1 /* 2 * SPDX-License-Identifier: CDDL 1.0 3 * 4 * Copyright (c) 2023 The FreeBSD Foundation 5 * 6 * This software was developed by Christos Margiolis <christos@FreeBSD.org> 7 * under sponsorship from the FreeBSD Foundation. 8 */ 9 10 #include <sys/param.h> 11 12 #include <sys/dtrace.h> 13 #include <cddl/dev/dtrace/dtrace_cddl.h> 14 15 #include "kinst.h" 16 17 DPCPU_DEFINE_STATIC(struct kinst_cpu_state, kinst_state); 18 19 #define _MATCH_REG(reg) \ 20 (offsetof(struct trapframe, tf_ ## reg) / sizeof(register_t)) 21 22 static int 23 kinst_regoff(struct trapframe *frame, int n) 24 { 25 switch (n) { 26 case 0: 27 /* There is no zero register in the trapframe structure. */ 28 return (-1); 29 case 1: 30 return (_MATCH_REG(ra)); 31 case 2: 32 return (_MATCH_REG(sp)); 33 case 3: 34 return (_MATCH_REG(gp)); 35 case 4: 36 return (_MATCH_REG(tp)); 37 case 5 ... 7: 38 return (_MATCH_REG(t[n - 5])); 39 case 8 ... 9: 40 return (_MATCH_REG(s[n - 8])); 41 case 10 ... 17: 42 return (_MATCH_REG(a[n - 10])); 43 case 18 ... 27: 44 return (_MATCH_REG(s[n - 18 + 2])); 45 case 28 ... 31: 46 return (_MATCH_REG(t[n - 28 + 3])); 47 default: 48 panic("%s: unhandled register index %d", __func__, n); 49 } 50 } 51 52 static int 53 kinst_c_regoff(struct trapframe *frame, int n) 54 { 55 switch (n) { 56 case 0 ... 1: 57 return (_MATCH_REG(s[n])); 58 case 2 ... 7: 59 return (_MATCH_REG(a[n - 2])); 60 default: 61 panic("%s: unhandled register index %d", __func__, n); 62 } 63 } 64 65 #undef _MATCH_REG 66 67 static int 68 kinst_emulate(struct trapframe *frame, const struct kinst_probe *kp) 69 { 70 kinst_patchval_t instr = kp->kp_savedval; 71 register_t prevpc; 72 uint64_t imm; 73 uint16_t off; 74 uint8_t funct; 75 76 if (kp->kp_md.instlen == INSN_SIZE) { 77 #define rs1_index ((instr & RS1_MASK) >> RS1_SHIFT) 78 #define rs2_index ((instr & RS2_MASK) >> RS2_SHIFT) 79 #define rd_index ((instr & RD_MASK) >> RD_SHIFT) 80 #define rs1 ((register_t *)frame)[kinst_regoff(frame, rs1_index)] 81 #define rs2 ((register_t *)frame)[kinst_regoff(frame, rs2_index)] 82 #define rd ((register_t *)frame)[kinst_regoff(frame, rd_index)] 83 #define rs1_lval (rs1_index != 0 ? rs1 : 0) 84 #define rs2_lval (rs2_index != 0 ? rs2 : 0) 85 switch (instr & 0x7f) { 86 case 0b1101111: /* jal */ 87 imm = 0; 88 imm |= ((instr >> 21) & 0x03ff) << 1; 89 imm |= ((instr >> 20) & 0x0001) << 11; 90 imm |= ((instr >> 12) & 0x00ff) << 12; 91 imm |= ((instr >> 31) & 0x0001) << 20; 92 if (imm & 0x0000000000100000) 93 imm |= 0xfffffffffff00000; 94 if (rd_index != 0) 95 rd = frame->tf_sepc + INSN_SIZE; 96 frame->tf_sepc += imm; 97 break; 98 case 0b1100111: /* jalr */ 99 prevpc = frame->tf_sepc; 100 imm = (instr & IMM_MASK) >> IMM_SHIFT; 101 if (imm & 0x0000000000000800) 102 imm |= 0xfffffffffffff000; 103 frame->tf_sepc = (rs1_lval + imm) & ~1; 104 if (rd_index != 0) 105 rd = prevpc + INSN_SIZE; 106 break; 107 case 0b1100011: /* branch */ 108 imm = 0; 109 imm |= ((instr >> 8) & 0x000f) << 1; 110 imm |= ((instr >> 25) & 0x003f) << 5; 111 imm |= ((instr >> 7) & 0x0001) << 11; 112 imm |= ((instr >> 31) & 0x0001) << 12; 113 if (imm & 0x0000000000001000) 114 imm |= 0xfffffffffffff000; 115 funct = (instr >> 12) & 0x07; 116 switch (funct) { 117 case 0b000: /* beq */ 118 if (rs1_lval == rs2_lval) 119 frame->tf_sepc += imm; 120 else 121 frame->tf_sepc += INSN_SIZE; 122 break; 123 case 0b001: /* bne */ 124 if (rs1_lval != rs2_lval) 125 frame->tf_sepc += imm; 126 else 127 frame->tf_sepc += INSN_SIZE; 128 break; 129 case 0b100: /* blt */ 130 if ((int64_t)rs1_lval < (int64_t)rs2_lval) 131 frame->tf_sepc += imm; 132 else 133 frame->tf_sepc += INSN_SIZE; 134 break; 135 case 0b110: /* bltu */ 136 if ((uint64_t)rs1_lval < (uint64_t)rs2_lval) 137 frame->tf_sepc += imm; 138 else 139 frame->tf_sepc += INSN_SIZE; 140 break; 141 case 0b101: /* bge */ 142 if ((int64_t)rs1_lval >= (int64_t)rs2_lval) 143 frame->tf_sepc += imm; 144 else 145 frame->tf_sepc += INSN_SIZE; 146 break; 147 case 0b111: /* bgeu */ 148 if ((uint64_t)rs1_lval >= (uint64_t)rs2_lval) 149 frame->tf_sepc += imm; 150 else 151 frame->tf_sepc += INSN_SIZE; 152 break; 153 } 154 break; 155 case 0b0010111: /* auipc */ 156 imm = instr & 0xfffff000; 157 rd = frame->tf_sepc + 158 (imm & 0x0000000080000000 ? 159 imm | 0xffffffff80000000 : imm); 160 frame->tf_sepc += INSN_SIZE; 161 break; 162 } 163 #undef rs1_lval 164 #undef rs2_lval 165 #undef rs1 166 #undef rs2 167 #undef rd 168 #undef rs1_index 169 #undef rs2_index 170 #undef rd_index 171 } else { 172 switch (instr & 0x03) { 173 #define rs1 \ 174 ((register_t *)frame)[kinst_c_regoff(frame, (instr >> 7) & 0x07)] 175 case 0b01: 176 funct = (instr >> 13) & 0x07; 177 switch (funct) { 178 case 0b101: /* c.j */ 179 off = (instr >> 2) & 0x07ff; 180 imm = 0; 181 imm |= ((off >> 1) & 0x07) << 1; 182 imm |= ((off >> 9) & 0x01) << 4; 183 imm |= ((off >> 0) & 0x01) << 5; 184 imm |= ((off >> 5) & 0x01) << 6; 185 imm |= ((off >> 4) & 0x01) << 7; 186 imm |= ((off >> 7) & 0x03) << 8; 187 imm |= ((off >> 6) & 0x01) << 10; 188 imm |= ((off >> 10) & 0x01) << 11; 189 if (imm & 0x0000000000000800) 190 imm |= 0xfffffffffffff000; 191 frame->tf_sepc += imm; 192 break; 193 case 0b110: /* c.beqz */ 194 case 0b111: /* c.bnez */ 195 imm = 0; 196 imm |= ((instr >> 3) & 0x03) << 1; 197 imm |= ((instr >> 10) & 0x03) << 3; 198 imm |= ((instr >> 2) & 0x01) << 5; 199 imm |= ((instr >> 5) & 0x03) << 6; 200 imm |= ((instr >> 12) & 0x01) << 8; 201 if (imm & 0x0000000000000100) 202 imm |= 0xffffffffffffff00; 203 if (funct == 0b110 && rs1 == 0) 204 frame->tf_sepc += imm; 205 else if (funct == 0b111 && rs1 != 0) 206 frame->tf_sepc += imm; 207 else 208 frame->tf_sepc += INSN_C_SIZE; 209 break; 210 } 211 break; 212 #undef rs1 213 #define rs1_index ((instr & RD_MASK) >> RD_SHIFT) 214 #define rs1 ((register_t *)frame)[kinst_regoff(frame, rs1_index)] 215 case 0b10: 216 funct = (instr >> 13) & 0x07; 217 if (funct == 0b100 && rs1_index != 0) { 218 /* c.jr/c.jalr */ 219 prevpc = frame->tf_sepc; 220 frame->tf_sepc = rs1; 221 if (((instr >> 12) & 0x01) != 0) 222 frame->tf_ra = prevpc + INSN_C_SIZE; 223 } 224 break; 225 #undef rs1 226 #undef rs1_index 227 } 228 } 229 230 return (MATCH_C_NOP); 231 } 232 233 static int 234 kinst_jump_next_instr(struct trapframe *frame, const struct kinst_probe *kp) 235 { 236 frame->tf_sepc = (register_t)((const uint8_t *)kp->kp_patchpoint + 237 kp->kp_md.instlen); 238 239 return (MATCH_C_NOP); 240 } 241 242 static void 243 kinst_trampoline_populate(struct kinst_probe *kp) 244 { 245 static uint16_t nop = MATCH_C_NOP; 246 static uint32_t ebreak = MATCH_EBREAK; 247 int ilen; 248 249 ilen = kp->kp_md.instlen; 250 kinst_memcpy(kp->kp_tramp, &kp->kp_savedval, ilen); 251 252 /* 253 * Since we cannot encode large displacements in a single instruction 254 * in order to encode a far-jump back to the next instruction, and we 255 * also cannot clobber a register inside the trampoline, we execute a 256 * breakpoint after the copied instruction. kinst_invop() is 257 * responsible for detecting this special case and performing the 258 * "jump" manually. 259 * 260 * Add a NOP after a compressed instruction for padding. 261 */ 262 if (ilen == INSN_C_SIZE) 263 kinst_memcpy(&kp->kp_tramp[ilen], &nop, INSN_C_SIZE); 264 265 kinst_memcpy(&kp->kp_tramp[INSN_SIZE], &ebreak, INSN_SIZE); 266 267 fence_i(); 268 } 269 270 /* 271 * There are two ways by which an instruction is traced: 272 * 273 * - By using the trampoline. 274 * - By emulating it in software (see kinst_emulate()). 275 * 276 * The trampoline is used for instructions that can be copied and executed 277 * as-is without additional modification. However, instructions that use 278 * PC-relative addressing have to be emulated, because RISC-V doesn't allow 279 * encoding of large displacements in a single instruction, and since we cannot 280 * clobber a register in order to encode the two-instruction sequence needed to 281 * create large displacements, we cannot use the trampoline at all. 282 * Fortunately, the instructions are simple enough to be emulated in just a few 283 * lines of code. 284 * 285 * The problem discussed above also means that, unlike amd64, we cannot encode 286 * a far-jump back from the trampoline to the next instruction. The mechanism 287 * employed to achieve this functionality, is to use a breakpoint instead of a 288 * jump after the copied instruction. This breakpoint is detected and handled 289 * by kinst_invop(), which performs the jump back to the next instruction 290 * manually (see kinst_jump_next_instr()). 291 */ 292 int 293 kinst_invop(uintptr_t addr, struct trapframe *frame, uintptr_t scratch) 294 { 295 solaris_cpu_t *cpu; 296 struct kinst_cpu_state *ks; 297 const struct kinst_probe *kp; 298 299 ks = DPCPU_PTR(kinst_state); 300 301 /* 302 * Detect if the breakpoint was triggered by the trampoline, and 303 * manually set the PC to the next instruction. 304 */ 305 if (ks->state == KINST_PROBE_FIRED && 306 addr == (uintptr_t)(ks->kp->kp_tramp + INSN_SIZE)) { 307 /* 308 * Restore interrupts if they were enabled prior to the first 309 * breakpoint. 310 */ 311 if ((ks->status & SSTATUS_SPIE) != 0) 312 frame->tf_sstatus |= SSTATUS_SPIE; 313 ks->state = KINST_PROBE_ARMED; 314 return (kinst_jump_next_instr(frame, ks->kp)); 315 } 316 317 LIST_FOREACH(kp, KINST_GETPROBE(addr), kp_hashnext) { 318 if ((uintptr_t)kp->kp_patchpoint == addr) 319 break; 320 } 321 if (kp == NULL) 322 return (0); 323 324 cpu = &solaris_cpu[curcpu]; 325 cpu->cpu_dtrace_caller = addr; 326 dtrace_probe(kp->kp_id, 0, 0, 0, 0, 0); 327 cpu->cpu_dtrace_caller = 0; 328 329 if (kp->kp_md.emulate) 330 return (kinst_emulate(frame, kp)); 331 332 ks->state = KINST_PROBE_FIRED; 333 ks->kp = kp; 334 335 /* 336 * Cache the current SSTATUS and clear interrupts for the 337 * duration of the double breakpoint. 338 */ 339 ks->status = frame->tf_sstatus; 340 frame->tf_sstatus &= ~SSTATUS_SPIE; 341 frame->tf_sepc = (register_t)kp->kp_tramp; 342 343 return (MATCH_C_NOP); 344 } 345 346 void 347 kinst_patch_tracepoint(struct kinst_probe *kp, kinst_patchval_t val) 348 { 349 switch (kp->kp_patchval) { 350 case KINST_C_PATCHVAL: 351 *(uint16_t *)kp->kp_patchpoint = (uint16_t)val; 352 fence_i(); 353 break; 354 case KINST_PATCHVAL: 355 *kp->kp_patchpoint = val; 356 fence_i(); 357 break; 358 } 359 } 360 361 static void 362 kinst_instr_dissect(struct kinst_probe *kp, int instrsize) 363 { 364 struct kinst_probe_md *kpmd; 365 kinst_patchval_t instr = kp->kp_savedval; 366 uint8_t funct; 367 368 kpmd = &kp->kp_md; 369 kpmd->instlen = instrsize; 370 kpmd->emulate = false; 371 372 /* 373 * The following instructions use PC-relative addressing and need to be 374 * emulated in software. 375 */ 376 if (kpmd->instlen == INSN_SIZE) { 377 switch (instr & 0x7f) { 378 case 0b1101111: /* jal */ 379 case 0b1100111: /* jalr */ 380 case 0b1100011: /* branch */ 381 case 0b0010111: /* auipc */ 382 kpmd->emulate = true; 383 break; 384 } 385 } else { 386 switch (instr & 0x03) { 387 case 0b01: 388 funct = (instr >> 13) & 0x07; 389 switch (funct) { 390 case 0b101: /* c.j */ 391 case 0b110: /* c.beqz */ 392 case 0b111: /* c.bnez */ 393 kpmd->emulate = true; 394 break; 395 } 396 break; 397 case 0b10: 398 funct = (instr >> 13) & 0x07; 399 if (funct == 0b100 && 400 ((instr >> 7) & 0x1f) != 0 && 401 ((instr >> 2) & 0x1f) == 0) 402 kpmd->emulate = true; /* c.jr/c.jalr */ 403 break; 404 } 405 } 406 407 if (!kpmd->emulate) 408 kinst_trampoline_populate(kp); 409 } 410 411 static bool 412 kinst_instr_system(kinst_patchval_t instr) 413 { 414 if (dtrace_match_opcode(instr, MATCH_C_EBREAK, MASK_C_EBREAK) || 415 (instr & 0x7f) == 0b1110011) 416 return (true); 417 418 return (false); 419 } 420 421 static bool 422 kinst_instr_lr(kinst_patchval_t instr) 423 { 424 if (dtrace_match_opcode(instr, MATCH_LR_W, MASK_LR_W) || 425 dtrace_match_opcode(instr, MATCH_LR_D, MASK_LR_D)) 426 return (true); 427 428 return (false); 429 } 430 431 static bool 432 kinst_instr_sc(kinst_patchval_t instr) 433 { 434 if (dtrace_match_opcode(instr, MATCH_SC_W, MASK_SC_W) || 435 dtrace_match_opcode(instr, MATCH_SC_D, MASK_SC_D)) 436 return (true); 437 438 return (false); 439 } 440 441 int 442 kinst_make_probe(linker_file_t lf, int symindx, linker_symval_t *symval, 443 void *opaque) 444 { 445 struct kinst_probe *kp; 446 dtrace_kinst_probedesc_t *pd; 447 const char *func; 448 kinst_patchval_t *insn, v; 449 uint8_t *instr, *limit; 450 int instrsize, n, off; 451 bool lrsc_block, store_found, ret_found; 452 453 pd = opaque; 454 func = symval->name; 455 456 if (kinst_excluded(func)) 457 return (0); 458 if (strcmp(func, pd->kpd_func) != 0) 459 return (0); 460 461 instr = (uint8_t *)(symval->value); 462 limit = (uint8_t *)(symval->value + symval->size); 463 if (instr >= limit) 464 return (0); 465 466 /* Check for the usual function prologue. */ 467 for (insn = (kinst_patchval_t *)instr; 468 insn < (kinst_patchval_t *)limit; insn++) { 469 if (dtrace_instr_sdsp(&insn) || dtrace_instr_c_sdsp(&insn)) 470 store_found = true; 471 else if (dtrace_instr_ret(&insn) || dtrace_instr_c_ret(&insn)) 472 ret_found = true; 473 if (store_found && ret_found) 474 break; 475 } 476 if (!store_found || !ret_found) 477 return (0); 478 479 n = 0; 480 lrsc_block = false; 481 while (instr < limit) { 482 instrsize = dtrace_instr_size(instr); 483 off = (int)(instr - (uint8_t *)symval->value); 484 485 /* 486 * Avoid undefined behavior (i.e simply casting `*instr` to 487 * `kinst_patchval_t`) in case the pointer is unaligned. 488 * memcpy() can safely operate on unaligned pointers. 489 */ 490 memcpy(&v, instr, sizeof(kinst_patchval_t)); 491 492 /* Skip SYSTEM instructions. */ 493 if (kinst_instr_system(v)) 494 goto cont; 495 496 /* 497 * Skip LR/SC blocks used to build atomic operations. If a 498 * breakpoint is placed in a LR/SC block, the loop becomes 499 * unconstrained. In this case we violate the operation and the 500 * loop might fail on some implementations (see section 8.3 of 501 * the RISC-V unprivileged spec). 502 */ 503 if (kinst_instr_lr(v)) 504 lrsc_block = true; 505 else if (kinst_instr_sc(v)) { 506 lrsc_block = false; 507 goto cont; 508 } 509 if (lrsc_block) 510 goto cont; 511 512 if (pd->kpd_off != -1 && off != pd->kpd_off) 513 goto cont; 514 515 /* 516 * Prevent separate dtrace(1) instances from creating copies of 517 * the same probe. 518 */ 519 LIST_FOREACH(kp, KINST_GETPROBE(instr), kp_hashnext) { 520 if (strcmp(kp->kp_func, func) == 0 && 521 strtol(kp->kp_name, NULL, 10) == off) 522 return (0); 523 } 524 if (++n > KINST_PROBETAB_MAX) { 525 KINST_LOG("probe list full: %d entries", n); 526 return (ENOMEM); 527 } 528 kp = malloc(sizeof(struct kinst_probe), M_KINST, 529 M_WAITOK | M_ZERO); 530 kp->kp_func = func; 531 snprintf(kp->kp_name, sizeof(kp->kp_name), "%d", off); 532 kp->kp_patchpoint = (kinst_patchval_t *)instr; 533 kp->kp_savedval = v; 534 if (instrsize == INSN_SIZE) 535 kp->kp_patchval = KINST_PATCHVAL; 536 else 537 kp->kp_patchval = KINST_C_PATCHVAL; 538 if ((kp->kp_tramp = kinst_trampoline_alloc(M_WAITOK)) == NULL) { 539 KINST_LOG("cannot allocate trampoline for %p", instr); 540 return (ENOMEM); 541 } 542 543 kinst_instr_dissect(kp, instrsize); 544 kinst_probe_create(kp, lf); 545 cont: 546 instr += instrsize; 547 } 548 if (lrsc_block) 549 KINST_LOG("warning: unterminated LR/SC block"); 550 551 return (0); 552 } 553 554 int 555 kinst_md_init(void) 556 { 557 struct kinst_cpu_state *ks; 558 int cpu; 559 560 CPU_FOREACH(cpu) { 561 ks = DPCPU_PTR(kinst_state); 562 ks->state = KINST_PROBE_ARMED; 563 } 564 565 return (0); 566 } 567 568 void 569 kinst_md_deinit(void) 570 { 571 } 572 573 /* 574 * Exclude machine-dependent functions that are not safe-to-trace. 575 */ 576 bool 577 kinst_md_excluded(const char *name) 578 { 579 if (strcmp(name, "cpu_exception_handler") == 0 || 580 strcmp(name, "cpu_exception_handler_supervisor") == 0 || 581 strcmp(name, "cpu_exception_handler_user") == 0 || 582 strcmp(name, "do_trap_supervisor") == 0 || 583 strcmp(name, "do_trap_user") == 0) 584 return (true); 585 586 return (false); 587 } 588