | |
| 88 UnmarkValidJumpTarget((current_position - data) - 1, valid_targets); | |
| 89 instruction_begin -= 3; | |
| 90 instruction_info_collected |= SPECIAL_INSTRUCTION; | |
| 91 } | | |
| 92 (0x65 0xa1 (0x00|0x04) 0x00 0x00 0x00 | # mov %gs:0x0/0x4,%eax | |
| 93 0x65 0x8b (0x05|0x0d|0x015|0x1d|0x25|0x2d|0x35|0x3d) | |
| 94 (0x00|0x04) 0x00 0x00 0x00); # mov %gs:0x0/0x4,%reg | |
| 95 | |
| 96 # Check if call is properly aligned | |
| 97 # | |
| 98 # For direct call we explicitly encode all variations. For indirect call | |
| 99 # we accept all the special instructions which ends with register-addressed | |
| 100 # indirect call. | |
| 101 call_alignment = | |
| 102 ((one_instruction & | |
| 103 # Direct call | |
| 104 ((data16 0xe8 rel16) | | |
| 105 (0xe8 rel32))) | | |
| 106 (special_instruction & | |
| 107 # Indirect call | |
| 108 (any* data16? 0xff ((opcode_2 | opcode_3) any* & | |
| 109 modrm_registers)))) | |
| 110 # Call instruction must aligned to the end of bundle. Previously this was | |
| 111 # strict requirement, today it's just warning to aid with debugging. | |
| 112 @{ | |
| 113 if (((current_position - data) & kBundleMask) != kBundleMask) | |
| 114 instruction_info_collected |= BAD_CALL_ALIGNMENT; | |
| 115 }; | |
| 116 | |
| 117 # This action calls user's callback (if needed) and cleans up validator's | |
| 118 # internal state. | |
| 119 # | |
| 120 # We call the user callback if there are validation errors or if the | |
| 121 # CALL_USER_CALLBACK_ON_EACH_INSTRUCTION option is used. | |
| 122 # | |
| 123 # After that we move instruction_begin and clean all the variables which | |
| 124 # only used in the processing of a single instruction (prefixes, operand | |
| 125 # states and instruction_info_collected). | |
| 126 action end_of_instruction_cleanup { | |
| 127 /* Mark start of this instruction as a valid target for jump. */ | |
| 128 MarkValidJumpTarget(instruction_begin - data, valid_targets); | |
| 129 | |
| 130 /* Call user-supplied callback. */ | |
| 131 instruction_end = current_position + 1; | |
| 132 if ((instruction_info_collected & VALIDATION_ERRORS_MASK) || | |
| 133 (options & CALL_USER_CALLBACK_ON_EACH_INSTRUCTION)) { | |
| 134 result &= user_callback(instruction_begin, instruction_end, | |
| 135 instruction_info_collected, callback_data); | |
| 136 } | |
| 137 | |
| 138 /* On successful match the instruction_begin must point to the next byte | |
| 139 * to be able to report the new offset as the start of instruction | |
| 140 * causing error. */ | |
| 141 instruction_begin = instruction_end; | |
| 142 | |
| 143 /* Clear variables (well, one variable currently). */ | |
| 144 instruction_info_collected = 0; | |
| 145 } | |
| 146 | |
| 147 # This action reports fatal error detected by DFA. | |
| 148 action report_fatal_error { | |
| 149 result &= user_callback(instruction_begin, current_position, | |
| 150 UNRECOGNIZED_INSTRUCTION, callback_data); | |
| 151 /* | |
| 152 * Process the next bundle: "continue" here is for the "for" cycle in | |
| 153 * the ValidateChunkIA32 function. | |
| 154 * | |
| 155 * It does not affect the case which we really care about (when code | |
| 156 * is validatable), but makes it possible to detect more errors in one | |
| 157 * run in tools like ncval. | |
| 158 */ | |
| 159 continue; | |
| 160 } | |
| 161 | |
| 162 # This is main ragel machine: it does 99% of validation work. There are only | |
| 163 # one thing to do if this machine accepts the bundles - check that direct | |
| 164 # jumps are correct. This is done in the following way: | |
| 165 # * DFA fills two arrays: valid_targets and jump_dests. | |
| 166 # * ProcessInvalidJumpTargets checks that "jump_dests & !valid_targets == 0". | |
| 167 # All other checks are done here. | |
| 168 main := ((call_alignment | one_instruction | special_instruction) | |
| 169 @end_of_instruction_cleanup)* | |
| 170 $!report_fatal_error; | |
| 171 | |
| 172 }%% | |
| 173 | |
| 174 %% write data; | |
| 175 | |
| 176 | |
| 177 Bool ValidateChunkIA32(const uint8_t *data, size_t size, | |
| 178 uint32_t options, | |
| 179 const NaClCPUFeaturesX86 *cpu_features, | |
| 180 ValidationCallbackFunc user_callback, | |
| 181 void *callback_data) { | |
| 182 bitmap_word valid_targets_small; | |
| 183 bitmap_word jump_dests_small; | |
| 184 bitmap_word *valid_targets; | |
| 185 bitmap_word *jump_dests; | |
| 186 const uint8_t *current_position; | |
| 187 const uint8_t *end_of_bundle; | |
| 188 int result = TRUE; | |
| 189 | |
| 190 CHECK(sizeof valid_targets_small == sizeof jump_dests_small); | |
| 191 CHECK(size % kBundleSize == 0); | |
| 192 | |
| 193 /* For a very small sequences (one bundle) malloc is too expensive. */ | |
| 194 if (size <= (sizeof valid_targets_small * 8)) { | |
| 195 valid_targets_small = 0; | |
| 196 valid_targets = &valid_targets_small; | |
| 197 jump_dests_small = 0; | |
| 198 jump_dests = &jump_dests_small; | |
| 199 } else { | |
| 200 valid_targets = BitmapAllocate(size); | |
| 201 jump_dests = BitmapAllocate(size); | |
| 202 if (!valid_targets || !jump_dests) { | |
| 203 free(jump_dests); | |
| 204 free(valid_targets); | |
| 205 errno = ENOMEM; | |
| 206 return FALSE; | |
| 207 } | |
| 208 } | |
| 209 | |
| 210 /* | |
| 211 * This option is usually used in tests: we will process the whole chunk | |
| 212 * in one pass. Usually each bundle is processed separately which means | |
| 213 * instructions (and super-instructions) can not cross borders of the bundle. | |
| 214 */ | |
| 215 if (options & PROCESS_CHUNK_AS_A_CONTIGUOUS_STREAM) | |
| 216 end_of_bundle = data + size; | |
| 217 else | |
| 218 end_of_bundle = data + kBundleSize; | |
| 219 | |
| 220 /* | |
| 221 * Main loop. Here we process the data array bundle-after-bundle. | |
| 222 * Ragel-produced DFA does all the checks with one exception: direct jumps. | |
| 223 * It collects the two arrays: valid_targets and jump_dests which are used | |
| 224 * to test direct jumps later. | |
| 225 */ | |
| 226 for (current_position = data; | |
| 227 current_position < data + size; | |
| 228 current_position = end_of_bundle, | |
| 229 end_of_bundle = current_position + kBundleSize) { | |
| 230 /* Start of the instruction being processed. */ | |
| 231 const uint8_t *instruction_begin = current_position; | |
| 232 /* Only used locally in the end_of_instruction_cleanup action. */ | |
| 233 const uint8_t *instruction_end; | |
| 234 uint32_t instruction_info_collected = 0; | |
| 235 int current_state; | |
| 236 | |
| 237 %% write init; | |
| 238 %% write exec; | |
| 239 } | |
| 240 | |
| 241 /* | |
| 242 * Check the direct jumps. All the targets from jump_dests must be in | |
| 243 * valid_targets. | |
| 244 */ | |
| 245 result &= ProcessInvalidJumpTargets(data, size, valid_targets, jump_dests, | |
| 246 user_callback, callback_data); | |
| 247 | |
| 248 /* We only use malloc for a large code sequences */ | |
| 249 if (jump_dests != &jump_dests_small) free(jump_dests); | |
| 250 if (valid_targets != &valid_targets_small) free(valid_targets); | |
| 251 if (!result) errno = EINVAL; | |
| 252 return result; | |
| 253 } | |
| OLD | NEW |