| 1 | /* |
| 2 | Copyright 2011, 2012 Kristian Nielsen and Monty Program Ab |
| 3 | |
| 4 | This file is free software; you can redistribute it and/or |
| 5 | modify it under the terms of the GNU Lesser General Public |
| 6 | License as published by the Free Software Foundation; either |
| 7 | version 2.1 of the License, or (at your option) any later version. |
| 8 | |
| 9 | This library is distributed in the hope that it will be useful, |
| 10 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 11 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 12 | Lesser General Public License for more details. |
| 13 | |
| 14 | You should have received a copy of the GNU General Public License |
| 15 | along with this. If not, see <http://www.gnu.org/licenses/>. |
| 16 | */ |
| 17 | |
| 18 | /* |
| 19 | Implementation of async context spawning using Posix ucontext and |
| 20 | swapcontext(). |
| 21 | */ |
| 22 | |
| 23 | #include "mysys_priv.h" |
| 24 | #include "m_string.h" |
| 25 | #include "my_context.h" |
| 26 | |
| 27 | #ifdef HAVE_VALGRIND_MEMCHECK_H |
| 28 | #include <valgrind/valgrind.h> |
| 29 | #endif |
| 30 | |
| 31 | #ifdef MY_CONTEXT_USE_UCONTEXT |
| 32 | /* |
| 33 | The makecontext() only allows to pass integers into the created context :-( |
| 34 | We want to pass pointers, so we do it this kinda hackish way. |
| 35 | Anyway, it should work everywhere, and at least it does not break strict |
| 36 | aliasing. |
| 37 | */ |
| 38 | union pass_void_ptr_as_2_int { |
| 39 | int a[2]; |
| 40 | void *p; |
| 41 | }; |
| 42 | |
| 43 | |
| 44 | /* |
| 45 | We use old-style function definition here, as this is passed to |
| 46 | makecontext(). And the type of the makecontext() argument does not match |
| 47 | the actual type (as the actual type can differ from call to call). |
| 48 | */ |
| 49 | static void |
| 50 | my_context_spawn_internal(i0, i1) |
| 51 | int i0, i1; |
| 52 | { |
| 53 | int err; |
| 54 | struct my_context *c; |
| 55 | union pass_void_ptr_as_2_int u; |
| 56 | |
| 57 | u.a[0]= i0; |
| 58 | u.a[1]= i1; |
| 59 | c= (struct my_context *)u.p; |
| 60 | |
| 61 | (*c->user_func)(c->user_data); |
| 62 | c->active= 0; |
| 63 | err= setcontext(&c->base_context); |
| 64 | fprintf(stderr, "Aieie, setcontext() failed: %d (errno=%d)\n" , err, errno); |
| 65 | } |
| 66 | |
| 67 | |
| 68 | int |
| 69 | my_context_continue(struct my_context *c) |
| 70 | { |
| 71 | int err; |
| 72 | |
| 73 | if (!c->active) |
| 74 | return 0; |
| 75 | |
| 76 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 77 | err= swapcontext(&c->base_context, &c->spawned_context); |
| 78 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 79 | if (err) |
| 80 | { |
| 81 | fprintf(stderr, "Aieie, swapcontext() failed: %d (errno=%d)\n" , |
| 82 | err, errno); |
| 83 | return -1; |
| 84 | } |
| 85 | |
| 86 | return c->active; |
| 87 | } |
| 88 | |
| 89 | |
| 90 | int |
| 91 | my_context_spawn(struct my_context *c, void (*f)(void *), void *d) |
| 92 | { |
| 93 | int err; |
| 94 | union pass_void_ptr_as_2_int u; |
| 95 | |
| 96 | err= getcontext(&c->spawned_context); |
| 97 | if (err) |
| 98 | return -1; |
| 99 | c->spawned_context.uc_stack.ss_sp= c->stack; |
| 100 | c->spawned_context.uc_stack.ss_size= c->stack_size; |
| 101 | c->spawned_context.uc_link= NULL; |
| 102 | c->user_func= f; |
| 103 | c->user_data= d; |
| 104 | c->active= 1; |
| 105 | u.p= c; |
| 106 | makecontext(&c->spawned_context, my_context_spawn_internal, 2, |
| 107 | u.a[0], u.a[1]); |
| 108 | |
| 109 | return my_context_continue(c); |
| 110 | } |
| 111 | |
| 112 | |
| 113 | int |
| 114 | my_context_yield(struct my_context *c) |
| 115 | { |
| 116 | int err; |
| 117 | |
| 118 | if (!c->active) |
| 119 | return -1; |
| 120 | |
| 121 | err= swapcontext(&c->spawned_context, &c->base_context); |
| 122 | if (err) |
| 123 | return -1; |
| 124 | return 0; |
| 125 | } |
| 126 | |
| 127 | int |
| 128 | my_context_init(struct my_context *c, size_t stack_size) |
| 129 | { |
| 130 | #if SIZEOF_CHARP > SIZEOF_INT*2 |
| 131 | #error Error: Unable to store pointer in 2 ints on this architecture |
| 132 | #endif |
| 133 | bzero(c, sizeof(*c)); |
| 134 | if (!(c->stack= malloc(stack_size))) |
| 135 | return -1; /* Out of memory */ |
| 136 | c->stack_size= stack_size; |
| 137 | #ifdef HAVE_VALGRIND_MEMCHECK_H |
| 138 | c->valgrind_stack_id= |
| 139 | VALGRIND_STACK_REGISTER(c->stack, ((unsigned char *)(c->stack))+stack_size); |
| 140 | #endif |
| 141 | return 0; |
| 142 | } |
| 143 | |
| 144 | void |
| 145 | my_context_destroy(struct my_context *c) |
| 146 | { |
| 147 | if (c->stack) |
| 148 | { |
| 149 | #ifdef HAVE_VALGRIND_MEMCHECK_H |
| 150 | VALGRIND_STACK_DEREGISTER(c->valgrind_stack_id); |
| 151 | #endif |
| 152 | free(c->stack); |
| 153 | } |
| 154 | DBUG_FREE_CODE_STATE(&c->dbug_state); |
| 155 | } |
| 156 | |
| 157 | #endif /* MY_CONTEXT_USE_UCONTEXT */ |
| 158 | |
| 159 | |
| 160 | #ifdef MY_CONTEXT_USE_X86_64_GCC_ASM |
| 161 | /* |
| 162 | GCC-amd64 implementation of my_context. |
| 163 | |
| 164 | This is slightly optimized in the common case where we never yield |
| 165 | (eg. fetch next row and it is already fully received in buffer). In this |
| 166 | case we do not need to restore registers at return (though we still need to |
| 167 | save them as we cannot know if we will yield or not in advance). |
| 168 | */ |
| 169 | |
| 170 | #include <stdint.h> |
| 171 | #include <stdlib.h> |
| 172 | |
| 173 | /* |
| 174 | Layout of saved registers etc. |
| 175 | Since this is accessed through gcc inline assembler, it is simpler to just |
| 176 | use numbers than to try to define nice constants or structs. |
| 177 | |
| 178 | 0 0 %rsp |
| 179 | 1 8 %rbp |
| 180 | 2 16 %rbx |
| 181 | 3 24 %r12 |
| 182 | 4 32 %r13 |
| 183 | 5 40 %r14 |
| 184 | 6 48 %r15 |
| 185 | 7 56 %rip for done |
| 186 | 8 64 %rip for yield/continue |
| 187 | */ |
| 188 | |
| 189 | int |
| 190 | my_context_spawn(struct my_context *c, void (*f)(void *), void *d) |
| 191 | { |
| 192 | int ret; |
| 193 | |
| 194 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 195 | |
| 196 | /* |
| 197 | There are 6 callee-save registers we need to save and restore when |
| 198 | suspending and continuing, plus stack pointer %rsp and instruction pointer |
| 199 | %rip. |
| 200 | |
| 201 | However, if we never suspend, the user-supplied function will in any case |
| 202 | restore the 6 callee-save registers, so we can avoid restoring them in |
| 203 | this case. |
| 204 | */ |
| 205 | __asm__ __volatile__ |
| 206 | ( |
| 207 | "movq %%rsp, (%[save])\n\t" |
| 208 | "movq %[stack], %%rsp\n\t" |
| 209 | #if (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 4) || __clang__) && \ |
| 210 | !defined(__INTEL_COMPILER) |
| 211 | /* |
| 212 | This emits a DWARF DW_CFA_undefined directive to make the return address |
| 213 | undefined. This indicates that this is the top of the stack frame, and |
| 214 | helps tools that use DWARF stack unwinding to obtain stack traces. |
| 215 | (I use numeric constant to avoid a dependency on libdwarf includes). |
| 216 | */ |
| 217 | ".cfi_escape 0x07, 16\n\t" |
| 218 | #endif |
| 219 | "movq %%rbp, 8(%[save])\n\t" |
| 220 | "movq %%rbx, 16(%[save])\n\t" |
| 221 | "movq %%r12, 24(%[save])\n\t" |
| 222 | "movq %%r13, 32(%[save])\n\t" |
| 223 | "movq %%r14, 40(%[save])\n\t" |
| 224 | "movq %%r15, 48(%[save])\n\t" |
| 225 | "leaq 1f(%%rip), %%rax\n\t" |
| 226 | "leaq 2f(%%rip), %%rcx\n\t" |
| 227 | "movq %%rax, 56(%[save])\n\t" |
| 228 | "movq %%rcx, 64(%[save])\n\t" |
| 229 | /* |
| 230 | Constraint below puts the argument to the user function into %rdi, as |
| 231 | needed for the calling convention. |
| 232 | */ |
| 233 | "callq *%[f]\n\t" |
| 234 | "jmpq *56(%[save])\n" |
| 235 | /* |
| 236 | Come here when operation is done. |
| 237 | We do not need to restore callee-save registers, as the called function |
| 238 | will do this for us if needed. |
| 239 | */ |
| 240 | "1:\n\t" |
| 241 | "movq (%[save]), %%rsp\n\t" |
| 242 | "xorl %[ret], %[ret]\n\t" |
| 243 | "jmp 3f\n" |
| 244 | /* Come here when operation was suspended. */ |
| 245 | "2:\n\t" |
| 246 | "movl $1, %[ret]\n" |
| 247 | "3:\n" |
| 248 | : [ret] "=a" (ret), |
| 249 | [f] "+S" (f), |
| 250 | /* Need this in %rdi to follow calling convention. */ |
| 251 | [d] "+D" (d) |
| 252 | : [stack] "a" (c->stack_top), |
| 253 | /* Need this in callee-save register to preserve in function call. */ |
| 254 | [save] "b" (&c->save[0]) |
| 255 | : "rcx" , "rdx" , "r8" , "r9" , "r10" , "r11" , "memory" , "cc" |
| 256 | ); |
| 257 | |
| 258 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 259 | |
| 260 | return ret; |
| 261 | } |
| 262 | |
| 263 | int |
| 264 | my_context_continue(struct my_context *c) |
| 265 | { |
| 266 | int ret; |
| 267 | |
| 268 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 269 | |
| 270 | __asm__ __volatile__ |
| 271 | ( |
| 272 | "movq (%[save]), %%rax\n\t" |
| 273 | "movq %%rsp, (%[save])\n\t" |
| 274 | "movq %%rax, %%rsp\n\t" |
| 275 | "movq 8(%[save]), %%rax\n\t" |
| 276 | "movq %%rbp, 8(%[save])\n\t" |
| 277 | "movq %%rax, %%rbp\n\t" |
| 278 | "movq 24(%[save]), %%rax\n\t" |
| 279 | "movq %%r12, 24(%[save])\n\t" |
| 280 | "movq %%rax, %%r12\n\t" |
| 281 | "movq 32(%[save]), %%rax\n\t" |
| 282 | "movq %%r13, 32(%[save])\n\t" |
| 283 | "movq %%rax, %%r13\n\t" |
| 284 | "movq 40(%[save]), %%rax\n\t" |
| 285 | "movq %%r14, 40(%[save])\n\t" |
| 286 | "movq %%rax, %%r14\n\t" |
| 287 | "movq 48(%[save]), %%rax\n\t" |
| 288 | "movq %%r15, 48(%[save])\n\t" |
| 289 | "movq %%rax, %%r15\n\t" |
| 290 | |
| 291 | "leaq 1f(%%rip), %%rax\n\t" |
| 292 | "leaq 2f(%%rip), %%rcx\n\t" |
| 293 | "movq %%rax, 56(%[save])\n\t" |
| 294 | "movq 64(%[save]), %%rax\n\t" |
| 295 | "movq %%rcx, 64(%[save])\n\t" |
| 296 | |
| 297 | "movq 16(%[save]), %%rcx\n\t" |
| 298 | "movq %%rbx, 16(%[save])\n\t" |
| 299 | "movq %%rcx, %%rbx\n\t" |
| 300 | |
| 301 | "jmpq *%%rax\n" |
| 302 | /* |
| 303 | Come here when operation is done. |
| 304 | Be sure to use the same callee-save register for %[save] here and in |
| 305 | my_context_spawn(), so we preserve the value correctly at this point. |
| 306 | */ |
| 307 | "1:\n\t" |
| 308 | "movq (%[save]), %%rsp\n\t" |
| 309 | "movq 8(%[save]), %%rbp\n\t" |
| 310 | /* %rbx is preserved from my_context_spawn() in this case. */ |
| 311 | "movq 24(%[save]), %%r12\n\t" |
| 312 | "movq 32(%[save]), %%r13\n\t" |
| 313 | "movq 40(%[save]), %%r14\n\t" |
| 314 | "movq 48(%[save]), %%r15\n\t" |
| 315 | "xorl %[ret], %[ret]\n\t" |
| 316 | "jmp 3f\n" |
| 317 | /* Come here when operation is suspended. */ |
| 318 | "2:\n\t" |
| 319 | "movl $1, %[ret]\n" |
| 320 | "3:\n" |
| 321 | : [ret] "=a" (ret) |
| 322 | : /* Need this in callee-save register to preserve in function call. */ |
| 323 | [save] "b" (&c->save[0]) |
| 324 | : "rcx" , "rdx" , "rsi" , "rdi" , "r8" , "r9" , "r10" , "r11" , "memory" , "cc" |
| 325 | ); |
| 326 | |
| 327 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 328 | |
| 329 | return ret; |
| 330 | } |
| 331 | |
| 332 | int |
| 333 | my_context_yield(struct my_context *c) |
| 334 | { |
| 335 | uint64_t *save= &c->save[0]; |
| 336 | __asm__ __volatile__ |
| 337 | ( |
| 338 | "movq (%[save]), %%rax\n\t" |
| 339 | "movq %%rsp, (%[save])\n\t" |
| 340 | "movq %%rax, %%rsp\n\t" |
| 341 | "movq 8(%[save]), %%rax\n\t" |
| 342 | "movq %%rbp, 8(%[save])\n\t" |
| 343 | "movq %%rax, %%rbp\n\t" |
| 344 | "movq 16(%[save]), %%rax\n\t" |
| 345 | "movq %%rbx, 16(%[save])\n\t" |
| 346 | "movq %%rax, %%rbx\n\t" |
| 347 | "movq 24(%[save]), %%rax\n\t" |
| 348 | "movq %%r12, 24(%[save])\n\t" |
| 349 | "movq %%rax, %%r12\n\t" |
| 350 | "movq 32(%[save]), %%rax\n\t" |
| 351 | "movq %%r13, 32(%[save])\n\t" |
| 352 | "movq %%rax, %%r13\n\t" |
| 353 | "movq 40(%[save]), %%rax\n\t" |
| 354 | "movq %%r14, 40(%[save])\n\t" |
| 355 | "movq %%rax, %%r14\n\t" |
| 356 | "movq 48(%[save]), %%rax\n\t" |
| 357 | "movq %%r15, 48(%[save])\n\t" |
| 358 | "movq %%rax, %%r15\n\t" |
| 359 | "movq 64(%[save]), %%rax\n\t" |
| 360 | "leaq 1f(%%rip), %%rcx\n\t" |
| 361 | "movq %%rcx, 64(%[save])\n\t" |
| 362 | |
| 363 | "jmpq *%%rax\n" |
| 364 | |
| 365 | "1:\n" |
| 366 | : [save] "+D" (save) |
| 367 | : |
| 368 | : "rax" , "rcx" , "rdx" , "rsi" , "r8" , "r9" , "r10" , "r11" , "memory" , "cc" |
| 369 | ); |
| 370 | return 0; |
| 371 | } |
| 372 | |
| 373 | int |
| 374 | my_context_init(struct my_context *c, size_t stack_size) |
| 375 | { |
| 376 | bzero(c, sizeof(*c)); |
| 377 | |
| 378 | if (!(c->stack_bot= malloc(stack_size))) |
| 379 | return -1; /* Out of memory */ |
| 380 | /* |
| 381 | The x86_64 ABI specifies 16-byte stack alignment. |
| 382 | Also put two zero words at the top of the stack. |
| 383 | */ |
| 384 | c->stack_top= (void *) |
| 385 | (( ((intptr)c->stack_bot + stack_size) & ~(intptr)0xf) - 16); |
| 386 | bzero(c->stack_top, 16); |
| 387 | |
| 388 | #ifdef HAVE_VALGRIND_MEMCHECK_H |
| 389 | c->valgrind_stack_id= |
| 390 | VALGRIND_STACK_REGISTER(c->stack_bot, c->stack_top); |
| 391 | #endif |
| 392 | return 0; |
| 393 | } |
| 394 | |
| 395 | void |
| 396 | my_context_destroy(struct my_context *c) |
| 397 | { |
| 398 | if (c->stack_bot) |
| 399 | { |
| 400 | free(c->stack_bot); |
| 401 | #ifdef HAVE_VALGRIND_MEMCHECK_H |
| 402 | VALGRIND_STACK_DEREGISTER(c->valgrind_stack_id); |
| 403 | #endif |
| 404 | } |
| 405 | DBUG_FREE_CODE_STATE(&c->dbug_state); |
| 406 | } |
| 407 | |
| 408 | #endif /* MY_CONTEXT_USE_X86_64_GCC_ASM */ |
| 409 | |
| 410 | |
| 411 | #ifdef MY_CONTEXT_USE_I386_GCC_ASM |
| 412 | /* |
| 413 | GCC-i386 implementation of my_context. |
| 414 | |
| 415 | This is slightly optimized in the common case where we never yield |
| 416 | (eg. fetch next row and it is already fully received in buffer). In this |
| 417 | case we do not need to restore registers at return (though we still need to |
| 418 | save them as we cannot know if we will yield or not in advance). |
| 419 | */ |
| 420 | |
| 421 | #include <stdint.h> |
| 422 | #include <stdlib.h> |
| 423 | |
| 424 | /* |
| 425 | Layout of saved registers etc. |
| 426 | Since this is accessed through gcc inline assembler, it is simpler to just |
| 427 | use numbers than to try to define nice constants or structs. |
| 428 | |
| 429 | 0 0 %esp |
| 430 | 1 4 %ebp |
| 431 | 2 8 %ebx |
| 432 | 3 12 %esi |
| 433 | 4 16 %edi |
| 434 | 5 20 %eip for done |
| 435 | 6 24 %eip for yield/continue |
| 436 | */ |
| 437 | |
| 438 | int |
| 439 | my_context_spawn(struct my_context *c, void (*f)(void *), void *d) |
| 440 | { |
| 441 | int ret; |
| 442 | |
| 443 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 444 | |
| 445 | /* |
| 446 | There are 4 callee-save registers we need to save and restore when |
| 447 | suspending and continuing, plus stack pointer %esp and instruction pointer |
| 448 | %eip. |
| 449 | |
| 450 | However, if we never suspend, the user-supplied function will in any case |
| 451 | restore the 4 callee-save registers, so we can avoid restoring them in |
| 452 | this case. |
| 453 | */ |
| 454 | __asm__ __volatile__ |
| 455 | ( |
| 456 | "movl %%esp, (%[save])\n\t" |
| 457 | "movl %[stack], %%esp\n\t" |
| 458 | #if (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 4)) && !defined(__INTEL_COMPILER) |
| 459 | /* |
| 460 | This emits a DWARF DW_CFA_undefined directive to make the return address |
| 461 | undefined. This indicates that this is the top of the stack frame, and |
| 462 | helps tools that use DWARF stack unwinding to obtain stack traces. |
| 463 | (I use numeric constant to avoid a dependency on libdwarf includes). |
| 464 | */ |
| 465 | ".cfi_escape 0x07, 8\n\t" |
| 466 | #endif |
| 467 | /* Push the parameter on the stack. */ |
| 468 | "pushl %[d]\n\t" |
| 469 | "movl %%ebp, 4(%[save])\n\t" |
| 470 | "movl %%ebx, 8(%[save])\n\t" |
| 471 | "movl %%esi, 12(%[save])\n\t" |
| 472 | "movl %%edi, 16(%[save])\n\t" |
| 473 | /* Get label addresses in -fPIC-compatible way (no pc-relative on 32bit) */ |
| 474 | "call 1f\n" |
| 475 | "1:\n\t" |
| 476 | "popl %%eax\n\t" |
| 477 | "addl $(2f-1b), %%eax\n\t" |
| 478 | "movl %%eax, 20(%[save])\n\t" |
| 479 | "addl $(3f-2f), %%eax\n\t" |
| 480 | "movl %%eax, 24(%[save])\n\t" |
| 481 | "call *%[f]\n\t" |
| 482 | "jmp *20(%[save])\n" |
| 483 | /* |
| 484 | Come here when operation is done. |
| 485 | We do not need to restore callee-save registers, as the called function |
| 486 | will do this for us if needed. |
| 487 | */ |
| 488 | "2:\n\t" |
| 489 | "movl (%[save]), %%esp\n\t" |
| 490 | "xorl %[ret], %[ret]\n\t" |
| 491 | "jmp 4f\n" |
| 492 | /* Come here when operation was suspended. */ |
| 493 | "3:\n\t" |
| 494 | "movl $1, %[ret]\n" |
| 495 | "4:\n" |
| 496 | : [ret] "=a" (ret), |
| 497 | [f] "+c" (f), |
| 498 | [d] "+d" (d) |
| 499 | : [stack] "a" (c->stack_top), |
| 500 | /* Need this in callee-save register to preserve across function call. */ |
| 501 | [save] "D" (&c->save[0]) |
| 502 | : "memory" , "cc" |
| 503 | ); |
| 504 | |
| 505 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 506 | |
| 507 | return ret; |
| 508 | } |
| 509 | |
| 510 | int |
| 511 | my_context_continue(struct my_context *c) |
| 512 | { |
| 513 | int ret; |
| 514 | |
| 515 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 516 | |
| 517 | __asm__ __volatile__ |
| 518 | ( |
| 519 | "movl (%[save]), %%eax\n\t" |
| 520 | "movl %%esp, (%[save])\n\t" |
| 521 | "movl %%eax, %%esp\n\t" |
| 522 | "movl 4(%[save]), %%eax\n\t" |
| 523 | "movl %%ebp, 4(%[save])\n\t" |
| 524 | "movl %%eax, %%ebp\n\t" |
| 525 | "movl 8(%[save]), %%eax\n\t" |
| 526 | "movl %%ebx, 8(%[save])\n\t" |
| 527 | "movl %%eax, %%ebx\n\t" |
| 528 | "movl 12(%[save]), %%eax\n\t" |
| 529 | "movl %%esi, 12(%[save])\n\t" |
| 530 | "movl %%eax, %%esi\n\t" |
| 531 | |
| 532 | "movl 24(%[save]), %%eax\n\t" |
| 533 | "call 1f\n" |
| 534 | "1:\n\t" |
| 535 | "popl %%ecx\n\t" |
| 536 | "addl $(2f-1b), %%ecx\n\t" |
| 537 | "movl %%ecx, 20(%[save])\n\t" |
| 538 | "addl $(3f-2f), %%ecx\n\t" |
| 539 | "movl %%ecx, 24(%[save])\n\t" |
| 540 | |
| 541 | /* Must restore %edi last as it is also our %[save] register. */ |
| 542 | "movl 16(%[save]), %%ecx\n\t" |
| 543 | "movl %%edi, 16(%[save])\n\t" |
| 544 | "movl %%ecx, %%edi\n\t" |
| 545 | |
| 546 | "jmp *%%eax\n" |
| 547 | /* |
| 548 | Come here when operation is done. |
| 549 | Be sure to use the same callee-save register for %[save] here and in |
| 550 | my_context_spawn(), so we preserve the value correctly at this point. |
| 551 | */ |
| 552 | "2:\n\t" |
| 553 | "movl (%[save]), %%esp\n\t" |
| 554 | "movl 4(%[save]), %%ebp\n\t" |
| 555 | "movl 8(%[save]), %%ebx\n\t" |
| 556 | "movl 12(%[save]), %%esi\n\t" |
| 557 | "movl 16(%[save]), %%edi\n\t" |
| 558 | "xorl %[ret], %[ret]\n\t" |
| 559 | "jmp 4f\n" |
| 560 | /* Come here when operation is suspended. */ |
| 561 | "3:\n\t" |
| 562 | "movl $1, %[ret]\n" |
| 563 | "4:\n" |
| 564 | : [ret] "=a" (ret) |
| 565 | : /* Need this in callee-save register to preserve in function call. */ |
| 566 | [save] "D" (&c->save[0]) |
| 567 | : "ecx" , "edx" , "memory" , "cc" |
| 568 | ); |
| 569 | |
| 570 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 571 | |
| 572 | return ret; |
| 573 | } |
| 574 | |
| 575 | int |
| 576 | my_context_yield(struct my_context *c) |
| 577 | { |
| 578 | uint64_t *save= &c->save[0]; |
| 579 | __asm__ __volatile__ |
| 580 | ( |
| 581 | "movl (%[save]), %%eax\n\t" |
| 582 | "movl %%esp, (%[save])\n\t" |
| 583 | "movl %%eax, %%esp\n\t" |
| 584 | "movl 4(%[save]), %%eax\n\t" |
| 585 | "movl %%ebp, 4(%[save])\n\t" |
| 586 | "movl %%eax, %%ebp\n\t" |
| 587 | "movl 8(%[save]), %%eax\n\t" |
| 588 | "movl %%ebx, 8(%[save])\n\t" |
| 589 | "movl %%eax, %%ebx\n\t" |
| 590 | "movl 12(%[save]), %%eax\n\t" |
| 591 | "movl %%esi, 12(%[save])\n\t" |
| 592 | "movl %%eax, %%esi\n\t" |
| 593 | "movl 16(%[save]), %%eax\n\t" |
| 594 | "movl %%edi, 16(%[save])\n\t" |
| 595 | "movl %%eax, %%edi\n\t" |
| 596 | |
| 597 | "movl 24(%[save]), %%eax\n\t" |
| 598 | "call 1f\n" |
| 599 | "1:\n\t" |
| 600 | "popl %%ecx\n\t" |
| 601 | "addl $(2f-1b), %%ecx\n\t" |
| 602 | "movl %%ecx, 24(%[save])\n\t" |
| 603 | |
| 604 | "jmp *%%eax\n" |
| 605 | |
| 606 | "2:\n" |
| 607 | : [save] "+d" (save) |
| 608 | : |
| 609 | : "eax" , "ecx" , "memory" , "cc" |
| 610 | ); |
| 611 | return 0; |
| 612 | } |
| 613 | |
| 614 | int |
| 615 | my_context_init(struct my_context *c, size_t stack_size) |
| 616 | { |
| 617 | bzero(c, sizeof(*c)); |
| 618 | if (!(c->stack_bot= malloc(stack_size))) |
| 619 | return -1; /* Out of memory */ |
| 620 | c->stack_top= (void *) |
| 621 | (( ((intptr)c->stack_bot + stack_size) & ~(intptr)0xf) - 16); |
| 622 | bzero(c->stack_top, 16); |
| 623 | |
| 624 | #ifdef HAVE_VALGRIND_MEMCHECK_H |
| 625 | c->valgrind_stack_id= |
| 626 | VALGRIND_STACK_REGISTER(c->stack_bot, c->stack_top); |
| 627 | #endif |
| 628 | return 0; |
| 629 | } |
| 630 | |
| 631 | void |
| 632 | my_context_destroy(struct my_context *c) |
| 633 | { |
| 634 | if (c->stack_bot) |
| 635 | { |
| 636 | free(c->stack_bot); |
| 637 | #ifdef HAVE_VALGRIND_MEMCHECK_H |
| 638 | VALGRIND_STACK_DEREGISTER(c->valgrind_stack_id); |
| 639 | #endif |
| 640 | } |
| 641 | DBUG_FREE_CODE_STATE(&c->dbug_state); |
| 642 | } |
| 643 | |
| 644 | #endif /* MY_CONTEXT_USE_I386_GCC_ASM */ |
| 645 | |
| 646 | |
| 647 | #ifdef MY_CONTEXT_USE_WIN32_FIBERS |
| 648 | int |
| 649 | my_context_yield(struct my_context *c) |
| 650 | { |
| 651 | c->return_value= 1; |
| 652 | SwitchToFiber(c->app_fiber); |
| 653 | return 0; |
| 654 | } |
| 655 | |
| 656 | |
| 657 | static void WINAPI |
| 658 | my_context_trampoline(void *p) |
| 659 | { |
| 660 | struct my_context *c= (struct my_context *)p; |
| 661 | /* |
| 662 | Reuse the Fiber by looping infinitely, each time we are scheduled we |
| 663 | spawn the appropriate function and switch back when it is done. |
| 664 | |
| 665 | This way we avoid the overhead of CreateFiber() for every asynchroneous |
| 666 | operation. |
| 667 | */ |
| 668 | for(;;) |
| 669 | { |
| 670 | (*(c->user_func))(c->user_arg); |
| 671 | c->return_value= 0; |
| 672 | SwitchToFiber(c->app_fiber); |
| 673 | } |
| 674 | } |
| 675 | |
| 676 | int |
| 677 | my_context_init(struct my_context *c, size_t stack_size) |
| 678 | { |
| 679 | bzero(c, sizeof(*c)); |
| 680 | c->lib_fiber= CreateFiber(stack_size, my_context_trampoline, c); |
| 681 | if (c->lib_fiber) |
| 682 | return 0; |
| 683 | return -1; |
| 684 | } |
| 685 | |
| 686 | void |
| 687 | my_context_destroy(struct my_context *c) |
| 688 | { |
| 689 | DBUG_FREE_CODE_STATE(&c->dbug_state); |
| 690 | if (c->lib_fiber) |
| 691 | { |
| 692 | DeleteFiber(c->lib_fiber); |
| 693 | c->lib_fiber= NULL; |
| 694 | } |
| 695 | } |
| 696 | |
| 697 | int |
| 698 | my_context_spawn(struct my_context *c, void (*f)(void *), void *d) |
| 699 | { |
| 700 | c->user_func= f; |
| 701 | c->user_arg= d; |
| 702 | return my_context_continue(c); |
| 703 | } |
| 704 | |
| 705 | int |
| 706 | my_context_continue(struct my_context *c) |
| 707 | { |
| 708 | /* |
| 709 | This seems to be a common trick to run ConvertThreadToFiber() only on the |
| 710 | first occurrence in a thread, in a way that works on multiple Windows |
| 711 | versions. |
| 712 | */ |
| 713 | void *current_fiber= GetCurrentFiber(); |
| 714 | if (current_fiber == NULL || current_fiber == (void *)0x1e00) |
| 715 | current_fiber= ConvertThreadToFiber(c); |
| 716 | c->app_fiber= current_fiber; |
| 717 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 718 | SwitchToFiber(c->lib_fiber); |
| 719 | DBUG_SWAP_CODE_STATE(&c->dbug_state); |
| 720 | |
| 721 | return c->return_value; |
| 722 | } |
| 723 | |
| 724 | #endif /* MY_CONTEXT_USE_WIN32_FIBERS */ |
| 725 | |
| 726 | #ifdef MY_CONTEXT_DISABLE |
| 727 | int |
| 728 | my_context_continue(struct my_context *c __attribute__((unused))) |
| 729 | { |
| 730 | return -1; |
| 731 | } |
| 732 | |
| 733 | |
| 734 | int |
| 735 | my_context_spawn(struct my_context *c __attribute__((unused)), |
| 736 | void (*f)(void *) __attribute__((unused)), |
| 737 | void *d __attribute__((unused))) |
| 738 | { |
| 739 | return -1; |
| 740 | } |
| 741 | |
| 742 | |
| 743 | int |
| 744 | my_context_yield(struct my_context *c __attribute__((unused))) |
| 745 | { |
| 746 | return -1; |
| 747 | } |
| 748 | |
| 749 | int |
| 750 | my_context_init(struct my_context *c __attribute__((unused)), |
| 751 | size_t stack_size __attribute__((unused))) |
| 752 | { |
| 753 | return -1; /* Out of memory */ |
| 754 | } |
| 755 | |
| 756 | void |
| 757 | my_context_destroy(struct my_context *c __attribute__((unused))) |
| 758 | { |
| 759 | } |
| 760 | |
| 761 | #endif |
| 762 | |