| | 1 | /** @file\r |
| | 2 | This module contains EBC support routines that are customized based on\r |
| | 3 | the target processor.\r |
| | 4 | \r |
| | 5 | Copyright (c) 2006 - 2012, Intel Corporation. All rights reserved.<BR>\r |
| | 6 | This program and the accompanying materials\r |
| | 7 | are licensed and made available under the terms and conditions of the BSD License\r |
| | 8 | which accompanies this distribution. The full text of the license may be found at\r |
| | 9 | http://opensource.org/licenses/bsd-license.php\r |
| | 10 | \r |
| | 11 | THE PROGRAM IS DISTRIBUTED UNDER THE BSD LICENSE ON AN "AS IS" BASIS,\r |
| | 12 | WITHOUT WARRANTIES OR REPRESENTATIONS OF ANY KIND, EITHER EXPRESS OR IMPLIED.\r |
| | 13 | \r |
| | 14 | **/\r |
| | 15 | \r |
| | 16 | #include "EbcInt.h"\r |
| | 17 | #include "EbcExecute.h"\r |
| | 18 | #include "EbcSupport.h"\r |
| | 19 | \r |
| | 20 | /**\r |
| | 21 | Given raw bytes of Itanium based code, format them into a bundle and\r |
| | 22 | write them out.\r |
| | 23 | \r |
| | 24 | @param MemPtr pointer to memory location to write the bundles\r |
| | 25 | to.\r |
| | 26 | @param Template 5-bit template.\r |
| | 27 | @param Slot0 Instruction slot 0 data for the bundle.\r |
| | 28 | @param Slot1 Instruction slot 1 data for the bundle.\r |
| | 29 | @param Slot2 Instruction slot 2 data for the bundle.\r |
| | 30 | \r |
| | 31 | @retval EFI_INVALID_PARAMETER Pointer is not aligned\r |
| | 32 | @retval EFI_INVALID_PARAMETER No more than 5 bits in template\r |
| | 33 | @retval EFI_INVALID_PARAMETER More than 41 bits used in code\r |
| | 34 | @retval EFI_SUCCESS All data is written.\r |
| | 35 | \r |
| | 36 | **/\r |
| | 37 | EFI_STATUS\r |
| | 38 | WriteBundle (\r |
| | 39 | IN VOID *MemPtr,\r |
| | 40 | IN UINT8 Template,\r |
| | 41 | IN UINT64 Slot0,\r |
| | 42 | IN UINT64 Slot1,\r |
| | 43 | IN UINT64 Slot2\r |
| | 44 | );\r |
| | 45 | \r |
| | 46 | /**\r |
| | 47 | Pushes a 64 bit unsigned value to the VM stack.\r |
| | 48 | \r |
| | 49 | @param VmPtr The pointer to current VM context.\r |
| | 50 | @param Arg The value to be pushed.\r |
| | 51 | \r |
| | 52 | **/\r |
| | 53 | VOID\r |
| | 54 | PushU64 (\r |
| | 55 | IN VM_CONTEXT *VmPtr,\r |
| | 56 | IN UINT64 Arg\r |
| | 57 | )\r |
| | 58 | {\r |
| | 59 | //\r |
| | 60 | // Advance the VM stack down, and then copy the argument to the stack.\r |
| | 61 | // Hope it's aligned.\r |
| | 62 | //\r |
| | 63 | VmPtr->Gpr[0] -= sizeof (UINT64);\r |
| | 64 | *(UINT64 *) VmPtr->Gpr[0] = Arg;\r |
| | 65 | }\r |
| | 66 | \r |
| | 67 | /**\r |
| | 68 | Begin executing an EBC image. The address of the entry point is passed\r |
| | 69 | in via a processor register, so we'll need to make a call to get the\r |
| | 70 | value.\r |
| | 71 | \r |
| | 72 | This is a thunk function. Microsoft x64 compiler only provide fast_call\r |
| | 73 | calling convention, so the first four arguments are passed by rcx, rdx,\r |
| | 74 | r8, and r9, while other arguments are passed in stack.\r |
| | 75 | \r |
| | 76 | @param Arg1 The 1st argument.\r |
| | 77 | @param ... The variable arguments list.\r |
| | 78 | \r |
| | 79 | @return The value returned by the EBC application we're going to run.\r |
| | 80 | \r |
| | 81 | **/\r |
| | 82 | UINT64\r |
| | 83 | EFIAPI\r |
| | 84 | EbcInterpret (\r |
| | 85 | UINT64 Arg1,\r |
| | 86 | ...\r |
| | 87 | )\r |
| | 88 | {\r |
| | 89 | //\r |
| | 90 | // Create a new VM context on the stack\r |
| | 91 | //\r |
| | 92 | VM_CONTEXT VmContext;\r |
| | 93 | UINTN Addr;\r |
| | 94 | EFI_STATUS Status;\r |
| | 95 | UINTN StackIndex;\r |
| | 96 | VA_LIST List;\r |
| | 97 | UINT64 Arg2;\r |
| | 98 | UINT64 Arg3;\r |
| | 99 | UINT64 Arg4;\r |
| | 100 | UINT64 Arg5;\r |
| | 101 | UINT64 Arg6;\r |
| | 102 | UINT64 Arg7;\r |
| | 103 | UINT64 Arg8;\r |
| | 104 | UINT64 Arg9;\r |
| | 105 | UINT64 Arg10;\r |
| | 106 | UINT64 Arg11;\r |
| | 107 | UINT64 Arg12;\r |
| | 108 | UINT64 Arg13;\r |
| | 109 | UINT64 Arg14;\r |
| | 110 | UINT64 Arg15;\r |
| | 111 | UINT64 Arg16;\r |
| | 112 | //\r |
| | 113 | // Get the EBC entry point from the processor register. Make sure you don't\r |
| | 114 | // call any functions before this or you could mess up the register the\r |
| | 115 | // entry point is passed in.\r |
| | 116 | //\r |
| | 117 | Addr = EbcLLGetEbcEntryPoint ();\r |
| | 118 | //\r |
| | 119 | // Need the args off the stack.\r |
| | 120 | //\r |
| | 121 | VA_START (List, Arg1);\r |
| | 122 | Arg2 = VA_ARG (List, UINT64);\r |
| | 123 | Arg3 = VA_ARG (List, UINT64);\r |
| | 124 | Arg4 = VA_ARG (List, UINT64);\r |
| | 125 | Arg5 = VA_ARG (List, UINT64);\r |
| | 126 | Arg6 = VA_ARG (List, UINT64);\r |
| | 127 | Arg7 = VA_ARG (List, UINT64);\r |
| | 128 | Arg8 = VA_ARG (List, UINT64);\r |
| | 129 | Arg9 = VA_ARG (List, UINT64);\r |
| | 130 | Arg10 = VA_ARG (List, UINT64);\r |
| | 131 | Arg11 = VA_ARG (List, UINT64);\r |
| | 132 | Arg12 = VA_ARG (List, UINT64);\r |
| | 133 | Arg13 = VA_ARG (List, UINT64);\r |
| | 134 | Arg14 = VA_ARG (List, UINT64);\r |
| | 135 | Arg15 = VA_ARG (List, UINT64);\r |
| | 136 | Arg16 = VA_ARG (List, UINT64);\r |
| | 137 | VA_END (List);\r |
| | 138 | //\r |
| | 139 | // Now clear out our context\r |
| | 140 | //\r |
| | 141 | ZeroMem ((VOID *) &VmContext, sizeof (VM_CONTEXT));\r |
| | 142 | //\r |
| | 143 | // Set the VM instruction pointer to the correct location in memory.\r |
| | 144 | //\r |
| | 145 | VmContext.Ip = (VMIP) Addr;\r |
| | 146 | //\r |
| | 147 | // Initialize the stack pointer for the EBC. Get the current system stack\r |
| | 148 | // pointer and adjust it down by the max needed for the interpreter.\r |
| | 149 | //\r |
| | 150 | //\r |
| | 151 | // NOTE: Eventually we should have the interpreter allocate memory\r |
| | 152 | // for stack space which it will use during its execution. This\r |
| | 153 | // would likely improve performance because the interpreter would\r |
| | 154 | // no longer be required to test each memory access and adjust\r |
| | 155 | // those reading from the stack gap.\r |
| | 156 | //\r |
| | 157 | // For IPF, the stack looks like (assuming 10 args passed)\r |
| | 158 | // arg10\r |
| | 159 | // arg9 (Bottom of high stack)\r |
| | 160 | // [ stack gap for interpreter execution ]\r |
| | 161 | // [ magic value for detection of stack corruption ]\r |
| | 162 | // arg8 (Top of low stack)\r |
| | 163 | // arg7....\r |
| | 164 | // arg1\r |
| | 165 | // [ 64-bit return address ]\r |
| | 166 | // [ ebc stack ]\r |
| | 167 | // If the EBC accesses memory in the stack gap, then we assume that it's\r |
| | 168 | // actually trying to access args9 and greater. Therefore we need to\r |
| | 169 | // adjust memory accesses in this region to point above the stack gap.\r |
| | 170 | //\r |
| | 171 | //\r |
| | 172 | // Now adjust the EBC stack pointer down to leave a gap for interpreter\r |
| | 173 | // execution. Then stuff a magic value there.\r |
| | 174 | //\r |
| | 175 | \r |
| | 176 | Status = GetEBCStack((EFI_HANDLE)(UINTN)-1, &VmContext.StackPool, &StackIndex);\r |
| | 177 | if (EFI_ERROR(Status)) {\r |
| | 178 | return Status;\r |
| | 179 | }\r |
| | 180 | VmContext.StackTop = (UINT8*)VmContext.StackPool + (STACK_REMAIN_SIZE);\r |
| | 181 | VmContext.Gpr[0] = (UINT64) ((UINT8*)VmContext.StackPool + STACK_POOL_SIZE);\r |
| | 182 | VmContext.HighStackBottom = (UINTN) VmContext.Gpr[0];\r |
| | 183 | VmContext.Gpr[0] -= sizeof (UINTN);\r |
| | 184 | \r |
| | 185 | \r |
| | 186 | PushU64 (&VmContext, (UINT64) VM_STACK_KEY_VALUE);\r |
| | 187 | VmContext.StackMagicPtr = (UINTN *) VmContext.Gpr[0];\r |
| | 188 | VmContext.LowStackTop = (UINTN) VmContext.Gpr[0];\r |
| | 189 | //\r |
| | 190 | // Push the EBC arguments on the stack. Does not matter that they may not\r |
| | 191 | // all be valid.\r |
| | 192 | //\r |
| | 193 | PushU64 (&VmContext, Arg16);\r |
| | 194 | PushU64 (&VmContext, Arg15);\r |
| | 195 | PushU64 (&VmContext, Arg14);\r |
| | 196 | PushU64 (&VmContext, Arg13);\r |
| | 197 | PushU64 (&VmContext, Arg12);\r |
| | 198 | PushU64 (&VmContext, Arg11);\r |
| | 199 | PushU64 (&VmContext, Arg10);\r |
| | 200 | PushU64 (&VmContext, Arg9);\r |
| | 201 | PushU64 (&VmContext, Arg8);\r |
| | 202 | PushU64 (&VmContext, Arg7);\r |
| | 203 | PushU64 (&VmContext, Arg6);\r |
| | 204 | PushU64 (&VmContext, Arg5);\r |
| | 205 | PushU64 (&VmContext, Arg4);\r |
| | 206 | PushU64 (&VmContext, Arg3);\r |
| | 207 | PushU64 (&VmContext, Arg2);\r |
| | 208 | PushU64 (&VmContext, Arg1);\r |
| | 209 | //\r |
| | 210 | // Push a bogus return address on the EBC stack because the\r |
| | 211 | // interpreter expects one there. For stack alignment purposes on IPF,\r |
| | 212 | // EBC return addresses are always 16 bytes. Push a bogus value as well.\r |
| | 213 | //\r |
| | 214 | PushU64 (&VmContext, 0);\r |
| | 215 | PushU64 (&VmContext, 0xDEADBEEFDEADBEEF);\r |
| | 216 | VmContext.StackRetAddr = (UINT64) VmContext.Gpr[0];\r |
| | 217 | //\r |
| | 218 | // Begin executing the EBC code\r |
| | 219 | //\r |
| | 220 | EbcExecute (&VmContext);\r |
| | 221 | //\r |
| | 222 | // Return the value in R[7] unless there was an error\r |
| | 223 | //\r |
| | 224 | ReturnEBCStack(StackIndex);\r |
| | 225 | return (UINT64) VmContext.Gpr[7];\r |
| | 226 | }\r |
| | 227 | \r |
| | 228 | \r |
| | 229 | /**\r |
| | 230 | Begin executing an EBC image. The address of the entry point is passed\r |
| | 231 | in via a processor register, so we'll need to make a call to get the\r |
| | 232 | value.\r |
| | 233 | \r |
| | 234 | @param ImageHandle image handle for the EBC application we're executing\r |
| | 235 | @param SystemTable standard system table passed into an driver's entry\r |
| | 236 | point\r |
| | 237 | \r |
| | 238 | @return The value returned by the EBC application we're going to run.\r |
| | 239 | \r |
| | 240 | **/\r |
| | 241 | UINT64\r |
| | 242 | EFIAPI\r |
| | 243 | ExecuteEbcImageEntryPoint (\r |
| | 244 | IN EFI_HANDLE ImageHandle,\r |
| | 245 | IN EFI_SYSTEM_TABLE *SystemTable\r |
| | 246 | )\r |
| | 247 | {\r |
| | 248 | //\r |
| | 249 | // Create a new VM context on the stack\r |
| | 250 | //\r |
| | 251 | VM_CONTEXT VmContext;\r |
| | 252 | UINTN Addr;\r |
| | 253 | EFI_STATUS Status;\r |
| | 254 | UINTN StackIndex;\r |
| | 255 | \r |
| | 256 | //\r |
| | 257 | // Get the EBC entry point from the processor register. Make sure you don't\r |
| | 258 | // call any functions before this or you could mess up the register the\r |
| | 259 | // entry point is passed in.\r |
| | 260 | //\r |
| | 261 | Addr = EbcLLGetEbcEntryPoint ();\r |
| | 262 | \r |
| | 263 | //\r |
| | 264 | // Now clear out our context\r |
| | 265 | //\r |
| | 266 | ZeroMem ((VOID *) &VmContext, sizeof (VM_CONTEXT));\r |
| | 267 | \r |
| | 268 | //\r |
| | 269 | // Save the image handle so we can track the thunks created for this image\r |
| | 270 | //\r |
| | 271 | VmContext.ImageHandle = ImageHandle;\r |
| | 272 | VmContext.SystemTable = SystemTable;\r |
| | 273 | \r |
| | 274 | //\r |
| | 275 | // Set the VM instruction pointer to the correct location in memory.\r |
| | 276 | //\r |
| | 277 | VmContext.Ip = (VMIP) Addr;\r |
| | 278 | \r |
| | 279 | //\r |
| | 280 | // Get the stack pointer. This is the bottom of the upper stack.\r |
| | 281 | //\r |
| | 282 | \r |
| | 283 | Status = GetEBCStack(ImageHandle, &VmContext.StackPool, &StackIndex);\r |
| | 284 | if (EFI_ERROR(Status)) {\r |
| | 285 | return Status;\r |
| | 286 | }\r |
| | 287 | VmContext.StackTop = (UINT8*)VmContext.StackPool + (STACK_REMAIN_SIZE);\r |
| | 288 | VmContext.Gpr[0] = (UINT64) ((UINT8*)VmContext.StackPool + STACK_POOL_SIZE);\r |
| | 289 | VmContext.HighStackBottom = (UINTN) VmContext.Gpr[0];\r |
| | 290 | VmContext.Gpr[0] -= sizeof (UINTN);\r |
| | 291 | \r |
| | 292 | \r |
| | 293 | //\r |
| | 294 | // Allocate stack space for the interpreter. Then put a magic value\r |
| | 295 | // at the bottom so we can detect stack corruption.\r |
| | 296 | //\r |
| | 297 | PushU64 (&VmContext, (UINT64) VM_STACK_KEY_VALUE);\r |
| | 298 | VmContext.StackMagicPtr = (UINTN *) (UINTN) VmContext.Gpr[0];\r |
| | 299 | \r |
| | 300 | //\r |
| | 301 | // When we thunk to external native code, we copy the last 8 qwords from\r |
| | 302 | // the EBC stack into the processor registers, and adjust the stack pointer\r |
| | 303 | // up. If the caller is not passing 8 parameters, then we've moved the\r |
| | 304 | // stack pointer up into the stack gap. If this happens, then the caller\r |
| | 305 | // can mess up the stack gap contents (in particular our magic value).\r |
| | 306 | // Therefore, leave another gap below the magic value. Pick 10 qwords down,\r |
| | 307 | // just as a starting point.\r |
| | 308 | //\r |
| | 309 | VmContext.Gpr[0] -= 10 * sizeof (UINT64);\r |
| | 310 | \r |
| | 311 | //\r |
| | 312 | // Align the stack pointer such that after pushing the system table,\r |
| | 313 | // image handle, and return address on the stack, it's aligned on a 16-byte\r |
| | 314 | // boundary as required for IPF.\r |
| | 315 | //\r |
| | 316 | VmContext.Gpr[0] &= (INT64)~0x0f;\r |
| | 317 | VmContext.LowStackTop = (UINTN) VmContext.Gpr[0];\r |
| | 318 | //\r |
| | 319 | // Simply copy the image handle and system table onto the EBC stack.\r |
| | 320 | // Greatly simplifies things by not having to spill the args\r |
| | 321 | //\r |
| | 322 | PushU64 (&VmContext, (UINT64) SystemTable);\r |
| | 323 | PushU64 (&VmContext, (UINT64) ImageHandle);\r |
| | 324 | \r |
| | 325 | //\r |
| | 326 | // Interpreter assumes 64-bit return address is pushed on the stack.\r |
| | 327 | // IPF does not do this so pad the stack accordingly. Also, a\r |
| | 328 | // "return address" is 16 bytes as required for IPF stack alignments.\r |
| | 329 | //\r |
| | 330 | PushU64 (&VmContext, (UINT64) 0);\r |
| | 331 | PushU64 (&VmContext, (UINT64) 0x1234567887654321);\r |
| | 332 | VmContext.StackRetAddr = (UINT64) VmContext.Gpr[0];\r |
| | 333 | \r |
| | 334 | //\r |
| | 335 | // Begin executing the EBC code\r |
| | 336 | //\r |
| | 337 | EbcExecute (&VmContext);\r |
| | 338 | \r |
| | 339 | //\r |
| | 340 | // Return the value in R[7] unless there was an error\r |
| | 341 | //\r |
| | 342 | ReturnEBCStack(StackIndex);\r |
| | 343 | return (UINT64) VmContext.Gpr[7];\r |
| | 344 | }\r |
| | 345 | \r |
| | 346 | \r |
| | 347 | /**\r |
| | 348 | Create thunks for an EBC image entry point, or an EBC protocol service.\r |
| | 349 | \r |
| | 350 | @param ImageHandle Image handle for the EBC image. If not null, then\r |
| | 351 | we're creating a thunk for an image entry point.\r |
| | 352 | @param EbcEntryPoint Address of the EBC code that the thunk is to call\r |
| | 353 | @param Thunk Returned thunk we create here\r |
| | 354 | @param Flags Flags indicating options for creating the thunk\r |
| | 355 | \r |
| | 356 | @retval EFI_SUCCESS The thunk was created successfully.\r |
| | 357 | @retval EFI_INVALID_PARAMETER The parameter of EbcEntryPoint is not 16-bit\r |
| | 358 | aligned.\r |
| | 359 | @retval EFI_OUT_OF_RESOURCES There is not enough memory to created the EBC\r |
| | 360 | Thunk.\r |
| | 361 | @retval EFI_BUFFER_TOO_SMALL EBC_THUNK_SIZE is not larger enough.\r |
| | 362 | \r |
| | 363 | **/\r |
| | 364 | EFI_STATUS\r |
| | 365 | EbcCreateThunks (\r |
| | 366 | IN EFI_HANDLE ImageHandle,\r |
| | 367 | IN VOID *EbcEntryPoint,\r |
| | 368 | OUT VOID **Thunk,\r |
| | 369 | IN UINT32 Flags\r |
| | 370 | )\r |
| | 371 | {\r |
| | 372 | UINT8 *Ptr;\r |
| | 373 | UINT8 *ThunkBase;\r |
| | 374 | UINT64 Addr;\r |
| | 375 | UINT64 Code[3]; // Code in a bundle\r |
| | 376 | UINT64 RegNum; // register number for MOVL\r |
| | 377 | UINT64 BitI; // bits of MOVL immediate data\r |
| | 378 | UINT64 BitIc; // bits of MOVL immediate data\r |
| | 379 | UINT64 BitImm5c; // bits of MOVL immediate data\r |
| | 380 | UINT64 BitImm9d; // bits of MOVL immediate data\r |
| | 381 | UINT64 BitImm7b; // bits of MOVL immediate data\r |
| | 382 | UINT64 Br; // branch register for loading and jumping\r |
| | 383 | UINT64 *Data64Ptr;\r |
| | 384 | UINT32 ThunkSize;\r |
| | 385 | UINT32 Size;\r |
| | 386 | \r |
| | 387 | //\r |
| | 388 | // Check alignment of pointer to EBC code, which must always be aligned\r |
| | 389 | // on a 2-byte boundary.\r |
| | 390 | //\r |
| | 391 | if ((UINT32) (UINTN) EbcEntryPoint & 0x01) {\r |
| | 392 | return EFI_INVALID_PARAMETER;\r |
| | 393 | }\r |
| | 394 | //\r |
| | 395 | // Allocate memory for the thunk. Make the (most likely incorrect) assumption\r |
| | 396 | // that the returned buffer is not aligned, so round up to the next\r |
| | 397 | // alignment size.\r |
| | 398 | //\r |
| | 399 | Size = EBC_THUNK_SIZE + EBC_THUNK_ALIGNMENT - 1;\r |
| | 400 | ThunkSize = Size;\r |
| | 401 | Ptr = AllocatePool (Size);\r |
| | 402 | \r |
| | 403 | if (Ptr == NULL) {\r |
| | 404 | return EFI_OUT_OF_RESOURCES;\r |
| | 405 | }\r |
| | 406 | //\r |
| | 407 | // Save the start address of the buffer.\r |
| | 408 | //\r |
| | 409 | ThunkBase = Ptr;\r |
| | 410 | \r |
| | 411 | //\r |
| | 412 | // Make sure it's aligned for code execution. If not, then\r |
| | 413 | // round up.\r |
| | 414 | //\r |
| | 415 | if ((UINT32) (UINTN) Ptr & (EBC_THUNK_ALIGNMENT - 1)) {\r |
| | 416 | Ptr = (UINT8 *) (((UINTN) Ptr + (EBC_THUNK_ALIGNMENT - 1)) &~ (UINT64) (EBC_THUNK_ALIGNMENT - 1));\r |
| | 417 | }\r |
| | 418 | //\r |
| | 419 | // Return the pointer to the thunk to the caller to user as the\r |
| | 420 | // image entry point.\r |
| | 421 | //\r |
| | 422 | *Thunk = (VOID *) Ptr;\r |
| | 423 | \r |
| | 424 | //\r |
| | 425 | // Clear out the thunk entry\r |
| | 426 | // ZeroMem(Ptr, Size);\r |
| | 427 | //\r |
| | 428 | // For IPF, when you do a call via a function pointer, the function pointer\r |
| | 429 | // actually points to a function descriptor which consists of a 64-bit\r |
| | 430 | // address of the function, followed by a 64-bit gp for the function being\r |
| | 431 | // called. See the the Software Conventions and Runtime Architecture Guide\r |
| | 432 | // for details.\r |
| | 433 | // So first off in our thunk, create a descriptor for our actual thunk code.\r |
| | 434 | // This means we need to create a pointer to the thunk code (which follows\r |
| | 435 | // the descriptor we're going to create), followed by the gp of the Vm\r |
| | 436 | // interpret function we're going to eventually execute.\r |
| | 437 | //\r |
| | 438 | Data64Ptr = (UINT64 *) Ptr;\r |
| | 439 | \r |
| | 440 | //\r |
| | 441 | // Write the function's entry point (which is our thunk code that follows\r |
| | 442 | // this descriptor we're creating).\r |
| | 443 | //\r |
| | 444 | *Data64Ptr = (UINT64) (Data64Ptr + 2);\r |
| | 445 | //\r |
| | 446 | // Get the gp from the descriptor for EbcInterpret and stuff it in our thunk\r |
| | 447 | // descriptor.\r |
| | 448 | //\r |
| | 449 | *(Data64Ptr + 1) = *(UINT64 *) ((UINT64 *) (UINTN) EbcInterpret + 1);\r |
| | 450 | //\r |
| | 451 | // Advance our thunk data pointer past the descriptor. Since the\r |
| | 452 | // descriptor consists of 16 bytes, the pointer is still aligned for\r |
| | 453 | // IPF code execution (on 16-byte boundary).\r |
| | 454 | //\r |
| | 455 | Ptr += sizeof (UINT64) * 2;\r |
| | 456 | \r |
| | 457 | //\r |
| | 458 | // *************************** MAGIC BUNDLE ********************************\r |
| | 459 | //\r |
| | 460 | // Write magic code bundle for: movl r8 = 0xca112ebcca112ebc to help the VM\r |
| | 461 | // to recognize it is a thunk.\r |
| | 462 | //\r |
| | 463 | Addr = (UINT64) 0xCA112EBCCA112EBC;\r |
| | 464 | \r |
| | 465 | //\r |
| | 466 | // Now generate the code bytes. First is nop.m 0x0\r |
| | 467 | //\r |
| | 468 | Code[0] = OPCODE_NOP;\r |
| | 469 | \r |
| | 470 | //\r |
| | 471 | // Next is simply Addr[62:22] (41 bits) of the address\r |
| | 472 | //\r |
| | 473 | Code[1] = RShiftU64 (Addr, 22) & 0x1ffffffffff;\r |
| | 474 | \r |
| | 475 | //\r |
| | 476 | // Extract bits from the address for insertion into the instruction\r |
| | 477 | // i = Addr[63:63]\r |
| | 478 | //\r |
| | 479 | BitI = RShiftU64 (Addr, 63) & 0x01;\r |
| | 480 | //\r |
| | 481 | // ic = Addr[21:21]\r |
| | 482 | //\r |
| | 483 | BitIc = RShiftU64 (Addr, 21) & 0x01;\r |
| | 484 | //\r |
| | 485 | // imm5c = Addr[20:16] for 5 bits\r |
| | 486 | //\r |
| | 487 | BitImm5c = RShiftU64 (Addr, 16) & 0x1F;\r |
| | 488 | //\r |
| | 489 | // imm9d = Addr[15:7] for 9 bits\r |
| | 490 | //\r |
| | 491 | BitImm9d = RShiftU64 (Addr, 7) & 0x1FF;\r |
| | 492 | //\r |
| | 493 | // imm7b = Addr[6:0] for 7 bits\r |
| | 494 | //\r |
| | 495 | BitImm7b = Addr & 0x7F;\r |
| | 496 | \r |
| | 497 | //\r |
| | 498 | // The EBC entry point will be put into r8, so r8 can be used here\r |
| | 499 | // temporary. R8 is general register and is auto-serialized.\r |
| | 500 | //\r |
| | 501 | RegNum = 8;\r |
| | 502 | \r |
| | 503 | //\r |
| | 504 | // Next is jumbled data, including opcode and rest of address\r |
| | 505 | //\r |
| | 506 | Code[2] = LShiftU64 (BitImm7b, 13);\r |
| | 507 | Code[2] = Code[2] | LShiftU64 (0x00, 20); // vc\r |
| | 508 | Code[2] = Code[2] | LShiftU64 (BitIc, 21);\r |
| | 509 | Code[2] = Code[2] | LShiftU64 (BitImm5c, 22);\r |
| | 510 | Code[2] = Code[2] | LShiftU64 (BitImm9d, 27);\r |
| | 511 | Code[2] = Code[2] | LShiftU64 (BitI, 36);\r |
| | 512 | Code[2] = Code[2] | LShiftU64 ((UINT64)MOVL_OPCODE, 37);\r |
| | 513 | Code[2] = Code[2] | LShiftU64 ((RegNum & 0x7F), 6);\r |
| | 514 | \r |
| | 515 | WriteBundle ((VOID *) Ptr, 0x05, Code[0], Code[1], Code[2]);\r |
| | 516 | \r |
| | 517 | //\r |
| | 518 | // *************************** FIRST BUNDLE ********************************\r |
| | 519 | //\r |
| | 520 | // Write code bundle for: movl r8 = EBC_ENTRY_POINT so we pass\r |
| | 521 | // the ebc entry point in to the interpreter function via a processor\r |
| | 522 | // register.\r |
| | 523 | // Note -- we could easily change this to pass in a pointer to a structure\r |
| | 524 | // that contained, among other things, the EBC image's entry point. But\r |
| | 525 | // for now pass it directly.\r |
| | 526 | //\r |
| | 527 | Ptr += 16;\r |
| | 528 | Addr = (UINT64) EbcEntryPoint;\r |
| | 529 | \r |
| | 530 | //\r |
| | 531 | // Now generate the code bytes. First is nop.m 0x0\r |
| | 532 | //\r |
| | 533 | Code[0] = OPCODE_NOP;\r |
| | 534 | \r |
| | 535 | //\r |
| | 536 | // Next is simply Addr[62:22] (41 bits) of the address\r |
| | 537 | //\r |
| | 538 | Code[1] = RShiftU64 (Addr, 22) & 0x1ffffffffff;\r |
| | 539 | \r |
| | 540 | //\r |
| | 541 | // Extract bits from the address for insertion into the instruction\r |
| | 542 | // i = Addr[63:63]\r |
| | 543 | //\r |
| | 544 | BitI = RShiftU64 (Addr, 63) & 0x01;\r |
| | 545 | //\r |
| | 546 | // ic = Addr[21:21]\r |
| | 547 | //\r |
| | 548 | BitIc = RShiftU64 (Addr, 21) & 0x01;\r |
| | 549 | //\r |
| | 550 | // imm5c = Addr[20:16] for 5 bits\r |
| | 551 | //\r |
| | 552 | BitImm5c = RShiftU64 (Addr, 16) & 0x1F;\r |
| | 553 | //\r |
| | 554 | // imm9d = Addr[15:7] for 9 bits\r |
| | 555 | //\r |
| | 556 | BitImm9d = RShiftU64 (Addr, 7) & 0x1FF;\r |
| | 557 | //\r |
| | 558 | // imm7b = Addr[6:0] for 7 bits\r |
| | 559 | //\r |
| | 560 | BitImm7b = Addr & 0x7F;\r |
| | 561 | \r |
| | 562 | //\r |
| | 563 | // Put the EBC entry point in r8, which is the location of the return value\r |
| | 564 | // for functions.\r |
| | 565 | //\r |
| | 566 | RegNum = 8;\r |
| | 567 | \r |
| | 568 | //\r |
| | 569 | // Next is jumbled data, including opcode and rest of address\r |
| | 570 | //\r |
| | 571 | Code[2] = LShiftU64 (BitImm7b, 13);\r |
| | 572 | Code[2] = Code[2] | LShiftU64 (0x00, 20); // vc\r |
| | 573 | Code[2] = Code[2] | LShiftU64 (BitIc, 21);\r |
| | 574 | Code[2] = Code[2] | LShiftU64 (BitImm5c, 22);\r |
| | 575 | Code[2] = Code[2] | LShiftU64 (BitImm9d, 27);\r |
| | 576 | Code[2] = Code[2] | LShiftU64 (BitI, 36);\r |
| | 577 | Code[2] = Code[2] | LShiftU64 ((UINT64)MOVL_OPCODE, 37);\r |
| | 578 | Code[2] = Code[2] | LShiftU64 ((RegNum & 0x7F), 6);\r |
| | 579 | \r |
| | 580 | WriteBundle ((VOID *) Ptr, 0x05, Code[0], Code[1], Code[2]);\r |
| | 581 | \r |
| | 582 | //\r |
| | 583 | // *************************** NEXT BUNDLE *********************************\r |
| | 584 | //\r |
| | 585 | // Write code bundle for:\r |
| | 586 | // movl rx = offset_of(EbcInterpret|ExecuteEbcImageEntryPoint)\r |
| | 587 | //\r |
| | 588 | // Advance pointer to next bundle, then compute the offset from this bundle\r |
| | 589 | // to the address of the entry point of the interpreter.\r |
| | 590 | //\r |
| | 591 | Ptr += 16;\r |
| | 592 | if ((Flags & FLAG_THUNK_ENTRY_POINT) != 0) {\r |
| | 593 | Addr = (UINT64) ExecuteEbcImageEntryPoint;\r |
| | 594 | } else {\r |
| | 595 | Addr = (UINT64) EbcInterpret;\r |
| | 596 | }\r |
| | 597 | //\r |
| | 598 | // Indirection on Itanium-based systems\r |
| | 599 | //\r |
| | 600 | Addr = *(UINT64 *) Addr;\r |
| | 601 | \r |
| | 602 | //\r |
| | 603 | // Now write the code to load the offset into a register\r |
| | 604 | //\r |
| | 605 | Code[0] = OPCODE_NOP;\r |
| | 606 | \r |
| | 607 | //\r |
| | 608 | // Next is simply Addr[62:22] (41 bits) of the address\r |
| | 609 | //\r |
| | 610 | Code[1] = RShiftU64 (Addr, 22) & 0x1ffffffffff;\r |
| | 611 | \r |
| | 612 | //\r |
| | 613 | // Extract bits from the address for insertion into the instruction\r |
| | 614 | // i = Addr[63:63]\r |
| | 615 | //\r |
| | 616 | BitI = RShiftU64 (Addr, 63) & 0x01;\r |
| | 617 | //\r |
| | 618 | // ic = Addr[21:21]\r |
| | 619 | //\r |
| | 620 | BitIc = RShiftU64 (Addr, 21) & 0x01;\r |
| | 621 | //\r |
| | 622 | // imm5c = Addr[20:16] for 5 bits\r |
| | 623 | //\r |
| | 624 | BitImm5c = RShiftU64 (Addr, 16) & 0x1F;\r |
| | 625 | //\r |
| | 626 | // imm9d = Addr[15:7] for 9 bits\r |
| | 627 | //\r |
| | 628 | BitImm9d = RShiftU64 (Addr, 7) & 0x1FF;\r |
| | 629 | //\r |
| | 630 | // imm7b = Addr[6:0] for 7 bits\r |
| | 631 | //\r |
| | 632 | BitImm7b = Addr & 0x7F;\r |
| | 633 | \r |
| | 634 | //\r |
| | 635 | // Put it in r31, a scratch register\r |
| | 636 | //\r |
| | 637 | RegNum = 31;\r |
| | 638 | \r |
| | 639 | //\r |
| | 640 | // Next is jumbled data, including opcode and rest of address\r |
| | 641 | //\r |
| | 642 | Code[2] = LShiftU64(BitImm7b, 13);\r |
| | 643 | Code[2] = Code[2] | LShiftU64 (0x00, 20); // vc\r |
| | 644 | Code[2] = Code[2] | LShiftU64 (BitIc, 21);\r |
| | 645 | Code[2] = Code[2] | LShiftU64 (BitImm5c, 22);\r |
| | 646 | Code[2] = Code[2] | LShiftU64 (BitImm9d, 27);\r |
| | 647 | Code[2] = Code[2] | LShiftU64 (BitI, 36);\r |
| | 648 | Code[2] = Code[2] | LShiftU64 ((UINT64)MOVL_OPCODE, 37);\r |
| | 649 | Code[2] = Code[2] | LShiftU64 ((RegNum & 0x7F), 6);\r |
| | 650 | \r |
| | 651 | WriteBundle ((VOID *) Ptr, 0x05, Code[0], Code[1], Code[2]);\r |
| | 652 | \r |
| | 653 | //\r |
| | 654 | // *************************** NEXT BUNDLE *********************************\r |
| | 655 | //\r |
| | 656 | // Load branch register with EbcInterpret() function offset from the bundle\r |
| | 657 | // address: mov b6 = RegNum\r |
| | 658 | //\r |
| | 659 | // See volume 3 page 4-29 of the Arch. Software Developer's Manual.\r |
| | 660 | //\r |
| | 661 | // Advance pointer to next bundle\r |
| | 662 | //\r |
| | 663 | Ptr += 16;\r |
| | 664 | Code[0] = OPCODE_NOP;\r |
| | 665 | Code[1] = OPCODE_NOP;\r |
| | 666 | Code[2] = OPCODE_MOV_BX_RX;\r |
| | 667 | \r |
| | 668 | //\r |
| | 669 | // Pick a branch register to use. Then fill in the bits for the branch\r |
| | 670 | // register and user register (same user register as previous bundle).\r |
| | 671 | //\r |
| | 672 | Br = 6;\r |
| | 673 | Code[2] |= LShiftU64 (Br, 6);\r |
| | 674 | Code[2] |= LShiftU64 (RegNum, 13);\r |
| | 675 | WriteBundle ((VOID *) Ptr, 0x0d, Code[0], Code[1], Code[2]);\r |
| | 676 | \r |
| | 677 | //\r |
| | 678 | // *************************** NEXT BUNDLE *********************************\r |
| | 679 | //\r |
| | 680 | // Now do the branch: (p0) br.cond.sptk.few b6\r |
| | 681 | //\r |
| | 682 | // Advance pointer to next bundle.\r |
| | 683 | // Fill in the bits for the branch register (same reg as previous bundle)\r |
| | 684 | //\r |
| | 685 | Ptr += 16;\r |
| | 686 | Code[0] = OPCODE_NOP;\r |
| | 687 | Code[1] = OPCODE_NOP;\r |
| | 688 | Code[2] = OPCODE_BR_COND_SPTK_FEW;\r |
| | 689 | Code[2] |= LShiftU64 (Br, 13);\r |
| | 690 | WriteBundle ((VOID *) Ptr, 0x1d, Code[0], Code[1], Code[2]);\r |
| | 691 | \r |
| | 692 | //\r |
| | 693 | // Add the thunk to our list of allocated thunks so we can do some cleanup\r |
| | 694 | // when the image is unloaded. Do this last since the Add function flushes\r |
| | 695 | // the instruction cache for us.\r |
| | 696 | //\r |
| | 697 | EbcAddImageThunk (ImageHandle, (VOID *) ThunkBase, ThunkSize);\r |
| | 698 | \r |
| | 699 | //\r |
| | 700 | // Done\r |
| | 701 | //\r |
| | 702 | return EFI_SUCCESS;\r |
| | 703 | }\r |
| | 704 | \r |
| | 705 | \r |
| | 706 | /**\r |
| | 707 | Given raw bytes of Itanium based code, format them into a bundle and\r |
| | 708 | write them out.\r |
| | 709 | \r |
| | 710 | @param MemPtr pointer to memory location to write the bundles\r |
| | 711 | to.\r |
| | 712 | @param Template 5-bit template.\r |
| | 713 | @param Slot0 Instruction slot 0 data for the bundle.\r |
| | 714 | @param Slot1 Instruction slot 1 data for the bundle.\r |
| | 715 | @param Slot2 Instruction slot 2 data for the bundle.\r |
| | 716 | \r |
| | 717 | @retval EFI_INVALID_PARAMETER Pointer is not aligned\r |
| | 718 | @retval EFI_INVALID_PARAMETER No more than 5 bits in template\r |
| | 719 | @retval EFI_INVALID_PARAMETER More than 41 bits used in code\r |
| | 720 | @retval EFI_SUCCESS All data is written.\r |
| | 721 | \r |
| | 722 | **/\r |
| | 723 | EFI_STATUS\r |
| | 724 | WriteBundle (\r |
| | 725 | IN VOID *MemPtr,\r |
| | 726 | IN UINT8 Template,\r |
| | 727 | IN UINT64 Slot0,\r |
| | 728 | IN UINT64 Slot1,\r |
| | 729 | IN UINT64 Slot2\r |
| | 730 | )\r |
| | 731 | {\r |
| | 732 | UINT8 *BPtr;\r |
| | 733 | UINT32 Index;\r |
| | 734 | UINT64 Low64;\r |
| | 735 | UINT64 High64;\r |
| | 736 | \r |
| | 737 | //\r |
| | 738 | // Verify pointer is aligned\r |
| | 739 | //\r |
| | 740 | if ((UINT64) MemPtr & 0xF) {\r |
| | 741 | return EFI_INVALID_PARAMETER;\r |
| | 742 | }\r |
| | 743 | //\r |
| | 744 | // Verify no more than 5 bits in template\r |
| | 745 | //\r |
| | 746 | if ((Template &~0x1F) != 0) {\r |
| | 747 | return EFI_INVALID_PARAMETER;\r |
| | 748 | }\r |
| | 749 | //\r |
| | 750 | // Verify max of 41 bits used in code\r |
| | 751 | //\r |
| | 752 | if (((Slot0 | Slot1 | Slot2) &~0x1ffffffffff) != 0) {\r |
| | 753 | return EFI_INVALID_PARAMETER;\r |
| | 754 | }\r |
| | 755 | \r |
| | 756 | Low64 = LShiftU64 (Slot1, 46);\r |
| | 757 | Low64 = Low64 | LShiftU64 (Slot0, 5) | Template;\r |
| | 758 | \r |
| | 759 | High64 = RShiftU64 (Slot1, 18);\r |
| | 760 | High64 = High64 | LShiftU64 (Slot2, 23);\r |
| | 761 | \r |
| | 762 | //\r |
| | 763 | // Now write it all out\r |
| | 764 | //\r |
| | 765 | BPtr = (UINT8 *) MemPtr;\r |
| | 766 | for (Index = 0; Index < 8; Index++) {\r |
| | 767 | *BPtr = (UINT8) Low64;\r |
| | 768 | Low64 = RShiftU64 (Low64, 8);\r |
| | 769 | BPtr++;\r |
| | 770 | }\r |
| | 771 | \r |
| | 772 | for (Index = 0; Index < 8; Index++) {\r |
| | 773 | *BPtr = (UINT8) High64;\r |
| | 774 | High64 = RShiftU64 (High64, 8);\r |
| | 775 | BPtr++;\r |
| | 776 | }\r |
| | 777 | \r |
| | 778 | return EFI_SUCCESS;\r |
| | 779 | }\r |
| | 780 | \r |
| | 781 | \r |
| | 782 | /**\r |
| | 783 | This function is called to execute an EBC CALLEX instruction.\r |
| | 784 | The function check the callee's content to see whether it is common native\r |
| | 785 | code or a thunk to another piece of EBC code.\r |
| | 786 | If the callee is common native code, use EbcLLCAllEXASM to manipulate,\r |
| | 787 | otherwise, set the VM->IP to target EBC code directly to avoid another VM\r |
| | 788 | be startup which cost time and stack space.\r |
| | 789 | \r |
| | 790 | @param VmPtr Pointer to a VM context.\r |
| | 791 | @param FuncAddr Callee's address\r |
| | 792 | @param NewStackPointer New stack pointer after the call\r |
| | 793 | @param FramePtr New frame pointer after the call\r |
| | 794 | @param Size The size of call instruction\r |
| | 795 | \r |
| | 796 | **/\r |
| | 797 | VOID\r |
| | 798 | EbcLLCALLEX (\r |
| | 799 | IN VM_CONTEXT *VmPtr,\r |
| | 800 | IN UINTN FuncAddr,\r |
| | 801 | IN UINTN NewStackPointer,\r |
| | 802 | IN VOID *FramePtr,\r |
| | 803 | IN UINT8 Size\r |
| | 804 | )\r |
| | 805 | {\r |
| | 806 | UINTN IsThunk;\r |
| | 807 | UINTN TargetEbcAddr;\r |
| | 808 | UINTN CodeOne18;\r |
| | 809 | UINTN CodeOne23;\r |
| | 810 | UINTN CodeTwoI;\r |
| | 811 | UINTN CodeTwoIc;\r |
| | 812 | UINTN CodeTwo7b;\r |
| | 813 | UINTN CodeTwo5c;\r |
| | 814 | UINTN CodeTwo9d;\r |
| | 815 | UINTN CalleeAddr;\r |
| | 816 | \r |
| | 817 | IsThunk = 1;\r |
| | 818 | TargetEbcAddr = 0;\r |
| | 819 | \r |
| | 820 | //\r |
| | 821 | // FuncAddr points to the descriptor of the target instructions.\r |
| | 822 | //\r |
| | 823 | CalleeAddr = *((UINT64 *)FuncAddr);\r |
| | 824 | \r |
| | 825 | //\r |
| | 826 | // Processor specific code to check whether the callee is a thunk to EBC.\r |
| | 827 | //\r |
| | 828 | if (*((UINT64 *)CalleeAddr) != 0xBCCA000100000005) {\r |
| | 829 | IsThunk = 0;\r |
| | 830 | goto Action;\r |
| | 831 | }\r |
| | 832 | if (*((UINT64 *)CalleeAddr + 1) != 0x697623C1004A112E) {\r |
| | 833 | IsThunk = 0;\r |
| | 834 | goto Action;\r |
| | 835 | }\r |
| | 836 | \r |
| | 837 | CodeOne18 = RShiftU64 (*((UINT64 *)CalleeAddr + 2), 46) & 0x3FFFF;\r |
| | 838 | CodeOne23 = (*((UINT64 *)CalleeAddr + 3)) & 0x7FFFFF;\r |
| | 839 | CodeTwoI = RShiftU64 (*((UINT64 *)CalleeAddr + 3), 59) & 0x1;\r |
| | 840 | CodeTwoIc = RShiftU64 (*((UINT64 *)CalleeAddr + 3), 44) & 0x1;\r |
| | 841 | CodeTwo7b = RShiftU64 (*((UINT64 *)CalleeAddr + 3), 36) & 0x7F;\r |
| | 842 | CodeTwo5c = RShiftU64 (*((UINT64 *)CalleeAddr + 3), 45) & 0x1F;\r |
| | 843 | CodeTwo9d = RShiftU64 (*((UINT64 *)CalleeAddr + 3), 50) & 0x1FF;\r |
| | 844 | \r |
| | 845 | TargetEbcAddr = CodeTwo7b;\r |
| | 846 | TargetEbcAddr = TargetEbcAddr | LShiftU64 (CodeTwo9d, 7);\r |
| | 847 | TargetEbcAddr = TargetEbcAddr | LShiftU64 (CodeTwo5c, 16);\r |
| | 848 | TargetEbcAddr = TargetEbcAddr | LShiftU64 (CodeTwoIc, 21);\r |
| | 849 | TargetEbcAddr = TargetEbcAddr | LShiftU64 (CodeOne18, 22);\r |
| | 850 | TargetEbcAddr = TargetEbcAddr | LShiftU64 (CodeOne23, 40);\r |
| | 851 | TargetEbcAddr = TargetEbcAddr | LShiftU64 (CodeTwoI, 63);\r |
| | 852 | \r |
| | 853 | Action:\r |
| | 854 | if (IsThunk == 1){\r |
| | 855 | //\r |
| | 856 | // The callee is a thunk to EBC, adjust the stack pointer down 16 bytes and\r |
| | 857 | // put our return address and frame pointer on the VM stack.\r |
| | 858 | // Then set the VM's IP to new EBC code.\r |
| | 859 | //\r |
| | 860 | VmPtr->Gpr[0] -= 8;\r |
| | 861 | VmWriteMemN (VmPtr, (UINTN) VmPtr->Gpr[0], (UINTN) FramePtr);\r |
| | 862 | VmPtr->FramePtr = (VOID *) (UINTN) VmPtr->Gpr[0];\r |
| | 863 | VmPtr->Gpr[0] -= 8;\r |
| | 864 | VmWriteMem64 (VmPtr, (UINTN) VmPtr->Gpr[0], (UINT64) (VmPtr->Ip + Size));\r |
| | 865 | \r |
| | 866 | VmPtr->Ip = (VMIP) (UINTN) TargetEbcAddr;\r |
| | 867 | } else {\r |
| | 868 | //\r |
| | 869 | // The callee is not a thunk to EBC, call native code,\r |
| | 870 | // and get return value.\r |
| | 871 | //\r |
| | 872 | VmPtr->Gpr[7] = EbcLLCALLEXNative (FuncAddr, NewStackPointer, FramePtr);\r |
| | 873 | \r |
| | 874 | //\r |
| | 875 | // Advance the IP.\r |
| | 876 | //\r |
| | 877 | VmPtr->Ip += Size;\r |
| | 878 | }\r |
| | 879 | }\r |
projects / mirror_edk2.git / blame_incremental