xref: /freebsd/sys/cddl/dev/kinst/amd64/kinst_isa.c (revision 9f27341c336aa12f6c7163c17e646e76c813b689)
1 /*
2  * SPDX-License-Identifier: CDDL 1.0
3  *
4  * Copyright 2022 Christos Margiolis <christos@FreeBSD.org>
5  * Copyright 2022 Mark Johnston <markj@FreeBSD.org>
6  */
7 
8 #include <sys/param.h>
9 #include <sys/pcpu.h>
10 
11 #include <machine/cpufunc.h>
12 #include <machine/md_var.h>
13 
14 #include <sys/dtrace.h>
15 #include <cddl/dev/dtrace/dtrace_cddl.h>
16 #include <dis_tables.h>
17 
18 #include "kinst.h"
19 
20 #define KINST_PUSHL_RBP		0x55
21 #define KINST_STI		0xfb
22 #define KINST_POPF		0x9d
23 
24 #define KINST_MODRM_MOD(b)	(((b) & 0xc0) >> 6)
25 #define KINST_MODRM_REG(b)	(((b) & 0x38) >> 3)
26 #define KINST_MODRM_RM(b)	((b) & 0x07)
27 
28 #define KINST_SIB_SCALE(s)	(((s) & 0xc0) >> 6)
29 #define KINST_SIB_INDEX(s)	(((s) & 0x38) >> 3)
30 #define KINST_SIB_BASE(s)	(((s) & 0x07) >> 0)
31 
32 #define KINST_REX_W(r)		(((r) & 0x08) >> 3)
33 #define KINST_REX_R(r)		(((r) & 0x04) >> 2)
34 #define KINST_REX_X(r)		(((r) & 0x02) >> 1)
35 #define KINST_REX_B(r)		(((r) & 0x01) >> 0)
36 
37 #define KINST_F_CALL		0x0001	/* instruction is a "call" */
38 #define KINST_F_DIRECT_CALL	0x0002	/* instruction is a direct call */
39 #define KINST_F_RIPREL		0x0004	/* instruction is position-dependent */
40 #define KINST_F_JMP		0x0008	/* instruction is a %rip-relative jmp */
41 #define KINST_F_MOD_DIRECT	0x0010	/* operand is not a memory address */
42 
43 /*
44  * Per-CPU trampolines used when the interrupted thread is executing with
45  * interrupts disabled.  If an interrupt is raised while executing a trampoline,
46  * the interrupt thread cannot safely overwrite its trampoline if it hits a
47  * kinst probe while executing the interrupt handler.
48  */
49 DPCPU_DEFINE_STATIC(uint8_t *, intr_tramp);
50 
51 /*
52  * Map ModR/M register bits to a trapframe offset.
53  */
54 static int
55 kinst_regoff(int reg)
56 {
57 #define	_MATCH_REG(i, reg)			\
58 	case i:					\
59 		return (offsetof(struct trapframe, tf_ ## reg) / \
60 		    sizeof(register_t))
61 	switch (reg) {
62 	_MATCH_REG( 0, rax);
63 	_MATCH_REG( 1, rcx);
64 	_MATCH_REG( 2, rdx);
65 	_MATCH_REG( 3, rbx);
66 	_MATCH_REG( 4, rsp); /* SIB when mod != 3 */
67 	_MATCH_REG( 5, rbp);
68 	_MATCH_REG( 6, rsi);
69 	_MATCH_REG( 7, rdi);
70 	_MATCH_REG( 8, r8); /* REX.R is set */
71 	_MATCH_REG( 9, r9);
72 	_MATCH_REG(10, r10);
73 	_MATCH_REG(11, r11);
74 	_MATCH_REG(12, r12);
75 	_MATCH_REG(13, r13);
76 	_MATCH_REG(14, r14);
77 	_MATCH_REG(15, r15);
78 	}
79 #undef _MATCH_REG
80 	panic("%s: unhandled register index %d", __func__, reg);
81 }
82 
83 /*
84  * Obtain the specified register's value.
85  */
86 static uint64_t
87 kinst_regval(struct trapframe *frame, int reg)
88 {
89 	if (reg == -1)
90 		return (0);
91 	return (((register_t *)frame)[kinst_regoff(reg)]);
92 }
93 
94 static uint32_t
95 kinst_riprel_disp(struct kinst_probe *kp, void *dst)
96 {
97 	return ((uint32_t)((intptr_t)kp->kp_patchpoint + kp->kp_md.disp -
98 	    (intptr_t)dst));
99 }
100 
101 static void
102 kinst_trampoline_populate(struct kinst_probe *kp, uint8_t *tramp)
103 {
104 	uint8_t *instr;
105 	uint32_t disp;
106 	int ilen;
107 
108 	ilen = kp->kp_md.tinstlen;
109 
110 	memcpy(tramp, kp->kp_md.template, ilen);
111 	if ((kp->kp_md.flags & KINST_F_RIPREL) != 0) {
112 		disp = kinst_riprel_disp(kp, tramp);
113 		memcpy(&tramp[kp->kp_md.dispoff], &disp, sizeof(uint32_t));
114 	}
115 
116 	/*
117 	 * The following position-independent jmp takes us back to the
118 	 * original code.  It is encoded as "jmp *0(%rip)" (six bytes),
119 	 * followed by the absolute address of the instruction following
120 	 * the one that was traced (eight bytes).
121 	 */
122 	tramp[ilen + 0] = 0xff;
123 	tramp[ilen + 1] = 0x25;
124 	tramp[ilen + 2] = 0x00;
125 	tramp[ilen + 3] = 0x00;
126 	tramp[ilen + 4] = 0x00;
127 	tramp[ilen + 5] = 0x00;
128 	instr = kp->kp_patchpoint + kp->kp_md.instlen;
129 	memcpy(&tramp[ilen + 6], &instr, sizeof(uintptr_t));
130 }
131 
132 int
133 kinst_invop(uintptr_t addr, struct trapframe *frame, uintptr_t scratch)
134 {
135 	solaris_cpu_t *cpu;
136 	uintptr_t *stack, retaddr;
137 	struct kinst_probe *kp;
138 	struct kinst_probe_md *kpmd;
139 	uint8_t *tramp;
140 
141 	stack = (uintptr_t *)frame->tf_rsp;
142 	cpu = &solaris_cpu[curcpu];
143 
144 	LIST_FOREACH(kp, KINST_GETPROBE(addr), kp_hashnext) {
145 		if ((uintptr_t)kp->kp_patchpoint == addr)
146 			break;
147 	}
148 	if (kp == NULL)
149 		return (0);
150 
151 	/*
152 	 * Report the address of the breakpoint for the benefit of consumers
153 	 * fetching register values with regs[].
154 	 */
155 	frame->tf_rip--;
156 
157 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
158 	cpu->cpu_dtrace_caller = stack[0];
159 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT | CPU_DTRACE_BADADDR);
160 	dtrace_probe(kp->kp_id, 0, 0, 0, 0, 0);
161 	cpu->cpu_dtrace_caller = 0;
162 
163 	kpmd = &kp->kp_md;
164 	if ((kpmd->flags & KINST_F_CALL) != 0) {
165 		/*
166 		 * dtrace_invop_start() reserves space on the stack to
167 		 * store the return address of the call instruction.
168 		 */
169 		retaddr = (uintptr_t)(kp->kp_patchpoint + kpmd->instlen);
170 		*(uintptr_t *)scratch = retaddr;
171 
172 		if ((kpmd->flags & KINST_F_DIRECT_CALL) != 0) {
173 			frame->tf_rip = (uintptr_t)(kp->kp_patchpoint +
174 			    kpmd->disp + kpmd->instlen);
175 		} else {
176 			register_t rval;
177 
178 			if (kpmd->reg1 == -1 && kpmd->reg2 == -1) {
179 				/* rip-relative */
180 				rval = frame->tf_rip + kpmd->instlen;
181 			} else {
182 				/* indirect */
183 				rval = kinst_regval(frame, kpmd->reg1) +
184 				    (kinst_regval(frame, kpmd->reg2) <<
185 				    kpmd->scale);
186 			}
187 
188 			if ((kpmd->flags & KINST_F_MOD_DIRECT) != 0) {
189 				frame->tf_rip = rval + kpmd->disp;
190 			} else {
191 				frame->tf_rip =
192 				    *(uintptr_t *)(rval + kpmd->disp);
193 			}
194 		}
195 		return (DTRACE_INVOP_CALL);
196 	} else {
197 		if ((frame->tf_rflags & PSL_I) == 0)
198 			tramp = DPCPU_GET(intr_tramp);
199 		else
200 			tramp = curthread->t_kinst;
201 		if (tramp == NULL) {
202 			/*
203 			 * A trampoline allocation failed, so this probe is
204 			 * effectively disabled.  Restore the original
205 			 * instruction.
206 			 *
207 			 * We can't safely print anything here, but the
208 			 * trampoline allocator should have left a breadcrumb in
209 			 * the dmesg.
210 			 */
211 			kinst_patch_tracepoint(kp, kp->kp_savedval);
212 			frame->tf_rip = (register_t)kp->kp_patchpoint;
213 		} else {
214 			kinst_trampoline_populate(kp, tramp);
215 			frame->tf_rip = (register_t)tramp;
216 		}
217 		return (DTRACE_INVOP_NOP);
218 	}
219 }
220 
221 void
222 kinst_patch_tracepoint(struct kinst_probe *kp, kinst_patchval_t val)
223 {
224 	register_t reg;
225 	int oldwp;
226 
227 	reg = intr_disable();
228 	oldwp = disable_wp();
229 	*kp->kp_patchpoint = val;
230 	restore_wp(oldwp);
231 	intr_restore(reg);
232 }
233 
234 static void
235 kinst_set_disp8(struct kinst_probe *kp, uint8_t byte)
236 {
237 	kp->kp_md.disp = (int64_t)(int8_t)byte;
238 }
239 
240 static void
241 kinst_set_disp32(struct kinst_probe *kp, uint8_t *bytes)
242 {
243 	int32_t disp32;
244 
245 	memcpy(&disp32, bytes, sizeof(disp32));
246 	kp->kp_md.disp = (int64_t)disp32;
247 }
248 
249 static int
250 kinst_dis_get_byte(void *p)
251 {
252 	int ret;
253 	uint8_t **instr = p;
254 
255 	ret = **instr;
256 	(*instr)++;
257 
258 	return (ret);
259 }
260 
261 /*
262  * Set up all of the state needed to faithfully execute a probed instruction.
263  *
264  * In the simple case, we copy the instruction unmodified to a per-thread
265  * trampoline, wherein it is followed by a jump back to the original code.
266  * - Instructions can have %rip as an operand:
267  *   - with %rip-relative addressing encoded in ModR/M, or
268  *   - implicitly as a part of the instruction definition (jmp, call).
269  * - Call instructions (which may be %rip-relative) need to push the correct
270  *   return address onto the stack.
271  *
272  * Call instructions are simple enough to be emulated in software, so we simply
273  * do not use the trampoline mechanism in that case.  kinst_invop() will compute
274  * the branch target using the address info computed here (register operands and
275  * displacement).
276  *
277  * %rip-relative operands encoded using the ModR/M byte always use a 32-bit
278  * displacement; when populating the trampoline the displacement is adjusted to
279  * be relative to the trampoline address.  Trampolines are always allocated
280  * above KERNBASE for this reason.
281  *
282  * For other %rip-relative operands (just jumps) we take the same approach.
283  * Instructions which specify an 8-bit displacement must be rewritten to use a
284  * 32-bit displacement.
285  */
286 static int
287 kinst_instr_dissect(struct kinst_probe *kp, uint8_t **instr)
288 {
289 	struct kinst_probe_md *kpmd;
290 	dis86_t d86;
291 	uint8_t *bytes, modrm, rex;
292 	int dispoff, i, ilen, opcidx;
293 
294 	kpmd = &kp->kp_md;
295 
296 	d86.d86_data = instr;
297 	d86.d86_get_byte = kinst_dis_get_byte;
298 	d86.d86_check_func = NULL;
299 	if (dtrace_disx86(&d86, SIZE64) != 0) {
300 		KINST_LOG("failed to disassemble instruction at: %p", *instr);
301 		return (EINVAL);
302 	}
303 	bytes = d86.d86_bytes;
304 	kpmd->instlen = kpmd->tinstlen = d86.d86_len;
305 
306 	/*
307 	 * Skip over prefixes, save REX.
308 	 */
309 	rex = 0;
310 	for (i = 0; i < kpmd->instlen; i++) {
311 		switch (bytes[i]) {
312 		case 0xf0 ... 0xf3:
313 			/* group 1 */
314 			continue;
315 		case 0x26:
316 		case 0x2e:
317 		case 0x36:
318 		case 0x3e:
319 		case 0x64:
320 		case 0x65:
321 			/* group 2 */
322 			continue;
323 		case 0x66:
324 			/* group 3 */
325 			continue;
326 		case 0x67:
327 			/* group 4 */
328 			continue;
329 		case 0x40 ... 0x4f:
330 			/* REX */
331 			rex = bytes[i];
332 			continue;
333 		}
334 		break;
335 	}
336 	KASSERT(i < kpmd->instlen,
337 	    ("%s: failed to disassemble instruction at %p", __func__, bytes));
338 	opcidx = i;
339 
340 	/*
341 	 * Identify instructions of interest by opcode: calls and jumps.
342 	 * Extract displacements.
343 	 */
344 	dispoff = -1;
345 	switch (bytes[opcidx]) {
346 	case 0x0f:
347 		switch (bytes[opcidx + 1]) {
348 		case 0x80 ... 0x8f:
349 			/* conditional jmp near */
350 			kpmd->flags |= KINST_F_JMP | KINST_F_RIPREL;
351 			dispoff = opcidx + 2;
352 			kinst_set_disp32(kp, &bytes[dispoff]);
353 			break;
354 		}
355 		break;
356 	case 0xe3:
357 		/*
358 		 * There is no straightforward way to translate this instruction
359 		 * to use a 32-bit displacement.  Fortunately, it is rarely
360 		 * used.
361 		 */
362 		return (EINVAL);
363 	case 0x70 ... 0x7f:
364 		/* conditional jmp short */
365 		kpmd->flags |= KINST_F_JMP | KINST_F_RIPREL;
366 		dispoff = opcidx + 1;
367 		kinst_set_disp8(kp, bytes[dispoff]);
368 		break;
369 	case 0xe9:
370 		/* unconditional jmp near */
371 		kpmd->flags |= KINST_F_JMP | KINST_F_RIPREL;
372 		dispoff = opcidx + 1;
373 		kinst_set_disp32(kp, &bytes[dispoff]);
374 		break;
375 	case 0xeb:
376 		/* unconditional jmp short */
377 		kpmd->flags |= KINST_F_JMP | KINST_F_RIPREL;
378 		dispoff = opcidx + 1;
379 		kinst_set_disp8(kp, bytes[dispoff]);
380 		break;
381 	case 0xe8:
382 	case 0x9a:
383 		/* direct call */
384 		kpmd->flags |= KINST_F_CALL | KINST_F_DIRECT_CALL;
385 		dispoff = opcidx + 1;
386 		kinst_set_disp32(kp, &bytes[dispoff]);
387 		break;
388 	case 0xff:
389 		KASSERT(d86.d86_got_modrm,
390 		    ("no ModR/M byte for instr at %p", *instr - kpmd->instlen));
391 		switch (KINST_MODRM_REG(bytes[d86.d86_rmindex])) {
392 		case 0x02:
393 		case 0x03:
394 			/* indirect call */
395 			kpmd->flags |= KINST_F_CALL;
396 			break;
397 		case 0x04:
398 		case 0x05:
399 			/* indirect jump */
400 			kpmd->flags |= KINST_F_JMP;
401 			break;
402 		}
403 	}
404 
405 	/*
406 	 * If there's a ModR/M byte, we need to check it to see if the operand
407 	 * is %rip-relative, and rewrite the displacement if so.  If not, we
408 	 * might still have to extract operand info if this is a call
409 	 * instruction.
410 	 */
411 	if (d86.d86_got_modrm) {
412 		uint8_t mod, rm, sib;
413 
414 		kpmd->reg1 = kpmd->reg2 = -1;
415 
416 		modrm = bytes[d86.d86_rmindex];
417 		mod = KINST_MODRM_MOD(modrm);
418 		rm = KINST_MODRM_RM(modrm);
419 		if (mod == 0 && rm == 5) {
420 			kpmd->flags |= KINST_F_RIPREL;
421 			dispoff = d86.d86_rmindex + 1;
422 			kinst_set_disp32(kp, &bytes[dispoff]);
423 		} else if ((kpmd->flags & KINST_F_CALL) != 0) {
424 			bool havesib;
425 
426 			havesib = (mod != 3 && rm == 4);
427 			dispoff = d86.d86_rmindex + (havesib ? 2 : 1);
428 			if (mod == 1)
429 				kinst_set_disp8(kp, bytes[dispoff]);
430 			else if (mod == 2)
431 				kinst_set_disp32(kp, &bytes[dispoff]);
432 			else if (mod == 3)
433 				kpmd->flags |= KINST_F_MOD_DIRECT;
434 
435 			if (havesib) {
436 				sib = bytes[d86.d86_rmindex + 1];
437 				if (KINST_SIB_BASE(sib) != 5) {
438 					kpmd->reg1 = KINST_SIB_BASE(sib) |
439 					    (KINST_REX_B(rex) << 3);
440 				}
441 				kpmd->scale = KINST_SIB_SCALE(sib);
442 				kpmd->reg2 = KINST_SIB_INDEX(sib) |
443 				    (KINST_REX_X(rex) << 3);
444 			} else {
445 				kpmd->reg1 = rm | (KINST_REX_B(rex) << 3);
446 			}
447 		}
448 	}
449 
450 	/*
451 	 * Calls are emulated in software; once operands are decoded we have
452 	 * nothing else to do.
453 	 */
454 	if ((kpmd->flags & KINST_F_CALL) != 0)
455 		return (0);
456 
457 	/*
458 	 * Allocate and populate an instruction trampoline template.
459 	 *
460 	 * Position-independent instructions can simply be copied, but
461 	 * position-dependent instructions require some surgery: jump
462 	 * instructions with an 8-bit displacement need to be converted to use a
463 	 * 32-bit displacement, and the adjusted displacement needs to be
464 	 * computed.
465 	 */
466 	ilen = kpmd->instlen;
467 	if ((kpmd->flags & KINST_F_RIPREL) != 0) {
468 		if ((kpmd->flags & KINST_F_JMP) == 0 ||
469 		    bytes[opcidx] == 0x0f ||
470 		    bytes[opcidx] == 0xe9 ||
471 		    bytes[opcidx] == 0xff) {
472 			memcpy(kpmd->template, bytes, dispoff);
473 			memcpy(&kpmd->template[dispoff + 4],
474 			    &bytes[dispoff + 4], ilen - (dispoff + 4));
475 			kpmd->dispoff = dispoff;
476 		} else if (bytes[opcidx] == 0xeb) {
477 			memcpy(kpmd->template, bytes, opcidx);
478 			kpmd->template[opcidx] = 0xe9;
479 			kpmd->dispoff = opcidx + 1;
480 
481 			/* Instruction length changes from 2 to 5. */
482 			kpmd->tinstlen = 5;
483 			kpmd->disp -= 3;
484 		} else if (bytes[opcidx] >= 0x70 && bytes[opcidx] <= 0x7f)  {
485 			memcpy(kpmd->template, bytes, opcidx);
486 			kpmd->template[opcidx] = 0x0f;
487 			kpmd->template[opcidx + 1] = bytes[opcidx] + 0x10;
488 			kpmd->dispoff = opcidx + 2;
489 
490 			/* Instruction length changes from 2 to 6. */
491 			kpmd->tinstlen = 6;
492 			kpmd->disp -= 4;
493 		} else {
494 			panic("unhandled opcode %#x", bytes[opcidx]);
495 		}
496 	} else {
497 		memcpy(kpmd->template, bytes, ilen);
498 	}
499 
500 	return (0);
501 }
502 
503 int
504 kinst_make_probe(linker_file_t lf, int symindx, linker_symval_t *symval,
505     void *opaque)
506 {
507 	struct kinst_probe *kp;
508 	dtrace_kinst_probedesc_t *pd;
509 	const char *func;
510 	int error, instrsize, n, off;
511 	uint8_t *instr, *limit;
512 
513 	pd = opaque;
514 	func = symval->name;
515 	if (strcmp(func, pd->kpd_func) != 0 || strcmp(func, "trap_check") == 0)
516 		return (0);
517 
518 	instr = (uint8_t *)symval->value;
519 	limit = (uint8_t *)symval->value + symval->size;
520 	if (instr >= limit)
521 		return (0);
522 
523 	/*
524 	 * Ignore functions not beginning with the usual function prologue.
525 	 * These might correspond to exception handlers with which we should not
526 	 * meddle.  This does however exclude functions which can be safely
527 	 * traced, such as cpu_switch().
528 	 */
529 	if (*instr != KINST_PUSHL_RBP)
530 		return (0);
531 
532 	n = 0;
533 	while (instr < limit) {
534 		instrsize = dtrace_instr_size(instr);
535 		off = (int)(instr - (uint8_t *)symval->value);
536 		if (pd->kpd_off != -1 && off != pd->kpd_off) {
537 			instr += instrsize;
538 			continue;
539 		}
540 
541 		/*
542 		 * Check for instructions which may enable interrupts.  Such
543 		 * instructions are tricky to trace since it is unclear whether
544 		 * to use the per-thread or per-CPU trampolines.  Since they are
545 		 * rare, we don't bother to implement special handling for them.
546 		 *
547 		 * If the caller specified an offset, return an error, otherwise
548 		 * silently ignore the instruction so that it remains possible
549 		 * to enable all instructions in a function.
550 		 */
551 		if (instrsize == 1 &&
552 		    (instr[0] == KINST_POPF || instr[0] == KINST_STI)) {
553 			if (pd->kpd_off != -1)
554 				return (EINVAL);
555 			instr += instrsize;
556 			continue;
557 		}
558 
559 		/*
560 		 * Prevent separate dtrace(1) instances from creating copies of
561 		 * the same probe.
562 		 */
563 		LIST_FOREACH(kp, KINST_GETPROBE(instr), kp_hashnext) {
564 			if (strcmp(kp->kp_func, func) == 0 &&
565 			    strtol(kp->kp_name, NULL, 10) == off)
566 				return (0);
567 		}
568 		if (++n > KINST_PROBETAB_MAX) {
569 			KINST_LOG("probe list full: %d entries", n);
570 			return (ENOMEM);
571 		}
572 		kp = malloc(sizeof(struct kinst_probe), M_KINST,
573 		    M_WAITOK | M_ZERO);
574 		kp->kp_func = func;
575 		snprintf(kp->kp_name, sizeof(kp->kp_name), "%d", off);
576 		kp->kp_savedval = *instr;
577 		kp->kp_patchval = KINST_PATCHVAL;
578 		kp->kp_patchpoint = instr;
579 
580 		error = kinst_instr_dissect(kp, &instr);
581 		if (error != 0)
582 			return (error);
583 
584 		kinst_probe_create(kp, lf);
585 	}
586 
587 	return (0);
588 }
589 
590 int
591 kinst_md_init(void)
592 {
593 	uint8_t *tramp;
594 	int cpu;
595 
596 	CPU_FOREACH(cpu) {
597 		tramp = kinst_trampoline_alloc(M_WAITOK);
598 		if (tramp == NULL)
599 			return (ENOMEM);
600 		DPCPU_ID_SET(cpu, intr_tramp, tramp);
601 	}
602 
603 	return (0);
604 }
605 
606 void
607 kinst_md_deinit(void)
608 {
609 	uint8_t *tramp;
610 	int cpu;
611 
612 	CPU_FOREACH(cpu) {
613 		tramp = DPCPU_ID_GET(cpu, intr_tramp);
614 		if (tramp != NULL) {
615 			kinst_trampoline_dealloc(tramp);
616 			DPCPU_ID_SET(cpu, intr_tramp, NULL);
617 		}
618 	}
619 }
620