]> CyberLeo.Net >> Repos - FreeBSD/stable/9.git/blob - sys/cddl/contrib/opensolaris/uts/intel/dtrace/fasttrap_isa.c
MFC r257679: Use suword32 and suword64 instead of copyout(9) in fasttrap
[FreeBSD/stable/9.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 #define r_rbp   r_ebp
108 #endif
109
110 /*
111  * Lossless User-Land Tracing on x86
112  * ---------------------------------
113  *
114  * The execution of most instructions is not dependent on the address; for
115  * these instructions it is sufficient to copy them into the user process's
116  * address space and execute them. To effectively single-step an instruction
117  * in user-land, we copy out the following sequence of instructions to scratch
118  * space in the user thread's ulwp_t structure.
119  *
120  * We then set the program counter (%eip or %rip) to point to this scratch
121  * space. Once execution resumes, the original instruction is executed and
122  * then control flow is redirected to what was originally the subsequent
123  * instruction. If the kernel attemps to deliver a signal while single-
124  * stepping, the signal is deferred and the program counter is moved into the
125  * second sequence of instructions. The second sequence ends in a trap into
126  * the kernel where the deferred signal is then properly handled and delivered.
127  *
128  * For instructions whose execute is position dependent, we perform simple
129  * emulation. These instructions are limited to control transfer
130  * instructions in 32-bit mode, but in 64-bit mode there's the added wrinkle
131  * of %rip-relative addressing that means that almost any instruction can be
132  * position dependent. For all the details on how we emulate generic
133  * instructions included %rip-relative instructions, see the code in
134  * fasttrap_pid_probe() below where we handle instructions of type
135  * FASTTRAP_T_COMMON (under the header: Generic Instruction Tracing).
136  */
137
138 #define FASTTRAP_MODRM_MOD(modrm)       (((modrm) >> 6) & 0x3)
139 #define FASTTRAP_MODRM_REG(modrm)       (((modrm) >> 3) & 0x7)
140 #define FASTTRAP_MODRM_RM(modrm)        ((modrm) & 0x7)
141 #define FASTTRAP_MODRM(mod, reg, rm)    (((mod) << 6) | ((reg) << 3) | (rm))
142
143 #define FASTTRAP_SIB_SCALE(sib)         (((sib) >> 6) & 0x3)
144 #define FASTTRAP_SIB_INDEX(sib)         (((sib) >> 3) & 0x7)
145 #define FASTTRAP_SIB_BASE(sib)          ((sib) & 0x7)
146
147 #define FASTTRAP_REX_W(rex)             (((rex) >> 3) & 1)
148 #define FASTTRAP_REX_R(rex)             (((rex) >> 2) & 1)
149 #define FASTTRAP_REX_X(rex)             (((rex) >> 1) & 1)
150 #define FASTTRAP_REX_B(rex)             ((rex) & 1)
151 #define FASTTRAP_REX(w, r, x, b)        \
152         (0x40 | ((w) << 3) | ((r) << 2) | ((x) << 1) | (b))
153
154 /*
155  * Single-byte op-codes.
156  */
157 #define FASTTRAP_PUSHL_EBP      0x55
158
159 #define FASTTRAP_JO             0x70
160 #define FASTTRAP_JNO            0x71
161 #define FASTTRAP_JB             0x72
162 #define FASTTRAP_JAE            0x73
163 #define FASTTRAP_JE             0x74
164 #define FASTTRAP_JNE            0x75
165 #define FASTTRAP_JBE            0x76
166 #define FASTTRAP_JA             0x77
167 #define FASTTRAP_JS             0x78
168 #define FASTTRAP_JNS            0x79
169 #define FASTTRAP_JP             0x7a
170 #define FASTTRAP_JNP            0x7b
171 #define FASTTRAP_JL             0x7c
172 #define FASTTRAP_JGE            0x7d
173 #define FASTTRAP_JLE            0x7e
174 #define FASTTRAP_JG             0x7f
175
176 #define FASTTRAP_NOP            0x90
177
178 #define FASTTRAP_MOV_EAX        0xb8
179 #define FASTTRAP_MOV_ECX        0xb9
180
181 #define FASTTRAP_RET16          0xc2
182 #define FASTTRAP_RET            0xc3
183
184 #define FASTTRAP_LOOPNZ         0xe0
185 #define FASTTRAP_LOOPZ          0xe1
186 #define FASTTRAP_LOOP           0xe2
187 #define FASTTRAP_JCXZ           0xe3
188
189 #define FASTTRAP_CALL           0xe8
190 #define FASTTRAP_JMP32          0xe9
191 #define FASTTRAP_JMP8           0xeb
192
193 #define FASTTRAP_INT3           0xcc
194 #define FASTTRAP_INT            0xcd
195
196 #define FASTTRAP_2_BYTE_OP      0x0f
197 #define FASTTRAP_GROUP5_OP      0xff
198
199 /*
200  * Two-byte op-codes (second byte only).
201  */
202 #define FASTTRAP_0F_JO          0x80
203 #define FASTTRAP_0F_JNO         0x81
204 #define FASTTRAP_0F_JB          0x82
205 #define FASTTRAP_0F_JAE         0x83
206 #define FASTTRAP_0F_JE          0x84
207 #define FASTTRAP_0F_JNE         0x85
208 #define FASTTRAP_0F_JBE         0x86
209 #define FASTTRAP_0F_JA          0x87
210 #define FASTTRAP_0F_JS          0x88
211 #define FASTTRAP_0F_JNS         0x89
212 #define FASTTRAP_0F_JP          0x8a
213 #define FASTTRAP_0F_JNP         0x8b
214 #define FASTTRAP_0F_JL          0x8c
215 #define FASTTRAP_0F_JGE         0x8d
216 #define FASTTRAP_0F_JLE         0x8e
217 #define FASTTRAP_0F_JG          0x8f
218
219 #define FASTTRAP_EFLAGS_OF      0x800
220 #define FASTTRAP_EFLAGS_DF      0x400
221 #define FASTTRAP_EFLAGS_SF      0x080
222 #define FASTTRAP_EFLAGS_ZF      0x040
223 #define FASTTRAP_EFLAGS_AF      0x010
224 #define FASTTRAP_EFLAGS_PF      0x004
225 #define FASTTRAP_EFLAGS_CF      0x001
226
227 /*
228  * Instruction prefixes.
229  */
230 #define FASTTRAP_PREFIX_OPERAND 0x66
231 #define FASTTRAP_PREFIX_ADDRESS 0x67
232 #define FASTTRAP_PREFIX_CS      0x2E
233 #define FASTTRAP_PREFIX_DS      0x3E
234 #define FASTTRAP_PREFIX_ES      0x26
235 #define FASTTRAP_PREFIX_FS      0x64
236 #define FASTTRAP_PREFIX_GS      0x65
237 #define FASTTRAP_PREFIX_SS      0x36
238 #define FASTTRAP_PREFIX_LOCK    0xF0
239 #define FASTTRAP_PREFIX_REP     0xF3
240 #define FASTTRAP_PREFIX_REPNE   0xF2
241
242 #define FASTTRAP_NOREG  0xff
243
244 /*
245  * Map between instruction register encodings and the kernel constants which
246  * correspond to indicies into struct regs.
247  */
248 #ifdef __amd64
249 static const uint8_t regmap[16] = {
250         REG_RAX, REG_RCX, REG_RDX, REG_RBX, REG_RSP, REG_RBP, REG_RSI, REG_RDI,
251         REG_R8, REG_R9, REG_R10, REG_R11, REG_R12, REG_R13, REG_R14, REG_R15,
252 };
253 #else
254 static const uint8_t regmap[8] = {
255         EAX, ECX, EDX, EBX, UESP, EBP, ESI, EDI
256 };
257 #endif
258
259 static ulong_t fasttrap_getreg(struct reg *, uint_t);
260
261 static uint64_t
262 fasttrap_anarg(struct reg *rp, int function_entry, int argno)
263 {
264         uint64_t value = 0;
265         int shift = function_entry ? 1 : 0;
266
267 #ifdef __amd64
268         if (curproc->p_model == DATAMODEL_LP64) {
269                 uintptr_t *stack;
270
271                 /*
272                  * In 64-bit mode, the first six arguments are stored in
273                  * registers.
274                  */
275                 if (argno < 6)
276                         return ((&rp->r_rdi)[argno]);
277
278                 stack = (uintptr_t *)rp->r_rsp;
279                 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
280                 value = dtrace_fulword(&stack[argno - 6 + shift]);
281                 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT | CPU_DTRACE_BADADDR);
282         } else {
283 #endif
284 #ifdef __i386
285                 uint32_t *stack = (uint32_t *)rp->r_esp;
286                 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
287                 value = dtrace_fuword32(&stack[argno + shift]);
288                 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT | CPU_DTRACE_BADADDR);
289 #endif
290 #ifdef __amd64
291         }
292 #endif
293
294         return (value);
295 }
296
297 /*ARGSUSED*/
298 int
299 fasttrap_tracepoint_init(proc_t *p, fasttrap_tracepoint_t *tp, uintptr_t pc,
300     fasttrap_probe_type_t type)
301 {
302         uint8_t instr[FASTTRAP_MAX_INSTR_SIZE + 10];
303         size_t len = FASTTRAP_MAX_INSTR_SIZE;
304         size_t first = MIN(len, PAGESIZE - (pc & PAGEOFFSET));
305         uint_t start = 0;
306         int rmindex, size;
307         uint8_t seg, rex = 0;
308
309         /*
310          * Read the instruction at the given address out of the process's
311          * address space. We don't have to worry about a debugger
312          * changing this instruction before we overwrite it with our trap
313          * instruction since P_PR_LOCK is set. Since instructions can span
314          * pages, we potentially read the instruction in two parts. If the
315          * second part fails, we just zero out that part of the instruction.
316          */
317         if (uread(p, &instr[0], first, pc) != 0)
318                 return (-1);
319         if (len > first &&
320             uread(p, &instr[first], len - first, pc + first) != 0) {
321                 bzero(&instr[first], len - first);
322                 len = first;
323         }
324
325         /*
326          * If the disassembly fails, then we have a malformed instruction.
327          */
328         if ((size = dtrace_instr_size_isa(instr, p->p_model, &rmindex)) <= 0)
329                 return (-1);
330
331         /*
332          * Make sure the disassembler isn't completely broken.
333          */
334         ASSERT(-1 <= rmindex && rmindex < size);
335
336         /*
337          * If the computed size is greater than the number of bytes read,
338          * then it was a malformed instruction possibly because it fell on a
339          * page boundary and the subsequent page was missing or because of
340          * some malicious user.
341          */
342         if (size > len)
343                 return (-1);
344
345         tp->ftt_size = (uint8_t)size;
346         tp->ftt_segment = FASTTRAP_SEG_NONE;
347
348         /*
349          * Find the start of the instruction's opcode by processing any
350          * legacy prefixes.
351          */
352         for (;;) {
353                 seg = 0;
354                 switch (instr[start]) {
355                 case FASTTRAP_PREFIX_SS:
356                         seg++;
357                         /*FALLTHRU*/
358                 case FASTTRAP_PREFIX_GS:
359                         seg++;
360                         /*FALLTHRU*/
361                 case FASTTRAP_PREFIX_FS:
362                         seg++;
363                         /*FALLTHRU*/
364                 case FASTTRAP_PREFIX_ES:
365                         seg++;
366                         /*FALLTHRU*/
367                 case FASTTRAP_PREFIX_DS:
368                         seg++;
369                         /*FALLTHRU*/
370                 case FASTTRAP_PREFIX_CS:
371                         seg++;
372                         /*FALLTHRU*/
373                 case FASTTRAP_PREFIX_OPERAND:
374                 case FASTTRAP_PREFIX_ADDRESS:
375                 case FASTTRAP_PREFIX_LOCK:
376                 case FASTTRAP_PREFIX_REP:
377                 case FASTTRAP_PREFIX_REPNE:
378                         if (seg != 0) {
379                                 /*
380                                  * It's illegal for an instruction to specify
381                                  * two segment prefixes -- give up on this
382                                  * illegal instruction.
383                                  */
384                                 if (tp->ftt_segment != FASTTRAP_SEG_NONE)
385                                         return (-1);
386
387                                 tp->ftt_segment = seg;
388                         }
389                         start++;
390                         continue;
391                 }
392                 break;
393         }
394
395 #ifdef __amd64
396         /*
397          * Identify the REX prefix on 64-bit processes.
398          */
399         if (p->p_model == DATAMODEL_LP64 && (instr[start] & 0xf0) == 0x40)
400                 rex = instr[start++];
401 #endif
402
403         /*
404          * Now that we're pretty sure that the instruction is okay, copy the
405          * valid part to the tracepoint.
406          */
407         bcopy(instr, tp->ftt_instr, FASTTRAP_MAX_INSTR_SIZE);
408
409         tp->ftt_type = FASTTRAP_T_COMMON;
410         if (instr[start] == FASTTRAP_2_BYTE_OP) {
411                 switch (instr[start + 1]) {
412                 case FASTTRAP_0F_JO:
413                 case FASTTRAP_0F_JNO:
414                 case FASTTRAP_0F_JB:
415                 case FASTTRAP_0F_JAE:
416                 case FASTTRAP_0F_JE:
417                 case FASTTRAP_0F_JNE:
418                 case FASTTRAP_0F_JBE:
419                 case FASTTRAP_0F_JA:
420                 case FASTTRAP_0F_JS:
421                 case FASTTRAP_0F_JNS:
422                 case FASTTRAP_0F_JP:
423                 case FASTTRAP_0F_JNP:
424                 case FASTTRAP_0F_JL:
425                 case FASTTRAP_0F_JGE:
426                 case FASTTRAP_0F_JLE:
427                 case FASTTRAP_0F_JG:
428                         tp->ftt_type = FASTTRAP_T_JCC;
429                         tp->ftt_code = (instr[start + 1] & 0x0f) | FASTTRAP_JO;
430                         tp->ftt_dest = pc + tp->ftt_size +
431                             /* LINTED - alignment */
432                             *(int32_t *)&instr[start + 2];
433                         break;
434                 }
435         } else if (instr[start] == FASTTRAP_GROUP5_OP) {
436                 uint_t mod = FASTTRAP_MODRM_MOD(instr[start + 1]);
437                 uint_t reg = FASTTRAP_MODRM_REG(instr[start + 1]);
438                 uint_t rm = FASTTRAP_MODRM_RM(instr[start + 1]);
439
440                 if (reg == 2 || reg == 4) {
441                         uint_t i, sz;
442
443                         if (reg == 2)
444                                 tp->ftt_type = FASTTRAP_T_CALL;
445                         else
446                                 tp->ftt_type = FASTTRAP_T_JMP;
447
448                         if (mod == 3)
449                                 tp->ftt_code = 2;
450                         else
451                                 tp->ftt_code = 1;
452
453                         ASSERT(p->p_model == DATAMODEL_LP64 || rex == 0);
454
455                         /*
456                          * See AMD x86-64 Architecture Programmer's Manual
457                          * Volume 3, Section 1.2.7, Table 1-12, and
458                          * Appendix A.3.1, Table A-15.
459                          */
460                         if (mod != 3 && rm == 4) {
461                                 uint8_t sib = instr[start + 2];
462                                 uint_t index = FASTTRAP_SIB_INDEX(sib);
463                                 uint_t base = FASTTRAP_SIB_BASE(sib);
464
465                                 tp->ftt_scale = FASTTRAP_SIB_SCALE(sib);
466
467                                 tp->ftt_index = (index == 4) ?
468                                     FASTTRAP_NOREG :
469                                     regmap[index | (FASTTRAP_REX_X(rex) << 3)];
470                                 tp->ftt_base = (mod == 0 && base == 5) ?
471                                     FASTTRAP_NOREG :
472                                     regmap[base | (FASTTRAP_REX_B(rex) << 3)];
473
474                                 i = 3;
475                                 sz = mod == 1 ? 1 : 4;
476                         } else {
477                                 /*
478                                  * In 64-bit mode, mod == 0 and r/m == 5
479                                  * denotes %rip-relative addressing; in 32-bit
480                                  * mode, the base register isn't used. In both
481                                  * modes, there is a 32-bit operand.
482                                  */
483                                 if (mod == 0 && rm == 5) {
484 #ifdef __amd64
485                                         if (p->p_model == DATAMODEL_LP64)
486                                                 tp->ftt_base = REG_RIP;
487                                         else
488 #endif
489                                                 tp->ftt_base = FASTTRAP_NOREG;
490                                         sz = 4;
491                                 } else  {
492                                         uint8_t base = rm |
493                                             (FASTTRAP_REX_B(rex) << 3);
494
495                                         tp->ftt_base = regmap[base];
496                                         sz = mod == 1 ? 1 : mod == 2 ? 4 : 0;
497                                 }
498                                 tp->ftt_index = FASTTRAP_NOREG;
499                                 i = 2;
500                         }
501
502                         if (sz == 1) {
503                                 tp->ftt_dest = *(int8_t *)&instr[start + i];
504                         } else if (sz == 4) {
505                                 /* LINTED - alignment */
506                                 tp->ftt_dest = *(int32_t *)&instr[start + i];
507                         } else {
508                                 tp->ftt_dest = 0;
509                         }
510                 }
511         } else {
512                 switch (instr[start]) {
513                 case FASTTRAP_RET:
514                         tp->ftt_type = FASTTRAP_T_RET;
515                         break;
516
517                 case FASTTRAP_RET16:
518                         tp->ftt_type = FASTTRAP_T_RET16;
519                         /* LINTED - alignment */
520                         tp->ftt_dest = *(uint16_t *)&instr[start + 1];
521                         break;
522
523                 case FASTTRAP_JO:
524                 case FASTTRAP_JNO:
525                 case FASTTRAP_JB:
526                 case FASTTRAP_JAE:
527                 case FASTTRAP_JE:
528                 case FASTTRAP_JNE:
529                 case FASTTRAP_JBE:
530                 case FASTTRAP_JA:
531                 case FASTTRAP_JS:
532                 case FASTTRAP_JNS:
533                 case FASTTRAP_JP:
534                 case FASTTRAP_JNP:
535                 case FASTTRAP_JL:
536                 case FASTTRAP_JGE:
537                 case FASTTRAP_JLE:
538                 case FASTTRAP_JG:
539                         tp->ftt_type = FASTTRAP_T_JCC;
540                         tp->ftt_code = instr[start];
541                         tp->ftt_dest = pc + tp->ftt_size +
542                             (int8_t)instr[start + 1];
543                         break;
544
545                 case FASTTRAP_LOOPNZ:
546                 case FASTTRAP_LOOPZ:
547                 case FASTTRAP_LOOP:
548                         tp->ftt_type = FASTTRAP_T_LOOP;
549                         tp->ftt_code = instr[start];
550                         tp->ftt_dest = pc + tp->ftt_size +
551                             (int8_t)instr[start + 1];
552                         break;
553
554                 case FASTTRAP_JCXZ:
555                         tp->ftt_type = FASTTRAP_T_JCXZ;
556                         tp->ftt_dest = pc + tp->ftt_size +
557                             (int8_t)instr[start + 1];
558                         break;
559
560                 case FASTTRAP_CALL:
561                         tp->ftt_type = FASTTRAP_T_CALL;
562                         tp->ftt_dest = pc + tp->ftt_size +
563                             /* LINTED - alignment */
564                             *(int32_t *)&instr[start + 1];
565                         tp->ftt_code = 0;
566                         break;
567
568                 case FASTTRAP_JMP32:
569                         tp->ftt_type = FASTTRAP_T_JMP;
570                         tp->ftt_dest = pc + tp->ftt_size +
571                             /* LINTED - alignment */
572                             *(int32_t *)&instr[start + 1];
573                         break;
574                 case FASTTRAP_JMP8:
575                         tp->ftt_type = FASTTRAP_T_JMP;
576                         tp->ftt_dest = pc + tp->ftt_size +
577                             (int8_t)instr[start + 1];
578                         break;
579
580                 case FASTTRAP_PUSHL_EBP:
581                         if (start == 0)
582                                 tp->ftt_type = FASTTRAP_T_PUSHL_EBP;
583                         break;
584
585                 case FASTTRAP_NOP:
586 #ifdef __amd64
587                         ASSERT(p->p_model == DATAMODEL_LP64 || rex == 0);
588
589                         /*
590                          * On amd64 we have to be careful not to confuse a nop
591                          * (actually xchgl %eax, %eax) with an instruction using
592                          * the same opcode, but that does something different
593                          * (e.g. xchgl %r8d, %eax or xcghq %r8, %rax).
594                          */
595                         if (FASTTRAP_REX_B(rex) == 0)
596 #endif
597                                 tp->ftt_type = FASTTRAP_T_NOP;
598                         break;
599
600                 case FASTTRAP_INT3:
601                         /*
602                          * The pid provider shares the int3 trap with debugger
603                          * breakpoints so we can't instrument them.
604                          */
605                         ASSERT(instr[start] == FASTTRAP_INSTR);
606                         return (-1);
607
608                 case FASTTRAP_INT:
609                         /*
610                          * Interrupts seem like they could be traced with
611                          * no negative implications, but it's possible that
612                          * a thread could be redirected by the trap handling
613                          * code which would eventually return to the
614                          * instruction after the interrupt. If the interrupt
615                          * were in our scratch space, the subsequent
616                          * instruction might be overwritten before we return.
617                          * Accordingly we refuse to instrument any interrupt.
618                          */
619                         return (-1);
620                 }
621         }
622
623 #ifdef __amd64
624         if (p->p_model == DATAMODEL_LP64 && tp->ftt_type == FASTTRAP_T_COMMON) {
625                 /*
626                  * If the process is 64-bit and the instruction type is still
627                  * FASTTRAP_T_COMMON -- meaning we're going to copy it out an
628                  * execute it -- we need to watch for %rip-relative
629                  * addressing mode. See the portion of fasttrap_pid_probe()
630                  * below where we handle tracepoints with type
631                  * FASTTRAP_T_COMMON for how we emulate instructions that
632                  * employ %rip-relative addressing.
633                  */
634                 if (rmindex != -1) {
635                         uint_t mod = FASTTRAP_MODRM_MOD(instr[rmindex]);
636                         uint_t reg = FASTTRAP_MODRM_REG(instr[rmindex]);
637                         uint_t rm = FASTTRAP_MODRM_RM(instr[rmindex]);
638
639                         ASSERT(rmindex > start);
640
641                         if (mod == 0 && rm == 5) {
642                                 /*
643                                  * We need to be sure to avoid other
644                                  * registers used by this instruction. While
645                                  * the reg field may determine the op code
646                                  * rather than denoting a register, assuming
647                                  * that it denotes a register is always safe.
648                                  * We leave the REX field intact and use
649                                  * whatever value's there for simplicity.
650                                  */
651                                 if (reg != 0) {
652                                         tp->ftt_ripmode = FASTTRAP_RIP_1 |
653                                             (FASTTRAP_RIP_X *
654                                             FASTTRAP_REX_B(rex));
655                                         rm = 0;
656                                 } else {
657                                         tp->ftt_ripmode = FASTTRAP_RIP_2 |
658                                             (FASTTRAP_RIP_X *
659                                             FASTTRAP_REX_B(rex));
660                                         rm = 1;
661                                 }
662
663                                 tp->ftt_modrm = tp->ftt_instr[rmindex];
664                                 tp->ftt_instr[rmindex] =
665                                     FASTTRAP_MODRM(2, reg, rm);
666                         }
667                 }
668         }
669 #endif
670
671         return (0);
672 }
673
674 int
675 fasttrap_tracepoint_install(proc_t *p, fasttrap_tracepoint_t *tp)
676 {
677         fasttrap_instr_t instr = FASTTRAP_INSTR;
678
679         if (uwrite(p, &instr, 1, tp->ftt_pc) != 0)
680                 return (-1);
681
682         return (0);
683 }
684
685 int
686 fasttrap_tracepoint_remove(proc_t *p, fasttrap_tracepoint_t *tp)
687 {
688         uint8_t instr;
689
690         /*
691          * Distinguish between read or write failures and a changed
692          * instruction.
693          */
694         if (uread(p, &instr, 1, tp->ftt_pc) != 0)
695                 return (0);
696         if (instr != FASTTRAP_INSTR)
697                 return (0);
698         if (uwrite(p, &tp->ftt_instr[0], 1, tp->ftt_pc) != 0)
699                 return (-1);
700
701         return (0);
702 }
703
704 #ifdef __amd64
705 static uintptr_t
706 fasttrap_fulword_noerr(const void *uaddr)
707 {
708         uintptr_t ret;
709
710         if ((ret = fasttrap_fulword(uaddr)) != -1)
711                 return (ret);
712
713         return (0);
714 }
715 #endif
716
717 #ifdef __i386__
718 static uint32_t
719 fasttrap_fuword32_noerr(const void *uaddr)
720 {
721         uint32_t ret;
722
723         if ((ret = fasttrap_fuword32(uaddr)) != -1)
724                 return (ret);
725
726         return (0);
727 }
728 #endif
729
730 static void
731 fasttrap_return_common(struct reg *rp, uintptr_t pc, pid_t pid,
732     uintptr_t new_pc)
733 {
734         fasttrap_tracepoint_t *tp;
735         fasttrap_bucket_t *bucket;
736         fasttrap_id_t *id;
737 #if defined(sun)
738         kmutex_t *pid_mtx;
739 #endif
740
741 #if defined(sun)
742         pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock;
743         mutex_enter(pid_mtx);
744 #endif
745         bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)];
746
747         for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
748                 if (pid == tp->ftt_pid && pc == tp->ftt_pc &&
749                     tp->ftt_proc->ftpc_acount != 0)
750                         break;
751         }
752
753         /*
754          * Don't sweat it if we can't find the tracepoint again; unlike
755          * when we're in fasttrap_pid_probe(), finding the tracepoint here
756          * is not essential to the correct execution of the process.
757          */
758         if (tp == NULL) {
759 #if defined(sun)
760                 mutex_exit(pid_mtx);
761 #endif
762                 return;
763         }
764
765         for (id = tp->ftt_retids; id != NULL; id = id->fti_next) {
766                 /*
767                  * If there's a branch that could act as a return site, we
768                  * need to trace it, and check here if the program counter is
769                  * external to the function.
770                  */
771                 if (tp->ftt_type != FASTTRAP_T_RET &&
772                     tp->ftt_type != FASTTRAP_T_RET16 &&
773                     new_pc - id->fti_probe->ftp_faddr <
774                     id->fti_probe->ftp_fsize)
775                         continue;
776
777                 dtrace_probe(id->fti_probe->ftp_id,
778                     pc - id->fti_probe->ftp_faddr,
779                     rp->r_rax, rp->r_rbx, 0, 0);
780         }
781
782 #if defined(sun)
783         mutex_exit(pid_mtx);
784 #endif
785 }
786
787 static void
788 fasttrap_sigsegv(proc_t *p, kthread_t *t, uintptr_t addr)
789 {
790 #if defined(sun)
791         sigqueue_t *sqp = kmem_zalloc(sizeof (sigqueue_t), KM_SLEEP);
792
793         sqp->sq_info.si_signo = SIGSEGV;
794         sqp->sq_info.si_code = SEGV_MAPERR;
795         sqp->sq_info.si_addr = (caddr_t)addr;
796
797         mutex_enter(&p->p_lock);
798         sigaddqa(p, t, sqp);
799         mutex_exit(&p->p_lock);
800
801         if (t != NULL)
802                 aston(t);
803 #else
804         ksiginfo_t *ksi = kmem_zalloc(sizeof (ksiginfo_t), KM_SLEEP);
805
806         ksiginfo_init(ksi);
807         ksi->ksi_signo = SIGSEGV;
808         ksi->ksi_code = SEGV_MAPERR;
809         ksi->ksi_addr = (caddr_t)addr;
810         (void) tdksignal(t, SIGSEGV, ksi);
811 #endif
812 }
813
814 #ifdef __amd64
815 static void
816 fasttrap_usdt_args64(fasttrap_probe_t *probe, struct reg *rp, int argc,
817     uintptr_t *argv)
818 {
819         int i, x, cap = MIN(argc, probe->ftp_nargs);
820         uintptr_t *stack = (uintptr_t *)rp->r_rsp;
821
822         for (i = 0; i < cap; i++) {
823                 x = probe->ftp_argmap[i];
824
825                 if (x < 6)
826                         argv[i] = (&rp->r_rdi)[x];
827                 else
828                         argv[i] = fasttrap_fulword_noerr(&stack[x]);
829         }
830
831         for (; i < argc; i++) {
832                 argv[i] = 0;
833         }
834 }
835 #endif
836
837 #ifdef __i386__
838 static void
839 fasttrap_usdt_args32(fasttrap_probe_t *probe, struct reg *rp, int argc,
840     uint32_t *argv)
841 {
842         int i, x, cap = MIN(argc, probe->ftp_nargs);
843         uint32_t *stack = (uint32_t *)rp->r_rsp;
844
845         for (i = 0; i < cap; i++) {
846                 x = probe->ftp_argmap[i];
847
848                 argv[i] = fasttrap_fuword32_noerr(&stack[x]);
849         }
850
851         for (; i < argc; i++) {
852                 argv[i] = 0;
853         }
854 }
855 #endif
856
857 static int
858 fasttrap_do_seg(fasttrap_tracepoint_t *tp, struct reg *rp, uintptr_t *addr)
859 {
860         proc_t *p = curproc;
861 #ifdef __i386__
862         struct segment_descriptor *desc;
863 #else
864         struct user_segment_descriptor *desc;
865 #endif
866         uint16_t sel = 0, ndx, type;
867         uintptr_t limit;
868
869         switch (tp->ftt_segment) {
870         case FASTTRAP_SEG_CS:
871                 sel = rp->r_cs;
872                 break;
873         case FASTTRAP_SEG_DS:
874                 sel = rp->r_ds;
875                 break;
876         case FASTTRAP_SEG_ES:
877                 sel = rp->r_es;
878                 break;
879         case FASTTRAP_SEG_FS:
880                 sel = rp->r_fs;
881                 break;
882         case FASTTRAP_SEG_GS:
883                 sel = rp->r_gs;
884                 break;
885         case FASTTRAP_SEG_SS:
886                 sel = rp->r_ss;
887                 break;
888         }
889
890         /*
891          * Make sure the given segment register specifies a user priority
892          * selector rather than a kernel selector.
893          */
894         if (ISPL(sel) != SEL_UPL)
895                 return (-1);
896
897         ndx = IDXSEL(sel);
898
899         /*
900          * Check the bounds and grab the descriptor out of the specified
901          * descriptor table.
902          */
903         if (ISLDT(sel)) {
904 #ifdef __i386__
905                 if (ndx > p->p_md.md_ldt->ldt_len)
906                         return (-1);
907
908                 desc = (struct segment_descriptor *)
909                     p->p_md.md_ldt[ndx].ldt_base;
910 #else
911                 if (ndx > max_ldt_segment)
912                         return (-1);
913
914                 desc = (struct user_segment_descriptor *)
915                     p->p_md.md_ldt[ndx].ldt_base;
916 #endif
917
918         } else {
919                 if (ndx >= NGDT)
920                         return (-1);
921
922 #ifdef __i386__
923                 desc = &gdt[ndx].sd;
924 #else
925                 desc = &gdt[ndx];
926 #endif
927         }
928
929         /*
930          * The descriptor must have user privilege level and it must be
931          * present in memory.
932          */
933         if (desc->sd_dpl != SEL_UPL || desc->sd_p != 1)
934                 return (-1);
935
936         type = desc->sd_type;
937
938         /*
939          * If the S bit in the type field is not set, this descriptor can
940          * only be used in system context.
941          */
942         if ((type & 0x10) != 0x10)
943                 return (-1);
944
945         limit = USD_GETLIMIT(desc) * (desc->sd_gran ? PAGESIZE : 1);
946
947         if (tp->ftt_segment == FASTTRAP_SEG_CS) {
948                 /*
949                  * The code/data bit and readable bit must both be set.
950                  */
951                 if ((type & 0xa) != 0xa)
952                         return (-1);
953
954                 if (*addr > limit)
955                         return (-1);
956         } else {
957                 /*
958                  * The code/data bit must be clear.
959                  */
960                 if ((type & 0x8) != 0)
961                         return (-1);
962
963                 /*
964                  * If the expand-down bit is clear, we just check the limit as
965                  * it would naturally be applied. Otherwise, we need to check
966                  * that the address is the range [limit + 1 .. 0xffff] or
967                  * [limit + 1 ... 0xffffffff] depending on if the default
968                  * operand size bit is set.
969                  */
970                 if ((type & 0x4) == 0) {
971                         if (*addr > limit)
972                                 return (-1);
973                 } else if (desc->sd_def32) {
974                         if (*addr < limit + 1 || 0xffff < *addr)
975                                 return (-1);
976                 } else {
977                         if (*addr < limit + 1 || 0xffffffff < *addr)
978                                 return (-1);
979                 }
980         }
981
982         *addr += USD_GETBASE(desc);
983
984         return (0);
985 }
986
987 int
988 fasttrap_pid_probe(struct reg *rp)
989 {
990         proc_t *p = curproc;
991         uintptr_t pc = rp->r_rip - 1;
992         uintptr_t new_pc = 0;
993         fasttrap_bucket_t *bucket;
994 #if defined(sun)
995         kmutex_t *pid_mtx;
996 #endif
997         fasttrap_tracepoint_t *tp, tp_local;
998         pid_t pid;
999         dtrace_icookie_t cookie;
1000         uint_t is_enabled = 0;
1001
1002         /*
1003          * It's possible that a user (in a veritable orgy of bad planning)
1004          * could redirect this thread's flow of control before it reached the
1005          * return probe fasttrap. In this case we need to kill the process
1006          * since it's in a unrecoverable state.
1007          */
1008         if (curthread->t_dtrace_step) {
1009                 ASSERT(curthread->t_dtrace_on);
1010                 fasttrap_sigtrap(p, curthread, pc);
1011                 return (0);
1012         }
1013
1014         /*
1015          * Clear all user tracing flags.
1016          */
1017         curthread->t_dtrace_ft = 0;
1018         curthread->t_dtrace_pc = 0;
1019         curthread->t_dtrace_npc = 0;
1020         curthread->t_dtrace_scrpc = 0;
1021         curthread->t_dtrace_astpc = 0;
1022 #ifdef __amd64
1023         curthread->t_dtrace_regv = 0;
1024 #endif
1025
1026 #if defined(sun)
1027         /*
1028          * Treat a child created by a call to vfork(2) as if it were its
1029          * parent. We know that there's only one thread of control in such a
1030          * process: this one.
1031          */
1032         while (p->p_flag & SVFORK) {
1033                 p = p->p_parent;
1034         }
1035 #endif
1036
1037         PROC_LOCK(p);
1038         _PHOLD(p);
1039         pid = p->p_pid;
1040 #if defined(sun)
1041         pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock;
1042         mutex_enter(pid_mtx);
1043 #endif
1044         bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)];
1045
1046         /*
1047          * Lookup the tracepoint that the process just hit.
1048          */
1049         for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
1050                 if (pid == tp->ftt_pid && pc == tp->ftt_pc &&
1051                     tp->ftt_proc->ftpc_acount != 0)
1052                         break;
1053         }
1054
1055         /*
1056          * If we couldn't find a matching tracepoint, either a tracepoint has
1057          * been inserted without using the pid<pid> ioctl interface (see
1058          * fasttrap_ioctl), or somehow we have mislaid this tracepoint.
1059          */
1060         if (tp == NULL) {
1061 #if defined(sun)
1062                 mutex_exit(pid_mtx);
1063 #endif
1064                 _PRELE(p);
1065                 PROC_UNLOCK(p);
1066                 return (-1);
1067         }
1068
1069         /*
1070          * Set the program counter to the address of the traced instruction
1071          * so that it looks right in ustack() output.
1072          */
1073         rp->r_rip = pc;
1074
1075         if (tp->ftt_ids != NULL) {
1076                 fasttrap_id_t *id;
1077
1078 #ifdef __amd64
1079                 if (p->p_model == DATAMODEL_LP64) {
1080                         for (id = tp->ftt_ids; id != NULL; id = id->fti_next) {
1081                                 fasttrap_probe_t *probe = id->fti_probe;
1082
1083                                 if (id->fti_ptype == DTFTP_ENTRY) {
1084                                         /*
1085                                          * We note that this was an entry
1086                                          * probe to help ustack() find the
1087                                          * first caller.
1088                                          */
1089                                         cookie = dtrace_interrupt_disable();
1090                                         DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY);
1091                                         dtrace_probe(probe->ftp_id, rp->r_rdi,
1092                                             rp->r_rsi, rp->r_rdx, rp->r_rcx,
1093                                             rp->r_r8);
1094                                         DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY);
1095                                         dtrace_interrupt_enable(cookie);
1096                                 } else if (id->fti_ptype == DTFTP_IS_ENABLED) {
1097                                         /*
1098                                          * Note that in this case, we don't
1099                                          * call dtrace_probe() since it's only
1100                                          * an artificial probe meant to change
1101                                          * the flow of control so that it
1102                                          * encounters the true probe.
1103                                          */
1104                                         is_enabled = 1;
1105                                 } else if (probe->ftp_argmap == NULL) {
1106                                         dtrace_probe(probe->ftp_id, rp->r_rdi,
1107                                             rp->r_rsi, rp->r_rdx, rp->r_rcx,
1108                                             rp->r_r8);
1109                                 } else {
1110                                         uintptr_t t[5];
1111
1112                                         fasttrap_usdt_args64(probe, rp,
1113                                             sizeof (t) / sizeof (t[0]), t);
1114
1115                                         dtrace_probe(probe->ftp_id, t[0], t[1],
1116                                             t[2], t[3], t[4]);
1117                                 }
1118                         }
1119                 } else {
1120 #else /* __amd64 */
1121                         uintptr_t s0, s1, s2, s3, s4, s5;
1122                         uint32_t *stack = (uint32_t *)rp->r_esp;
1123
1124                         /*
1125                          * In 32-bit mode, all arguments are passed on the
1126                          * stack. If this is a function entry probe, we need
1127                          * to skip the first entry on the stack as it
1128                          * represents the return address rather than a
1129                          * parameter to the function.
1130                          */
1131                         s0 = fasttrap_fuword32_noerr(&stack[0]);
1132                         s1 = fasttrap_fuword32_noerr(&stack[1]);
1133                         s2 = fasttrap_fuword32_noerr(&stack[2]);
1134                         s3 = fasttrap_fuword32_noerr(&stack[3]);
1135                         s4 = fasttrap_fuword32_noerr(&stack[4]);
1136                         s5 = fasttrap_fuword32_noerr(&stack[5]);
1137
1138                         for (id = tp->ftt_ids; id != NULL; id = id->fti_next) {
1139                                 fasttrap_probe_t *probe = id->fti_probe;
1140
1141                                 if (id->fti_ptype == DTFTP_ENTRY) {
1142                                         /*
1143                                          * We note that this was an entry
1144                                          * probe to help ustack() find the
1145                                          * first caller.
1146                                          */
1147                                         cookie = dtrace_interrupt_disable();
1148                                         DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY);
1149                                         dtrace_probe(probe->ftp_id, s1, s2,
1150                                             s3, s4, s5);
1151                                         DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY);
1152                                         dtrace_interrupt_enable(cookie);
1153                                 } else if (id->fti_ptype == DTFTP_IS_ENABLED) {
1154                                         /*
1155                                          * Note that in this case, we don't
1156                                          * call dtrace_probe() since it's only
1157                                          * an artificial probe meant to change
1158                                          * the flow of control so that it
1159                                          * encounters the true probe.
1160                                          */
1161                                         is_enabled = 1;
1162                                 } else if (probe->ftp_argmap == NULL) {
1163                                         dtrace_probe(probe->ftp_id, s0, s1,
1164                                             s2, s3, s4);
1165                                 } else {
1166                                         uint32_t t[5];
1167
1168                                         fasttrap_usdt_args32(probe, rp,
1169                                             sizeof (t) / sizeof (t[0]), t);
1170
1171                                         dtrace_probe(probe->ftp_id, t[0], t[1],
1172                                             t[2], t[3], t[4]);
1173                                 }
1174                         }
1175 #endif /* __amd64 */
1176 #ifdef __amd64
1177                 }
1178 #endif
1179         }
1180
1181         /*
1182          * We're about to do a bunch of work so we cache a local copy of
1183          * the tracepoint to emulate the instruction, and then find the
1184          * tracepoint again later if we need to light up any return probes.
1185          */
1186         tp_local = *tp;
1187         PROC_UNLOCK(p);
1188 #if defined(sun)
1189         mutex_exit(pid_mtx);
1190 #endif
1191         tp = &tp_local;
1192
1193         /*
1194          * Set the program counter to appear as though the traced instruction
1195          * had completely executed. This ensures that fasttrap_getreg() will
1196          * report the expected value for REG_RIP.
1197          */
1198         rp->r_rip = pc + tp->ftt_size;
1199
1200         /*
1201          * If there's an is-enabled probe connected to this tracepoint it
1202          * means that there was a 'xorl %eax, %eax' or 'xorq %rax, %rax'
1203          * instruction that was placed there by DTrace when the binary was
1204          * linked. As this probe is, in fact, enabled, we need to stuff 1
1205          * into %eax or %rax. Accordingly, we can bypass all the instruction
1206          * emulation logic since we know the inevitable result. It's possible
1207          * that a user could construct a scenario where the 'is-enabled'
1208          * probe was on some other instruction, but that would be a rather
1209          * exotic way to shoot oneself in the foot.
1210          */
1211         if (is_enabled) {
1212                 rp->r_rax = 1;
1213                 new_pc = rp->r_rip;
1214                 goto done;
1215         }
1216
1217         /*
1218          * We emulate certain types of instructions to ensure correctness
1219          * (in the case of position dependent instructions) or optimize
1220          * common cases. The rest we have the thread execute back in user-
1221          * land.
1222          */
1223         switch (tp->ftt_type) {
1224         case FASTTRAP_T_RET:
1225         case FASTTRAP_T_RET16:
1226         {
1227                 uintptr_t dst = 0;
1228                 uintptr_t addr = 0;
1229                 int ret = 0;
1230
1231                 /*
1232                  * We have to emulate _every_ facet of the behavior of a ret
1233                  * instruction including what happens if the load from %esp
1234                  * fails; in that case, we send a SIGSEGV.
1235                  */
1236 #ifdef __amd64
1237                 if (p->p_model == DATAMODEL_NATIVE) {
1238                         ret = dst = fasttrap_fulword((void *)rp->r_rsp);
1239                         addr = rp->r_rsp + sizeof (uintptr_t);
1240                 } else {
1241 #endif
1242 #ifdef __i386__
1243                         uint32_t dst32;
1244                         ret = dst32 = fasttrap_fuword32((void *)rp->r_esp);
1245                         dst = dst32;
1246                         addr = rp->r_esp + sizeof (uint32_t);
1247 #endif
1248 #ifdef __amd64
1249                 }
1250 #endif
1251
1252                 if (ret == -1) {
1253                         fasttrap_sigsegv(p, curthread, rp->r_rsp);
1254                         new_pc = pc;
1255                         break;
1256                 }
1257
1258                 if (tp->ftt_type == FASTTRAP_T_RET16)
1259                         addr += tp->ftt_dest;
1260
1261                 rp->r_rsp = addr;
1262                 new_pc = dst;
1263                 break;
1264         }
1265
1266         case FASTTRAP_T_JCC:
1267         {
1268                 uint_t taken = 0;
1269
1270                 switch (tp->ftt_code) {
1271                 case FASTTRAP_JO:
1272                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_OF) != 0;
1273                         break;
1274                 case FASTTRAP_JNO:
1275                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0;
1276                         break;
1277                 case FASTTRAP_JB:
1278                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) != 0;
1279                         break;
1280                 case FASTTRAP_JAE:
1281                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) == 0;
1282                         break;
1283                 case FASTTRAP_JE:
1284                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0;
1285                         break;
1286                 case FASTTRAP_JNE:
1287                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0;
1288                         break;
1289                 case FASTTRAP_JBE:
1290                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) != 0 ||
1291                             (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0;
1292                         break;
1293                 case FASTTRAP_JA:
1294                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) == 0 &&
1295                             (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0;
1296                         break;
1297                 case FASTTRAP_JS:
1298                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_SF) != 0;
1299                         break;
1300                 case FASTTRAP_JNS:
1301                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0;
1302                         break;
1303                 case FASTTRAP_JP:
1304                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_PF) != 0;
1305                         break;
1306                 case FASTTRAP_JNP:
1307                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_PF) == 0;
1308                         break;
1309                 case FASTTRAP_JL:
1310                         taken = ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) !=
1311                             ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
1312                         break;
1313                 case FASTTRAP_JGE:
1314                         taken = ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) ==
1315                             ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
1316                         break;
1317                 case FASTTRAP_JLE:
1318                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0 ||
1319                             ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) !=
1320                             ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
1321                         break;
1322                 case FASTTRAP_JG:
1323                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0 &&
1324                             ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) ==
1325                             ((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
1326                         break;
1327
1328                 }
1329
1330                 if (taken)
1331                         new_pc = tp->ftt_dest;
1332                 else
1333                         new_pc = pc + tp->ftt_size;
1334                 break;
1335         }
1336
1337         case FASTTRAP_T_LOOP:
1338         {
1339                 uint_t taken = 0;
1340 #ifdef __amd64
1341                 greg_t cx = rp->r_rcx--;
1342 #else
1343                 greg_t cx = rp->r_ecx--;
1344 #endif
1345
1346                 switch (tp->ftt_code) {
1347                 case FASTTRAP_LOOPNZ:
1348                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0 &&
1349                             cx != 0;
1350                         break;
1351                 case FASTTRAP_LOOPZ:
1352                         taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0 &&
1353                             cx != 0;
1354                         break;
1355                 case FASTTRAP_LOOP:
1356                         taken = (cx != 0);
1357                         break;
1358                 }
1359
1360                 if (taken)
1361                         new_pc = tp->ftt_dest;
1362                 else
1363                         new_pc = pc + tp->ftt_size;
1364                 break;
1365         }
1366
1367         case FASTTRAP_T_JCXZ:
1368         {
1369 #ifdef __amd64
1370                 greg_t cx = rp->r_rcx;
1371 #else
1372                 greg_t cx = rp->r_ecx;
1373 #endif
1374
1375                 if (cx == 0)
1376                         new_pc = tp->ftt_dest;
1377                 else
1378                         new_pc = pc + tp->ftt_size;
1379                 break;
1380         }
1381
1382         case FASTTRAP_T_PUSHL_EBP:
1383         {
1384                 int ret = 0;
1385
1386 #ifdef __amd64
1387                 if (p->p_model == DATAMODEL_NATIVE) {
1388                         rp->r_rsp -= sizeof (uintptr_t);
1389                         ret = fasttrap_sulword((void *)rp->r_rsp, rp->r_rbp);
1390                 } else {
1391 #endif
1392 #ifdef __i386__
1393                         rp->r_rsp -= sizeof (uint32_t);
1394                         ret = fasttrap_suword32((void *)rp->r_rsp, rp->r_rbp);
1395 #endif
1396 #ifdef __amd64
1397                 }
1398 #endif
1399
1400                 if (ret == -1) {
1401                         fasttrap_sigsegv(p, curthread, rp->r_rsp);
1402                         new_pc = pc;
1403                         break;
1404                 }
1405
1406                 new_pc = pc + tp->ftt_size;
1407                 break;
1408         }
1409
1410         case FASTTRAP_T_NOP:
1411                 new_pc = pc + tp->ftt_size;
1412                 break;
1413
1414         case FASTTRAP_T_JMP:
1415         case FASTTRAP_T_CALL:
1416                 if (tp->ftt_code == 0) {
1417                         new_pc = tp->ftt_dest;
1418                 } else {
1419 #ifdef __amd64
1420                         uintptr_t value;
1421 #endif
1422                         uintptr_t addr = tp->ftt_dest;
1423
1424                         if (tp->ftt_base != FASTTRAP_NOREG)
1425                                 addr += fasttrap_getreg(rp, tp->ftt_base);
1426                         if (tp->ftt_index != FASTTRAP_NOREG)
1427                                 addr += fasttrap_getreg(rp, tp->ftt_index) <<
1428                                     tp->ftt_scale;
1429
1430                         if (tp->ftt_code == 1) {
1431                                 /*
1432                                  * If there's a segment prefix for this
1433                                  * instruction, we'll need to check permissions
1434                                  * and bounds on the given selector, and adjust
1435                                  * the address accordingly.
1436                                  */
1437                                 if (tp->ftt_segment != FASTTRAP_SEG_NONE &&
1438                                     fasttrap_do_seg(tp, rp, &addr) != 0) {
1439                                         fasttrap_sigsegv(p, curthread, addr);
1440                                         new_pc = pc;
1441                                         break;
1442                                 }
1443
1444 #ifdef __amd64
1445                                 if (p->p_model == DATAMODEL_NATIVE) {
1446                                         if ((value = fasttrap_fulword((void *)addr))
1447                                              == -1) {
1448                                                 fasttrap_sigsegv(p, curthread,
1449                                                     addr);
1450                                                 new_pc = pc;
1451                                                 break;
1452                                         }
1453                                         new_pc = value;
1454                                 } else {
1455 #endif
1456 #ifdef __i386__
1457                                         uint32_t value32;
1458                                         addr = (uintptr_t)(uint32_t)addr;
1459                                         if ((value32 = fasttrap_fuword32((void *)addr))
1460                                             == -1) {
1461                                                 fasttrap_sigsegv(p, curthread,
1462                                                     addr);
1463                                                 new_pc = pc;
1464                                                 break;
1465                                         }
1466                                         new_pc = value32;
1467 #endif
1468                                 }
1469 #ifdef __amd64
1470                         } else {
1471                                 new_pc = addr;
1472                         }
1473 #endif
1474                 }
1475
1476                 /*
1477                  * If this is a call instruction, we need to push the return
1478                  * address onto the stack. If this fails, we send the process
1479                  * a SIGSEGV and reset the pc to emulate what would happen if
1480                  * this instruction weren't traced.
1481                  */
1482                 if (tp->ftt_type == FASTTRAP_T_CALL) {
1483                         int ret = 0;
1484                         uintptr_t addr = 0, pcps;
1485 #ifdef __amd64
1486                         if (p->p_model == DATAMODEL_NATIVE) {
1487                                 addr = rp->r_rsp - sizeof (uintptr_t);
1488                                 pcps = pc + tp->ftt_size;
1489                                 ret = fasttrap_sulword((void *)addr, pcps);
1490                         } else {
1491 #endif
1492 #ifdef __i386__
1493                                 addr = rp->r_rsp - sizeof (uint32_t);
1494                                 pcps = (uint32_t)(pc + tp->ftt_size);
1495                                 ret = fasttrap_suword32((void *)addr, pcps);
1496 #endif
1497 #ifdef __amd64
1498                         }
1499 #endif
1500
1501                         if (ret == -1) {
1502                                 fasttrap_sigsegv(p, curthread, addr);
1503                                 new_pc = pc;
1504                                 break;
1505                         }
1506
1507                         rp->r_rsp = addr;
1508                 }
1509
1510                 break;
1511
1512         case FASTTRAP_T_COMMON:
1513         {
1514                 uintptr_t addr;
1515 #if defined(__amd64)
1516                 uint8_t scratch[2 * FASTTRAP_MAX_INSTR_SIZE + 22];
1517 #else
1518                 uint8_t scratch[2 * FASTTRAP_MAX_INSTR_SIZE + 7];
1519 #endif
1520                 uint_t i = 0;
1521 #if defined(sun)
1522                 klwp_t *lwp = ttolwp(curthread);
1523 #endif
1524
1525                 /*
1526                  * Compute the address of the ulwp_t and step over the
1527                  * ul_self pointer. The method used to store the user-land
1528                  * thread pointer is very different on 32- and 64-bit
1529                  * kernels.
1530                  */
1531 #if defined(sun)
1532 #if defined(__amd64)
1533                 if (p->p_model == DATAMODEL_LP64) {
1534                         addr = lwp->lwp_pcb.pcb_fsbase;
1535                         addr += sizeof (void *);
1536                 } else {
1537                         addr = lwp->lwp_pcb.pcb_gsbase;
1538                         addr += sizeof (caddr32_t);
1539                 }
1540 #else
1541                 addr = USD_GETBASE(&lwp->lwp_pcb.pcb_gsdesc);
1542                 addr += sizeof (void *);
1543 #endif
1544 #endif /* sun */
1545 #ifdef __i386__
1546                 addr = USD_GETBASE(&curthread->td_pcb->pcb_gsd);
1547 #else
1548                 addr = curthread->td_pcb->pcb_gsbase;
1549 #endif
1550                 addr += sizeof (void *);
1551
1552                 /*
1553                  * Generic Instruction Tracing
1554                  * ---------------------------
1555                  *
1556                  * This is the layout of the scratch space in the user-land
1557                  * thread structure for our generated instructions.
1558                  *
1559                  *      32-bit mode                     bytes
1560                  *      ------------------------        -----
1561                  * a:   <original instruction>          <= 15
1562                  *      jmp     <pc + tp->ftt_size>         5
1563                  * b:   <original instruction>          <= 15
1564                  *      int     T_DTRACE_RET                2
1565                  *                                      -----
1566                  *                                      <= 37
1567                  *
1568                  *      64-bit mode                     bytes
1569                  *      ------------------------        -----
1570                  * a:   <original instruction>          <= 15
1571                  *      jmp     0(%rip)                     6
1572                  *      <pc + tp->ftt_size>                 8
1573                  * b:   <original instruction>          <= 15
1574                  *      int     T_DTRACE_RET                2
1575                  *                                      -----
1576                  *                                      <= 46
1577                  *
1578                  * The %pc is set to a, and curthread->t_dtrace_astpc is set
1579                  * to b. If we encounter a signal on the way out of the
1580                  * kernel, trap() will set %pc to curthread->t_dtrace_astpc
1581                  * so that we execute the original instruction and re-enter
1582                  * the kernel rather than redirecting to the next instruction.
1583                  *
1584                  * If there are return probes (so we know that we're going to
1585                  * need to reenter the kernel after executing the original
1586                  * instruction), the scratch space will just contain the
1587                  * original instruction followed by an interrupt -- the same
1588                  * data as at b.
1589                  *
1590                  * %rip-relative Addressing
1591                  * ------------------------
1592                  *
1593                  * There's a further complication in 64-bit mode due to %rip-
1594                  * relative addressing. While this is clearly a beneficial
1595                  * architectural decision for position independent code, it's
1596                  * hard not to see it as a personal attack against the pid
1597                  * provider since before there was a relatively small set of
1598                  * instructions to emulate; with %rip-relative addressing,
1599                  * almost every instruction can potentially depend on the
1600                  * address at which it's executed. Rather than emulating
1601                  * the broad spectrum of instructions that can now be
1602                  * position dependent, we emulate jumps and others as in
1603                  * 32-bit mode, and take a different tack for instructions
1604                  * using %rip-relative addressing.
1605                  *
1606                  * For every instruction that uses the ModRM byte, the
1607                  * in-kernel disassembler reports its location. We use the
1608                  * ModRM byte to identify that an instruction uses
1609                  * %rip-relative addressing and to see what other registers
1610                  * the instruction uses. To emulate those instructions,
1611                  * we modify the instruction to be %rax-relative rather than
1612                  * %rip-relative (or %rcx-relative if the instruction uses
1613                  * %rax; or %r8- or %r9-relative if the REX.B is present so
1614                  * we don't have to rewrite the REX prefix). We then load
1615                  * the value that %rip would have been into the scratch
1616                  * register and generate an instruction to reset the scratch
1617                  * register back to its original value. The instruction
1618                  * sequence looks like this:
1619                  *
1620                  *      64-mode %rip-relative           bytes
1621                  *      ------------------------        -----
1622                  * a:   <modified instruction>          <= 15
1623                  *      movq    $<value>, %<scratch>        6
1624                  *      jmp     0(%rip)                     6
1625                  *      <pc + tp->ftt_size>                 8
1626                  * b:   <modified instruction>          <= 15
1627                  *      int     T_DTRACE_RET                2
1628                  *                                      -----
1629                  *                                         52
1630                  *
1631                  * We set curthread->t_dtrace_regv so that upon receiving
1632                  * a signal we can reset the value of the scratch register.
1633                  */
1634
1635                 ASSERT(tp->ftt_size < FASTTRAP_MAX_INSTR_SIZE);
1636
1637                 curthread->t_dtrace_scrpc = addr;
1638                 bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size);
1639                 i += tp->ftt_size;
1640
1641 #ifdef __amd64
1642                 if (tp->ftt_ripmode != 0) {
1643                         greg_t *reg = NULL;
1644
1645                         ASSERT(p->p_model == DATAMODEL_LP64);
1646                         ASSERT(tp->ftt_ripmode &
1647                             (FASTTRAP_RIP_1 | FASTTRAP_RIP_2));
1648
1649                         /*
1650                          * If this was a %rip-relative instruction, we change
1651                          * it to be either a %rax- or %rcx-relative
1652                          * instruction (depending on whether those registers
1653                          * are used as another operand; or %r8- or %r9-
1654                          * relative depending on the value of REX.B). We then
1655                          * set that register and generate a movq instruction
1656                          * to reset the value.
1657                          */
1658                         if (tp->ftt_ripmode & FASTTRAP_RIP_X)
1659                                 scratch[i++] = FASTTRAP_REX(1, 0, 0, 1);
1660                         else
1661                                 scratch[i++] = FASTTRAP_REX(1, 0, 0, 0);
1662
1663                         if (tp->ftt_ripmode & FASTTRAP_RIP_1)
1664                                 scratch[i++] = FASTTRAP_MOV_EAX;
1665                         else
1666                                 scratch[i++] = FASTTRAP_MOV_ECX;
1667
1668                         switch (tp->ftt_ripmode) {
1669                         case FASTTRAP_RIP_1:
1670                                 reg = &rp->r_rax;
1671                                 curthread->t_dtrace_reg = REG_RAX;
1672                                 break;
1673                         case FASTTRAP_RIP_2:
1674                                 reg = &rp->r_rcx;
1675                                 curthread->t_dtrace_reg = REG_RCX;
1676                                 break;
1677                         case FASTTRAP_RIP_1 | FASTTRAP_RIP_X:
1678                                 reg = &rp->r_r8;
1679                                 curthread->t_dtrace_reg = REG_R8;
1680                                 break;
1681                         case FASTTRAP_RIP_2 | FASTTRAP_RIP_X:
1682                                 reg = &rp->r_r9;
1683                                 curthread->t_dtrace_reg = REG_R9;
1684                                 break;
1685                         }
1686
1687                         /* LINTED - alignment */
1688                         *(uint64_t *)&scratch[i] = *reg;
1689                         curthread->t_dtrace_regv = *reg;
1690                         *reg = pc + tp->ftt_size;
1691                         i += sizeof (uint64_t);
1692                 }
1693 #endif
1694
1695                 /*
1696                  * Generate the branch instruction to what would have
1697                  * normally been the subsequent instruction. In 32-bit mode,
1698                  * this is just a relative branch; in 64-bit mode this is a
1699                  * %rip-relative branch that loads the 64-bit pc value
1700                  * immediately after the jmp instruction.
1701                  */
1702 #ifdef __amd64
1703                 if (p->p_model == DATAMODEL_LP64) {
1704                         scratch[i++] = FASTTRAP_GROUP5_OP;
1705                         scratch[i++] = FASTTRAP_MODRM(0, 4, 5);
1706                         /* LINTED - alignment */
1707                         *(uint32_t *)&scratch[i] = 0;
1708                         i += sizeof (uint32_t);
1709                         /* LINTED - alignment */
1710                         *(uint64_t *)&scratch[i] = pc + tp->ftt_size;
1711                         i += sizeof (uint64_t);
1712                 } else {
1713 #endif
1714 #ifdef __i386__
1715                         /*
1716                          * Set up the jmp to the next instruction; note that
1717                          * the size of the traced instruction cancels out.
1718                          */
1719                         scratch[i++] = FASTTRAP_JMP32;
1720                         /* LINTED - alignment */
1721                         *(uint32_t *)&scratch[i] = pc - addr - 5;
1722                         i += sizeof (uint32_t);
1723 #endif
1724 #ifdef __amd64
1725                 }
1726 #endif
1727
1728                 curthread->t_dtrace_astpc = addr + i;
1729                 bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size);
1730                 i += tp->ftt_size;
1731                 scratch[i++] = FASTTRAP_INT;
1732                 scratch[i++] = T_DTRACE_RET;
1733
1734                 ASSERT(i <= sizeof (scratch));
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
1798         PROC_LOCK(p);
1799         proc_write_regs(curthread, rp);
1800         _PRELE(p);
1801         PROC_UNLOCK(p);
1802
1803         return (0);
1804 }
1805
1806 int
1807 fasttrap_return_probe(struct reg *rp)
1808 {
1809         proc_t *p = curproc;
1810         uintptr_t pc = curthread->t_dtrace_pc;
1811         uintptr_t npc = curthread->t_dtrace_npc;
1812
1813         curthread->t_dtrace_pc = 0;
1814         curthread->t_dtrace_npc = 0;
1815         curthread->t_dtrace_scrpc = 0;
1816         curthread->t_dtrace_astpc = 0;
1817
1818 #if defined(sun)
1819         /*
1820          * Treat a child created by a call to vfork(2) as if it were its
1821          * parent. We know that there's only one thread of control in such a
1822          * process: this one.
1823          */
1824         while (p->p_flag & SVFORK) {
1825                 p = p->p_parent;
1826         }
1827 #endif
1828
1829         /*
1830          * We set rp->r_rip to the address of the traced instruction so
1831          * that it appears to dtrace_probe() that we're on the original
1832          * instruction, and so that the user can't easily detect our
1833          * complex web of lies. dtrace_return_probe() (our caller)
1834          * will correctly set %pc after we return.
1835          */
1836         rp->r_rip = pc;
1837
1838         fasttrap_return_common(rp, pc, p->p_pid, npc);
1839
1840         return (0);
1841 }
1842
1843 /*ARGSUSED*/
1844 uint64_t
1845 fasttrap_pid_getarg(void *arg, dtrace_id_t id, void *parg, int argno,
1846     int aframes)
1847 {
1848         struct reg r;
1849
1850         fill_regs(curthread, &r);
1851
1852         return (fasttrap_anarg(&r, 1, argno));
1853 }
1854
1855 /*ARGSUSED*/
1856 uint64_t
1857 fasttrap_usdt_getarg(void *arg, dtrace_id_t id, void *parg, int argno,
1858     int aframes)
1859 {
1860         struct reg r;
1861
1862         fill_regs(curthread, &r);
1863
1864         return (fasttrap_anarg(&r, 0, argno));
1865 }
1866
1867 static ulong_t
1868 fasttrap_getreg(struct reg *rp, uint_t reg)
1869 {
1870 #ifdef __amd64
1871         switch (reg) {
1872         case REG_R15:           return (rp->r_r15);
1873         case REG_R14:           return (rp->r_r14);
1874         case REG_R13:           return (rp->r_r13);
1875         case REG_R12:           return (rp->r_r12);
1876         case REG_R11:           return (rp->r_r11);
1877         case REG_R10:           return (rp->r_r10);
1878         case REG_R9:            return (rp->r_r9);
1879         case REG_R8:            return (rp->r_r8);
1880         case REG_RDI:           return (rp->r_rdi);
1881         case REG_RSI:           return (rp->r_rsi);
1882         case REG_RBP:           return (rp->r_rbp);
1883         case REG_RBX:           return (rp->r_rbx);
1884         case REG_RDX:           return (rp->r_rdx);
1885         case REG_RCX:           return (rp->r_rcx);
1886         case REG_RAX:           return (rp->r_rax);
1887         case REG_TRAPNO:        return (rp->r_trapno);
1888         case REG_ERR:           return (rp->r_err);
1889         case REG_RIP:           return (rp->r_rip);
1890         case REG_CS:            return (rp->r_cs);
1891 #if defined(sun)
1892         case REG_RFL:           return (rp->r_rfl);
1893 #endif
1894         case REG_RSP:           return (rp->r_rsp);
1895         case REG_SS:            return (rp->r_ss);
1896         case REG_FS:            return (rp->r_fs);
1897         case REG_GS:            return (rp->r_gs);
1898         case REG_DS:            return (rp->r_ds);
1899         case REG_ES:            return (rp->r_es);
1900         case REG_FSBASE:        return (rdmsr(MSR_FSBASE));
1901         case REG_GSBASE:        return (rdmsr(MSR_GSBASE));
1902         }
1903
1904         panic("dtrace: illegal register constant");
1905         /*NOTREACHED*/
1906 #else
1907 #define _NGREG 19
1908         if (reg >= _NGREG)
1909                 panic("dtrace: illegal register constant");
1910
1911         return (((greg_t *)&rp->r_gs)[reg]);
1912 #endif
1913 }