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