1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 22 /* 23 * Copyright 2007 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 #pragma ident "%Z%%M% %I% %E% SMI" 28 29 #include <sys/fasttrap_isa.h> 30 #include <sys/fasttrap_impl.h> 31 #include <sys/dtrace.h> 32 #include <sys/dtrace_impl.h> 33 #include <sys/cmn_err.h> 34 #include <sys/regset.h> 35 #include <sys/privregs.h> 36 #include <sys/segments.h> 37 #include <sys/x86_archext.h> 38 #include <sys/sysmacros.h> 39 #include <sys/trap.h> 40 #include <sys/archsystm.h> 41 42 /* 43 * Lossless User-Land Tracing on x86 44 * --------------------------------- 45 * 46 * The execution of most instructions is not dependent on the address; for 47 * these instructions it is sufficient to copy them into the user process's 48 * address space and execute them. To effectively single-step an instruction 49 * in user-land, we copy out the following sequence of instructions to scratch 50 * space in the user thread's ulwp_t structure. 51 * 52 * We then set the program counter (%eip or %rip) to point to this scratch 53 * space. Once execution resumes, the original instruction is executed and 54 * then control flow is redirected to what was originally the subsequent 55 * instruction. If the kernel attemps to deliver a signal while single- 56 * stepping, the signal is deferred and the program counter is moved into the 57 * second sequence of instructions. The second sequence ends in a trap into 58 * the kernel where the deferred signal is then properly handled and delivered. 59 * 60 * For instructions whose execute is position dependent, we perform simple 61 * emulation. These instructions are limited to control transfer 62 * instructions in 32-bit mode, but in 64-bit mode there's the added wrinkle 63 * of %rip-relative addressing that means that almost any instruction can be 64 * position dependent. For all the details on how we emulate generic 65 * instructions included %rip-relative instructions, see the code in 66 * fasttrap_pid_probe() below where we handle instructions of type 67 * FASTTRAP_T_COMMON (under the header: Generic Instruction Tracing). 68 */ 69 70 #define FASTTRAP_MODRM_MOD(modrm) (((modrm) >> 6) & 0x3) 71 #define FASTTRAP_MODRM_REG(modrm) (((modrm) >> 3) & 0x7) 72 #define FASTTRAP_MODRM_RM(modrm) ((modrm) & 0x7) 73 #define FASTTRAP_MODRM(mod, reg, rm) (((mod) << 6) | ((reg) << 3) | (rm)) 74 75 #define FASTTRAP_SIB_SCALE(sib) (((sib) >> 6) & 0x3) 76 #define FASTTRAP_SIB_INDEX(sib) (((sib) >> 3) & 0x7) 77 #define FASTTRAP_SIB_BASE(sib) ((sib) & 0x7) 78 79 #define FASTTRAP_REX_W(rex) (((rex) >> 3) & 1) 80 #define FASTTRAP_REX_R(rex) (((rex) >> 2) & 1) 81 #define FASTTRAP_REX_X(rex) (((rex) >> 1) & 1) 82 #define FASTTRAP_REX_B(rex) ((rex) & 1) 83 #define FASTTRAP_REX(w, r, x, b) \ 84 (0x40 | ((w) << 3) | ((r) << 2) | ((x) << 1) | (b)) 85 86 /* 87 * Single-byte op-codes. 88 */ 89 #define FASTTRAP_PUSHL_EBP 0x55 90 91 #define FASTTRAP_JO 0x70 92 #define FASTTRAP_JNO 0x71 93 #define FASTTRAP_JB 0x72 94 #define FASTTRAP_JAE 0x73 95 #define FASTTRAP_JE 0x74 96 #define FASTTRAP_JNE 0x75 97 #define FASTTRAP_JBE 0x76 98 #define FASTTRAP_JA 0x77 99 #define FASTTRAP_JS 0x78 100 #define FASTTRAP_JNS 0x79 101 #define FASTTRAP_JP 0x7a 102 #define FASTTRAP_JNP 0x7b 103 #define FASTTRAP_JL 0x7c 104 #define FASTTRAP_JGE 0x7d 105 #define FASTTRAP_JLE 0x7e 106 #define FASTTRAP_JG 0x7f 107 108 #define FASTTRAP_NOP 0x90 109 110 #define FASTTRAP_MOV_EAX 0xb8 111 #define FASTTRAP_MOV_ECX 0xb9 112 113 #define FASTTRAP_RET16 0xc2 114 #define FASTTRAP_RET 0xc3 115 116 #define FASTTRAP_LOOPNZ 0xe0 117 #define FASTTRAP_LOOPZ 0xe1 118 #define FASTTRAP_LOOP 0xe2 119 #define FASTTRAP_JCXZ 0xe3 120 121 #define FASTTRAP_CALL 0xe8 122 #define FASTTRAP_JMP32 0xe9 123 #define FASTTRAP_JMP8 0xeb 124 125 #define FASTTRAP_INT3 0xcc 126 #define FASTTRAP_INT 0xcd 127 128 #define FASTTRAP_2_BYTE_OP 0x0f 129 #define FASTTRAP_GROUP5_OP 0xff 130 131 /* 132 * Two-byte op-codes (second byte only). 133 */ 134 #define FASTTRAP_0F_JO 0x80 135 #define FASTTRAP_0F_JNO 0x81 136 #define FASTTRAP_0F_JB 0x82 137 #define FASTTRAP_0F_JAE 0x83 138 #define FASTTRAP_0F_JE 0x84 139 #define FASTTRAP_0F_JNE 0x85 140 #define FASTTRAP_0F_JBE 0x86 141 #define FASTTRAP_0F_JA 0x87 142 #define FASTTRAP_0F_JS 0x88 143 #define FASTTRAP_0F_JNS 0x89 144 #define FASTTRAP_0F_JP 0x8a 145 #define FASTTRAP_0F_JNP 0x8b 146 #define FASTTRAP_0F_JL 0x8c 147 #define FASTTRAP_0F_JGE 0x8d 148 #define FASTTRAP_0F_JLE 0x8e 149 #define FASTTRAP_0F_JG 0x8f 150 151 #define FASTTRAP_EFLAGS_OF 0x800 152 #define FASTTRAP_EFLAGS_DF 0x400 153 #define FASTTRAP_EFLAGS_SF 0x080 154 #define FASTTRAP_EFLAGS_ZF 0x040 155 #define FASTTRAP_EFLAGS_AF 0x010 156 #define FASTTRAP_EFLAGS_PF 0x004 157 #define FASTTRAP_EFLAGS_CF 0x001 158 159 /* 160 * Instruction prefixes. 161 */ 162 #define FASTTRAP_PREFIX_OPERAND 0x66 163 #define FASTTRAP_PREFIX_ADDRESS 0x67 164 #define FASTTRAP_PREFIX_CS 0x2E 165 #define FASTTRAP_PREFIX_DS 0x3E 166 #define FASTTRAP_PREFIX_ES 0x26 167 #define FASTTRAP_PREFIX_FS 0x64 168 #define FASTTRAP_PREFIX_GS 0x65 169 #define FASTTRAP_PREFIX_SS 0x36 170 #define FASTTRAP_PREFIX_LOCK 0xF0 171 #define FASTTRAP_PREFIX_REP 0xF3 172 #define FASTTRAP_PREFIX_REPNE 0xF2 173 174 #define FASTTRAP_NOREG 0xff 175 176 /* 177 * Map between instruction register encodings and the kernel constants which 178 * correspond to indicies into struct regs. 179 */ 180 #ifdef __amd64 181 static const uint8_t regmap[16] = { 182 REG_RAX, REG_RCX, REG_RDX, REG_RBX, REG_RSP, REG_RBP, REG_RSI, REG_RDI, 183 REG_R8, REG_R9, REG_R10, REG_R11, REG_R12, REG_R13, REG_R14, REG_R15, 184 }; 185 #else 186 static const uint8_t regmap[8] = { 187 EAX, ECX, EDX, EBX, UESP, EBP, ESI, EDI 188 }; 189 #endif 190 191 static ulong_t fasttrap_getreg(struct regs *, uint_t); 192 193 static uint64_t 194 fasttrap_anarg(struct regs *rp, int function_entry, int argno) 195 { 196 uint64_t value; 197 int shift = function_entry ? 1 : 0; 198 199 #ifdef __amd64 200 if (curproc->p_model == DATAMODEL_LP64) { 201 uintptr_t *stack; 202 203 /* 204 * In 64-bit mode, the first six arguments are stored in 205 * registers. 206 */ 207 if (argno < 6) 208 return ((&rp->r_rdi)[argno]); 209 210 stack = (uintptr_t *)rp->r_sp; 211 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 212 value = dtrace_fulword(&stack[argno - 6 + shift]); 213 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT | CPU_DTRACE_BADADDR); 214 } else { 215 #endif 216 uint32_t *stack = (uint32_t *)rp->r_sp; 217 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 218 value = dtrace_fuword32(&stack[argno + shift]); 219 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT | CPU_DTRACE_BADADDR); 220 #ifdef __amd64 221 } 222 #endif 223 224 return (value); 225 } 226 227 /*ARGSUSED*/ 228 int 229 fasttrap_tracepoint_init(proc_t *p, fasttrap_tracepoint_t *tp, uintptr_t pc, 230 fasttrap_probe_type_t type) 231 { 232 uint8_t instr[FASTTRAP_MAX_INSTR_SIZE + 10]; 233 size_t len = FASTTRAP_MAX_INSTR_SIZE; 234 size_t first = MIN(len, PAGESIZE - (pc & PAGEOFFSET)); 235 uint_t start = 0; 236 int rmindex, size; 237 uint8_t seg, rex = 0; 238 239 /* 240 * Read the instruction at the given address out of the process's 241 * address space. We don't have to worry about a debugger 242 * changing this instruction before we overwrite it with our trap 243 * instruction since P_PR_LOCK is set. Since instructions can span 244 * pages, we potentially read the instruction in two parts. If the 245 * second part fails, we just zero out that part of the instruction. 246 */ 247 if (uread(p, &instr[0], first, pc) != 0) 248 return (-1); 249 if (len > first && 250 uread(p, &instr[first], len - first, pc + first) != 0) { 251 bzero(&instr[first], len - first); 252 len = first; 253 } 254 255 /* 256 * If the disassembly fails, then we have a malformed instruction. 257 */ 258 if ((size = dtrace_instr_size_isa(instr, p->p_model, &rmindex)) <= 0) 259 return (-1); 260 261 /* 262 * Make sure the disassembler isn't completely broken. 263 */ 264 ASSERT(-1 <= rmindex && rmindex < size); 265 266 /* 267 * If the computed size is greater than the number of bytes read, 268 * then it was a malformed instruction possibly because it fell on a 269 * page boundary and the subsequent page was missing or because of 270 * some malicious user. 271 */ 272 if (size > len) 273 return (-1); 274 275 tp->ftt_size = (uint8_t)size; 276 tp->ftt_segment = FASTTRAP_SEG_NONE; 277 278 /* 279 * Find the start of the instruction's opcode by processing any 280 * legacy prefixes. 281 */ 282 for (;;) { 283 seg = 0; 284 switch (instr[start]) { 285 case FASTTRAP_PREFIX_SS: 286 seg++; 287 /*FALLTHRU*/ 288 case FASTTRAP_PREFIX_GS: 289 seg++; 290 /*FALLTHRU*/ 291 case FASTTRAP_PREFIX_FS: 292 seg++; 293 /*FALLTHRU*/ 294 case FASTTRAP_PREFIX_ES: 295 seg++; 296 /*FALLTHRU*/ 297 case FASTTRAP_PREFIX_DS: 298 seg++; 299 /*FALLTHRU*/ 300 case FASTTRAP_PREFIX_CS: 301 seg++; 302 /*FALLTHRU*/ 303 case FASTTRAP_PREFIX_OPERAND: 304 case FASTTRAP_PREFIX_ADDRESS: 305 case FASTTRAP_PREFIX_LOCK: 306 case FASTTRAP_PREFIX_REP: 307 case FASTTRAP_PREFIX_REPNE: 308 if (seg != 0) { 309 /* 310 * It's illegal for an instruction to specify 311 * two segment prefixes -- give up on this 312 * illegal instruction. 313 */ 314 if (tp->ftt_segment != FASTTRAP_SEG_NONE) 315 return (-1); 316 317 tp->ftt_segment = seg; 318 } 319 start++; 320 continue; 321 } 322 break; 323 } 324 325 #ifdef __amd64 326 /* 327 * Identify the REX prefix on 64-bit processes. 328 */ 329 if (p->p_model == DATAMODEL_LP64 && (instr[start] & 0xf0) == 0x40) 330 rex = instr[start++]; 331 #endif 332 333 /* 334 * Now that we're pretty sure that the instruction is okay, copy the 335 * valid part to the tracepoint. 336 */ 337 bcopy(instr, tp->ftt_instr, FASTTRAP_MAX_INSTR_SIZE); 338 339 tp->ftt_type = FASTTRAP_T_COMMON; 340 if (instr[start] == FASTTRAP_2_BYTE_OP) { 341 switch (instr[start + 1]) { 342 case FASTTRAP_0F_JO: 343 case FASTTRAP_0F_JNO: 344 case FASTTRAP_0F_JB: 345 case FASTTRAP_0F_JAE: 346 case FASTTRAP_0F_JE: 347 case FASTTRAP_0F_JNE: 348 case FASTTRAP_0F_JBE: 349 case FASTTRAP_0F_JA: 350 case FASTTRAP_0F_JS: 351 case FASTTRAP_0F_JNS: 352 case FASTTRAP_0F_JP: 353 case FASTTRAP_0F_JNP: 354 case FASTTRAP_0F_JL: 355 case FASTTRAP_0F_JGE: 356 case FASTTRAP_0F_JLE: 357 case FASTTRAP_0F_JG: 358 tp->ftt_type = FASTTRAP_T_JCC; 359 tp->ftt_code = (instr[start + 1] & 0x0f) | FASTTRAP_JO; 360 tp->ftt_dest = pc + tp->ftt_size + 361 /* LINTED - alignment */ 362 *(int32_t *)&instr[start + 2]; 363 break; 364 } 365 } else if (instr[start] == FASTTRAP_GROUP5_OP) { 366 uint_t mod = FASTTRAP_MODRM_MOD(instr[start + 1]); 367 uint_t reg = FASTTRAP_MODRM_REG(instr[start + 1]); 368 uint_t rm = FASTTRAP_MODRM_RM(instr[start + 1]); 369 370 if (reg == 2 || reg == 4) { 371 uint_t i, sz; 372 373 if (reg == 2) 374 tp->ftt_type = FASTTRAP_T_CALL; 375 else 376 tp->ftt_type = FASTTRAP_T_JMP; 377 378 if (mod == 3) 379 tp->ftt_code = 2; 380 else 381 tp->ftt_code = 1; 382 383 ASSERT(p->p_model == DATAMODEL_LP64 || rex == 0); 384 385 /* 386 * See AMD x86-64 Architecture Programmer's Manual 387 * Volume 3, Section 1.2.7, Table 1-12, and 388 * Appendix A.3.1, Table A-15. 389 */ 390 if (mod != 3 && rm == 4) { 391 uint8_t sib = instr[start + 2]; 392 uint_t index = FASTTRAP_SIB_INDEX(sib); 393 uint_t base = FASTTRAP_SIB_BASE(sib); 394 395 tp->ftt_scale = FASTTRAP_SIB_SCALE(sib); 396 397 tp->ftt_index = (index == 4) ? 398 FASTTRAP_NOREG : 399 regmap[index | (FASTTRAP_REX_X(rex) << 3)]; 400 tp->ftt_base = (mod == 0 && base == 5) ? 401 FASTTRAP_NOREG : 402 regmap[base | (FASTTRAP_REX_B(rex) << 3)]; 403 404 i = 3; 405 sz = mod == 1 ? 1 : 4; 406 } else { 407 /* 408 * In 64-bit mode, mod == 0 and r/m == 5 409 * denotes %rip-relative addressing; in 32-bit 410 * mode, the base register isn't used. In both 411 * modes, there is a 32-bit operand. 412 */ 413 if (mod == 0 && rm == 5) { 414 #ifdef __amd64 415 if (p->p_model == DATAMODEL_LP64) 416 tp->ftt_base = REG_RIP; 417 else 418 #endif 419 tp->ftt_base = FASTTRAP_NOREG; 420 sz = 4; 421 } else { 422 uint8_t base = rm | 423 (FASTTRAP_REX_B(rex) << 3); 424 425 tp->ftt_base = regmap[base]; 426 sz = mod == 1 ? 1 : mod == 2 ? 4 : 0; 427 } 428 tp->ftt_index = FASTTRAP_NOREG; 429 i = 2; 430 } 431 432 if (sz == 1) { 433 tp->ftt_dest = *(int8_t *)&instr[start + i]; 434 } else if (sz == 4) { 435 /* LINTED - alignment */ 436 tp->ftt_dest = *(int32_t *)&instr[start + i]; 437 } else { 438 tp->ftt_dest = 0; 439 } 440 } 441 } else { 442 switch (instr[start]) { 443 case FASTTRAP_RET: 444 tp->ftt_type = FASTTRAP_T_RET; 445 break; 446 447 case FASTTRAP_RET16: 448 tp->ftt_type = FASTTRAP_T_RET16; 449 /* LINTED - alignment */ 450 tp->ftt_dest = *(uint16_t *)&instr[start + 1]; 451 break; 452 453 case FASTTRAP_JO: 454 case FASTTRAP_JNO: 455 case FASTTRAP_JB: 456 case FASTTRAP_JAE: 457 case FASTTRAP_JE: 458 case FASTTRAP_JNE: 459 case FASTTRAP_JBE: 460 case FASTTRAP_JA: 461 case FASTTRAP_JS: 462 case FASTTRAP_JNS: 463 case FASTTRAP_JP: 464 case FASTTRAP_JNP: 465 case FASTTRAP_JL: 466 case FASTTRAP_JGE: 467 case FASTTRAP_JLE: 468 case FASTTRAP_JG: 469 tp->ftt_type = FASTTRAP_T_JCC; 470 tp->ftt_code = instr[start]; 471 tp->ftt_dest = pc + tp->ftt_size + 472 (int8_t)instr[start + 1]; 473 break; 474 475 case FASTTRAP_LOOPNZ: 476 case FASTTRAP_LOOPZ: 477 case FASTTRAP_LOOP: 478 tp->ftt_type = FASTTRAP_T_LOOP; 479 tp->ftt_code = instr[start]; 480 tp->ftt_dest = pc + tp->ftt_size + 481 (int8_t)instr[start + 1]; 482 break; 483 484 case FASTTRAP_JCXZ: 485 tp->ftt_type = FASTTRAP_T_JCXZ; 486 tp->ftt_dest = pc + tp->ftt_size + 487 (int8_t)instr[start + 1]; 488 break; 489 490 case FASTTRAP_CALL: 491 tp->ftt_type = FASTTRAP_T_CALL; 492 tp->ftt_dest = pc + tp->ftt_size + 493 /* LINTED - alignment */ 494 *(int32_t *)&instr[start + 1]; 495 tp->ftt_code = 0; 496 break; 497 498 case FASTTRAP_JMP32: 499 tp->ftt_type = FASTTRAP_T_JMP; 500 tp->ftt_dest = pc + tp->ftt_size + 501 /* LINTED - alignment */ 502 *(int32_t *)&instr[start + 1]; 503 break; 504 case FASTTRAP_JMP8: 505 tp->ftt_type = FASTTRAP_T_JMP; 506 tp->ftt_dest = pc + tp->ftt_size + 507 (int8_t)instr[start + 1]; 508 break; 509 510 case FASTTRAP_PUSHL_EBP: 511 if (start == 0) 512 tp->ftt_type = FASTTRAP_T_PUSHL_EBP; 513 break; 514 515 case FASTTRAP_NOP: 516 #ifdef __amd64 517 ASSERT(p->p_model == DATAMODEL_LP64 || rex == 0); 518 519 /* 520 * On amd64 we have to be careful not to confuse a nop 521 * (actually xchgl %eax, %eax) with an instruction using 522 * the same opcode, but that does something different 523 * (e.g. xchgl %r8d, %eax or xcghq %r8, %rax). 524 */ 525 if (FASTTRAP_REX_B(rex) == 0) 526 #endif 527 tp->ftt_type = FASTTRAP_T_NOP; 528 break; 529 530 case FASTTRAP_INT3: 531 /* 532 * The pid provider shares the int3 trap with debugger 533 * breakpoints so we can't instrument them. 534 */ 535 ASSERT(instr[start] == FASTTRAP_INSTR); 536 return (-1); 537 538 case FASTTRAP_INT: 539 /* 540 * Interrupts seem like they could be traced with 541 * no negative implications, but it's possible that 542 * a thread could be redirected by the trap handling 543 * code which would eventually return to the 544 * instruction after the interrupt. If the interrupt 545 * were in our scratch space, the subsequent 546 * instruction might be overwritten before we return. 547 * Accordingly we refuse to instrument any interrupt. 548 */ 549 return (-1); 550 } 551 } 552 553 #ifdef __amd64 554 if (p->p_model == DATAMODEL_LP64 && tp->ftt_type == FASTTRAP_T_COMMON) { 555 /* 556 * If the process is 64-bit and the instruction type is still 557 * FASTTRAP_T_COMMON -- meaning we're going to copy it out an 558 * execute it -- we need to watch for %rip-relative 559 * addressing mode. See the portion of fasttrap_pid_probe() 560 * below where we handle tracepoints with type 561 * FASTTRAP_T_COMMON for how we emulate instructions that 562 * employ %rip-relative addressing. 563 */ 564 if (rmindex != -1) { 565 uint_t mod = FASTTRAP_MODRM_MOD(instr[rmindex]); 566 uint_t reg = FASTTRAP_MODRM_REG(instr[rmindex]); 567 uint_t rm = FASTTRAP_MODRM_RM(instr[rmindex]); 568 569 ASSERT(rmindex > start); 570 571 if (mod == 0 && rm == 5) { 572 /* 573 * We need to be sure to avoid other 574 * registers used by this instruction. While 575 * the reg field may determine the op code 576 * rather than denoting a register, assuming 577 * that it denotes a register is always safe. 578 * We leave the REX field intact and use 579 * whatever value's there for simplicity. 580 */ 581 if (reg != 0) { 582 tp->ftt_ripmode = FASTTRAP_RIP_1 | 583 (FASTTRAP_RIP_X * 584 FASTTRAP_REX_B(rex)); 585 rm = 0; 586 } else { 587 tp->ftt_ripmode = FASTTRAP_RIP_2 | 588 (FASTTRAP_RIP_X * 589 FASTTRAP_REX_B(rex)); 590 rm = 1; 591 } 592 593 tp->ftt_modrm = tp->ftt_instr[rmindex]; 594 tp->ftt_instr[rmindex] = 595 FASTTRAP_MODRM(2, reg, rm); 596 } 597 } 598 } 599 #endif 600 601 return (0); 602 } 603 604 int 605 fasttrap_tracepoint_install(proc_t *p, fasttrap_tracepoint_t *tp) 606 { 607 fasttrap_instr_t instr = FASTTRAP_INSTR; 608 609 if (uwrite(p, &instr, 1, tp->ftt_pc) != 0) 610 return (-1); 611 612 return (0); 613 } 614 615 int 616 fasttrap_tracepoint_remove(proc_t *p, fasttrap_tracepoint_t *tp) 617 { 618 uint8_t instr; 619 620 /* 621 * Distinguish between read or write failures and a changed 622 * instruction. 623 */ 624 if (uread(p, &instr, 1, tp->ftt_pc) != 0) 625 return (0); 626 if (instr != FASTTRAP_INSTR) 627 return (0); 628 if (uwrite(p, &tp->ftt_instr[0], 1, tp->ftt_pc) != 0) 629 return (-1); 630 631 return (0); 632 } 633 634 #ifdef __amd64 635 static uintptr_t 636 fasttrap_fulword_noerr(const void *uaddr) 637 { 638 uintptr_t ret; 639 640 if (fasttrap_fulword(uaddr, &ret) == 0) 641 return (ret); 642 643 return (0); 644 } 645 #endif 646 647 static uint32_t 648 fasttrap_fuword32_noerr(const void *uaddr) 649 { 650 uint32_t ret; 651 652 if (fasttrap_fuword32(uaddr, &ret) == 0) 653 return (ret); 654 655 return (0); 656 } 657 658 static void 659 fasttrap_return_common(struct regs *rp, uintptr_t pc, pid_t pid, 660 uintptr_t new_pc) 661 { 662 fasttrap_tracepoint_t *tp; 663 fasttrap_bucket_t *bucket; 664 fasttrap_id_t *id; 665 kmutex_t *pid_mtx; 666 667 pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock; 668 mutex_enter(pid_mtx); 669 bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)]; 670 671 for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) { 672 if (pid == tp->ftt_pid && pc == tp->ftt_pc && 673 tp->ftt_proc->ftpc_acount != 0) 674 break; 675 } 676 677 /* 678 * Don't sweat it if we can't find the tracepoint again; unlike 679 * when we're in fasttrap_pid_probe(), finding the tracepoint here 680 * is not essential to the correct execution of the process. 681 */ 682 if (tp == NULL) { 683 mutex_exit(pid_mtx); 684 return; 685 } 686 687 for (id = tp->ftt_retids; id != NULL; id = id->fti_next) { 688 /* 689 * If there's a branch that could act as a return site, we 690 * need to trace it, and check here if the program counter is 691 * external to the function. 692 */ 693 if (tp->ftt_type != FASTTRAP_T_RET && 694 tp->ftt_type != FASTTRAP_T_RET16 && 695 new_pc - id->fti_probe->ftp_faddr < 696 id->fti_probe->ftp_fsize) 697 continue; 698 699 dtrace_probe(id->fti_probe->ftp_id, 700 pc - id->fti_probe->ftp_faddr, 701 rp->r_r0, rp->r_r1, 0, 0); 702 } 703 704 mutex_exit(pid_mtx); 705 } 706 707 static void 708 fasttrap_sigsegv(proc_t *p, kthread_t *t, uintptr_t addr) 709 { 710 sigqueue_t *sqp = kmem_zalloc(sizeof (sigqueue_t), KM_SLEEP); 711 712 sqp->sq_info.si_signo = SIGSEGV; 713 sqp->sq_info.si_code = SEGV_MAPERR; 714 sqp->sq_info.si_addr = (caddr_t)addr; 715 716 mutex_enter(&p->p_lock); 717 sigaddqa(p, t, sqp); 718 mutex_exit(&p->p_lock); 719 720 if (t != NULL) 721 aston(t); 722 } 723 724 #ifdef __amd64 725 static void 726 fasttrap_usdt_args64(fasttrap_probe_t *probe, struct regs *rp, int argc, 727 uintptr_t *argv) 728 { 729 int i, x, cap = MIN(argc, probe->ftp_nargs); 730 uintptr_t *stack = (uintptr_t *)rp->r_sp; 731 732 for (i = 0; i < cap; i++) { 733 x = probe->ftp_argmap[i]; 734 735 if (x < 6) 736 argv[i] = (&rp->r_rdi)[x]; 737 else 738 argv[i] = fasttrap_fulword_noerr(&stack[x]); 739 } 740 741 for (; i < argc; i++) { 742 argv[i] = 0; 743 } 744 } 745 #endif 746 747 static void 748 fasttrap_usdt_args32(fasttrap_probe_t *probe, struct regs *rp, int argc, 749 uint32_t *argv) 750 { 751 int i, x, cap = MIN(argc, probe->ftp_nargs); 752 uint32_t *stack = (uint32_t *)rp->r_sp; 753 754 for (i = 0; i < cap; i++) { 755 x = probe->ftp_argmap[i]; 756 757 argv[i] = fasttrap_fuword32_noerr(&stack[x]); 758 } 759 760 for (; i < argc; i++) { 761 argv[i] = 0; 762 } 763 } 764 765 static int 766 fasttrap_do_seg(fasttrap_tracepoint_t *tp, struct regs *rp, uintptr_t *addr) 767 { 768 proc_t *p = curproc; 769 user_desc_t *desc; 770 uint16_t sel, ndx, type; 771 uintptr_t limit; 772 773 switch (tp->ftt_segment) { 774 case FASTTRAP_SEG_CS: 775 sel = rp->r_cs; 776 break; 777 case FASTTRAP_SEG_DS: 778 sel = rp->r_ds; 779 break; 780 case FASTTRAP_SEG_ES: 781 sel = rp->r_es; 782 break; 783 case FASTTRAP_SEG_FS: 784 sel = rp->r_fs; 785 break; 786 case FASTTRAP_SEG_GS: 787 sel = rp->r_gs; 788 break; 789 case FASTTRAP_SEG_SS: 790 sel = rp->r_ss; 791 break; 792 } 793 794 /* 795 * Make sure the given segment register specifies a user priority 796 * selector rather than a kernel selector. 797 */ 798 if (!SELISUPL(sel)) 799 return (-1); 800 801 ndx = SELTOIDX(sel); 802 803 /* 804 * Check the bounds and grab the descriptor out of the specified 805 * descriptor table. 806 */ 807 if (SELISLDT(sel)) { 808 if (ndx > p->p_ldtlimit) 809 return (-1); 810 811 desc = p->p_ldt + ndx; 812 813 } else { 814 if (ndx >= NGDT) 815 return (-1); 816 817 desc = cpu_get_gdt() + ndx; 818 } 819 820 /* 821 * The descriptor must have user privilege level and it must be 822 * present in memory. 823 */ 824 if (desc->usd_dpl != SEL_UPL || desc->usd_p != 1) 825 return (-1); 826 827 type = desc->usd_type; 828 829 /* 830 * If the S bit in the type field is not set, this descriptor can 831 * only be used in system context. 832 */ 833 if ((type & 0x10) != 0x10) 834 return (-1); 835 836 limit = USEGD_GETLIMIT(desc) * (desc->usd_gran ? PAGESIZE : 1); 837 838 if (tp->ftt_segment == FASTTRAP_SEG_CS) { 839 /* 840 * The code/data bit and readable bit must both be set. 841 */ 842 if ((type & 0xa) != 0xa) 843 return (-1); 844 845 if (*addr > limit) 846 return (-1); 847 } else { 848 /* 849 * The code/data bit must be clear. 850 */ 851 if ((type & 0x8) != 0) 852 return (-1); 853 854 /* 855 * If the expand-down bit is clear, we just check the limit as 856 * it would naturally be applied. Otherwise, we need to check 857 * that the address is the range [limit + 1 .. 0xffff] or 858 * [limit + 1 ... 0xffffffff] depending on if the default 859 * operand size bit is set. 860 */ 861 if ((type & 0x4) == 0) { 862 if (*addr > limit) 863 return (-1); 864 } else if (desc->usd_def32) { 865 if (*addr < limit + 1 || 0xffff < *addr) 866 return (-1); 867 } else { 868 if (*addr < limit + 1 || 0xffffffff < *addr) 869 return (-1); 870 } 871 } 872 873 *addr += USEGD_GETBASE(desc); 874 875 return (0); 876 } 877 878 int 879 fasttrap_pid_probe(struct regs *rp) 880 { 881 proc_t *p = curproc; 882 uintptr_t pc = rp->r_pc - 1, new_pc = 0; 883 fasttrap_bucket_t *bucket; 884 kmutex_t *pid_mtx; 885 fasttrap_tracepoint_t *tp, tp_local; 886 pid_t pid; 887 dtrace_icookie_t cookie; 888 uint_t is_enabled = 0; 889 890 /* 891 * It's possible that a user (in a veritable orgy of bad planning) 892 * could redirect this thread's flow of control before it reached the 893 * return probe fasttrap. In this case we need to kill the process 894 * since it's in a unrecoverable state. 895 */ 896 if (curthread->t_dtrace_step) { 897 ASSERT(curthread->t_dtrace_on); 898 fasttrap_sigtrap(p, curthread, pc); 899 return (0); 900 } 901 902 /* 903 * Clear all user tracing flags. 904 */ 905 curthread->t_dtrace_ft = 0; 906 curthread->t_dtrace_pc = 0; 907 curthread->t_dtrace_npc = 0; 908 curthread->t_dtrace_scrpc = 0; 909 curthread->t_dtrace_astpc = 0; 910 #ifdef __amd64 911 curthread->t_dtrace_regv = 0; 912 #endif 913 914 /* 915 * Treat a child created by a call to vfork(2) as if it were its 916 * parent. We know that there's only one thread of control in such a 917 * process: this one. 918 */ 919 while (p->p_flag & SVFORK) { 920 p = p->p_parent; 921 } 922 923 pid = p->p_pid; 924 pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock; 925 mutex_enter(pid_mtx); 926 bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)]; 927 928 /* 929 * Lookup the tracepoint that the process just hit. 930 */ 931 for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) { 932 if (pid == tp->ftt_pid && pc == tp->ftt_pc && 933 tp->ftt_proc->ftpc_acount != 0) 934 break; 935 } 936 937 /* 938 * If we couldn't find a matching tracepoint, either a tracepoint has 939 * been inserted without using the pid<pid> ioctl interface (see 940 * fasttrap_ioctl), or somehow we have mislaid this tracepoint. 941 */ 942 if (tp == NULL) { 943 mutex_exit(pid_mtx); 944 return (-1); 945 } 946 947 /* 948 * Set the program counter to the address of the traced instruction 949 * so that it looks right in ustack() output. 950 */ 951 rp->r_pc = pc; 952 953 if (tp->ftt_ids != NULL) { 954 fasttrap_id_t *id; 955 956 #ifdef __amd64 957 if (p->p_model == DATAMODEL_LP64) { 958 for (id = tp->ftt_ids; id != NULL; id = id->fti_next) { 959 fasttrap_probe_t *probe = id->fti_probe; 960 961 if (id->fti_ptype == DTFTP_ENTRY) { 962 /* 963 * We note that this was an entry 964 * probe to help ustack() find the 965 * first caller. 966 */ 967 cookie = dtrace_interrupt_disable(); 968 DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY); 969 dtrace_probe(probe->ftp_id, rp->r_rdi, 970 rp->r_rsi, rp->r_rdx, rp->r_rcx, 971 rp->r_r8); 972 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY); 973 dtrace_interrupt_enable(cookie); 974 } else if (id->fti_ptype == DTFTP_IS_ENABLED) { 975 /* 976 * Note that in this case, we don't 977 * call dtrace_probe() since it's only 978 * an artificial probe meant to change 979 * the flow of control so that it 980 * encounters the true probe. 981 */ 982 is_enabled = 1; 983 } else if (probe->ftp_argmap == NULL) { 984 dtrace_probe(probe->ftp_id, rp->r_rdi, 985 rp->r_rsi, rp->r_rdx, rp->r_rcx, 986 rp->r_r8); 987 } else { 988 uintptr_t t[5]; 989 990 fasttrap_usdt_args64(probe, rp, 991 sizeof (t) / sizeof (t[0]), t); 992 993 dtrace_probe(probe->ftp_id, t[0], t[1], 994 t[2], t[3], t[4]); 995 } 996 } 997 } else { 998 #endif 999 uintptr_t s0, s1, s2, s3, s4, s5; 1000 uint32_t *stack = (uint32_t *)rp->r_sp; 1001 1002 /* 1003 * In 32-bit mode, all arguments are passed on the 1004 * stack. If this is a function entry probe, we need 1005 * to skip the first entry on the stack as it 1006 * represents the return address rather than a 1007 * parameter to the function. 1008 */ 1009 s0 = fasttrap_fuword32_noerr(&stack[0]); 1010 s1 = fasttrap_fuword32_noerr(&stack[1]); 1011 s2 = fasttrap_fuword32_noerr(&stack[2]); 1012 s3 = fasttrap_fuword32_noerr(&stack[3]); 1013 s4 = fasttrap_fuword32_noerr(&stack[4]); 1014 s5 = fasttrap_fuword32_noerr(&stack[5]); 1015 1016 for (id = tp->ftt_ids; id != NULL; id = id->fti_next) { 1017 fasttrap_probe_t *probe = id->fti_probe; 1018 1019 if (id->fti_ptype == DTFTP_ENTRY) { 1020 /* 1021 * We note that this was an entry 1022 * probe to help ustack() find the 1023 * first caller. 1024 */ 1025 cookie = dtrace_interrupt_disable(); 1026 DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY); 1027 dtrace_probe(probe->ftp_id, s1, s2, 1028 s3, s4, s5); 1029 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY); 1030 dtrace_interrupt_enable(cookie); 1031 } else if (id->fti_ptype == DTFTP_IS_ENABLED) { 1032 /* 1033 * Note that in this case, we don't 1034 * call dtrace_probe() since it's only 1035 * an artificial probe meant to change 1036 * the flow of control so that it 1037 * encounters the true probe. 1038 */ 1039 is_enabled = 1; 1040 } else if (probe->ftp_argmap == NULL) { 1041 dtrace_probe(probe->ftp_id, s0, s1, 1042 s2, s3, s4); 1043 } else { 1044 uint32_t t[5]; 1045 1046 fasttrap_usdt_args32(probe, rp, 1047 sizeof (t) / sizeof (t[0]), t); 1048 1049 dtrace_probe(probe->ftp_id, t[0], t[1], 1050 t[2], t[3], t[4]); 1051 } 1052 } 1053 #ifdef __amd64 1054 } 1055 #endif 1056 } 1057 1058 /* 1059 * We're about to do a bunch of work so we cache a local copy of 1060 * the tracepoint to emulate the instruction, and then find the 1061 * tracepoint again later if we need to light up any return probes. 1062 */ 1063 tp_local = *tp; 1064 mutex_exit(pid_mtx); 1065 tp = &tp_local; 1066 1067 /* 1068 * Set the program counter to appear as though the traced instruction 1069 * had completely executed. This ensures that fasttrap_getreg() will 1070 * report the expected value for REG_RIP. 1071 */ 1072 rp->r_pc = pc + tp->ftt_size; 1073 1074 /* 1075 * If there's an is-enabled probe connected to this tracepoint it 1076 * means that there was a 'xorl %eax, %eax' or 'xorq %rax, %rax' 1077 * instruction that was placed there by DTrace when the binary was 1078 * linked. As this probe is, in fact, enabled, we need to stuff 1 1079 * into %eax or %rax. Accordingly, we can bypass all the instruction 1080 * emulation logic since we know the inevitable result. It's possible 1081 * that a user could construct a scenario where the 'is-enabled' 1082 * probe was on some other instruction, but that would be a rather 1083 * exotic way to shoot oneself in the foot. 1084 */ 1085 if (is_enabled) { 1086 rp->r_r0 = 1; 1087 new_pc = rp->r_pc; 1088 goto done; 1089 } 1090 1091 /* 1092 * We emulate certain types of instructions to ensure correctness 1093 * (in the case of position dependent instructions) or optimize 1094 * common cases. The rest we have the thread execute back in user- 1095 * land. 1096 */ 1097 switch (tp->ftt_type) { 1098 case FASTTRAP_T_RET: 1099 case FASTTRAP_T_RET16: 1100 { 1101 uintptr_t dst; 1102 uintptr_t addr; 1103 int ret; 1104 1105 /* 1106 * We have to emulate _every_ facet of the behavior of a ret 1107 * instruction including what happens if the load from %esp 1108 * fails; in that case, we send a SIGSEGV. 1109 */ 1110 #ifdef __amd64 1111 if (p->p_model == DATAMODEL_NATIVE) { 1112 #endif 1113 ret = fasttrap_fulword((void *)rp->r_sp, &dst); 1114 addr = rp->r_sp + sizeof (uintptr_t); 1115 #ifdef __amd64 1116 } else { 1117 uint32_t dst32; 1118 ret = fasttrap_fuword32((void *)rp->r_sp, &dst32); 1119 dst = dst32; 1120 addr = rp->r_sp + sizeof (uint32_t); 1121 } 1122 #endif 1123 1124 if (ret == -1) { 1125 fasttrap_sigsegv(p, curthread, rp->r_sp); 1126 new_pc = pc; 1127 break; 1128 } 1129 1130 if (tp->ftt_type == FASTTRAP_T_RET16) 1131 addr += tp->ftt_dest; 1132 1133 rp->r_sp = addr; 1134 new_pc = dst; 1135 break; 1136 } 1137 1138 case FASTTRAP_T_JCC: 1139 { 1140 uint_t taken; 1141 1142 switch (tp->ftt_code) { 1143 case FASTTRAP_JO: 1144 taken = (rp->r_ps & FASTTRAP_EFLAGS_OF) != 0; 1145 break; 1146 case FASTTRAP_JNO: 1147 taken = (rp->r_ps & FASTTRAP_EFLAGS_OF) == 0; 1148 break; 1149 case FASTTRAP_JB: 1150 taken = (rp->r_ps & FASTTRAP_EFLAGS_CF) != 0; 1151 break; 1152 case FASTTRAP_JAE: 1153 taken = (rp->r_ps & FASTTRAP_EFLAGS_CF) == 0; 1154 break; 1155 case FASTTRAP_JE: 1156 taken = (rp->r_ps & FASTTRAP_EFLAGS_ZF) != 0; 1157 break; 1158 case FASTTRAP_JNE: 1159 taken = (rp->r_ps & FASTTRAP_EFLAGS_ZF) == 0; 1160 break; 1161 case FASTTRAP_JBE: 1162 taken = (rp->r_ps & FASTTRAP_EFLAGS_CF) != 0 || 1163 (rp->r_ps & FASTTRAP_EFLAGS_ZF) != 0; 1164 break; 1165 case FASTTRAP_JA: 1166 taken = (rp->r_ps & FASTTRAP_EFLAGS_CF) == 0 && 1167 (rp->r_ps & FASTTRAP_EFLAGS_ZF) == 0; 1168 break; 1169 case FASTTRAP_JS: 1170 taken = (rp->r_ps & FASTTRAP_EFLAGS_SF) != 0; 1171 break; 1172 case FASTTRAP_JNS: 1173 taken = (rp->r_ps & FASTTRAP_EFLAGS_SF) == 0; 1174 break; 1175 case FASTTRAP_JP: 1176 taken = (rp->r_ps & FASTTRAP_EFLAGS_PF) != 0; 1177 break; 1178 case FASTTRAP_JNP: 1179 taken = (rp->r_ps & FASTTRAP_EFLAGS_PF) == 0; 1180 break; 1181 case FASTTRAP_JL: 1182 taken = ((rp->r_ps & FASTTRAP_EFLAGS_SF) == 0) != 1183 ((rp->r_ps & FASTTRAP_EFLAGS_OF) == 0); 1184 break; 1185 case FASTTRAP_JGE: 1186 taken = ((rp->r_ps & FASTTRAP_EFLAGS_SF) == 0) == 1187 ((rp->r_ps & FASTTRAP_EFLAGS_OF) == 0); 1188 break; 1189 case FASTTRAP_JLE: 1190 taken = (rp->r_ps & FASTTRAP_EFLAGS_ZF) != 0 || 1191 ((rp->r_ps & FASTTRAP_EFLAGS_SF) == 0) != 1192 ((rp->r_ps & FASTTRAP_EFLAGS_OF) == 0); 1193 break; 1194 case FASTTRAP_JG: 1195 taken = (rp->r_ps & FASTTRAP_EFLAGS_ZF) == 0 && 1196 ((rp->r_ps & FASTTRAP_EFLAGS_SF) == 0) == 1197 ((rp->r_ps & FASTTRAP_EFLAGS_OF) == 0); 1198 break; 1199 1200 } 1201 1202 if (taken) 1203 new_pc = tp->ftt_dest; 1204 else 1205 new_pc = pc + tp->ftt_size; 1206 break; 1207 } 1208 1209 case FASTTRAP_T_LOOP: 1210 { 1211 uint_t taken; 1212 #ifdef __amd64 1213 greg_t cx = rp->r_rcx--; 1214 #else 1215 greg_t cx = rp->r_ecx--; 1216 #endif 1217 1218 switch (tp->ftt_code) { 1219 case FASTTRAP_LOOPNZ: 1220 taken = (rp->r_ps & FASTTRAP_EFLAGS_ZF) == 0 && 1221 cx != 0; 1222 break; 1223 case FASTTRAP_LOOPZ: 1224 taken = (rp->r_ps & FASTTRAP_EFLAGS_ZF) != 0 && 1225 cx != 0; 1226 break; 1227 case FASTTRAP_LOOP: 1228 taken = (cx != 0); 1229 break; 1230 } 1231 1232 if (taken) 1233 new_pc = tp->ftt_dest; 1234 else 1235 new_pc = pc + tp->ftt_size; 1236 break; 1237 } 1238 1239 case FASTTRAP_T_JCXZ: 1240 { 1241 #ifdef __amd64 1242 greg_t cx = rp->r_rcx; 1243 #else 1244 greg_t cx = rp->r_ecx; 1245 #endif 1246 1247 if (cx == 0) 1248 new_pc = tp->ftt_dest; 1249 else 1250 new_pc = pc + tp->ftt_size; 1251 break; 1252 } 1253 1254 case FASTTRAP_T_PUSHL_EBP: 1255 { 1256 int ret; 1257 uintptr_t addr; 1258 #ifdef __amd64 1259 if (p->p_model == DATAMODEL_NATIVE) { 1260 #endif 1261 addr = rp->r_sp - sizeof (uintptr_t); 1262 ret = fasttrap_sulword((void *)addr, rp->r_fp); 1263 #ifdef __amd64 1264 } else { 1265 addr = rp->r_sp - sizeof (uint32_t); 1266 ret = fasttrap_suword32((void *)addr, 1267 (uint32_t)rp->r_fp); 1268 } 1269 #endif 1270 1271 if (ret == -1) { 1272 fasttrap_sigsegv(p, curthread, addr); 1273 new_pc = pc; 1274 break; 1275 } 1276 1277 rp->r_sp = addr; 1278 new_pc = pc + tp->ftt_size; 1279 break; 1280 } 1281 1282 case FASTTRAP_T_NOP: 1283 new_pc = pc + tp->ftt_size; 1284 break; 1285 1286 case FASTTRAP_T_JMP: 1287 case FASTTRAP_T_CALL: 1288 if (tp->ftt_code == 0) { 1289 new_pc = tp->ftt_dest; 1290 } else { 1291 uintptr_t value, addr = tp->ftt_dest; 1292 1293 if (tp->ftt_base != FASTTRAP_NOREG) 1294 addr += fasttrap_getreg(rp, tp->ftt_base); 1295 if (tp->ftt_index != FASTTRAP_NOREG) 1296 addr += fasttrap_getreg(rp, tp->ftt_index) << 1297 tp->ftt_scale; 1298 1299 if (tp->ftt_code == 1) { 1300 /* 1301 * If there's a segment prefix for this 1302 * instruction, we'll need to check permissions 1303 * and bounds on the given selector, and adjust 1304 * the address accordingly. 1305 */ 1306 if (tp->ftt_segment != FASTTRAP_SEG_NONE && 1307 fasttrap_do_seg(tp, rp, &addr) != 0) { 1308 fasttrap_sigsegv(p, curthread, addr); 1309 new_pc = pc; 1310 break; 1311 } 1312 1313 #ifdef __amd64 1314 if (p->p_model == DATAMODEL_NATIVE) { 1315 #endif 1316 if (fasttrap_fulword((void *)addr, 1317 &value) == -1) { 1318 fasttrap_sigsegv(p, curthread, 1319 addr); 1320 new_pc = pc; 1321 break; 1322 } 1323 new_pc = value; 1324 #ifdef __amd64 1325 } else { 1326 uint32_t value32; 1327 addr = (uintptr_t)(uint32_t)addr; 1328 if (fasttrap_fuword32((void *)addr, 1329 &value32) == -1) { 1330 fasttrap_sigsegv(p, curthread, 1331 addr); 1332 new_pc = pc; 1333 break; 1334 } 1335 new_pc = value32; 1336 } 1337 #endif 1338 } else { 1339 new_pc = addr; 1340 } 1341 } 1342 1343 /* 1344 * If this is a call instruction, we need to push the return 1345 * address onto the stack. If this fails, we send the process 1346 * a SIGSEGV and reset the pc to emulate what would happen if 1347 * this instruction weren't traced. 1348 */ 1349 if (tp->ftt_type == FASTTRAP_T_CALL) { 1350 int ret; 1351 uintptr_t addr; 1352 #ifdef __amd64 1353 if (p->p_model == DATAMODEL_NATIVE) { 1354 addr = rp->r_sp - sizeof (uintptr_t); 1355 ret = fasttrap_sulword((void *)addr, 1356 pc + tp->ftt_size); 1357 } else { 1358 #endif 1359 addr = rp->r_sp - sizeof (uint32_t); 1360 ret = fasttrap_suword32((void *)addr, 1361 (uint32_t)(pc + tp->ftt_size)); 1362 #ifdef __amd64 1363 } 1364 #endif 1365 1366 if (ret == -1) { 1367 fasttrap_sigsegv(p, curthread, addr); 1368 new_pc = pc; 1369 break; 1370 } 1371 1372 rp->r_sp = addr; 1373 } 1374 1375 break; 1376 1377 case FASTTRAP_T_COMMON: 1378 { 1379 uintptr_t addr; 1380 uint8_t scratch[2 * FASTTRAP_MAX_INSTR_SIZE + 5 + 2]; 1381 uint_t i = 0; 1382 klwp_t *lwp = ttolwp(curthread); 1383 1384 /* 1385 * Compute the address of the ulwp_t and step over the 1386 * ul_self pointer. The method used to store the user-land 1387 * thread pointer is very different on 32- and 64-bit 1388 * kernels. 1389 */ 1390 #if defined(__amd64) 1391 if (p->p_model == DATAMODEL_LP64) { 1392 addr = lwp->lwp_pcb.pcb_fsbase; 1393 addr += sizeof (void *); 1394 } else { 1395 addr = lwp->lwp_pcb.pcb_gsbase; 1396 addr += sizeof (caddr32_t); 1397 } 1398 #elif defined(__i386) 1399 addr = USEGD_GETBASE(&lwp->lwp_pcb.pcb_gsdesc); 1400 addr += sizeof (void *); 1401 #endif 1402 1403 /* 1404 * Generic Instruction Tracing 1405 * --------------------------- 1406 * 1407 * This is the layout of the scratch space in the user-land 1408 * thread structure for our generated instructions. 1409 * 1410 * 32-bit mode bytes 1411 * ------------------------ ----- 1412 * a: <original instruction> <= 15 1413 * jmp <pc + tp->ftt_size> 5 1414 * b: <original instrction> <= 15 1415 * int T_DTRACE_RET 2 1416 * ----- 1417 * <= 37 1418 * 1419 * 64-bit mode bytes 1420 * ------------------------ ----- 1421 * a: <original instruction> <= 15 1422 * jmp 0(%rip) 6 1423 * <pc + tp->ftt_size> 8 1424 * b: <original instruction> <= 15 1425 * int T_DTRACE_RET 2 1426 * ----- 1427 * <= 46 1428 * 1429 * The %pc is set to a, and curthread->t_dtrace_astpc is set 1430 * to b. If we encounter a signal on the way out of the 1431 * kernel, trap() will set %pc to curthread->t_dtrace_astpc 1432 * so that we execute the original instruction and re-enter 1433 * the kernel rather than redirecting to the next instruction. 1434 * 1435 * If there are return probes (so we know that we're going to 1436 * need to reenter the kernel after executing the original 1437 * instruction), the scratch space will just contain the 1438 * original instruction followed by an interrupt -- the same 1439 * data as at b. 1440 * 1441 * %rip-relative Addressing 1442 * ------------------------ 1443 * 1444 * There's a further complication in 64-bit mode due to %rip- 1445 * relative addressing. While this is clearly a beneficial 1446 * architectural decision for position independent code, it's 1447 * hard not to see it as a personal attack against the pid 1448 * provider since before there was a relatively small set of 1449 * instructions to emulate; with %rip-relative addressing, 1450 * almost every instruction can potentially depend on the 1451 * address at which it's executed. Rather than emulating 1452 * the broad spectrum of instructions that can now be 1453 * position dependent, we emulate jumps and others as in 1454 * 32-bit mode, and take a different tack for instructions 1455 * using %rip-relative addressing. 1456 * 1457 * For every instruction that uses the ModRM byte, the 1458 * in-kernel disassembler reports its location. We use the 1459 * ModRM byte to identify that an instruction uses 1460 * %rip-relative addressing and to see what other registers 1461 * the instruction uses. To emulate those instructions, 1462 * we modify the instruction to be %rax-relative rather than 1463 * %rip-relative (or %rcx-relative if the instruction uses 1464 * %rax; or %r8- or %r9-relative if the REX.B is present so 1465 * we don't have to rewrite the REX prefix). We then load 1466 * the value that %rip would have been into the scratch 1467 * register and generate an instruction to reset the scratch 1468 * register back to its original value. The instruction 1469 * sequence looks like this: 1470 * 1471 * 64-mode %rip-relative bytes 1472 * ------------------------ ----- 1473 * a: <modified instruction> <= 15 1474 * movq $<value>, %<scratch> 6 1475 * jmp 0(%rip) 6 1476 * <pc + tp->ftt_size> 8 1477 * b: <modified instruction> <= 15 1478 * int T_DTRACE_RET 2 1479 * ----- 1480 * 52 1481 * 1482 * We set curthread->t_dtrace_regv so that upon receiving 1483 * a signal we can reset the value of the scratch register. 1484 */ 1485 1486 ASSERT(tp->ftt_size < FASTTRAP_MAX_INSTR_SIZE); 1487 1488 curthread->t_dtrace_scrpc = addr; 1489 bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size); 1490 i += tp->ftt_size; 1491 1492 #ifdef __amd64 1493 if (tp->ftt_ripmode != 0) { 1494 greg_t *reg; 1495 1496 ASSERT(p->p_model == DATAMODEL_LP64); 1497 ASSERT(tp->ftt_ripmode & 1498 (FASTTRAP_RIP_1 | FASTTRAP_RIP_2)); 1499 1500 /* 1501 * If this was a %rip-relative instruction, we change 1502 * it to be either a %rax- or %rcx-relative 1503 * instruction (depending on whether those registers 1504 * are used as another operand; or %r8- or %r9- 1505 * relative depending on the value of REX.B). We then 1506 * set that register and generate a movq instruction 1507 * to reset the value. 1508 */ 1509 if (tp->ftt_ripmode & FASTTRAP_RIP_X) 1510 scratch[i++] = FASTTRAP_REX(1, 0, 0, 1); 1511 else 1512 scratch[i++] = FASTTRAP_REX(1, 0, 0, 0); 1513 1514 if (tp->ftt_ripmode & FASTTRAP_RIP_1) 1515 scratch[i++] = FASTTRAP_MOV_EAX; 1516 else 1517 scratch[i++] = FASTTRAP_MOV_ECX; 1518 1519 switch (tp->ftt_ripmode) { 1520 case FASTTRAP_RIP_1: 1521 reg = &rp->r_rax; 1522 curthread->t_dtrace_reg = REG_RAX; 1523 break; 1524 case FASTTRAP_RIP_2: 1525 reg = &rp->r_rcx; 1526 curthread->t_dtrace_reg = REG_RCX; 1527 break; 1528 case FASTTRAP_RIP_1 | FASTTRAP_RIP_X: 1529 reg = &rp->r_r8; 1530 curthread->t_dtrace_reg = REG_R8; 1531 break; 1532 case FASTTRAP_RIP_2 | FASTTRAP_RIP_X: 1533 reg = &rp->r_r9; 1534 curthread->t_dtrace_reg = REG_R9; 1535 break; 1536 } 1537 1538 /* LINTED - alignment */ 1539 *(uint64_t *)&scratch[i] = *reg; 1540 curthread->t_dtrace_regv = *reg; 1541 *reg = pc + tp->ftt_size; 1542 i += sizeof (uint64_t); 1543 } 1544 #endif 1545 1546 /* 1547 * Generate the branch instruction to what would have 1548 * normally been the subsequent instruction. In 32-bit mode, 1549 * this is just a relative branch; in 64-bit mode this is a 1550 * %rip-relative branch that loads the 64-bit pc value 1551 * immediately after the jmp instruction. 1552 */ 1553 #ifdef __amd64 1554 if (p->p_model == DATAMODEL_LP64) { 1555 scratch[i++] = FASTTRAP_GROUP5_OP; 1556 scratch[i++] = FASTTRAP_MODRM(0, 4, 5); 1557 /* LINTED - alignment */ 1558 *(uint32_t *)&scratch[i] = 0; 1559 i += sizeof (uint32_t); 1560 /* LINTED - alignment */ 1561 *(uint64_t *)&scratch[i] = pc + tp->ftt_size; 1562 i += sizeof (uint64_t); 1563 } else { 1564 #endif 1565 /* 1566 * Set up the jmp to the next instruction; note that 1567 * the size of the traced instruction cancels out. 1568 */ 1569 scratch[i++] = FASTTRAP_JMP32; 1570 /* LINTED - alignment */ 1571 *(uint32_t *)&scratch[i] = pc - addr - 5; 1572 i += sizeof (uint32_t); 1573 #ifdef __amd64 1574 } 1575 #endif 1576 1577 curthread->t_dtrace_astpc = addr + i; 1578 bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size); 1579 i += tp->ftt_size; 1580 scratch[i++] = FASTTRAP_INT; 1581 scratch[i++] = T_DTRACE_RET; 1582 1583 if (fasttrap_copyout(scratch, (char *)addr, i)) { 1584 fasttrap_sigtrap(p, curthread, pc); 1585 new_pc = pc; 1586 break; 1587 } 1588 1589 if (tp->ftt_retids != NULL) { 1590 curthread->t_dtrace_step = 1; 1591 curthread->t_dtrace_ret = 1; 1592 new_pc = curthread->t_dtrace_astpc; 1593 } else { 1594 new_pc = curthread->t_dtrace_scrpc; 1595 } 1596 1597 curthread->t_dtrace_pc = pc; 1598 curthread->t_dtrace_npc = pc + tp->ftt_size; 1599 curthread->t_dtrace_on = 1; 1600 break; 1601 } 1602 1603 default: 1604 panic("fasttrap: mishandled an instruction"); 1605 } 1606 1607 done: 1608 /* 1609 * If there were no return probes when we first found the tracepoint, 1610 * we should feel no obligation to honor any return probes that were 1611 * subsequently enabled -- they'll just have to wait until the next 1612 * time around. 1613 */ 1614 if (tp->ftt_retids != NULL) { 1615 /* 1616 * We need to wait until the results of the instruction are 1617 * apparent before invoking any return probes. If this 1618 * instruction was emulated we can just call 1619 * fasttrap_return_common(); if it needs to be executed, we 1620 * need to wait until the user thread returns to the kernel. 1621 */ 1622 if (tp->ftt_type != FASTTRAP_T_COMMON) { 1623 /* 1624 * Set the program counter to the address of the traced 1625 * instruction so that it looks right in ustack() 1626 * output. We had previously set it to the end of the 1627 * instruction to simplify %rip-relative addressing. 1628 */ 1629 rp->r_pc = pc; 1630 1631 fasttrap_return_common(rp, pc, pid, new_pc); 1632 } else { 1633 ASSERT(curthread->t_dtrace_ret != 0); 1634 ASSERT(curthread->t_dtrace_pc == pc); 1635 ASSERT(curthread->t_dtrace_scrpc != 0); 1636 ASSERT(new_pc == curthread->t_dtrace_astpc); 1637 } 1638 } 1639 1640 rp->r_pc = new_pc; 1641 1642 return (0); 1643 } 1644 1645 int 1646 fasttrap_return_probe(struct regs *rp) 1647 { 1648 proc_t *p = curproc; 1649 uintptr_t pc = curthread->t_dtrace_pc; 1650 uintptr_t npc = curthread->t_dtrace_npc; 1651 1652 curthread->t_dtrace_pc = 0; 1653 curthread->t_dtrace_npc = 0; 1654 curthread->t_dtrace_scrpc = 0; 1655 curthread->t_dtrace_astpc = 0; 1656 1657 /* 1658 * Treat a child created by a call to vfork(2) as if it were its 1659 * parent. We know that there's only one thread of control in such a 1660 * process: this one. 1661 */ 1662 while (p->p_flag & SVFORK) { 1663 p = p->p_parent; 1664 } 1665 1666 /* 1667 * We set rp->r_pc to the address of the traced instruction so 1668 * that it appears to dtrace_probe() that we're on the original 1669 * instruction, and so that the user can't easily detect our 1670 * complex web of lies. dtrace_return_probe() (our caller) 1671 * will correctly set %pc after we return. 1672 */ 1673 rp->r_pc = pc; 1674 1675 fasttrap_return_common(rp, pc, p->p_pid, npc); 1676 1677 return (0); 1678 } 1679 1680 /*ARGSUSED*/ 1681 uint64_t 1682 fasttrap_pid_getarg(void *arg, dtrace_id_t id, void *parg, int argno, 1683 int aframes) 1684 { 1685 return (fasttrap_anarg(ttolwp(curthread)->lwp_regs, 1, argno)); 1686 } 1687 1688 /*ARGSUSED*/ 1689 uint64_t 1690 fasttrap_usdt_getarg(void *arg, dtrace_id_t id, void *parg, int argno, 1691 int aframes) 1692 { 1693 return (fasttrap_anarg(ttolwp(curthread)->lwp_regs, 0, argno)); 1694 } 1695 1696 static ulong_t 1697 fasttrap_getreg(struct regs *rp, uint_t reg) 1698 { 1699 #ifdef __amd64 1700 switch (reg) { 1701 case REG_R15: return (rp->r_r15); 1702 case REG_R14: return (rp->r_r14); 1703 case REG_R13: return (rp->r_r13); 1704 case REG_R12: return (rp->r_r12); 1705 case REG_R11: return (rp->r_r11); 1706 case REG_R10: return (rp->r_r10); 1707 case REG_R9: return (rp->r_r9); 1708 case REG_R8: return (rp->r_r8); 1709 case REG_RDI: return (rp->r_rdi); 1710 case REG_RSI: return (rp->r_rsi); 1711 case REG_RBP: return (rp->r_rbp); 1712 case REG_RBX: return (rp->r_rbx); 1713 case REG_RDX: return (rp->r_rdx); 1714 case REG_RCX: return (rp->r_rcx); 1715 case REG_RAX: return (rp->r_rax); 1716 case REG_TRAPNO: return (rp->r_trapno); 1717 case REG_ERR: return (rp->r_err); 1718 case REG_RIP: return (rp->r_rip); 1719 case REG_CS: return (rp->r_cs); 1720 case REG_RFL: return (rp->r_rfl); 1721 case REG_RSP: return (rp->r_rsp); 1722 case REG_SS: return (rp->r_ss); 1723 case REG_FS: return (rp->r_fs); 1724 case REG_GS: return (rp->r_gs); 1725 case REG_DS: return (rp->r_ds); 1726 case REG_ES: return (rp->r_es); 1727 case REG_FSBASE: return (rdmsr(MSR_AMD_FSBASE)); 1728 case REG_GSBASE: return (rdmsr(MSR_AMD_GSBASE)); 1729 } 1730 1731 panic("dtrace: illegal register constant"); 1732 /*NOTREACHED*/ 1733 #else 1734 if (reg >= _NGREG) 1735 panic("dtrace: illegal register constant"); 1736 1737 return (((greg_t *)&rp->r_gs)[reg]); 1738 #endif 1739 } 1740