| 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/insn-eval.h> |
| 21 | #include <asm/mmu_context.h> |
| 22 | #include <asm/nops.h> |
| 23 | |
| 24 | /* Post-execution fixups. */ |
| 25 | |
| 26 | /* Adjust IP back to vicinity of actual insn */ |
| 27 | #define UPROBE_FIX_IP 0x01 |
| 28 | |
| 29 | /* Adjust the return address of a call insn */ |
| 30 | #define UPROBE_FIX_CALL 0x02 |
| 31 | |
| 32 | /* Instruction will modify TF, don't change it */ |
| 33 | #define UPROBE_FIX_SETF 0x04 |
| 34 | |
| 35 | #define UPROBE_FIX_RIP_SI 0x08 |
| 36 | #define UPROBE_FIX_RIP_DI 0x10 |
| 37 | #define UPROBE_FIX_RIP_BX 0x20 |
| 38 | #define UPROBE_FIX_RIP_MASK \ |
| 39 | (UPROBE_FIX_RIP_SI | UPROBE_FIX_RIP_DI | UPROBE_FIX_RIP_BX) |
| 40 | |
| 41 | #define UPROBE_TRAP_NR UINT_MAX |
| 42 | |
| 43 | /* Adaptations for mhiramat x86 decoder v14. */ |
| 44 | #define OPCODE1(insn) ((insn)->opcode.bytes[0]) |
| 45 | #define OPCODE2(insn) ((insn)->opcode.bytes[1]) |
| 46 | #define OPCODE3(insn) ((insn)->opcode.bytes[2]) |
| 47 | #define MODRM_REG(insn) X86_MODRM_REG((insn)->modrm.value) |
| 48 | |
| 49 | #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\ |
| 50 | (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) | \ |
| 51 | (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) | \ |
| 52 | (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) | \ |
| 53 | (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf)) \ |
| 54 | << (row % 32)) |
| 55 | |
| 56 | /* |
| 57 | * Good-instruction tables for 32-bit apps. This is non-const and volatile |
| 58 | * to keep gcc from statically optimizing it out, as variable_test_bit makes |
| 59 | * some versions of gcc to think only *(unsigned long*) is used. |
| 60 | * |
| 61 | * Opcodes we'll probably never support: |
| 62 | * 6c-6f - ins,outs. SEGVs if used in userspace |
| 63 | * e4-e7 - in,out imm. SEGVs if used in userspace |
| 64 | * ec-ef - in,out acc. SEGVs if used in userspace |
| 65 | * cc - int3. SIGTRAP if used in userspace |
| 66 | * ce - into. Not used in userspace - no kernel support to make it useful. SEGVs |
| 67 | * (why we support bound (62) then? it's similar, and similarly unused...) |
| 68 | * f1 - int1. SIGTRAP if used in userspace |
| 69 | * f4 - hlt. SEGVs if used in userspace |
| 70 | * fa - cli. SEGVs if used in userspace |
| 71 | * fb - sti. SEGVs if used in userspace |
| 72 | * |
| 73 | * Opcodes which need some work to be supported: |
| 74 | * 07,17,1f - pop es/ss/ds |
| 75 | * Normally not used in userspace, but would execute if used. |
| 76 | * Can cause GP or stack exception if tries to load wrong segment descriptor. |
| 77 | * We hesitate to run them under single step since kernel's handling |
| 78 | * of userspace single-stepping (TF flag) is fragile. |
| 79 | * We can easily refuse to support push es/cs/ss/ds (06/0e/16/1e) |
| 80 | * on the same grounds that they are never used. |
| 81 | * cd - int N. |
| 82 | * Used by userspace for "int 80" syscall entry. (Other "int N" |
| 83 | * cause GP -> SEGV since their IDT gates don't allow calls from CPL 3). |
| 84 | * Not supported since kernel's handling of userspace single-stepping |
| 85 | * (TF flag) is fragile. |
| 86 | * cf - iret. Normally not used in userspace. Doesn't SEGV unless arguments are bad |
| 87 | */ |
| 88 | #if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION) |
| 89 | static volatile u32 good_insns_32[256 / 32] = { |
| 90 | /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */ |
| 91 | /* ---------------------------------------------- */ |
| 92 | W(0x00, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* 00 */ |
| 93 | W(0x10, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0) , /* 10 */ |
| 94 | W(0x20, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 20 */ |
| 95 | W(0x30, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 30 */ |
| 96 | W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */ |
| 97 | W(0x50, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 50 */ |
| 98 | W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* 60 */ |
| 99 | W(0x70, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 70 */ |
| 100 | W(0x80, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 80 */ |
| 101 | W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */ |
| 102 | W(0xa0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* a0 */ |
| 103 | W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* b0 */ |
| 104 | W(0xc0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* c0 */ |
| 105 | W(0xd0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* d0 */ |
| 106 | W(0xe0, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0) | /* e0 */ |
| 107 | W(0xf0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1) /* f0 */ |
| 108 | /* ---------------------------------------------- */ |
| 109 | /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */ |
| 110 | }; |
| 111 | #else |
| 112 | #define good_insns_32 NULL |
| 113 | #endif |
| 114 | |
| 115 | /* Good-instruction tables for 64-bit apps. |
| 116 | * |
| 117 | * Genuinely invalid opcodes: |
| 118 | * 06,07 - formerly push/pop es |
| 119 | * 0e - formerly push cs |
| 120 | * 16,17 - formerly push/pop ss |
| 121 | * 1e,1f - formerly push/pop ds |
| 122 | * 27,2f,37,3f - formerly daa/das/aaa/aas |
| 123 | * 60,61 - formerly pusha/popa |
| 124 | * 62 - formerly bound. EVEX prefix for AVX512 (not yet supported) |
| 125 | * 82 - formerly redundant encoding of Group1 |
| 126 | * 9a - formerly call seg:ofs |
| 127 | * ce - formerly into |
| 128 | * d4,d5 - formerly aam/aad |
| 129 | * d6 - formerly undocumented salc |
| 130 | * ea - formerly jmp seg:ofs |
| 131 | * |
| 132 | * Opcodes we'll probably never support: |
| 133 | * 6c-6f - ins,outs. SEGVs if used in userspace |
| 134 | * e4-e7 - in,out imm. SEGVs if used in userspace |
| 135 | * ec-ef - in,out acc. SEGVs if used in userspace |
| 136 | * cc - int3. SIGTRAP if used in userspace |
| 137 | * f1 - int1. SIGTRAP if used in userspace |
| 138 | * f4 - hlt. SEGVs if used in userspace |
| 139 | * fa - cli. SEGVs if used in userspace |
| 140 | * fb - sti. SEGVs if used in userspace |
| 141 | * |
| 142 | * Opcodes which need some work to be supported: |
| 143 | * cd - int N. |
| 144 | * Used by userspace for "int 80" syscall entry. (Other "int N" |
| 145 | * cause GP -> SEGV since their IDT gates don't allow calls from CPL 3). |
| 146 | * Not supported since kernel's handling of userspace single-stepping |
| 147 | * (TF flag) is fragile. |
| 148 | * cf - iret. Normally not used in userspace. Doesn't SEGV unless arguments are bad |
| 149 | */ |
| 150 | #if defined(CONFIG_X86_64) |
| 151 | static volatile u32 good_insns_64[256 / 32] = { |
| 152 | /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */ |
| 153 | /* ---------------------------------------------- */ |
| 154 | W(0x00, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 1) | /* 00 */ |
| 155 | W(0x10, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0) , /* 10 */ |
| 156 | W(0x20, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0) | /* 20 */ |
| 157 | W(0x30, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0) , /* 30 */ |
| 158 | W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */ |
| 159 | W(0x50, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 50 */ |
| 160 | W(0x60, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* 60 */ |
| 161 | W(0x70, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 70 */ |
| 162 | W(0x80, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 80 */ |
| 163 | W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1) , /* 90 */ |
| 164 | W(0xa0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* a0 */ |
| 165 | W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* b0 */ |
| 166 | W(0xc0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0) | /* c0 */ |
| 167 | W(0xd0, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* d0 */ |
| 168 | W(0xe0, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0) | /* e0 */ |
| 169 | W(0xf0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1) /* f0 */ |
| 170 | /* ---------------------------------------------- */ |
| 171 | /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */ |
| 172 | }; |
| 173 | #else |
| 174 | #define good_insns_64 NULL |
| 175 | #endif |
| 176 | |
| 177 | /* Using this for both 64-bit and 32-bit apps. |
| 178 | * Opcodes we don't support: |
| 179 | * 0f 00 - SLDT/STR/LLDT/LTR/VERR/VERW/-/- group. System insns |
| 180 | * 0f 01 - SGDT/SIDT/LGDT/LIDT/SMSW/-/LMSW/INVLPG group. |
| 181 | * Also encodes tons of other system insns if mod=11. |
| 182 | * Some are in fact non-system: xend, xtest, rdtscp, maybe more |
| 183 | * 0f 05 - syscall |
| 184 | * 0f 06 - clts (CPL0 insn) |
| 185 | * 0f 07 - sysret |
| 186 | * 0f 08 - invd (CPL0 insn) |
| 187 | * 0f 09 - wbinvd (CPL0 insn) |
| 188 | * 0f 0b - ud2 |
| 189 | * 0f 30 - wrmsr (CPL0 insn) (then why rdmsr is allowed, it's also CPL0 insn?) |
| 190 | * 0f 34 - sysenter |
| 191 | * 0f 35 - sysexit |
| 192 | * 0f 37 - getsec |
| 193 | * 0f 78 - vmread (Intel VMX. CPL0 insn) |
| 194 | * 0f 79 - vmwrite (Intel VMX. CPL0 insn) |
| 195 | * Note: with prefixes, these two opcodes are |
| 196 | * extrq/insertq/AVX512 convert vector ops. |
| 197 | * 0f ae - group15: [f]xsave,[f]xrstor,[v]{ld,st}mxcsr,clflush[opt], |
| 198 | * {rd,wr}{fs,gs}base,{s,l,m}fence. |
| 199 | * Why? They are all user-executable. |
| 200 | */ |
| 201 | static volatile u32 good_2byte_insns[256 / 32] = { |
| 202 | /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */ |
| 203 | /* ---------------------------------------------- */ |
| 204 | W(0x00, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1) | /* 00 */ |
| 205 | W(0x10, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 10 */ |
| 206 | W(0x20, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 20 */ |
| 207 | W(0x30, 0, 1, 1, 1, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1) , /* 30 */ |
| 208 | W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */ |
| 209 | W(0x50, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 50 */ |
| 210 | W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 60 */ |
| 211 | W(0x70, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1) , /* 70 */ |
| 212 | W(0x80, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 80 */ |
| 213 | W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */ |
| 214 | W(0xa0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1) | /* a0 */ |
| 215 | W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* b0 */ |
| 216 | W(0xc0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */ |
| 217 | W(0xd0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* d0 */ |
| 218 | W(0xe0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* e0 */ |
| 219 | W(0xf0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) /* f0 */ |
| 220 | /* ---------------------------------------------- */ |
| 221 | /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */ |
| 222 | }; |
| 223 | #undef W |
| 224 | |
| 225 | /* |
| 226 | * opcodes we may need to refine support for: |
| 227 | * |
| 228 | * 0f - 2-byte instructions: For many of these instructions, the validity |
| 229 | * depends on the prefix and/or the reg field. On such instructions, we |
| 230 | * just consider the opcode combination valid if it corresponds to any |
| 231 | * valid instruction. |
| 232 | * |
| 233 | * 8f - Group 1 - only reg = 0 is OK |
| 234 | * c6-c7 - Group 11 - only reg = 0 is OK |
| 235 | * d9-df - fpu insns with some illegal encodings |
| 236 | * f2, f3 - repnz, repz prefixes. These are also the first byte for |
| 237 | * certain floating-point instructions, such as addsd. |
| 238 | * |
| 239 | * fe - Group 4 - only reg = 0 or 1 is OK |
| 240 | * ff - Group 5 - only reg = 0-6 is OK |
| 241 | * |
| 242 | * others -- Do we need to support these? |
| 243 | * |
| 244 | * 0f - (floating-point?) prefetch instructions |
| 245 | * 07, 17, 1f - pop es, pop ss, pop ds |
| 246 | * 26, 2e, 36, 3e - es:, cs:, ss:, ds: segment prefixes -- |
| 247 | * but 64 and 65 (fs: and gs:) seem to be used, so we support them |
| 248 | * 67 - addr16 prefix |
| 249 | * ce - into |
| 250 | * f0 - lock prefix |
| 251 | */ |
| 252 | |
| 253 | /* |
| 254 | * TODO: |
| 255 | * - Where necessary, examine the modrm byte and allow only valid instructions |
| 256 | * in the different Groups and fpu instructions. |
| 257 | */ |
| 258 | |
| 259 | static bool is_prefix_bad(struct insn *insn) |
| 260 | { |
| 261 | insn_byte_t p; |
| 262 | |
| 263 | for_each_insn_prefix(insn, p) { |
| 264 | insn_attr_t attr; |
| 265 | |
| 266 | attr = inat_get_opcode_attribute(opcode: p); |
| 267 | switch (attr) { |
| 268 | case INAT_MAKE_PREFIX(INAT_PFX_ES): |
| 269 | case INAT_MAKE_PREFIX(INAT_PFX_CS): |
| 270 | case INAT_MAKE_PREFIX(INAT_PFX_DS): |
| 271 | case INAT_MAKE_PREFIX(INAT_PFX_SS): |
| 272 | case INAT_MAKE_PREFIX(INAT_PFX_LOCK): |
| 273 | return true; |
| 274 | } |
| 275 | } |
| 276 | return false; |
| 277 | } |
| 278 | |
| 279 | static int uprobe_init_insn(struct arch_uprobe *auprobe, struct insn *insn, bool x86_64) |
| 280 | { |
| 281 | enum insn_mode m = x86_64 ? INSN_MODE_64 : INSN_MODE_32; |
| 282 | u32 volatile *good_insns; |
| 283 | int ret; |
| 284 | |
| 285 | ret = insn_decode(insn, kaddr: auprobe->insn, buf_len: sizeof(auprobe->insn), m); |
| 286 | if (ret < 0) |
| 287 | return -ENOEXEC; |
| 288 | |
| 289 | if (is_prefix_bad(insn)) |
| 290 | return -ENOTSUPP; |
| 291 | |
| 292 | /* We should not singlestep on the exception masking instructions */ |
| 293 | if (insn_masking_exception(insn)) |
| 294 | return -ENOTSUPP; |
| 295 | |
| 296 | if (x86_64) |
| 297 | good_insns = good_insns_64; |
| 298 | else |
| 299 | good_insns = good_insns_32; |
| 300 | |
| 301 | if (test_bit(OPCODE1(insn), (unsigned long *)good_insns)) |
| 302 | return 0; |
| 303 | |
| 304 | if (insn->opcode.nbytes == 2) { |
| 305 | if (test_bit(OPCODE2(insn), (unsigned long *)good_2byte_insns)) |
| 306 | return 0; |
| 307 | } |
| 308 | |
| 309 | return -ENOTSUPP; |
| 310 | } |
| 311 | |
| 312 | #ifdef CONFIG_X86_64 |
| 313 | |
| 314 | struct uretprobe_syscall_args { |
| 315 | unsigned long r11; |
| 316 | unsigned long cx; |
| 317 | unsigned long ax; |
| 318 | }; |
| 319 | |
| 320 | asm ( |
| 321 | ".pushsection .rodata\n" |
| 322 | ".global uretprobe_trampoline_entry\n" |
| 323 | "uretprobe_trampoline_entry:\n" |
| 324 | "push %rax\n" |
| 325 | "push %rcx\n" |
| 326 | "push %r11\n" |
| 327 | "mov $" __stringify(__NR_uretprobe) ", %rax\n" |
| 328 | "syscall\n" |
| 329 | ".global uretprobe_syscall_check\n" |
| 330 | "uretprobe_syscall_check:\n" |
| 331 | "pop %r11\n" |
| 332 | "pop %rcx\n" |
| 333 | /* |
| 334 | * The uretprobe syscall replaces stored %rax value with final |
| 335 | * return address, so we don't restore %rax in here and just |
| 336 | * call ret. |
| 337 | */ |
| 338 | "ret\n" |
| 339 | "int3\n" |
| 340 | ".global uretprobe_trampoline_end\n" |
| 341 | "uretprobe_trampoline_end:\n" |
| 342 | ".popsection\n" |
| 343 | ); |
| 344 | |
| 345 | extern u8 uretprobe_trampoline_entry[]; |
| 346 | extern u8 uretprobe_trampoline_end[]; |
| 347 | extern u8 uretprobe_syscall_check[]; |
| 348 | |
| 349 | void *arch_uretprobe_trampoline(unsigned long *psize) |
| 350 | { |
| 351 | static uprobe_opcode_t insn = UPROBE_SWBP_INSN; |
| 352 | struct pt_regs *regs = task_pt_regs(current); |
| 353 | |
| 354 | /* |
| 355 | * At the moment the uretprobe syscall trampoline is supported |
| 356 | * only for native 64-bit process, the compat process still uses |
| 357 | * standard breakpoint. |
| 358 | */ |
| 359 | if (user_64bit_mode(regs)) { |
| 360 | *psize = uretprobe_trampoline_end - uretprobe_trampoline_entry; |
| 361 | return uretprobe_trampoline_entry; |
| 362 | } |
| 363 | |
| 364 | *psize = UPROBE_SWBP_INSN_SIZE; |
| 365 | return &insn; |
| 366 | } |
| 367 | |
| 368 | static unsigned long trampoline_check_ip(unsigned long tramp) |
| 369 | { |
| 370 | return tramp + (uretprobe_syscall_check - uretprobe_trampoline_entry); |
| 371 | } |
| 372 | |
| 373 | SYSCALL_DEFINE0(uretprobe) |
| 374 | { |
| 375 | struct pt_regs *regs = task_pt_regs(current); |
| 376 | struct uretprobe_syscall_args args; |
| 377 | unsigned long err, ip, sp, tramp; |
| 378 | |
| 379 | /* If there's no trampoline, we are called from wrong place. */ |
| 380 | tramp = uprobe_get_trampoline_vaddr(); |
| 381 | if (unlikely(tramp == UPROBE_NO_TRAMPOLINE_VADDR)) |
| 382 | goto sigill; |
| 383 | |
| 384 | /* Make sure the ip matches the only allowed sys_uretprobe caller. */ |
| 385 | if (unlikely(regs->ip != trampoline_check_ip(tramp))) |
| 386 | goto sigill; |
| 387 | |
| 388 | err = copy_from_user(to: &args, from: (void __user *)regs->sp, n: sizeof(args)); |
| 389 | if (err) |
| 390 | goto sigill; |
| 391 | |
| 392 | /* expose the "right" values of r11/cx/ax/sp to uprobe_consumer/s */ |
| 393 | regs->r11 = args.r11; |
| 394 | regs->cx = args.cx; |
| 395 | regs->ax = args.ax; |
| 396 | regs->sp += sizeof(args); |
| 397 | regs->orig_ax = -1; |
| 398 | |
| 399 | ip = regs->ip; |
| 400 | sp = regs->sp; |
| 401 | |
| 402 | uprobe_handle_trampoline(regs); |
| 403 | |
| 404 | /* |
| 405 | * Some of the uprobe consumers has changed sp, we can do nothing, |
| 406 | * just return via iret. |
| 407 | * .. or shadow stack is enabled, in which case we need to skip |
| 408 | * return through the user space stack address. |
| 409 | */ |
| 410 | if (regs->sp != sp || shstk_is_enabled()) |
| 411 | return regs->ax; |
| 412 | regs->sp -= sizeof(args); |
| 413 | |
| 414 | /* for the case uprobe_consumer has changed r11/cx */ |
| 415 | args.r11 = regs->r11; |
| 416 | args.cx = regs->cx; |
| 417 | |
| 418 | /* |
| 419 | * ax register is passed through as return value, so we can use |
| 420 | * its space on stack for ip value and jump to it through the |
| 421 | * trampoline's ret instruction |
| 422 | */ |
| 423 | args.ax = regs->ip; |
| 424 | regs->ip = ip; |
| 425 | |
| 426 | err = copy_to_user(to: (void __user *)regs->sp, from: &args, n: sizeof(args)); |
| 427 | if (err) |
| 428 | goto sigill; |
| 429 | |
| 430 | /* ensure sysret, see do_syscall_64() */ |
| 431 | regs->r11 = regs->flags; |
| 432 | regs->cx = regs->ip; |
| 433 | |
| 434 | return regs->ax; |
| 435 | |
| 436 | sigill: |
| 437 | force_sig(SIGILL); |
| 438 | return -1; |
| 439 | } |
| 440 | |
| 441 | /* |
| 442 | * If arch_uprobe->insn doesn't use rip-relative addressing, return |
| 443 | * immediately. Otherwise, rewrite the instruction so that it accesses |
| 444 | * its memory operand indirectly through a scratch register. Set |
| 445 | * defparam->fixups accordingly. (The contents of the scratch register |
| 446 | * will be saved before we single-step the modified instruction, |
| 447 | * and restored afterward). |
| 448 | * |
| 449 | * We do this because a rip-relative instruction can access only a |
| 450 | * relatively small area (+/- 2 GB from the instruction), and the XOL |
| 451 | * area typically lies beyond that area. At least for instructions |
| 452 | * that store to memory, we can't execute the original instruction |
| 453 | * and "fix things up" later, because the misdirected store could be |
| 454 | * disastrous. |
| 455 | * |
| 456 | * Some useful facts about rip-relative instructions: |
| 457 | * |
| 458 | * - There's always a modrm byte with bit layout "00 reg 101". |
| 459 | * - There's never a SIB byte. |
| 460 | * - The displacement is always 4 bytes. |
| 461 | * - REX.B=1 bit in REX prefix, which normally extends r/m field, |
| 462 | * has no effect on rip-relative mode. It doesn't make modrm byte |
| 463 | * with r/m=101 refer to register 1101 = R13. |
| 464 | */ |
| 465 | static void riprel_analyze(struct arch_uprobe *auprobe, struct insn *insn) |
| 466 | { |
| 467 | u8 *cursor; |
| 468 | u8 reg; |
| 469 | u8 reg2; |
| 470 | |
| 471 | if (!insn_rip_relative(insn)) |
| 472 | return; |
| 473 | |
| 474 | /* |
| 475 | * insn_rip_relative() would have decoded rex_prefix, vex_prefix, modrm. |
| 476 | * Clear REX.b bit (extension of MODRM.rm field): |
| 477 | * we want to encode low numbered reg, not r8+. |
| 478 | */ |
| 479 | if (insn->rex_prefix.nbytes) { |
| 480 | cursor = auprobe->insn + insn_offset_rex_prefix(insn); |
| 481 | /* REX byte has 0100wrxb layout, clearing REX.b bit */ |
| 482 | *cursor &= 0xfe; |
| 483 | } |
| 484 | /* |
| 485 | * Similar treatment for VEX3/EVEX prefix. |
| 486 | * TODO: add XOP treatment when insn decoder supports them |
| 487 | */ |
| 488 | if (insn->vex_prefix.nbytes >= 3) { |
| 489 | /* |
| 490 | * vex2: c5 rvvvvLpp (has no b bit) |
| 491 | * vex3/xop: c4/8f rxbmmmmm wvvvvLpp |
| 492 | * evex: 62 rxbR00mm wvvvv1pp zllBVaaa |
| 493 | * Setting VEX3.b (setting because it has inverted meaning). |
| 494 | * Setting EVEX.x since (in non-SIB encoding) EVEX.x |
| 495 | * is the 4th bit of MODRM.rm, and needs the same treatment. |
| 496 | * For VEX3-encoded insns, VEX3.x value has no effect in |
| 497 | * non-SIB encoding, the change is superfluous but harmless. |
| 498 | */ |
| 499 | cursor = auprobe->insn + insn_offset_vex_prefix(insn) + 1; |
| 500 | *cursor |= 0x60; |
| 501 | } |
| 502 | |
| 503 | /* |
| 504 | * Convert from rip-relative addressing to register-relative addressing |
| 505 | * via a scratch register. |
| 506 | * |
| 507 | * This is tricky since there are insns with modrm byte |
| 508 | * which also use registers not encoded in modrm byte: |
| 509 | * [i]div/[i]mul: implicitly use dx:ax |
| 510 | * shift ops: implicitly use cx |
| 511 | * cmpxchg: implicitly uses ax |
| 512 | * cmpxchg8/16b: implicitly uses dx:ax and bx:cx |
| 513 | * Encoding: 0f c7/1 modrm |
| 514 | * The code below thinks that reg=1 (cx), chooses si as scratch. |
| 515 | * mulx: implicitly uses dx: mulx r/m,r1,r2 does r1:r2 = dx * r/m. |
| 516 | * First appeared in Haswell (BMI2 insn). It is vex-encoded. |
| 517 | * Example where none of bx,cx,dx can be used as scratch reg: |
| 518 | * c4 e2 63 f6 0d disp32 mulx disp32(%rip),%ebx,%ecx |
| 519 | * [v]pcmpistri: implicitly uses cx, xmm0 |
| 520 | * [v]pcmpistrm: implicitly uses xmm0 |
| 521 | * [v]pcmpestri: implicitly uses ax, dx, cx, xmm0 |
| 522 | * [v]pcmpestrm: implicitly uses ax, dx, xmm0 |
| 523 | * Evil SSE4.2 string comparison ops from hell. |
| 524 | * maskmovq/[v]maskmovdqu: implicitly uses (ds:rdi) as destination. |
| 525 | * Encoding: 0f f7 modrm, 66 0f f7 modrm, vex-encoded: c5 f9 f7 modrm. |
| 526 | * Store op1, byte-masked by op2 msb's in each byte, to (ds:rdi). |
| 527 | * AMD says it has no 3-operand form (vex.vvvv must be 1111) |
| 528 | * and that it can have only register operands, not mem |
| 529 | * (its modrm byte must have mode=11). |
| 530 | * If these restrictions will ever be lifted, |
| 531 | * we'll need code to prevent selection of di as scratch reg! |
| 532 | * |
| 533 | * Summary: I don't know any insns with modrm byte which |
| 534 | * use SI register implicitly. DI register is used only |
| 535 | * by one insn (maskmovq) and BX register is used |
| 536 | * only by one too (cmpxchg8b). |
| 537 | * BP is stack-segment based (may be a problem?). |
| 538 | * AX, DX, CX are off-limits (many implicit users). |
| 539 | * SP is unusable (it's stack pointer - think about "pop mem"; |
| 540 | * also, rsp+disp32 needs sib encoding -> insn length change). |
| 541 | */ |
| 542 | |
| 543 | reg = MODRM_REG(insn); /* Fetch modrm.reg */ |
| 544 | reg2 = 0xff; /* Fetch vex.vvvv */ |
| 545 | if (insn->vex_prefix.nbytes) |
| 546 | reg2 = insn->vex_prefix.bytes[2]; |
| 547 | /* |
| 548 | * TODO: add XOP vvvv reading. |
| 549 | * |
| 550 | * vex.vvvv field is in bits 6-3, bits are inverted. |
| 551 | * But in 32-bit mode, high-order bit may be ignored. |
| 552 | * Therefore, let's consider only 3 low-order bits. |
| 553 | */ |
| 554 | reg2 = ((reg2 >> 3) & 0x7) ^ 0x7; |
| 555 | /* |
| 556 | * Register numbering is ax,cx,dx,bx, sp,bp,si,di, r8..r15. |
| 557 | * |
| 558 | * Choose scratch reg. Order is important: must not select bx |
| 559 | * if we can use si (cmpxchg8b case!) |
| 560 | */ |
| 561 | if (reg != 6 && reg2 != 6) { |
| 562 | reg2 = 6; |
| 563 | auprobe->defparam.fixups |= UPROBE_FIX_RIP_SI; |
| 564 | } else if (reg != 7 && reg2 != 7) { |
| 565 | reg2 = 7; |
| 566 | auprobe->defparam.fixups |= UPROBE_FIX_RIP_DI; |
| 567 | /* TODO (paranoia): force maskmovq to not use di */ |
| 568 | } else { |
| 569 | reg2 = 3; |
| 570 | auprobe->defparam.fixups |= UPROBE_FIX_RIP_BX; |
| 571 | } |
| 572 | /* |
| 573 | * Point cursor at the modrm byte. The next 4 bytes are the |
| 574 | * displacement. Beyond the displacement, for some instructions, |
| 575 | * is the immediate operand. |
| 576 | */ |
| 577 | cursor = auprobe->insn + insn_offset_modrm(insn); |
| 578 | /* |
| 579 | * Change modrm from "00 reg 101" to "10 reg reg2". Example: |
| 580 | * 89 05 disp32 mov %eax,disp32(%rip) becomes |
| 581 | * 89 86 disp32 mov %eax,disp32(%rsi) |
| 582 | */ |
| 583 | *cursor = 0x80 | (reg << 3) | reg2; |
| 584 | } |
| 585 | |
| 586 | static inline unsigned long * |
| 587 | scratch_reg(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 588 | { |
| 589 | if (auprobe->defparam.fixups & UPROBE_FIX_RIP_SI) |
| 590 | return ®s->si; |
| 591 | if (auprobe->defparam.fixups & UPROBE_FIX_RIP_DI) |
| 592 | return ®s->di; |
| 593 | return ®s->bx; |
| 594 | } |
| 595 | |
| 596 | /* |
| 597 | * If we're emulating a rip-relative instruction, save the contents |
| 598 | * of the scratch register and store the target address in that register. |
| 599 | */ |
| 600 | static void riprel_pre_xol(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 601 | { |
| 602 | if (auprobe->defparam.fixups & UPROBE_FIX_RIP_MASK) { |
| 603 | struct uprobe_task *utask = current->utask; |
| 604 | unsigned long *sr = scratch_reg(auprobe, regs); |
| 605 | |
| 606 | utask->autask.saved_scratch_register = *sr; |
| 607 | *sr = utask->vaddr + auprobe->defparam.ilen; |
| 608 | } |
| 609 | } |
| 610 | |
| 611 | static void riprel_post_xol(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 612 | { |
| 613 | if (auprobe->defparam.fixups & UPROBE_FIX_RIP_MASK) { |
| 614 | struct uprobe_task *utask = current->utask; |
| 615 | unsigned long *sr = scratch_reg(auprobe, regs); |
| 616 | |
| 617 | *sr = utask->autask.saved_scratch_register; |
| 618 | } |
| 619 | } |
| 620 | |
| 621 | static int tramp_mremap(const struct vm_special_mapping *sm, struct vm_area_struct *new_vma) |
| 622 | { |
| 623 | return -EPERM; |
| 624 | } |
| 625 | |
| 626 | static struct page *tramp_mapping_pages[2] __ro_after_init; |
| 627 | |
| 628 | static struct vm_special_mapping tramp_mapping = { |
| 629 | .name = "[uprobes-trampoline]" , |
| 630 | .mremap = tramp_mremap, |
| 631 | .pages = tramp_mapping_pages, |
| 632 | }; |
| 633 | |
| 634 | struct uprobe_trampoline { |
| 635 | struct hlist_node node; |
| 636 | unsigned long vaddr; |
| 637 | }; |
| 638 | |
| 639 | static bool is_reachable_by_call(unsigned long vtramp, unsigned long vaddr) |
| 640 | { |
| 641 | long delta = (long)(vaddr + 5 - vtramp); |
| 642 | |
| 643 | return delta >= INT_MIN && delta <= INT_MAX; |
| 644 | } |
| 645 | |
| 646 | static unsigned long find_nearest_trampoline(unsigned long vaddr) |
| 647 | { |
| 648 | struct vm_unmapped_area_info info = { |
| 649 | .length = PAGE_SIZE, |
| 650 | .align_mask = ~PAGE_MASK, |
| 651 | }; |
| 652 | unsigned long low_limit, high_limit; |
| 653 | unsigned long low_tramp, high_tramp; |
| 654 | unsigned long call_end = vaddr + 5; |
| 655 | |
| 656 | if (check_add_overflow(call_end, INT_MIN, &low_limit)) |
| 657 | low_limit = PAGE_SIZE; |
| 658 | |
| 659 | high_limit = call_end + INT_MAX; |
| 660 | |
| 661 | /* Search up from the caller address. */ |
| 662 | info.low_limit = call_end; |
| 663 | info.high_limit = min(high_limit, TASK_SIZE); |
| 664 | high_tramp = vm_unmapped_area(info: &info); |
| 665 | |
| 666 | /* Search down from the caller address. */ |
| 667 | info.low_limit = max(low_limit, PAGE_SIZE); |
| 668 | info.high_limit = call_end; |
| 669 | info.flags = VM_UNMAPPED_AREA_TOPDOWN; |
| 670 | low_tramp = vm_unmapped_area(info: &info); |
| 671 | |
| 672 | if (IS_ERR_VALUE(high_tramp) && IS_ERR_VALUE(low_tramp)) |
| 673 | return -ENOMEM; |
| 674 | if (IS_ERR_VALUE(high_tramp)) |
| 675 | return low_tramp; |
| 676 | if (IS_ERR_VALUE(low_tramp)) |
| 677 | return high_tramp; |
| 678 | |
| 679 | /* Return address that's closest to the caller address. */ |
| 680 | if (call_end - low_tramp < high_tramp - call_end) |
| 681 | return low_tramp; |
| 682 | return high_tramp; |
| 683 | } |
| 684 | |
| 685 | static struct uprobe_trampoline *create_uprobe_trampoline(unsigned long vaddr) |
| 686 | { |
| 687 | struct pt_regs *regs = task_pt_regs(current); |
| 688 | struct mm_struct *mm = current->mm; |
| 689 | struct uprobe_trampoline *tramp; |
| 690 | struct vm_area_struct *vma; |
| 691 | |
| 692 | if (!user_64bit_mode(regs)) |
| 693 | return NULL; |
| 694 | |
| 695 | vaddr = find_nearest_trampoline(vaddr); |
| 696 | if (IS_ERR_VALUE(vaddr)) |
| 697 | return NULL; |
| 698 | |
| 699 | tramp = kzalloc(sizeof(*tramp), GFP_KERNEL); |
| 700 | if (unlikely(!tramp)) |
| 701 | return NULL; |
| 702 | |
| 703 | tramp->vaddr = vaddr; |
| 704 | vma = _install_special_mapping(mm, addr: tramp->vaddr, PAGE_SIZE, |
| 705 | VM_READ|VM_EXEC|VM_MAYEXEC|VM_MAYREAD|VM_DONTCOPY|VM_IO, |
| 706 | spec: &tramp_mapping); |
| 707 | if (IS_ERR(ptr: vma)) { |
| 708 | kfree(objp: tramp); |
| 709 | return NULL; |
| 710 | } |
| 711 | return tramp; |
| 712 | } |
| 713 | |
| 714 | static struct uprobe_trampoline *get_uprobe_trampoline(unsigned long vaddr, bool *new) |
| 715 | { |
| 716 | struct uprobes_state *state = ¤t->mm->uprobes_state; |
| 717 | struct uprobe_trampoline *tramp = NULL; |
| 718 | |
| 719 | if (vaddr > TASK_SIZE || vaddr < PAGE_SIZE) |
| 720 | return NULL; |
| 721 | |
| 722 | hlist_for_each_entry(tramp, &state->head_tramps, node) { |
| 723 | if (is_reachable_by_call(vtramp: tramp->vaddr, vaddr)) { |
| 724 | *new = false; |
| 725 | return tramp; |
| 726 | } |
| 727 | } |
| 728 | |
| 729 | tramp = create_uprobe_trampoline(vaddr); |
| 730 | if (!tramp) |
| 731 | return NULL; |
| 732 | |
| 733 | *new = true; |
| 734 | hlist_add_head(n: &tramp->node, h: &state->head_tramps); |
| 735 | return tramp; |
| 736 | } |
| 737 | |
| 738 | static void destroy_uprobe_trampoline(struct uprobe_trampoline *tramp) |
| 739 | { |
| 740 | /* |
| 741 | * We do not unmap and release uprobe trampoline page itself, |
| 742 | * because there's no easy way to make sure none of the threads |
| 743 | * is still inside the trampoline. |
| 744 | */ |
| 745 | hlist_del(n: &tramp->node); |
| 746 | kfree(objp: tramp); |
| 747 | } |
| 748 | |
| 749 | void arch_uprobe_init_state(struct mm_struct *mm) |
| 750 | { |
| 751 | INIT_HLIST_HEAD(&mm->uprobes_state.head_tramps); |
| 752 | } |
| 753 | |
| 754 | void arch_uprobe_clear_state(struct mm_struct *mm) |
| 755 | { |
| 756 | struct uprobes_state *state = &mm->uprobes_state; |
| 757 | struct uprobe_trampoline *tramp; |
| 758 | struct hlist_node *n; |
| 759 | |
| 760 | hlist_for_each_entry_safe(tramp, n, &state->head_tramps, node) |
| 761 | destroy_uprobe_trampoline(tramp); |
| 762 | } |
| 763 | |
| 764 | static bool __in_uprobe_trampoline(unsigned long ip) |
| 765 | { |
| 766 | struct vm_area_struct *vma = vma_lookup(current->mm, addr: ip); |
| 767 | |
| 768 | return vma && vma_is_special_mapping(vma, sm: &tramp_mapping); |
| 769 | } |
| 770 | |
| 771 | static bool in_uprobe_trampoline(unsigned long ip) |
| 772 | { |
| 773 | struct mm_struct *mm = current->mm; |
| 774 | bool found, retry = true; |
| 775 | unsigned int seq; |
| 776 | |
| 777 | rcu_read_lock(); |
| 778 | if (mmap_lock_speculate_try_begin(mm, seq: &seq)) { |
| 779 | found = __in_uprobe_trampoline(ip); |
| 780 | retry = mmap_lock_speculate_retry(mm, seq); |
| 781 | } |
| 782 | rcu_read_unlock(); |
| 783 | |
| 784 | if (retry) { |
| 785 | mmap_read_lock(mm); |
| 786 | found = __in_uprobe_trampoline(ip); |
| 787 | mmap_read_unlock(mm); |
| 788 | } |
| 789 | return found; |
| 790 | } |
| 791 | |
| 792 | /* |
| 793 | * See uprobe syscall trampoline; the call to the trampoline will push |
| 794 | * the return address on the stack, the trampoline itself then pushes |
| 795 | * cx, r11 and ax. |
| 796 | */ |
| 797 | struct uprobe_syscall_args { |
| 798 | unsigned long ax; |
| 799 | unsigned long r11; |
| 800 | unsigned long cx; |
| 801 | unsigned long retaddr; |
| 802 | }; |
| 803 | |
| 804 | SYSCALL_DEFINE0(uprobe) |
| 805 | { |
| 806 | struct pt_regs *regs = task_pt_regs(current); |
| 807 | struct uprobe_syscall_args args; |
| 808 | unsigned long ip, sp, sret; |
| 809 | int err; |
| 810 | |
| 811 | /* Allow execution only from uprobe trampolines. */ |
| 812 | if (!in_uprobe_trampoline(ip: regs->ip)) |
| 813 | return -ENXIO; |
| 814 | |
| 815 | err = copy_from_user(to: &args, from: (void __user *)regs->sp, n: sizeof(args)); |
| 816 | if (err) |
| 817 | goto sigill; |
| 818 | |
| 819 | ip = regs->ip; |
| 820 | |
| 821 | /* |
| 822 | * expose the "right" values of ax/r11/cx/ip/sp to uprobe_consumer/s, plus: |
| 823 | * - adjust ip to the probe address, call saved next instruction address |
| 824 | * - adjust sp to the probe's stack frame (check trampoline code) |
| 825 | */ |
| 826 | regs->ax = args.ax; |
| 827 | regs->r11 = args.r11; |
| 828 | regs->cx = args.cx; |
| 829 | regs->ip = args.retaddr - 5; |
| 830 | regs->sp += sizeof(args); |
| 831 | regs->orig_ax = -1; |
| 832 | |
| 833 | sp = regs->sp; |
| 834 | |
| 835 | err = shstk_pop(val: (u64 *)&sret); |
| 836 | if (err == -EFAULT || (!err && sret != args.retaddr)) |
| 837 | goto sigill; |
| 838 | |
| 839 | handle_syscall_uprobe(regs, bp_vaddr: regs->ip); |
| 840 | |
| 841 | /* |
| 842 | * Some of the uprobe consumers has changed sp, we can do nothing, |
| 843 | * just return via iret. |
| 844 | */ |
| 845 | if (regs->sp != sp) { |
| 846 | /* skip the trampoline call */ |
| 847 | if (args.retaddr - 5 == regs->ip) |
| 848 | regs->ip += 5; |
| 849 | return regs->ax; |
| 850 | } |
| 851 | |
| 852 | regs->sp -= sizeof(args); |
| 853 | |
| 854 | /* for the case uprobe_consumer has changed ax/r11/cx */ |
| 855 | args.ax = regs->ax; |
| 856 | args.r11 = regs->r11; |
| 857 | args.cx = regs->cx; |
| 858 | |
| 859 | /* keep return address unless we are instructed otherwise */ |
| 860 | if (args.retaddr - 5 != regs->ip) |
| 861 | args.retaddr = regs->ip; |
| 862 | |
| 863 | if (shstk_push(val: args.retaddr) == -EFAULT) |
| 864 | goto sigill; |
| 865 | |
| 866 | regs->ip = ip; |
| 867 | |
| 868 | err = copy_to_user(to: (void __user *)regs->sp, from: &args, n: sizeof(args)); |
| 869 | if (err) |
| 870 | goto sigill; |
| 871 | |
| 872 | /* ensure sysret, see do_syscall_64() */ |
| 873 | regs->r11 = regs->flags; |
| 874 | regs->cx = regs->ip; |
| 875 | return 0; |
| 876 | |
| 877 | sigill: |
| 878 | force_sig(SIGILL); |
| 879 | return -1; |
| 880 | } |
| 881 | |
| 882 | asm ( |
| 883 | ".pushsection .rodata\n" |
| 884 | ".balign " __stringify(PAGE_SIZE) "\n" |
| 885 | "uprobe_trampoline_entry:\n" |
| 886 | "push %rcx\n" |
| 887 | "push %r11\n" |
| 888 | "push %rax\n" |
| 889 | "mov $" __stringify(__NR_uprobe) ", %rax\n" |
| 890 | "syscall\n" |
| 891 | "pop %rax\n" |
| 892 | "pop %r11\n" |
| 893 | "pop %rcx\n" |
| 894 | "ret\n" |
| 895 | "int3\n" |
| 896 | ".balign " __stringify(PAGE_SIZE) "\n" |
| 897 | ".popsection\n" |
| 898 | ); |
| 899 | |
| 900 | extern u8 uprobe_trampoline_entry[]; |
| 901 | |
| 902 | static int __init arch_uprobes_init(void) |
| 903 | { |
| 904 | tramp_mapping_pages[0] = virt_to_page(uprobe_trampoline_entry); |
| 905 | return 0; |
| 906 | } |
| 907 | |
| 908 | late_initcall(arch_uprobes_init); |
| 909 | |
| 910 | enum { |
| 911 | EXPECT_SWBP, |
| 912 | EXPECT_CALL, |
| 913 | }; |
| 914 | |
| 915 | struct write_opcode_ctx { |
| 916 | unsigned long base; |
| 917 | int expect; |
| 918 | }; |
| 919 | |
| 920 | static int is_call_insn(uprobe_opcode_t *insn) |
| 921 | { |
| 922 | return *insn == CALL_INSN_OPCODE; |
| 923 | } |
| 924 | |
| 925 | /* |
| 926 | * Verification callback used by int3_update uprobe_write calls to make sure |
| 927 | * the underlying instruction is as expected - either int3 or call. |
| 928 | */ |
| 929 | static int verify_insn(struct page *page, unsigned long vaddr, uprobe_opcode_t *new_opcode, |
| 930 | int nbytes, void *data) |
| 931 | { |
| 932 | struct write_opcode_ctx *ctx = data; |
| 933 | uprobe_opcode_t old_opcode[5]; |
| 934 | |
| 935 | uprobe_copy_from_page(page, vaddr: ctx->base, dst: (uprobe_opcode_t *) &old_opcode, len: 5); |
| 936 | |
| 937 | switch (ctx->expect) { |
| 938 | case EXPECT_SWBP: |
| 939 | if (is_swbp_insn(insn: &old_opcode[0])) |
| 940 | return 1; |
| 941 | break; |
| 942 | case EXPECT_CALL: |
| 943 | if (is_call_insn(insn: &old_opcode[0])) |
| 944 | return 1; |
| 945 | break; |
| 946 | } |
| 947 | |
| 948 | return -1; |
| 949 | } |
| 950 | |
| 951 | /* |
| 952 | * Modify multi-byte instructions by using INT3 breakpoints on SMP. |
| 953 | * We completely avoid using stop_machine() here, and achieve the |
| 954 | * synchronization using INT3 breakpoints and SMP cross-calls. |
| 955 | * (borrowed comment from smp_text_poke_batch_finish) |
| 956 | * |
| 957 | * The way it is done: |
| 958 | * - Add an INT3 trap to the address that will be patched |
| 959 | * - SMP sync all CPUs |
| 960 | * - Update all but the first byte of the patched range |
| 961 | * - SMP sync all CPUs |
| 962 | * - Replace the first byte (INT3) by the first byte of the replacing opcode |
| 963 | * - SMP sync all CPUs |
| 964 | */ |
| 965 | static int int3_update(struct arch_uprobe *auprobe, struct vm_area_struct *vma, |
| 966 | unsigned long vaddr, char *insn, bool optimize) |
| 967 | { |
| 968 | uprobe_opcode_t int3 = UPROBE_SWBP_INSN; |
| 969 | struct write_opcode_ctx ctx = { |
| 970 | .base = vaddr, |
| 971 | }; |
| 972 | int err; |
| 973 | |
| 974 | /* |
| 975 | * Write int3 trap. |
| 976 | * |
| 977 | * The swbp_optimize path comes with breakpoint already installed, |
| 978 | * so we can skip this step for optimize == true. |
| 979 | */ |
| 980 | if (!optimize) { |
| 981 | ctx.expect = EXPECT_CALL; |
| 982 | err = uprobe_write(auprobe, vma, opcode_vaddr: vaddr, insn: &int3, nbytes: 1, verify: verify_insn, |
| 983 | is_register: true /* is_register */, do_update_ref_ctr: false /* do_update_ref_ctr */, |
| 984 | data: &ctx); |
| 985 | if (err) |
| 986 | return err; |
| 987 | } |
| 988 | |
| 989 | smp_text_poke_sync_each_cpu(); |
| 990 | |
| 991 | /* Write all but the first byte of the patched range. */ |
| 992 | ctx.expect = EXPECT_SWBP; |
| 993 | err = uprobe_write(auprobe, vma, opcode_vaddr: vaddr + 1, insn: insn + 1, nbytes: 4, verify: verify_insn, |
| 994 | is_register: true /* is_register */, do_update_ref_ctr: false /* do_update_ref_ctr */, |
| 995 | data: &ctx); |
| 996 | if (err) |
| 997 | return err; |
| 998 | |
| 999 | smp_text_poke_sync_each_cpu(); |
| 1000 | |
| 1001 | /* |
| 1002 | * Write first byte. |
| 1003 | * |
| 1004 | * The swbp_unoptimize needs to finish uprobe removal together |
| 1005 | * with ref_ctr update, using uprobe_write with proper flags. |
| 1006 | */ |
| 1007 | err = uprobe_write(auprobe, vma, opcode_vaddr: vaddr, insn, nbytes: 1, verify: verify_insn, |
| 1008 | is_register: optimize /* is_register */, do_update_ref_ctr: !optimize /* do_update_ref_ctr */, |
| 1009 | data: &ctx); |
| 1010 | if (err) |
| 1011 | return err; |
| 1012 | |
| 1013 | smp_text_poke_sync_each_cpu(); |
| 1014 | return 0; |
| 1015 | } |
| 1016 | |
| 1017 | static int swbp_optimize(struct arch_uprobe *auprobe, struct vm_area_struct *vma, |
| 1018 | unsigned long vaddr, unsigned long tramp) |
| 1019 | { |
| 1020 | u8 call[5]; |
| 1021 | |
| 1022 | __text_gen_insn(buf: call, CALL_INSN_OPCODE, addr: (const void *) vaddr, |
| 1023 | dest: (const void *) tramp, CALL_INSN_SIZE); |
| 1024 | return int3_update(auprobe, vma, vaddr, insn: call, optimize: true /* optimize */); |
| 1025 | } |
| 1026 | |
| 1027 | static int swbp_unoptimize(struct arch_uprobe *auprobe, struct vm_area_struct *vma, |
| 1028 | unsigned long vaddr) |
| 1029 | { |
| 1030 | return int3_update(auprobe, vma, vaddr, insn: auprobe->insn, optimize: false /* optimize */); |
| 1031 | } |
| 1032 | |
| 1033 | static int copy_from_vaddr(struct mm_struct *mm, unsigned long vaddr, void *dst, int len) |
| 1034 | { |
| 1035 | unsigned int gup_flags = FOLL_FORCE|FOLL_SPLIT_PMD; |
| 1036 | struct vm_area_struct *vma; |
| 1037 | struct page *page; |
| 1038 | |
| 1039 | page = get_user_page_vma_remote(mm, addr: vaddr, gup_flags, vmap: &vma); |
| 1040 | if (IS_ERR(ptr: page)) |
| 1041 | return PTR_ERR(ptr: page); |
| 1042 | uprobe_copy_from_page(page, vaddr, dst, len); |
| 1043 | put_page(page); |
| 1044 | return 0; |
| 1045 | } |
| 1046 | |
| 1047 | static bool __is_optimized(uprobe_opcode_t *insn, unsigned long vaddr) |
| 1048 | { |
| 1049 | struct __packed __arch_relative_insn { |
| 1050 | u8 op; |
| 1051 | s32 raddr; |
| 1052 | } *call = (struct __arch_relative_insn *) insn; |
| 1053 | |
| 1054 | if (!is_call_insn(insn)) |
| 1055 | return false; |
| 1056 | return __in_uprobe_trampoline(ip: vaddr + 5 + call->raddr); |
| 1057 | } |
| 1058 | |
| 1059 | static int is_optimized(struct mm_struct *mm, unsigned long vaddr) |
| 1060 | { |
| 1061 | uprobe_opcode_t insn[5]; |
| 1062 | int err; |
| 1063 | |
| 1064 | err = copy_from_vaddr(mm, vaddr, dst: &insn, len: 5); |
| 1065 | if (err) |
| 1066 | return err; |
| 1067 | return __is_optimized(insn: (uprobe_opcode_t *)&insn, vaddr); |
| 1068 | } |
| 1069 | |
| 1070 | static bool should_optimize(struct arch_uprobe *auprobe) |
| 1071 | { |
| 1072 | return !test_bit(ARCH_UPROBE_FLAG_OPTIMIZE_FAIL, &auprobe->flags) && |
| 1073 | test_bit(ARCH_UPROBE_FLAG_CAN_OPTIMIZE, &auprobe->flags); |
| 1074 | } |
| 1075 | |
| 1076 | int set_swbp(struct arch_uprobe *auprobe, struct vm_area_struct *vma, |
| 1077 | unsigned long vaddr) |
| 1078 | { |
| 1079 | if (should_optimize(auprobe)) { |
| 1080 | /* |
| 1081 | * We could race with another thread that already optimized the probe, |
| 1082 | * so let's not overwrite it with int3 again in this case. |
| 1083 | */ |
| 1084 | int ret = is_optimized(mm: vma->vm_mm, vaddr); |
| 1085 | if (ret < 0) |
| 1086 | return ret; |
| 1087 | if (ret) |
| 1088 | return 0; |
| 1089 | } |
| 1090 | return uprobe_write_opcode(auprobe, vma, vaddr, UPROBE_SWBP_INSN, |
| 1091 | is_register: true /* is_register */); |
| 1092 | } |
| 1093 | |
| 1094 | int set_orig_insn(struct arch_uprobe *auprobe, struct vm_area_struct *vma, |
| 1095 | unsigned long vaddr) |
| 1096 | { |
| 1097 | if (test_bit(ARCH_UPROBE_FLAG_CAN_OPTIMIZE, &auprobe->flags)) { |
| 1098 | int ret = is_optimized(mm: vma->vm_mm, vaddr); |
| 1099 | if (ret < 0) |
| 1100 | return ret; |
| 1101 | if (ret) { |
| 1102 | ret = swbp_unoptimize(auprobe, vma, vaddr); |
| 1103 | WARN_ON_ONCE(ret); |
| 1104 | return ret; |
| 1105 | } |
| 1106 | } |
| 1107 | return uprobe_write_opcode(auprobe, vma, vaddr, *(uprobe_opcode_t *)&auprobe->insn, |
| 1108 | is_register: false /* is_register */); |
| 1109 | } |
| 1110 | |
| 1111 | static int __arch_uprobe_optimize(struct arch_uprobe *auprobe, struct mm_struct *mm, |
| 1112 | unsigned long vaddr) |
| 1113 | { |
| 1114 | struct uprobe_trampoline *tramp; |
| 1115 | struct vm_area_struct *vma; |
| 1116 | bool new = false; |
| 1117 | int err = 0; |
| 1118 | |
| 1119 | vma = find_vma(mm, addr: vaddr); |
| 1120 | if (!vma) |
| 1121 | return -EINVAL; |
| 1122 | tramp = get_uprobe_trampoline(vaddr, new: &new); |
| 1123 | if (!tramp) |
| 1124 | return -EINVAL; |
| 1125 | err = swbp_optimize(auprobe, vma, vaddr, tramp: tramp->vaddr); |
| 1126 | if (WARN_ON_ONCE(err) && new) |
| 1127 | destroy_uprobe_trampoline(tramp); |
| 1128 | return err; |
| 1129 | } |
| 1130 | |
| 1131 | void arch_uprobe_optimize(struct arch_uprobe *auprobe, unsigned long vaddr) |
| 1132 | { |
| 1133 | struct mm_struct *mm = current->mm; |
| 1134 | uprobe_opcode_t insn[5]; |
| 1135 | |
| 1136 | if (!should_optimize(auprobe)) |
| 1137 | return; |
| 1138 | |
| 1139 | mmap_write_lock(mm); |
| 1140 | |
| 1141 | /* |
| 1142 | * Check if some other thread already optimized the uprobe for us, |
| 1143 | * if it's the case just go away silently. |
| 1144 | */ |
| 1145 | if (copy_from_vaddr(mm, vaddr, dst: &insn, len: 5)) |
| 1146 | goto unlock; |
| 1147 | if (!is_swbp_insn(insn: (uprobe_opcode_t*) &insn)) |
| 1148 | goto unlock; |
| 1149 | |
| 1150 | /* |
| 1151 | * If we fail to optimize the uprobe we set the fail bit so the |
| 1152 | * above should_optimize will fail from now on. |
| 1153 | */ |
| 1154 | if (__arch_uprobe_optimize(auprobe, mm, vaddr)) |
| 1155 | set_bit(nr: ARCH_UPROBE_FLAG_OPTIMIZE_FAIL, addr: &auprobe->flags); |
| 1156 | |
| 1157 | unlock: |
| 1158 | mmap_write_unlock(mm); |
| 1159 | } |
| 1160 | |
| 1161 | static bool can_optimize(struct insn *insn, unsigned long vaddr) |
| 1162 | { |
| 1163 | if (!insn->x86_64 || insn->length != 5) |
| 1164 | return false; |
| 1165 | |
| 1166 | if (!insn_is_nop(insn)) |
| 1167 | return false; |
| 1168 | |
| 1169 | /* We can't do cross page atomic writes yet. */ |
| 1170 | return PAGE_SIZE - (vaddr & ~PAGE_MASK) >= 5; |
| 1171 | } |
| 1172 | #else /* 32-bit: */ |
| 1173 | /* |
| 1174 | * No RIP-relative addressing on 32-bit |
| 1175 | */ |
| 1176 | static void riprel_analyze(struct arch_uprobe *auprobe, struct insn *insn) |
| 1177 | { |
| 1178 | } |
| 1179 | static void riprel_pre_xol(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1180 | { |
| 1181 | } |
| 1182 | static void riprel_post_xol(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1183 | { |
| 1184 | } |
| 1185 | static bool can_optimize(struct insn *insn, unsigned long vaddr) |
| 1186 | { |
| 1187 | return false; |
| 1188 | } |
| 1189 | #endif /* CONFIG_X86_64 */ |
| 1190 | |
| 1191 | struct uprobe_xol_ops { |
| 1192 | bool (*emulate)(struct arch_uprobe *, struct pt_regs *); |
| 1193 | int (*pre_xol)(struct arch_uprobe *, struct pt_regs *); |
| 1194 | int (*post_xol)(struct arch_uprobe *, struct pt_regs *); |
| 1195 | void (*abort)(struct arch_uprobe *, struct pt_regs *); |
| 1196 | }; |
| 1197 | |
| 1198 | static inline int sizeof_long(struct pt_regs *regs) |
| 1199 | { |
| 1200 | /* |
| 1201 | * Check registers for mode as in_xxx_syscall() does not apply here. |
| 1202 | */ |
| 1203 | return user_64bit_mode(regs) ? 8 : 4; |
| 1204 | } |
| 1205 | |
| 1206 | static int default_pre_xol_op(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1207 | { |
| 1208 | riprel_pre_xol(auprobe, regs); |
| 1209 | return 0; |
| 1210 | } |
| 1211 | |
| 1212 | static int emulate_push_stack(struct pt_regs *regs, unsigned long val) |
| 1213 | { |
| 1214 | unsigned long new_sp = regs->sp - sizeof_long(regs); |
| 1215 | |
| 1216 | if (copy_to_user(to: (void __user *)new_sp, from: &val, n: sizeof_long(regs))) |
| 1217 | return -EFAULT; |
| 1218 | |
| 1219 | regs->sp = new_sp; |
| 1220 | return 0; |
| 1221 | } |
| 1222 | |
| 1223 | /* |
| 1224 | * We have to fix things up as follows: |
| 1225 | * |
| 1226 | * Typically, the new ip is relative to the copied instruction. We need |
| 1227 | * to make it relative to the original instruction (FIX_IP). Exceptions |
| 1228 | * are return instructions and absolute or indirect jump or call instructions. |
| 1229 | * |
| 1230 | * If the single-stepped instruction was a call, the return address that |
| 1231 | * is atop the stack is the address following the copied instruction. We |
| 1232 | * need to make it the address following the original instruction (FIX_CALL). |
| 1233 | * |
| 1234 | * If the original instruction was a rip-relative instruction such as |
| 1235 | * "movl %edx,0xnnnn(%rip)", we have instead executed an equivalent |
| 1236 | * instruction using a scratch register -- e.g., "movl %edx,0xnnnn(%rsi)". |
| 1237 | * We need to restore the contents of the scratch register |
| 1238 | * (FIX_RIP_reg). |
| 1239 | */ |
| 1240 | static int default_post_xol_op(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1241 | { |
| 1242 | struct uprobe_task *utask = current->utask; |
| 1243 | |
| 1244 | riprel_post_xol(auprobe, regs); |
| 1245 | if (auprobe->defparam.fixups & UPROBE_FIX_IP) { |
| 1246 | long correction = utask->vaddr - utask->xol_vaddr; |
| 1247 | regs->ip += correction; |
| 1248 | } else if (auprobe->defparam.fixups & UPROBE_FIX_CALL) { |
| 1249 | regs->sp += sizeof_long(regs); /* Pop incorrect return address */ |
| 1250 | if (emulate_push_stack(regs, val: utask->vaddr + auprobe->defparam.ilen)) |
| 1251 | return -ERESTART; |
| 1252 | } |
| 1253 | /* popf; tell the caller to not touch TF */ |
| 1254 | if (auprobe->defparam.fixups & UPROBE_FIX_SETF) |
| 1255 | utask->autask.saved_tf = true; |
| 1256 | |
| 1257 | return 0; |
| 1258 | } |
| 1259 | |
| 1260 | static void default_abort_op(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1261 | { |
| 1262 | riprel_post_xol(auprobe, regs); |
| 1263 | } |
| 1264 | |
| 1265 | static const struct uprobe_xol_ops default_xol_ops = { |
| 1266 | .pre_xol = default_pre_xol_op, |
| 1267 | .post_xol = default_post_xol_op, |
| 1268 | .abort = default_abort_op, |
| 1269 | }; |
| 1270 | |
| 1271 | static bool branch_is_call(struct arch_uprobe *auprobe) |
| 1272 | { |
| 1273 | return auprobe->branch.opc1 == 0xe8; |
| 1274 | } |
| 1275 | |
| 1276 | #define CASE_COND \ |
| 1277 | COND(70, 71, XF(OF)) \ |
| 1278 | COND(72, 73, XF(CF)) \ |
| 1279 | COND(74, 75, XF(ZF)) \ |
| 1280 | COND(78, 79, XF(SF)) \ |
| 1281 | COND(7a, 7b, XF(PF)) \ |
| 1282 | COND(76, 77, XF(CF) || XF(ZF)) \ |
| 1283 | COND(7c, 7d, XF(SF) != XF(OF)) \ |
| 1284 | COND(7e, 7f, XF(ZF) || XF(SF) != XF(OF)) |
| 1285 | |
| 1286 | #define COND(op_y, op_n, expr) \ |
| 1287 | case 0x ## op_y: DO((expr) != 0) \ |
| 1288 | case 0x ## op_n: DO((expr) == 0) |
| 1289 | |
| 1290 | #define XF(xf) (!!(flags & X86_EFLAGS_ ## xf)) |
| 1291 | |
| 1292 | static bool is_cond_jmp_opcode(u8 opcode) |
| 1293 | { |
| 1294 | switch (opcode) { |
| 1295 | #define DO(expr) \ |
| 1296 | return true; |
| 1297 | CASE_COND |
| 1298 | #undef DO |
| 1299 | |
| 1300 | default: |
| 1301 | return false; |
| 1302 | } |
| 1303 | } |
| 1304 | |
| 1305 | static bool check_jmp_cond(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1306 | { |
| 1307 | unsigned long flags = regs->flags; |
| 1308 | |
| 1309 | switch (auprobe->branch.opc1) { |
| 1310 | #define DO(expr) \ |
| 1311 | return expr; |
| 1312 | CASE_COND |
| 1313 | #undef DO |
| 1314 | |
| 1315 | default: /* not a conditional jmp */ |
| 1316 | return true; |
| 1317 | } |
| 1318 | } |
| 1319 | |
| 1320 | #undef XF |
| 1321 | #undef COND |
| 1322 | #undef CASE_COND |
| 1323 | |
| 1324 | static bool branch_emulate_op(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1325 | { |
| 1326 | unsigned long new_ip = regs->ip += auprobe->branch.ilen; |
| 1327 | unsigned long offs = (long)auprobe->branch.offs; |
| 1328 | |
| 1329 | if (branch_is_call(auprobe)) { |
| 1330 | /* |
| 1331 | * If it fails we execute this (mangled, see the comment in |
| 1332 | * branch_clear_offset) insn out-of-line. In the likely case |
| 1333 | * this should trigger the trap, and the probed application |
| 1334 | * should die or restart the same insn after it handles the |
| 1335 | * signal, arch_uprobe_post_xol() won't be even called. |
| 1336 | * |
| 1337 | * But there is corner case, see the comment in ->post_xol(). |
| 1338 | */ |
| 1339 | if (emulate_push_stack(regs, val: new_ip)) |
| 1340 | return false; |
| 1341 | } else if (!check_jmp_cond(auprobe, regs)) { |
| 1342 | offs = 0; |
| 1343 | } |
| 1344 | |
| 1345 | regs->ip = new_ip + offs; |
| 1346 | return true; |
| 1347 | } |
| 1348 | |
| 1349 | static bool push_emulate_op(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1350 | { |
| 1351 | unsigned long *src_ptr = (void *)regs + auprobe->push.reg_offset; |
| 1352 | |
| 1353 | if (emulate_push_stack(regs, val: *src_ptr)) |
| 1354 | return false; |
| 1355 | regs->ip += auprobe->push.ilen; |
| 1356 | return true; |
| 1357 | } |
| 1358 | |
| 1359 | static int branch_post_xol_op(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1360 | { |
| 1361 | BUG_ON(!branch_is_call(auprobe)); |
| 1362 | /* |
| 1363 | * We can only get here if branch_emulate_op() failed to push the ret |
| 1364 | * address _and_ another thread expanded our stack before the (mangled) |
| 1365 | * "call" insn was executed out-of-line. Just restore ->sp and restart. |
| 1366 | * We could also restore ->ip and try to call branch_emulate_op() again. |
| 1367 | */ |
| 1368 | regs->sp += sizeof_long(regs); |
| 1369 | return -ERESTART; |
| 1370 | } |
| 1371 | |
| 1372 | static void branch_clear_offset(struct arch_uprobe *auprobe, struct insn *insn) |
| 1373 | { |
| 1374 | /* |
| 1375 | * Turn this insn into "call 1f; 1:", this is what we will execute |
| 1376 | * out-of-line if ->emulate() fails. We only need this to generate |
| 1377 | * a trap, so that the probed task receives the correct signal with |
| 1378 | * the properly filled siginfo. |
| 1379 | * |
| 1380 | * But see the comment in ->post_xol(), in the unlikely case it can |
| 1381 | * succeed. So we need to ensure that the new ->ip can not fall into |
| 1382 | * the non-canonical area and trigger #GP. |
| 1383 | * |
| 1384 | * We could turn it into (say) "pushf", but then we would need to |
| 1385 | * divorce ->insn[] and ->ixol[]. We need to preserve the 1st byte |
| 1386 | * of ->insn[] for set_orig_insn(). |
| 1387 | */ |
| 1388 | memset(auprobe->insn + insn_offset_immediate(insn), |
| 1389 | 0, insn->immediate.nbytes); |
| 1390 | } |
| 1391 | |
| 1392 | static const struct uprobe_xol_ops branch_xol_ops = { |
| 1393 | .emulate = branch_emulate_op, |
| 1394 | .post_xol = branch_post_xol_op, |
| 1395 | }; |
| 1396 | |
| 1397 | static const struct uprobe_xol_ops push_xol_ops = { |
| 1398 | .emulate = push_emulate_op, |
| 1399 | }; |
| 1400 | |
| 1401 | /* Returns -ENOSYS if branch_xol_ops doesn't handle this insn */ |
| 1402 | static int branch_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn) |
| 1403 | { |
| 1404 | u8 opc1 = OPCODE1(insn); |
| 1405 | insn_byte_t p; |
| 1406 | |
| 1407 | if (insn_is_nop(insn)) |
| 1408 | goto setup; |
| 1409 | |
| 1410 | switch (opc1) { |
| 1411 | case 0xeb: /* jmp 8 */ |
| 1412 | case 0xe9: /* jmp 32 */ |
| 1413 | break; |
| 1414 | |
| 1415 | case 0xe8: /* call relative */ |
| 1416 | branch_clear_offset(auprobe, insn); |
| 1417 | break; |
| 1418 | |
| 1419 | case 0x0f: |
| 1420 | if (insn->opcode.nbytes != 2) |
| 1421 | return -ENOSYS; |
| 1422 | /* |
| 1423 | * If it is a "near" conditional jmp, OPCODE2() - 0x10 matches |
| 1424 | * OPCODE1() of the "short" jmp which checks the same condition. |
| 1425 | */ |
| 1426 | opc1 = OPCODE2(insn) - 0x10; |
| 1427 | fallthrough; |
| 1428 | default: |
| 1429 | if (!is_cond_jmp_opcode(opcode: opc1)) |
| 1430 | return -ENOSYS; |
| 1431 | } |
| 1432 | |
| 1433 | /* |
| 1434 | * 16-bit overrides such as CALLW (66 e8 nn nn) are not supported. |
| 1435 | * Intel and AMD behavior differ in 64-bit mode: Intel ignores 66 prefix. |
| 1436 | * No one uses these insns, reject any branch insns with such prefix. |
| 1437 | */ |
| 1438 | for_each_insn_prefix(insn, p) { |
| 1439 | if (p == 0x66) |
| 1440 | return -ENOTSUPP; |
| 1441 | } |
| 1442 | |
| 1443 | setup: |
| 1444 | auprobe->branch.opc1 = opc1; |
| 1445 | auprobe->branch.ilen = insn->length; |
| 1446 | auprobe->branch.offs = insn->immediate.value; |
| 1447 | |
| 1448 | auprobe->ops = &branch_xol_ops; |
| 1449 | return 0; |
| 1450 | } |
| 1451 | |
| 1452 | /* Returns -ENOSYS if push_xol_ops doesn't handle this insn */ |
| 1453 | static int push_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn) |
| 1454 | { |
| 1455 | u8 opc1 = OPCODE1(insn), reg_offset = 0; |
| 1456 | |
| 1457 | if (opc1 < 0x50 || opc1 > 0x57) |
| 1458 | return -ENOSYS; |
| 1459 | |
| 1460 | if (insn->length > 2) |
| 1461 | return -ENOSYS; |
| 1462 | if (insn->length == 2) { |
| 1463 | /* only support rex_prefix 0x41 (x64 only) */ |
| 1464 | #ifdef CONFIG_X86_64 |
| 1465 | if (insn->rex_prefix.nbytes != 1 || |
| 1466 | insn->rex_prefix.bytes[0] != 0x41) |
| 1467 | return -ENOSYS; |
| 1468 | |
| 1469 | switch (opc1) { |
| 1470 | case 0x50: |
| 1471 | reg_offset = offsetof(struct pt_regs, r8); |
| 1472 | break; |
| 1473 | case 0x51: |
| 1474 | reg_offset = offsetof(struct pt_regs, r9); |
| 1475 | break; |
| 1476 | case 0x52: |
| 1477 | reg_offset = offsetof(struct pt_regs, r10); |
| 1478 | break; |
| 1479 | case 0x53: |
| 1480 | reg_offset = offsetof(struct pt_regs, r11); |
| 1481 | break; |
| 1482 | case 0x54: |
| 1483 | reg_offset = offsetof(struct pt_regs, r12); |
| 1484 | break; |
| 1485 | case 0x55: |
| 1486 | reg_offset = offsetof(struct pt_regs, r13); |
| 1487 | break; |
| 1488 | case 0x56: |
| 1489 | reg_offset = offsetof(struct pt_regs, r14); |
| 1490 | break; |
| 1491 | case 0x57: |
| 1492 | reg_offset = offsetof(struct pt_regs, r15); |
| 1493 | break; |
| 1494 | } |
| 1495 | #else |
| 1496 | return -ENOSYS; |
| 1497 | #endif |
| 1498 | } else { |
| 1499 | switch (opc1) { |
| 1500 | case 0x50: |
| 1501 | reg_offset = offsetof(struct pt_regs, ax); |
| 1502 | break; |
| 1503 | case 0x51: |
| 1504 | reg_offset = offsetof(struct pt_regs, cx); |
| 1505 | break; |
| 1506 | case 0x52: |
| 1507 | reg_offset = offsetof(struct pt_regs, dx); |
| 1508 | break; |
| 1509 | case 0x53: |
| 1510 | reg_offset = offsetof(struct pt_regs, bx); |
| 1511 | break; |
| 1512 | case 0x54: |
| 1513 | reg_offset = offsetof(struct pt_regs, sp); |
| 1514 | break; |
| 1515 | case 0x55: |
| 1516 | reg_offset = offsetof(struct pt_regs, bp); |
| 1517 | break; |
| 1518 | case 0x56: |
| 1519 | reg_offset = offsetof(struct pt_regs, si); |
| 1520 | break; |
| 1521 | case 0x57: |
| 1522 | reg_offset = offsetof(struct pt_regs, di); |
| 1523 | break; |
| 1524 | } |
| 1525 | } |
| 1526 | |
| 1527 | auprobe->push.reg_offset = reg_offset; |
| 1528 | auprobe->push.ilen = insn->length; |
| 1529 | auprobe->ops = &push_xol_ops; |
| 1530 | return 0; |
| 1531 | } |
| 1532 | |
| 1533 | /** |
| 1534 | * arch_uprobe_analyze_insn - instruction analysis including validity and fixups. |
| 1535 | * @auprobe: the probepoint information. |
| 1536 | * @mm: the probed address space. |
| 1537 | * @addr: virtual address at which to install the probepoint |
| 1538 | * Return 0 on success or a -ve number on error. |
| 1539 | */ |
| 1540 | int arch_uprobe_analyze_insn(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long addr) |
| 1541 | { |
| 1542 | u8 fix_ip_or_call = UPROBE_FIX_IP; |
| 1543 | struct insn insn; |
| 1544 | int ret; |
| 1545 | |
| 1546 | ret = uprobe_init_insn(auprobe, insn: &insn, x86_64: is_64bit_mm(mm)); |
| 1547 | if (ret) |
| 1548 | return ret; |
| 1549 | |
| 1550 | if (can_optimize(insn: &insn, vaddr: addr)) |
| 1551 | set_bit(nr: ARCH_UPROBE_FLAG_CAN_OPTIMIZE, addr: &auprobe->flags); |
| 1552 | |
| 1553 | ret = branch_setup_xol_ops(auprobe, insn: &insn); |
| 1554 | if (ret != -ENOSYS) |
| 1555 | return ret; |
| 1556 | |
| 1557 | ret = push_setup_xol_ops(auprobe, insn: &insn); |
| 1558 | if (ret != -ENOSYS) |
| 1559 | return ret; |
| 1560 | |
| 1561 | /* |
| 1562 | * Figure out which fixups default_post_xol_op() will need to perform, |
| 1563 | * and annotate defparam->fixups accordingly. |
| 1564 | */ |
| 1565 | switch (OPCODE1(&insn)) { |
| 1566 | case 0x9d: /* popf */ |
| 1567 | auprobe->defparam.fixups |= UPROBE_FIX_SETF; |
| 1568 | break; |
| 1569 | case 0xc3: /* ret or lret -- ip is correct */ |
| 1570 | case 0xcb: |
| 1571 | case 0xc2: |
| 1572 | case 0xca: |
| 1573 | case 0xea: /* jmp absolute -- ip is correct */ |
| 1574 | fix_ip_or_call = 0; |
| 1575 | break; |
| 1576 | case 0x9a: /* call absolute - Fix return addr, not ip */ |
| 1577 | fix_ip_or_call = UPROBE_FIX_CALL; |
| 1578 | break; |
| 1579 | case 0xff: |
| 1580 | switch (MODRM_REG(&insn)) { |
| 1581 | case 2: case 3: /* call or lcall, indirect */ |
| 1582 | fix_ip_or_call = UPROBE_FIX_CALL; |
| 1583 | break; |
| 1584 | case 4: case 5: /* jmp or ljmp, indirect */ |
| 1585 | fix_ip_or_call = 0; |
| 1586 | break; |
| 1587 | } |
| 1588 | fallthrough; |
| 1589 | default: |
| 1590 | riprel_analyze(auprobe, insn: &insn); |
| 1591 | } |
| 1592 | |
| 1593 | auprobe->defparam.ilen = insn.length; |
| 1594 | auprobe->defparam.fixups |= fix_ip_or_call; |
| 1595 | |
| 1596 | auprobe->ops = &default_xol_ops; |
| 1597 | return 0; |
| 1598 | } |
| 1599 | |
| 1600 | /* |
| 1601 | * arch_uprobe_pre_xol - prepare to execute out of line. |
| 1602 | * @auprobe: the probepoint information. |
| 1603 | * @regs: reflects the saved user state of current task. |
| 1604 | */ |
| 1605 | int arch_uprobe_pre_xol(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1606 | { |
| 1607 | struct uprobe_task *utask = current->utask; |
| 1608 | |
| 1609 | if (auprobe->ops->pre_xol) { |
| 1610 | int err = auprobe->ops->pre_xol(auprobe, regs); |
| 1611 | if (err) |
| 1612 | return err; |
| 1613 | } |
| 1614 | |
| 1615 | regs->ip = utask->xol_vaddr; |
| 1616 | utask->autask.saved_trap_nr = current->thread.trap_nr; |
| 1617 | current->thread.trap_nr = UPROBE_TRAP_NR; |
| 1618 | |
| 1619 | utask->autask.saved_tf = !!(regs->flags & X86_EFLAGS_TF); |
| 1620 | regs->flags |= X86_EFLAGS_TF; |
| 1621 | if (test_tsk_thread_flag(current, TIF_BLOCKSTEP)) |
| 1622 | set_task_blockstep(current, on: false); |
| 1623 | |
| 1624 | return 0; |
| 1625 | } |
| 1626 | |
| 1627 | /* |
| 1628 | * If xol insn itself traps and generates a signal(Say, |
| 1629 | * SIGILL/SIGSEGV/etc), then detect the case where a singlestepped |
| 1630 | * instruction jumps back to its own address. It is assumed that anything |
| 1631 | * like do_page_fault/do_trap/etc sets thread.trap_nr != -1. |
| 1632 | * |
| 1633 | * arch_uprobe_pre_xol/arch_uprobe_post_xol save/restore thread.trap_nr, |
| 1634 | * arch_uprobe_xol_was_trapped() simply checks that ->trap_nr is not equal to |
| 1635 | * UPROBE_TRAP_NR == -1 set by arch_uprobe_pre_xol(). |
| 1636 | */ |
| 1637 | bool arch_uprobe_xol_was_trapped(struct task_struct *t) |
| 1638 | { |
| 1639 | if (t->thread.trap_nr != UPROBE_TRAP_NR) |
| 1640 | return true; |
| 1641 | |
| 1642 | return false; |
| 1643 | } |
| 1644 | |
| 1645 | /* |
| 1646 | * Called after single-stepping. To avoid the SMP problems that can |
| 1647 | * occur when we temporarily put back the original opcode to |
| 1648 | * single-step, we single-stepped a copy of the instruction. |
| 1649 | * |
| 1650 | * This function prepares to resume execution after the single-step. |
| 1651 | */ |
| 1652 | int arch_uprobe_post_xol(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1653 | { |
| 1654 | struct uprobe_task *utask = current->utask; |
| 1655 | bool send_sigtrap = utask->autask.saved_tf; |
| 1656 | int err = 0; |
| 1657 | |
| 1658 | WARN_ON_ONCE(current->thread.trap_nr != UPROBE_TRAP_NR); |
| 1659 | current->thread.trap_nr = utask->autask.saved_trap_nr; |
| 1660 | |
| 1661 | if (auprobe->ops->post_xol) { |
| 1662 | err = auprobe->ops->post_xol(auprobe, regs); |
| 1663 | if (err) { |
| 1664 | /* |
| 1665 | * Restore ->ip for restart or post mortem analysis. |
| 1666 | * ->post_xol() must not return -ERESTART unless this |
| 1667 | * is really possible. |
| 1668 | */ |
| 1669 | regs->ip = utask->vaddr; |
| 1670 | if (err == -ERESTART) |
| 1671 | err = 0; |
| 1672 | send_sigtrap = false; |
| 1673 | } |
| 1674 | } |
| 1675 | /* |
| 1676 | * arch_uprobe_pre_xol() doesn't save the state of TIF_BLOCKSTEP |
| 1677 | * so we can get an extra SIGTRAP if we do not clear TF. We need |
| 1678 | * to examine the opcode to make it right. |
| 1679 | */ |
| 1680 | if (send_sigtrap) |
| 1681 | send_sig(SIGTRAP, current, 0); |
| 1682 | |
| 1683 | if (!utask->autask.saved_tf) |
| 1684 | regs->flags &= ~X86_EFLAGS_TF; |
| 1685 | |
| 1686 | return err; |
| 1687 | } |
| 1688 | |
| 1689 | /* callback routine for handling exceptions. */ |
| 1690 | int arch_uprobe_exception_notify(struct notifier_block *self, unsigned long val, void *data) |
| 1691 | { |
| 1692 | struct die_args *args = data; |
| 1693 | struct pt_regs *regs = args->regs; |
| 1694 | int ret = NOTIFY_DONE; |
| 1695 | |
| 1696 | /* We are only interested in userspace traps */ |
| 1697 | if (regs && !user_mode(regs)) |
| 1698 | return NOTIFY_DONE; |
| 1699 | |
| 1700 | switch (val) { |
| 1701 | case DIE_INT3: |
| 1702 | if (uprobe_pre_sstep_notifier(regs)) |
| 1703 | ret = NOTIFY_STOP; |
| 1704 | |
| 1705 | break; |
| 1706 | |
| 1707 | case DIE_DEBUG: |
| 1708 | if (uprobe_post_sstep_notifier(regs)) |
| 1709 | ret = NOTIFY_STOP; |
| 1710 | |
| 1711 | break; |
| 1712 | |
| 1713 | default: |
| 1714 | break; |
| 1715 | } |
| 1716 | |
| 1717 | return ret; |
| 1718 | } |
| 1719 | |
| 1720 | /* |
| 1721 | * This function gets called when XOL instruction either gets trapped or |
| 1722 | * the thread has a fatal signal. Reset the instruction pointer to its |
| 1723 | * probed address for the potential restart or for post mortem analysis. |
| 1724 | */ |
| 1725 | void arch_uprobe_abort_xol(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1726 | { |
| 1727 | struct uprobe_task *utask = current->utask; |
| 1728 | |
| 1729 | if (auprobe->ops->abort) |
| 1730 | auprobe->ops->abort(auprobe, regs); |
| 1731 | |
| 1732 | current->thread.trap_nr = utask->autask.saved_trap_nr; |
| 1733 | regs->ip = utask->vaddr; |
| 1734 | /* clear TF if it was set by us in arch_uprobe_pre_xol() */ |
| 1735 | if (!utask->autask.saved_tf) |
| 1736 | regs->flags &= ~X86_EFLAGS_TF; |
| 1737 | } |
| 1738 | |
| 1739 | static bool __skip_sstep(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1740 | { |
| 1741 | if (auprobe->ops->emulate) |
| 1742 | return auprobe->ops->emulate(auprobe, regs); |
| 1743 | return false; |
| 1744 | } |
| 1745 | |
| 1746 | bool arch_uprobe_skip_sstep(struct arch_uprobe *auprobe, struct pt_regs *regs) |
| 1747 | { |
| 1748 | bool ret = __skip_sstep(auprobe, regs); |
| 1749 | if (ret && (regs->flags & X86_EFLAGS_TF)) |
| 1750 | send_sig(SIGTRAP, current, 0); |
| 1751 | return ret; |
| 1752 | } |
| 1753 | |
| 1754 | unsigned long |
| 1755 | arch_uretprobe_hijack_return_addr(unsigned long trampoline_vaddr, struct pt_regs *regs) |
| 1756 | { |
| 1757 | int rasize = sizeof_long(regs), nleft; |
| 1758 | unsigned long orig_ret_vaddr = 0; /* clear high bits for 32-bit apps */ |
| 1759 | |
| 1760 | if (copy_from_user(to: &orig_ret_vaddr, from: (void __user *)regs->sp, n: rasize)) |
| 1761 | return -1; |
| 1762 | |
| 1763 | /* check whether address has been already hijacked */ |
| 1764 | if (orig_ret_vaddr == trampoline_vaddr) |
| 1765 | return orig_ret_vaddr; |
| 1766 | |
| 1767 | nleft = copy_to_user(to: (void __user *)regs->sp, from: &trampoline_vaddr, n: rasize); |
| 1768 | if (likely(!nleft)) { |
| 1769 | if (shstk_update_last_frame(val: trampoline_vaddr)) { |
| 1770 | force_sig(SIGSEGV); |
| 1771 | return -1; |
| 1772 | } |
| 1773 | return orig_ret_vaddr; |
| 1774 | } |
| 1775 | |
| 1776 | if (nleft != rasize) { |
| 1777 | pr_err("return address clobbered: pid=%d, %%sp=%#lx, %%ip=%#lx\n" , |
| 1778 | current->pid, regs->sp, regs->ip); |
| 1779 | |
| 1780 | force_sig(SIGSEGV); |
| 1781 | } |
| 1782 | |
| 1783 | return -1; |
| 1784 | } |
| 1785 | |
| 1786 | bool arch_uretprobe_is_alive(struct return_instance *ret, enum rp_check ctx, |
| 1787 | struct pt_regs *regs) |
| 1788 | { |
| 1789 | if (ctx == RP_CHECK_CALL) /* sp was just decremented by "call" insn */ |
| 1790 | return regs->sp < ret->stack; |
| 1791 | else |
| 1792 | return regs->sp <= ret->stack; |
| 1793 | } |
| 1794 | |
| 1795 | /* |
| 1796 | * Heuristic-based check if uprobe is installed at the function entry. |
| 1797 | * |
| 1798 | * Under assumption of user code being compiled with frame pointers, |
| 1799 | * `push %rbp/%ebp` is a good indicator that we indeed are. |
| 1800 | * |
| 1801 | * Similarly, `endbr64` (assuming 64-bit mode) is also a common pattern. |
| 1802 | * If we get this wrong, captured stack trace might have one extra bogus |
| 1803 | * entry, but the rest of stack trace will still be meaningful. |
| 1804 | */ |
| 1805 | bool is_uprobe_at_func_entry(struct pt_regs *regs) |
| 1806 | { |
| 1807 | struct arch_uprobe *auprobe; |
| 1808 | |
| 1809 | if (!current->utask) |
| 1810 | return false; |
| 1811 | |
| 1812 | auprobe = current->utask->auprobe; |
| 1813 | if (!auprobe) |
| 1814 | return false; |
| 1815 | |
| 1816 | /* push %rbp/%ebp */ |
| 1817 | if (auprobe->insn[0] == 0x55) |
| 1818 | return true; |
| 1819 | |
| 1820 | /* endbr64 (64-bit only) */ |
| 1821 | if (user_64bit_mode(regs) && is_endbr(val: (u32 *)auprobe->insn)) |
| 1822 | return true; |
| 1823 | |
| 1824 | return false; |
| 1825 | } |
| 1826 | |