]>
Commit | Line | Data |
---|---|---|
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 2008 Sun Microsystems, Inc. All rights reserved. | |
24 | * Use is subject to license terms. | |
25 | */ | |
26 | ||
27 | #include <sys/fasttrap_isa.h> | |
28 | #include <sys/fasttrap_impl.h> | |
29 | #include <sys/dtrace.h> | |
30 | #include <sys/dtrace_impl.h> | |
31 | extern dtrace_id_t dtrace_probeid_error; | |
32 | ||
33 | #include "fasttrap_regset.h" | |
34 | ||
35 | #include <sys/dtrace_ptss.h> | |
36 | #include <kern/debug.h> | |
37 | ||
38 | #include <machine/pal_routines.h> | |
39 | ||
40 | /* Solaris proc_t is the struct. Darwin's proc_t is a pointer to it. */ | |
41 | #define proc_t struct proc /* Steer clear of the Darwin typedef for proc_t */ | |
42 | ||
43 | /* | |
44 | * Lossless User-Land Tracing on x86 | |
45 | * --------------------------------- | |
46 | * | |
47 | * The execution of most instructions is not dependent on the address; for | |
48 | * these instructions it is sufficient to copy them into the user process's | |
49 | * address space and execute them. To effectively single-step an instruction | |
50 | * in user-land, we copy out the following sequence of instructions to scratch | |
51 | * space in the user thread's ulwp_t structure. | |
52 | * | |
53 | * We then set the program counter (%eip or %rip) to point to this scratch | |
54 | * space. Once execution resumes, the original instruction is executed and | |
55 | * then control flow is redirected to what was originally the subsequent | |
56 | * instruction. If the kernel attemps to deliver a signal while single- | |
57 | * stepping, the signal is deferred and the program counter is moved into the | |
58 | * second sequence of instructions. The second sequence ends in a trap into | |
59 | * the kernel where the deferred signal is then properly handled and delivered. | |
60 | * | |
61 | * For instructions whose execute is position dependent, we perform simple | |
62 | * emulation. These instructions are limited to control transfer | |
63 | * instructions in 32-bit mode, but in 64-bit mode there's the added wrinkle | |
64 | * of %rip-relative addressing that means that almost any instruction can be | |
65 | * position dependent. For all the details on how we emulate generic | |
66 | * instructions included %rip-relative instructions, see the code in | |
67 | * fasttrap_pid_probe() below where we handle instructions of type | |
68 | * FASTTRAP_T_COMMON (under the header: Generic Instruction Tracing). | |
69 | */ | |
70 | ||
71 | #define FASTTRAP_MODRM_MOD(modrm) (((modrm) >> 6) & 0x3) | |
72 | #define FASTTRAP_MODRM_REG(modrm) (((modrm) >> 3) & 0x7) | |
73 | #define FASTTRAP_MODRM_RM(modrm) ((modrm) & 0x7) | |
74 | #define FASTTRAP_MODRM(mod, reg, rm) (((mod) << 6) | ((reg) << 3) | (rm)) | |
75 | ||
76 | #define FASTTRAP_SIB_SCALE(sib) (((sib) >> 6) & 0x3) | |
77 | #define FASTTRAP_SIB_INDEX(sib) (((sib) >> 3) & 0x7) | |
78 | #define FASTTRAP_SIB_BASE(sib) ((sib) & 0x7) | |
79 | ||
80 | #define FASTTRAP_REX_W(rex) (((rex) >> 3) & 1) | |
81 | #define FASTTRAP_REX_R(rex) (((rex) >> 2) & 1) | |
82 | #define FASTTRAP_REX_X(rex) (((rex) >> 1) & 1) | |
83 | #define FASTTRAP_REX_B(rex) ((rex) & 1) | |
84 | #define FASTTRAP_REX(w, r, x, b) \ | |
85 | (0x40 | ((w) << 3) | ((r) << 2) | ((x) << 1) | (b)) | |
86 | ||
87 | /* | |
88 | * Single-byte op-codes. | |
89 | */ | |
90 | #define FASTTRAP_PUSHL_EBP 0x55 | |
91 | ||
92 | #define FASTTRAP_JO 0x70 | |
93 | #define FASTTRAP_JNO 0x71 | |
94 | #define FASTTRAP_JB 0x72 | |
95 | #define FASTTRAP_JAE 0x73 | |
96 | #define FASTTRAP_JE 0x74 | |
97 | #define FASTTRAP_JNE 0x75 | |
98 | #define FASTTRAP_JBE 0x76 | |
99 | #define FASTTRAP_JA 0x77 | |
100 | #define FASTTRAP_JS 0x78 | |
101 | #define FASTTRAP_JNS 0x79 | |
102 | #define FASTTRAP_JP 0x7a | |
103 | #define FASTTRAP_JNP 0x7b | |
104 | #define FASTTRAP_JL 0x7c | |
105 | #define FASTTRAP_JGE 0x7d | |
106 | #define FASTTRAP_JLE 0x7e | |
107 | #define FASTTRAP_JG 0x7f | |
108 | ||
109 | #define FASTTRAP_NOP 0x90 | |
110 | ||
111 | #define FASTTRAP_MOV_EAX 0xb8 | |
112 | #define FASTTRAP_MOV_ECX 0xb9 | |
113 | ||
114 | #define FASTTRAP_RET16 0xc2 | |
115 | #define FASTTRAP_RET 0xc3 | |
116 | ||
117 | #define FASTTRAP_LOOPNZ 0xe0 | |
118 | #define FASTTRAP_LOOPZ 0xe1 | |
119 | #define FASTTRAP_LOOP 0xe2 | |
120 | #define FASTTRAP_JCXZ 0xe3 | |
121 | ||
122 | #define FASTTRAP_CALL 0xe8 | |
123 | #define FASTTRAP_JMP32 0xe9 | |
124 | #define FASTTRAP_JMP8 0xeb | |
125 | ||
126 | #define FASTTRAP_INT3 0xcc | |
127 | #define FASTTRAP_INT 0xcd | |
128 | #define T_DTRACE_RET 0x7f | |
129 | ||
130 | #define FASTTRAP_2_BYTE_OP 0x0f | |
131 | #define FASTTRAP_GROUP5_OP 0xff | |
132 | ||
133 | /* | |
134 | * Two-byte op-codes (second byte only). | |
135 | */ | |
136 | #define FASTTRAP_0F_JO 0x80 | |
137 | #define FASTTRAP_0F_JNO 0x81 | |
138 | #define FASTTRAP_0F_JB 0x82 | |
139 | #define FASTTRAP_0F_JAE 0x83 | |
140 | #define FASTTRAP_0F_JE 0x84 | |
141 | #define FASTTRAP_0F_JNE 0x85 | |
142 | #define FASTTRAP_0F_JBE 0x86 | |
143 | #define FASTTRAP_0F_JA 0x87 | |
144 | #define FASTTRAP_0F_JS 0x88 | |
145 | #define FASTTRAP_0F_JNS 0x89 | |
146 | #define FASTTRAP_0F_JP 0x8a | |
147 | #define FASTTRAP_0F_JNP 0x8b | |
148 | #define FASTTRAP_0F_JL 0x8c | |
149 | #define FASTTRAP_0F_JGE 0x8d | |
150 | #define FASTTRAP_0F_JLE 0x8e | |
151 | #define FASTTRAP_0F_JG 0x8f | |
152 | ||
153 | #define FASTTRAP_EFLAGS_OF 0x800 | |
154 | #define FASTTRAP_EFLAGS_DF 0x400 | |
155 | #define FASTTRAP_EFLAGS_SF 0x080 | |
156 | #define FASTTRAP_EFLAGS_ZF 0x040 | |
157 | #define FASTTRAP_EFLAGS_AF 0x010 | |
158 | #define FASTTRAP_EFLAGS_PF 0x004 | |
159 | #define FASTTRAP_EFLAGS_CF 0x001 | |
160 | ||
161 | /* | |
162 | * Instruction prefixes. | |
163 | */ | |
164 | #define FASTTRAP_PREFIX_OPERAND 0x66 | |
165 | #define FASTTRAP_PREFIX_ADDRESS 0x67 | |
166 | #define FASTTRAP_PREFIX_CS 0x2E | |
167 | #define FASTTRAP_PREFIX_DS 0x3E | |
168 | #define FASTTRAP_PREFIX_ES 0x26 | |
169 | #define FASTTRAP_PREFIX_FS 0x64 | |
170 | #define FASTTRAP_PREFIX_GS 0x65 | |
171 | #define FASTTRAP_PREFIX_SS 0x36 | |
172 | #define FASTTRAP_PREFIX_LOCK 0xF0 | |
173 | #define FASTTRAP_PREFIX_REP 0xF3 | |
174 | #define FASTTRAP_PREFIX_REPNE 0xF2 | |
175 | ||
176 | #define FASTTRAP_NOREG 0xff | |
177 | ||
178 | /* | |
179 | * Map between instruction register encodings and the kernel constants which | |
180 | * correspond to indicies into struct regs. | |
181 | */ | |
182 | ||
183 | /* | |
184 | * APPLE NOTE: We are cheating here. The regmap is used to decode which register | |
185 | * a given instruction is trying to reference. OS X does not have extended registers | |
186 | * for 32 bit apps, but the *order* is the same. So for 32 bit state, we will return: | |
187 | * | |
188 | * REG_RAX -> EAX | |
189 | * REG_RCX -> ECX | |
190 | * REG_RDX -> EDX | |
191 | * REG_RBX -> EBX | |
192 | * REG_RSP -> UESP | |
193 | * REG_RBP -> EBP | |
194 | * REG_RSI -> ESI | |
195 | * REG_RDI -> EDI | |
196 | * | |
197 | * The fasttrap_getreg function knows how to make the correct transformation. | |
198 | */ | |
199 | static const uint8_t regmap[16] = { | |
200 | REG_RAX, REG_RCX, REG_RDX, REG_RBX, REG_RSP, REG_RBP, REG_RSI, REG_RDI, | |
201 | REG_R8, REG_R9, REG_R10, REG_R11, REG_R12, REG_R13, REG_R14, REG_R15, | |
202 | }; | |
203 | ||
204 | static user_addr_t fasttrap_getreg(x86_saved_state_t *, uint_t); | |
205 | ||
206 | static uint64_t | |
207 | fasttrap_anarg(x86_saved_state_t *regs, int function_entry, int argno) | |
208 | { | |
209 | uint64_t value; | |
210 | int shift = function_entry ? 1 : 0; | |
211 | ||
212 | x86_saved_state64_t *regs64; | |
213 | x86_saved_state32_t *regs32; | |
214 | unsigned int p_model; | |
215 | ||
216 | if (is_saved_state64(regs)) { | |
217 | regs64 = saved_state64(regs); | |
218 | regs32 = NULL; | |
219 | p_model = DATAMODEL_LP64; | |
220 | } else { | |
221 | regs64 = NULL; | |
222 | regs32 = saved_state32(regs); | |
223 | p_model = DATAMODEL_ILP32; | |
224 | } | |
225 | ||
226 | if (p_model == DATAMODEL_LP64) { | |
227 | user_addr_t stack; | |
228 | ||
229 | /* | |
230 | * In 64-bit mode, the first six arguments are stored in | |
231 | * registers. | |
232 | */ | |
233 | if (argno < 6) | |
234 | return ((®s64->rdi)[argno]); | |
235 | ||
236 | stack = regs64->isf.rsp + sizeof(uint64_t) * (argno - 6 + shift); | |
237 | DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); | |
238 | value = dtrace_fuword64(stack); | |
239 | DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT | CPU_DTRACE_BADADDR); | |
240 | } else { | |
241 | uint32_t *stack = (uint32_t *)(uintptr_t)(regs32->uesp); | |
242 | DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT); | |
243 | value = dtrace_fuword32((user_addr_t)(unsigned long)&stack[argno + shift]); | |
244 | DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT | CPU_DTRACE_BADADDR); | |
245 | } | |
246 | ||
247 | return (value); | |
248 | } | |
249 | ||
250 | /*ARGSUSED*/ | |
251 | int | |
252 | fasttrap_tracepoint_init(proc_t *p, fasttrap_tracepoint_t *tp, user_addr_t pc, | |
253 | fasttrap_probe_type_t type) | |
254 | { | |
255 | #pragma unused(type) | |
256 | uint8_t instr[FASTTRAP_MAX_INSTR_SIZE + 10]; | |
257 | size_t len = FASTTRAP_MAX_INSTR_SIZE; | |
258 | size_t first = MIN(len, PAGE_SIZE - (pc & PAGE_MASK)); | |
259 | uint_t start = 0; | |
260 | size_t size; | |
261 | int rmindex; | |
262 | uint8_t seg, rex = 0; | |
263 | unsigned int p_model = (p->p_flag & P_LP64) ? DATAMODEL_LP64 : DATAMODEL_ILP32; | |
264 | ||
265 | /* | |
266 | * Read the instruction at the given address out of the process's | |
267 | * address space. We don't have to worry about a debugger | |
268 | * changing this instruction before we overwrite it with our trap | |
269 | * instruction since P_PR_LOCK is set. Since instructions can span | |
270 | * pages, we potentially read the instruction in two parts. If the | |
271 | * second part fails, we just zero out that part of the instruction. | |
272 | */ | |
273 | /* | |
274 | * APPLE NOTE: Of course, we do not have a P_PR_LOCK, so this is racey... | |
275 | */ | |
276 | if (uread(p, &instr[0], first, pc) != 0) | |
277 | return (-1); | |
278 | if (len > first && | |
279 | uread(p, &instr[first], len - first, pc + first) != 0) { | |
280 | bzero(&instr[first], len - first); | |
281 | len = first; | |
282 | } | |
283 | ||
284 | /* | |
285 | * If the disassembly fails, then we have a malformed instruction. | |
286 | */ | |
287 | if ((size = dtrace_instr_size_isa(instr, p_model, &rmindex)) <= 0) | |
288 | return (-1); | |
289 | ||
290 | /* | |
291 | * Make sure the disassembler isn't completely broken. | |
292 | */ | |
293 | ASSERT(-1 <= rmindex && rmindex < (int)size); | |
294 | ||
295 | /* | |
296 | * If the computed size is greater than the number of bytes read, | |
297 | * then it was a malformed instruction possibly because it fell on a | |
298 | * page boundary and the subsequent page was missing or because of | |
299 | * some malicious user. | |
300 | */ | |
301 | if (size > len) | |
302 | return (-1); | |
303 | ||
304 | tp->ftt_size = (uint8_t)size; | |
305 | tp->ftt_segment = FASTTRAP_SEG_NONE; | |
306 | ||
307 | /* | |
308 | * Find the start of the instruction's opcode by processing any | |
309 | * legacy prefixes. | |
310 | */ | |
311 | for (;;) { | |
312 | seg = 0; | |
313 | switch (instr[start]) { | |
314 | case FASTTRAP_PREFIX_SS: | |
315 | seg++; | |
316 | OS_FALLTHROUGH; | |
317 | case FASTTRAP_PREFIX_GS: | |
318 | seg++; | |
319 | OS_FALLTHROUGH; | |
320 | case FASTTRAP_PREFIX_FS: | |
321 | seg++; | |
322 | OS_FALLTHROUGH; | |
323 | case FASTTRAP_PREFIX_ES: | |
324 | seg++; | |
325 | OS_FALLTHROUGH; | |
326 | case FASTTRAP_PREFIX_DS: | |
327 | seg++; | |
328 | OS_FALLTHROUGH; | |
329 | case FASTTRAP_PREFIX_CS: | |
330 | seg++; | |
331 | OS_FALLTHROUGH; | |
332 | case FASTTRAP_PREFIX_OPERAND: | |
333 | case FASTTRAP_PREFIX_ADDRESS: | |
334 | case FASTTRAP_PREFIX_LOCK: | |
335 | case FASTTRAP_PREFIX_REP: | |
336 | case FASTTRAP_PREFIX_REPNE: | |
337 | if (seg != 0) { | |
338 | /* | |
339 | * It's illegal for an instruction to specify | |
340 | * two segment prefixes -- give up on this | |
341 | * illegal instruction. | |
342 | */ | |
343 | if (tp->ftt_segment != FASTTRAP_SEG_NONE) | |
344 | return (-1); | |
345 | ||
346 | tp->ftt_segment = seg; | |
347 | } | |
348 | start++; | |
349 | continue; | |
350 | } | |
351 | break; | |
352 | } | |
353 | ||
354 | /* | |
355 | * Identify the REX prefix on 64-bit processes. | |
356 | */ | |
357 | if (p_model == DATAMODEL_LP64 && (instr[start] & 0xf0) == 0x40) | |
358 | rex = instr[start++]; | |
359 | ||
360 | /* | |
361 | * Now that we're pretty sure that the instruction is okay, copy the | |
362 | * valid part to the tracepoint. | |
363 | */ | |
364 | bcopy(instr, tp->ftt_instr, FASTTRAP_MAX_INSTR_SIZE); | |
365 | ||
366 | tp->ftt_type = FASTTRAP_T_COMMON; | |
367 | if (instr[start] == FASTTRAP_2_BYTE_OP) { | |
368 | switch (instr[start + 1]) { | |
369 | case FASTTRAP_0F_JO: | |
370 | case FASTTRAP_0F_JNO: | |
371 | case FASTTRAP_0F_JB: | |
372 | case FASTTRAP_0F_JAE: | |
373 | case FASTTRAP_0F_JE: | |
374 | case FASTTRAP_0F_JNE: | |
375 | case FASTTRAP_0F_JBE: | |
376 | case FASTTRAP_0F_JA: | |
377 | case FASTTRAP_0F_JS: | |
378 | case FASTTRAP_0F_JNS: | |
379 | case FASTTRAP_0F_JP: | |
380 | case FASTTRAP_0F_JNP: | |
381 | case FASTTRAP_0F_JL: | |
382 | case FASTTRAP_0F_JGE: | |
383 | case FASTTRAP_0F_JLE: | |
384 | case FASTTRAP_0F_JG: | |
385 | tp->ftt_type = FASTTRAP_T_JCC; | |
386 | tp->ftt_code = (instr[start + 1] & 0x0f) | FASTTRAP_JO; | |
387 | tp->ftt_dest = pc + tp->ftt_size + | |
388 | /* LINTED - alignment */ | |
389 | *(int32_t *)&instr[start + 2]; | |
390 | break; | |
391 | } | |
392 | } else if (instr[start] == FASTTRAP_GROUP5_OP) { | |
393 | uint_t mod = FASTTRAP_MODRM_MOD(instr[start + 1]); | |
394 | uint_t reg = FASTTRAP_MODRM_REG(instr[start + 1]); | |
395 | uint_t rm = FASTTRAP_MODRM_RM(instr[start + 1]); | |
396 | ||
397 | if (reg == 2 || reg == 4) { | |
398 | uint_t i, sz; | |
399 | ||
400 | if (reg == 2) | |
401 | tp->ftt_type = FASTTRAP_T_CALL; | |
402 | else | |
403 | tp->ftt_type = FASTTRAP_T_JMP; | |
404 | ||
405 | if (mod == 3) | |
406 | tp->ftt_code = 2; | |
407 | else | |
408 | tp->ftt_code = 1; | |
409 | ||
410 | ASSERT(p_model == DATAMODEL_LP64 || rex == 0); | |
411 | ||
412 | /* | |
413 | * See AMD x86-64 Architecture Programmer's Manual | |
414 | * Volume 3, Section 1.2.7, Table 1-12, and | |
415 | * Appendix A.3.1, Table A-15. | |
416 | */ | |
417 | if (mod != 3 && rm == 4) { | |
418 | uint8_t sib = instr[start + 2]; | |
419 | uint_t index = FASTTRAP_SIB_INDEX(sib); | |
420 | uint_t base = FASTTRAP_SIB_BASE(sib); | |
421 | ||
422 | tp->ftt_scale = FASTTRAP_SIB_SCALE(sib); | |
423 | ||
424 | tp->ftt_index = (index == 4) ? | |
425 | FASTTRAP_NOREG : | |
426 | regmap[index | (FASTTRAP_REX_X(rex) << 3)]; | |
427 | tp->ftt_base = (mod == 0 && base == 5) ? | |
428 | FASTTRAP_NOREG : | |
429 | regmap[base | (FASTTRAP_REX_B(rex) << 3)]; | |
430 | ||
431 | i = 3; | |
432 | sz = mod == 1 ? 1 : 4; | |
433 | } else { | |
434 | /* | |
435 | * In 64-bit mode, mod == 0 and r/m == 5 | |
436 | * denotes %rip-relative addressing; in 32-bit | |
437 | * mode, the base register isn't used. In both | |
438 | * modes, there is a 32-bit operand. | |
439 | */ | |
440 | if (mod == 0 && rm == 5) { | |
441 | if (p_model == DATAMODEL_LP64) | |
442 | tp->ftt_base = REG_RIP; | |
443 | else | |
444 | tp->ftt_base = FASTTRAP_NOREG; | |
445 | sz = 4; | |
446 | } else { | |
447 | uint8_t base = rm | | |
448 | (FASTTRAP_REX_B(rex) << 3); | |
449 | ||
450 | tp->ftt_base = regmap[base]; | |
451 | sz = mod == 1 ? 1 : mod == 2 ? 4 : 0; | |
452 | } | |
453 | tp->ftt_index = FASTTRAP_NOREG; | |
454 | i = 2; | |
455 | } | |
456 | ||
457 | if (sz == 1) { | |
458 | tp->ftt_dest = *(int8_t *)&instr[start + i]; | |
459 | } else if (sz == 4) { | |
460 | /* LINTED - alignment */ | |
461 | tp->ftt_dest = *(int32_t *)&instr[start + i]; | |
462 | } else { | |
463 | tp->ftt_dest = 0; | |
464 | } | |
465 | } | |
466 | } else { | |
467 | switch (instr[start]) { | |
468 | case FASTTRAP_RET: | |
469 | tp->ftt_type = FASTTRAP_T_RET; | |
470 | break; | |
471 | ||
472 | case FASTTRAP_RET16: | |
473 | tp->ftt_type = FASTTRAP_T_RET16; | |
474 | /* LINTED - alignment */ | |
475 | tp->ftt_dest = *(uint16_t *)&instr[start + 1]; | |
476 | break; | |
477 | ||
478 | case FASTTRAP_JO: | |
479 | case FASTTRAP_JNO: | |
480 | case FASTTRAP_JB: | |
481 | case FASTTRAP_JAE: | |
482 | case FASTTRAP_JE: | |
483 | case FASTTRAP_JNE: | |
484 | case FASTTRAP_JBE: | |
485 | case FASTTRAP_JA: | |
486 | case FASTTRAP_JS: | |
487 | case FASTTRAP_JNS: | |
488 | case FASTTRAP_JP: | |
489 | case FASTTRAP_JNP: | |
490 | case FASTTRAP_JL: | |
491 | case FASTTRAP_JGE: | |
492 | case FASTTRAP_JLE: | |
493 | case FASTTRAP_JG: | |
494 | tp->ftt_type = FASTTRAP_T_JCC; | |
495 | tp->ftt_code = instr[start]; | |
496 | tp->ftt_dest = pc + tp->ftt_size + | |
497 | (int8_t)instr[start + 1]; | |
498 | break; | |
499 | ||
500 | case FASTTRAP_LOOPNZ: | |
501 | case FASTTRAP_LOOPZ: | |
502 | case FASTTRAP_LOOP: | |
503 | tp->ftt_type = FASTTRAP_T_LOOP; | |
504 | tp->ftt_code = instr[start]; | |
505 | tp->ftt_dest = pc + tp->ftt_size + | |
506 | (int8_t)instr[start + 1]; | |
507 | break; | |
508 | ||
509 | case FASTTRAP_JCXZ: | |
510 | tp->ftt_type = FASTTRAP_T_JCXZ; | |
511 | tp->ftt_dest = pc + tp->ftt_size + | |
512 | (int8_t)instr[start + 1]; | |
513 | break; | |
514 | ||
515 | case FASTTRAP_CALL: | |
516 | tp->ftt_type = FASTTRAP_T_CALL; | |
517 | tp->ftt_dest = pc + tp->ftt_size + | |
518 | /* LINTED - alignment */ | |
519 | *(int32_t *)&instr[start + 1]; | |
520 | tp->ftt_code = 0; | |
521 | break; | |
522 | ||
523 | case FASTTRAP_JMP32: | |
524 | tp->ftt_type = FASTTRAP_T_JMP; | |
525 | tp->ftt_dest = pc + tp->ftt_size + | |
526 | /* LINTED - alignment */ | |
527 | *(int32_t *)&instr[start + 1]; | |
528 | break; | |
529 | case FASTTRAP_JMP8: | |
530 | tp->ftt_type = FASTTRAP_T_JMP; | |
531 | tp->ftt_dest = pc + tp->ftt_size + | |
532 | (int8_t)instr[start + 1]; | |
533 | break; | |
534 | ||
535 | case FASTTRAP_PUSHL_EBP: | |
536 | if (start == 0) | |
537 | tp->ftt_type = FASTTRAP_T_PUSHL_EBP; | |
538 | break; | |
539 | ||
540 | case FASTTRAP_NOP: | |
541 | ASSERT(p_model == DATAMODEL_LP64 || rex == 0); | |
542 | ||
543 | /* | |
544 | * On sol64 we have to be careful not to confuse a nop | |
545 | * (actually xchgl %eax, %eax) with an instruction using | |
546 | * the same opcode, but that does something different | |
547 | * (e.g. xchgl %r8d, %eax or xcghq %r8, %rax). | |
548 | */ | |
549 | if (FASTTRAP_REX_B(rex) == 0) | |
550 | tp->ftt_type = FASTTRAP_T_NOP; | |
551 | break; | |
552 | ||
553 | case FASTTRAP_INT3: | |
554 | /* | |
555 | * The pid provider shares the int3 trap with debugger | |
556 | * breakpoints so we can't instrument them. | |
557 | */ | |
558 | ASSERT(instr[start] == FASTTRAP_INSTR); | |
559 | return (-1); | |
560 | ||
561 | case FASTTRAP_INT: | |
562 | /* | |
563 | * Interrupts seem like they could be traced with | |
564 | * no negative implications, but it's possible that | |
565 | * a thread could be redirected by the trap handling | |
566 | * code which would eventually return to the | |
567 | * instruction after the interrupt. If the interrupt | |
568 | * were in our scratch space, the subsequent | |
569 | * instruction might be overwritten before we return. | |
570 | * Accordingly we refuse to instrument any interrupt. | |
571 | */ | |
572 | return (-1); | |
573 | } | |
574 | } | |
575 | ||
576 | if (p_model == DATAMODEL_LP64 && tp->ftt_type == FASTTRAP_T_COMMON) { | |
577 | /* | |
578 | * If the process is 64-bit and the instruction type is still | |
579 | * FASTTRAP_T_COMMON -- meaning we're going to copy it out an | |
580 | * execute it -- we need to watch for %rip-relative | |
581 | * addressing mode. See the portion of fasttrap_pid_probe() | |
582 | * below where we handle tracepoints with type | |
583 | * FASTTRAP_T_COMMON for how we emulate instructions that | |
584 | * employ %rip-relative addressing. | |
585 | */ | |
586 | if (rmindex != -1) { | |
587 | uint_t mod = FASTTRAP_MODRM_MOD(instr[rmindex]); | |
588 | uint_t reg = FASTTRAP_MODRM_REG(instr[rmindex]); | |
589 | uint_t rm = FASTTRAP_MODRM_RM(instr[rmindex]); | |
590 | ||
591 | ASSERT(rmindex > (int)start); | |
592 | ||
593 | if (mod == 0 && rm == 5) { | |
594 | /* | |
595 | * We need to be sure to avoid other | |
596 | * registers used by this instruction. While | |
597 | * the reg field may determine the op code | |
598 | * rather than denoting a register, assuming | |
599 | * that it denotes a register is always safe. | |
600 | * We leave the REX field intact and use | |
601 | * whatever value's there for simplicity. | |
602 | */ | |
603 | if (reg != 0) { | |
604 | tp->ftt_ripmode = FASTTRAP_RIP_1 | | |
605 | (FASTTRAP_RIP_X * | |
606 | FASTTRAP_REX_B(rex)); | |
607 | rm = 0; | |
608 | } else { | |
609 | tp->ftt_ripmode = FASTTRAP_RIP_2 | | |
610 | (FASTTRAP_RIP_X * | |
611 | FASTTRAP_REX_B(rex)); | |
612 | rm = 1; | |
613 | } | |
614 | ||
615 | tp->ftt_modrm = tp->ftt_instr[rmindex]; | |
616 | tp->ftt_instr[rmindex] = | |
617 | FASTTRAP_MODRM(2, reg, rm); | |
618 | } | |
619 | } | |
620 | } | |
621 | ||
622 | return (0); | |
623 | } | |
624 | ||
625 | int | |
626 | fasttrap_tracepoint_install(proc_t *p, fasttrap_tracepoint_t *tp) | |
627 | { | |
628 | fasttrap_instr_t instr = FASTTRAP_INSTR; | |
629 | ||
630 | if (uwrite(p, &instr, 1, tp->ftt_pc) != 0) | |
631 | return (-1); | |
632 | ||
633 | tp->ftt_installed = 1; | |
634 | ||
635 | return (0); | |
636 | } | |
637 | ||
638 | int | |
639 | fasttrap_tracepoint_remove(proc_t *p, fasttrap_tracepoint_t *tp) | |
640 | { | |
641 | uint8_t instr; | |
642 | ||
643 | /* | |
644 | * Distinguish between read or write failures and a changed | |
645 | * instruction. | |
646 | */ | |
647 | if (uread(p, &instr, 1, tp->ftt_pc) != 0) | |
648 | goto end; | |
649 | if (instr != FASTTRAP_INSTR) | |
650 | goto end; | |
651 | if (uwrite(p, &tp->ftt_instr[0], 1, tp->ftt_pc) != 0) | |
652 | return (-1); | |
653 | end: | |
654 | tp->ftt_installed = 0; | |
655 | ||
656 | return (0); | |
657 | } | |
658 | ||
659 | static void | |
660 | fasttrap_return_common(x86_saved_state_t *regs, user_addr_t pc, pid_t pid, | |
661 | user_addr_t new_pc) | |
662 | { | |
663 | x86_saved_state64_t *regs64; | |
664 | x86_saved_state32_t *regs32; | |
665 | unsigned int p_model; | |
666 | int retire_tp = 1; | |
667 | ||
668 | dtrace_icookie_t cookie; | |
669 | ||
670 | if (is_saved_state64(regs)) { | |
671 | regs64 = saved_state64(regs); | |
672 | regs32 = NULL; | |
673 | p_model = DATAMODEL_LP64; | |
674 | } else { | |
675 | regs64 = NULL; | |
676 | regs32 = saved_state32(regs); | |
677 | p_model = DATAMODEL_ILP32; | |
678 | } | |
679 | ||
680 | fasttrap_tracepoint_t *tp; | |
681 | fasttrap_bucket_t *bucket; | |
682 | fasttrap_id_t *id; | |
683 | lck_mtx_t *pid_mtx; | |
684 | ||
685 | pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock; | |
686 | lck_mtx_lock(pid_mtx); | |
687 | bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)]; | |
688 | ||
689 | for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) { | |
690 | if (pid == tp->ftt_pid && pc == tp->ftt_pc && | |
691 | tp->ftt_proc->ftpc_acount != 0) | |
692 | break; | |
693 | } | |
694 | ||
695 | /* | |
696 | * Don't sweat it if we can't find the tracepoint again; unlike | |
697 | * when we're in fasttrap_pid_probe(), finding the tracepoint here | |
698 | * is not essential to the correct execution of the process. | |
699 | */ | |
700 | if (tp == NULL) { | |
701 | lck_mtx_unlock(pid_mtx); | |
702 | return; | |
703 | } | |
704 | ||
705 | for (id = tp->ftt_retids; id != NULL; id = id->fti_next) { | |
706 | fasttrap_probe_t *probe = id->fti_probe; | |
707 | /* | |
708 | * If there's a branch that could act as a return site, we | |
709 | * need to trace it, and check here if the program counter is | |
710 | * external to the function. | |
711 | */ | |
712 | if (tp->ftt_type != FASTTRAP_T_RET && | |
713 | tp->ftt_type != FASTTRAP_T_RET16 && | |
714 | new_pc - probe->ftp_faddr < probe->ftp_fsize) | |
715 | continue; | |
716 | ||
717 | if (probe->ftp_prov->ftp_provider_type == DTFTP_PROVIDER_ONESHOT) { | |
718 | if (os_atomic_xchg(&probe->ftp_triggered, 1, relaxed)) { | |
719 | /* already triggered */ | |
720 | continue; | |
721 | } | |
722 | } | |
723 | /* | |
724 | * If we have at least one probe associated that | |
725 | * is not a oneshot probe, don't remove the | |
726 | * tracepoint | |
727 | */ | |
728 | else { | |
729 | retire_tp = 0; | |
730 | } | |
731 | /* | |
732 | * Provide a hint to the stack trace functions to add the | |
733 | * following pc to the top of the stack since it's missing | |
734 | * on a return probe yet highly desirable for consistency. | |
735 | */ | |
736 | cookie = dtrace_interrupt_disable(); | |
737 | cpu_core[CPU->cpu_id].cpuc_missing_tos = pc; | |
738 | if (ISSET(current_proc()->p_lflag, P_LNOATTACH)) { | |
739 | dtrace_probe(dtrace_probeid_error, 0 /* state */, probe->ftp_id, | |
740 | 1 /* ndx */, -1 /* offset */, DTRACEFLT_UPRIV); | |
741 | } else if (p_model == DATAMODEL_LP64) { | |
742 | dtrace_probe(probe->ftp_id, | |
743 | pc - id->fti_probe->ftp_faddr, | |
744 | regs64->rax, regs64->rdx, 0, 0); | |
745 | } else { | |
746 | dtrace_probe(probe->ftp_id, | |
747 | pc - id->fti_probe->ftp_faddr, | |
748 | regs32->eax, regs32->edx, 0, 0); | |
749 | } | |
750 | /* remove the hint */ | |
751 | cpu_core[CPU->cpu_id].cpuc_missing_tos = 0; | |
752 | dtrace_interrupt_enable(cookie); | |
753 | } | |
754 | ||
755 | lck_mtx_unlock(pid_mtx); | |
756 | } | |
757 | ||
758 | static void | |
759 | fasttrap_sigsegv(proc_t *p, uthread_t t, user_addr_t addr) | |
760 | { | |
761 | proc_lock(p); | |
762 | ||
763 | /* Set fault address and mark signal */ | |
764 | t->uu_code = addr; | |
765 | t->uu_siglist |= sigmask(SIGSEGV); | |
766 | ||
767 | /* | |
768 | * XXX These two line may be redundant; if not, then we need | |
769 | * XXX to potentially set the data address in the machine | |
770 | * XXX specific thread state structure to indicate the address. | |
771 | */ | |
772 | t->uu_exception = KERN_INVALID_ADDRESS; /* SIGSEGV */ | |
773 | t->uu_subcode = 0; /* XXX pad */ | |
774 | ||
775 | proc_unlock(p); | |
776 | ||
777 | /* raise signal */ | |
778 | signal_setast(t->uu_context.vc_thread); | |
779 | } | |
780 | ||
781 | static void | |
782 | fasttrap_usdt_args64(fasttrap_probe_t *probe, x86_saved_state64_t *regs64, int argc, | |
783 | uint64_t *argv) | |
784 | { | |
785 | int i, x, cap = MIN(argc, probe->ftp_nargs); | |
786 | user_addr_t stack = (user_addr_t)regs64->isf.rsp; | |
787 | ||
788 | for (i = 0; i < cap; i++) { | |
789 | x = probe->ftp_argmap[i]; | |
790 | ||
791 | if (x < 6) { | |
792 | /* FIXME! This may be broken, needs testing */ | |
793 | argv[i] = (®s64->rdi)[x]; | |
794 | } else { | |
795 | fasttrap_fuword64_noerr(stack + (x * sizeof(uint64_t)), &argv[i]); | |
796 | } | |
797 | } | |
798 | ||
799 | for (; i < argc; i++) { | |
800 | argv[i] = 0; | |
801 | } | |
802 | } | |
803 | ||
804 | static void | |
805 | fasttrap_usdt_args32(fasttrap_probe_t *probe, x86_saved_state32_t *regs32, int argc, | |
806 | uint32_t *argv) | |
807 | { | |
808 | int i, x, cap = MIN(argc, probe->ftp_nargs); | |
809 | uint32_t *stack = (uint32_t *)(uintptr_t)(regs32->uesp); | |
810 | ||
811 | for (i = 0; i < cap; i++) { | |
812 | x = probe->ftp_argmap[i]; | |
813 | ||
814 | fasttrap_fuword32_noerr((user_addr_t)(unsigned long)&stack[x], &argv[i]); | |
815 | } | |
816 | ||
817 | for (; i < argc; i++) { | |
818 | argv[i] = 0; | |
819 | } | |
820 | } | |
821 | ||
822 | /* | |
823 | * FIXME! | |
824 | */ | |
825 | static int | |
826 | fasttrap_do_seg(fasttrap_tracepoint_t *tp, x86_saved_state_t *rp, user_addr_t *addr) // 64 bit | |
827 | { | |
828 | #pragma unused(tp, rp, addr) | |
829 | printf("fasttrap_do_seg() called while unimplemented.\n"); | |
830 | #if 0 | |
831 | proc_t *p = curproc; | |
832 | user_desc_t *desc; | |
833 | uint16_t sel, ndx, type; | |
834 | uintptr_t limit; | |
835 | ||
836 | switch (tp->ftt_segment) { | |
837 | case FASTTRAP_SEG_CS: | |
838 | sel = rp->r_cs; | |
839 | break; | |
840 | case FASTTRAP_SEG_DS: | |
841 | sel = rp->r_ds; | |
842 | break; | |
843 | case FASTTRAP_SEG_ES: | |
844 | sel = rp->r_es; | |
845 | break; | |
846 | case FASTTRAP_SEG_FS: | |
847 | sel = rp->r_fs; | |
848 | break; | |
849 | case FASTTRAP_SEG_GS: | |
850 | sel = rp->r_gs; | |
851 | break; | |
852 | case FASTTRAP_SEG_SS: | |
853 | sel = rp->r_ss; | |
854 | break; | |
855 | } | |
856 | ||
857 | /* | |
858 | * Make sure the given segment register specifies a user priority | |
859 | * selector rather than a kernel selector. | |
860 | */ | |
861 | if (!SELISUPL(sel)) | |
862 | return (-1); | |
863 | ||
864 | ndx = SELTOIDX(sel); | |
865 | ||
866 | /* | |
867 | * Check the bounds and grab the descriptor out of the specified | |
868 | * descriptor table. | |
869 | */ | |
870 | if (SELISLDT(sel)) { | |
871 | if (ndx > p->p_ldtlimit) | |
872 | return (-1); | |
873 | ||
874 | desc = p->p_ldt + ndx; | |
875 | ||
876 | } else { | |
877 | if (ndx >= NGDT) | |
878 | return (-1); | |
879 | ||
880 | desc = cpu_get_gdt() + ndx; | |
881 | } | |
882 | ||
883 | /* | |
884 | * The descriptor must have user privilege level and it must be | |
885 | * present in memory. | |
886 | */ | |
887 | if (desc->usd_dpl != SEL_UPL || desc->usd_p != 1) | |
888 | return (-1); | |
889 | ||
890 | type = desc->usd_type; | |
891 | ||
892 | /* | |
893 | * If the S bit in the type field is not set, this descriptor can | |
894 | * only be used in system context. | |
895 | */ | |
896 | if ((type & 0x10) != 0x10) | |
897 | return (-1); | |
898 | ||
899 | limit = USEGD_GETLIMIT(desc) * (desc->usd_gran ? PAGESIZE : 1); | |
900 | ||
901 | if (tp->ftt_segment == FASTTRAP_SEG_CS) { | |
902 | /* | |
903 | * The code/data bit and readable bit must both be set. | |
904 | */ | |
905 | if ((type & 0xa) != 0xa) | |
906 | return (-1); | |
907 | ||
908 | if (*addr > limit) | |
909 | return (-1); | |
910 | } else { | |
911 | /* | |
912 | * The code/data bit must be clear. | |
913 | */ | |
914 | if ((type & 0x8) != 0) | |
915 | return (-1); | |
916 | ||
917 | /* | |
918 | * If the expand-down bit is clear, we just check the limit as | |
919 | * it would naturally be applied. Otherwise, we need to check | |
920 | * that the address is the range [limit + 1 .. 0xffff] or | |
921 | * [limit + 1 ... 0xffffffff] depending on if the default | |
922 | * operand size bit is set. | |
923 | */ | |
924 | if ((type & 0x4) == 0) { | |
925 | if (*addr > limit) | |
926 | return (-1); | |
927 | } else if (desc->usd_def32) { | |
928 | if (*addr < limit + 1 || 0xffff < *addr) | |
929 | return (-1); | |
930 | } else { | |
931 | if (*addr < limit + 1 || 0xffffffff < *addr) | |
932 | return (-1); | |
933 | } | |
934 | } | |
935 | ||
936 | *addr += USEGD_GETBASE(desc); | |
937 | #endif /* 0 */ | |
938 | return (0); | |
939 | } | |
940 | ||
941 | /* | |
942 | * Due to variances between Solaris and xnu, I have split this into a 32 bit and 64 bit | |
943 | * code path. It still takes an x86_saved_state_t* argument, because it must sometimes | |
944 | * call other methods that require a x86_saved_state_t. | |
945 | * | |
946 | * NOTE!!!! | |
947 | * | |
948 | * Any changes made to this method must be echo'd in fasttrap_pid_probe64! | |
949 | * | |
950 | */ | |
951 | static int | |
952 | fasttrap_pid_probe32(x86_saved_state_t *regs) | |
953 | { | |
954 | ASSERT(is_saved_state32(regs)); | |
955 | ||
956 | x86_saved_state32_t *regs32 = saved_state32(regs); | |
957 | user_addr_t pc = regs32->eip - 1; | |
958 | proc_t *p = current_proc(); | |
959 | user_addr_t new_pc = 0; | |
960 | fasttrap_bucket_t *bucket; | |
961 | lck_mtx_t *pid_mtx; | |
962 | fasttrap_tracepoint_t *tp, tp_local; | |
963 | pid_t pid; | |
964 | dtrace_icookie_t cookie; | |
965 | uint_t is_enabled = 0, retire_tp = 1; | |
966 | ||
967 | uthread_t uthread = (uthread_t)get_bsdthread_info(current_thread()); | |
968 | ||
969 | /* | |
970 | * It's possible that a user (in a veritable orgy of bad planning) | |
971 | * could redirect this thread's flow of control before it reached the | |
972 | * return probe fasttrap. In this case we need to kill the process | |
973 | * since it's in a unrecoverable state. | |
974 | */ | |
975 | if (uthread->t_dtrace_step) { | |
976 | ASSERT(uthread->t_dtrace_on); | |
977 | fasttrap_sigtrap(p, uthread, pc); | |
978 | return (0); | |
979 | } | |
980 | ||
981 | /* | |
982 | * Clear all user tracing flags. | |
983 | */ | |
984 | uthread->t_dtrace_ft = 0; | |
985 | uthread->t_dtrace_pc = 0; | |
986 | uthread->t_dtrace_npc = 0; | |
987 | uthread->t_dtrace_scrpc = 0; | |
988 | uthread->t_dtrace_astpc = 0; | |
989 | ||
990 | /* | |
991 | * Treat a child created by a call to vfork(2) as if it were its | |
992 | * parent. We know that there's only one thread of control in such a | |
993 | * process: this one. | |
994 | */ | |
995 | if (p->p_lflag & P_LINVFORK) { | |
996 | proc_list_lock(); | |
997 | while (p->p_lflag & P_LINVFORK) | |
998 | p = p->p_pptr; | |
999 | proc_list_unlock(); | |
1000 | } | |
1001 | ||
1002 | pid = p->p_pid; | |
1003 | pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock; | |
1004 | lck_mtx_lock(pid_mtx); | |
1005 | bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)]; | |
1006 | ||
1007 | /* | |
1008 | * Lookup the tracepoint that the process just hit. | |
1009 | */ | |
1010 | for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) { | |
1011 | if (pid == tp->ftt_pid && pc == tp->ftt_pc && | |
1012 | tp->ftt_proc->ftpc_acount != 0) | |
1013 | break; | |
1014 | } | |
1015 | ||
1016 | /* | |
1017 | * If we couldn't find a matching tracepoint, either a tracepoint has | |
1018 | * been inserted without using the pid<pid> ioctl interface (see | |
1019 | * fasttrap_ioctl), or somehow we have mislaid this tracepoint. | |
1020 | */ | |
1021 | if (tp == NULL) { | |
1022 | lck_mtx_unlock(pid_mtx); | |
1023 | return (-1); | |
1024 | } | |
1025 | ||
1026 | /* | |
1027 | * Set the program counter to the address of the traced instruction | |
1028 | * so that it looks right in ustack() output. | |
1029 | */ | |
1030 | regs32->eip = pc; | |
1031 | ||
1032 | if (tp->ftt_ids != NULL) { | |
1033 | fasttrap_id_t *id; | |
1034 | ||
1035 | uint32_t s0, s1, s2, s3, s4, s5; | |
1036 | uint32_t *stack = (uint32_t *)(uintptr_t)(regs32->uesp); | |
1037 | ||
1038 | /* | |
1039 | * In 32-bit mode, all arguments are passed on the | |
1040 | * stack. If this is a function entry probe, we need | |
1041 | * to skip the first entry on the stack as it | |
1042 | * represents the return address rather than a | |
1043 | * parameter to the function. | |
1044 | */ | |
1045 | fasttrap_fuword32_noerr((user_addr_t)(unsigned long)&stack[0], &s0); | |
1046 | fasttrap_fuword32_noerr((user_addr_t)(unsigned long)&stack[1], &s1); | |
1047 | fasttrap_fuword32_noerr((user_addr_t)(unsigned long)&stack[2], &s2); | |
1048 | fasttrap_fuword32_noerr((user_addr_t)(unsigned long)&stack[3], &s3); | |
1049 | fasttrap_fuword32_noerr((user_addr_t)(unsigned long)&stack[4], &s4); | |
1050 | fasttrap_fuword32_noerr((user_addr_t)(unsigned long)&stack[5], &s5); | |
1051 | ||
1052 | for (id = tp->ftt_ids; id != NULL; id = id->fti_next) { | |
1053 | fasttrap_probe_t *probe = id->fti_probe; | |
1054 | ||
1055 | if (ISSET(current_proc()->p_lflag, P_LNOATTACH)) { | |
1056 | dtrace_probe(dtrace_probeid_error, 0 /* state */, probe->ftp_id, | |
1057 | 1 /* ndx */, -1 /* offset */, DTRACEFLT_UPRIV); | |
1058 | } else { | |
1059 | if (probe->ftp_prov->ftp_provider_type == DTFTP_PROVIDER_ONESHOT) { | |
1060 | if (os_atomic_xchg(&probe->ftp_triggered, 1, relaxed)) { | |
1061 | /* already triggered */ | |
1062 | continue; | |
1063 | } | |
1064 | } | |
1065 | /* | |
1066 | * If we have at least one probe associated that | |
1067 | * is not a oneshot probe, don't remove the | |
1068 | * tracepoint | |
1069 | */ | |
1070 | else { | |
1071 | retire_tp = 0; | |
1072 | } | |
1073 | if (id->fti_ptype == DTFTP_ENTRY) { | |
1074 | /* | |
1075 | * We note that this was an entry | |
1076 | * probe to help ustack() find the | |
1077 | * first caller. | |
1078 | */ | |
1079 | cookie = dtrace_interrupt_disable(); | |
1080 | DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY); | |
1081 | dtrace_probe(probe->ftp_id, s1, s2, | |
1082 | s3, s4, s5); | |
1083 | DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY); | |
1084 | dtrace_interrupt_enable(cookie); | |
1085 | } else if (id->fti_ptype == DTFTP_IS_ENABLED) { | |
1086 | /* | |
1087 | * Note that in this case, we don't | |
1088 | * call dtrace_probe() since it's only | |
1089 | * an artificial probe meant to change | |
1090 | * the flow of control so that it | |
1091 | * encounters the true probe. | |
1092 | */ | |
1093 | is_enabled = 1; | |
1094 | } else if (probe->ftp_argmap == NULL) { | |
1095 | dtrace_probe(probe->ftp_id, s0, s1, | |
1096 | s2, s3, s4); | |
1097 | } else { | |
1098 | uint32_t t[5]; | |
1099 | ||
1100 | fasttrap_usdt_args32(probe, regs32, | |
1101 | sizeof (t) / sizeof (t[0]), t); | |
1102 | ||
1103 | dtrace_probe(probe->ftp_id, t[0], t[1], | |
1104 | t[2], t[3], t[4]); | |
1105 | } | |
1106 | } | |
1107 | } | |
1108 | if (retire_tp) { | |
1109 | fasttrap_tracepoint_retire(p, tp); | |
1110 | } | |
1111 | } | |
1112 | ||
1113 | /* | |
1114 | * We're about to do a bunch of work so we cache a local copy of | |
1115 | * the tracepoint to emulate the instruction, and then find the | |
1116 | * tracepoint again later if we need to light up any return probes. | |
1117 | */ | |
1118 | tp_local = *tp; | |
1119 | lck_mtx_unlock(pid_mtx); | |
1120 | tp = &tp_local; | |
1121 | ||
1122 | /* | |
1123 | * Set the program counter to appear as though the traced instruction | |
1124 | * had completely executed. This ensures that fasttrap_getreg() will | |
1125 | * report the expected value for REG_RIP. | |
1126 | */ | |
1127 | regs32->eip = pc + tp->ftt_size; | |
1128 | ||
1129 | /* | |
1130 | * If there's an is-enabled probe connected to this tracepoint it | |
1131 | * means that there was a 'xorl %eax, %eax' or 'xorq %rax, %rax' | |
1132 | * instruction that was placed there by DTrace when the binary was | |
1133 | * linked. As this probe is, in fact, enabled, we need to stuff 1 | |
1134 | * into %eax or %rax. Accordingly, we can bypass all the instruction | |
1135 | * emulation logic since we know the inevitable result. It's possible | |
1136 | * that a user could construct a scenario where the 'is-enabled' | |
1137 | * probe was on some other instruction, but that would be a rather | |
1138 | * exotic way to shoot oneself in the foot. | |
1139 | */ | |
1140 | if (is_enabled) { | |
1141 | regs32->eax = 1; | |
1142 | new_pc = regs32->eip; | |
1143 | goto done; | |
1144 | } | |
1145 | ||
1146 | /* | |
1147 | * We emulate certain types of instructions to ensure correctness | |
1148 | * (in the case of position dependent instructions) or optimize | |
1149 | * common cases. The rest we have the thread execute back in user- | |
1150 | * land. | |
1151 | */ | |
1152 | switch (tp->ftt_type) { | |
1153 | case FASTTRAP_T_RET: | |
1154 | case FASTTRAP_T_RET16: | |
1155 | { | |
1156 | user_addr_t dst; | |
1157 | user_addr_t addr; | |
1158 | int ret; | |
1159 | ||
1160 | /* | |
1161 | * We have to emulate _every_ facet of the behavior of a ret | |
1162 | * instruction including what happens if the load from %esp | |
1163 | * fails; in that case, we send a SIGSEGV. | |
1164 | */ | |
1165 | uint32_t dst32; | |
1166 | ret = fasttrap_fuword32((user_addr_t)regs32->uesp, &dst32); | |
1167 | dst = dst32; | |
1168 | addr = regs32->uesp + sizeof (uint32_t); | |
1169 | ||
1170 | if (ret == -1) { | |
1171 | fasttrap_sigsegv(p, uthread, (user_addr_t)regs32->uesp); | |
1172 | new_pc = pc; | |
1173 | break; | |
1174 | } | |
1175 | ||
1176 | if (tp->ftt_type == FASTTRAP_T_RET16) | |
1177 | addr += tp->ftt_dest; | |
1178 | ||
1179 | regs32->uesp = addr; | |
1180 | new_pc = dst; | |
1181 | break; | |
1182 | } | |
1183 | ||
1184 | case FASTTRAP_T_JCC: | |
1185 | { | |
1186 | uint_t taken; | |
1187 | ||
1188 | switch (tp->ftt_code) { | |
1189 | case FASTTRAP_JO: | |
1190 | taken = (regs32->efl & FASTTRAP_EFLAGS_OF) != 0; | |
1191 | break; | |
1192 | case FASTTRAP_JNO: | |
1193 | taken = (regs32->efl & FASTTRAP_EFLAGS_OF) == 0; | |
1194 | break; | |
1195 | case FASTTRAP_JB: | |
1196 | taken = (regs32->efl & FASTTRAP_EFLAGS_CF) != 0; | |
1197 | break; | |
1198 | case FASTTRAP_JAE: | |
1199 | taken = (regs32->efl & FASTTRAP_EFLAGS_CF) == 0; | |
1200 | break; | |
1201 | case FASTTRAP_JE: | |
1202 | taken = (regs32->efl & FASTTRAP_EFLAGS_ZF) != 0; | |
1203 | break; | |
1204 | case FASTTRAP_JNE: | |
1205 | taken = (regs32->efl & FASTTRAP_EFLAGS_ZF) == 0; | |
1206 | break; | |
1207 | case FASTTRAP_JBE: | |
1208 | taken = (regs32->efl & FASTTRAP_EFLAGS_CF) != 0 || | |
1209 | (regs32->efl & FASTTRAP_EFLAGS_ZF) != 0; | |
1210 | break; | |
1211 | case FASTTRAP_JA: | |
1212 | taken = (regs32->efl & FASTTRAP_EFLAGS_CF) == 0 && | |
1213 | (regs32->efl & FASTTRAP_EFLAGS_ZF) == 0; | |
1214 | break; | |
1215 | case FASTTRAP_JS: | |
1216 | taken = (regs32->efl & FASTTRAP_EFLAGS_SF) != 0; | |
1217 | break; | |
1218 | case FASTTRAP_JNS: | |
1219 | taken = (regs32->efl & FASTTRAP_EFLAGS_SF) == 0; | |
1220 | break; | |
1221 | case FASTTRAP_JP: | |
1222 | taken = (regs32->efl & FASTTRAP_EFLAGS_PF) != 0; | |
1223 | break; | |
1224 | case FASTTRAP_JNP: | |
1225 | taken = (regs32->efl & FASTTRAP_EFLAGS_PF) == 0; | |
1226 | break; | |
1227 | case FASTTRAP_JL: | |
1228 | taken = ((regs32->efl & FASTTRAP_EFLAGS_SF) == 0) != | |
1229 | ((regs32->efl & FASTTRAP_EFLAGS_OF) == 0); | |
1230 | break; | |
1231 | case FASTTRAP_JGE: | |
1232 | taken = ((regs32->efl & FASTTRAP_EFLAGS_SF) == 0) == | |
1233 | ((regs32->efl & FASTTRAP_EFLAGS_OF) == 0); | |
1234 | break; | |
1235 | case FASTTRAP_JLE: | |
1236 | taken = (regs32->efl & FASTTRAP_EFLAGS_ZF) != 0 || | |
1237 | ((regs32->efl & FASTTRAP_EFLAGS_SF) == 0) != | |
1238 | ((regs32->efl & FASTTRAP_EFLAGS_OF) == 0); | |
1239 | break; | |
1240 | case FASTTRAP_JG: | |
1241 | taken = (regs32->efl & FASTTRAP_EFLAGS_ZF) == 0 && | |
1242 | ((regs32->efl & FASTTRAP_EFLAGS_SF) == 0) == | |
1243 | ((regs32->efl & FASTTRAP_EFLAGS_OF) == 0); | |
1244 | break; | |
1245 | default: | |
1246 | taken = FALSE; | |
1247 | } | |
1248 | ||
1249 | if (taken) | |
1250 | new_pc = tp->ftt_dest; | |
1251 | else | |
1252 | new_pc = pc + tp->ftt_size; | |
1253 | break; | |
1254 | } | |
1255 | ||
1256 | case FASTTRAP_T_LOOP: | |
1257 | { | |
1258 | uint_t taken; | |
1259 | greg_t cx = regs32->ecx--; | |
1260 | ||
1261 | switch (tp->ftt_code) { | |
1262 | case FASTTRAP_LOOPNZ: | |
1263 | taken = (regs32->efl & FASTTRAP_EFLAGS_ZF) == 0 && | |
1264 | cx != 0; | |
1265 | break; | |
1266 | case FASTTRAP_LOOPZ: | |
1267 | taken = (regs32->efl & FASTTRAP_EFLAGS_ZF) != 0 && | |
1268 | cx != 0; | |
1269 | break; | |
1270 | case FASTTRAP_LOOP: | |
1271 | taken = (cx != 0); | |
1272 | break; | |
1273 | default: | |
1274 | taken = FALSE; | |
1275 | } | |
1276 | ||
1277 | if (taken) | |
1278 | new_pc = tp->ftt_dest; | |
1279 | else | |
1280 | new_pc = pc + tp->ftt_size; | |
1281 | break; | |
1282 | } | |
1283 | ||
1284 | case FASTTRAP_T_JCXZ: | |
1285 | { | |
1286 | greg_t cx = regs32->ecx; | |
1287 | ||
1288 | if (cx == 0) | |
1289 | new_pc = tp->ftt_dest; | |
1290 | else | |
1291 | new_pc = pc + tp->ftt_size; | |
1292 | break; | |
1293 | } | |
1294 | ||
1295 | case FASTTRAP_T_PUSHL_EBP: | |
1296 | { | |
1297 | user_addr_t addr = regs32->uesp - sizeof (uint32_t); | |
1298 | int ret = fasttrap_suword32(addr, (uint32_t)regs32->ebp); | |
1299 | ||
1300 | if (ret == -1) { | |
1301 | fasttrap_sigsegv(p, uthread, addr); | |
1302 | new_pc = pc; | |
1303 | break; | |
1304 | } | |
1305 | ||
1306 | regs32->uesp = addr; | |
1307 | new_pc = pc + tp->ftt_size; | |
1308 | break; | |
1309 | } | |
1310 | ||
1311 | case FASTTRAP_T_NOP: | |
1312 | new_pc = pc + tp->ftt_size; | |
1313 | break; | |
1314 | ||
1315 | case FASTTRAP_T_JMP: | |
1316 | case FASTTRAP_T_CALL: | |
1317 | if (tp->ftt_code == 0) { | |
1318 | new_pc = tp->ftt_dest; | |
1319 | } else { | |
1320 | user_addr_t /* value ,*/ addr = tp->ftt_dest; | |
1321 | ||
1322 | if (tp->ftt_base != FASTTRAP_NOREG) | |
1323 | addr += fasttrap_getreg(regs, tp->ftt_base); | |
1324 | if (tp->ftt_index != FASTTRAP_NOREG) | |
1325 | addr += fasttrap_getreg(regs, tp->ftt_index) << | |
1326 | tp->ftt_scale; | |
1327 | ||
1328 | if (tp->ftt_code == 1) { | |
1329 | /* | |
1330 | * If there's a segment prefix for this | |
1331 | * instruction, we'll need to check permissions | |
1332 | * and bounds on the given selector, and adjust | |
1333 | * the address accordingly. | |
1334 | */ | |
1335 | if (tp->ftt_segment != FASTTRAP_SEG_NONE && | |
1336 | fasttrap_do_seg(tp, regs, &addr) != 0) { | |
1337 | fasttrap_sigsegv(p, uthread, addr); | |
1338 | new_pc = pc; | |
1339 | break; | |
1340 | } | |
1341 | ||
1342 | uint32_t value32; | |
1343 | addr = (user_addr_t)(uint32_t)addr; | |
1344 | if (fasttrap_fuword32(addr, &value32) == -1) { | |
1345 | fasttrap_sigsegv(p, uthread, addr); | |
1346 | new_pc = pc; | |
1347 | break; | |
1348 | } | |
1349 | new_pc = value32; | |
1350 | } else { | |
1351 | new_pc = addr; | |
1352 | } | |
1353 | } | |
1354 | ||
1355 | /* | |
1356 | * If this is a call instruction, we need to push the return | |
1357 | * address onto the stack. If this fails, we send the process | |
1358 | * a SIGSEGV and reset the pc to emulate what would happen if | |
1359 | * this instruction weren't traced. | |
1360 | */ | |
1361 | if (tp->ftt_type == FASTTRAP_T_CALL) { | |
1362 | user_addr_t addr = regs32->uesp - sizeof (uint32_t); | |
1363 | int ret = fasttrap_suword32(addr, (uint32_t)(pc + tp->ftt_size)); | |
1364 | ||
1365 | if (ret == -1) { | |
1366 | fasttrap_sigsegv(p, uthread, addr); | |
1367 | new_pc = pc; | |
1368 | break; | |
1369 | } | |
1370 | ||
1371 | regs32->uesp = addr; | |
1372 | } | |
1373 | break; | |
1374 | ||
1375 | case FASTTRAP_T_COMMON: | |
1376 | { | |
1377 | user_addr_t addr, write_addr; | |
1378 | uint8_t scratch[2 * FASTTRAP_MAX_INSTR_SIZE + 7]; | |
1379 | uint_t i = 0; | |
1380 | ||
1381 | /* | |
1382 | * Generic Instruction Tracing | |
1383 | * --------------------------- | |
1384 | * | |
1385 | * This is the layout of the scratch space in the user-land | |
1386 | * thread structure for our generated instructions. | |
1387 | * | |
1388 | * 32-bit mode bytes | |
1389 | * ------------------------ ----- | |
1390 | * a: <original instruction> <= 15 | |
1391 | * jmp <pc + tp->ftt_size> 5 | |
1392 | * b: <original instrction> <= 15 | |
1393 | * int T_DTRACE_RET 2 | |
1394 | * ----- | |
1395 | * <= 37 | |
1396 | * | |
1397 | * 64-bit mode bytes | |
1398 | * ------------------------ ----- | |
1399 | * a: <original instruction> <= 15 | |
1400 | * jmp 0(%rip) 6 | |
1401 | * <pc + tp->ftt_size> 8 | |
1402 | * b: <original instruction> <= 15 | |
1403 | * int T_DTRACE_RET 2 | |
1404 | * ----- | |
1405 | * <= 46 | |
1406 | * | |
1407 | * The %pc is set to a, and curthread->t_dtrace_astpc is set | |
1408 | * to b. If we encounter a signal on the way out of the | |
1409 | * kernel, trap() will set %pc to curthread->t_dtrace_astpc | |
1410 | * so that we execute the original instruction and re-enter | |
1411 | * the kernel rather than redirecting to the next instruction. | |
1412 | * | |
1413 | * If there are return probes (so we know that we're going to | |
1414 | * need to reenter the kernel after executing the original | |
1415 | * instruction), the scratch space will just contain the | |
1416 | * original instruction followed by an interrupt -- the same | |
1417 | * data as at b. | |
1418 | */ | |
1419 | ||
1420 | addr = uthread->t_dtrace_scratch->addr; | |
1421 | write_addr = uthread->t_dtrace_scratch->write_addr; | |
1422 | ||
1423 | if (addr == 0LL || write_addr == 0LL) { | |
1424 | fasttrap_sigtrap(p, uthread, pc); // Should be killing target proc | |
1425 | new_pc = pc; | |
1426 | break; | |
1427 | } | |
1428 | ||
1429 | ASSERT(tp->ftt_size < FASTTRAP_MAX_INSTR_SIZE); | |
1430 | ||
1431 | uthread->t_dtrace_scrpc = addr; | |
1432 | bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size); | |
1433 | i += tp->ftt_size; | |
1434 | ||
1435 | /* | |
1436 | * Set up the jmp to the next instruction; note that | |
1437 | * the size of the traced instruction cancels out. | |
1438 | */ | |
1439 | scratch[i++] = FASTTRAP_JMP32; | |
1440 | /* LINTED - alignment */ | |
1441 | *(uint32_t *)&scratch[i] = pc - addr - 5; | |
1442 | i += sizeof (uint32_t); | |
1443 | ||
1444 | uthread->t_dtrace_astpc = addr + i; | |
1445 | bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size); | |
1446 | i += tp->ftt_size; | |
1447 | scratch[i++] = FASTTRAP_INT; | |
1448 | scratch[i++] = T_DTRACE_RET; | |
1449 | ||
1450 | ASSERT(i <= sizeof (scratch)); | |
1451 | ||
1452 | if (fasttrap_copyout(scratch, write_addr, i)) { | |
1453 | fasttrap_sigtrap(p, uthread, pc); | |
1454 | new_pc = pc; | |
1455 | break; | |
1456 | } | |
1457 | ||
1458 | if (tp->ftt_retids != NULL) { | |
1459 | uthread->t_dtrace_step = 1; | |
1460 | uthread->t_dtrace_ret = 1; | |
1461 | new_pc = uthread->t_dtrace_astpc; | |
1462 | } else { | |
1463 | new_pc = uthread->t_dtrace_scrpc; | |
1464 | } | |
1465 | ||
1466 | uthread->t_dtrace_pc = pc; | |
1467 | uthread->t_dtrace_npc = pc + tp->ftt_size; | |
1468 | uthread->t_dtrace_on = 1; | |
1469 | break; | |
1470 | } | |
1471 | ||
1472 | default: | |
1473 | panic("fasttrap: mishandled an instruction"); | |
1474 | } | |
1475 | ||
1476 | done: | |
1477 | /* | |
1478 | * APPLE NOTE: | |
1479 | * | |
1480 | * We're setting this earlier than Solaris does, to get a "correct" | |
1481 | * ustack() output. In the Sun code, a() -> b() -> c() -> d() is | |
1482 | * reported at: d, b, a. The new way gives c, b, a, which is closer | |
1483 | * to correct, as the return instruction has already exectued. | |
1484 | */ | |
1485 | regs32->eip = new_pc; | |
1486 | ||
1487 | /* | |
1488 | * If there were no return probes when we first found the tracepoint, | |
1489 | * we should feel no obligation to honor any return probes that were | |
1490 | * subsequently enabled -- they'll just have to wait until the next | |
1491 | * time around. | |
1492 | */ | |
1493 | if (tp->ftt_retids != NULL) { | |
1494 | /* | |
1495 | * We need to wait until the results of the instruction are | |
1496 | * apparent before invoking any return probes. If this | |
1497 | * instruction was emulated we can just call | |
1498 | * fasttrap_return_common(); if it needs to be executed, we | |
1499 | * need to wait until the user thread returns to the kernel. | |
1500 | */ | |
1501 | if (tp->ftt_type != FASTTRAP_T_COMMON) { | |
1502 | fasttrap_return_common(regs, pc, pid, new_pc); | |
1503 | } else { | |
1504 | ASSERT(uthread->t_dtrace_ret != 0); | |
1505 | ASSERT(uthread->t_dtrace_pc == pc); | |
1506 | ASSERT(uthread->t_dtrace_scrpc != 0); | |
1507 | ASSERT(new_pc == uthread->t_dtrace_astpc); | |
1508 | } | |
1509 | } | |
1510 | ||
1511 | return (0); | |
1512 | } | |
1513 | ||
1514 | /* | |
1515 | * Due to variances between Solaris and xnu, I have split this into a 32 bit and 64 bit | |
1516 | * code path. It still takes an x86_saved_state_t* argument, because it must sometimes | |
1517 | * call other methods that require a x86_saved_state_t. | |
1518 | * | |
1519 | * NOTE!!!! | |
1520 | * | |
1521 | * Any changes made to this method must be echo'd in fasttrap_pid_probe32! | |
1522 | * | |
1523 | */ | |
1524 | static int | |
1525 | fasttrap_pid_probe64(x86_saved_state_t *regs) | |
1526 | { | |
1527 | ASSERT(is_saved_state64(regs)); | |
1528 | ||
1529 | x86_saved_state64_t *regs64 = saved_state64(regs); | |
1530 | user_addr_t pc = regs64->isf.rip - 1; | |
1531 | proc_t *p = current_proc(); | |
1532 | user_addr_t new_pc = 0; | |
1533 | fasttrap_bucket_t *bucket; | |
1534 | lck_mtx_t *pid_mtx; | |
1535 | fasttrap_tracepoint_t *tp, tp_local; | |
1536 | pid_t pid; | |
1537 | dtrace_icookie_t cookie; | |
1538 | uint_t is_enabled = 0; | |
1539 | int retire_tp = 1; | |
1540 | ||
1541 | uthread_t uthread = (uthread_t)get_bsdthread_info(current_thread()); | |
1542 | ||
1543 | /* | |
1544 | * It's possible that a user (in a veritable orgy of bad planning) | |
1545 | * could redirect this thread's flow of control before it reached the | |
1546 | * return probe fasttrap. In this case we need to kill the process | |
1547 | * since it's in a unrecoverable state. | |
1548 | */ | |
1549 | if (uthread->t_dtrace_step) { | |
1550 | ASSERT(uthread->t_dtrace_on); | |
1551 | fasttrap_sigtrap(p, uthread, pc); | |
1552 | return (0); | |
1553 | } | |
1554 | ||
1555 | /* | |
1556 | * Clear all user tracing flags. | |
1557 | */ | |
1558 | uthread->t_dtrace_ft = 0; | |
1559 | uthread->t_dtrace_pc = 0; | |
1560 | uthread->t_dtrace_npc = 0; | |
1561 | uthread->t_dtrace_scrpc = 0; | |
1562 | uthread->t_dtrace_astpc = 0; | |
1563 | uthread->t_dtrace_regv = 0; | |
1564 | ||
1565 | /* | |
1566 | * Treat a child created by a call to vfork(2) as if it were its | |
1567 | * parent. We know that there's only one thread of control in such a | |
1568 | * process: this one. | |
1569 | */ | |
1570 | if (p->p_lflag & P_LINVFORK) { | |
1571 | proc_list_lock(); | |
1572 | while (p->p_lflag & P_LINVFORK) | |
1573 | p = p->p_pptr; | |
1574 | proc_list_unlock(); | |
1575 | } | |
1576 | ||
1577 | pid = p->p_pid; | |
1578 | pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock; | |
1579 | lck_mtx_lock(pid_mtx); | |
1580 | bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)]; | |
1581 | ||
1582 | /* | |
1583 | * Lookup the tracepoint that the process just hit. | |
1584 | */ | |
1585 | for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) { | |
1586 | if (pid == tp->ftt_pid && pc == tp->ftt_pc && | |
1587 | tp->ftt_proc->ftpc_acount != 0) | |
1588 | break; | |
1589 | } | |
1590 | ||
1591 | /* | |
1592 | * If we couldn't find a matching tracepoint, either a tracepoint has | |
1593 | * been inserted without using the pid<pid> ioctl interface (see | |
1594 | * fasttrap_ioctl), or somehow we have mislaid this tracepoint. | |
1595 | */ | |
1596 | if (tp == NULL) { | |
1597 | lck_mtx_unlock(pid_mtx); | |
1598 | return (-1); | |
1599 | } | |
1600 | ||
1601 | /* | |
1602 | * Set the program counter to the address of the traced instruction | |
1603 | * so that it looks right in ustack() output. | |
1604 | */ | |
1605 | regs64->isf.rip = pc; | |
1606 | ||
1607 | if (tp->ftt_ids != NULL) { | |
1608 | fasttrap_id_t *id; | |
1609 | ||
1610 | for (id = tp->ftt_ids; id != NULL; id = id->fti_next) { | |
1611 | fasttrap_probe_t *probe = id->fti_probe; | |
1612 | ||
1613 | if (probe->ftp_prov->ftp_provider_type == DTFTP_PROVIDER_ONESHOT) { | |
1614 | if (os_atomic_xchg(&probe->ftp_triggered, 1, relaxed)) { | |
1615 | /* already triggered */ | |
1616 | continue; | |
1617 | } | |
1618 | } | |
1619 | /* | |
1620 | * If we have at least probe associated that | |
1621 | * is not a oneshot probe, don't remove the | |
1622 | * tracepoint | |
1623 | */ | |
1624 | else { | |
1625 | retire_tp = 0; | |
1626 | } | |
1627 | if (ISSET(current_proc()->p_lflag, P_LNOATTACH)) { | |
1628 | dtrace_probe(dtrace_probeid_error, 0 /* state */, probe->ftp_id, | |
1629 | 1 /* ndx */, -1 /* offset */, DTRACEFLT_UPRIV); | |
1630 | } else if (id->fti_ptype == DTFTP_ENTRY) { | |
1631 | /* | |
1632 | * We note that this was an entry | |
1633 | * probe to help ustack() find the | |
1634 | * first caller. | |
1635 | */ | |
1636 | cookie = dtrace_interrupt_disable(); | |
1637 | DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY); | |
1638 | dtrace_probe(probe->ftp_id, regs64->rdi, | |
1639 | regs64->rsi, regs64->rdx, regs64->rcx, | |
1640 | regs64->r8); | |
1641 | DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY); | |
1642 | dtrace_interrupt_enable(cookie); | |
1643 | } else if (id->fti_ptype == DTFTP_IS_ENABLED) { | |
1644 | /* | |
1645 | * Note that in this case, we don't | |
1646 | * call dtrace_probe() since it's only | |
1647 | * an artificial probe meant to change | |
1648 | * the flow of control so that it | |
1649 | * encounters the true probe. | |
1650 | */ | |
1651 | is_enabled = 1; | |
1652 | } else if (probe->ftp_argmap == NULL) { | |
1653 | dtrace_probe(probe->ftp_id, regs64->rdi, | |
1654 | regs64->rsi, regs64->rdx, regs64->rcx, | |
1655 | regs64->r8); | |
1656 | } else { | |
1657 | uint64_t t[5]; | |
1658 | ||
1659 | fasttrap_usdt_args64(probe, regs64, | |
1660 | sizeof (t) / sizeof (t[0]), t); | |
1661 | ||
1662 | dtrace_probe(probe->ftp_id, t[0], t[1], | |
1663 | t[2], t[3], t[4]); | |
1664 | } | |
1665 | ||
1666 | } | |
1667 | if (retire_tp) { | |
1668 | fasttrap_tracepoint_retire(p, tp); | |
1669 | } | |
1670 | } | |
1671 | ||
1672 | /* | |
1673 | * We're about to do a bunch of work so we cache a local copy of | |
1674 | * the tracepoint to emulate the instruction, and then find the | |
1675 | * tracepoint again later if we need to light up any return probes. | |
1676 | */ | |
1677 | tp_local = *tp; | |
1678 | lck_mtx_unlock(pid_mtx); | |
1679 | tp = &tp_local; | |
1680 | ||
1681 | /* | |
1682 | * Set the program counter to appear as though the traced instruction | |
1683 | * had completely executed. This ensures that fasttrap_getreg() will | |
1684 | * report the expected value for REG_RIP. | |
1685 | */ | |
1686 | regs64->isf.rip = pc + tp->ftt_size; | |
1687 | ||
1688 | /* | |
1689 | * If there's an is-enabled probe connected to this tracepoint it | |
1690 | * means that there was a 'xorl %eax, %eax' or 'xorq %rax, %rax' | |
1691 | * instruction that was placed there by DTrace when the binary was | |
1692 | * linked. As this probe is, in fact, enabled, we need to stuff 1 | |
1693 | * into %eax or %rax. Accordingly, we can bypass all the instruction | |
1694 | * emulation logic since we know the inevitable result. It's possible | |
1695 | * that a user could construct a scenario where the 'is-enabled' | |
1696 | * probe was on some other instruction, but that would be a rather | |
1697 | * exotic way to shoot oneself in the foot. | |
1698 | */ | |
1699 | if (is_enabled) { | |
1700 | regs64->rax = 1; | |
1701 | new_pc = regs64->isf.rip; | |
1702 | goto done; | |
1703 | } | |
1704 | ||
1705 | /* | |
1706 | * We emulate certain types of instructions to ensure correctness | |
1707 | * (in the case of position dependent instructions) or optimize | |
1708 | * common cases. The rest we have the thread execute back in user- | |
1709 | * land. | |
1710 | */ | |
1711 | switch (tp->ftt_type) { | |
1712 | case FASTTRAP_T_RET: | |
1713 | case FASTTRAP_T_RET16: | |
1714 | { | |
1715 | user_addr_t dst; | |
1716 | user_addr_t addr; | |
1717 | int ret; | |
1718 | ||
1719 | /* | |
1720 | * We have to emulate _every_ facet of the behavior of a ret | |
1721 | * instruction including what happens if the load from %esp | |
1722 | * fails; in that case, we send a SIGSEGV. | |
1723 | */ | |
1724 | ret = fasttrap_fuword64((user_addr_t)regs64->isf.rsp, &dst); | |
1725 | addr = regs64->isf.rsp + sizeof (uint64_t); | |
1726 | ||
1727 | if (ret == -1) { | |
1728 | fasttrap_sigsegv(p, uthread, (user_addr_t)regs64->isf.rsp); | |
1729 | new_pc = pc; | |
1730 | break; | |
1731 | } | |
1732 | ||
1733 | if (tp->ftt_type == FASTTRAP_T_RET16) | |
1734 | addr += tp->ftt_dest; | |
1735 | ||
1736 | regs64->isf.rsp = addr; | |
1737 | new_pc = dst; | |
1738 | break; | |
1739 | } | |
1740 | ||
1741 | case FASTTRAP_T_JCC: | |
1742 | { | |
1743 | uint_t taken; | |
1744 | ||
1745 | switch (tp->ftt_code) { | |
1746 | case FASTTRAP_JO: | |
1747 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_OF) != 0; | |
1748 | break; | |
1749 | case FASTTRAP_JNO: | |
1750 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_OF) == 0; | |
1751 | break; | |
1752 | case FASTTRAP_JB: | |
1753 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_CF) != 0; | |
1754 | break; | |
1755 | case FASTTRAP_JAE: | |
1756 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_CF) == 0; | |
1757 | break; | |
1758 | case FASTTRAP_JE: | |
1759 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_ZF) != 0; | |
1760 | break; | |
1761 | case FASTTRAP_JNE: | |
1762 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_ZF) == 0; | |
1763 | break; | |
1764 | case FASTTRAP_JBE: | |
1765 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_CF) != 0 || | |
1766 | (regs64->isf.rflags & FASTTRAP_EFLAGS_ZF) != 0; | |
1767 | break; | |
1768 | case FASTTRAP_JA: | |
1769 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_CF) == 0 && | |
1770 | (regs64->isf.rflags & FASTTRAP_EFLAGS_ZF) == 0; | |
1771 | break; | |
1772 | case FASTTRAP_JS: | |
1773 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_SF) != 0; | |
1774 | break; | |
1775 | case FASTTRAP_JNS: | |
1776 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_SF) == 0; | |
1777 | break; | |
1778 | case FASTTRAP_JP: | |
1779 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_PF) != 0; | |
1780 | break; | |
1781 | case FASTTRAP_JNP: | |
1782 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_PF) == 0; | |
1783 | break; | |
1784 | case FASTTRAP_JL: | |
1785 | taken = ((regs64->isf.rflags & FASTTRAP_EFLAGS_SF) == 0) != | |
1786 | ((regs64->isf.rflags & FASTTRAP_EFLAGS_OF) == 0); | |
1787 | break; | |
1788 | case FASTTRAP_JGE: | |
1789 | taken = ((regs64->isf.rflags & FASTTRAP_EFLAGS_SF) == 0) == | |
1790 | ((regs64->isf.rflags & FASTTRAP_EFLAGS_OF) == 0); | |
1791 | break; | |
1792 | case FASTTRAP_JLE: | |
1793 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_ZF) != 0 || | |
1794 | ((regs64->isf.rflags & FASTTRAP_EFLAGS_SF) == 0) != | |
1795 | ((regs64->isf.rflags & FASTTRAP_EFLAGS_OF) == 0); | |
1796 | break; | |
1797 | case FASTTRAP_JG: | |
1798 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_ZF) == 0 && | |
1799 | ((regs64->isf.rflags & FASTTRAP_EFLAGS_SF) == 0) == | |
1800 | ((regs64->isf.rflags & FASTTRAP_EFLAGS_OF) == 0); | |
1801 | break; | |
1802 | default: | |
1803 | taken = FALSE; | |
1804 | } | |
1805 | ||
1806 | if (taken) | |
1807 | new_pc = tp->ftt_dest; | |
1808 | else | |
1809 | new_pc = pc + tp->ftt_size; | |
1810 | break; | |
1811 | } | |
1812 | ||
1813 | case FASTTRAP_T_LOOP: | |
1814 | { | |
1815 | uint_t taken; | |
1816 | uint64_t cx = regs64->rcx--; | |
1817 | ||
1818 | switch (tp->ftt_code) { | |
1819 | case FASTTRAP_LOOPNZ: | |
1820 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_ZF) == 0 && | |
1821 | cx != 0; | |
1822 | break; | |
1823 | case FASTTRAP_LOOPZ: | |
1824 | taken = (regs64->isf.rflags & FASTTRAP_EFLAGS_ZF) != 0 && | |
1825 | cx != 0; | |
1826 | break; | |
1827 | case FASTTRAP_LOOP: | |
1828 | taken = (cx != 0); | |
1829 | break; | |
1830 | default: | |
1831 | taken = FALSE; | |
1832 | } | |
1833 | ||
1834 | if (taken) | |
1835 | new_pc = tp->ftt_dest; | |
1836 | else | |
1837 | new_pc = pc + tp->ftt_size; | |
1838 | break; | |
1839 | } | |
1840 | ||
1841 | case FASTTRAP_T_JCXZ: | |
1842 | { | |
1843 | uint64_t cx = regs64->rcx; | |
1844 | ||
1845 | if (cx == 0) | |
1846 | new_pc = tp->ftt_dest; | |
1847 | else | |
1848 | new_pc = pc + tp->ftt_size; | |
1849 | break; | |
1850 | } | |
1851 | ||
1852 | case FASTTRAP_T_PUSHL_EBP: | |
1853 | { | |
1854 | user_addr_t addr = regs64->isf.rsp - sizeof (uint64_t); | |
1855 | int ret = fasttrap_suword64(addr, (uint64_t)regs64->rbp); | |
1856 | ||
1857 | if (ret == -1) { | |
1858 | fasttrap_sigsegv(p, uthread, addr); | |
1859 | new_pc = pc; | |
1860 | break; | |
1861 | } | |
1862 | ||
1863 | regs64->isf.rsp = addr; | |
1864 | new_pc = pc + tp->ftt_size; | |
1865 | break; | |
1866 | } | |
1867 | ||
1868 | case FASTTRAP_T_NOP: | |
1869 | new_pc = pc + tp->ftt_size; | |
1870 | break; | |
1871 | ||
1872 | case FASTTRAP_T_JMP: | |
1873 | case FASTTRAP_T_CALL: | |
1874 | if (tp->ftt_code == 0) { | |
1875 | new_pc = tp->ftt_dest; | |
1876 | } else { | |
1877 | user_addr_t value, addr = tp->ftt_dest; | |
1878 | ||
1879 | if (tp->ftt_base != FASTTRAP_NOREG) | |
1880 | addr += fasttrap_getreg(regs, tp->ftt_base); | |
1881 | if (tp->ftt_index != FASTTRAP_NOREG) | |
1882 | addr += fasttrap_getreg(regs, tp->ftt_index) << | |
1883 | tp->ftt_scale; | |
1884 | ||
1885 | if (tp->ftt_code == 1) { | |
1886 | /* | |
1887 | * If there's a segment prefix for this | |
1888 | * instruction, we'll need to check permissions | |
1889 | * and bounds on the given selector, and adjust | |
1890 | * the address accordingly. | |
1891 | */ | |
1892 | if (tp->ftt_segment != FASTTRAP_SEG_NONE && | |
1893 | fasttrap_do_seg(tp, regs, &addr) != 0) { | |
1894 | fasttrap_sigsegv(p, uthread, addr); | |
1895 | new_pc = pc; | |
1896 | break; | |
1897 | } | |
1898 | ||
1899 | if (fasttrap_fuword64(addr, &value) == -1) { | |
1900 | fasttrap_sigsegv(p, uthread, addr); | |
1901 | new_pc = pc; | |
1902 | break; | |
1903 | } | |
1904 | new_pc = value; | |
1905 | } else { | |
1906 | new_pc = addr; | |
1907 | } | |
1908 | } | |
1909 | ||
1910 | /* | |
1911 | * If this is a call instruction, we need to push the return | |
1912 | * address onto the stack. If this fails, we send the process | |
1913 | * a SIGSEGV and reset the pc to emulate what would happen if | |
1914 | * this instruction weren't traced. | |
1915 | */ | |
1916 | if (tp->ftt_type == FASTTRAP_T_CALL) { | |
1917 | user_addr_t addr = regs64->isf.rsp - sizeof (uint64_t); | |
1918 | int ret = fasttrap_suword64(addr, pc + tp->ftt_size); | |
1919 | ||
1920 | if (ret == -1) { | |
1921 | fasttrap_sigsegv(p, uthread, addr); | |
1922 | new_pc = pc; | |
1923 | break; | |
1924 | } | |
1925 | ||
1926 | regs64->isf.rsp = addr; | |
1927 | } | |
1928 | break; | |
1929 | ||
1930 | case FASTTRAP_T_COMMON: | |
1931 | { | |
1932 | user_addr_t addr, write_addr; | |
1933 | uint8_t scratch[2 * FASTTRAP_MAX_INSTR_SIZE + 22]; | |
1934 | uint_t i = 0; | |
1935 | ||
1936 | /* | |
1937 | * Generic Instruction Tracing | |
1938 | * --------------------------- | |
1939 | * | |
1940 | * This is the layout of the scratch space in the user-land | |
1941 | * thread structure for our generated instructions. | |
1942 | * | |
1943 | * 32-bit mode bytes | |
1944 | * ------------------------ ----- | |
1945 | * a: <original instruction> <= 15 | |
1946 | * jmp <pc + tp->ftt_size> 5 | |
1947 | * b: <original instrction> <= 15 | |
1948 | * int T_DTRACE_RET 2 | |
1949 | * ----- | |
1950 | * <= 37 | |
1951 | * | |
1952 | * 64-bit mode bytes | |
1953 | * ------------------------ ----- | |
1954 | * a: <original instruction> <= 15 | |
1955 | * jmp 0(%rip) 6 | |
1956 | * <pc + tp->ftt_size> 8 | |
1957 | * b: <original instruction> <= 15 | |
1958 | * int T_DTRACE_RET 2 | |
1959 | * ----- | |
1960 | * <= 46 | |
1961 | * | |
1962 | * The %pc is set to a, and curthread->t_dtrace_astpc is set | |
1963 | * to b. If we encounter a signal on the way out of the | |
1964 | * kernel, trap() will set %pc to curthread->t_dtrace_astpc | |
1965 | * so that we execute the original instruction and re-enter | |
1966 | * the kernel rather than redirecting to the next instruction. | |
1967 | * | |
1968 | * If there are return probes (so we know that we're going to | |
1969 | * need to reenter the kernel after executing the original | |
1970 | * instruction), the scratch space will just contain the | |
1971 | * original instruction followed by an interrupt -- the same | |
1972 | * data as at b. | |
1973 | * | |
1974 | * %rip-relative Addressing | |
1975 | * ------------------------ | |
1976 | * | |
1977 | * There's a further complication in 64-bit mode due to %rip- | |
1978 | * relative addressing. While this is clearly a beneficial | |
1979 | * architectural decision for position independent code, it's | |
1980 | * hard not to see it as a personal attack against the pid | |
1981 | * provider since before there was a relatively small set of | |
1982 | * instructions to emulate; with %rip-relative addressing, | |
1983 | * almost every instruction can potentially depend on the | |
1984 | * address at which it's executed. Rather than emulating | |
1985 | * the broad spectrum of instructions that can now be | |
1986 | * position dependent, we emulate jumps and others as in | |
1987 | * 32-bit mode, and take a different tack for instructions | |
1988 | * using %rip-relative addressing. | |
1989 | * | |
1990 | * For every instruction that uses the ModRM byte, the | |
1991 | * in-kernel disassembler reports its location. We use the | |
1992 | * ModRM byte to identify that an instruction uses | |
1993 | * %rip-relative addressing and to see what other registers | |
1994 | * the instruction uses. To emulate those instructions, | |
1995 | * we modify the instruction to be %rax-relative rather than | |
1996 | * %rip-relative (or %rcx-relative if the instruction uses | |
1997 | * %rax; or %r8- or %r9-relative if the REX.B is present so | |
1998 | * we don't have to rewrite the REX prefix). We then load | |
1999 | * the value that %rip would have been into the scratch | |
2000 | * register and generate an instruction to reset the scratch | |
2001 | * register back to its original value. The instruction | |
2002 | * sequence looks like this: | |
2003 | * | |
2004 | * 64-mode %rip-relative bytes | |
2005 | * ------------------------ ----- | |
2006 | * a: <modified instruction> <= 15 | |
2007 | * movq $<value>, %<scratch> 6 | |
2008 | * jmp 0(%rip) 6 | |
2009 | * <pc + tp->ftt_size> 8 | |
2010 | * b: <modified instruction> <= 15 | |
2011 | * int T_DTRACE_RET 2 | |
2012 | * ----- | |
2013 | * 52 | |
2014 | * | |
2015 | * We set curthread->t_dtrace_regv so that upon receiving | |
2016 | * a signal we can reset the value of the scratch register. | |
2017 | */ | |
2018 | ||
2019 | addr = uthread->t_dtrace_scratch->addr; | |
2020 | write_addr = uthread->t_dtrace_scratch->write_addr; | |
2021 | ||
2022 | if (addr == 0LL || write_addr == 0LL) { | |
2023 | fasttrap_sigtrap(p, uthread, pc); // Should be killing target proc | |
2024 | new_pc = pc; | |
2025 | break; | |
2026 | } | |
2027 | ||
2028 | ASSERT(tp->ftt_size < FASTTRAP_MAX_INSTR_SIZE); | |
2029 | ||
2030 | uthread->t_dtrace_scrpc = addr; | |
2031 | bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size); | |
2032 | i += tp->ftt_size; | |
2033 | ||
2034 | if (tp->ftt_ripmode != 0) { | |
2035 | uint64_t* reg; | |
2036 | ||
2037 | ASSERT(tp->ftt_ripmode & | |
2038 | (FASTTRAP_RIP_1 | FASTTRAP_RIP_2)); | |
2039 | ||
2040 | /* | |
2041 | * If this was a %rip-relative instruction, we change | |
2042 | * it to be either a %rax- or %rcx-relative | |
2043 | * instruction (depending on whether those registers | |
2044 | * are used as another operand; or %r8- or %r9- | |
2045 | * relative depending on the value of REX.B). We then | |
2046 | * set that register and generate a movq instruction | |
2047 | * to reset the value. | |
2048 | */ | |
2049 | if (tp->ftt_ripmode & FASTTRAP_RIP_X) | |
2050 | scratch[i++] = FASTTRAP_REX(1, 0, 0, 1); | |
2051 | else | |
2052 | scratch[i++] = FASTTRAP_REX(1, 0, 0, 0); | |
2053 | ||
2054 | if (tp->ftt_ripmode & FASTTRAP_RIP_1) | |
2055 | scratch[i++] = FASTTRAP_MOV_EAX; | |
2056 | else | |
2057 | scratch[i++] = FASTTRAP_MOV_ECX; | |
2058 | ||
2059 | switch (tp->ftt_ripmode) { | |
2060 | case FASTTRAP_RIP_1: | |
2061 | reg = ®s64->rax; | |
2062 | uthread->t_dtrace_reg = REG_RAX; | |
2063 | break; | |
2064 | case FASTTRAP_RIP_2: | |
2065 | reg = ®s64->rcx; | |
2066 | uthread->t_dtrace_reg = REG_RCX; | |
2067 | break; | |
2068 | case FASTTRAP_RIP_1 | FASTTRAP_RIP_X: | |
2069 | reg = ®s64->r8; | |
2070 | uthread->t_dtrace_reg = REG_R8; | |
2071 | break; | |
2072 | case FASTTRAP_RIP_2 | FASTTRAP_RIP_X: | |
2073 | reg = ®s64->r9; | |
2074 | uthread->t_dtrace_reg = REG_R9; | |
2075 | break; | |
2076 | default: | |
2077 | reg = NULL; | |
2078 | panic("unhandled ripmode in fasttrap_pid_probe64"); | |
2079 | } | |
2080 | ||
2081 | /* LINTED - alignment */ | |
2082 | *(uint64_t *)&scratch[i] = *reg; | |
2083 | uthread->t_dtrace_regv = *reg; | |
2084 | *reg = pc + tp->ftt_size; | |
2085 | i += sizeof (uint64_t); | |
2086 | } | |
2087 | ||
2088 | /* | |
2089 | * Generate the branch instruction to what would have | |
2090 | * normally been the subsequent instruction. In 32-bit mode, | |
2091 | * this is just a relative branch; in 64-bit mode this is a | |
2092 | * %rip-relative branch that loads the 64-bit pc value | |
2093 | * immediately after the jmp instruction. | |
2094 | */ | |
2095 | scratch[i++] = FASTTRAP_GROUP5_OP; | |
2096 | scratch[i++] = FASTTRAP_MODRM(0, 4, 5); | |
2097 | /* LINTED - alignment */ | |
2098 | *(uint32_t *)&scratch[i] = 0; | |
2099 | i += sizeof (uint32_t); | |
2100 | /* LINTED - alignment */ | |
2101 | *(uint64_t *)&scratch[i] = pc + tp->ftt_size; | |
2102 | i += sizeof (uint64_t); | |
2103 | ||
2104 | uthread->t_dtrace_astpc = addr + i; | |
2105 | bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size); | |
2106 | i += tp->ftt_size; | |
2107 | scratch[i++] = FASTTRAP_INT; | |
2108 | scratch[i++] = T_DTRACE_RET; | |
2109 | ||
2110 | ASSERT(i <= sizeof (scratch)); | |
2111 | ||
2112 | if (fasttrap_copyout(scratch, write_addr, i)) { | |
2113 | fasttrap_sigtrap(p, uthread, pc); | |
2114 | new_pc = pc; | |
2115 | break; | |
2116 | } | |
2117 | ||
2118 | if (tp->ftt_retids != NULL) { | |
2119 | uthread->t_dtrace_step = 1; | |
2120 | uthread->t_dtrace_ret = 1; | |
2121 | new_pc = uthread->t_dtrace_astpc; | |
2122 | } else { | |
2123 | new_pc = uthread->t_dtrace_scrpc; | |
2124 | } | |
2125 | ||
2126 | uthread->t_dtrace_pc = pc; | |
2127 | uthread->t_dtrace_npc = pc + tp->ftt_size; | |
2128 | uthread->t_dtrace_on = 1; | |
2129 | break; | |
2130 | } | |
2131 | ||
2132 | default: | |
2133 | panic("fasttrap: mishandled an instruction"); | |
2134 | } | |
2135 | ||
2136 | done: | |
2137 | /* | |
2138 | * APPLE NOTE: | |
2139 | * | |
2140 | * We're setting this earlier than Solaris does, to get a "correct" | |
2141 | * ustack() output. In the Sun code, a() -> b() -> c() -> d() is | |
2142 | * reported at: d, b, a. The new way gives c, b, a, which is closer | |
2143 | * to correct, as the return instruction has already exectued. | |
2144 | */ | |
2145 | regs64->isf.rip = new_pc; | |
2146 | ||
2147 | ||
2148 | /* | |
2149 | * If there were no return probes when we first found the tracepoint, | |
2150 | * we should feel no obligation to honor any return probes that were | |
2151 | * subsequently enabled -- they'll just have to wait until the next | |
2152 | * time around. | |
2153 | */ | |
2154 | if (tp->ftt_retids != NULL) { | |
2155 | /* | |
2156 | * We need to wait until the results of the instruction are | |
2157 | * apparent before invoking any return probes. If this | |
2158 | * instruction was emulated we can just call | |
2159 | * fasttrap_return_common(); if it needs to be executed, we | |
2160 | * need to wait until the user thread returns to the kernel. | |
2161 | */ | |
2162 | if (tp->ftt_type != FASTTRAP_T_COMMON) { | |
2163 | fasttrap_return_common(regs, pc, pid, new_pc); | |
2164 | } else { | |
2165 | ASSERT(uthread->t_dtrace_ret != 0); | |
2166 | ASSERT(uthread->t_dtrace_pc == pc); | |
2167 | ASSERT(uthread->t_dtrace_scrpc != 0); | |
2168 | ASSERT(new_pc == uthread->t_dtrace_astpc); | |
2169 | } | |
2170 | } | |
2171 | ||
2172 | return (0); | |
2173 | } | |
2174 | ||
2175 | int | |
2176 | fasttrap_pid_probe(x86_saved_state_t *regs) | |
2177 | { | |
2178 | if (is_saved_state64(regs)) | |
2179 | return fasttrap_pid_probe64(regs); | |
2180 | ||
2181 | return fasttrap_pid_probe32(regs); | |
2182 | } | |
2183 | ||
2184 | int | |
2185 | fasttrap_return_probe(x86_saved_state_t *regs) | |
2186 | { | |
2187 | x86_saved_state64_t *regs64; | |
2188 | x86_saved_state32_t *regs32; | |
2189 | unsigned int p_model; | |
2190 | ||
2191 | if (is_saved_state64(regs)) { | |
2192 | regs64 = saved_state64(regs); | |
2193 | regs32 = NULL; | |
2194 | p_model = DATAMODEL_LP64; | |
2195 | } else { | |
2196 | regs64 = NULL; | |
2197 | regs32 = saved_state32(regs); | |
2198 | p_model = DATAMODEL_ILP32; | |
2199 | } | |
2200 | ||
2201 | proc_t *p = current_proc(); | |
2202 | uthread_t uthread = (uthread_t)get_bsdthread_info(current_thread()); | |
2203 | user_addr_t pc = uthread->t_dtrace_pc; | |
2204 | user_addr_t npc = uthread->t_dtrace_npc; | |
2205 | ||
2206 | uthread->t_dtrace_pc = 0; | |
2207 | uthread->t_dtrace_npc = 0; | |
2208 | uthread->t_dtrace_scrpc = 0; | |
2209 | uthread->t_dtrace_astpc = 0; | |
2210 | ||
2211 | /* | |
2212 | * Treat a child created by a call to vfork(2) as if it were its | |
2213 | * parent. We know that there's only one thread of control in such a | |
2214 | * process: this one. | |
2215 | */ | |
2216 | proc_list_lock(); | |
2217 | while (p->p_lflag & P_LINVFORK) | |
2218 | p = p->p_pptr; | |
2219 | proc_list_unlock(); | |
2220 | ||
2221 | /* | |
2222 | * We set rp->r_pc to the address of the traced instruction so | |
2223 | * that it appears to dtrace_probe() that we're on the original | |
2224 | * instruction, and so that the user can't easily detect our | |
2225 | * complex web of lies. dtrace_return_probe() (our caller) | |
2226 | * will correctly set %pc after we return. | |
2227 | */ | |
2228 | if (p_model == DATAMODEL_LP64) | |
2229 | regs64->isf.rip = pc; | |
2230 | else | |
2231 | regs32->eip = pc; | |
2232 | ||
2233 | fasttrap_return_common(regs, pc, p->p_pid, npc); | |
2234 | ||
2235 | return (0); | |
2236 | } | |
2237 | ||
2238 | uint64_t | |
2239 | fasttrap_pid_getarg(void *arg, dtrace_id_t id, void *parg, int argno, | |
2240 | int aframes) | |
2241 | { | |
2242 | pal_register_cache_state(current_thread(), VALID); | |
2243 | #pragma unused(arg, id, parg, aframes) | |
2244 | return (fasttrap_anarg((x86_saved_state_t *)find_user_regs(current_thread()), 1, argno)); | |
2245 | } | |
2246 | ||
2247 | uint64_t | |
2248 | fasttrap_usdt_getarg(void *arg, dtrace_id_t id, void *parg, int argno, | |
2249 | int aframes) | |
2250 | { | |
2251 | pal_register_cache_state(current_thread(), VALID); | |
2252 | #pragma unused(arg, id, parg, aframes) | |
2253 | return (fasttrap_anarg((x86_saved_state_t *)find_user_regs(current_thread()), 0, argno)); | |
2254 | } | |
2255 | ||
2256 | /* | |
2257 | * APPLE NOTE: See comments by regmap array definition. We are cheating | |
2258 | * when returning 32 bit registers. | |
2259 | */ | |
2260 | static user_addr_t | |
2261 | fasttrap_getreg(x86_saved_state_t *regs, uint_t reg) | |
2262 | { | |
2263 | if (is_saved_state64(regs)) { | |
2264 | x86_saved_state64_t *regs64 = saved_state64(regs); | |
2265 | ||
2266 | switch (reg) { | |
2267 | case REG_RAX: return regs64->rax; | |
2268 | case REG_RCX: return regs64->rcx; | |
2269 | case REG_RDX: return regs64->rdx; | |
2270 | case REG_RBX: return regs64->rbx; | |
2271 | case REG_RSP: return regs64->isf.rsp; | |
2272 | case REG_RBP: return regs64->rbp; | |
2273 | case REG_RSI: return regs64->rsi; | |
2274 | case REG_RDI: return regs64->rdi; | |
2275 | case REG_R8: return regs64->r8; | |
2276 | case REG_R9: return regs64->r9; | |
2277 | case REG_R10: return regs64->r10; | |
2278 | case REG_R11: return regs64->r11; | |
2279 | case REG_R12: return regs64->r12; | |
2280 | case REG_R13: return regs64->r13; | |
2281 | case REG_R14: return regs64->r14; | |
2282 | case REG_R15: return regs64->r15; | |
2283 | case REG_TRAPNO: return regs64->isf.trapno; | |
2284 | case REG_ERR: return regs64->isf.err; | |
2285 | case REG_RIP: return regs64->isf.rip; | |
2286 | case REG_CS: return regs64->isf.cs; | |
2287 | case REG_RFL: return regs64->isf.rflags; | |
2288 | case REG_SS: return regs64->isf.ss; | |
2289 | case REG_FS: return regs64->fs; | |
2290 | case REG_GS: return regs64->gs; | |
2291 | case REG_ES: | |
2292 | case REG_DS: | |
2293 | case REG_FSBASE: | |
2294 | case REG_GSBASE: | |
2295 | // Important to distinguish these requests (which should be legal) from other values. | |
2296 | panic("dtrace: unimplemented x86_64 getreg()"); | |
2297 | } | |
2298 | ||
2299 | panic("dtrace: unhandled x86_64 getreg() constant"); | |
2300 | } else { | |
2301 | x86_saved_state32_t *regs32 = saved_state32(regs); | |
2302 | ||
2303 | switch (reg) { | |
2304 | case REG_RAX: return regs32->eax; | |
2305 | case REG_RCX: return regs32->ecx; | |
2306 | case REG_RDX: return regs32->edx; | |
2307 | case REG_RBX: return regs32->ebx; | |
2308 | case REG_RSP: return regs32->uesp; | |
2309 | case REG_RBP: return regs32->ebp; | |
2310 | case REG_RSI: return regs32->esi; | |
2311 | case REG_RDI: return regs32->edi; | |
2312 | } | |
2313 | ||
2314 | panic("dtrace: unhandled i386 getreg() constant"); | |
2315 | } | |
2316 | ||
2317 | return 0; | |
2318 | } |