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 * Portions Copyright 2010 The FreeBSD Foundation 22 * 23 * $FreeBSD$ 24 */ 25 26 /* 27 * Copyright 2008 Sun Microsystems, Inc. All rights reserved. 28 * Use is subject to license terms. 29 */ 30 31 #if defined(sun) 32 #pragma ident "%Z%%M% %I% %E% SMI" 33 #endif 34 35 #include <sys/fasttrap_isa.h> 36 #include <sys/fasttrap_impl.h> 37 #include <sys/dtrace.h> 38 #include <sys/dtrace_impl.h> 39 #include <sys/cmn_err.h> 40 #if defined(sun) 41 #include <sys/regset.h> 42 #include <sys/privregs.h> 43 #include <sys/segments.h> 44 #include <sys/x86_archext.h> 45 #else 46 #include <cddl/dev/dtrace/dtrace_cddl.h> 47 #include <sys/types.h> 48 #include <sys/proc.h> 49 #include <sys/dtrace_bsd.h> 50 #include <cddl/dev/dtrace/x86/regset.h> 51 #include <machine/segments.h> 52 #include <machine/reg.h> 53 #include <machine/pcb.h> 54 #endif 55 #include <sys/sysmacros.h> 56 #if defined(sun) 57 #include <sys/trap.h> 58 #include <sys/archsystm.h> 59 #else 60 #include <sys/ptrace.h> 61 62 static int 63 proc_ops(int op, proc_t *p, void *kaddr, off_t uaddr, size_t len) 64 { 65 struct iovec iov; 66 struct uio uio; 67 68 iov.iov_base = kaddr; 69 iov.iov_len = len; 70 uio.uio_offset = uaddr; 71 uio.uio_iov = &iov; 72 uio.uio_resid = len; 73 uio.uio_iovcnt = 1; 74 uio.uio_segflg = UIO_SYSSPACE; 75 uio.uio_td = curthread; 76 uio.uio_rw = op; 77 PHOLD(p); 78 if (proc_rwmem(p, &uio) != 0) { 79 PRELE(p); 80 return (-1); 81 } 82 PRELE(p); 83 84 return (0); 85 } 86 87 static int 88 uread(proc_t *p, void *kaddr, size_t len, uintptr_t uaddr) 89 { 90 91 return (proc_ops(UIO_READ, p, kaddr, uaddr, len)); 92 } 93 94 static int 95 uwrite(proc_t *p, void *kaddr, size_t len, uintptr_t uaddr) 96 { 97 98 return (proc_ops(UIO_WRITE, p, kaddr, uaddr, len)); 99 } 100 #endif /* sun */ 101 #ifdef __i386__ 102 #define r_rax r_eax 103 #define r_rbx r_ebx 104 #define r_rip r_eip 105 #define r_rflags r_eflags 106 #define r_rsp r_esp 107 #define r_rbp r_ebp 108 #endif 109 110 /* 111 * Lossless User-Land Tracing on x86 112 * --------------------------------- 113 * 114 * The execution of most instructions is not dependent on the address; for 115 * these instructions it is sufficient to copy them into the user process's 116 * address space and execute them. To effectively single-step an instruction 117 * in user-land, we copy out the following sequence of instructions to scratch 118 * space in the user thread's ulwp_t structure. 119 * 120 * We then set the program counter (%eip or %rip) to point to this scratch 121 * space. Once execution resumes, the original instruction is executed and 122 * then control flow is redirected to what was originally the subsequent 123 * instruction. If the kernel attemps to deliver a signal while single- 124 * stepping, the signal is deferred and the program counter is moved into the 125 * second sequence of instructions. The second sequence ends in a trap into 126 * the kernel where the deferred signal is then properly handled and delivered. 127 * 128 * For instructions whose execute is position dependent, we perform simple 129 * emulation. These instructions are limited to control transfer 130 * instructions in 32-bit mode, but in 64-bit mode there's the added wrinkle 131 * of %rip-relative addressing that means that almost any instruction can be 132 * position dependent. For all the details on how we emulate generic 133 * instructions included %rip-relative instructions, see the code in 134 * fasttrap_pid_probe() below where we handle instructions of type 135 * FASTTRAP_T_COMMON (under the header: Generic Instruction Tracing). 136 */ 137 138 #define FASTTRAP_MODRM_MOD(modrm) (((modrm) >> 6) & 0x3) 139 #define FASTTRAP_MODRM_REG(modrm) (((modrm) >> 3) & 0x7) 140 #define FASTTRAP_MODRM_RM(modrm) ((modrm) & 0x7) 141 #define FASTTRAP_MODRM(mod, reg, rm) (((mod) << 6) | ((reg) << 3) | (rm)) 142 143 #define FASTTRAP_SIB_SCALE(sib) (((sib) >> 6) & 0x3) 144 #define FASTTRAP_SIB_INDEX(sib) (((sib) >> 3) & 0x7) 145 #define FASTTRAP_SIB_BASE(sib) ((sib) & 0x7) 146 147 #define FASTTRAP_REX_W(rex) (((rex) >> 3) & 1) 148 #define FASTTRAP_REX_R(rex) (((rex) >> 2) & 1) 149 #define FASTTRAP_REX_X(rex) (((rex) >> 1) & 1) 150 #define FASTTRAP_REX_B(rex) ((rex) & 1) 151 #define FASTTRAP_REX(w, r, x, b) \ 152 (0x40 | ((w) << 3) | ((r) << 2) | ((x) << 1) | (b)) 153 154 /* 155 * Single-byte op-codes. 156 */ 157 #define FASTTRAP_PUSHL_EBP 0x55 158 159 #define FASTTRAP_JO 0x70 160 #define FASTTRAP_JNO 0x71 161 #define FASTTRAP_JB 0x72 162 #define FASTTRAP_JAE 0x73 163 #define FASTTRAP_JE 0x74 164 #define FASTTRAP_JNE 0x75 165 #define FASTTRAP_JBE 0x76 166 #define FASTTRAP_JA 0x77 167 #define FASTTRAP_JS 0x78 168 #define FASTTRAP_JNS 0x79 169 #define FASTTRAP_JP 0x7a 170 #define FASTTRAP_JNP 0x7b 171 #define FASTTRAP_JL 0x7c 172 #define FASTTRAP_JGE 0x7d 173 #define FASTTRAP_JLE 0x7e 174 #define FASTTRAP_JG 0x7f 175 176 #define FASTTRAP_NOP 0x90 177 178 #define FASTTRAP_MOV_EAX 0xb8 179 #define FASTTRAP_MOV_ECX 0xb9 180 181 #define FASTTRAP_RET16 0xc2 182 #define FASTTRAP_RET 0xc3 183 184 #define FASTTRAP_LOOPNZ 0xe0 185 #define FASTTRAP_LOOPZ 0xe1 186 #define FASTTRAP_LOOP 0xe2 187 #define FASTTRAP_JCXZ 0xe3 188 189 #define FASTTRAP_CALL 0xe8 190 #define FASTTRAP_JMP32 0xe9 191 #define FASTTRAP_JMP8 0xeb 192 193 #define FASTTRAP_INT3 0xcc 194 #define FASTTRAP_INT 0xcd 195 196 #define FASTTRAP_2_BYTE_OP 0x0f 197 #define FASTTRAP_GROUP5_OP 0xff 198 199 /* 200 * Two-byte op-codes (second byte only). 201 */ 202 #define FASTTRAP_0F_JO 0x80 203 #define FASTTRAP_0F_JNO 0x81 204 #define FASTTRAP_0F_JB 0x82 205 #define FASTTRAP_0F_JAE 0x83 206 #define FASTTRAP_0F_JE 0x84 207 #define FASTTRAP_0F_JNE 0x85 208 #define FASTTRAP_0F_JBE 0x86 209 #define FASTTRAP_0F_JA 0x87 210 #define FASTTRAP_0F_JS 0x88 211 #define FASTTRAP_0F_JNS 0x89 212 #define FASTTRAP_0F_JP 0x8a 213 #define FASTTRAP_0F_JNP 0x8b 214 #define FASTTRAP_0F_JL 0x8c 215 #define FASTTRAP_0F_JGE 0x8d 216 #define FASTTRAP_0F_JLE 0x8e 217 #define FASTTRAP_0F_JG 0x8f 218 219 #define FASTTRAP_EFLAGS_OF 0x800 220 #define FASTTRAP_EFLAGS_DF 0x400 221 #define FASTTRAP_EFLAGS_SF 0x080 222 #define FASTTRAP_EFLAGS_ZF 0x040 223 #define FASTTRAP_EFLAGS_AF 0x010 224 #define FASTTRAP_EFLAGS_PF 0x004 225 #define FASTTRAP_EFLAGS_CF 0x001 226 227 /* 228 * Instruction prefixes. 229 */ 230 #define FASTTRAP_PREFIX_OPERAND 0x66 231 #define FASTTRAP_PREFIX_ADDRESS 0x67 232 #define FASTTRAP_PREFIX_CS 0x2E 233 #define FASTTRAP_PREFIX_DS 0x3E 234 #define FASTTRAP_PREFIX_ES 0x26 235 #define FASTTRAP_PREFIX_FS 0x64 236 #define FASTTRAP_PREFIX_GS 0x65 237 #define FASTTRAP_PREFIX_SS 0x36 238 #define FASTTRAP_PREFIX_LOCK 0xF0 239 #define FASTTRAP_PREFIX_REP 0xF3 240 #define FASTTRAP_PREFIX_REPNE 0xF2 241 242 #define FASTTRAP_NOREG 0xff 243 244 /* 245 * Map between instruction register encodings and the kernel constants which 246 * correspond to indicies into struct regs. 247 */ 248 #ifdef __amd64 249 static const uint8_t regmap[16] = { 250 REG_RAX, REG_RCX, REG_RDX, REG_RBX, REG_RSP, REG_RBP, REG_RSI, REG_RDI, 251 REG_R8, REG_R9, REG_R10, REG_R11, REG_R12, REG_R13, REG_R14, REG_R15, 252 }; 253 #else 254 static const uint8_t regmap[8] = { 255 EAX, ECX, EDX, EBX, UESP, EBP, ESI, EDI 256 }; 257 #endif 258 259 static ulong_t fasttrap_getreg(struct reg *, uint_t); 260 261 static uint64_t 262 fasttrap_anarg(struct reg *rp, int function_entry, int argno) 263 { 264 uint64_t value = 0; 265 int shift = function_entry ? 1 : 0; 266 267 #ifdef __amd64 268 if (curproc->p_model == DATAMODEL_LP64) { 269 uintptr_t *stack; 270 271 /* 272 * In 64-bit mode, the first six arguments are stored in 273 * registers. 274 */ 275 if (argno < 6) 276 switch (argno) { 277 case 0: 278 return (rp->r_rdi); 279 case 1: 280 return (rp->r_rsi); 281 case 2: 282 return (rp->r_rdx); 283 case 3: 284 return (rp->r_rcx); 285 case 4: 286 return (rp->r_r8); 287 case 5: 288 return (rp->r_r9); 289 } 290 291 stack = (uintptr_t *)rp->r_rsp; 292 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 293 value = dtrace_fulword(&stack[argno - 6 + shift]); 294 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT | CPU_DTRACE_BADADDR); 295 } else { 296 #endif 297 #ifdef __i386 298 uint32_t *stack = (uint32_t *)rp->r_esp; 299 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); 300 value = dtrace_fuword32(&stack[argno + shift]); 301 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT | CPU_DTRACE_BADADDR); 302 #endif 303 #ifdef __amd64 304 } 305 #endif 306 307 return (value); 308 } 309 310 /*ARGSUSED*/ 311 int 312 fasttrap_tracepoint_init(proc_t *p, fasttrap_tracepoint_t *tp, uintptr_t pc, 313 fasttrap_probe_type_t type) 314 { 315 uint8_t instr[FASTTRAP_MAX_INSTR_SIZE + 10]; 316 size_t len = FASTTRAP_MAX_INSTR_SIZE; 317 size_t first = MIN(len, PAGESIZE - (pc & PAGEOFFSET)); 318 uint_t start = 0; 319 int rmindex, size; 320 uint8_t seg, rex = 0; 321 322 /* 323 * Read the instruction at the given address out of the process's 324 * address space. We don't have to worry about a debugger 325 * changing this instruction before we overwrite it with our trap 326 * instruction since P_PR_LOCK is set. Since instructions can span 327 * pages, we potentially read the instruction in two parts. If the 328 * second part fails, we just zero out that part of the instruction. 329 */ 330 if (uread(p, &instr[0], first, pc) != 0) 331 return (-1); 332 if (len > first && 333 uread(p, &instr[first], len - first, pc + first) != 0) { 334 bzero(&instr[first], len - first); 335 len = first; 336 } 337 338 /* 339 * If the disassembly fails, then we have a malformed instruction. 340 */ 341 if ((size = dtrace_instr_size_isa(instr, p->p_model, &rmindex)) <= 0) 342 return (-1); 343 344 /* 345 * Make sure the disassembler isn't completely broken. 346 */ 347 ASSERT(-1 <= rmindex && rmindex < size); 348 349 /* 350 * If the computed size is greater than the number of bytes read, 351 * then it was a malformed instruction possibly because it fell on a 352 * page boundary and the subsequent page was missing or because of 353 * some malicious user. 354 */ 355 if (size > len) 356 return (-1); 357 358 tp->ftt_size = (uint8_t)size; 359 tp->ftt_segment = FASTTRAP_SEG_NONE; 360 361 /* 362 * Find the start of the instruction's opcode by processing any 363 * legacy prefixes. 364 */ 365 for (;;) { 366 seg = 0; 367 switch (instr[start]) { 368 case FASTTRAP_PREFIX_SS: 369 seg++; 370 /*FALLTHRU*/ 371 case FASTTRAP_PREFIX_GS: 372 seg++; 373 /*FALLTHRU*/ 374 case FASTTRAP_PREFIX_FS: 375 seg++; 376 /*FALLTHRU*/ 377 case FASTTRAP_PREFIX_ES: 378 seg++; 379 /*FALLTHRU*/ 380 case FASTTRAP_PREFIX_DS: 381 seg++; 382 /*FALLTHRU*/ 383 case FASTTRAP_PREFIX_CS: 384 seg++; 385 /*FALLTHRU*/ 386 case FASTTRAP_PREFIX_OPERAND: 387 case FASTTRAP_PREFIX_ADDRESS: 388 case FASTTRAP_PREFIX_LOCK: 389 case FASTTRAP_PREFIX_REP: 390 case FASTTRAP_PREFIX_REPNE: 391 if (seg != 0) { 392 /* 393 * It's illegal for an instruction to specify 394 * two segment prefixes -- give up on this 395 * illegal instruction. 396 */ 397 if (tp->ftt_segment != FASTTRAP_SEG_NONE) 398 return (-1); 399 400 tp->ftt_segment = seg; 401 } 402 start++; 403 continue; 404 } 405 break; 406 } 407 408 #ifdef __amd64 409 /* 410 * Identify the REX prefix on 64-bit processes. 411 */ 412 if (p->p_model == DATAMODEL_LP64 && (instr[start] & 0xf0) == 0x40) 413 rex = instr[start++]; 414 #endif 415 416 /* 417 * Now that we're pretty sure that the instruction is okay, copy the 418 * valid part to the tracepoint. 419 */ 420 bcopy(instr, tp->ftt_instr, FASTTRAP_MAX_INSTR_SIZE); 421 422 tp->ftt_type = FASTTRAP_T_COMMON; 423 if (instr[start] == FASTTRAP_2_BYTE_OP) { 424 switch (instr[start + 1]) { 425 case FASTTRAP_0F_JO: 426 case FASTTRAP_0F_JNO: 427 case FASTTRAP_0F_JB: 428 case FASTTRAP_0F_JAE: 429 case FASTTRAP_0F_JE: 430 case FASTTRAP_0F_JNE: 431 case FASTTRAP_0F_JBE: 432 case FASTTRAP_0F_JA: 433 case FASTTRAP_0F_JS: 434 case FASTTRAP_0F_JNS: 435 case FASTTRAP_0F_JP: 436 case FASTTRAP_0F_JNP: 437 case FASTTRAP_0F_JL: 438 case FASTTRAP_0F_JGE: 439 case FASTTRAP_0F_JLE: 440 case FASTTRAP_0F_JG: 441 tp->ftt_type = FASTTRAP_T_JCC; 442 tp->ftt_code = (instr[start + 1] & 0x0f) | FASTTRAP_JO; 443 tp->ftt_dest = pc + tp->ftt_size + 444 /* LINTED - alignment */ 445 *(int32_t *)&instr[start + 2]; 446 break; 447 } 448 } else if (instr[start] == FASTTRAP_GROUP5_OP) { 449 uint_t mod = FASTTRAP_MODRM_MOD(instr[start + 1]); 450 uint_t reg = FASTTRAP_MODRM_REG(instr[start + 1]); 451 uint_t rm = FASTTRAP_MODRM_RM(instr[start + 1]); 452 453 if (reg == 2 || reg == 4) { 454 uint_t i, sz; 455 456 if (reg == 2) 457 tp->ftt_type = FASTTRAP_T_CALL; 458 else 459 tp->ftt_type = FASTTRAP_T_JMP; 460 461 if (mod == 3) 462 tp->ftt_code = 2; 463 else 464 tp->ftt_code = 1; 465 466 ASSERT(p->p_model == DATAMODEL_LP64 || rex == 0); 467 468 /* 469 * See AMD x86-64 Architecture Programmer's Manual 470 * Volume 3, Section 1.2.7, Table 1-12, and 471 * Appendix A.3.1, Table A-15. 472 */ 473 if (mod != 3 && rm == 4) { 474 uint8_t sib = instr[start + 2]; 475 uint_t index = FASTTRAP_SIB_INDEX(sib); 476 uint_t base = FASTTRAP_SIB_BASE(sib); 477 478 tp->ftt_scale = FASTTRAP_SIB_SCALE(sib); 479 480 tp->ftt_index = (index == 4) ? 481 FASTTRAP_NOREG : 482 regmap[index | (FASTTRAP_REX_X(rex) << 3)]; 483 tp->ftt_base = (mod == 0 && base == 5) ? 484 FASTTRAP_NOREG : 485 regmap[base | (FASTTRAP_REX_B(rex) << 3)]; 486 487 i = 3; 488 sz = mod == 1 ? 1 : 4; 489 } else { 490 /* 491 * In 64-bit mode, mod == 0 and r/m == 5 492 * denotes %rip-relative addressing; in 32-bit 493 * mode, the base register isn't used. In both 494 * modes, there is a 32-bit operand. 495 */ 496 if (mod == 0 && rm == 5) { 497 #ifdef __amd64 498 if (p->p_model == DATAMODEL_LP64) 499 tp->ftt_base = REG_RIP; 500 else 501 #endif 502 tp->ftt_base = FASTTRAP_NOREG; 503 sz = 4; 504 } else { 505 uint8_t base = rm | 506 (FASTTRAP_REX_B(rex) << 3); 507 508 tp->ftt_base = regmap[base]; 509 sz = mod == 1 ? 1 : mod == 2 ? 4 : 0; 510 } 511 tp->ftt_index = FASTTRAP_NOREG; 512 i = 2; 513 } 514 515 if (sz == 1) { 516 tp->ftt_dest = *(int8_t *)&instr[start + i]; 517 } else if (sz == 4) { 518 /* LINTED - alignment */ 519 tp->ftt_dest = *(int32_t *)&instr[start + i]; 520 } else { 521 tp->ftt_dest = 0; 522 } 523 } 524 } else { 525 switch (instr[start]) { 526 case FASTTRAP_RET: 527 tp->ftt_type = FASTTRAP_T_RET; 528 break; 529 530 case FASTTRAP_RET16: 531 tp->ftt_type = FASTTRAP_T_RET16; 532 /* LINTED - alignment */ 533 tp->ftt_dest = *(uint16_t *)&instr[start + 1]; 534 break; 535 536 case FASTTRAP_JO: 537 case FASTTRAP_JNO: 538 case FASTTRAP_JB: 539 case FASTTRAP_JAE: 540 case FASTTRAP_JE: 541 case FASTTRAP_JNE: 542 case FASTTRAP_JBE: 543 case FASTTRAP_JA: 544 case FASTTRAP_JS: 545 case FASTTRAP_JNS: 546 case FASTTRAP_JP: 547 case FASTTRAP_JNP: 548 case FASTTRAP_JL: 549 case FASTTRAP_JGE: 550 case FASTTRAP_JLE: 551 case FASTTRAP_JG: 552 tp->ftt_type = FASTTRAP_T_JCC; 553 tp->ftt_code = instr[start]; 554 tp->ftt_dest = pc + tp->ftt_size + 555 (int8_t)instr[start + 1]; 556 break; 557 558 case FASTTRAP_LOOPNZ: 559 case FASTTRAP_LOOPZ: 560 case FASTTRAP_LOOP: 561 tp->ftt_type = FASTTRAP_T_LOOP; 562 tp->ftt_code = instr[start]; 563 tp->ftt_dest = pc + tp->ftt_size + 564 (int8_t)instr[start + 1]; 565 break; 566 567 case FASTTRAP_JCXZ: 568 tp->ftt_type = FASTTRAP_T_JCXZ; 569 tp->ftt_dest = pc + tp->ftt_size + 570 (int8_t)instr[start + 1]; 571 break; 572 573 case FASTTRAP_CALL: 574 tp->ftt_type = FASTTRAP_T_CALL; 575 tp->ftt_dest = pc + tp->ftt_size + 576 /* LINTED - alignment */ 577 *(int32_t *)&instr[start + 1]; 578 tp->ftt_code = 0; 579 break; 580 581 case FASTTRAP_JMP32: 582 tp->ftt_type = FASTTRAP_T_JMP; 583 tp->ftt_dest = pc + tp->ftt_size + 584 /* LINTED - alignment */ 585 *(int32_t *)&instr[start + 1]; 586 break; 587 case FASTTRAP_JMP8: 588 tp->ftt_type = FASTTRAP_T_JMP; 589 tp->ftt_dest = pc + tp->ftt_size + 590 (int8_t)instr[start + 1]; 591 break; 592 593 case FASTTRAP_PUSHL_EBP: 594 if (start == 0) 595 tp->ftt_type = FASTTRAP_T_PUSHL_EBP; 596 break; 597 598 case FASTTRAP_NOP: 599 #ifdef __amd64 600 ASSERT(p->p_model == DATAMODEL_LP64 || rex == 0); 601 602 /* 603 * On amd64 we have to be careful not to confuse a nop 604 * (actually xchgl %eax, %eax) with an instruction using 605 * the same opcode, but that does something different 606 * (e.g. xchgl %r8d, %eax or xcghq %r8, %rax). 607 */ 608 if (FASTTRAP_REX_B(rex) == 0) 609 #endif 610 tp->ftt_type = FASTTRAP_T_NOP; 611 break; 612 613 case FASTTRAP_INT3: 614 /* 615 * The pid provider shares the int3 trap with debugger 616 * breakpoints so we can't instrument them. 617 */ 618 ASSERT(instr[start] == FASTTRAP_INSTR); 619 return (-1); 620 621 case FASTTRAP_INT: 622 /* 623 * Interrupts seem like they could be traced with 624 * no negative implications, but it's possible that 625 * a thread could be redirected by the trap handling 626 * code which would eventually return to the 627 * instruction after the interrupt. If the interrupt 628 * were in our scratch space, the subsequent 629 * instruction might be overwritten before we return. 630 * Accordingly we refuse to instrument any interrupt. 631 */ 632 return (-1); 633 } 634 } 635 636 #ifdef __amd64 637 if (p->p_model == DATAMODEL_LP64 && tp->ftt_type == FASTTRAP_T_COMMON) { 638 /* 639 * If the process is 64-bit and the instruction type is still 640 * FASTTRAP_T_COMMON -- meaning we're going to copy it out an 641 * execute it -- we need to watch for %rip-relative 642 * addressing mode. See the portion of fasttrap_pid_probe() 643 * below where we handle tracepoints with type 644 * FASTTRAP_T_COMMON for how we emulate instructions that 645 * employ %rip-relative addressing. 646 */ 647 if (rmindex != -1) { 648 uint_t mod = FASTTRAP_MODRM_MOD(instr[rmindex]); 649 uint_t reg = FASTTRAP_MODRM_REG(instr[rmindex]); 650 uint_t rm = FASTTRAP_MODRM_RM(instr[rmindex]); 651 652 ASSERT(rmindex > start); 653 654 if (mod == 0 && rm == 5) { 655 /* 656 * We need to be sure to avoid other 657 * registers used by this instruction. While 658 * the reg field may determine the op code 659 * rather than denoting a register, assuming 660 * that it denotes a register is always safe. 661 * We leave the REX field intact and use 662 * whatever value's there for simplicity. 663 */ 664 if (reg != 0) { 665 tp->ftt_ripmode = FASTTRAP_RIP_1 | 666 (FASTTRAP_RIP_X * 667 FASTTRAP_REX_B(rex)); 668 rm = 0; 669 } else { 670 tp->ftt_ripmode = FASTTRAP_RIP_2 | 671 (FASTTRAP_RIP_X * 672 FASTTRAP_REX_B(rex)); 673 rm = 1; 674 } 675 676 tp->ftt_modrm = tp->ftt_instr[rmindex]; 677 tp->ftt_instr[rmindex] = 678 FASTTRAP_MODRM(2, reg, rm); 679 } 680 } 681 } 682 #endif 683 684 return (0); 685 } 686 687 int 688 fasttrap_tracepoint_install(proc_t *p, fasttrap_tracepoint_t *tp) 689 { 690 fasttrap_instr_t instr = FASTTRAP_INSTR; 691 692 if (uwrite(p, &instr, 1, tp->ftt_pc) != 0) 693 return (-1); 694 695 return (0); 696 } 697 698 int 699 fasttrap_tracepoint_remove(proc_t *p, fasttrap_tracepoint_t *tp) 700 { 701 uint8_t instr; 702 703 /* 704 * Distinguish between read or write failures and a changed 705 * instruction. 706 */ 707 if (uread(p, &instr, 1, tp->ftt_pc) != 0) 708 return (0); 709 if (instr != FASTTRAP_INSTR) 710 return (0); 711 if (uwrite(p, &tp->ftt_instr[0], 1, tp->ftt_pc) != 0) 712 return (-1); 713 714 return (0); 715 } 716 717 #ifdef __amd64 718 static uintptr_t 719 fasttrap_fulword_noerr(const void *uaddr) 720 { 721 uintptr_t ret; 722 723 if ((ret = fasttrap_fulword(uaddr)) != -1) 724 return (ret); 725 726 return (0); 727 } 728 #endif 729 730 #ifdef __i386__ 731 static uint32_t 732 fasttrap_fuword32_noerr(const void *uaddr) 733 { 734 uint32_t ret; 735 736 if ((ret = fasttrap_fuword32(uaddr)) != -1) 737 return (ret); 738 739 return (0); 740 } 741 #endif 742 743 static void 744 fasttrap_return_common(struct reg *rp, uintptr_t pc, pid_t pid, 745 uintptr_t new_pc) 746 { 747 fasttrap_tracepoint_t *tp; 748 fasttrap_bucket_t *bucket; 749 fasttrap_id_t *id; 750 #if defined(sun) 751 kmutex_t *pid_mtx; 752 #endif 753 754 #if defined(sun) 755 pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock; 756 mutex_enter(pid_mtx); 757 #endif 758 bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)]; 759 760 for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) { 761 if (pid == tp->ftt_pid && pc == tp->ftt_pc && 762 tp->ftt_proc->ftpc_acount != 0) 763 break; 764 } 765 766 /* 767 * Don't sweat it if we can't find the tracepoint again; unlike 768 * when we're in fasttrap_pid_probe(), finding the tracepoint here 769 * is not essential to the correct execution of the process. 770 */ 771 if (tp == NULL) { 772 #if defined(sun) 773 mutex_exit(pid_mtx); 774 #endif 775 return; 776 } 777 778 for (id = tp->ftt_retids; id != NULL; id = id->fti_next) { 779 /* 780 * If there's a branch that could act as a return site, we 781 * need to trace it, and check here if the program counter is 782 * external to the function. 783 */ 784 if (tp->ftt_type != FASTTRAP_T_RET && 785 tp->ftt_type != FASTTRAP_T_RET16 && 786 new_pc - id->fti_probe->ftp_faddr < 787 id->fti_probe->ftp_fsize) 788 continue; 789 790 dtrace_probe(id->fti_probe->ftp_id, 791 pc - id->fti_probe->ftp_faddr, 792 rp->r_rax, rp->r_rbx, 0, 0); 793 } 794 795 #if defined(sun) 796 mutex_exit(pid_mtx); 797 #endif 798 } 799 800 static void 801 fasttrap_sigsegv(proc_t *p, kthread_t *t, uintptr_t addr) 802 { 803 #if defined(sun) 804 sigqueue_t *sqp = kmem_zalloc(sizeof (sigqueue_t), KM_SLEEP); 805 806 sqp->sq_info.si_signo = SIGSEGV; 807 sqp->sq_info.si_code = SEGV_MAPERR; 808 sqp->sq_info.si_addr = (caddr_t)addr; 809 810 mutex_enter(&p->p_lock); 811 sigaddqa(p, t, sqp); 812 mutex_exit(&p->p_lock); 813 814 if (t != NULL) 815 aston(t); 816 #else 817 ksiginfo_t *ksi = kmem_zalloc(sizeof (ksiginfo_t), KM_SLEEP); 818 819 ksiginfo_init(ksi); 820 ksi->ksi_signo = SIGSEGV; 821 ksi->ksi_code = SEGV_MAPERR; 822 ksi->ksi_addr = (caddr_t)addr; 823 (void) tdksignal(t, SIGSEGV, ksi); 824 #endif 825 } 826 827 #ifdef __amd64 828 static void 829 fasttrap_usdt_args64(fasttrap_probe_t *probe, struct reg *rp, int argc, 830 uintptr_t *argv) 831 { 832 int i, x, cap = MIN(argc, probe->ftp_nargs); 833 uintptr_t *stack = (uintptr_t *)rp->r_rsp; 834 835 for (i = 0; i < cap; i++) { 836 x = probe->ftp_argmap[i]; 837 838 if (x < 6) 839 argv[i] = (&rp->r_rdi)[x]; 840 else 841 argv[i] = fasttrap_fulword_noerr(&stack[x]); 842 } 843 844 for (; i < argc; i++) { 845 argv[i] = 0; 846 } 847 } 848 #endif 849 850 #ifdef __i386__ 851 static void 852 fasttrap_usdt_args32(fasttrap_probe_t *probe, struct reg *rp, int argc, 853 uint32_t *argv) 854 { 855 int i, x, cap = MIN(argc, probe->ftp_nargs); 856 uint32_t *stack = (uint32_t *)rp->r_rsp; 857 858 for (i = 0; i < cap; i++) { 859 x = probe->ftp_argmap[i]; 860 861 argv[i] = fasttrap_fuword32_noerr(&stack[x]); 862 } 863 864 for (; i < argc; i++) { 865 argv[i] = 0; 866 } 867 } 868 #endif 869 870 static int 871 fasttrap_do_seg(fasttrap_tracepoint_t *tp, struct reg *rp, uintptr_t *addr) 872 { 873 proc_t *p = curproc; 874 #ifdef __i386__ 875 struct segment_descriptor *desc; 876 #else 877 struct user_segment_descriptor *desc; 878 #endif 879 uint16_t sel = 0, ndx, type; 880 uintptr_t limit; 881 882 switch (tp->ftt_segment) { 883 case FASTTRAP_SEG_CS: 884 sel = rp->r_cs; 885 break; 886 case FASTTRAP_SEG_DS: 887 sel = rp->r_ds; 888 break; 889 case FASTTRAP_SEG_ES: 890 sel = rp->r_es; 891 break; 892 case FASTTRAP_SEG_FS: 893 sel = rp->r_fs; 894 break; 895 case FASTTRAP_SEG_GS: 896 sel = rp->r_gs; 897 break; 898 case FASTTRAP_SEG_SS: 899 sel = rp->r_ss; 900 break; 901 } 902 903 /* 904 * Make sure the given segment register specifies a user priority 905 * selector rather than a kernel selector. 906 */ 907 if (ISPL(sel) != SEL_UPL) 908 return (-1); 909 910 ndx = IDXSEL(sel); 911 912 /* 913 * Check the bounds and grab the descriptor out of the specified 914 * descriptor table. 915 */ 916 if (ISLDT(sel)) { 917 #ifdef __i386__ 918 if (ndx > p->p_md.md_ldt->ldt_len) 919 return (-1); 920 921 desc = (struct segment_descriptor *) 922 p->p_md.md_ldt[ndx].ldt_base; 923 #else 924 if (ndx > max_ldt_segment) 925 return (-1); 926 927 desc = (struct user_segment_descriptor *) 928 p->p_md.md_ldt[ndx].ldt_base; 929 #endif 930 931 } else { 932 if (ndx >= NGDT) 933 return (-1); 934 935 #ifdef __i386__ 936 desc = &gdt[ndx].sd; 937 #else 938 desc = &gdt[ndx]; 939 #endif 940 } 941 942 /* 943 * The descriptor must have user privilege level and it must be 944 * present in memory. 945 */ 946 if (desc->sd_dpl != SEL_UPL || desc->sd_p != 1) 947 return (-1); 948 949 type = desc->sd_type; 950 951 /* 952 * If the S bit in the type field is not set, this descriptor can 953 * only be used in system context. 954 */ 955 if ((type & 0x10) != 0x10) 956 return (-1); 957 958 limit = USD_GETLIMIT(desc) * (desc->sd_gran ? PAGESIZE : 1); 959 960 if (tp->ftt_segment == FASTTRAP_SEG_CS) { 961 /* 962 * The code/data bit and readable bit must both be set. 963 */ 964 if ((type & 0xa) != 0xa) 965 return (-1); 966 967 if (*addr > limit) 968 return (-1); 969 } else { 970 /* 971 * The code/data bit must be clear. 972 */ 973 if ((type & 0x8) != 0) 974 return (-1); 975 976 /* 977 * If the expand-down bit is clear, we just check the limit as 978 * it would naturally be applied. Otherwise, we need to check 979 * that the address is the range [limit + 1 .. 0xffff] or 980 * [limit + 1 ... 0xffffffff] depending on if the default 981 * operand size bit is set. 982 */ 983 if ((type & 0x4) == 0) { 984 if (*addr > limit) 985 return (-1); 986 } else if (desc->sd_def32) { 987 if (*addr < limit + 1 || 0xffff < *addr) 988 return (-1); 989 } else { 990 if (*addr < limit + 1 || 0xffffffff < *addr) 991 return (-1); 992 } 993 } 994 995 *addr += USD_GETBASE(desc); 996 997 return (0); 998 } 999 1000 int 1001 fasttrap_pid_probe(struct reg *rp) 1002 { 1003 proc_t *p = curproc; 1004 #if !defined(sun) 1005 proc_t *pp; 1006 #endif 1007 uintptr_t pc = rp->r_rip - 1; 1008 uintptr_t new_pc = 0; 1009 fasttrap_bucket_t *bucket; 1010 #if defined(sun) 1011 kmutex_t *pid_mtx; 1012 #endif 1013 fasttrap_tracepoint_t *tp, tp_local; 1014 pid_t pid; 1015 dtrace_icookie_t cookie; 1016 uint_t is_enabled = 0; 1017 1018 /* 1019 * It's possible that a user (in a veritable orgy of bad planning) 1020 * could redirect this thread's flow of control before it reached the 1021 * return probe fasttrap. In this case we need to kill the process 1022 * since it's in a unrecoverable state. 1023 */ 1024 if (curthread->t_dtrace_step) { 1025 ASSERT(curthread->t_dtrace_on); 1026 fasttrap_sigtrap(p, curthread, pc); 1027 return (0); 1028 } 1029 1030 /* 1031 * Clear all user tracing flags. 1032 */ 1033 curthread->t_dtrace_ft = 0; 1034 curthread->t_dtrace_pc = 0; 1035 curthread->t_dtrace_npc = 0; 1036 curthread->t_dtrace_scrpc = 0; 1037 curthread->t_dtrace_astpc = 0; 1038 #ifdef __amd64 1039 curthread->t_dtrace_regv = 0; 1040 #endif 1041 1042 /* 1043 * Treat a child created by a call to vfork(2) as if it were its 1044 * parent. We know that there's only one thread of control in such a 1045 * process: this one. 1046 */ 1047 #if defined(sun) 1048 while (p->p_flag & SVFORK) { 1049 p = p->p_parent; 1050 } 1051 1052 pid = p->p_pid; 1053 pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock; 1054 mutex_enter(pid_mtx); 1055 #else 1056 pp = p; 1057 sx_slock(&proctree_lock); 1058 while (pp->p_vmspace == pp->p_pptr->p_vmspace) 1059 pp = pp->p_pptr; 1060 pid = pp->p_pid; 1061 sx_sunlock(&proctree_lock); 1062 pp = NULL; 1063 1064 PROC_LOCK(p); 1065 _PHOLD(p); 1066 #endif 1067 1068 bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)]; 1069 1070 /* 1071 * Lookup the tracepoint that the process just hit. 1072 */ 1073 for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) { 1074 if (pid == tp->ftt_pid && pc == tp->ftt_pc && 1075 tp->ftt_proc->ftpc_acount != 0) 1076 break; 1077 } 1078 1079 /* 1080 * If we couldn't find a matching tracepoint, either a tracepoint has 1081 * been inserted without using the pid<pid> ioctl interface (see 1082 * fasttrap_ioctl), or somehow we have mislaid this tracepoint. 1083 */ 1084 if (tp == NULL) { 1085 #if defined(sun) 1086 mutex_exit(pid_mtx); 1087 #else 1088 _PRELE(p); 1089 PROC_UNLOCK(p); 1090 #endif 1091 return (-1); 1092 } 1093 1094 /* 1095 * Set the program counter to the address of the traced instruction 1096 * so that it looks right in ustack() output. 1097 */ 1098 rp->r_rip = pc; 1099 1100 if (tp->ftt_ids != NULL) { 1101 fasttrap_id_t *id; 1102 1103 #ifdef __amd64 1104 if (p->p_model == DATAMODEL_LP64) { 1105 for (id = tp->ftt_ids; id != NULL; id = id->fti_next) { 1106 fasttrap_probe_t *probe = id->fti_probe; 1107 1108 if (id->fti_ptype == DTFTP_ENTRY) { 1109 /* 1110 * We note that this was an entry 1111 * probe to help ustack() find the 1112 * first caller. 1113 */ 1114 cookie = dtrace_interrupt_disable(); 1115 DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY); 1116 dtrace_probe(probe->ftp_id, rp->r_rdi, 1117 rp->r_rsi, rp->r_rdx, rp->r_rcx, 1118 rp->r_r8); 1119 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY); 1120 dtrace_interrupt_enable(cookie); 1121 } else if (id->fti_ptype == DTFTP_IS_ENABLED) { 1122 /* 1123 * Note that in this case, we don't 1124 * call dtrace_probe() since it's only 1125 * an artificial probe meant to change 1126 * the flow of control so that it 1127 * encounters the true probe. 1128 */ 1129 is_enabled = 1; 1130 } else if (probe->ftp_argmap == NULL) { 1131 dtrace_probe(probe->ftp_id, rp->r_rdi, 1132 rp->r_rsi, rp->r_rdx, rp->r_rcx, 1133 rp->r_r8); 1134 } else { 1135 uintptr_t t[5]; 1136 1137 fasttrap_usdt_args64(probe, rp, 1138 sizeof (t) / sizeof (t[0]), t); 1139 1140 dtrace_probe(probe->ftp_id, t[0], t[1], 1141 t[2], t[3], t[4]); 1142 } 1143 } 1144 } else { 1145 #else /* __amd64 */ 1146 uintptr_t s0, s1, s2, s3, s4, s5; 1147 uint32_t *stack = (uint32_t *)rp->r_esp; 1148 1149 /* 1150 * In 32-bit mode, all arguments are passed on the 1151 * stack. If this is a function entry probe, we need 1152 * to skip the first entry on the stack as it 1153 * represents the return address rather than a 1154 * parameter to the function. 1155 */ 1156 s0 = fasttrap_fuword32_noerr(&stack[0]); 1157 s1 = fasttrap_fuword32_noerr(&stack[1]); 1158 s2 = fasttrap_fuword32_noerr(&stack[2]); 1159 s3 = fasttrap_fuword32_noerr(&stack[3]); 1160 s4 = fasttrap_fuword32_noerr(&stack[4]); 1161 s5 = fasttrap_fuword32_noerr(&stack[5]); 1162 1163 for (id = tp->ftt_ids; id != NULL; id = id->fti_next) { 1164 fasttrap_probe_t *probe = id->fti_probe; 1165 1166 if (id->fti_ptype == DTFTP_ENTRY) { 1167 /* 1168 * We note that this was an entry 1169 * probe to help ustack() find the 1170 * first caller. 1171 */ 1172 cookie = dtrace_interrupt_disable(); 1173 DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY); 1174 dtrace_probe(probe->ftp_id, s1, s2, 1175 s3, s4, s5); 1176 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY); 1177 dtrace_interrupt_enable(cookie); 1178 } else if (id->fti_ptype == DTFTP_IS_ENABLED) { 1179 /* 1180 * Note that in this case, we don't 1181 * call dtrace_probe() since it's only 1182 * an artificial probe meant to change 1183 * the flow of control so that it 1184 * encounters the true probe. 1185 */ 1186 is_enabled = 1; 1187 } else if (probe->ftp_argmap == NULL) { 1188 dtrace_probe(probe->ftp_id, s0, s1, 1189 s2, s3, s4); 1190 } else { 1191 uint32_t t[5]; 1192 1193 fasttrap_usdt_args32(probe, rp, 1194 sizeof (t) / sizeof (t[0]), t); 1195 1196 dtrace_probe(probe->ftp_id, t[0], t[1], 1197 t[2], t[3], t[4]); 1198 } 1199 } 1200 #endif /* __amd64 */ 1201 #ifdef __amd64 1202 } 1203 #endif 1204 } 1205 1206 /* 1207 * We're about to do a bunch of work so we cache a local copy of 1208 * the tracepoint to emulate the instruction, and then find the 1209 * tracepoint again later if we need to light up any return probes. 1210 */ 1211 tp_local = *tp; 1212 #if defined(sun) 1213 mutex_exit(pid_mtx); 1214 #else 1215 PROC_UNLOCK(p); 1216 #endif 1217 tp = &tp_local; 1218 1219 /* 1220 * Set the program counter to appear as though the traced instruction 1221 * had completely executed. This ensures that fasttrap_getreg() will 1222 * report the expected value for REG_RIP. 1223 */ 1224 rp->r_rip = pc + tp->ftt_size; 1225 1226 /* 1227 * If there's an is-enabled probe connected to this tracepoint it 1228 * means that there was a 'xorl %eax, %eax' or 'xorq %rax, %rax' 1229 * instruction that was placed there by DTrace when the binary was 1230 * linked. As this probe is, in fact, enabled, we need to stuff 1 1231 * into %eax or %rax. Accordingly, we can bypass all the instruction 1232 * emulation logic since we know the inevitable result. It's possible 1233 * that a user could construct a scenario where the 'is-enabled' 1234 * probe was on some other instruction, but that would be a rather 1235 * exotic way to shoot oneself in the foot. 1236 */ 1237 if (is_enabled) { 1238 rp->r_rax = 1; 1239 new_pc = rp->r_rip; 1240 goto done; 1241 } 1242 1243 /* 1244 * We emulate certain types of instructions to ensure correctness 1245 * (in the case of position dependent instructions) or optimize 1246 * common cases. The rest we have the thread execute back in user- 1247 * land. 1248 */ 1249 switch (tp->ftt_type) { 1250 case FASTTRAP_T_RET: 1251 case FASTTRAP_T_RET16: 1252 { 1253 uintptr_t dst = 0; 1254 uintptr_t addr = 0; 1255 int ret = 0; 1256 1257 /* 1258 * We have to emulate _every_ facet of the behavior of a ret 1259 * instruction including what happens if the load from %esp 1260 * fails; in that case, we send a SIGSEGV. 1261 */ 1262 #ifdef __amd64 1263 if (p->p_model == DATAMODEL_NATIVE) { 1264 ret = dst = fasttrap_fulword((void *)rp->r_rsp); 1265 addr = rp->r_rsp + sizeof (uintptr_t); 1266 } else { 1267 #endif 1268 #ifdef __i386__ 1269 uint32_t dst32; 1270 ret = dst32 = fasttrap_fuword32((void *)rp->r_esp); 1271 dst = dst32; 1272 addr = rp->r_esp + sizeof (uint32_t); 1273 #endif 1274 #ifdef __amd64 1275 } 1276 #endif 1277 1278 if (ret == -1) { 1279 fasttrap_sigsegv(p, curthread, rp->r_rsp); 1280 new_pc = pc; 1281 break; 1282 } 1283 1284 if (tp->ftt_type == FASTTRAP_T_RET16) 1285 addr += tp->ftt_dest; 1286 1287 rp->r_rsp = addr; 1288 new_pc = dst; 1289 break; 1290 } 1291 1292 case FASTTRAP_T_JCC: 1293 { 1294 uint_t taken = 0; 1295 1296 switch (tp->ftt_code) { 1297 case FASTTRAP_JO: 1298 taken = (rp->r_rflags & FASTTRAP_EFLAGS_OF) != 0; 1299 break; 1300 case FASTTRAP_JNO: 1301 taken = (rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0; 1302 break; 1303 case FASTTRAP_JB: 1304 taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) != 0; 1305 break; 1306 case FASTTRAP_JAE: 1307 taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) == 0; 1308 break; 1309 case FASTTRAP_JE: 1310 taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0; 1311 break; 1312 case FASTTRAP_JNE: 1313 taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0; 1314 break; 1315 case FASTTRAP_JBE: 1316 taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) != 0 || 1317 (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0; 1318 break; 1319 case FASTTRAP_JA: 1320 taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) == 0 && 1321 (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0; 1322 break; 1323 case FASTTRAP_JS: 1324 taken = (rp->r_rflags & FASTTRAP_EFLAGS_SF) != 0; 1325 break; 1326 case FASTTRAP_JNS: 1327 taken = (rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0; 1328 break; 1329 case FASTTRAP_JP: 1330 taken = (rp->r_rflags & FASTTRAP_EFLAGS_PF) != 0; 1331 break; 1332 case FASTTRAP_JNP: 1333 taken = (rp->r_rflags & FASTTRAP_EFLAGS_PF) == 0; 1334 break; 1335 case FASTTRAP_JL: 1336 taken = ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) != 1337 ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0); 1338 break; 1339 case FASTTRAP_JGE: 1340 taken = ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) == 1341 ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0); 1342 break; 1343 case FASTTRAP_JLE: 1344 taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0 || 1345 ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) != 1346 ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0); 1347 break; 1348 case FASTTRAP_JG: 1349 taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0 && 1350 ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) == 1351 ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0); 1352 break; 1353 1354 } 1355 1356 if (taken) 1357 new_pc = tp->ftt_dest; 1358 else 1359 new_pc = pc + tp->ftt_size; 1360 break; 1361 } 1362 1363 case FASTTRAP_T_LOOP: 1364 { 1365 uint_t taken = 0; 1366 #ifdef __amd64 1367 greg_t cx = rp->r_rcx--; 1368 #else 1369 greg_t cx = rp->r_ecx--; 1370 #endif 1371 1372 switch (tp->ftt_code) { 1373 case FASTTRAP_LOOPNZ: 1374 taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0 && 1375 cx != 0; 1376 break; 1377 case FASTTRAP_LOOPZ: 1378 taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0 && 1379 cx != 0; 1380 break; 1381 case FASTTRAP_LOOP: 1382 taken = (cx != 0); 1383 break; 1384 } 1385 1386 if (taken) 1387 new_pc = tp->ftt_dest; 1388 else 1389 new_pc = pc + tp->ftt_size; 1390 break; 1391 } 1392 1393 case FASTTRAP_T_JCXZ: 1394 { 1395 #ifdef __amd64 1396 greg_t cx = rp->r_rcx; 1397 #else 1398 greg_t cx = rp->r_ecx; 1399 #endif 1400 1401 if (cx == 0) 1402 new_pc = tp->ftt_dest; 1403 else 1404 new_pc = pc + tp->ftt_size; 1405 break; 1406 } 1407 1408 case FASTTRAP_T_PUSHL_EBP: 1409 { 1410 int ret = 0; 1411 1412 #ifdef __amd64 1413 if (p->p_model == DATAMODEL_NATIVE) { 1414 rp->r_rsp -= sizeof (uintptr_t); 1415 ret = fasttrap_sulword((void *)rp->r_rsp, rp->r_rbp); 1416 } else { 1417 #endif 1418 #ifdef __i386__ 1419 rp->r_rsp -= sizeof (uint32_t); 1420 ret = fasttrap_suword32((void *)rp->r_rsp, rp->r_rbp); 1421 #endif 1422 #ifdef __amd64 1423 } 1424 #endif 1425 1426 if (ret == -1) { 1427 fasttrap_sigsegv(p, curthread, rp->r_rsp); 1428 new_pc = pc; 1429 break; 1430 } 1431 1432 new_pc = pc + tp->ftt_size; 1433 break; 1434 } 1435 1436 case FASTTRAP_T_NOP: 1437 new_pc = pc + tp->ftt_size; 1438 break; 1439 1440 case FASTTRAP_T_JMP: 1441 case FASTTRAP_T_CALL: 1442 if (tp->ftt_code == 0) { 1443 new_pc = tp->ftt_dest; 1444 } else { 1445 #ifdef __amd64 1446 uintptr_t value; 1447 #endif 1448 uintptr_t addr = tp->ftt_dest; 1449 1450 if (tp->ftt_base != FASTTRAP_NOREG) 1451 addr += fasttrap_getreg(rp, tp->ftt_base); 1452 if (tp->ftt_index != FASTTRAP_NOREG) 1453 addr += fasttrap_getreg(rp, tp->ftt_index) << 1454 tp->ftt_scale; 1455 1456 if (tp->ftt_code == 1) { 1457 /* 1458 * If there's a segment prefix for this 1459 * instruction, we'll need to check permissions 1460 * and bounds on the given selector, and adjust 1461 * the address accordingly. 1462 */ 1463 if (tp->ftt_segment != FASTTRAP_SEG_NONE && 1464 fasttrap_do_seg(tp, rp, &addr) != 0) { 1465 fasttrap_sigsegv(p, curthread, addr); 1466 new_pc = pc; 1467 break; 1468 } 1469 1470 #ifdef __amd64 1471 if (p->p_model == DATAMODEL_NATIVE) { 1472 if ((value = fasttrap_fulword((void *)addr)) 1473 == -1) { 1474 fasttrap_sigsegv(p, curthread, 1475 addr); 1476 new_pc = pc; 1477 break; 1478 } 1479 new_pc = value; 1480 } else { 1481 #endif 1482 #ifdef __i386__ 1483 uint32_t value32; 1484 addr = (uintptr_t)(uint32_t)addr; 1485 if ((value32 = fasttrap_fuword32((void *)addr)) 1486 == -1) { 1487 fasttrap_sigsegv(p, curthread, 1488 addr); 1489 new_pc = pc; 1490 break; 1491 } 1492 new_pc = value32; 1493 #endif 1494 } 1495 #ifdef __amd64 1496 } else { 1497 new_pc = addr; 1498 } 1499 #endif 1500 } 1501 1502 /* 1503 * If this is a call instruction, we need to push the return 1504 * address onto the stack. If this fails, we send the process 1505 * a SIGSEGV and reset the pc to emulate what would happen if 1506 * this instruction weren't traced. 1507 */ 1508 if (tp->ftt_type == FASTTRAP_T_CALL) { 1509 int ret = 0; 1510 uintptr_t addr = 0, pcps; 1511 #ifdef __amd64 1512 if (p->p_model == DATAMODEL_NATIVE) { 1513 addr = rp->r_rsp - sizeof (uintptr_t); 1514 pcps = pc + tp->ftt_size; 1515 ret = fasttrap_sulword((void *)addr, pcps); 1516 } else { 1517 #endif 1518 #ifdef __i386__ 1519 addr = rp->r_rsp - sizeof (uint32_t); 1520 pcps = (uint32_t)(pc + tp->ftt_size); 1521 ret = fasttrap_suword32((void *)addr, pcps); 1522 #endif 1523 #ifdef __amd64 1524 } 1525 #endif 1526 1527 if (ret == -1) { 1528 fasttrap_sigsegv(p, curthread, addr); 1529 new_pc = pc; 1530 break; 1531 } 1532 1533 rp->r_rsp = addr; 1534 } 1535 1536 break; 1537 1538 case FASTTRAP_T_COMMON: 1539 { 1540 uintptr_t addr; 1541 #if defined(__amd64) 1542 uint8_t scratch[2 * FASTTRAP_MAX_INSTR_SIZE + 22]; 1543 #else 1544 uint8_t scratch[2 * FASTTRAP_MAX_INSTR_SIZE + 7]; 1545 #endif 1546 uint_t i = 0; 1547 #if defined(sun) 1548 klwp_t *lwp = ttolwp(curthread); 1549 #endif 1550 1551 /* 1552 * Compute the address of the ulwp_t and step over the 1553 * ul_self pointer. The method used to store the user-land 1554 * thread pointer is very different on 32- and 64-bit 1555 * kernels. 1556 */ 1557 #if defined(sun) 1558 #if defined(__amd64) 1559 if (p->p_model == DATAMODEL_LP64) { 1560 addr = lwp->lwp_pcb.pcb_fsbase; 1561 addr += sizeof (void *); 1562 } else { 1563 addr = lwp->lwp_pcb.pcb_gsbase; 1564 addr += sizeof (caddr32_t); 1565 } 1566 #else 1567 addr = USD_GETBASE(&lwp->lwp_pcb.pcb_gsdesc); 1568 addr += sizeof (void *); 1569 #endif 1570 #endif /* sun */ 1571 #ifdef __i386__ 1572 addr = USD_GETBASE(&curthread->td_pcb->pcb_gsd); 1573 #else 1574 addr = curthread->td_pcb->pcb_gsbase; 1575 #endif 1576 addr += sizeof (void *); 1577 1578 /* 1579 * Generic Instruction Tracing 1580 * --------------------------- 1581 * 1582 * This is the layout of the scratch space in the user-land 1583 * thread structure for our generated instructions. 1584 * 1585 * 32-bit mode bytes 1586 * ------------------------ ----- 1587 * a: <original instruction> <= 15 1588 * jmp <pc + tp->ftt_size> 5 1589 * b: <original instruction> <= 15 1590 * int T_DTRACE_RET 2 1591 * ----- 1592 * <= 37 1593 * 1594 * 64-bit mode bytes 1595 * ------------------------ ----- 1596 * a: <original instruction> <= 15 1597 * jmp 0(%rip) 6 1598 * <pc + tp->ftt_size> 8 1599 * b: <original instruction> <= 15 1600 * int T_DTRACE_RET 2 1601 * ----- 1602 * <= 46 1603 * 1604 * The %pc is set to a, and curthread->t_dtrace_astpc is set 1605 * to b. If we encounter a signal on the way out of the 1606 * kernel, trap() will set %pc to curthread->t_dtrace_astpc 1607 * so that we execute the original instruction and re-enter 1608 * the kernel rather than redirecting to the next instruction. 1609 * 1610 * If there are return probes (so we know that we're going to 1611 * need to reenter the kernel after executing the original 1612 * instruction), the scratch space will just contain the 1613 * original instruction followed by an interrupt -- the same 1614 * data as at b. 1615 * 1616 * %rip-relative Addressing 1617 * ------------------------ 1618 * 1619 * There's a further complication in 64-bit mode due to %rip- 1620 * relative addressing. While this is clearly a beneficial 1621 * architectural decision for position independent code, it's 1622 * hard not to see it as a personal attack against the pid 1623 * provider since before there was a relatively small set of 1624 * instructions to emulate; with %rip-relative addressing, 1625 * almost every instruction can potentially depend on the 1626 * address at which it's executed. Rather than emulating 1627 * the broad spectrum of instructions that can now be 1628 * position dependent, we emulate jumps and others as in 1629 * 32-bit mode, and take a different tack for instructions 1630 * using %rip-relative addressing. 1631 * 1632 * For every instruction that uses the ModRM byte, the 1633 * in-kernel disassembler reports its location. We use the 1634 * ModRM byte to identify that an instruction uses 1635 * %rip-relative addressing and to see what other registers 1636 * the instruction uses. To emulate those instructions, 1637 * we modify the instruction to be %rax-relative rather than 1638 * %rip-relative (or %rcx-relative if the instruction uses 1639 * %rax; or %r8- or %r9-relative if the REX.B is present so 1640 * we don't have to rewrite the REX prefix). We then load 1641 * the value that %rip would have been into the scratch 1642 * register and generate an instruction to reset the scratch 1643 * register back to its original value. The instruction 1644 * sequence looks like this: 1645 * 1646 * 64-mode %rip-relative bytes 1647 * ------------------------ ----- 1648 * a: <modified instruction> <= 15 1649 * movq $<value>, %<scratch> 6 1650 * jmp 0(%rip) 6 1651 * <pc + tp->ftt_size> 8 1652 * b: <modified instruction> <= 15 1653 * int T_DTRACE_RET 2 1654 * ----- 1655 * 52 1656 * 1657 * We set curthread->t_dtrace_regv so that upon receiving 1658 * a signal we can reset the value of the scratch register. 1659 */ 1660 1661 ASSERT(tp->ftt_size < FASTTRAP_MAX_INSTR_SIZE); 1662 1663 curthread->t_dtrace_scrpc = addr; 1664 bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size); 1665 i += tp->ftt_size; 1666 1667 #ifdef __amd64 1668 if (tp->ftt_ripmode != 0) { 1669 greg_t *reg = NULL; 1670 1671 ASSERT(p->p_model == DATAMODEL_LP64); 1672 ASSERT(tp->ftt_ripmode & 1673 (FASTTRAP_RIP_1 | FASTTRAP_RIP_2)); 1674 1675 /* 1676 * If this was a %rip-relative instruction, we change 1677 * it to be either a %rax- or %rcx-relative 1678 * instruction (depending on whether those registers 1679 * are used as another operand; or %r8- or %r9- 1680 * relative depending on the value of REX.B). We then 1681 * set that register and generate a movq instruction 1682 * to reset the value. 1683 */ 1684 if (tp->ftt_ripmode & FASTTRAP_RIP_X) 1685 scratch[i++] = FASTTRAP_REX(1, 0, 0, 1); 1686 else 1687 scratch[i++] = FASTTRAP_REX(1, 0, 0, 0); 1688 1689 if (tp->ftt_ripmode & FASTTRAP_RIP_1) 1690 scratch[i++] = FASTTRAP_MOV_EAX; 1691 else 1692 scratch[i++] = FASTTRAP_MOV_ECX; 1693 1694 switch (tp->ftt_ripmode) { 1695 case FASTTRAP_RIP_1: 1696 reg = &rp->r_rax; 1697 curthread->t_dtrace_reg = REG_RAX; 1698 break; 1699 case FASTTRAP_RIP_2: 1700 reg = &rp->r_rcx; 1701 curthread->t_dtrace_reg = REG_RCX; 1702 break; 1703 case FASTTRAP_RIP_1 | FASTTRAP_RIP_X: 1704 reg = &rp->r_r8; 1705 curthread->t_dtrace_reg = REG_R8; 1706 break; 1707 case FASTTRAP_RIP_2 | FASTTRAP_RIP_X: 1708 reg = &rp->r_r9; 1709 curthread->t_dtrace_reg = REG_R9; 1710 break; 1711 } 1712 1713 /* LINTED - alignment */ 1714 *(uint64_t *)&scratch[i] = *reg; 1715 curthread->t_dtrace_regv = *reg; 1716 *reg = pc + tp->ftt_size; 1717 i += sizeof (uint64_t); 1718 } 1719 #endif 1720 1721 /* 1722 * Generate the branch instruction to what would have 1723 * normally been the subsequent instruction. In 32-bit mode, 1724 * this is just a relative branch; in 64-bit mode this is a 1725 * %rip-relative branch that loads the 64-bit pc value 1726 * immediately after the jmp instruction. 1727 */ 1728 #ifdef __amd64 1729 if (p->p_model == DATAMODEL_LP64) { 1730 scratch[i++] = FASTTRAP_GROUP5_OP; 1731 scratch[i++] = FASTTRAP_MODRM(0, 4, 5); 1732 /* LINTED - alignment */ 1733 *(uint32_t *)&scratch[i] = 0; 1734 i += sizeof (uint32_t); 1735 /* LINTED - alignment */ 1736 *(uint64_t *)&scratch[i] = pc + tp->ftt_size; 1737 i += sizeof (uint64_t); 1738 } else { 1739 #endif 1740 #ifdef __i386__ 1741 /* 1742 * Set up the jmp to the next instruction; note that 1743 * the size of the traced instruction cancels out. 1744 */ 1745 scratch[i++] = FASTTRAP_JMP32; 1746 /* LINTED - alignment */ 1747 *(uint32_t *)&scratch[i] = pc - addr - 5; 1748 i += sizeof (uint32_t); 1749 #endif 1750 #ifdef __amd64 1751 } 1752 #endif 1753 1754 curthread->t_dtrace_astpc = addr + i; 1755 bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size); 1756 i += tp->ftt_size; 1757 scratch[i++] = FASTTRAP_INT; 1758 scratch[i++] = T_DTRACE_RET; 1759 1760 ASSERT(i <= sizeof (scratch)); 1761 1762 #if defined(sun) 1763 if (fasttrap_copyout(scratch, (char *)addr, i)) { 1764 #else 1765 if (uwrite(p, scratch, i, addr)) { 1766 #endif 1767 fasttrap_sigtrap(p, curthread, pc); 1768 new_pc = pc; 1769 break; 1770 } 1771 if (tp->ftt_retids != NULL) { 1772 curthread->t_dtrace_step = 1; 1773 curthread->t_dtrace_ret = 1; 1774 new_pc = curthread->t_dtrace_astpc; 1775 } else { 1776 new_pc = curthread->t_dtrace_scrpc; 1777 } 1778 1779 curthread->t_dtrace_pc = pc; 1780 curthread->t_dtrace_npc = pc + tp->ftt_size; 1781 curthread->t_dtrace_on = 1; 1782 break; 1783 } 1784 1785 default: 1786 panic("fasttrap: mishandled an instruction"); 1787 } 1788 1789 done: 1790 /* 1791 * If there were no return probes when we first found the tracepoint, 1792 * we should feel no obligation to honor any return probes that were 1793 * subsequently enabled -- they'll just have to wait until the next 1794 * time around. 1795 */ 1796 if (tp->ftt_retids != NULL) { 1797 /* 1798 * We need to wait until the results of the instruction are 1799 * apparent before invoking any return probes. If this 1800 * instruction was emulated we can just call 1801 * fasttrap_return_common(); if it needs to be executed, we 1802 * need to wait until the user thread returns to the kernel. 1803 */ 1804 if (tp->ftt_type != FASTTRAP_T_COMMON) { 1805 /* 1806 * Set the program counter to the address of the traced 1807 * instruction so that it looks right in ustack() 1808 * output. We had previously set it to the end of the 1809 * instruction to simplify %rip-relative addressing. 1810 */ 1811 rp->r_rip = pc; 1812 1813 fasttrap_return_common(rp, pc, pid, new_pc); 1814 } else { 1815 ASSERT(curthread->t_dtrace_ret != 0); 1816 ASSERT(curthread->t_dtrace_pc == pc); 1817 ASSERT(curthread->t_dtrace_scrpc != 0); 1818 ASSERT(new_pc == curthread->t_dtrace_astpc); 1819 } 1820 } 1821 1822 rp->r_rip = new_pc; 1823 1824 #if !defined(sun) 1825 PROC_LOCK(p); 1826 proc_write_regs(curthread, rp); 1827 _PRELE(p); 1828 PROC_UNLOCK(p); 1829 #endif 1830 1831 return (0); 1832 } 1833 1834 int 1835 fasttrap_return_probe(struct reg *rp) 1836 { 1837 proc_t *p = curproc; 1838 uintptr_t pc = curthread->t_dtrace_pc; 1839 uintptr_t npc = curthread->t_dtrace_npc; 1840 1841 curthread->t_dtrace_pc = 0; 1842 curthread->t_dtrace_npc = 0; 1843 curthread->t_dtrace_scrpc = 0; 1844 curthread->t_dtrace_astpc = 0; 1845 1846 #if defined(sun) 1847 /* 1848 * Treat a child created by a call to vfork(2) as if it were its 1849 * parent. We know that there's only one thread of control in such a 1850 * process: this one. 1851 */ 1852 while (p->p_flag & SVFORK) { 1853 p = p->p_parent; 1854 } 1855 #endif 1856 1857 /* 1858 * We set rp->r_rip to the address of the traced instruction so 1859 * that it appears to dtrace_probe() that we're on the original 1860 * instruction, and so that the user can't easily detect our 1861 * complex web of lies. dtrace_return_probe() (our caller) 1862 * will correctly set %pc after we return. 1863 */ 1864 rp->r_rip = pc; 1865 1866 fasttrap_return_common(rp, pc, p->p_pid, npc); 1867 1868 return (0); 1869 } 1870 1871 /*ARGSUSED*/ 1872 uint64_t 1873 fasttrap_pid_getarg(void *arg, dtrace_id_t id, void *parg, int argno, 1874 int aframes) 1875 { 1876 struct reg r; 1877 1878 fill_regs(curthread, &r); 1879 1880 return (fasttrap_anarg(&r, 1, argno)); 1881 } 1882 1883 /*ARGSUSED*/ 1884 uint64_t 1885 fasttrap_usdt_getarg(void *arg, dtrace_id_t id, void *parg, int argno, 1886 int aframes) 1887 { 1888 struct reg r; 1889 1890 fill_regs(curthread, &r); 1891 1892 return (fasttrap_anarg(&r, 0, argno)); 1893 } 1894 1895 static ulong_t 1896 fasttrap_getreg(struct reg *rp, uint_t reg) 1897 { 1898 #ifdef __amd64 1899 switch (reg) { 1900 case REG_R15: return (rp->r_r15); 1901 case REG_R14: return (rp->r_r14); 1902 case REG_R13: return (rp->r_r13); 1903 case REG_R12: return (rp->r_r12); 1904 case REG_R11: return (rp->r_r11); 1905 case REG_R10: return (rp->r_r10); 1906 case REG_R9: return (rp->r_r9); 1907 case REG_R8: return (rp->r_r8); 1908 case REG_RDI: return (rp->r_rdi); 1909 case REG_RSI: return (rp->r_rsi); 1910 case REG_RBP: return (rp->r_rbp); 1911 case REG_RBX: return (rp->r_rbx); 1912 case REG_RDX: return (rp->r_rdx); 1913 case REG_RCX: return (rp->r_rcx); 1914 case REG_RAX: return (rp->r_rax); 1915 case REG_TRAPNO: return (rp->r_trapno); 1916 case REG_ERR: return (rp->r_err); 1917 case REG_RIP: return (rp->r_rip); 1918 case REG_CS: return (rp->r_cs); 1919 #if defined(sun) 1920 case REG_RFL: return (rp->r_rfl); 1921 #endif 1922 case REG_RSP: return (rp->r_rsp); 1923 case REG_SS: return (rp->r_ss); 1924 case REG_FS: return (rp->r_fs); 1925 case REG_GS: return (rp->r_gs); 1926 case REG_DS: return (rp->r_ds); 1927 case REG_ES: return (rp->r_es); 1928 case REG_FSBASE: return (rdmsr(MSR_FSBASE)); 1929 case REG_GSBASE: return (rdmsr(MSR_GSBASE)); 1930 } 1931 1932 panic("dtrace: illegal register constant"); 1933 /*NOTREACHED*/ 1934 #else 1935 #define _NGREG 19 1936 if (reg >= _NGREG) 1937 panic("dtrace: illegal register constant"); 1938 1939 return (((greg_t *)&rp->r_gs)[reg]); 1940 #endif 1941 } 1942