| 1 | #ifndef UST_PROCESSOR_H |
| 2 | #define UST_PROCESSOR_H |
| 3 | |
| 4 | #include <stddef.h> |
| 5 | #include <string.h> |
| 6 | |
| 7 | extern __thread long ust_reg_stack[500]; |
| 8 | extern volatile __thread long *ust_reg_stack_ptr; |
| 9 | |
| 10 | #ifndef __x86_64 |
| 11 | |
| 12 | struct registers { |
| 13 | short ss; |
| 14 | short cs; |
| 15 | long esi; |
| 16 | long ebp; |
| 17 | long edx; |
| 18 | long edi; |
| 19 | long ecx; |
| 20 | long ebx; |
| 21 | long eax; |
| 22 | long eflags; |
| 23 | long esp; |
| 24 | }; |
| 25 | |
| 26 | #ifdef CONFIG_UST_GDB_INTEGRATION |
| 27 | |
| 28 | /* save_registers - saves most of the processor's registers so |
| 29 | * they are available to the probe. gdb uses this to give the |
| 30 | * value of local variables. |
| 31 | * |
| 32 | * Saving all registers without losing any of their values is |
| 33 | * tricky. |
| 34 | * |
| 35 | * We cannot pass to the asm stub the address of a registers structure |
| 36 | * on the stack, because it will use a register and override its value. |
| 37 | * |
| 38 | * We don't want to use a stub to push the regs on the stack and then |
| 39 | * another stub to copy them to a structure because changing %sp in asm |
| 40 | * and then returning to C (even briefly) can have unexpected results. |
| 41 | * Also, gcc might modify %sp between the stubs in reaction to the |
| 42 | * register needs of the second stub that needs to know where to copy |
| 43 | * the register values. |
| 44 | * |
| 45 | * So the chosen approach is to use another stack, declared in thread- |
| 46 | * local storage, to push the registers. They are subsequently copied |
| 47 | * to the stack, by C code. |
| 48 | */ |
| 49 | |
| 50 | #define save_registers(regsptr) \ |
| 51 | asm volatile ( \ |
| 52 | /* save original esp */ \ |
| 53 | "pushl %%esp\n\t" \ |
| 54 | /* push original eflags */ \ |
| 55 | "pushfl\n\t" \ |
| 56 | /* eax will hold the ptr to the private stack bottom */ \ |
| 57 | "pushl %%eax\n\t" \ |
| 58 | /* ebx is used for TLS access */ \ |
| 59 | "pushl %%ebx\n\t" \ |
| 60 | /* ecx will be used to temporarily hold the stack bottom addr */\ |
| 61 | "pushl %%ecx\n\t" \ |
| 62 | /* rdi is the input to __tls_get_addr, and also a temp var */ \ |
| 63 | "pushl %%edi\n\t" \ |
| 64 | /* For TLS access, we have to do function calls. However, \ |
| 65 | * we must not lose the original value of: \ |
| 66 | * esp, eflags, eax, ebx, ecx, edx, esi, edi, ebp, cs, ss \ |
| 67 | * \ |
| 68 | * Some registers' original values have already been saved: \ |
| 69 | * esp, eflags, eax, ebx, ecx, edi \ |
| 70 | * \ |
| 71 | * In addition, the i386 ABI says the following registers belong\ |
| 72 | * to the caller function: \ |
| 73 | * esp, ebp, esi, edi, ebx \ |
| 74 | * \ |
| 75 | * The following registers should not be changed by the callee: \ |
| 76 | * cs, ss \ |
| 77 | * \ |
| 78 | * Therefore, the following registers must be explicitly \ |
| 79 | * preserved: \ |
| 80 | * edx \ |
| 81 | */ \ |
| 82 | "pushl %%edx\n\t" \ |
| 83 | /* Get GOT address */ \ |
| 84 | "call __i686.get_pc_thunk.bx\n\t" \ |
| 85 | "addl $_GLOBAL_OFFSET_TABLE_, %%ebx\n\t" \ |
| 86 | /* Start TLS access of private reg stack pointer */ \ |
| 87 | "leal ust_reg_stack_ptr@tlsgd(,%%ebx,1),%%eax\n\t" \ |
| 88 | "call ___tls_get_addr@plt\n\t" \ |
| 89 | /* --- End TLS access */ \ |
| 90 | /* check if ust_reg_stack_ptr has been initialized */ \ |
| 91 | "movl (%%eax),%%ecx\n\t" \ |
| 92 | "testl %%ecx,%%ecx\n\t" \ |
| 93 | "jne 1f\n\t" \ |
| 94 | "movl %%eax,%%ecx\n\t" \ |
| 95 | /* Save ecx because we are using it. */ \ |
| 96 | "pushl %%ecx\n\t" \ |
| 97 | /* Start TLS access of private reg stack */ \ |
| 98 | "leal ust_reg_stack@tlsgd(,%%ebx,1),%%eax\n\t" \ |
| 99 | "call ___tls_get_addr@plt\n\t" \ |
| 100 | /* --- End TLS access */ \ |
| 101 | "popl %%ecx\n\t" \ |
| 102 | "addl $500,%%eax\n\t" \ |
| 103 | "movl %%eax,(%%ecx)\n\t" \ |
| 104 | "movl %%ecx,%%eax\n\t" \ |
| 105 | /* now the pointer to the private stack is in eax. \ |
| 106 | must add stack size so the ptr points to the stack bottom. */ \ |
| 107 | "1:\n\t" \ |
| 108 | /* edx was pushed for function calls */ \ |
| 109 | "popl %%edx\n\t" \ |
| 110 | /* Manually push esp to private stack */ \ |
| 111 | "addl $-4,(%%eax)\n\t" \ |
| 112 | "movl 20(%%esp), %%edi\n\t" \ |
| 113 | "movl (%%eax), %%ebx\n\t" \ |
| 114 | "movl %%edi, (%%ebx)\n\t" \ |
| 115 | /* Manually push eflags to private stack */ \ |
| 116 | "addl $-4,(%%eax)\n\t" \ |
| 117 | "movl 16(%%esp), %%edi\n\t" \ |
| 118 | "movl (%%eax), %%ebx\n\t" \ |
| 119 | "movl %%edi, (%%ebx)\n\t" \ |
| 120 | /* Manually push eax to private stack */ \ |
| 121 | "addl $-4,(%%eax)\n\t" \ |
| 122 | "movl 12(%%esp), %%edi\n\t" \ |
| 123 | "movl (%%eax), %%ebx\n\t" \ |
| 124 | "movl %%edi, (%%ebx)\n\t" \ |
| 125 | /* Manually push ebx to private stack */ \ |
| 126 | "addl $-4,(%%eax)\n\t" \ |
| 127 | "movl 8(%%esp), %%edi\n\t" \ |
| 128 | "movl (%%eax), %%ebx\n\t" \ |
| 129 | "movl %%edi, (%%ebx)\n\t" \ |
| 130 | /* Manually push ecx to private stack */ \ |
| 131 | "addl $-4,(%%eax)\n\t" \ |
| 132 | "movl 4(%%esp), %%edi\n\t" \ |
| 133 | "movl (%%eax), %%ebx\n\t" \ |
| 134 | "movl %%edi, (%%ebx)\n\t" \ |
| 135 | /* Manually push edi to private stack */ \ |
| 136 | "addl $-4,(%%eax)\n\t" \ |
| 137 | "movl 0(%%esp), %%edi\n\t" \ |
| 138 | "movl (%%eax), %%ebx\n\t" \ |
| 139 | "movl %%edi, (%%ebx)\n\t" \ |
| 140 | /* now push regs to tls */ \ |
| 141 | /* -- esp already pushed -- */ \ |
| 142 | /* -- eax already pushed -- */ \ |
| 143 | /* -- ebx already pushed -- */ \ |
| 144 | /* -- ecx already pushed -- */ \ |
| 145 | /* -- edi already pushed -- */ \ |
| 146 | "addl $-4,(%%eax)\n\t" \ |
| 147 | "movl (%%eax), %%ebx\n\t" \ |
| 148 | "movl %%edx,(%%ebx)\n\t" \ |
| 149 | "addl $-4,(%%eax)\n\t" \ |
| 150 | "movl (%%eax), %%ebx\n\t" \ |
| 151 | "movl %%ebp,(%%ebx)\n\t" \ |
| 152 | "addl $-4,(%%eax)\n\t" \ |
| 153 | "movl (%%eax), %%ebx\n\t" \ |
| 154 | "movl %%esi,(%%ebx)\n\t" \ |
| 155 | /* push cs */ \ |
| 156 | "addl $-2,(%%eax)\n\t" \ |
| 157 | "movl (%%eax), %%ebx\n\t" \ |
| 158 | "movw %%cs, (%%ebx)\n\t" \ |
| 159 | /* push ss */ \ |
| 160 | "addl $-2,(%%eax)\n\t" \ |
| 161 | "movl (%%eax), %%ebx\n\t" \ |
| 162 | "movw %%ss, (%%ebx)\n\t" \ |
| 163 | /* restore original values of regs that were used internally */ \ |
| 164 | "popl %%edi\n\t" \ |
| 165 | "popl %%ecx\n\t" \ |
| 166 | "popl %%ebx\n\t" \ |
| 167 | "popl %%eax\n\t" \ |
| 168 | /* cancel push of rsp */ \ |
| 169 | "addl $4,%%esp\n\t" \ |
| 170 | /* cancel push of eflags */ \ |
| 171 | "addl $4,%%esp\n\t" \ |
| 172 | ::: "memory"); \ |
| 173 | memcpy(regsptr, (void *)ust_reg_stack_ptr, sizeof(struct registers)); \ |
| 174 | ust_reg_stack_ptr = (void *)(((long)ust_reg_stack_ptr) + sizeof(struct registers)); |
| 175 | |
| 176 | #else /* CONFIG_UST_GDB_INTEGRATION */ |
| 177 | |
| 178 | #define save_registers(a) |
| 179 | |
| 180 | #endif /* CONFIG_UST_GDB_INTEGRATION */ |
| 181 | |
| 182 | #define RELATIVE_ADDRESS(__rel_label__) __rel_label__ |
| 183 | |
| 184 | #define ARCH_COPY_ADDR(src, dst) "lea " src "," dst |
| 185 | |
| 186 | #define _ASM_PTR ".long " |
| 187 | |
| 188 | #else /* below is code for x86-64 */ |
| 189 | |
| 190 | struct registers { |
| 191 | int padding; /* 4 bytes */ |
| 192 | short ss; |
| 193 | short cs; |
| 194 | unsigned long r15; |
| 195 | unsigned long r14; |
| 196 | unsigned long r13; |
| 197 | unsigned long r12; |
| 198 | unsigned long r11; |
| 199 | unsigned long r10; |
| 200 | unsigned long r9; |
| 201 | unsigned long r8; |
| 202 | unsigned long rsi; |
| 203 | unsigned long rbp; |
| 204 | unsigned long rdx; |
| 205 | unsigned long rcx; |
| 206 | unsigned long rdi; |
| 207 | unsigned long rbx; |
| 208 | unsigned long rax; |
| 209 | unsigned long rflags; |
| 210 | unsigned long rsp; |
| 211 | }; |
| 212 | |
| 213 | #ifdef CONFIG_UST_GDB_INTEGRATION |
| 214 | |
| 215 | #define save_registers(regsptr) \ |
| 216 | asm volatile ( \ |
| 217 | /* save original rsp */ \ |
| 218 | "pushq %%rsp\n\t" \ |
| 219 | /* push original rflags */ \ |
| 220 | "pushfq\n\t" \ |
| 221 | /* rax will hold the ptr to the private stack bottom */ \ |
| 222 | "pushq %%rax\n\t" \ |
| 223 | /* rbx will be used to temporarily hold the stack bottom addr */ \ |
| 224 | "pushq %%rbx\n\t" \ |
| 225 | /* rdi is the input to __tls_get_addr, and also a temp var */ \ |
| 226 | "pushq %%rdi\n\t" \ |
| 227 | /* For TLS access, we have to do function calls. However, \ |
| 228 | * we must not lose the original value of: \ |
| 229 | * rsp, rflags, rax, rbx, rcx, rdx, rsi, rdi, rbp, r8, r9 \ |
| 230 | * r10, r11, r12, r13, r14, r15, cs, ss \ |
| 231 | * \ |
| 232 | * Some registers' original values have already been saved: \ |
| 233 | * rsp, rflags, rax, rbx, rdi \ |
| 234 | * \ |
| 235 | * In addition, the x86-64 ABI says the following registers \ |
| 236 | * belong to the caller function: \ |
| 237 | * rbp, rbx, r12, r13, r14, r15 \ |
| 238 | * \ |
| 239 | * The following registers should not be changed by the callee: \ |
| 240 | * cs, ss \ |
| 241 | * \ |
| 242 | * Therefore, the following registers must be explicitly \ |
| 243 | * preserved: \ |
| 244 | * rcx, rdx, rsi, r8, r9, r10, r11 \ |
| 245 | */ \ |
| 246 | "pushq %%rcx\n\t" \ |
| 247 | "pushq %%rdx\n\t" \ |
| 248 | "pushq %%rsi\n\t" \ |
| 249 | "pushq %%r8\n\t" \ |
| 250 | "pushq %%r9\n\t" \ |
| 251 | "pushq %%r10\n\t" \ |
| 252 | "pushq %%r11\n\t" \ |
| 253 | /* Start TLS access of private reg stack pointer */ \ |
| 254 | ".byte 0x66\n\t" \ |
| 255 | "leaq ust_reg_stack_ptr@tlsgd(%%rip), %%rdi\n\t" \ |
| 256 | ".word 0x6666\n\t" \ |
| 257 | "rex64\n\t" \ |
| 258 | "call __tls_get_addr@plt\n\t" \ |
| 259 | /* --- End TLS access */ \ |
| 260 | /* check if ust_reg_stack_ptr has been initialized */ \ |
| 261 | "movq (%%rax),%%rbx\n\t" \ |
| 262 | "testq %%rbx,%%rbx\n\t" \ |
| 263 | "jne 1f\n\t" \ |
| 264 | "movq %%rax,%%rbx\n\t" \ |
| 265 | /* Start TLS access of private reg stack */ \ |
| 266 | ".byte 0x66\n\t" \ |
| 267 | "leaq ust_reg_stack@tlsgd(%%rip), %%rdi\n\t" \ |
| 268 | ".word 0x6666\n\t" \ |
| 269 | "rex64\n\t" \ |
| 270 | "call __tls_get_addr@plt\n\t" \ |
| 271 | /* --- End TLS access */ \ |
| 272 | "addq $500,%%rax\n\t" \ |
| 273 | "movq %%rax,(%%rbx)\n\t" \ |
| 274 | "movq %%rbx,%%rax\n\t" \ |
| 275 | /* now the pointer to the private stack is in rax. |
| 276 | must add stack size so the ptr points to the stack bottom. */ \ |
| 277 | "1:\n\t" \ |
| 278 | /* Pop regs that were pushed for function calls */ \ |
| 279 | "popq %%r11\n\t" \ |
| 280 | "popq %%r10\n\t" \ |
| 281 | "popq %%r9\n\t" \ |
| 282 | "popq %%r8\n\t" \ |
| 283 | "popq %%rsi\n\t" \ |
| 284 | "popq %%rdx\n\t" \ |
| 285 | "popq %%rcx\n\t" \ |
| 286 | /* Manually push rsp to private stack */ \ |
| 287 | "addq $-8,(%%rax)\n\t" \ |
| 288 | "movq 32(%%rsp), %%rdi\n\t" \ |
| 289 | "movq (%%rax), %%rbx\n\t" \ |
| 290 | "movq %%rdi, (%%rbx)\n\t" \ |
| 291 | /* Manually push eflags to private stack */ \ |
| 292 | "addq $-8,(%%rax)\n\t" \ |
| 293 | "movq 24(%%rsp), %%rdi\n\t" \ |
| 294 | "movq (%%rax), %%rbx\n\t" \ |
| 295 | "movq %%rdi, (%%rbx)\n\t" \ |
| 296 | /* Manually push rax to private stack */ \ |
| 297 | "addq $-8,(%%rax)\n\t" \ |
| 298 | "movq 16(%%rsp), %%rdi\n\t" \ |
| 299 | "movq (%%rax), %%rbx\n\t" \ |
| 300 | "movq %%rdi, (%%rbx)\n\t" \ |
| 301 | /* Manually push rbx to private stack */ \ |
| 302 | "addq $-8,(%%rax)\n\t" \ |
| 303 | "movq 8(%%rsp), %%rdi\n\t" \ |
| 304 | "movq (%%rax), %%rbx\n\t" \ |
| 305 | "movq %%rdi, (%%rbx)\n\t" \ |
| 306 | /* Manually push rdi to private stack */ \ |
| 307 | "addq $-8,(%%rax)\n\t" \ |
| 308 | "movq 0(%%rsp), %%rdi\n\t" \ |
| 309 | "movq (%%rax), %%rbx\n\t" \ |
| 310 | "movq %%rdi, (%%rbx)\n\t" \ |
| 311 | /* now push regs to tls */ \ |
| 312 | /* -- rsp already pushed -- */ \ |
| 313 | /* -- rax already pushed -- */ \ |
| 314 | /* -- rbx already pushed -- */ \ |
| 315 | /* -- rdi already pushed -- */ \ |
| 316 | "addq $-8,(%%rax)\n\t" \ |
| 317 | "movq (%%rax), %%rbx\n\t" \ |
| 318 | "movq %%rcx,(%%rbx)\n\t" \ |
| 319 | "addq $-8,(%%rax)\n\t" \ |
| 320 | "movq (%%rax), %%rbx\n\t" \ |
| 321 | "movq %%rdx,(%%rbx)\n\t" \ |
| 322 | "addq $-8,(%%rax)\n\t" \ |
| 323 | "movq (%%rax), %%rbx\n\t" \ |
| 324 | "movq %%rbp,(%%rbx)\n\t" \ |
| 325 | "addq $-8,(%%rax)\n\t" \ |
| 326 | "movq (%%rax), %%rbx\n\t" \ |
| 327 | "movq %%rsi,(%%rbx)\n\t" \ |
| 328 | "addq $-8,(%%rax)\n\t" \ |
| 329 | "movq (%%rax), %%rbx\n\t" \ |
| 330 | "movq %%r8,(%%rbx)\n\t" \ |
| 331 | "addq $-8,(%%rax)\n\t" \ |
| 332 | "movq (%%rax), %%rbx\n\t" \ |
| 333 | "movq %%r9,(%%rbx)\n\t" \ |
| 334 | "addq $-8,(%%rax)\n\t" \ |
| 335 | "movq (%%rax), %%rbx\n\t" \ |
| 336 | "movq %%r10,(%%rbx)\n\t" \ |
| 337 | "addq $-8,(%%rax)\n\t" \ |
| 338 | "movq (%%rax), %%rbx\n\t" \ |
| 339 | "movq %%r11,(%%rbx)\n\t" \ |
| 340 | "addq $-8,(%%rax)\n\t" \ |
| 341 | "movq (%%rax), %%rbx\n\t" \ |
| 342 | "movq %%r12,(%%rbx)\n\t" \ |
| 343 | "addq $-8,(%%rax)\n\t" \ |
| 344 | "movq (%%rax), %%rbx\n\t" \ |
| 345 | "movq %%r13,(%%rbx)\n\t" \ |
| 346 | "addq $-8,(%%rax)\n\t" \ |
| 347 | "movq (%%rax), %%rbx\n\t" \ |
| 348 | "movq %%r14,(%%rbx)\n\t" \ |
| 349 | "addq $-8,(%%rax)\n\t" \ |
| 350 | "movq (%%rax), %%rbx\n\t" \ |
| 351 | "movq %%r15,(%%rbx)\n\t" \ |
| 352 | /* push cs */ \ |
| 353 | "addq $-2,(%%rax)\n\t" \ |
| 354 | "movq (%%rax), %%rbx\n\t" \ |
| 355 | "movw %%cs, (%%rbx)\n\t" \ |
| 356 | /* push ss */ \ |
| 357 | "addq $-2,(%%rax)\n\t" \ |
| 358 | "movq (%%rax), %%rbx\n\t" \ |
| 359 | "movw %%ss, (%%rbx)\n\t" \ |
| 360 | /* add padding for struct registers */ \ |
| 361 | "addq $-4,(%%rax)\n\t" \ |
| 362 | /* restore original values of regs that were used internally */ \ |
| 363 | "popq %%rdi\n\t" \ |
| 364 | "popq %%rbx\n\t" \ |
| 365 | "popq %%rax\n\t" \ |
| 366 | /* cancel push of rsp */ \ |
| 367 | "addq $8,%%rsp\n\t" \ |
| 368 | /* cancel push of rflags */ \ |
| 369 | "addq $8,%%rsp\n\t" \ |
| 370 | ::); \ |
| 371 | memcpy(regsptr, (void *)ust_reg_stack_ptr, sizeof(struct registers)); \ |
| 372 | ust_reg_stack_ptr = (void *)(((long)ust_reg_stack_ptr) + sizeof(struct registers)); |
| 373 | |
| 374 | #else /* CONFIG_UST_GDB_INTEGRATION */ |
| 375 | |
| 376 | #define save_registers(a) |
| 377 | |
| 378 | #endif /* CONFIG_UST_GDB_INTEGRATION */ |
| 379 | |
| 380 | /* Macro to insert the address of a relative jump in an assembly stub, |
| 381 | * in a relocatable way. On x86-64, this uses a special (%rip) notation. */ |
| 382 | #define RELATIVE_ADDRESS(__rel_label__) __rel_label__(%%rip) |
| 383 | |
| 384 | #define ARCH_COPY_ADDR(src, dst) "lea " src "(%%rip)," dst |
| 385 | |
| 386 | #define _ASM_PTR ".quad " |
| 387 | |
| 388 | #endif |
| 389 | |
| 390 | #endif /* UST_PROCESSOR_H */ |