xref: /openbmc/linux/arch/x86/kernel/uprobes.c (revision 293d5b43)
1 /*
2  * User-space Probes (UProbes) for x86
3  *
4  * This program is free software; you can redistribute it and/or modify
5  * it under the terms of the GNU General Public License as published by
6  * the Free Software Foundation; either version 2 of the License, or
7  * (at your option) any later version.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write to the Free Software
16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
17  *
18  * Copyright (C) IBM Corporation, 2008-2011
19  * Authors:
20  *	Srikar Dronamraju
21  *	Jim Keniston
22  */
23 #include <linux/kernel.h>
24 #include <linux/sched.h>
25 #include <linux/ptrace.h>
26 #include <linux/uprobes.h>
27 #include <linux/uaccess.h>
28 
29 #include <linux/kdebug.h>
30 #include <asm/processor.h>
31 #include <asm/insn.h>
32 #include <asm/mmu_context.h>
33 
34 /* Post-execution fixups. */
35 
36 /* Adjust IP back to vicinity of actual insn */
37 #define UPROBE_FIX_IP		0x01
38 
39 /* Adjust the return address of a call insn */
40 #define UPROBE_FIX_CALL		0x02
41 
42 /* Instruction will modify TF, don't change it */
43 #define UPROBE_FIX_SETF		0x04
44 
45 #define UPROBE_FIX_RIP_SI	0x08
46 #define UPROBE_FIX_RIP_DI	0x10
47 #define UPROBE_FIX_RIP_BX	0x20
48 #define UPROBE_FIX_RIP_MASK	\
49 	(UPROBE_FIX_RIP_SI | UPROBE_FIX_RIP_DI | UPROBE_FIX_RIP_BX)
50 
51 #define	UPROBE_TRAP_NR		UINT_MAX
52 
53 /* Adaptations for mhiramat x86 decoder v14. */
54 #define OPCODE1(insn)		((insn)->opcode.bytes[0])
55 #define OPCODE2(insn)		((insn)->opcode.bytes[1])
56 #define OPCODE3(insn)		((insn)->opcode.bytes[2])
57 #define MODRM_REG(insn)		X86_MODRM_REG((insn)->modrm.value)
58 
59 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
60 	(((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
61 	  (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
62 	  (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
63 	  (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
64 	 << (row % 32))
65 
66 /*
67  * Good-instruction tables for 32-bit apps.  This is non-const and volatile
68  * to keep gcc from statically optimizing it out, as variable_test_bit makes
69  * some versions of gcc to think only *(unsigned long*) is used.
70  *
71  * Opcodes we'll probably never support:
72  * 6c-6f - ins,outs. SEGVs if used in userspace
73  * e4-e7 - in,out imm. SEGVs if used in userspace
74  * ec-ef - in,out acc. SEGVs if used in userspace
75  * cc - int3. SIGTRAP if used in userspace
76  * ce - into. Not used in userspace - no kernel support to make it useful. SEGVs
77  *	(why we support bound (62) then? it's similar, and similarly unused...)
78  * f1 - int1. SIGTRAP if used in userspace
79  * f4 - hlt. SEGVs if used in userspace
80  * fa - cli. SEGVs if used in userspace
81  * fb - sti. SEGVs if used in userspace
82  *
83  * Opcodes which need some work to be supported:
84  * 07,17,1f - pop es/ss/ds
85  *	Normally not used in userspace, but would execute if used.
86  *	Can cause GP or stack exception if tries to load wrong segment descriptor.
87  *	We hesitate to run them under single step since kernel's handling
88  *	of userspace single-stepping (TF flag) is fragile.
89  *	We can easily refuse to support push es/cs/ss/ds (06/0e/16/1e)
90  *	on the same grounds that they are never used.
91  * cd - int N.
92  *	Used by userspace for "int 80" syscall entry. (Other "int N"
93  *	cause GP -> SEGV since their IDT gates don't allow calls from CPL 3).
94  *	Not supported since kernel's handling of userspace single-stepping
95  *	(TF flag) is fragile.
96  * cf - iret. Normally not used in userspace. Doesn't SEGV unless arguments are bad
97  */
98 #if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION)
99 static volatile u32 good_insns_32[256 / 32] = {
100 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
101 	/*      ----------------------------------------------         */
102 	W(0x00, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* 00 */
103 	W(0x10, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0) , /* 10 */
104 	W(0x20, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 20 */
105 	W(0x30, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 30 */
106 	W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
107 	W(0x50, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 50 */
108 	W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* 60 */
109 	W(0x70, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 70 */
110 	W(0x80, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 80 */
111 	W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
112 	W(0xa0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* a0 */
113 	W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* b0 */
114 	W(0xc0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* c0 */
115 	W(0xd0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* d0 */
116 	W(0xe0, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0) | /* e0 */
117 	W(0xf0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1)   /* f0 */
118 	/*      ----------------------------------------------         */
119 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
120 };
121 #else
122 #define good_insns_32	NULL
123 #endif
124 
125 /* Good-instruction tables for 64-bit apps.
126  *
127  * Genuinely invalid opcodes:
128  * 06,07 - formerly push/pop es
129  * 0e - formerly push cs
130  * 16,17 - formerly push/pop ss
131  * 1e,1f - formerly push/pop ds
132  * 27,2f,37,3f - formerly daa/das/aaa/aas
133  * 60,61 - formerly pusha/popa
134  * 62 - formerly bound. EVEX prefix for AVX512 (not yet supported)
135  * 82 - formerly redundant encoding of Group1
136  * 9a - formerly call seg:ofs
137  * ce - formerly into
138  * d4,d5 - formerly aam/aad
139  * d6 - formerly undocumented salc
140  * ea - formerly jmp seg:ofs
141  *
142  * Opcodes we'll probably never support:
143  * 6c-6f - ins,outs. SEGVs if used in userspace
144  * e4-e7 - in,out imm. SEGVs if used in userspace
145  * ec-ef - in,out acc. SEGVs if used in userspace
146  * cc - int3. SIGTRAP if used in userspace
147  * f1 - int1. SIGTRAP if used in userspace
148  * f4 - hlt. SEGVs if used in userspace
149  * fa - cli. SEGVs if used in userspace
150  * fb - sti. SEGVs if used in userspace
151  *
152  * Opcodes which need some work to be supported:
153  * cd - int N.
154  *	Used by userspace for "int 80" syscall entry. (Other "int N"
155  *	cause GP -> SEGV since their IDT gates don't allow calls from CPL 3).
156  *	Not supported since kernel's handling of userspace single-stepping
157  *	(TF flag) is fragile.
158  * cf - iret. Normally not used in userspace. Doesn't SEGV unless arguments are bad
159  */
160 #if defined(CONFIG_X86_64)
161 static volatile u32 good_insns_64[256 / 32] = {
162 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
163 	/*      ----------------------------------------------         */
164 	W(0x00, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 1) | /* 00 */
165 	W(0x10, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0) , /* 10 */
166 	W(0x20, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0) | /* 20 */
167 	W(0x30, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0) , /* 30 */
168 	W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
169 	W(0x50, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 50 */
170 	W(0x60, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* 60 */
171 	W(0x70, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 70 */
172 	W(0x80, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 80 */
173 	W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1) , /* 90 */
174 	W(0xa0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* a0 */
175 	W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* b0 */
176 	W(0xc0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* c0 */
177 	W(0xd0, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* d0 */
178 	W(0xe0, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0) | /* e0 */
179 	W(0xf0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1)   /* f0 */
180 	/*      ----------------------------------------------         */
181 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
182 };
183 #else
184 #define good_insns_64	NULL
185 #endif
186 
187 /* Using this for both 64-bit and 32-bit apps.
188  * Opcodes we don't support:
189  * 0f 00 - SLDT/STR/LLDT/LTR/VERR/VERW/-/- group. System insns
190  * 0f 01 - SGDT/SIDT/LGDT/LIDT/SMSW/-/LMSW/INVLPG group.
191  *	Also encodes tons of other system insns if mod=11.
192  *	Some are in fact non-system: xend, xtest, rdtscp, maybe more
193  * 0f 05 - syscall
194  * 0f 06 - clts (CPL0 insn)
195  * 0f 07 - sysret
196  * 0f 08 - invd (CPL0 insn)
197  * 0f 09 - wbinvd (CPL0 insn)
198  * 0f 0b - ud2
199  * 0f 30 - wrmsr (CPL0 insn) (then why rdmsr is allowed, it's also CPL0 insn?)
200  * 0f 34 - sysenter
201  * 0f 35 - sysexit
202  * 0f 37 - getsec
203  * 0f 78 - vmread (Intel VMX. CPL0 insn)
204  * 0f 79 - vmwrite (Intel VMX. CPL0 insn)
205  *	Note: with prefixes, these two opcodes are
206  *	extrq/insertq/AVX512 convert vector ops.
207  * 0f ae - group15: [f]xsave,[f]xrstor,[v]{ld,st}mxcsr,clflush[opt],
208  *	{rd,wr}{fs,gs}base,{s,l,m}fence.
209  *	Why? They are all user-executable.
210  */
211 static volatile u32 good_2byte_insns[256 / 32] = {
212 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
213 	/*      ----------------------------------------------         */
214 	W(0x00, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1) | /* 00 */
215 	W(0x10, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 10 */
216 	W(0x20, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 20 */
217 	W(0x30, 0, 1, 1, 1, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1) , /* 30 */
218 	W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
219 	W(0x50, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 50 */
220 	W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 60 */
221 	W(0x70, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1) , /* 70 */
222 	W(0x80, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 80 */
223 	W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
224 	W(0xa0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1) | /* a0 */
225 	W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* b0 */
226 	W(0xc0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
227 	W(0xd0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* d0 */
228 	W(0xe0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* e0 */
229 	W(0xf0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1)   /* f0 */
230 	/*      ----------------------------------------------         */
231 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
232 };
233 #undef W
234 
235 /*
236  * opcodes we may need to refine support for:
237  *
238  *  0f - 2-byte instructions: For many of these instructions, the validity
239  *  depends on the prefix and/or the reg field.  On such instructions, we
240  *  just consider the opcode combination valid if it corresponds to any
241  *  valid instruction.
242  *
243  *  8f - Group 1 - only reg = 0 is OK
244  *  c6-c7 - Group 11 - only reg = 0 is OK
245  *  d9-df - fpu insns with some illegal encodings
246  *  f2, f3 - repnz, repz prefixes.  These are also the first byte for
247  *  certain floating-point instructions, such as addsd.
248  *
249  *  fe - Group 4 - only reg = 0 or 1 is OK
250  *  ff - Group 5 - only reg = 0-6 is OK
251  *
252  * others -- Do we need to support these?
253  *
254  *  0f - (floating-point?) prefetch instructions
255  *  07, 17, 1f - pop es, pop ss, pop ds
256  *  26, 2e, 36, 3e - es:, cs:, ss:, ds: segment prefixes --
257  *	but 64 and 65 (fs: and gs:) seem to be used, so we support them
258  *  67 - addr16 prefix
259  *  ce - into
260  *  f0 - lock prefix
261  */
262 
263 /*
264  * TODO:
265  * - Where necessary, examine the modrm byte and allow only valid instructions
266  * in the different Groups and fpu instructions.
267  */
268 
269 static bool is_prefix_bad(struct insn *insn)
270 {
271 	int i;
272 
273 	for (i = 0; i < insn->prefixes.nbytes; i++) {
274 		switch (insn->prefixes.bytes[i]) {
275 		case 0x26:	/* INAT_PFX_ES   */
276 		case 0x2E:	/* INAT_PFX_CS   */
277 		case 0x36:	/* INAT_PFX_DS   */
278 		case 0x3E:	/* INAT_PFX_SS   */
279 		case 0xF0:	/* INAT_PFX_LOCK */
280 			return true;
281 		}
282 	}
283 	return false;
284 }
285 
286 static int uprobe_init_insn(struct arch_uprobe *auprobe, struct insn *insn, bool x86_64)
287 {
288 	u32 volatile *good_insns;
289 
290 	insn_init(insn, auprobe->insn, sizeof(auprobe->insn), x86_64);
291 	/* has the side-effect of processing the entire instruction */
292 	insn_get_length(insn);
293 	if (WARN_ON_ONCE(!insn_complete(insn)))
294 		return -ENOEXEC;
295 
296 	if (is_prefix_bad(insn))
297 		return -ENOTSUPP;
298 
299 	if (x86_64)
300 		good_insns = good_insns_64;
301 	else
302 		good_insns = good_insns_32;
303 
304 	if (test_bit(OPCODE1(insn), (unsigned long *)good_insns))
305 		return 0;
306 
307 	if (insn->opcode.nbytes == 2) {
308 		if (test_bit(OPCODE2(insn), (unsigned long *)good_2byte_insns))
309 			return 0;
310 	}
311 
312 	return -ENOTSUPP;
313 }
314 
315 #ifdef CONFIG_X86_64
316 /*
317  * If arch_uprobe->insn doesn't use rip-relative addressing, return
318  * immediately.  Otherwise, rewrite the instruction so that it accesses
319  * its memory operand indirectly through a scratch register.  Set
320  * defparam->fixups accordingly. (The contents of the scratch register
321  * will be saved before we single-step the modified instruction,
322  * and restored afterward).
323  *
324  * We do this because a rip-relative instruction can access only a
325  * relatively small area (+/- 2 GB from the instruction), and the XOL
326  * area typically lies beyond that area.  At least for instructions
327  * that store to memory, we can't execute the original instruction
328  * and "fix things up" later, because the misdirected store could be
329  * disastrous.
330  *
331  * Some useful facts about rip-relative instructions:
332  *
333  *  - There's always a modrm byte with bit layout "00 reg 101".
334  *  - There's never a SIB byte.
335  *  - The displacement is always 4 bytes.
336  *  - REX.B=1 bit in REX prefix, which normally extends r/m field,
337  *    has no effect on rip-relative mode. It doesn't make modrm byte
338  *    with r/m=101 refer to register 1101 = R13.
339  */
340 static void riprel_analyze(struct arch_uprobe *auprobe, struct insn *insn)
341 {
342 	u8 *cursor;
343 	u8 reg;
344 	u8 reg2;
345 
346 	if (!insn_rip_relative(insn))
347 		return;
348 
349 	/*
350 	 * insn_rip_relative() would have decoded rex_prefix, vex_prefix, modrm.
351 	 * Clear REX.b bit (extension of MODRM.rm field):
352 	 * we want to encode low numbered reg, not r8+.
353 	 */
354 	if (insn->rex_prefix.nbytes) {
355 		cursor = auprobe->insn + insn_offset_rex_prefix(insn);
356 		/* REX byte has 0100wrxb layout, clearing REX.b bit */
357 		*cursor &= 0xfe;
358 	}
359 	/*
360 	 * Similar treatment for VEX3/EVEX prefix.
361 	 * TODO: add XOP treatment when insn decoder supports them
362 	 */
363 	if (insn->vex_prefix.nbytes >= 3) {
364 		/*
365 		 * vex2:     c5    rvvvvLpp   (has no b bit)
366 		 * vex3/xop: c4/8f rxbmmmmm wvvvvLpp
367 		 * evex:     62    rxbR00mm wvvvv1pp zllBVaaa
368 		 * Setting VEX3.b (setting because it has inverted meaning).
369 		 * Setting EVEX.x since (in non-SIB encoding) EVEX.x
370 		 * is the 4th bit of MODRM.rm, and needs the same treatment.
371 		 * For VEX3-encoded insns, VEX3.x value has no effect in
372 		 * non-SIB encoding, the change is superfluous but harmless.
373 		 */
374 		cursor = auprobe->insn + insn_offset_vex_prefix(insn) + 1;
375 		*cursor |= 0x60;
376 	}
377 
378 	/*
379 	 * Convert from rip-relative addressing to register-relative addressing
380 	 * via a scratch register.
381 	 *
382 	 * This is tricky since there are insns with modrm byte
383 	 * which also use registers not encoded in modrm byte:
384 	 * [i]div/[i]mul: implicitly use dx:ax
385 	 * shift ops: implicitly use cx
386 	 * cmpxchg: implicitly uses ax
387 	 * cmpxchg8/16b: implicitly uses dx:ax and bx:cx
388 	 *   Encoding: 0f c7/1 modrm
389 	 *   The code below thinks that reg=1 (cx), chooses si as scratch.
390 	 * mulx: implicitly uses dx: mulx r/m,r1,r2 does r1:r2 = dx * r/m.
391 	 *   First appeared in Haswell (BMI2 insn). It is vex-encoded.
392 	 *   Example where none of bx,cx,dx can be used as scratch reg:
393 	 *   c4 e2 63 f6 0d disp32   mulx disp32(%rip),%ebx,%ecx
394 	 * [v]pcmpistri: implicitly uses cx, xmm0
395 	 * [v]pcmpistrm: implicitly uses xmm0
396 	 * [v]pcmpestri: implicitly uses ax, dx, cx, xmm0
397 	 * [v]pcmpestrm: implicitly uses ax, dx, xmm0
398 	 *   Evil SSE4.2 string comparison ops from hell.
399 	 * maskmovq/[v]maskmovdqu: implicitly uses (ds:rdi) as destination.
400 	 *   Encoding: 0f f7 modrm, 66 0f f7 modrm, vex-encoded: c5 f9 f7 modrm.
401 	 *   Store op1, byte-masked by op2 msb's in each byte, to (ds:rdi).
402 	 *   AMD says it has no 3-operand form (vex.vvvv must be 1111)
403 	 *   and that it can have only register operands, not mem
404 	 *   (its modrm byte must have mode=11).
405 	 *   If these restrictions will ever be lifted,
406 	 *   we'll need code to prevent selection of di as scratch reg!
407 	 *
408 	 * Summary: I don't know any insns with modrm byte which
409 	 * use SI register implicitly. DI register is used only
410 	 * by one insn (maskmovq) and BX register is used
411 	 * only by one too (cmpxchg8b).
412 	 * BP is stack-segment based (may be a problem?).
413 	 * AX, DX, CX are off-limits (many implicit users).
414 	 * SP is unusable (it's stack pointer - think about "pop mem";
415 	 * also, rsp+disp32 needs sib encoding -> insn length change).
416 	 */
417 
418 	reg = MODRM_REG(insn);	/* Fetch modrm.reg */
419 	reg2 = 0xff;		/* Fetch vex.vvvv */
420 	if (insn->vex_prefix.nbytes)
421 		reg2 = insn->vex_prefix.bytes[2];
422 	/*
423 	 * TODO: add XOP vvvv reading.
424 	 *
425 	 * vex.vvvv field is in bits 6-3, bits are inverted.
426 	 * But in 32-bit mode, high-order bit may be ignored.
427 	 * Therefore, let's consider only 3 low-order bits.
428 	 */
429 	reg2 = ((reg2 >> 3) & 0x7) ^ 0x7;
430 	/*
431 	 * Register numbering is ax,cx,dx,bx, sp,bp,si,di, r8..r15.
432 	 *
433 	 * Choose scratch reg. Order is important: must not select bx
434 	 * if we can use si (cmpxchg8b case!)
435 	 */
436 	if (reg != 6 && reg2 != 6) {
437 		reg2 = 6;
438 		auprobe->defparam.fixups |= UPROBE_FIX_RIP_SI;
439 	} else if (reg != 7 && reg2 != 7) {
440 		reg2 = 7;
441 		auprobe->defparam.fixups |= UPROBE_FIX_RIP_DI;
442 		/* TODO (paranoia): force maskmovq to not use di */
443 	} else {
444 		reg2 = 3;
445 		auprobe->defparam.fixups |= UPROBE_FIX_RIP_BX;
446 	}
447 	/*
448 	 * Point cursor at the modrm byte.  The next 4 bytes are the
449 	 * displacement.  Beyond the displacement, for some instructions,
450 	 * is the immediate operand.
451 	 */
452 	cursor = auprobe->insn + insn_offset_modrm(insn);
453 	/*
454 	 * Change modrm from "00 reg 101" to "10 reg reg2". Example:
455 	 * 89 05 disp32  mov %eax,disp32(%rip) becomes
456 	 * 89 86 disp32  mov %eax,disp32(%rsi)
457 	 */
458 	*cursor = 0x80 | (reg << 3) | reg2;
459 }
460 
461 static inline unsigned long *
462 scratch_reg(struct arch_uprobe *auprobe, struct pt_regs *regs)
463 {
464 	if (auprobe->defparam.fixups & UPROBE_FIX_RIP_SI)
465 		return &regs->si;
466 	if (auprobe->defparam.fixups & UPROBE_FIX_RIP_DI)
467 		return &regs->di;
468 	return &regs->bx;
469 }
470 
471 /*
472  * If we're emulating a rip-relative instruction, save the contents
473  * of the scratch register and store the target address in that register.
474  */
475 static void riprel_pre_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
476 {
477 	if (auprobe->defparam.fixups & UPROBE_FIX_RIP_MASK) {
478 		struct uprobe_task *utask = current->utask;
479 		unsigned long *sr = scratch_reg(auprobe, regs);
480 
481 		utask->autask.saved_scratch_register = *sr;
482 		*sr = utask->vaddr + auprobe->defparam.ilen;
483 	}
484 }
485 
486 static void riprel_post_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
487 {
488 	if (auprobe->defparam.fixups & UPROBE_FIX_RIP_MASK) {
489 		struct uprobe_task *utask = current->utask;
490 		unsigned long *sr = scratch_reg(auprobe, regs);
491 
492 		*sr = utask->autask.saved_scratch_register;
493 	}
494 }
495 #else /* 32-bit: */
496 /*
497  * No RIP-relative addressing on 32-bit
498  */
499 static void riprel_analyze(struct arch_uprobe *auprobe, struct insn *insn)
500 {
501 }
502 static void riprel_pre_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
503 {
504 }
505 static void riprel_post_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
506 {
507 }
508 #endif /* CONFIG_X86_64 */
509 
510 struct uprobe_xol_ops {
511 	bool	(*emulate)(struct arch_uprobe *, struct pt_regs *);
512 	int	(*pre_xol)(struct arch_uprobe *, struct pt_regs *);
513 	int	(*post_xol)(struct arch_uprobe *, struct pt_regs *);
514 	void	(*abort)(struct arch_uprobe *, struct pt_regs *);
515 };
516 
517 static inline int sizeof_long(void)
518 {
519 	return in_ia32_syscall() ? 4 : 8;
520 }
521 
522 static int default_pre_xol_op(struct arch_uprobe *auprobe, struct pt_regs *regs)
523 {
524 	riprel_pre_xol(auprobe, regs);
525 	return 0;
526 }
527 
528 static int push_ret_address(struct pt_regs *regs, unsigned long ip)
529 {
530 	unsigned long new_sp = regs->sp - sizeof_long();
531 
532 	if (copy_to_user((void __user *)new_sp, &ip, sizeof_long()))
533 		return -EFAULT;
534 
535 	regs->sp = new_sp;
536 	return 0;
537 }
538 
539 /*
540  * We have to fix things up as follows:
541  *
542  * Typically, the new ip is relative to the copied instruction.  We need
543  * to make it relative to the original instruction (FIX_IP).  Exceptions
544  * are return instructions and absolute or indirect jump or call instructions.
545  *
546  * If the single-stepped instruction was a call, the return address that
547  * is atop the stack is the address following the copied instruction.  We
548  * need to make it the address following the original instruction (FIX_CALL).
549  *
550  * If the original instruction was a rip-relative instruction such as
551  * "movl %edx,0xnnnn(%rip)", we have instead executed an equivalent
552  * instruction using a scratch register -- e.g., "movl %edx,0xnnnn(%rsi)".
553  * We need to restore the contents of the scratch register
554  * (FIX_RIP_reg).
555  */
556 static int default_post_xol_op(struct arch_uprobe *auprobe, struct pt_regs *regs)
557 {
558 	struct uprobe_task *utask = current->utask;
559 
560 	riprel_post_xol(auprobe, regs);
561 	if (auprobe->defparam.fixups & UPROBE_FIX_IP) {
562 		long correction = utask->vaddr - utask->xol_vaddr;
563 		regs->ip += correction;
564 	} else if (auprobe->defparam.fixups & UPROBE_FIX_CALL) {
565 		regs->sp += sizeof_long(); /* Pop incorrect return address */
566 		if (push_ret_address(regs, utask->vaddr + auprobe->defparam.ilen))
567 			return -ERESTART;
568 	}
569 	/* popf; tell the caller to not touch TF */
570 	if (auprobe->defparam.fixups & UPROBE_FIX_SETF)
571 		utask->autask.saved_tf = true;
572 
573 	return 0;
574 }
575 
576 static void default_abort_op(struct arch_uprobe *auprobe, struct pt_regs *regs)
577 {
578 	riprel_post_xol(auprobe, regs);
579 }
580 
581 static const struct uprobe_xol_ops default_xol_ops = {
582 	.pre_xol  = default_pre_xol_op,
583 	.post_xol = default_post_xol_op,
584 	.abort	  = default_abort_op,
585 };
586 
587 static bool branch_is_call(struct arch_uprobe *auprobe)
588 {
589 	return auprobe->branch.opc1 == 0xe8;
590 }
591 
592 #define CASE_COND					\
593 	COND(70, 71, XF(OF))				\
594 	COND(72, 73, XF(CF))				\
595 	COND(74, 75, XF(ZF))				\
596 	COND(78, 79, XF(SF))				\
597 	COND(7a, 7b, XF(PF))				\
598 	COND(76, 77, XF(CF) || XF(ZF))			\
599 	COND(7c, 7d, XF(SF) != XF(OF))			\
600 	COND(7e, 7f, XF(ZF) || XF(SF) != XF(OF))
601 
602 #define COND(op_y, op_n, expr)				\
603 	case 0x ## op_y: DO((expr) != 0)		\
604 	case 0x ## op_n: DO((expr) == 0)
605 
606 #define XF(xf)	(!!(flags & X86_EFLAGS_ ## xf))
607 
608 static bool is_cond_jmp_opcode(u8 opcode)
609 {
610 	switch (opcode) {
611 	#define DO(expr)	\
612 		return true;
613 	CASE_COND
614 	#undef	DO
615 
616 	default:
617 		return false;
618 	}
619 }
620 
621 static bool check_jmp_cond(struct arch_uprobe *auprobe, struct pt_regs *regs)
622 {
623 	unsigned long flags = regs->flags;
624 
625 	switch (auprobe->branch.opc1) {
626 	#define DO(expr)	\
627 		return expr;
628 	CASE_COND
629 	#undef	DO
630 
631 	default:	/* not a conditional jmp */
632 		return true;
633 	}
634 }
635 
636 #undef	XF
637 #undef	COND
638 #undef	CASE_COND
639 
640 static bool branch_emulate_op(struct arch_uprobe *auprobe, struct pt_regs *regs)
641 {
642 	unsigned long new_ip = regs->ip += auprobe->branch.ilen;
643 	unsigned long offs = (long)auprobe->branch.offs;
644 
645 	if (branch_is_call(auprobe)) {
646 		/*
647 		 * If it fails we execute this (mangled, see the comment in
648 		 * branch_clear_offset) insn out-of-line. In the likely case
649 		 * this should trigger the trap, and the probed application
650 		 * should die or restart the same insn after it handles the
651 		 * signal, arch_uprobe_post_xol() won't be even called.
652 		 *
653 		 * But there is corner case, see the comment in ->post_xol().
654 		 */
655 		if (push_ret_address(regs, new_ip))
656 			return false;
657 	} else if (!check_jmp_cond(auprobe, regs)) {
658 		offs = 0;
659 	}
660 
661 	regs->ip = new_ip + offs;
662 	return true;
663 }
664 
665 static int branch_post_xol_op(struct arch_uprobe *auprobe, struct pt_regs *regs)
666 {
667 	BUG_ON(!branch_is_call(auprobe));
668 	/*
669 	 * We can only get here if branch_emulate_op() failed to push the ret
670 	 * address _and_ another thread expanded our stack before the (mangled)
671 	 * "call" insn was executed out-of-line. Just restore ->sp and restart.
672 	 * We could also restore ->ip and try to call branch_emulate_op() again.
673 	 */
674 	regs->sp += sizeof_long();
675 	return -ERESTART;
676 }
677 
678 static void branch_clear_offset(struct arch_uprobe *auprobe, struct insn *insn)
679 {
680 	/*
681 	 * Turn this insn into "call 1f; 1:", this is what we will execute
682 	 * out-of-line if ->emulate() fails. We only need this to generate
683 	 * a trap, so that the probed task receives the correct signal with
684 	 * the properly filled siginfo.
685 	 *
686 	 * But see the comment in ->post_xol(), in the unlikely case it can
687 	 * succeed. So we need to ensure that the new ->ip can not fall into
688 	 * the non-canonical area and trigger #GP.
689 	 *
690 	 * We could turn it into (say) "pushf", but then we would need to
691 	 * divorce ->insn[] and ->ixol[]. We need to preserve the 1st byte
692 	 * of ->insn[] for set_orig_insn().
693 	 */
694 	memset(auprobe->insn + insn_offset_immediate(insn),
695 		0, insn->immediate.nbytes);
696 }
697 
698 static const struct uprobe_xol_ops branch_xol_ops = {
699 	.emulate  = branch_emulate_op,
700 	.post_xol = branch_post_xol_op,
701 };
702 
703 /* Returns -ENOSYS if branch_xol_ops doesn't handle this insn */
704 static int branch_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn)
705 {
706 	u8 opc1 = OPCODE1(insn);
707 	int i;
708 
709 	switch (opc1) {
710 	case 0xeb:	/* jmp 8 */
711 	case 0xe9:	/* jmp 32 */
712 	case 0x90:	/* prefix* + nop; same as jmp with .offs = 0 */
713 		break;
714 
715 	case 0xe8:	/* call relative */
716 		branch_clear_offset(auprobe, insn);
717 		break;
718 
719 	case 0x0f:
720 		if (insn->opcode.nbytes != 2)
721 			return -ENOSYS;
722 		/*
723 		 * If it is a "near" conditional jmp, OPCODE2() - 0x10 matches
724 		 * OPCODE1() of the "short" jmp which checks the same condition.
725 		 */
726 		opc1 = OPCODE2(insn) - 0x10;
727 	default:
728 		if (!is_cond_jmp_opcode(opc1))
729 			return -ENOSYS;
730 	}
731 
732 	/*
733 	 * 16-bit overrides such as CALLW (66 e8 nn nn) are not supported.
734 	 * Intel and AMD behavior differ in 64-bit mode: Intel ignores 66 prefix.
735 	 * No one uses these insns, reject any branch insns with such prefix.
736 	 */
737 	for (i = 0; i < insn->prefixes.nbytes; i++) {
738 		if (insn->prefixes.bytes[i] == 0x66)
739 			return -ENOTSUPP;
740 	}
741 
742 	auprobe->branch.opc1 = opc1;
743 	auprobe->branch.ilen = insn->length;
744 	auprobe->branch.offs = insn->immediate.value;
745 
746 	auprobe->ops = &branch_xol_ops;
747 	return 0;
748 }
749 
750 /**
751  * arch_uprobe_analyze_insn - instruction analysis including validity and fixups.
752  * @mm: the probed address space.
753  * @arch_uprobe: the probepoint information.
754  * @addr: virtual address at which to install the probepoint
755  * Return 0 on success or a -ve number on error.
756  */
757 int arch_uprobe_analyze_insn(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long addr)
758 {
759 	struct insn insn;
760 	u8 fix_ip_or_call = UPROBE_FIX_IP;
761 	int ret;
762 
763 	ret = uprobe_init_insn(auprobe, &insn, is_64bit_mm(mm));
764 	if (ret)
765 		return ret;
766 
767 	ret = branch_setup_xol_ops(auprobe, &insn);
768 	if (ret != -ENOSYS)
769 		return ret;
770 
771 	/*
772 	 * Figure out which fixups default_post_xol_op() will need to perform,
773 	 * and annotate defparam->fixups accordingly.
774 	 */
775 	switch (OPCODE1(&insn)) {
776 	case 0x9d:		/* popf */
777 		auprobe->defparam.fixups |= UPROBE_FIX_SETF;
778 		break;
779 	case 0xc3:		/* ret or lret -- ip is correct */
780 	case 0xcb:
781 	case 0xc2:
782 	case 0xca:
783 	case 0xea:		/* jmp absolute -- ip is correct */
784 		fix_ip_or_call = 0;
785 		break;
786 	case 0x9a:		/* call absolute - Fix return addr, not ip */
787 		fix_ip_or_call = UPROBE_FIX_CALL;
788 		break;
789 	case 0xff:
790 		switch (MODRM_REG(&insn)) {
791 		case 2: case 3:			/* call or lcall, indirect */
792 			fix_ip_or_call = UPROBE_FIX_CALL;
793 			break;
794 		case 4: case 5:			/* jmp or ljmp, indirect */
795 			fix_ip_or_call = 0;
796 			break;
797 		}
798 		/* fall through */
799 	default:
800 		riprel_analyze(auprobe, &insn);
801 	}
802 
803 	auprobe->defparam.ilen = insn.length;
804 	auprobe->defparam.fixups |= fix_ip_or_call;
805 
806 	auprobe->ops = &default_xol_ops;
807 	return 0;
808 }
809 
810 /*
811  * arch_uprobe_pre_xol - prepare to execute out of line.
812  * @auprobe: the probepoint information.
813  * @regs: reflects the saved user state of current task.
814  */
815 int arch_uprobe_pre_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
816 {
817 	struct uprobe_task *utask = current->utask;
818 
819 	if (auprobe->ops->pre_xol) {
820 		int err = auprobe->ops->pre_xol(auprobe, regs);
821 		if (err)
822 			return err;
823 	}
824 
825 	regs->ip = utask->xol_vaddr;
826 	utask->autask.saved_trap_nr = current->thread.trap_nr;
827 	current->thread.trap_nr = UPROBE_TRAP_NR;
828 
829 	utask->autask.saved_tf = !!(regs->flags & X86_EFLAGS_TF);
830 	regs->flags |= X86_EFLAGS_TF;
831 	if (test_tsk_thread_flag(current, TIF_BLOCKSTEP))
832 		set_task_blockstep(current, false);
833 
834 	return 0;
835 }
836 
837 /*
838  * If xol insn itself traps and generates a signal(Say,
839  * SIGILL/SIGSEGV/etc), then detect the case where a singlestepped
840  * instruction jumps back to its own address. It is assumed that anything
841  * like do_page_fault/do_trap/etc sets thread.trap_nr != -1.
842  *
843  * arch_uprobe_pre_xol/arch_uprobe_post_xol save/restore thread.trap_nr,
844  * arch_uprobe_xol_was_trapped() simply checks that ->trap_nr is not equal to
845  * UPROBE_TRAP_NR == -1 set by arch_uprobe_pre_xol().
846  */
847 bool arch_uprobe_xol_was_trapped(struct task_struct *t)
848 {
849 	if (t->thread.trap_nr != UPROBE_TRAP_NR)
850 		return true;
851 
852 	return false;
853 }
854 
855 /*
856  * Called after single-stepping. To avoid the SMP problems that can
857  * occur when we temporarily put back the original opcode to
858  * single-step, we single-stepped a copy of the instruction.
859  *
860  * This function prepares to resume execution after the single-step.
861  */
862 int arch_uprobe_post_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
863 {
864 	struct uprobe_task *utask = current->utask;
865 	bool send_sigtrap = utask->autask.saved_tf;
866 	int err = 0;
867 
868 	WARN_ON_ONCE(current->thread.trap_nr != UPROBE_TRAP_NR);
869 	current->thread.trap_nr = utask->autask.saved_trap_nr;
870 
871 	if (auprobe->ops->post_xol) {
872 		err = auprobe->ops->post_xol(auprobe, regs);
873 		if (err) {
874 			/*
875 			 * Restore ->ip for restart or post mortem analysis.
876 			 * ->post_xol() must not return -ERESTART unless this
877 			 * is really possible.
878 			 */
879 			regs->ip = utask->vaddr;
880 			if (err == -ERESTART)
881 				err = 0;
882 			send_sigtrap = false;
883 		}
884 	}
885 	/*
886 	 * arch_uprobe_pre_xol() doesn't save the state of TIF_BLOCKSTEP
887 	 * so we can get an extra SIGTRAP if we do not clear TF. We need
888 	 * to examine the opcode to make it right.
889 	 */
890 	if (send_sigtrap)
891 		send_sig(SIGTRAP, current, 0);
892 
893 	if (!utask->autask.saved_tf)
894 		regs->flags &= ~X86_EFLAGS_TF;
895 
896 	return err;
897 }
898 
899 /* callback routine for handling exceptions. */
900 int arch_uprobe_exception_notify(struct notifier_block *self, unsigned long val, void *data)
901 {
902 	struct die_args *args = data;
903 	struct pt_regs *regs = args->regs;
904 	int ret = NOTIFY_DONE;
905 
906 	/* We are only interested in userspace traps */
907 	if (regs && !user_mode(regs))
908 		return NOTIFY_DONE;
909 
910 	switch (val) {
911 	case DIE_INT3:
912 		if (uprobe_pre_sstep_notifier(regs))
913 			ret = NOTIFY_STOP;
914 
915 		break;
916 
917 	case DIE_DEBUG:
918 		if (uprobe_post_sstep_notifier(regs))
919 			ret = NOTIFY_STOP;
920 
921 	default:
922 		break;
923 	}
924 
925 	return ret;
926 }
927 
928 /*
929  * This function gets called when XOL instruction either gets trapped or
930  * the thread has a fatal signal. Reset the instruction pointer to its
931  * probed address for the potential restart or for post mortem analysis.
932  */
933 void arch_uprobe_abort_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
934 {
935 	struct uprobe_task *utask = current->utask;
936 
937 	if (auprobe->ops->abort)
938 		auprobe->ops->abort(auprobe, regs);
939 
940 	current->thread.trap_nr = utask->autask.saved_trap_nr;
941 	regs->ip = utask->vaddr;
942 	/* clear TF if it was set by us in arch_uprobe_pre_xol() */
943 	if (!utask->autask.saved_tf)
944 		regs->flags &= ~X86_EFLAGS_TF;
945 }
946 
947 static bool __skip_sstep(struct arch_uprobe *auprobe, struct pt_regs *regs)
948 {
949 	if (auprobe->ops->emulate)
950 		return auprobe->ops->emulate(auprobe, regs);
951 	return false;
952 }
953 
954 bool arch_uprobe_skip_sstep(struct arch_uprobe *auprobe, struct pt_regs *regs)
955 {
956 	bool ret = __skip_sstep(auprobe, regs);
957 	if (ret && (regs->flags & X86_EFLAGS_TF))
958 		send_sig(SIGTRAP, current, 0);
959 	return ret;
960 }
961 
962 unsigned long
963 arch_uretprobe_hijack_return_addr(unsigned long trampoline_vaddr, struct pt_regs *regs)
964 {
965 	int rasize = sizeof_long(), nleft;
966 	unsigned long orig_ret_vaddr = 0; /* clear high bits for 32-bit apps */
967 
968 	if (copy_from_user(&orig_ret_vaddr, (void __user *)regs->sp, rasize))
969 		return -1;
970 
971 	/* check whether address has been already hijacked */
972 	if (orig_ret_vaddr == trampoline_vaddr)
973 		return orig_ret_vaddr;
974 
975 	nleft = copy_to_user((void __user *)regs->sp, &trampoline_vaddr, rasize);
976 	if (likely(!nleft))
977 		return orig_ret_vaddr;
978 
979 	if (nleft != rasize) {
980 		pr_err("uprobe: return address clobbered: pid=%d, %%sp=%#lx, "
981 			"%%ip=%#lx\n", current->pid, regs->sp, regs->ip);
982 
983 		force_sig_info(SIGSEGV, SEND_SIG_FORCED, current);
984 	}
985 
986 	return -1;
987 }
988 
989 bool arch_uretprobe_is_alive(struct return_instance *ret, enum rp_check ctx,
990 				struct pt_regs *regs)
991 {
992 	if (ctx == RP_CHECK_CALL) /* sp was just decremented by "call" insn */
993 		return regs->sp < ret->stack;
994 	else
995 		return regs->sp <= ret->stack;
996 }
997