xref: /linux/arch/x86/kernel/uprobes.c (revision 6e7fd890f1d6ac83805409e9c346240de2705584)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * User-space Probes (UProbes) for x86
4  *
5  * Copyright (C) IBM Corporation, 2008-2011
6  * Authors:
7  *	Srikar Dronamraju
8  *	Jim Keniston
9  */
10 #include <linux/kernel.h>
11 #include <linux/sched.h>
12 #include <linux/ptrace.h>
13 #include <linux/uprobes.h>
14 #include <linux/uaccess.h>
15 #include <linux/syscalls.h>
16 
17 #include <linux/kdebug.h>
18 #include <asm/processor.h>
19 #include <asm/insn.h>
20 #include <asm/mmu_context.h>
21 
22 /* Post-execution fixups. */
23 
24 /* Adjust IP back to vicinity of actual insn */
25 #define UPROBE_FIX_IP		0x01
26 
27 /* Adjust the return address of a call insn */
28 #define UPROBE_FIX_CALL		0x02
29 
30 /* Instruction will modify TF, don't change it */
31 #define UPROBE_FIX_SETF		0x04
32 
33 #define UPROBE_FIX_RIP_SI	0x08
34 #define UPROBE_FIX_RIP_DI	0x10
35 #define UPROBE_FIX_RIP_BX	0x20
36 #define UPROBE_FIX_RIP_MASK	\
37 	(UPROBE_FIX_RIP_SI | UPROBE_FIX_RIP_DI | UPROBE_FIX_RIP_BX)
38 
39 #define	UPROBE_TRAP_NR		UINT_MAX
40 
41 /* Adaptations for mhiramat x86 decoder v14. */
42 #define OPCODE1(insn)		((insn)->opcode.bytes[0])
43 #define OPCODE2(insn)		((insn)->opcode.bytes[1])
44 #define OPCODE3(insn)		((insn)->opcode.bytes[2])
45 #define MODRM_REG(insn)		X86_MODRM_REG((insn)->modrm.value)
46 
47 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
48 	(((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
49 	  (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
50 	  (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
51 	  (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
52 	 << (row % 32))
53 
54 /*
55  * Good-instruction tables for 32-bit apps.  This is non-const and volatile
56  * to keep gcc from statically optimizing it out, as variable_test_bit makes
57  * some versions of gcc to think only *(unsigned long*) is used.
58  *
59  * Opcodes we'll probably never support:
60  * 6c-6f - ins,outs. SEGVs if used in userspace
61  * e4-e7 - in,out imm. SEGVs if used in userspace
62  * ec-ef - in,out acc. SEGVs if used in userspace
63  * cc - int3. SIGTRAP if used in userspace
64  * ce - into. Not used in userspace - no kernel support to make it useful. SEGVs
65  *	(why we support bound (62) then? it's similar, and similarly unused...)
66  * f1 - int1. SIGTRAP if used in userspace
67  * f4 - hlt. SEGVs if used in userspace
68  * fa - cli. SEGVs if used in userspace
69  * fb - sti. SEGVs if used in userspace
70  *
71  * Opcodes which need some work to be supported:
72  * 07,17,1f - pop es/ss/ds
73  *	Normally not used in userspace, but would execute if used.
74  *	Can cause GP or stack exception if tries to load wrong segment descriptor.
75  *	We hesitate to run them under single step since kernel's handling
76  *	of userspace single-stepping (TF flag) is fragile.
77  *	We can easily refuse to support push es/cs/ss/ds (06/0e/16/1e)
78  *	on the same grounds that they are never used.
79  * cd - int N.
80  *	Used by userspace for "int 80" syscall entry. (Other "int N"
81  *	cause GP -> SEGV since their IDT gates don't allow calls from CPL 3).
82  *	Not supported since kernel's handling of userspace single-stepping
83  *	(TF flag) is fragile.
84  * cf - iret. Normally not used in userspace. Doesn't SEGV unless arguments are bad
85  */
86 #if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION)
87 static volatile u32 good_insns_32[256 / 32] = {
88 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
89 	/*      ----------------------------------------------         */
90 	W(0x00, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* 00 */
91 	W(0x10, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0) , /* 10 */
92 	W(0x20, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 20 */
93 	W(0x30, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 30 */
94 	W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
95 	W(0x50, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 50 */
96 	W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* 60 */
97 	W(0x70, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 70 */
98 	W(0x80, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 80 */
99 	W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
100 	W(0xa0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* a0 */
101 	W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* b0 */
102 	W(0xc0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* c0 */
103 	W(0xd0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* d0 */
104 	W(0xe0, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0) | /* e0 */
105 	W(0xf0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1)   /* f0 */
106 	/*      ----------------------------------------------         */
107 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
108 };
109 #else
110 #define good_insns_32	NULL
111 #endif
112 
113 /* Good-instruction tables for 64-bit apps.
114  *
115  * Genuinely invalid opcodes:
116  * 06,07 - formerly push/pop es
117  * 0e - formerly push cs
118  * 16,17 - formerly push/pop ss
119  * 1e,1f - formerly push/pop ds
120  * 27,2f,37,3f - formerly daa/das/aaa/aas
121  * 60,61 - formerly pusha/popa
122  * 62 - formerly bound. EVEX prefix for AVX512 (not yet supported)
123  * 82 - formerly redundant encoding of Group1
124  * 9a - formerly call seg:ofs
125  * ce - formerly into
126  * d4,d5 - formerly aam/aad
127  * d6 - formerly undocumented salc
128  * ea - formerly jmp seg:ofs
129  *
130  * Opcodes we'll probably never support:
131  * 6c-6f - ins,outs. SEGVs if used in userspace
132  * e4-e7 - in,out imm. SEGVs if used in userspace
133  * ec-ef - in,out acc. SEGVs if used in userspace
134  * cc - int3. SIGTRAP if used in userspace
135  * f1 - int1. SIGTRAP if used in userspace
136  * f4 - hlt. SEGVs if used in userspace
137  * fa - cli. SEGVs if used in userspace
138  * fb - sti. SEGVs if used in userspace
139  *
140  * Opcodes which need some work to be supported:
141  * cd - int N.
142  *	Used by userspace for "int 80" syscall entry. (Other "int N"
143  *	cause GP -> SEGV since their IDT gates don't allow calls from CPL 3).
144  *	Not supported since kernel's handling of userspace single-stepping
145  *	(TF flag) is fragile.
146  * cf - iret. Normally not used in userspace. Doesn't SEGV unless arguments are bad
147  */
148 #if defined(CONFIG_X86_64)
149 static volatile u32 good_insns_64[256 / 32] = {
150 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
151 	/*      ----------------------------------------------         */
152 	W(0x00, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 1) | /* 00 */
153 	W(0x10, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0) , /* 10 */
154 	W(0x20, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0) | /* 20 */
155 	W(0x30, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0) , /* 30 */
156 	W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
157 	W(0x50, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 50 */
158 	W(0x60, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* 60 */
159 	W(0x70, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 70 */
160 	W(0x80, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 80 */
161 	W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1) , /* 90 */
162 	W(0xa0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* a0 */
163 	W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* b0 */
164 	W(0xc0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* c0 */
165 	W(0xd0, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* d0 */
166 	W(0xe0, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0) | /* e0 */
167 	W(0xf0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1)   /* f0 */
168 	/*      ----------------------------------------------         */
169 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
170 };
171 #else
172 #define good_insns_64	NULL
173 #endif
174 
175 /* Using this for both 64-bit and 32-bit apps.
176  * Opcodes we don't support:
177  * 0f 00 - SLDT/STR/LLDT/LTR/VERR/VERW/-/- group. System insns
178  * 0f 01 - SGDT/SIDT/LGDT/LIDT/SMSW/-/LMSW/INVLPG group.
179  *	Also encodes tons of other system insns if mod=11.
180  *	Some are in fact non-system: xend, xtest, rdtscp, maybe more
181  * 0f 05 - syscall
182  * 0f 06 - clts (CPL0 insn)
183  * 0f 07 - sysret
184  * 0f 08 - invd (CPL0 insn)
185  * 0f 09 - wbinvd (CPL0 insn)
186  * 0f 0b - ud2
187  * 0f 30 - wrmsr (CPL0 insn) (then why rdmsr is allowed, it's also CPL0 insn?)
188  * 0f 34 - sysenter
189  * 0f 35 - sysexit
190  * 0f 37 - getsec
191  * 0f 78 - vmread (Intel VMX. CPL0 insn)
192  * 0f 79 - vmwrite (Intel VMX. CPL0 insn)
193  *	Note: with prefixes, these two opcodes are
194  *	extrq/insertq/AVX512 convert vector ops.
195  * 0f ae - group15: [f]xsave,[f]xrstor,[v]{ld,st}mxcsr,clflush[opt],
196  *	{rd,wr}{fs,gs}base,{s,l,m}fence.
197  *	Why? They are all user-executable.
198  */
199 static volatile u32 good_2byte_insns[256 / 32] = {
200 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
201 	/*      ----------------------------------------------         */
202 	W(0x00, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1) | /* 00 */
203 	W(0x10, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 10 */
204 	W(0x20, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 20 */
205 	W(0x30, 0, 1, 1, 1, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1) , /* 30 */
206 	W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
207 	W(0x50, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 50 */
208 	W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 60 */
209 	W(0x70, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1) , /* 70 */
210 	W(0x80, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 80 */
211 	W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
212 	W(0xa0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1) | /* a0 */
213 	W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* b0 */
214 	W(0xc0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
215 	W(0xd0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* d0 */
216 	W(0xe0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* e0 */
217 	W(0xf0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1)   /* f0 */
218 	/*      ----------------------------------------------         */
219 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
220 };
221 #undef W
222 
223 /*
224  * opcodes we may need to refine support for:
225  *
226  *  0f - 2-byte instructions: For many of these instructions, the validity
227  *  depends on the prefix and/or the reg field.  On such instructions, we
228  *  just consider the opcode combination valid if it corresponds to any
229  *  valid instruction.
230  *
231  *  8f - Group 1 - only reg = 0 is OK
232  *  c6-c7 - Group 11 - only reg = 0 is OK
233  *  d9-df - fpu insns with some illegal encodings
234  *  f2, f3 - repnz, repz prefixes.  These are also the first byte for
235  *  certain floating-point instructions, such as addsd.
236  *
237  *  fe - Group 4 - only reg = 0 or 1 is OK
238  *  ff - Group 5 - only reg = 0-6 is OK
239  *
240  * others -- Do we need to support these?
241  *
242  *  0f - (floating-point?) prefetch instructions
243  *  07, 17, 1f - pop es, pop ss, pop ds
244  *  26, 2e, 36, 3e - es:, cs:, ss:, ds: segment prefixes --
245  *	but 64 and 65 (fs: and gs:) seem to be used, so we support them
246  *  67 - addr16 prefix
247  *  ce - into
248  *  f0 - lock prefix
249  */
250 
251 /*
252  * TODO:
253  * - Where necessary, examine the modrm byte and allow only valid instructions
254  * in the different Groups and fpu instructions.
255  */
256 
257 static bool is_prefix_bad(struct insn *insn)
258 {
259 	insn_byte_t p;
260 	int i;
261 
262 	for_each_insn_prefix(insn, i, p) {
263 		insn_attr_t attr;
264 
265 		attr = inat_get_opcode_attribute(p);
266 		switch (attr) {
267 		case INAT_MAKE_PREFIX(INAT_PFX_ES):
268 		case INAT_MAKE_PREFIX(INAT_PFX_CS):
269 		case INAT_MAKE_PREFIX(INAT_PFX_DS):
270 		case INAT_MAKE_PREFIX(INAT_PFX_SS):
271 		case INAT_MAKE_PREFIX(INAT_PFX_LOCK):
272 			return true;
273 		}
274 	}
275 	return false;
276 }
277 
278 static int uprobe_init_insn(struct arch_uprobe *auprobe, struct insn *insn, bool x86_64)
279 {
280 	enum insn_mode m = x86_64 ? INSN_MODE_64 : INSN_MODE_32;
281 	u32 volatile *good_insns;
282 	int ret;
283 
284 	ret = insn_decode(insn, auprobe->insn, sizeof(auprobe->insn), m);
285 	if (ret < 0)
286 		return -ENOEXEC;
287 
288 	if (is_prefix_bad(insn))
289 		return -ENOTSUPP;
290 
291 	/* We should not singlestep on the exception masking instructions */
292 	if (insn_masking_exception(insn))
293 		return -ENOTSUPP;
294 
295 	if (x86_64)
296 		good_insns = good_insns_64;
297 	else
298 		good_insns = good_insns_32;
299 
300 	if (test_bit(OPCODE1(insn), (unsigned long *)good_insns))
301 		return 0;
302 
303 	if (insn->opcode.nbytes == 2) {
304 		if (test_bit(OPCODE2(insn), (unsigned long *)good_2byte_insns))
305 			return 0;
306 	}
307 
308 	return -ENOTSUPP;
309 }
310 
311 #ifdef CONFIG_X86_64
312 
313 asm (
314 	".pushsection .rodata\n"
315 	".global uretprobe_trampoline_entry\n"
316 	"uretprobe_trampoline_entry:\n"
317 	"pushq %rax\n"
318 	"pushq %rcx\n"
319 	"pushq %r11\n"
320 	"movq $" __stringify(__NR_uretprobe) ", %rax\n"
321 	"syscall\n"
322 	".global uretprobe_syscall_check\n"
323 	"uretprobe_syscall_check:\n"
324 	"popq %r11\n"
325 	"popq %rcx\n"
326 
327 	/* The uretprobe syscall replaces stored %rax value with final
328 	 * return address, so we don't restore %rax in here and just
329 	 * call ret.
330 	 */
331 	"retq\n"
332 	".global uretprobe_trampoline_end\n"
333 	"uretprobe_trampoline_end:\n"
334 	".popsection\n"
335 );
336 
337 extern u8 uretprobe_trampoline_entry[];
338 extern u8 uretprobe_trampoline_end[];
339 extern u8 uretprobe_syscall_check[];
340 
341 void *arch_uprobe_trampoline(unsigned long *psize)
342 {
343 	static uprobe_opcode_t insn = UPROBE_SWBP_INSN;
344 	struct pt_regs *regs = task_pt_regs(current);
345 
346 	/*
347 	 * At the moment the uretprobe syscall trampoline is supported
348 	 * only for native 64-bit process, the compat process still uses
349 	 * standard breakpoint.
350 	 */
351 	if (user_64bit_mode(regs)) {
352 		*psize = uretprobe_trampoline_end - uretprobe_trampoline_entry;
353 		return uretprobe_trampoline_entry;
354 	}
355 
356 	*psize = UPROBE_SWBP_INSN_SIZE;
357 	return &insn;
358 }
359 
360 static unsigned long trampoline_check_ip(void)
361 {
362 	unsigned long tramp = uprobe_get_trampoline_vaddr();
363 
364 	return tramp + (uretprobe_syscall_check - uretprobe_trampoline_entry);
365 }
366 
367 SYSCALL_DEFINE0(uretprobe)
368 {
369 	struct pt_regs *regs = task_pt_regs(current);
370 	unsigned long err, ip, sp, r11_cx_ax[3];
371 
372 	if (regs->ip != trampoline_check_ip())
373 		goto sigill;
374 
375 	err = copy_from_user(r11_cx_ax, (void __user *)regs->sp, sizeof(r11_cx_ax));
376 	if (err)
377 		goto sigill;
378 
379 	/* expose the "right" values of r11/cx/ax/sp to uprobe_consumer/s */
380 	regs->r11 = r11_cx_ax[0];
381 	regs->cx  = r11_cx_ax[1];
382 	regs->ax  = r11_cx_ax[2];
383 	regs->sp += sizeof(r11_cx_ax);
384 	regs->orig_ax = -1;
385 
386 	ip = regs->ip;
387 	sp = regs->sp;
388 
389 	uprobe_handle_trampoline(regs);
390 
391 	/*
392 	 * Some of the uprobe consumers has changed sp, we can do nothing,
393 	 * just return via iret.
394 	 * .. or shadow stack is enabled, in which case we need to skip
395 	 * return through the user space stack address.
396 	 */
397 	if (regs->sp != sp || shstk_is_enabled())
398 		return regs->ax;
399 	regs->sp -= sizeof(r11_cx_ax);
400 
401 	/* for the case uprobe_consumer has changed r11/cx */
402 	r11_cx_ax[0] = regs->r11;
403 	r11_cx_ax[1] = regs->cx;
404 
405 	/*
406 	 * ax register is passed through as return value, so we can use
407 	 * its space on stack for ip value and jump to it through the
408 	 * trampoline's ret instruction
409 	 */
410 	r11_cx_ax[2] = regs->ip;
411 	regs->ip = ip;
412 
413 	err = copy_to_user((void __user *)regs->sp, r11_cx_ax, sizeof(r11_cx_ax));
414 	if (err)
415 		goto sigill;
416 
417 	/* ensure sysret, see do_syscall_64() */
418 	regs->r11 = regs->flags;
419 	regs->cx  = regs->ip;
420 
421 	return regs->ax;
422 
423 sigill:
424 	force_sig(SIGILL);
425 	return -1;
426 }
427 
428 /*
429  * If arch_uprobe->insn doesn't use rip-relative addressing, return
430  * immediately.  Otherwise, rewrite the instruction so that it accesses
431  * its memory operand indirectly through a scratch register.  Set
432  * defparam->fixups accordingly. (The contents of the scratch register
433  * will be saved before we single-step the modified instruction,
434  * and restored afterward).
435  *
436  * We do this because a rip-relative instruction can access only a
437  * relatively small area (+/- 2 GB from the instruction), and the XOL
438  * area typically lies beyond that area.  At least for instructions
439  * that store to memory, we can't execute the original instruction
440  * and "fix things up" later, because the misdirected store could be
441  * disastrous.
442  *
443  * Some useful facts about rip-relative instructions:
444  *
445  *  - There's always a modrm byte with bit layout "00 reg 101".
446  *  - There's never a SIB byte.
447  *  - The displacement is always 4 bytes.
448  *  - REX.B=1 bit in REX prefix, which normally extends r/m field,
449  *    has no effect on rip-relative mode. It doesn't make modrm byte
450  *    with r/m=101 refer to register 1101 = R13.
451  */
452 static void riprel_analyze(struct arch_uprobe *auprobe, struct insn *insn)
453 {
454 	u8 *cursor;
455 	u8 reg;
456 	u8 reg2;
457 
458 	if (!insn_rip_relative(insn))
459 		return;
460 
461 	/*
462 	 * insn_rip_relative() would have decoded rex_prefix, vex_prefix, modrm.
463 	 * Clear REX.b bit (extension of MODRM.rm field):
464 	 * we want to encode low numbered reg, not r8+.
465 	 */
466 	if (insn->rex_prefix.nbytes) {
467 		cursor = auprobe->insn + insn_offset_rex_prefix(insn);
468 		/* REX byte has 0100wrxb layout, clearing REX.b bit */
469 		*cursor &= 0xfe;
470 	}
471 	/*
472 	 * Similar treatment for VEX3/EVEX prefix.
473 	 * TODO: add XOP treatment when insn decoder supports them
474 	 */
475 	if (insn->vex_prefix.nbytes >= 3) {
476 		/*
477 		 * vex2:     c5    rvvvvLpp   (has no b bit)
478 		 * vex3/xop: c4/8f rxbmmmmm wvvvvLpp
479 		 * evex:     62    rxbR00mm wvvvv1pp zllBVaaa
480 		 * Setting VEX3.b (setting because it has inverted meaning).
481 		 * Setting EVEX.x since (in non-SIB encoding) EVEX.x
482 		 * is the 4th bit of MODRM.rm, and needs the same treatment.
483 		 * For VEX3-encoded insns, VEX3.x value has no effect in
484 		 * non-SIB encoding, the change is superfluous but harmless.
485 		 */
486 		cursor = auprobe->insn + insn_offset_vex_prefix(insn) + 1;
487 		*cursor |= 0x60;
488 	}
489 
490 	/*
491 	 * Convert from rip-relative addressing to register-relative addressing
492 	 * via a scratch register.
493 	 *
494 	 * This is tricky since there are insns with modrm byte
495 	 * which also use registers not encoded in modrm byte:
496 	 * [i]div/[i]mul: implicitly use dx:ax
497 	 * shift ops: implicitly use cx
498 	 * cmpxchg: implicitly uses ax
499 	 * cmpxchg8/16b: implicitly uses dx:ax and bx:cx
500 	 *   Encoding: 0f c7/1 modrm
501 	 *   The code below thinks that reg=1 (cx), chooses si as scratch.
502 	 * mulx: implicitly uses dx: mulx r/m,r1,r2 does r1:r2 = dx * r/m.
503 	 *   First appeared in Haswell (BMI2 insn). It is vex-encoded.
504 	 *   Example where none of bx,cx,dx can be used as scratch reg:
505 	 *   c4 e2 63 f6 0d disp32   mulx disp32(%rip),%ebx,%ecx
506 	 * [v]pcmpistri: implicitly uses cx, xmm0
507 	 * [v]pcmpistrm: implicitly uses xmm0
508 	 * [v]pcmpestri: implicitly uses ax, dx, cx, xmm0
509 	 * [v]pcmpestrm: implicitly uses ax, dx, xmm0
510 	 *   Evil SSE4.2 string comparison ops from hell.
511 	 * maskmovq/[v]maskmovdqu: implicitly uses (ds:rdi) as destination.
512 	 *   Encoding: 0f f7 modrm, 66 0f f7 modrm, vex-encoded: c5 f9 f7 modrm.
513 	 *   Store op1, byte-masked by op2 msb's in each byte, to (ds:rdi).
514 	 *   AMD says it has no 3-operand form (vex.vvvv must be 1111)
515 	 *   and that it can have only register operands, not mem
516 	 *   (its modrm byte must have mode=11).
517 	 *   If these restrictions will ever be lifted,
518 	 *   we'll need code to prevent selection of di as scratch reg!
519 	 *
520 	 * Summary: I don't know any insns with modrm byte which
521 	 * use SI register implicitly. DI register is used only
522 	 * by one insn (maskmovq) and BX register is used
523 	 * only by one too (cmpxchg8b).
524 	 * BP is stack-segment based (may be a problem?).
525 	 * AX, DX, CX are off-limits (many implicit users).
526 	 * SP is unusable (it's stack pointer - think about "pop mem";
527 	 * also, rsp+disp32 needs sib encoding -> insn length change).
528 	 */
529 
530 	reg = MODRM_REG(insn);	/* Fetch modrm.reg */
531 	reg2 = 0xff;		/* Fetch vex.vvvv */
532 	if (insn->vex_prefix.nbytes)
533 		reg2 = insn->vex_prefix.bytes[2];
534 	/*
535 	 * TODO: add XOP vvvv reading.
536 	 *
537 	 * vex.vvvv field is in bits 6-3, bits are inverted.
538 	 * But in 32-bit mode, high-order bit may be ignored.
539 	 * Therefore, let's consider only 3 low-order bits.
540 	 */
541 	reg2 = ((reg2 >> 3) & 0x7) ^ 0x7;
542 	/*
543 	 * Register numbering is ax,cx,dx,bx, sp,bp,si,di, r8..r15.
544 	 *
545 	 * Choose scratch reg. Order is important: must not select bx
546 	 * if we can use si (cmpxchg8b case!)
547 	 */
548 	if (reg != 6 && reg2 != 6) {
549 		reg2 = 6;
550 		auprobe->defparam.fixups |= UPROBE_FIX_RIP_SI;
551 	} else if (reg != 7 && reg2 != 7) {
552 		reg2 = 7;
553 		auprobe->defparam.fixups |= UPROBE_FIX_RIP_DI;
554 		/* TODO (paranoia): force maskmovq to not use di */
555 	} else {
556 		reg2 = 3;
557 		auprobe->defparam.fixups |= UPROBE_FIX_RIP_BX;
558 	}
559 	/*
560 	 * Point cursor at the modrm byte.  The next 4 bytes are the
561 	 * displacement.  Beyond the displacement, for some instructions,
562 	 * is the immediate operand.
563 	 */
564 	cursor = auprobe->insn + insn_offset_modrm(insn);
565 	/*
566 	 * Change modrm from "00 reg 101" to "10 reg reg2". Example:
567 	 * 89 05 disp32  mov %eax,disp32(%rip) becomes
568 	 * 89 86 disp32  mov %eax,disp32(%rsi)
569 	 */
570 	*cursor = 0x80 | (reg << 3) | reg2;
571 }
572 
573 static inline unsigned long *
574 scratch_reg(struct arch_uprobe *auprobe, struct pt_regs *regs)
575 {
576 	if (auprobe->defparam.fixups & UPROBE_FIX_RIP_SI)
577 		return &regs->si;
578 	if (auprobe->defparam.fixups & UPROBE_FIX_RIP_DI)
579 		return &regs->di;
580 	return &regs->bx;
581 }
582 
583 /*
584  * If we're emulating a rip-relative instruction, save the contents
585  * of the scratch register and store the target address in that register.
586  */
587 static void riprel_pre_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
588 {
589 	if (auprobe->defparam.fixups & UPROBE_FIX_RIP_MASK) {
590 		struct uprobe_task *utask = current->utask;
591 		unsigned long *sr = scratch_reg(auprobe, regs);
592 
593 		utask->autask.saved_scratch_register = *sr;
594 		*sr = utask->vaddr + auprobe->defparam.ilen;
595 	}
596 }
597 
598 static void riprel_post_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
599 {
600 	if (auprobe->defparam.fixups & UPROBE_FIX_RIP_MASK) {
601 		struct uprobe_task *utask = current->utask;
602 		unsigned long *sr = scratch_reg(auprobe, regs);
603 
604 		*sr = utask->autask.saved_scratch_register;
605 	}
606 }
607 #else /* 32-bit: */
608 /*
609  * No RIP-relative addressing on 32-bit
610  */
611 static void riprel_analyze(struct arch_uprobe *auprobe, struct insn *insn)
612 {
613 }
614 static void riprel_pre_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
615 {
616 }
617 static void riprel_post_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
618 {
619 }
620 #endif /* CONFIG_X86_64 */
621 
622 struct uprobe_xol_ops {
623 	bool	(*emulate)(struct arch_uprobe *, struct pt_regs *);
624 	int	(*pre_xol)(struct arch_uprobe *, struct pt_regs *);
625 	int	(*post_xol)(struct arch_uprobe *, struct pt_regs *);
626 	void	(*abort)(struct arch_uprobe *, struct pt_regs *);
627 };
628 
629 static inline int sizeof_long(struct pt_regs *regs)
630 {
631 	/*
632 	 * Check registers for mode as in_xxx_syscall() does not apply here.
633 	 */
634 	return user_64bit_mode(regs) ? 8 : 4;
635 }
636 
637 static int default_pre_xol_op(struct arch_uprobe *auprobe, struct pt_regs *regs)
638 {
639 	riprel_pre_xol(auprobe, regs);
640 	return 0;
641 }
642 
643 static int emulate_push_stack(struct pt_regs *regs, unsigned long val)
644 {
645 	unsigned long new_sp = regs->sp - sizeof_long(regs);
646 
647 	if (copy_to_user((void __user *)new_sp, &val, sizeof_long(regs)))
648 		return -EFAULT;
649 
650 	regs->sp = new_sp;
651 	return 0;
652 }
653 
654 /*
655  * We have to fix things up as follows:
656  *
657  * Typically, the new ip is relative to the copied instruction.  We need
658  * to make it relative to the original instruction (FIX_IP).  Exceptions
659  * are return instructions and absolute or indirect jump or call instructions.
660  *
661  * If the single-stepped instruction was a call, the return address that
662  * is atop the stack is the address following the copied instruction.  We
663  * need to make it the address following the original instruction (FIX_CALL).
664  *
665  * If the original instruction was a rip-relative instruction such as
666  * "movl %edx,0xnnnn(%rip)", we have instead executed an equivalent
667  * instruction using a scratch register -- e.g., "movl %edx,0xnnnn(%rsi)".
668  * We need to restore the contents of the scratch register
669  * (FIX_RIP_reg).
670  */
671 static int default_post_xol_op(struct arch_uprobe *auprobe, struct pt_regs *regs)
672 {
673 	struct uprobe_task *utask = current->utask;
674 
675 	riprel_post_xol(auprobe, regs);
676 	if (auprobe->defparam.fixups & UPROBE_FIX_IP) {
677 		long correction = utask->vaddr - utask->xol_vaddr;
678 		regs->ip += correction;
679 	} else if (auprobe->defparam.fixups & UPROBE_FIX_CALL) {
680 		regs->sp += sizeof_long(regs); /* Pop incorrect return address */
681 		if (emulate_push_stack(regs, utask->vaddr + auprobe->defparam.ilen))
682 			return -ERESTART;
683 	}
684 	/* popf; tell the caller to not touch TF */
685 	if (auprobe->defparam.fixups & UPROBE_FIX_SETF)
686 		utask->autask.saved_tf = true;
687 
688 	return 0;
689 }
690 
691 static void default_abort_op(struct arch_uprobe *auprobe, struct pt_regs *regs)
692 {
693 	riprel_post_xol(auprobe, regs);
694 }
695 
696 static const struct uprobe_xol_ops default_xol_ops = {
697 	.pre_xol  = default_pre_xol_op,
698 	.post_xol = default_post_xol_op,
699 	.abort	  = default_abort_op,
700 };
701 
702 static bool branch_is_call(struct arch_uprobe *auprobe)
703 {
704 	return auprobe->branch.opc1 == 0xe8;
705 }
706 
707 #define CASE_COND					\
708 	COND(70, 71, XF(OF))				\
709 	COND(72, 73, XF(CF))				\
710 	COND(74, 75, XF(ZF))				\
711 	COND(78, 79, XF(SF))				\
712 	COND(7a, 7b, XF(PF))				\
713 	COND(76, 77, XF(CF) || XF(ZF))			\
714 	COND(7c, 7d, XF(SF) != XF(OF))			\
715 	COND(7e, 7f, XF(ZF) || XF(SF) != XF(OF))
716 
717 #define COND(op_y, op_n, expr)				\
718 	case 0x ## op_y: DO((expr) != 0)		\
719 	case 0x ## op_n: DO((expr) == 0)
720 
721 #define XF(xf)	(!!(flags & X86_EFLAGS_ ## xf))
722 
723 static bool is_cond_jmp_opcode(u8 opcode)
724 {
725 	switch (opcode) {
726 	#define DO(expr)	\
727 		return true;
728 	CASE_COND
729 	#undef	DO
730 
731 	default:
732 		return false;
733 	}
734 }
735 
736 static bool check_jmp_cond(struct arch_uprobe *auprobe, struct pt_regs *regs)
737 {
738 	unsigned long flags = regs->flags;
739 
740 	switch (auprobe->branch.opc1) {
741 	#define DO(expr)	\
742 		return expr;
743 	CASE_COND
744 	#undef	DO
745 
746 	default:	/* not a conditional jmp */
747 		return true;
748 	}
749 }
750 
751 #undef	XF
752 #undef	COND
753 #undef	CASE_COND
754 
755 static bool branch_emulate_op(struct arch_uprobe *auprobe, struct pt_regs *regs)
756 {
757 	unsigned long new_ip = regs->ip += auprobe->branch.ilen;
758 	unsigned long offs = (long)auprobe->branch.offs;
759 
760 	if (branch_is_call(auprobe)) {
761 		/*
762 		 * If it fails we execute this (mangled, see the comment in
763 		 * branch_clear_offset) insn out-of-line. In the likely case
764 		 * this should trigger the trap, and the probed application
765 		 * should die or restart the same insn after it handles the
766 		 * signal, arch_uprobe_post_xol() won't be even called.
767 		 *
768 		 * But there is corner case, see the comment in ->post_xol().
769 		 */
770 		if (emulate_push_stack(regs, new_ip))
771 			return false;
772 	} else if (!check_jmp_cond(auprobe, regs)) {
773 		offs = 0;
774 	}
775 
776 	regs->ip = new_ip + offs;
777 	return true;
778 }
779 
780 static bool push_emulate_op(struct arch_uprobe *auprobe, struct pt_regs *regs)
781 {
782 	unsigned long *src_ptr = (void *)regs + auprobe->push.reg_offset;
783 
784 	if (emulate_push_stack(regs, *src_ptr))
785 		return false;
786 	regs->ip += auprobe->push.ilen;
787 	return true;
788 }
789 
790 static int branch_post_xol_op(struct arch_uprobe *auprobe, struct pt_regs *regs)
791 {
792 	BUG_ON(!branch_is_call(auprobe));
793 	/*
794 	 * We can only get here if branch_emulate_op() failed to push the ret
795 	 * address _and_ another thread expanded our stack before the (mangled)
796 	 * "call" insn was executed out-of-line. Just restore ->sp and restart.
797 	 * We could also restore ->ip and try to call branch_emulate_op() again.
798 	 */
799 	regs->sp += sizeof_long(regs);
800 	return -ERESTART;
801 }
802 
803 static void branch_clear_offset(struct arch_uprobe *auprobe, struct insn *insn)
804 {
805 	/*
806 	 * Turn this insn into "call 1f; 1:", this is what we will execute
807 	 * out-of-line if ->emulate() fails. We only need this to generate
808 	 * a trap, so that the probed task receives the correct signal with
809 	 * the properly filled siginfo.
810 	 *
811 	 * But see the comment in ->post_xol(), in the unlikely case it can
812 	 * succeed. So we need to ensure that the new ->ip can not fall into
813 	 * the non-canonical area and trigger #GP.
814 	 *
815 	 * We could turn it into (say) "pushf", but then we would need to
816 	 * divorce ->insn[] and ->ixol[]. We need to preserve the 1st byte
817 	 * of ->insn[] for set_orig_insn().
818 	 */
819 	memset(auprobe->insn + insn_offset_immediate(insn),
820 		0, insn->immediate.nbytes);
821 }
822 
823 static const struct uprobe_xol_ops branch_xol_ops = {
824 	.emulate  = branch_emulate_op,
825 	.post_xol = branch_post_xol_op,
826 };
827 
828 static const struct uprobe_xol_ops push_xol_ops = {
829 	.emulate  = push_emulate_op,
830 };
831 
832 /* Returns -ENOSYS if branch_xol_ops doesn't handle this insn */
833 static int branch_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn)
834 {
835 	u8 opc1 = OPCODE1(insn);
836 	insn_byte_t p;
837 	int i;
838 
839 	switch (opc1) {
840 	case 0xeb:	/* jmp 8 */
841 	case 0xe9:	/* jmp 32 */
842 		break;
843 	case 0x90:	/* prefix* + nop; same as jmp with .offs = 0 */
844 		goto setup;
845 
846 	case 0xe8:	/* call relative */
847 		branch_clear_offset(auprobe, insn);
848 		break;
849 
850 	case 0x0f:
851 		if (insn->opcode.nbytes != 2)
852 			return -ENOSYS;
853 		/*
854 		 * If it is a "near" conditional jmp, OPCODE2() - 0x10 matches
855 		 * OPCODE1() of the "short" jmp which checks the same condition.
856 		 */
857 		opc1 = OPCODE2(insn) - 0x10;
858 		fallthrough;
859 	default:
860 		if (!is_cond_jmp_opcode(opc1))
861 			return -ENOSYS;
862 	}
863 
864 	/*
865 	 * 16-bit overrides such as CALLW (66 e8 nn nn) are not supported.
866 	 * Intel and AMD behavior differ in 64-bit mode: Intel ignores 66 prefix.
867 	 * No one uses these insns, reject any branch insns with such prefix.
868 	 */
869 	for_each_insn_prefix(insn, i, p) {
870 		if (p == 0x66)
871 			return -ENOTSUPP;
872 	}
873 
874 setup:
875 	auprobe->branch.opc1 = opc1;
876 	auprobe->branch.ilen = insn->length;
877 	auprobe->branch.offs = insn->immediate.value;
878 
879 	auprobe->ops = &branch_xol_ops;
880 	return 0;
881 }
882 
883 /* Returns -ENOSYS if push_xol_ops doesn't handle this insn */
884 static int push_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn)
885 {
886 	u8 opc1 = OPCODE1(insn), reg_offset = 0;
887 
888 	if (opc1 < 0x50 || opc1 > 0x57)
889 		return -ENOSYS;
890 
891 	if (insn->length > 2)
892 		return -ENOSYS;
893 	if (insn->length == 2) {
894 		/* only support rex_prefix 0x41 (x64 only) */
895 #ifdef CONFIG_X86_64
896 		if (insn->rex_prefix.nbytes != 1 ||
897 		    insn->rex_prefix.bytes[0] != 0x41)
898 			return -ENOSYS;
899 
900 		switch (opc1) {
901 		case 0x50:
902 			reg_offset = offsetof(struct pt_regs, r8);
903 			break;
904 		case 0x51:
905 			reg_offset = offsetof(struct pt_regs, r9);
906 			break;
907 		case 0x52:
908 			reg_offset = offsetof(struct pt_regs, r10);
909 			break;
910 		case 0x53:
911 			reg_offset = offsetof(struct pt_regs, r11);
912 			break;
913 		case 0x54:
914 			reg_offset = offsetof(struct pt_regs, r12);
915 			break;
916 		case 0x55:
917 			reg_offset = offsetof(struct pt_regs, r13);
918 			break;
919 		case 0x56:
920 			reg_offset = offsetof(struct pt_regs, r14);
921 			break;
922 		case 0x57:
923 			reg_offset = offsetof(struct pt_regs, r15);
924 			break;
925 		}
926 #else
927 		return -ENOSYS;
928 #endif
929 	} else {
930 		switch (opc1) {
931 		case 0x50:
932 			reg_offset = offsetof(struct pt_regs, ax);
933 			break;
934 		case 0x51:
935 			reg_offset = offsetof(struct pt_regs, cx);
936 			break;
937 		case 0x52:
938 			reg_offset = offsetof(struct pt_regs, dx);
939 			break;
940 		case 0x53:
941 			reg_offset = offsetof(struct pt_regs, bx);
942 			break;
943 		case 0x54:
944 			reg_offset = offsetof(struct pt_regs, sp);
945 			break;
946 		case 0x55:
947 			reg_offset = offsetof(struct pt_regs, bp);
948 			break;
949 		case 0x56:
950 			reg_offset = offsetof(struct pt_regs, si);
951 			break;
952 		case 0x57:
953 			reg_offset = offsetof(struct pt_regs, di);
954 			break;
955 		}
956 	}
957 
958 	auprobe->push.reg_offset = reg_offset;
959 	auprobe->push.ilen = insn->length;
960 	auprobe->ops = &push_xol_ops;
961 	return 0;
962 }
963 
964 /**
965  * arch_uprobe_analyze_insn - instruction analysis including validity and fixups.
966  * @auprobe: the probepoint information.
967  * @mm: the probed address space.
968  * @addr: virtual address at which to install the probepoint
969  * Return 0 on success or a -ve number on error.
970  */
971 int arch_uprobe_analyze_insn(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long addr)
972 {
973 	struct insn insn;
974 	u8 fix_ip_or_call = UPROBE_FIX_IP;
975 	int ret;
976 
977 	ret = uprobe_init_insn(auprobe, &insn, is_64bit_mm(mm));
978 	if (ret)
979 		return ret;
980 
981 	ret = branch_setup_xol_ops(auprobe, &insn);
982 	if (ret != -ENOSYS)
983 		return ret;
984 
985 	ret = push_setup_xol_ops(auprobe, &insn);
986 	if (ret != -ENOSYS)
987 		return ret;
988 
989 	/*
990 	 * Figure out which fixups default_post_xol_op() will need to perform,
991 	 * and annotate defparam->fixups accordingly.
992 	 */
993 	switch (OPCODE1(&insn)) {
994 	case 0x9d:		/* popf */
995 		auprobe->defparam.fixups |= UPROBE_FIX_SETF;
996 		break;
997 	case 0xc3:		/* ret or lret -- ip is correct */
998 	case 0xcb:
999 	case 0xc2:
1000 	case 0xca:
1001 	case 0xea:		/* jmp absolute -- ip is correct */
1002 		fix_ip_or_call = 0;
1003 		break;
1004 	case 0x9a:		/* call absolute - Fix return addr, not ip */
1005 		fix_ip_or_call = UPROBE_FIX_CALL;
1006 		break;
1007 	case 0xff:
1008 		switch (MODRM_REG(&insn)) {
1009 		case 2: case 3:			/* call or lcall, indirect */
1010 			fix_ip_or_call = UPROBE_FIX_CALL;
1011 			break;
1012 		case 4: case 5:			/* jmp or ljmp, indirect */
1013 			fix_ip_or_call = 0;
1014 			break;
1015 		}
1016 		fallthrough;
1017 	default:
1018 		riprel_analyze(auprobe, &insn);
1019 	}
1020 
1021 	auprobe->defparam.ilen = insn.length;
1022 	auprobe->defparam.fixups |= fix_ip_or_call;
1023 
1024 	auprobe->ops = &default_xol_ops;
1025 	return 0;
1026 }
1027 
1028 /*
1029  * arch_uprobe_pre_xol - prepare to execute out of line.
1030  * @auprobe: the probepoint information.
1031  * @regs: reflects the saved user state of current task.
1032  */
1033 int arch_uprobe_pre_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
1034 {
1035 	struct uprobe_task *utask = current->utask;
1036 
1037 	if (auprobe->ops->pre_xol) {
1038 		int err = auprobe->ops->pre_xol(auprobe, regs);
1039 		if (err)
1040 			return err;
1041 	}
1042 
1043 	regs->ip = utask->xol_vaddr;
1044 	utask->autask.saved_trap_nr = current->thread.trap_nr;
1045 	current->thread.trap_nr = UPROBE_TRAP_NR;
1046 
1047 	utask->autask.saved_tf = !!(regs->flags & X86_EFLAGS_TF);
1048 	regs->flags |= X86_EFLAGS_TF;
1049 	if (test_tsk_thread_flag(current, TIF_BLOCKSTEP))
1050 		set_task_blockstep(current, false);
1051 
1052 	return 0;
1053 }
1054 
1055 /*
1056  * If xol insn itself traps and generates a signal(Say,
1057  * SIGILL/SIGSEGV/etc), then detect the case where a singlestepped
1058  * instruction jumps back to its own address. It is assumed that anything
1059  * like do_page_fault/do_trap/etc sets thread.trap_nr != -1.
1060  *
1061  * arch_uprobe_pre_xol/arch_uprobe_post_xol save/restore thread.trap_nr,
1062  * arch_uprobe_xol_was_trapped() simply checks that ->trap_nr is not equal to
1063  * UPROBE_TRAP_NR == -1 set by arch_uprobe_pre_xol().
1064  */
1065 bool arch_uprobe_xol_was_trapped(struct task_struct *t)
1066 {
1067 	if (t->thread.trap_nr != UPROBE_TRAP_NR)
1068 		return true;
1069 
1070 	return false;
1071 }
1072 
1073 /*
1074  * Called after single-stepping. To avoid the SMP problems that can
1075  * occur when we temporarily put back the original opcode to
1076  * single-step, we single-stepped a copy of the instruction.
1077  *
1078  * This function prepares to resume execution after the single-step.
1079  */
1080 int arch_uprobe_post_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
1081 {
1082 	struct uprobe_task *utask = current->utask;
1083 	bool send_sigtrap = utask->autask.saved_tf;
1084 	int err = 0;
1085 
1086 	WARN_ON_ONCE(current->thread.trap_nr != UPROBE_TRAP_NR);
1087 	current->thread.trap_nr = utask->autask.saved_trap_nr;
1088 
1089 	if (auprobe->ops->post_xol) {
1090 		err = auprobe->ops->post_xol(auprobe, regs);
1091 		if (err) {
1092 			/*
1093 			 * Restore ->ip for restart or post mortem analysis.
1094 			 * ->post_xol() must not return -ERESTART unless this
1095 			 * is really possible.
1096 			 */
1097 			regs->ip = utask->vaddr;
1098 			if (err == -ERESTART)
1099 				err = 0;
1100 			send_sigtrap = false;
1101 		}
1102 	}
1103 	/*
1104 	 * arch_uprobe_pre_xol() doesn't save the state of TIF_BLOCKSTEP
1105 	 * so we can get an extra SIGTRAP if we do not clear TF. We need
1106 	 * to examine the opcode to make it right.
1107 	 */
1108 	if (send_sigtrap)
1109 		send_sig(SIGTRAP, current, 0);
1110 
1111 	if (!utask->autask.saved_tf)
1112 		regs->flags &= ~X86_EFLAGS_TF;
1113 
1114 	return err;
1115 }
1116 
1117 /* callback routine for handling exceptions. */
1118 int arch_uprobe_exception_notify(struct notifier_block *self, unsigned long val, void *data)
1119 {
1120 	struct die_args *args = data;
1121 	struct pt_regs *regs = args->regs;
1122 	int ret = NOTIFY_DONE;
1123 
1124 	/* We are only interested in userspace traps */
1125 	if (regs && !user_mode(regs))
1126 		return NOTIFY_DONE;
1127 
1128 	switch (val) {
1129 	case DIE_INT3:
1130 		if (uprobe_pre_sstep_notifier(regs))
1131 			ret = NOTIFY_STOP;
1132 
1133 		break;
1134 
1135 	case DIE_DEBUG:
1136 		if (uprobe_post_sstep_notifier(regs))
1137 			ret = NOTIFY_STOP;
1138 
1139 		break;
1140 
1141 	default:
1142 		break;
1143 	}
1144 
1145 	return ret;
1146 }
1147 
1148 /*
1149  * This function gets called when XOL instruction either gets trapped or
1150  * the thread has a fatal signal. Reset the instruction pointer to its
1151  * probed address for the potential restart or for post mortem analysis.
1152  */
1153 void arch_uprobe_abort_xol(struct arch_uprobe *auprobe, struct pt_regs *regs)
1154 {
1155 	struct uprobe_task *utask = current->utask;
1156 
1157 	if (auprobe->ops->abort)
1158 		auprobe->ops->abort(auprobe, regs);
1159 
1160 	current->thread.trap_nr = utask->autask.saved_trap_nr;
1161 	regs->ip = utask->vaddr;
1162 	/* clear TF if it was set by us in arch_uprobe_pre_xol() */
1163 	if (!utask->autask.saved_tf)
1164 		regs->flags &= ~X86_EFLAGS_TF;
1165 }
1166 
1167 static bool __skip_sstep(struct arch_uprobe *auprobe, struct pt_regs *regs)
1168 {
1169 	if (auprobe->ops->emulate)
1170 		return auprobe->ops->emulate(auprobe, regs);
1171 	return false;
1172 }
1173 
1174 bool arch_uprobe_skip_sstep(struct arch_uprobe *auprobe, struct pt_regs *regs)
1175 {
1176 	bool ret = __skip_sstep(auprobe, regs);
1177 	if (ret && (regs->flags & X86_EFLAGS_TF))
1178 		send_sig(SIGTRAP, current, 0);
1179 	return ret;
1180 }
1181 
1182 unsigned long
1183 arch_uretprobe_hijack_return_addr(unsigned long trampoline_vaddr, struct pt_regs *regs)
1184 {
1185 	int rasize = sizeof_long(regs), nleft;
1186 	unsigned long orig_ret_vaddr = 0; /* clear high bits for 32-bit apps */
1187 
1188 	if (copy_from_user(&orig_ret_vaddr, (void __user *)regs->sp, rasize))
1189 		return -1;
1190 
1191 	/* check whether address has been already hijacked */
1192 	if (orig_ret_vaddr == trampoline_vaddr)
1193 		return orig_ret_vaddr;
1194 
1195 	nleft = copy_to_user((void __user *)regs->sp, &trampoline_vaddr, rasize);
1196 	if (likely(!nleft)) {
1197 		if (shstk_update_last_frame(trampoline_vaddr)) {
1198 			force_sig(SIGSEGV);
1199 			return -1;
1200 		}
1201 		return orig_ret_vaddr;
1202 	}
1203 
1204 	if (nleft != rasize) {
1205 		pr_err("return address clobbered: pid=%d, %%sp=%#lx, %%ip=%#lx\n",
1206 		       current->pid, regs->sp, regs->ip);
1207 
1208 		force_sig(SIGSEGV);
1209 	}
1210 
1211 	return -1;
1212 }
1213 
1214 bool arch_uretprobe_is_alive(struct return_instance *ret, enum rp_check ctx,
1215 				struct pt_regs *regs)
1216 {
1217 	if (ctx == RP_CHECK_CALL) /* sp was just decremented by "call" insn */
1218 		return regs->sp < ret->stack;
1219 	else
1220 		return regs->sp <= ret->stack;
1221 }
1222