]> CyberLeo.Net >> Repos - FreeBSD/releng/9.0.git/blob - sys/cddl/contrib/opensolaris/uts/intel/dtrace/fasttrap_isa.c
MFC r227290 and r227291:
[FreeBSD/releng/9.0.git] / sys / cddl / contrib / opensolaris / uts / intel / dtrace / fasttrap_isa.c
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  *
21  * Portions Copyright 2010 The FreeBSD Foundation
22  *
23  * $FreeBSD$
24  */
25
26 /*
27  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
28  * Use is subject to license terms.
29  */
30
31 #if defined(sun)
32 #pragma ident   "%Z%%M% %I%     %E% SMI"
33 #endif
34
35 #include <sys/fasttrap_isa.h>
36 #include <sys/fasttrap_impl.h>
37 #include <sys/dtrace.h>
38 #include <sys/dtrace_impl.h>
39 #include <sys/cmn_err.h>
40 #if defined(sun)
41 #include <sys/regset.h>
42 #include <sys/privregs.h>
43 #include <sys/segments.h>
44 #include <sys/x86_archext.h>
45 #else
46 #include <cddl/dev/dtrace/dtrace_cddl.h>
47 #include <sys/types.h>
48 #include <sys/proc.h>
49 #include <sys/dtrace_bsd.h>
50 #include <cddl/dev/dtrace/i386/regset.h>
51 #include <machine/segments.h>
52 #include <machine/reg.h>
53 #include <machine/pcb.h>
54 #endif
55 #include <sys/sysmacros.h>
56 #if defined(sun)
57 #include <sys/trap.h>
58 #include <sys/archsystm.h>
59 #else
60 #include <sys/ptrace.h>
61
62 static int
63 proc_ops(int op, proc_t *p, void *kaddr, off_t uaddr, size_t len)
64 {
65         struct iovec iov;
66         struct uio uio;
67
68         iov.iov_base = kaddr;
69         iov.iov_len = len;
70         uio.uio_offset = uaddr;
71         uio.uio_iov = &iov;
72         uio.uio_resid = len;
73         uio.uio_iovcnt = 1;
74         uio.uio_segflg = UIO_SYSSPACE;
75         uio.uio_td = curthread;
76         uio.uio_rw = op;
77         PHOLD(p);
78         if (proc_rwmem(p, &uio) < 0) {
79                 PRELE(p);
80                 return (-1);
81         }
82         PRELE(p);
83
84         return (0);
85 }
86
87 static int
88 uread(proc_t *p, void *kaddr, size_t len, uintptr_t uaddr)
89 {
90
91         return (proc_ops(UIO_READ, p, kaddr, uaddr, len));
92 }
93
94 static int
95 uwrite(proc_t *p, void *kaddr, size_t len, uintptr_t uaddr)
96 {
97
98         return (proc_ops(UIO_WRITE, p, kaddr, uaddr, len));
99 }
100 #endif /* sun */
101 #ifdef __i386__
102 #define r_rax   r_eax
103 #define r_rbx   r_ebx
104 #define r_rip   r_eip
105 #define r_rflags r_eflags
106 #define r_rsp   r_esp
107 #endif
108
109 /*
110  * Lossless User-Land Tracing on x86
111  * ---------------------------------
112  *
113  * The execution of most instructions is not dependent on the address; for
114  * these instructions it is sufficient to copy them into the user process's
115  * address space and execute them. To effectively single-step an instruction
116  * in user-land, we copy out the following sequence of instructions to scratch
117  * space in the user thread's ulwp_t structure.
118  *
119  * We then set the program counter (%eip or %rip) to point to this scratch
120  * space. Once execution resumes, the original instruction is executed and
121  * then control flow is redirected to what was originally the subsequent
122  * instruction. If the kernel attemps to deliver a signal while single-
123  * stepping, the signal is deferred and the program counter is moved into the
124  * second sequence of instructions. The second sequence ends in a trap into
125  * the kernel where the deferred signal is then properly handled and delivered.
126  *
127  * For instructions whose execute is position dependent, we perform simple
128  * emulation. These instructions are limited to control transfer
129  * instructions in 32-bit mode, but in 64-bit mode there's the added wrinkle
130  * of %rip-relative addressing that means that almost any instruction can be
131  * position dependent. For all the details on how we emulate generic
132  * instructions included %rip-relative instructions, see the code in
133  * fasttrap_pid_probe() below where we handle instructions of type
134  * FASTTRAP_T_COMMON (under the header: Generic Instruction Tracing).
135  */
136
137 #define FASTTRAP_MODRM_MOD(modrm)       (((modrm) >> 6) & 0x3)
138 #define FASTTRAP_MODRM_REG(modrm)       (((modrm) >> 3) & 0x7)
139 #define FASTTRAP_MODRM_RM(modrm)        ((modrm) & 0x7)
140 #define FASTTRAP_MODRM(mod, reg, rm)    (((mod) << 6) | ((reg) << 3) | (rm))
141
142 #define FASTTRAP_SIB_SCALE(sib)         (((sib) >> 6) & 0x3)
143 #define FASTTRAP_SIB_INDEX(sib)         (((sib) >> 3) & 0x7)
144 #define FASTTRAP_SIB_BASE(sib)          ((sib) & 0x7)
145
146 #define FASTTRAP_REX_W(rex)             (((rex) >> 3) & 1)
147 #define FASTTRAP_REX_R(rex)             (((rex) >> 2) & 1)
148 #define FASTTRAP_REX_X(rex)             (((rex) >> 1) & 1)
149 #define FASTTRAP_REX_B(rex)             ((rex) & 1)
150 #define FASTTRAP_REX(w, r, x, b)        \
151         (0x40 | ((w) << 3) | ((r) << 2) | ((x) << 1) | (b))
152
153 /*
154  * Single-byte op-codes.
155  */
156 #define FASTTRAP_PUSHL_EBP      0x55
157
158 #define FASTTRAP_JO             0x70
159 #define FASTTRAP_JNO            0x71
160 #define FASTTRAP_JB             0x72
161 #define FASTTRAP_JAE            0x73
162 #define FASTTRAP_JE             0x74
163 #define FASTTRAP_JNE            0x75
164 #define FASTTRAP_JBE            0x76
165 #define FASTTRAP_JA             0x77
166 #define FASTTRAP_JS             0x78
167 #define FASTTRAP_JNS            0x79
168 #define FASTTRAP_JP             0x7a
169 #define FASTTRAP_JNP            0x7b
170 #define FASTTRAP_JL             0x7c
171 #define FASTTRAP_JGE            0x7d
172 #define FASTTRAP_JLE            0x7e
173 #define FASTTRAP_JG             0x7f
174
175 #define FASTTRAP_NOP            0x90
176
177 #define FASTTRAP_MOV_EAX        0xb8
178 #define FASTTRAP_MOV_ECX        0xb9
179
180 #define FASTTRAP_RET16          0xc2
181 #define FASTTRAP_RET            0xc3
182
183 #define FASTTRAP_LOOPNZ         0xe0
184 #define FASTTRAP_LOOPZ          0xe1
185 #define FASTTRAP_LOOP           0xe2
186 #define FASTTRAP_JCXZ           0xe3
187
188 #define FASTTRAP_CALL           0xe8
189 #define FASTTRAP_JMP32          0xe9
190 #define FASTTRAP_JMP8           0xeb
191
192 #define FASTTRAP_INT3           0xcc
193 #define FASTTRAP_INT            0xcd
194
195 #define FASTTRAP_2_BYTE_OP      0x0f
196 #define FASTTRAP_GROUP5_OP      0xff
197
198 /*
199  * Two-byte op-codes (second byte only).
200  */
201 #define FASTTRAP_0F_JO          0x80
202 #define FASTTRAP_0F_JNO         0x81
203 #define FASTTRAP_0F_JB          0x82
204 #define FASTTRAP_0F_JAE         0x83
205 #define FASTTRAP_0F_JE          0x84
206 #define FASTTRAP_0F_JNE         0x85
207 #define FASTTRAP_0F_JBE         0x86
208 #define FASTTRAP_0F_JA          0x87
209 #define FASTTRAP_0F_JS          0x88
210 #define FASTTRAP_0F_JNS         0x89
211 #define FASTTRAP_0F_JP          0x8a
212 #define FASTTRAP_0F_JNP         0x8b
213 #define FASTTRAP_0F_JL          0x8c
214 #define FASTTRAP_0F_JGE         0x8d
215 #define FASTTRAP_0F_JLE         0x8e
216 #define FASTTRAP_0F_JG          0x8f
217
218 #define FASTTRAP_EFLAGS_OF      0x800
219 #define FASTTRAP_EFLAGS_DF      0x400
220 #define FASTTRAP_EFLAGS_SF      0x080
221 #define FASTTRAP_EFLAGS_ZF      0x040
222 #define FASTTRAP_EFLAGS_AF      0x010
223 #define FASTTRAP_EFLAGS_PF      0x004
224 #define FASTTRAP_EFLAGS_CF      0x001
225
226 /*
227  * Instruction prefixes.
228  */
229 #define FASTTRAP_PREFIX_OPERAND 0x66
230 #define FASTTRAP_PREFIX_ADDRESS 0x67
231 #define FASTTRAP_PREFIX_CS      0x2E
232 #define FASTTRAP_PREFIX_DS      0x3E
233 #define FASTTRAP_PREFIX_ES      0x26
234 #define FASTTRAP_PREFIX_FS      0x64
235 #define FASTTRAP_PREFIX_GS      0x65
236 #define FASTTRAP_PREFIX_SS      0x36
237 #define FASTTRAP_PREFIX_LOCK    0xF0
238 #define FASTTRAP_PREFIX_REP     0xF3
239 #define FASTTRAP_PREFIX_REPNE   0xF2
240
241 #define FASTTRAP_NOREG  0xff
242
243 /*
244  * Map between instruction register encodings and the kernel constants which
245  * correspond to indicies into struct regs.
246  */
247 #ifdef __amd64
248 static const uint8_t regmap[16] = {
249         REG_RAX, REG_RCX, REG_RDX, REG_RBX, REG_RSP, REG_RBP, REG_RSI, REG_RDI,
250         REG_R8, REG_R9, REG_R10, REG_R11, REG_R12, REG_R13, REG_R14, REG_R15,
251 };
252 #else
253 static const uint8_t regmap[8] = {
254         EAX, ECX, EDX, EBX, UESP, EBP, ESI, EDI
255 };
256 #endif
257
258 static ulong_t fasttrap_getreg(struct reg *, uint_t);
259
260 static uint64_t
261 fasttrap_anarg(struct reg *rp, int function_entry, int argno)
262 {
263         uint64_t value = 0;
264         int shift = function_entry ? 1 : 0;
265
266 #ifdef __amd64
267         if (curproc->p_model == DATAMODEL_LP64) {
268                 uintptr_t *stack;
269
270                 /*
271                  * In 64-bit mode, the first six arguments are stored in
272                  * registers.
273                  */
274                 if (argno < 6)
275                         return ((&rp->r_rdi)[argno]);
276
277                 stack = (uintptr_t *)rp->r_rsp;
278                 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
279                 value = dtrace_fulword(&stack[argno - 6 + shift]);
280                 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT | CPU_DTRACE_BADADDR);
281         } else {
282 #endif
283 #ifdef __i386
284                 uint32_t *stack = (uint32_t *)rp->r_esp;
285                 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
286                 value = dtrace_fuword32(&stack[argno + shift]);
287                 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT | CPU_DTRACE_BADADDR);
288 #endif
289 #ifdef __amd64
290         }
291 #endif
292
293         return (value);
294 }
295
296 /*ARGSUSED*/
297 int
298 fasttrap_tracepoint_init(proc_t *p, fasttrap_tracepoint_t *tp, uintptr_t pc,
299     fasttrap_probe_type_t type)
300 {
301         uint8_t instr[FASTTRAP_MAX_INSTR_SIZE + 10];
302         size_t len = FASTTRAP_MAX_INSTR_SIZE;
303         size_t first = MIN(len, PAGESIZE - (pc & PAGEOFFSET));
304         uint_t start = 0;
305         int rmindex, size;
306         uint8_t seg, rex = 0;
307
308         /*
309          * Read the instruction at the given address out of the process's
310          * address space. We don't have to worry about a debugger
311          * changing this instruction before we overwrite it with our trap
312          * instruction since P_PR_LOCK is set. Since instructions can span
313          * pages, we potentially read the instruction in two parts. If the
314          * second part fails, we just zero out that part of the instruction.
315          */
316         if (uread(p, &instr[0], first, pc) != 0)
317                 return (-1);
318         if (len > first &&
319             uread(p, &instr[first], len - first, pc + first) != 0) {
320                 bzero(&instr[first], len - first);
321                 len = first;
322         }
323
324         /*
325          * If the disassembly fails, then we have a malformed instruction.
326          */
327         if ((size = dtrace_instr_size_isa(instr, p->p_model, &rmindex)) <= 0)
328                 return (-1);
329
330         /*
331          * Make sure the disassembler isn't completely broken.
332          */
333         ASSERT(-1 <= rmindex && rmindex < size);
334
335         /*
336          * If the computed size is greater than the number of bytes read,
337          * then it was a malformed instruction possibly because it fell on a
338          * page boundary and the subsequent page was missing or because of
339          * some malicious user.
340          */
341         if (size > len)
342                 return (-1);
343
344         tp->ftt_size = (uint8_t)size;
345         tp->ftt_segment = FASTTRAP_SEG_NONE;
346
347         /*
348          * Find the start of the instruction's opcode by processing any
349          * legacy prefixes.
350          */
351         for (;;) {
352                 seg = 0;
353                 switch (instr[start]) {
354                 case FASTTRAP_PREFIX_SS:
355                         seg++;
356                         /*FALLTHRU*/
357                 case FASTTRAP_PREFIX_GS:
358                         seg++;
359                         /*FALLTHRU*/
360                 case FASTTRAP_PREFIX_FS:
361                         seg++;
362                         /*FALLTHRU*/
363                 case FASTTRAP_PREFIX_ES:
364                         seg++;
365                         /*FALLTHRU*/
366                 case FASTTRAP_PREFIX_DS:
367                         seg++;
368                         /*FALLTHRU*/
369                 case FASTTRAP_PREFIX_CS:
370                         seg++;
371                         /*FALLTHRU*/
372                 case FASTTRAP_PREFIX_OPERAND:
373                 case FASTTRAP_PREFIX_ADDRESS:
374                 case FASTTRAP_PREFIX_LOCK:
375                 case FASTTRAP_PREFIX_REP:
376                 case FASTTRAP_PREFIX_REPNE:
377                         if (seg != 0) {
378                                 /*
379                                  * It's illegal for an instruction to specify
380                                  * two segment prefixes -- give up on this
381                                  * illegal instruction.
382                                  */
383                                 if (tp->ftt_segment != FASTTRAP_SEG_NONE)
384                                         return (-1);
385
386                                 tp->ftt_segment = seg;
387                         }
388                         start++;
389                         continue;
390                 }
391                 break;
392         }
393
394 #ifdef __amd64
395         /*
396          * Identify the REX prefix on 64-bit processes.
397          */
398         if (p->p_model == DATAMODEL_LP64 && (instr[start] & 0xf0) == 0x40)
399                 rex = instr[start++];
400 #endif
401
402         /*
403          * Now that we're pretty sure that the instruction is okay, copy the
404          * valid part to the tracepoint.
405          */
406         bcopy(instr, tp->ftt_instr, FASTTRAP_MAX_INSTR_SIZE);
407
408         tp->ftt_type = FASTTRAP_T_COMMON;
409         if (instr[start] == FASTTRAP_2_BYTE_OP) {
410                 switch (instr[start + 1]) {
411                 case FASTTRAP_0F_JO:
412                 case FASTTRAP_0F_JNO:
413                 case FASTTRAP_0F_JB:
414                 case FASTTRAP_0F_JAE:
415                 case FASTTRAP_0F_JE:
416                 case FASTTRAP_0F_JNE:
417                 case FASTTRAP_0F_JBE:
418                 case FASTTRAP_0F_JA:
419                 case FASTTRAP_0F_JS:
420                 case FASTTRAP_0F_JNS:
421                 case FASTTRAP_0F_JP:
422                 case FASTTRAP_0F_JNP:
423                 case FASTTRAP_0F_JL:
424                 case FASTTRAP_0F_JGE:
425                 case FASTTRAP_0F_JLE:
426                 case FASTTRAP_0F_JG:
427                         tp->ftt_type = FASTTRAP_T_JCC;
428                         tp->ftt_code = (instr[start + 1] & 0x0f) | FASTTRAP_JO;
429                         tp->ftt_dest = pc + tp->ftt_size +
430                             /* LINTED - alignment */
431                             *(int32_t *)&instr[start + 2];
432                         break;
433                 }
434         } else if (instr[start] == FASTTRAP_GROUP5_OP) {
435                 uint_t mod = FASTTRAP_MODRM_MOD(instr[start + 1]);
436                 uint_t reg = FASTTRAP_MODRM_REG(instr[start + 1]);
437                 uint_t rm = FASTTRAP_MODRM_RM(instr[start + 1]);
438
439                 if (reg == 2 || reg == 4) {
440                         uint_t i, sz;
441
442                         if (reg == 2)
443                                 tp->ftt_type = FASTTRAP_T_CALL;
444                         else
445                                 tp->ftt_type = FASTTRAP_T_JMP;
446
447                         if (mod == 3)
448                                 tp->ftt_code = 2;
449                         else
450                                 tp->ftt_code = 1;
451
452                         ASSERT(p->p_model == DATAMODEL_LP64 || rex == 0);
453
454                         /*
455                          * See AMD x86-64 Architecture Programmer's Manual
456                          * Volume 3, Section 1.2.7, Table 1-12, and
457                          * Appendix A.3.1, Table A-15.
458                          */
459                         if (mod != 3 && rm == 4) {
460                                 uint8_t sib = instr[start + 2];
461                                 uint_t index = FASTTRAP_SIB_INDEX(sib);
462                                 uint_t base = FASTTRAP_SIB_BASE(sib);
463
464                                 tp->ftt_scale = FASTTRAP_SIB_SCALE(sib);
465
466                                 tp->ftt_index = (index == 4) ?
467                                     FASTTRAP_NOREG :
468                                     regmap[index | (FASTTRAP_REX_X(rex) << 3)];
469                                 tp->ftt_base = (mod == 0 && base == 5) ?
470                                     FASTTRAP_NOREG :
471                                     regmap[base | (FASTTRAP_REX_B(rex) << 3)];
472
473                                 i = 3;
474                                 sz = mod == 1 ? 1 : 4;
475                         } else {
476                                 /*
477                                  * In 64-bit mode, mod == 0 and r/m == 5
478                                  * denotes %rip-relative addressing; in 32-bit
479                                  * mode, the base register isn't used. In both
480                                  * modes, there is a 32-bit operand.
481                                  */
482                                 if (mod == 0 && rm == 5) {
483 #ifdef __amd64
484                                         if (p->p_model == DATAMODEL_LP64)
485                                                 tp->ftt_base = REG_RIP;
486                                         else
487 #endif
488                                                 tp->ftt_base = FASTTRAP_NOREG;
489                                         sz = 4;
490                                 } else  {
491                                         uint8_t base = rm |
492                                             (FASTTRAP_REX_B(rex) << 3);
493
494                                         tp->ftt_base = regmap[base];
495                                         sz = mod == 1 ? 1 : mod == 2 ? 4 : 0;
496                                 }
497                                 tp->ftt_index = FASTTRAP_NOREG;
498                                 i = 2;
499                         }
500
501                         if (sz == 1) {
502                                 tp->ftt_dest = *(int8_t *)&instr[start + i];
503                         } else if (sz == 4) {
504                                 /* LINTED - alignment */
505                                 tp->ftt_dest = *(int32_t *)&instr[start + i];
506                         } else {
507                                 tp->ftt_dest = 0;
508                         }
509                 }
510         } else {
511                 switch (instr[start]) {
512                 case FASTTRAP_RET:
513                         tp->ftt_type = FASTTRAP_T_RET;
514                         break;
515
516                 case FASTTRAP_RET16:
517                         tp->ftt_type = FASTTRAP_T_RET16;
518                         /* LINTED - alignment */
519                         tp->ftt_dest = *(uint16_t *)&instr[start + 1];
520                         break;
521
522                 case FASTTRAP_JO:
523                 case FASTTRAP_JNO:
524                 case FASTTRAP_JB:
525                 case FASTTRAP_JAE:
526                 case FASTTRAP_JE:
527                 case FASTTRAP_JNE:
528                 case FASTTRAP_JBE:
529                 case FASTTRAP_JA:
530                 case FASTTRAP_JS:
531                 case FASTTRAP_JNS:
532                 case FASTTRAP_JP:
533                 case FASTTRAP_JNP:
534                 case FASTTRAP_JL:
535                 case FASTTRAP_JGE:
536                 case FASTTRAP_JLE:
537                 case FASTTRAP_JG:
538                         tp->ftt_type = FASTTRAP_T_JCC;
539                         tp->ftt_code = instr[start];
540                         tp->ftt_dest = pc + tp->ftt_size +
541                             (int8_t)instr[start + 1];
542                         break;
543
544                 case FASTTRAP_LOOPNZ:
545                 case FASTTRAP_LOOPZ:
546                 case FASTTRAP_LOOP:
547                         tp->ftt_type = FASTTRAP_T_LOOP;
548                         tp->ftt_code = instr[start];
549                         tp->ftt_dest = pc + tp->ftt_size +
550                             (int8_t)instr[start + 1];
551                         break;
552
553                 case FASTTRAP_JCXZ:
554                         tp->ftt_type = FASTTRAP_T_JCXZ;
555                         tp->ftt_dest = pc + tp->ftt_size +
556                             (int8_t)instr[start + 1];
557                         break;
558
559                 case FASTTRAP_CALL:
560                         tp->ftt_type = FASTTRAP_T_CALL;
561                         tp->ftt_dest = pc + tp->ftt_size +
562                             /* LINTED - alignment */
563                             *(int32_t *)&instr[start + 1];
564                         tp->ftt_code = 0;
565                         break;
566
567                 case FASTTRAP_JMP32:
568                         tp->ftt_type = FASTTRAP_T_JMP;
569                         tp->ftt_dest = pc + tp->ftt_size +
570                             /* LINTED - alignment */
571                             *(int32_t *)&instr[start + 1];
572                         break;
573                 case FASTTRAP_JMP8:
574                         tp->ftt_type = FASTTRAP_T_JMP;
575                         tp->ftt_dest = pc + tp->ftt_size +
576                             (int8_t)instr[start + 1];
577                         break;
578
579                 case FASTTRAP_PUSHL_EBP:
580                         if (start == 0)
581                                 tp->ftt_type = FASTTRAP_T_PUSHL_EBP;
582                         break;
583
584                 case FASTTRAP_NOP:
585 #ifdef __amd64
586                         ASSERT(p->p_model == DATAMODEL_LP64 || rex == 0);
587
588                         /*
589                          * On amd64 we have to be careful not to confuse a nop
590                          * (actually xchgl %eax, %eax) with an instruction using
591                          * the same opcode, but that does something different
592                          * (e.g. xchgl %r8d, %eax or xcghq %r8, %rax).
593                          */
594                         if (FASTTRAP_REX_B(rex) == 0)
595 #endif
596                                 tp->ftt_type = FASTTRAP_T_NOP;
597                         break;
598
599                 case FASTTRAP_INT3:
600                         /*
601                          * The pid provider shares the int3 trap with debugger
602                          * breakpoints so we can't instrument them.
603                          */
604                         ASSERT(instr[start] == FASTTRAP_INSTR);
605                         return (-1);
606
607                 case FASTTRAP_INT:
608                         /*
609                          * Interrupts seem like they could be traced with
610                          * no negative implications, but it's possible that
611                          * a thread could be redirected by the trap handling
612                          * code which would eventually return to the
613                          * instruction after the interrupt. If the interrupt
614                          * were in our scratch space, the subsequent
615                          * instruction might be overwritten before we return.
616                          * Accordingly we refuse to instrument any interrupt.
617                          */
618                         return (-1);
619                 }
620         }
621
622 #ifdef __amd64
623         if (p->p_model == DATAMODEL_LP64 && tp->ftt_type == FASTTRAP_T_COMMON) {
624                 /*
625                  * If the process is 64-bit and the instruction type is still
626                  * FASTTRAP_T_COMMON -- meaning we're going to copy it out an
627                  * execute it -- we need to watch for %rip-relative
628                  * addressing mode. See the portion of fasttrap_pid_probe()
629                  * below where we handle tracepoints with type
630                  * FASTTRAP_T_COMMON for how we emulate instructions that
631                  * employ %rip-relative addressing.
632                  */
633                 if (rmindex != -1) {
634                         uint_t mod = FASTTRAP_MODRM_MOD(instr[rmindex]);
635                         uint_t reg = FASTTRAP_MODRM_REG(instr[rmindex]);
636                         uint_t rm = FASTTRAP_MODRM_RM(instr[rmindex]);
637
638                         ASSERT(rmindex > start);
639
640                         if (mod == 0 && rm == 5) {
641                                 /*
642                                  * We need to be sure to avoid other
643                                  * registers used by this instruction. While
644                                  * the reg field may determine the op code
645                                  * rather than denoting a register, assuming
646                                  * that it denotes a register is always safe.
647                                  * We leave the REX field intact and use
648                                  * whatever value's there for simplicity.
649                                  */
650                                 if (reg != 0) {
651                                         tp->ftt_ripmode = FASTTRAP_RIP_1 |
652                                             (FASTTRAP_RIP_X *
653                                             FASTTRAP_REX_B(rex));
654                                         rm = 0;
655                                 } else {
656                                         tp->ftt_ripmode = FASTTRAP_RIP_2 |
657                                             (FASTTRAP_RIP_X *
658                                             FASTTRAP_REX_B(rex));
659                                         rm = 1;
660                                 }
661
662                                 tp->ftt_modrm = tp->ftt_instr[rmindex];
663                                 tp->ftt_instr[rmindex] =
664                                     FASTTRAP_MODRM(2, reg, rm);
665                         }
666                 }
667         }
668 #endif
669
670         return (0);
671 }
672
673 int
674 fasttrap_tracepoint_install(proc_t *p, fasttrap_tracepoint_t *tp)
675 {
676         fasttrap_instr_t instr = FASTTRAP_INSTR;
677
678         if (uwrite(p, &instr, 1, tp->ftt_pc) != 0)
679                 return (-1);
680
681         return (0);
682 }
683
684 int
685 fasttrap_tracepoint_remove(proc_t *p, fasttrap_tracepoint_t *tp)
686 {
687         uint8_t instr;
688
689         /*
690          * Distinguish between read or write failures and a changed
691          * instruction.
692          */
693         if (uread(p, &instr, 1, tp->ftt_pc) != 0)
694                 return (0);
695         if (instr != FASTTRAP_INSTR)
696                 return (0);
697         if (uwrite(p, &tp->ftt_instr[0], 1, tp->ftt_pc) != 0)
698                 return (-1);
699
700         return (0);
701 }
702
703 #ifdef __amd64
704 static uintptr_t
705 fasttrap_fulword_noerr(const void *uaddr)
706 {
707         uintptr_t ret;
708
709         if ((ret = fasttrap_fulword(uaddr)) != -1)
710                 return (ret);
711
712         return (0);
713 }
714 #endif
715
716 #ifdef __i386__
717 static uint32_t
718 fasttrap_fuword32_noerr(const void *uaddr)
719 {
720         uint32_t ret;
721
722         if ((ret = fasttrap_fuword32(uaddr)) != -1)
723                 return (ret);
724
725         return (0);
726 }
727 #endif
728
729 static void
730 fasttrap_return_common(struct reg *rp, uintptr_t pc, pid_t pid,
731     uintptr_t new_pc)
732 {
733         fasttrap_tracepoint_t *tp;
734         fasttrap_bucket_t *bucket;
735         fasttrap_id_t *id;
736 #if defined(sun)
737         kmutex_t *pid_mtx;
738 #endif
739
740 #if defined(sun)
741         pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock;
742         mutex_enter(pid_mtx);
743 #endif
744         bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)];
745
746         for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
747                 if (pid == tp->ftt_pid && pc == tp->ftt_pc &&
748                     tp->ftt_proc->ftpc_acount != 0)
749                         break;
750         }
751
752         /*
753          * Don't sweat it if we can't find the tracepoint again; unlike
754          * when we're in fasttrap_pid_probe(), finding the tracepoint here
755          * is not essential to the correct execution of the process.
756          */
757         if (tp == NULL) {
758 #if defined(sun)
759                 mutex_exit(pid_mtx);
760 #endif
761                 return;
762         }
763
764         for (id = tp->ftt_retids; id != NULL; id = id->fti_next) {
765                 /*
766                  * If there's a branch that could act as a return site, we
767                  * need to trace it, and check here if the program counter is
768                  * external to the function.
769                  */
770                 if (tp->ftt_type != FASTTRAP_T_RET &&
771                     tp->ftt_type != FASTTRAP_T_RET16 &&
772                     new_pc - id->fti_probe->ftp_faddr <
773                     id->fti_probe->ftp_fsize)
774                         continue;
775
776                 dtrace_probe(id->fti_probe->ftp_id,
777                     pc - id->fti_probe->ftp_faddr,
778                     rp->r_rax, rp->r_rbx, 0, 0);
779         }
780
781 #if defined(sun)
782         mutex_exit(pid_mtx);
783 #endif
784 }
785
786 static void
787 fasttrap_sigsegv(proc_t *p, kthread_t *t, uintptr_t addr)
788 {
789 #if defined(sun)
790         sigqueue_t *sqp = kmem_zalloc(sizeof (sigqueue_t), KM_SLEEP);
791
792         sqp->sq_info.si_signo = SIGSEGV;
793         sqp->sq_info.si_code = SEGV_MAPERR;
794         sqp->sq_info.si_addr = (caddr_t)addr;
795
796         mutex_enter(&p->p_lock);
797         sigaddqa(p, t, sqp);
798         mutex_exit(&p->p_lock);
799
800         if (t != NULL)
801                 aston(t);
802 #else
803         ksiginfo_t *ksi = kmem_zalloc(sizeof (ksiginfo_t), KM_SLEEP);
804
805         ksiginfo_init(ksi);
806         ksi->ksi_signo = SIGSEGV;
807         ksi->ksi_code = SEGV_MAPERR;
808         ksi->ksi_addr = (caddr_t)addr;
809         (void) tdksignal(t, SIGSEGV, ksi);
810 #endif
811 }
812
813 #ifdef __amd64
814 static void
815 fasttrap_usdt_args64(fasttrap_probe_t *probe, struct reg *rp, int argc,
816     uintptr_t *argv)
817 {
818         int i, x, cap = MIN(argc, probe->ftp_nargs);
819         uintptr_t *stack = (uintptr_t *)rp->r_rsp;
820
821         for (i = 0; i < cap; i++) {
822                 x = probe->ftp_argmap[i];
823
824                 if (x < 6)
825                         argv[i] = (&rp->r_rdi)[x];
826                 else
827                         argv[i] = fasttrap_fulword_noerr(&stack[x]);
828         }
829
830         for (; i < argc; i++) {
831                 argv[i] = 0;
832         }
833 }
834 #endif
835
836 #ifdef __i386__
837 static void
838 fasttrap_usdt_args32(fasttrap_probe_t *probe, struct reg *rp, int argc,
839     uint32_t *argv)
840 {
841         int i, x, cap = MIN(argc, probe->ftp_nargs);
842         uint32_t *stack = (uint32_t *)rp->r_rsp;
843
844         for (i = 0; i < cap; i++) {
845                 x = probe->ftp_argmap[i];
846
847                 argv[i] = fasttrap_fuword32_noerr(&stack[x]);
848         }
849
850         for (; i < argc; i++) {
851                 argv[i] = 0;
852         }
853 }
854 #endif
855
856 static int
857 fasttrap_do_seg(fasttrap_tracepoint_t *tp, struct reg *rp, uintptr_t *addr)
858 {
859         proc_t *p = curproc;
860 #ifdef __i386__
861         struct segment_descriptor *desc;
862 #else
863         struct user_segment_descriptor *desc;
864 #endif
865         uint16_t sel = 0, ndx, type;
866         uintptr_t limit;
867
868         switch (tp->ftt_segment) {
869         case FASTTRAP_SEG_CS:
870                 sel = rp->r_cs;
871                 break;
872         case FASTTRAP_SEG_DS:
873                 sel = rp->r_ds;
874                 break;
875         case FASTTRAP_SEG_ES:
876                 sel = rp->r_es;
877                 break;
878         case FASTTRAP_SEG_FS:
879                 sel = rp->r_fs;
880                 break;
881         case FASTTRAP_SEG_GS:
882                 sel = rp->r_gs;
883                 break;
884         case FASTTRAP_SEG_SS:
885                 sel = rp->r_ss;
886                 break;
887         }
888
889         /*
890          * Make sure the given segment register specifies a user priority
891          * selector rather than a kernel selector.
892          */
893         if (ISPL(sel) != SEL_UPL)
894                 return (-1);
895
896         ndx = IDXSEL(sel);
897
898         /*
899          * Check the bounds and grab the descriptor out of the specified
900          * descriptor table.
901          */
902         if (ISLDT(sel)) {
903 #ifdef __i386__
904                 if (ndx > p->p_md.md_ldt->ldt_len)
905                         return (-1);
906
907                 desc = (struct segment_descriptor *)
908                     p->p_md.md_ldt[ndx].ldt_base;
909 #else
910                 if (ndx > max_ldt_segment)
911                         return (-1);
912
913                 desc = (struct user_segment_descriptor *)
914                     p->p_md.md_ldt[ndx].ldt_base;
915 #endif
916
917         } else {
918                 if (ndx >= NGDT)
919                         return (-1);
920
921 #ifdef __i386__
922                 desc = &gdt[ndx].sd;
923 #else
924                 desc = &gdt[ndx];
925 #endif
926         }
927
928         /*
929          * The descriptor must have user privilege level and it must be
930          * present in memory.
931          */
932         if (desc->sd_dpl != SEL_UPL || desc->sd_p != 1)
933                 return (-1);
934
935         type = desc->sd_type;
936
937         /*
938          * If the S bit in the type field is not set, this descriptor can
939          * only be used in system context.
940          */
941         if ((type & 0x10) != 0x10)
942                 return (-1);
943
944         limit = USD_GETLIMIT(desc) * (desc->sd_gran ? PAGESIZE : 1);
945
946         if (tp->ftt_segment == FASTTRAP_SEG_CS) {
947                 /*
948                  * The code/data bit and readable bit must both be set.
949                  */
950                 if ((type & 0xa) != 0xa)
951                         return (-1);
952
953                 if (*addr > limit)
954                         return (-1);
955         } else {
956                 /*
957                  * The code/data bit must be clear.
958                  */
959                 if ((type & 0x8) != 0)
960                         return (-1);
961
962                 /*
963                  * If the expand-down bit is clear, we just check the limit as
964                  * it would naturally be applied. Otherwise, we need to check
965                  * that the address is the range [limit + 1 .. 0xffff] or
966                  * [limit + 1 ... 0xffffffff] depending on if the default
967                  * operand size bit is set.
968                  */
969                 if ((type & 0x4) == 0) {
970                         if (*addr > limit)
971                                 return (-1);
972                 } else if (desc->sd_def32) {
973                         if (*addr < limit + 1 || 0xffff < *addr)
974                                 return (-1);
975                 } else {
976                         if (*addr < limit + 1 || 0xffffffff < *addr)
977                                 return (-1);
978                 }
979         }
980
981         *addr += USD_GETBASE(desc);
982
983         return (0);
984 }
985
986 int
987 fasttrap_pid_probe(struct reg *rp)
988 {
989         proc_t *p = curproc;
990         uintptr_t pc = rp->r_rip - 1;
991         uintptr_t new_pc = 0;
992         fasttrap_bucket_t *bucket;
993 #if defined(sun)
994         kmutex_t *pid_mtx;
995 #endif
996         fasttrap_tracepoint_t *tp, tp_local;
997         pid_t pid;
998         dtrace_icookie_t cookie;
999         uint_t is_enabled = 0;
1000
1001         /*
1002          * It's possible that a user (in a veritable orgy of bad planning)
1003          * could redirect this thread's flow of control before it reached the
1004          * return probe fasttrap. In this case we need to kill the process
1005          * since it's in a unrecoverable state.
1006          */
1007         if (curthread->t_dtrace_step) {
1008                 ASSERT(curthread->t_dtrace_on);
1009                 fasttrap_sigtrap(p, curthread, pc);
1010                 return (0);
1011         }
1012
1013         /*
1014          * Clear all user tracing flags.
1015          */
1016         curthread->t_dtrace_ft = 0;
1017         curthread->t_dtrace_pc = 0;
1018         curthread->t_dtrace_npc = 0;
1019         curthread->t_dtrace_scrpc = 0;
1020         curthread->t_dtrace_astpc = 0;
1021 #ifdef __amd64
1022         curthread->t_dtrace_regv = 0;
1023 #endif
1024
1025 #if defined(sun)
1026         /*
1027          * Treat a child created by a call to vfork(2) as if it were its
1028          * parent. We know that there's only one thread of control in such a
1029          * process: this one.
1030          */
1031         while (p->p_flag & SVFORK) {
1032                 p = p->p_parent;
1033         }
1034 #endif
1035
1036         PROC_LOCK(p);
1037         pid = p->p_pid;
1038 #if defined(sun)
1039         pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock;
1040         mutex_enter(pid_mtx);
1041 #endif
1042         bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)];
1043
1044         /*
1045          * Lookup the tracepoint that the process just hit.
1046          */
1047         for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
1048                 if (pid == tp->ftt_pid && pc == tp->ftt_pc &&
1049                     tp->ftt_proc->ftpc_acount != 0)
1050                         break;
1051         }
1052
1053         /*
1054          * If we couldn't find a matching tracepoint, either a tracepoint has
1055          * been inserted without using the pid<pid> ioctl interface (see
1056          * fasttrap_ioctl), or somehow we have mislaid this tracepoint.
1057          */
1058         if (tp == NULL) {
1059 #if defined(sun)
1060                 mutex_exit(pid_mtx);
1061 #endif
1062                 PROC_UNLOCK(p);
1063                 return (-1);
1064         }
1065
1066         /*
1067          * Set the program counter to the address of the traced instruction
1068          * so that it looks right in ustack() output.
1069          */
1070         rp->r_rip = pc;
1071
1072         if (tp->ftt_ids != NULL) {
1073                 fasttrap_id_t *id;
1074
1075 #ifdef __amd64
1076                 if (p->p_model == DATAMODEL_LP64) {
1077                         for (id = tp->ftt_ids; id != NULL; id = id->fti_next) {
1078                                 fasttrap_probe_t *probe = id->fti_probe;
1079
1080                                 if (id->fti_ptype == DTFTP_ENTRY) {
1081                                         /*
1082                                          * We note that this was an entry
1083                                          * probe to help ustack() find the
1084                                          * first caller.
1085                                          */
1086                                         cookie = dtrace_interrupt_disable();
1087                                         DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY);
1088                                         dtrace_probe(probe->ftp_id, rp->r_rdi,
1089                                             rp->r_rsi, rp->r_rdx, rp->r_rcx,
1090                                             rp->r_r8);
1091                                         DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY);
1092                                         dtrace_interrupt_enable(cookie);
1093                                 } else if (id->fti_ptype == DTFTP_IS_ENABLED) {
1094                                         /*
1095                                          * Note that in this case, we don't
1096                                          * call dtrace_probe() since it's only
1097                                          * an artificial probe meant to change
1098                                          * the flow of control so that it
1099                                          * encounters the true probe.
1100                                          */
1101                                         is_enabled = 1;
1102                                 } else if (probe->ftp_argmap == NULL) {
1103                                         dtrace_probe(probe->ftp_id, rp->r_rdi,
1104                                             rp->r_rsi, rp->r_rdx, rp->r_rcx,
1105                                             rp->r_r8);
1106                                 } else {
1107                                         uintptr_t t[5];
1108
1109                                         fasttrap_usdt_args64(probe, rp,
1110                                             sizeof (t) / sizeof (t[0]), t);
1111
1112                                         dtrace_probe(probe->ftp_id, t[0], t[1],
1113                                             t[2], t[3], t[4]);
1114                                 }
1115                         }
1116                 } else {
1117 #else /* __amd64 */
1118                         uintptr_t s0, s1, s2, s3, s4, s5;
1119                         uint32_t *stack = (uint32_t *)rp->r_esp;
1120
1121                         /*
1122                          * In 32-bit mode, all arguments are passed on the
1123                          * stack. If this is a function entry probe, we need
1124                          * to skip the first entry on the stack as it
1125                          * represents the return address rather than a
1126                          * parameter to the function.
1127                          */
1128                         s0 = fasttrap_fuword32_noerr(&stack[0]);
1129                         s1 = fasttrap_fuword32_noerr(&stack[1]);
1130                         s2 = fasttrap_fuword32_noerr(&stack[2]);
1131                         s3 = fasttrap_fuword32_noerr(&stack[3]);
1132                         s4 = fasttrap_fuword32_noerr(&stack[4]);
1133                         s5 = fasttrap_fuword32_noerr(&stack[5]);
1134
1135                         for (id = tp->ftt_ids; id != NULL; id = id->fti_next) {
1136                                 fasttrap_probe_t *probe = id->fti_probe;
1137
1138                                 if (id->fti_ptype == DTFTP_ENTRY) {
1139                                         /*
1140                                          * We note that this was an entry
1141                                          * probe to help ustack() find the
1142                                          * first caller.
1143                                          */
1144                                         cookie = dtrace_interrupt_disable();
1145                                         DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY);
1146                                         dtrace_probe(probe->ftp_id, s1, s2,
1147                                             s3, s4, s5);
1148                                         DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY);
1149                                         dtrace_interrupt_enable(cookie);
1150                                 } else if (id->fti_ptype == DTFTP_IS_ENABLED) {
1151                                         /*
1152                                          * Note that in this case, we don't
1153                                          * call dtrace_probe() since it's only
1154                                          * an artificial probe meant to change
1155                                          * the flow of control so that it
1156                                          * encounters the true probe.
1157                                          */
1158                                         is_enabled = 1;
1159                                 } else if (probe->ftp_argmap == NULL) {
1160                                         dtrace_probe(probe->ftp_id, s0, s1,
1161                                             s2, s3, s4);
1162                                 } else {
1163                                         uint32_t t[5];
1164
1165                                         fasttrap_usdt_args32(probe, rp,
1166                                             sizeof (t) / sizeof (t[0]), t);
1167
1168                                         dtrace_probe(probe->ftp_id, t[0], t[1],
1169                                             t[2], t[3], t[4]);
1170                                 }
1171                         }
1172 #endif /* __amd64 */
1173 #ifdef __amd64
1174                 }
1175 #endif
1176         }
1177
1178         /*
1179          * We're about to do a bunch of work so we cache a local copy of
1180          * the tracepoint to emulate the instruction, and then find the
1181          * tracepoint again later if we need to light up any return probes.
1182          */
1183         tp_local = *tp;
1184         PROC_UNLOCK(p);
1185 #if defined(sun)
1186         mutex_exit(pid_mtx);
1187 #endif
1188         tp = &tp_local;
1189
1190         /*
1191          * Set the program counter to appear as though the traced instruction
1192          * had completely executed. This ensures that fasttrap_getreg() will
1193          * report the expected value for REG_RIP.
1194          */
1195         rp->r_rip = pc + tp->ftt_size;
1196
1197         /*
1198          * If there's an is-enabled probe connected to this tracepoint it
1199          * means that there was a 'xorl %eax, %eax' or 'xorq %rax, %rax'
1200          * instruction that was placed there by DTrace when the binary was
1201          * linked. As this probe is, in fact, enabled, we need to stuff 1
1202          * into %eax or %rax. Accordingly, we can bypass all the instruction
1203          * emulation logic since we know the inevitable result. It's possible
1204          * that a user could construct a scenario where the 'is-enabled'
1205          * probe was on some other instruction, but that would be a rather
1206          * exotic way to shoot oneself in the foot.
1207          */
1208         if (is_enabled) {
1209                 rp->r_rax = 1;
1210                 new_pc = rp->r_rip;
1211                 goto done;
1212         }
1213
1214         /*
1215          * We emulate certain types of instructions to ensure correctness
1216          * (in the case of position dependent instructions) or optimize
1217          * common cases. The rest we have the thread execute back in user-
1218          * land.
1219          */
1220         switch (tp->ftt_type) {
1221         case FASTTRAP_T_RET:
1222         case FASTTRAP_T_RET16:
1223         {
1224                 uintptr_t dst = 0;
1225                 uintptr_t addr = 0;
1226                 int ret = 0;
1227
1228                 /*
1229                  * We have to emulate _every_ facet of the behavior of a ret
1230                  * instruction including what happens if the load from %esp
1231                  * fails; in that case, we send a SIGSEGV.
1232                  */
1233 #ifdef __amd64
1234                 if (p->p_model == DATAMODEL_NATIVE) {
1235                         ret = dst = fasttrap_fulword((void *)rp->r_rsp);
1236                         addr = rp->r_rsp + sizeof (uintptr_t);
1237                 } else {
1238 #endif
1239 #ifdef __i386__
1240                         uint32_t dst32;
1241                         ret = dst32 = fasttrap_fuword32((void *)rp->r_esp);
1242                         dst = dst32;
1243                         addr = rp->r_esp + sizeof (uint32_t);
1244 #endif
1245 #ifdef __amd64
1246                 }
1247 #endif
1248
1249                 if (ret == -1) {
1250                         fasttrap_sigsegv(p, curthread, rp->r_rsp);
1251                         new_pc = pc;
1252                         break;
1253                 }
1254
1255                 if (tp->ftt_type == FASTTRAP_T_RET16)
1256                         addr += tp->ftt_dest;
1257
1258                 rp->r_rsp = addr;
1259                 new_pc = dst;
1260                 break;
1261         }
1262
1263         case FASTTRAP_T_JCC:
1264         {
1265                 uint_t taken = 0;
1266
1267                 switch (tp->ftt_code) {
1268                 case FASTTRAP_JO:
1269                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_OF) != 0;
1270                         break;
1271                 case FASTTRAP_JNO:
1272                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0;
1273                         break;
1274                 case FASTTRAP_JB:
1275                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) != 0;
1276                         break;
1277                 case FASTTRAP_JAE:
1278                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) == 0;
1279                         break;
1280                 case FASTTRAP_JE:
1281                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0;
1282                         break;
1283                 case FASTTRAP_JNE:
1284                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0;
1285                         break;
1286                 case FASTTRAP_JBE:
1287                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) != 0 ||
1288                             (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0;
1289                         break;
1290                 case FASTTRAP_JA:
1291                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) == 0 &&
1292                             (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0;
1293                         break;
1294                 case FASTTRAP_JS:
1295                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_SF) != 0;
1296                         break;
1297                 case FASTTRAP_JNS:
1298                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0;
1299                         break;
1300                 case FASTTRAP_JP:
1301                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_PF) != 0;
1302                         break;
1303                 case FASTTRAP_JNP:
1304                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_PF) == 0;
1305                         break;
1306                 case FASTTRAP_JL:
1307                         taken = ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) !=
1308                             ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
1309                         break;
1310                 case FASTTRAP_JGE:
1311                         taken = ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) ==
1312                             ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
1313                         break;
1314                 case FASTTRAP_JLE:
1315                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0 ||
1316                             ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) !=
1317                             ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
1318                         break;
1319                 case FASTTRAP_JG:
1320                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0 &&
1321                             ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) ==
1322                             ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
1323                         break;
1324
1325                 }
1326
1327                 if (taken)
1328                         new_pc = tp->ftt_dest;
1329                 else
1330                         new_pc = pc + tp->ftt_size;
1331                 break;
1332         }
1333
1334         case FASTTRAP_T_LOOP:
1335         {
1336                 uint_t taken = 0;
1337 #ifdef __amd64
1338                 greg_t cx = rp->r_rcx--;
1339 #else
1340                 greg_t cx = rp->r_ecx--;
1341 #endif
1342
1343                 switch (tp->ftt_code) {
1344                 case FASTTRAP_LOOPNZ:
1345                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0 &&
1346                             cx != 0;
1347                         break;
1348                 case FASTTRAP_LOOPZ:
1349                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0 &&
1350                             cx != 0;
1351                         break;
1352                 case FASTTRAP_LOOP:
1353                         taken = (cx != 0);
1354                         break;
1355                 }
1356
1357                 if (taken)
1358                         new_pc = tp->ftt_dest;
1359                 else
1360                         new_pc = pc + tp->ftt_size;
1361                 break;
1362         }
1363
1364         case FASTTRAP_T_JCXZ:
1365         {
1366 #ifdef __amd64
1367                 greg_t cx = rp->r_rcx;
1368 #else
1369                 greg_t cx = rp->r_ecx;
1370 #endif
1371
1372                 if (cx == 0)
1373                         new_pc = tp->ftt_dest;
1374                 else
1375                         new_pc = pc + tp->ftt_size;
1376                 break;
1377         }
1378
1379         case FASTTRAP_T_PUSHL_EBP:
1380         {
1381                 int ret = 0;
1382                 uintptr_t addr = 0;
1383
1384 #ifdef __amd64
1385                 if (p->p_model == DATAMODEL_NATIVE) {
1386                         addr = rp->r_rsp - sizeof (uintptr_t);
1387                         ret = fasttrap_sulword((void *)addr, &rp->r_rsp);
1388                 } else {
1389 #endif
1390 #ifdef __i386__
1391                         addr = rp->r_rsp - sizeof (uint32_t);
1392                         ret = fasttrap_suword32((void *)addr, &rp->r_rsp);
1393 #endif
1394 #ifdef __amd64
1395                 }
1396 #endif
1397
1398                 if (ret == -1) {
1399                         fasttrap_sigsegv(p, curthread, addr);
1400                         new_pc = pc;
1401                         break;
1402                 }
1403
1404                 rp->r_rsp = addr;
1405                 new_pc = pc + tp->ftt_size;
1406                 break;
1407         }
1408
1409         case FASTTRAP_T_NOP:
1410                 new_pc = pc + tp->ftt_size;
1411                 break;
1412
1413         case FASTTRAP_T_JMP:
1414         case FASTTRAP_T_CALL:
1415                 if (tp->ftt_code == 0) {
1416                         new_pc = tp->ftt_dest;
1417                 } else {
1418 #ifdef __amd64
1419                         uintptr_t value;
1420 #endif
1421                         uintptr_t addr = tp->ftt_dest;
1422
1423                         if (tp->ftt_base != FASTTRAP_NOREG)
1424                                 addr += fasttrap_getreg(rp, tp->ftt_base);
1425                         if (tp->ftt_index != FASTTRAP_NOREG)
1426                                 addr += fasttrap_getreg(rp, tp->ftt_index) <<
1427                                     tp->ftt_scale;
1428
1429                         if (tp->ftt_code == 1) {
1430                                 /*
1431                                  * If there's a segment prefix for this
1432                                  * instruction, we'll need to check permissions
1433                                  * and bounds on the given selector, and adjust
1434                                  * the address accordingly.
1435                                  */
1436                                 if (tp->ftt_segment != FASTTRAP_SEG_NONE &&
1437                                     fasttrap_do_seg(tp, rp, &addr) != 0) {
1438                                         fasttrap_sigsegv(p, curthread, addr);
1439                                         new_pc = pc;
1440                                         break;
1441                                 }
1442
1443 #ifdef __amd64
1444                                 if (p->p_model == DATAMODEL_NATIVE) {
1445                                         if ((value = fasttrap_fulword((void *)addr))
1446                                              == -1) {
1447                                                 fasttrap_sigsegv(p, curthread,
1448                                                     addr);
1449                                                 new_pc = pc;
1450                                                 break;
1451                                         }
1452                                         new_pc = value;
1453                                 } else {
1454 #endif
1455 #ifdef __i386__
1456                                         uint32_t value32;
1457                                         addr = (uintptr_t)(uint32_t)addr;
1458                                         if ((value32 = fasttrap_fuword32((void *)addr))
1459                                             == -1) {
1460                                                 fasttrap_sigsegv(p, curthread,
1461                                                     addr);
1462                                                 new_pc = pc;
1463                                                 break;
1464                                         }
1465                                         new_pc = value32;
1466 #endif
1467                                 }
1468 #ifdef __amd64
1469                         } else {
1470                                 new_pc = addr;
1471                         }
1472 #endif
1473                 }
1474
1475                 /*
1476                  * If this is a call instruction, we need to push the return
1477                  * address onto the stack. If this fails, we send the process
1478                  * a SIGSEGV and reset the pc to emulate what would happen if
1479                  * this instruction weren't traced.
1480                  */
1481                 if (tp->ftt_type == FASTTRAP_T_CALL) {
1482                         int ret = 0;
1483                         uintptr_t addr = 0, pcps;
1484 #ifdef __amd64
1485                         if (p->p_model == DATAMODEL_NATIVE) {
1486                                 addr = rp->r_rsp - sizeof (uintptr_t);
1487                                 pcps = pc + tp->ftt_size;
1488                                 ret = fasttrap_sulword((void *)addr, &pcps);
1489                         } else {
1490 #endif
1491 #ifdef __i386__
1492                                 addr = rp->r_rsp - sizeof (uint32_t);
1493                                 pcps = (uint32_t)(pc + tp->ftt_size);
1494                                 ret = fasttrap_suword32((void *)addr, &pcps);
1495 #endif
1496 #ifdef __amd64
1497                         }
1498 #endif
1499
1500                         if (ret == -1) {
1501                                 fasttrap_sigsegv(p, curthread, addr);
1502                                 new_pc = pc;
1503                                 break;
1504                         }
1505
1506                         rp->r_rsp = addr;
1507                 }
1508
1509                 break;
1510
1511         case FASTTRAP_T_COMMON:
1512         {
1513                 uintptr_t addr;
1514 #if defined(__amd64)
1515                 uint8_t scratch[2 * FASTTRAP_MAX_INSTR_SIZE + 22];
1516 #else
1517                 uint8_t scratch[2 * FASTTRAP_MAX_INSTR_SIZE + 7];
1518 #endif
1519                 uint_t i = 0;
1520 #if defined(sun)
1521                 klwp_t *lwp = ttolwp(curthread);
1522 #endif
1523
1524                 /*
1525                  * Compute the address of the ulwp_t and step over the
1526                  * ul_self pointer. The method used to store the user-land
1527                  * thread pointer is very different on 32- and 64-bit
1528                  * kernels.
1529                  */
1530 #if defined(sun)
1531 #if defined(__amd64)
1532                 if (p->p_model == DATAMODEL_LP64) {
1533                         addr = lwp->lwp_pcb.pcb_fsbase;
1534                         addr += sizeof (void *);
1535                 } else {
1536                         addr = lwp->lwp_pcb.pcb_gsbase;
1537                         addr += sizeof (caddr32_t);
1538                 }
1539 #else
1540                 addr = USD_GETBASE(&lwp->lwp_pcb.pcb_gsdesc);
1541                 addr += sizeof (void *);
1542 #endif
1543 #endif /* sun */
1544 #ifdef __i386__
1545                 addr = USD_GETBASE(&curthread->td_pcb->pcb_gsd);
1546 #else
1547                 addr = curthread->td_pcb->pcb_gsbase;
1548 #endif
1549                 addr += sizeof (void *);
1550
1551                 /*
1552                  * Generic Instruction Tracing
1553                  * ---------------------------
1554                  *
1555                  * This is the layout of the scratch space in the user-land
1556                  * thread structure for our generated instructions.
1557                  *
1558                  *      32-bit mode                     bytes
1559                  *      ------------------------        -----
1560                  * a:   <original instruction>          <= 15
1561                  *      jmp     <pc + tp->ftt_size>         5
1562                  * b:   <original instruction>          <= 15
1563                  *      int     T_DTRACE_RET                2
1564                  *                                      -----
1565                  *                                      <= 37
1566                  *
1567                  *      64-bit mode                     bytes
1568                  *      ------------------------        -----
1569                  * a:   <original instruction>          <= 15
1570                  *      jmp     0(%rip)                     6
1571                  *      <pc + tp->ftt_size>                 8
1572                  * b:   <original instruction>          <= 15
1573                  *      int     T_DTRACE_RET                2
1574                  *                                      -----
1575                  *                                      <= 46
1576                  *
1577                  * The %pc is set to a, and curthread->t_dtrace_astpc is set
1578                  * to b. If we encounter a signal on the way out of the
1579                  * kernel, trap() will set %pc to curthread->t_dtrace_astpc
1580                  * so that we execute the original instruction and re-enter
1581                  * the kernel rather than redirecting to the next instruction.
1582                  *
1583                  * If there are return probes (so we know that we're going to
1584                  * need to reenter the kernel after executing the original
1585                  * instruction), the scratch space will just contain the
1586                  * original instruction followed by an interrupt -- the same
1587                  * data as at b.
1588                  *
1589                  * %rip-relative Addressing
1590                  * ------------------------
1591                  *
1592                  * There's a further complication in 64-bit mode due to %rip-
1593                  * relative addressing. While this is clearly a beneficial
1594                  * architectural decision for position independent code, it's
1595                  * hard not to see it as a personal attack against the pid
1596                  * provider since before there was a relatively small set of
1597                  * instructions to emulate; with %rip-relative addressing,
1598                  * almost every instruction can potentially depend on the
1599                  * address at which it's executed. Rather than emulating
1600                  * the broad spectrum of instructions that can now be
1601                  * position dependent, we emulate jumps and others as in
1602                  * 32-bit mode, and take a different tack for instructions
1603                  * using %rip-relative addressing.
1604                  *
1605                  * For every instruction that uses the ModRM byte, the
1606                  * in-kernel disassembler reports its location. We use the
1607                  * ModRM byte to identify that an instruction uses
1608                  * %rip-relative addressing and to see what other registers
1609                  * the instruction uses. To emulate those instructions,
1610                  * we modify the instruction to be %rax-relative rather than
1611                  * %rip-relative (or %rcx-relative if the instruction uses
1612                  * %rax; or %r8- or %r9-relative if the REX.B is present so
1613                  * we don't have to rewrite the REX prefix). We then load
1614                  * the value that %rip would have been into the scratch
1615                  * register and generate an instruction to reset the scratch
1616                  * register back to its original value. The instruction
1617                  * sequence looks like this:
1618                  *
1619                  *      64-mode %rip-relative           bytes
1620                  *      ------------------------        -----
1621                  * a:   <modified instruction>          <= 15
1622                  *      movq    $<value>, %<scratch>        6
1623                  *      jmp     0(%rip)                     6
1624                  *      <pc + tp->ftt_size>                 8
1625                  * b:   <modified instruction>          <= 15
1626                  *      int     T_DTRACE_RET                2
1627                  *                                      -----
1628                  *                                         52
1629                  *
1630                  * We set curthread->t_dtrace_regv so that upon receiving
1631                  * a signal we can reset the value of the scratch register.
1632                  */
1633
1634                 ASSERT(tp->ftt_size < FASTTRAP_MAX_INSTR_SIZE);
1635
1636                 curthread->t_dtrace_scrpc = addr;
1637                 bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size);
1638                 i += tp->ftt_size;
1639
1640 #ifdef __amd64
1641                 if (tp->ftt_ripmode != 0) {
1642                         greg_t *reg = NULL;
1643
1644                         ASSERT(p->p_model == DATAMODEL_LP64);
1645                         ASSERT(tp->ftt_ripmode &
1646                             (FASTTRAP_RIP_1 | FASTTRAP_RIP_2));
1647
1648                         /*
1649                          * If this was a %rip-relative instruction, we change
1650                          * it to be either a %rax- or %rcx-relative
1651                          * instruction (depending on whether those registers
1652                          * are used as another operand; or %r8- or %r9-
1653                          * relative depending on the value of REX.B). We then
1654                          * set that register and generate a movq instruction
1655                          * to reset the value.
1656                          */
1657                         if (tp->ftt_ripmode & FASTTRAP_RIP_X)
1658                                 scratch[i++] = FASTTRAP_REX(1, 0, 0, 1);
1659                         else
1660                                 scratch[i++] = FASTTRAP_REX(1, 0, 0, 0);
1661
1662                         if (tp->ftt_ripmode & FASTTRAP_RIP_1)
1663                                 scratch[i++] = FASTTRAP_MOV_EAX;
1664                         else
1665                                 scratch[i++] = FASTTRAP_MOV_ECX;
1666
1667                         switch (tp->ftt_ripmode) {
1668                         case FASTTRAP_RIP_1:
1669                                 reg = &rp->r_rax;
1670                                 curthread->t_dtrace_reg = REG_RAX;
1671                                 break;
1672                         case FASTTRAP_RIP_2:
1673                                 reg = &rp->r_rcx;
1674                                 curthread->t_dtrace_reg = REG_RCX;
1675                                 break;
1676                         case FASTTRAP_RIP_1 | FASTTRAP_RIP_X:
1677                                 reg = &rp->r_r8;
1678                                 curthread->t_dtrace_reg = REG_R8;
1679                                 break;
1680                         case FASTTRAP_RIP_2 | FASTTRAP_RIP_X:
1681                                 reg = &rp->r_r9;
1682                                 curthread->t_dtrace_reg = REG_R9;
1683                                 break;
1684                         }
1685
1686                         /* LINTED - alignment */
1687                         *(uint64_t *)&scratch[i] = *reg;
1688                         curthread->t_dtrace_regv = *reg;
1689                         *reg = pc + tp->ftt_size;
1690                         i += sizeof (uint64_t);
1691                 }
1692 #endif
1693
1694                 /*
1695                  * Generate the branch instruction to what would have
1696                  * normally been the subsequent instruction. In 32-bit mode,
1697                  * this is just a relative branch; in 64-bit mode this is a
1698                  * %rip-relative branch that loads the 64-bit pc value
1699                  * immediately after the jmp instruction.
1700                  */
1701 #ifdef __amd64
1702                 if (p->p_model == DATAMODEL_LP64) {
1703                         scratch[i++] = FASTTRAP_GROUP5_OP;
1704                         scratch[i++] = FASTTRAP_MODRM(0, 4, 5);
1705                         /* LINTED - alignment */
1706                         *(uint32_t *)&scratch[i] = 0;
1707                         i += sizeof (uint32_t);
1708                         /* LINTED - alignment */
1709                         *(uint64_t *)&scratch[i] = pc + tp->ftt_size;
1710                         i += sizeof (uint64_t);
1711                 } else {
1712 #endif
1713 #ifdef __i386__
1714                         /*
1715                          * Set up the jmp to the next instruction; note that
1716                          * the size of the traced instruction cancels out.
1717                          */
1718                         scratch[i++] = FASTTRAP_JMP32;
1719                         /* LINTED - alignment */
1720                         *(uint32_t *)&scratch[i] = pc - addr - 5;
1721                         i += sizeof (uint32_t);
1722 #endif
1723 #ifdef __amd64
1724                 }
1725 #endif
1726
1727                 curthread->t_dtrace_astpc = addr + i;
1728                 bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size);
1729                 i += tp->ftt_size;
1730                 scratch[i++] = FASTTRAP_INT;
1731                 scratch[i++] = T_DTRACE_RET;
1732
1733                 ASSERT(i <= sizeof (scratch));
1734
1735
1736 #if defined(sun)
1737                 if (fasttrap_copyout(scratch, (char *)addr, i)) {
1738 #else
1739                 if (uwrite(curproc, scratch, i, addr)) {
1740 #endif
1741                         fasttrap_sigtrap(p, curthread, pc);
1742                         new_pc = pc;
1743                         break;
1744                 }
1745                 if (tp->ftt_retids != NULL) {
1746                         curthread->t_dtrace_step = 1;
1747                         curthread->t_dtrace_ret = 1;
1748                         new_pc = curthread->t_dtrace_astpc;
1749                 } else {
1750                         new_pc = curthread->t_dtrace_scrpc;
1751                 }
1752
1753                 curthread->t_dtrace_pc = pc;
1754                 curthread->t_dtrace_npc = pc + tp->ftt_size;
1755                 curthread->t_dtrace_on = 1;
1756                 break;
1757         }
1758
1759         default:
1760                 panic("fasttrap: mishandled an instruction");
1761         }
1762
1763 done:
1764         /*
1765          * If there were no return probes when we first found the tracepoint,
1766          * we should feel no obligation to honor any return probes that were
1767          * subsequently enabled -- they'll just have to wait until the next
1768          * time around.
1769          */
1770         if (tp->ftt_retids != NULL) {
1771                 /*
1772                  * We need to wait until the results of the instruction are
1773                  * apparent before invoking any return probes. If this
1774                  * instruction was emulated we can just call
1775                  * fasttrap_return_common(); if it needs to be executed, we
1776                  * need to wait until the user thread returns to the kernel.
1777                  */
1778                 if (tp->ftt_type != FASTTRAP_T_COMMON) {
1779                         /*
1780                          * Set the program counter to the address of the traced
1781                          * instruction so that it looks right in ustack()
1782                          * output. We had previously set it to the end of the
1783                          * instruction to simplify %rip-relative addressing.
1784                          */
1785                         rp->r_rip = pc;
1786
1787                         fasttrap_return_common(rp, pc, pid, new_pc);
1788                 } else {
1789                         ASSERT(curthread->t_dtrace_ret != 0);
1790                         ASSERT(curthread->t_dtrace_pc == pc);
1791                         ASSERT(curthread->t_dtrace_scrpc != 0);
1792                         ASSERT(new_pc == curthread->t_dtrace_astpc);
1793                 }
1794         }
1795
1796         rp->r_rip = new_pc;
1797         set_regs(curthread, rp);
1798
1799         return (0);
1800 }
1801
1802 int
1803 fasttrap_return_probe(struct reg *rp)
1804 {
1805         proc_t *p = curproc;
1806         uintptr_t pc = curthread->t_dtrace_pc;
1807         uintptr_t npc = curthread->t_dtrace_npc;
1808
1809         curthread->t_dtrace_pc = 0;
1810         curthread->t_dtrace_npc = 0;
1811         curthread->t_dtrace_scrpc = 0;
1812         curthread->t_dtrace_astpc = 0;
1813
1814 #if defined(sun)
1815         /*
1816          * Treat a child created by a call to vfork(2) as if it were its
1817          * parent. We know that there's only one thread of control in such a
1818          * process: this one.
1819          */
1820         while (p->p_flag & SVFORK) {
1821                 p = p->p_parent;
1822         }
1823 #endif
1824
1825         /*
1826          * We set rp->r_rip to the address of the traced instruction so
1827          * that it appears to dtrace_probe() that we're on the original
1828          * instruction, and so that the user can't easily detect our
1829          * complex web of lies. dtrace_return_probe() (our caller)
1830          * will correctly set %pc after we return.
1831          */
1832         rp->r_rip = pc;
1833
1834         fasttrap_return_common(rp, pc, p->p_pid, npc);
1835
1836         return (0);
1837 }
1838
1839 /*ARGSUSED*/
1840 uint64_t
1841 fasttrap_pid_getarg(void *arg, dtrace_id_t id, void *parg, int argno,
1842     int aframes)
1843 {
1844         struct reg r;
1845
1846         fill_regs(curthread, &r);
1847
1848         return (fasttrap_anarg(&r, 1, argno));
1849 }
1850
1851 /*ARGSUSED*/
1852 uint64_t
1853 fasttrap_usdt_getarg(void *arg, dtrace_id_t id, void *parg, int argno,
1854     int aframes)
1855 {
1856         struct reg r;
1857
1858         fill_regs(curthread, &r);
1859
1860         return (fasttrap_anarg(&r, 0, argno));
1861 }
1862
1863 static ulong_t
1864 fasttrap_getreg(struct reg *rp, uint_t reg)
1865 {
1866 #ifdef __amd64
1867         switch (reg) {
1868         case REG_R15:           return (rp->r_r15);
1869         case REG_R14:           return (rp->r_r14);
1870         case REG_R13:           return (rp->r_r13);
1871         case REG_R12:           return (rp->r_r12);
1872         case REG_R11:           return (rp->r_r11);
1873         case REG_R10:           return (rp->r_r10);
1874         case REG_R9:            return (rp->r_r9);
1875         case REG_R8:            return (rp->r_r8);
1876         case REG_RDI:           return (rp->r_rdi);
1877         case REG_RSI:           return (rp->r_rsi);
1878         case REG_RBP:           return (rp->r_rbp);
1879         case REG_RBX:           return (rp->r_rbx);
1880         case REG_RDX:           return (rp->r_rdx);
1881         case REG_RCX:           return (rp->r_rcx);
1882         case REG_RAX:           return (rp->r_rax);
1883         case REG_TRAPNO:        return (rp->r_trapno);
1884         case REG_ERR:           return (rp->r_err);
1885         case REG_RIP:           return (rp->r_rip);
1886         case REG_CS:            return (rp->r_cs);
1887 #if defined(sun)
1888         case REG_RFL:           return (rp->r_rfl);
1889 #endif
1890         case REG_RSP:           return (rp->r_rsp);
1891         case REG_SS:            return (rp->r_ss);
1892         case REG_FS:            return (rp->r_fs);
1893         case REG_GS:            return (rp->r_gs);
1894         case REG_DS:            return (rp->r_ds);
1895         case REG_ES:            return (rp->r_es);
1896         case REG_FSBASE:        return (rdmsr(MSR_FSBASE));
1897         case REG_GSBASE:        return (rdmsr(MSR_GSBASE));
1898         }
1899
1900         panic("dtrace: illegal register constant");
1901         /*NOTREACHED*/
1902 #else
1903 #define _NGREG 19
1904         if (reg >= _NGREG)
1905                 panic("dtrace: illegal register constant");
1906
1907         return (((greg_t *)&rp->r_gs)[reg]);
1908 #endif
1909 }