rempas 修订了这个 Gist . 跳至此修订
1 file changed, 146 insertions
stack_head_clone3_waitid.c(文件已创建)
| @@ -0,0 +1,146 @@ | |||
| 1 | + | // libc-free x86-64 Linux multi-threading example | |
| 2 | + | // $ cc -nostdlib stack_head.c | |
| 3 | + | // Ref: https://nullprogram.com/blog/2023/03/23/ | |
| 4 | + | // This is free and unencumbered software released into the public domain. | |
| 5 | + | // | |
| 6 | + | #include "clone_args.h" | |
| 7 | + | ||
| 8 | + | #define SYS_write 1 | |
| 9 | + | #define SYS_mmap 9 | |
| 10 | + | #define SYS_nanosleep 35 | |
| 11 | + | #define SYS_clone 56 | |
| 12 | + | #define SYS_exit 60 | |
| 13 | + | #define SYS_waitid 247 | |
| 14 | + | #define SYS_exit_group 231 | |
| 15 | + | ||
| 16 | + | #define SYSCALL1(n, a) \ | |
| 17 | + | syscall6(n,(long)(a),0,0,0,0,0) | |
| 18 | + | #define SYSCALL2(n, a, b) \ | |
| 19 | + | syscall6(n,(long)(a),(long)(b),0,0,0,0) | |
| 20 | + | #define SYSCALL3(n, a, b, c) \ | |
| 21 | + | syscall6(n,(long)(a),(long)(b),(long)(c),0,0,0) | |
| 22 | + | #define SYSCALL4(n, a, b, c, d) \ | |
| 23 | + | syscall6(n,(long)(a),(long)(b),(long)(c),(long)(d),0,0) | |
| 24 | + | #define SYSCALL5(n, a, b, c, d, e) \ | |
| 25 | + | syscall6(n,(long)(a),(long)(b),(long)(c),(long)(d),(long)(e),0) | |
| 26 | + | #define SYSCALL6(n, a, b, c, d, e, f) \ | |
| 27 | + | syscall6(n,(long)(a),(long)(b),(long)(c),(long)(d),(long)(e),(long)(f)) | |
| 28 | + | ||
| 29 | + | static long syscall6(long n, long a, long b, long c, long d, long e, long f) { | |
| 30 | + | register long ret; | |
| 31 | + | register long r10 asm("r10") = d; | |
| 32 | + | register long r8 asm("r8") = e; | |
| 33 | + | register long r9 asm("r9") = f; | |
| 34 | + | __asm volatile ( | |
| 35 | + | "syscall" | |
| 36 | + | : "=a"(ret) | |
| 37 | + | : "a"(n), "D"(a), "S"(b), "d"(c), "r"(r10), "r"(r8), "r"(r9) | |
| 38 | + | : "rcx", "r11", "memory" | |
| 39 | + | ); | |
| 40 | + | return ret; | |
| 41 | + | } | |
| 42 | + | ||
| 43 | + | static void millisleep(int ms) { | |
| 44 | + | long ts[] = {ms/1000, ms%1000 * 1000000L}; | |
| 45 | + | SYSCALL2(SYS_nanosleep, ts, ts); | |
| 46 | + | } | |
| 47 | + | ||
| 48 | + | /* int num = 65; */ // Ignore this, used for my own personal test | |
| 49 | + | ||
| 50 | + | static long fullwrite(int fd, void *buf, long len) { | |
| 51 | + | for (long off = 0; off < len;) { | |
| 52 | + | long r = SYSCALL3(SYS_write, fd, buf+off, len-off); | |
| 53 | + | if (r < 0) { return r; } | |
| 54 | + | off += r; | |
| 55 | + | } | |
| 56 | + | ||
| 57 | + | return len; | |
| 58 | + | } | |
| 59 | + | ||
| 60 | + | __attribute((noreturn)) static void exit(int status) { | |
| 61 | + | SYSCALL1(SYS_exit, status); | |
| 62 | + | __builtin_unreachable(); | |
| 63 | + | } | |
| 64 | + | ||
| 65 | + | __attribute((noreturn)) static void exit_group(int status) { | |
| 66 | + | SYSCALL1(SYS_exit_group, status); | |
| 67 | + | __builtin_unreachable(); | |
| 68 | + | } | |
| 69 | + | ||
| 70 | + | long sys_waitid() { | |
| 71 | + | struct siginfo s; | |
| 72 | + | return SYSCALL4(SYS_waitid, 0 /* P_ALL */, 0, &s, 0x40000000 /* __WALL */); | |
| 73 | + | } | |
| 74 | + | ||
| 75 | + | /* The structure must have a 16-byte alignment on all architectures */ | |
| 76 | + | struct __attribute((aligned(16))) stack_head { | |
| 77 | + | void (*entry)(struct stack_head*); // The entry point pointer. Will receive a pointer to its own stack_head | |
| 78 | + | ||
| 79 | + | /* The rest of the arguments can be filled with any thread-local | |
| 80 | + | data we want. The following is just an example! */ | |
| 81 | + | char *message; | |
| 82 | + | long message_length; | |
| 83 | + | int print_count; | |
| 84 | + | }; | |
| 85 | + | ||
| 86 | + | typedef struct stack_head stack_head; | |
| 87 | + | ||
| 88 | + | __attribute((naked)) static long newthread(CloneArgs* args) { | |
| 89 | + | __asm volatile ( | |
| 90 | + | /* "args" is already in "rdi" */ | |
| 91 | + | "mov $88, %%esi\n" // arg2 = size (always "88" until further notice) | |
| 92 | + | "mov $435, %%eax\n" // SYS_clone3 | |
| 93 | + | "syscall\n" | |
| 94 | + | "mov %%rsp, %%rdi\n" // entry point argument | |
| 95 | + | "ret\n" | |
| 96 | + | : : : "rax", "rcx", "rsi", "rdi", "r11", "memory" | |
| 97 | + | ); | |
| 98 | + | } | |
| 99 | + | ||
| 100 | + | static void threadentry(stack_head *stack) { | |
| 101 | + | char *message = stack->message; | |
| 102 | + | int length = stack->message_length; | |
| 103 | + | int count = stack->print_count; | |
| 104 | + | for (int i = 0; i < count; i++) { | |
| 105 | + | fullwrite(1, message, length); | |
| 106 | + | millisleep(25); | |
| 107 | + | } | |
| 108 | + | ||
| 109 | + | exit(0); | |
| 110 | + | } | |
| 111 | + | ||
| 112 | + | static stack_head *newstack(long size) { | |
| 113 | + | unsigned long p = SYSCALL6(SYS_mmap, 0, size, 3, 0x22, -1, 0); | |
| 114 | + | if (p > -4096UL) { return 0; } | |
| 115 | + | ||
| 116 | + | long count = size / sizeof(stack_head); | |
| 117 | + | return (stack_head*)p + count - 1; | |
| 118 | + | } | |
| 119 | + | ||
| 120 | + | __attribute((force_align_arg_pointer)) void _start(void) { | |
| 121 | + | stack_head* stack = newstack(1<<16); | |
| 122 | + | stack->entry = threadentry; | |
| 123 | + | ||
| 124 | + | // Thread data | |
| 125 | + | stack->message = "hello world\n"; | |
| 126 | + | stack->message_length = 12; | |
| 127 | + | stack->print_count = 20; | |
| 128 | + | ||
| 129 | + | CloneArgs args = { | |
| 130 | + | CLONE_VM, | |
| 131 | + | 0, 0, 0, SIGCHLD /* I have also tried using "0" here */, | |
| 132 | + | (__aligned_u64)stack, /* stack */ | |
| 133 | + | 1 << 16, /* stack_size */ | |
| 134 | + | 0, 0, 0, 0 | |
| 135 | + | }; | |
| 136 | + | ||
| 137 | + | newthread(&args); | |
| 138 | + | ||
| 139 | + | /* Fails */ | |
| 140 | + | if (sys_waitid() != 0) { exit_group(10); }; | |
| 141 | + | ||
| 142 | + | // Try to sleep and it indeed sleeps but the thread doesn't write the message | |
| 143 | + | /* millisleep(2000); */ | |
| 144 | + | ||
| 145 | + | exit_group(0); | |
| 146 | + | } | |
更新
更早