The material available from web, mail-list, books like Mac OS X Internals, and even source code is quite limited.
Now I know that xnu kernel raise an EXC_CRASH,
Every Mach thread and/or task (the underlying kernel object on top of which the BSD layer process is implemented) has exception ports. three port levels are available: Thread, Task, and host. When an exception occurs, a Mach message is sent - first to the thread port, then - if none caught it - the task'S, and finally the host. If you obtain the port, you can catch the exception, debug it (as does gdb on OS X) or generate a crash dump (as does Crash Reporter). Specifically, launchd - the parent of all OS X system tasks - registers their exception ports, so it gets the messages, and then triggers CrashReporter (as you can see from the launchd's ReportCrash plist:
<key>MachServices</key>
<dict>
<key>com.apple.ReportCrash.DirectoryService</key>
<dict>
<key>DrainMessagesOnCrash</key>
<string>All</string>
<key>ExceptionServer</key>
<dict/>
</dict>
</dict>
The XNU kernel code is responsible for sending the message on EXC_CRASH. Specifically, proc_prepareexit does that:
/* If a core should be generated, notify crash reporter */
if (hassigprop(WTERMSIG(rv), SA_CORE) || ((p->p_csflags & CS_KILLED) != 0)) {
/*
* Workaround for processes checking up on PT_DENY_ATTACH:
* should be backed out post-Leopard (details in 5431025).
*/
if ((SIGSEGV == WTERMSIG(rv)) &&
(p->p_pptr->p_lflag & P_LNOATTACH)) {
goto skipcheck;
}
/*
* Crash Reporter looks for the signal value, original exception
* type, and low 20 bits of the original code in code[0]
* (8, 4, and 20 bits respectively). code[1] is unmodified.
*/
code = ((WTERMSIG(rv) & 0xff) << 24) |
((ut->uu_exception & 0x0f) << 20) |
((int)ut->uu_code & 0xfffff);
subcode = ut->uu_subcode;
(void) task_exception_notify(EXC_CRASH, code, subcode); // <-- Sends the msg
}