1.0.3.16: experimental x86-64/darwin suport
[sbcl.git] / src / runtime / x86-64-darwin-os.c
1
2 #ifdef LISP_FEATURE_SB_THREAD
3 #include <architecture/i386/table.h>
4 #include <i386/user_ldt.h>
5 #include <mach/mach_init.h>
6 #endif
7
8 #include "thread.h"
9 #include "validate.h"
10 #include "runtime.h"
11 #include "interrupt.h"
12 #include "x86-64-darwin-os.h"
13 #include "genesis/fdefn.h"
14
15 #include <mach/mach.h>
16 #include <mach/mach_error.h>
17 #include <mach/mach_types.h>
18 #include <mach/sync_policy.h>
19 #include <mach/machine/thread_state.h>
20 #include <mach/machine/thread_status.h>
21 #include <sys/_types.h>
22 #include <sys/ucontext.h>
23 #include <pthread.h>
24 #include <assert.h>
25 #include <stdlib.h>
26 #include <stdio.h>
27
28 #ifdef LISP_FEATURE_SB_THREAD
29 pthread_mutex_t mach_exception_lock = PTHREAD_MUTEX_INITIALIZER;
30 #endif
31
32 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
33
34 kern_return_t mach_thread_init(mach_port_t thread_exception_port);
35
36 void sigill_handler(int signal, siginfo_t *siginfo, void *void_context);
37 void sigtrap_handler(int signal, siginfo_t *siginfo, void *void_context);
38 void memory_fault_handler(int signal, siginfo_t *siginfo, void *void_context);
39
40 /* exc_server handles mach exception messages from the kernel and
41  * calls catch exception raise. We use the system-provided
42  * mach_msg_server, which, I assume, calls exc_server in a loop.
43  *
44  */
45 extern boolean_t exc_server();
46
47 /* This executes in the faulting thread as part of the signal
48  * emulation.  It is passed a context with the uc_mcontext field
49  * pointing to a valid block of memory. */
50 void build_fake_signal_context(struct ucontext *context,
51                                x86_thread_state64_t *thread_state,
52                                x86_float_state64_t *float_state) {
53     pthread_sigmask(0, NULL, &context->uc_sigmask);
54     context->uc_mcontext->ss = *thread_state;
55     context->uc_mcontext->fs = *float_state;
56 }
57
58 /* This executes in the faulting thread as part of the signal
59  * emulation.  It is effectively the inverse operation from above. */
60 void update_thread_state_from_context(x86_thread_state64_t *thread_state,
61                                       x86_float_state64_t *float_state,
62                                       struct ucontext *context) {
63     *thread_state = context->uc_mcontext->ss;
64     *float_state = context->uc_mcontext->fs;
65     pthread_sigmask(SIG_SETMASK, &context->uc_sigmask, NULL);
66 }
67
68 /* Modify a context to push new data on its stack. */
69 void push_context(u64 data, x86_thread_state64_t *context)
70 {
71     u64 *stack_pointer;
72
73     stack_pointer = (u64*) context->rsp;
74     *(--stack_pointer) = data;
75     context->rsp = (u64) stack_pointer;
76 }
77
78 void align_context_stack(x86_thread_state64_t *context)
79 {
80     /* 16byte align the stack (provided that the stack is, as it
81      * should be, 8byte aligned. */
82     while (context->rsp & 15) push_context(0, context);
83 }
84
85 /* Stack allocation starts with a context that has a mod-4 ESP value
86  * and needs to leave a context with a mod-16 ESP that will restore
87  * the old ESP value and other register state when activated.  The
88  * first part of this is the recovery trampoline, which loads ESP from
89  * EBP, pops EBP, and returns. */
90 asm(".globl _stack_allocation_recover; .align 4; _stack_allocation_recover: mov %rbp, %rsp; pop %rsi; pop %rdi; pop \
91 %rdx; pop %rcx; pop %r8; pop %r9; pop %rbp; ret;");
92
93 void open_stack_allocation(x86_thread_state64_t *context)
94 {
95     void stack_allocation_recover(void);
96
97     push_context(context->rip, context);
98     push_context(context->rbp, context);
99
100     push_context(context->r9, context);
101     push_context(context->r8, context);
102     push_context(context->rcx, context);
103     push_context(context->rdx, context);
104     push_context(context->rsi, context);
105     push_context(context->rdi, context);
106
107     context->rbp = context->rsp;
108     context->rip = (u64) stack_allocation_recover;
109
110     align_context_stack(context);
111 }
112
113 /* Stack allocation of data starts with a context with a mod-16 ESP
114  * value and reserves some space on it by manipulating the ESP
115  * register. */
116 void *stack_allocate(x86_thread_state64_t *context, size_t size)
117 {
118     /* round up size to 16byte multiple */
119     size = (size + 15) & -16;
120
121     context->rsp = ((u64)context->rsp) - size;
122
123     return (void *)context->rsp;
124 }
125
126 /* Arranging to invoke a C function is tricky, as we have to assume
127  * cdecl calling conventions (caller removes args) and x86/darwin
128  * alignment requirements.  The simplest way to arrange this,
129  * actually, is to open a new stack allocation.
130  * WARNING!!! THIS DOES NOT PRESERVE REGISTERS! */
131 void call_c_function_in_context(x86_thread_state64_t *context,
132                                 void *function,
133                                 int nargs,
134                                 ...)
135 {
136     va_list ap;
137     int i;
138     u64 *stack_pointer;
139
140     /* Set up to restore stack on exit. */
141     open_stack_allocation(context);
142
143     /* Have to keep stack 16byte aligned on x86/darwin. */
144     for (i = (1 & -nargs); i; i--) {
145         push_context(0, context);
146     }
147
148     context->rsp = ((u64)context->rsp) - nargs * 8;
149     stack_pointer = (u64 *)context->rsp;
150
151     va_start(ap, nargs);
152     if (nargs > 0) context->rdi = va_arg(ap, u64);
153     if (nargs > 1) context->rsi = va_arg(ap, u64);
154     if (nargs > 2) context->rdx = va_arg(ap, u64);
155     if (nargs > 3) context->rcx = va_arg(ap, u64);
156     if (nargs > 4) context->r8 = va_arg(ap, u64);
157     if (nargs > 5) context->r9 = va_arg(ap, u64);
158     for (i = 6; i < nargs; i++) {
159         stack_pointer[i] = va_arg(ap, u64);
160     }
161     va_end(ap);
162
163     push_context(context->rip, context);
164     context->rip = (u64) function;
165 }
166
167 void signal_emulation_wrapper(x86_thread_state64_t *thread_state,
168                               x86_float_state64_t *float_state,
169                               int signal,
170                               siginfo_t *siginfo,
171                               void (*handler)(int, siginfo_t *, void *))
172 {
173
174     /* CLH: FIXME **NOTE: HACK ALERT!** Ideally, we would allocate
175      * context and regs on the stack as local variables, but this
176      * causes problems for the lisp debugger. When it walks the stack
177      * for a back trace, it sees the 1) address of the local variable
178      * on the stack and thinks that is a frame pointer to a lisp
179      * frame, and, 2) the address of the sap that we alloc'ed in
180      * dynamic space and thinks that is a return address, so it,
181      * heuristicly (and wrongly), chooses that this should be
182      * interpreted as a lisp frame instead of as a C frame.
183      * We can work around this in this case by os_validating the
184      * context (and regs just for symmetry).
185      */
186
187     struct ucontext *context;
188     struct mcontext *regs;
189
190     context = (struct ucontext*) os_validate(0, sizeof(struct ucontext));
191     regs = (struct mcontext*) os_validate(0, sizeof(struct mcontext));
192     context->uc_mcontext = regs;
193
194     /* when BSD signals are fired, they mask they signals in sa_mask
195        which always seem to be the blockable_sigset, for us, so we
196        need to:
197        1) save the current sigmask
198        2) block blockable signals
199        3) call the signal handler
200        4) restore the sigmask */
201
202     build_fake_signal_context(context, thread_state, float_state);
203
204     block_blockable_signals();
205
206     handler(signal, siginfo, context);
207
208     update_thread_state_from_context(thread_state, float_state, context);
209
210     os_invalidate((os_vm_address_t)context, sizeof(struct ucontext));
211     os_invalidate((os_vm_address_t)regs, sizeof(struct mcontext));
212
213     /* Trap to restore the signal context. */
214     asm volatile ("mov %0, %%rax; mov %1, %%rbx; .quad 0xffffffffffff0b0f"
215                   : : "r" (thread_state), "r" (float_state));
216 }
217
218 #if defined DUMP_CONTEXT
219 void dump_context(x86_thread_state64_t *context)
220 {
221     int i;
222     u64 *stack_pointer;
223
224     printf("rax: %08lx  rcx: %08lx  rdx: %08lx  rbx: %08lx\n",
225            context->rax, context->rcx, context->rdx, context->rbx);
226     printf("rsp: %08lx  rbp: %08lx  rsi: %08lx  rdi: %08lx\n",
227            context->rsp, context->rbp, context->rsi, context->rdi);
228     printf("rip: %08lx  eflags: %08lx\n",
229            context->rip, context->rflags);
230     printf("cs: %04hx  ds: %04hx  es: %04hx  "
231            "ss: %04hx  fs: %04hx  gs: %04hx\n",
232            context->cs, context->ds, context->rs,
233            context->ss, context->fs, context->gs);
234
235     stack_pointer = (u64 *)context->rsp;
236     for (i = 0; i < 48; i+=4) {
237         printf("%08x:  %08x %08x %08x %08x\n",
238                context->rsp + (i * 4),
239                stack_pointer[i],
240                stack_pointer[i+1],
241                stack_pointer[i+2],
242                stack_pointer[i+3]);
243     }
244 }
245 #endif
246
247 void
248 control_stack_exhausted_handler(int signal, siginfo_t *siginfo, void *void_context) {
249     os_context_t *context = arch_os_get_context(&void_context);
250
251     arrange_return_to_lisp_function
252         (context, SymbolFunction(CONTROL_STACK_EXHAUSTED_ERROR));
253 }
254
255 void
256 undefined_alien_handler(int signal, siginfo_t *siginfo, void *void_context) {
257     os_context_t *context = arch_os_get_context(&void_context);
258
259     arrange_return_to_lisp_function
260         (context, SymbolFunction(UNDEFINED_ALIEN_VARIABLE_ERROR));
261 }
262
263 kern_return_t
264 catch_exception_raise(mach_port_t exception_port,
265                       mach_port_t thread,
266                       mach_port_t task,
267                       exception_type_t exception,
268                       exception_data_t code_vector,
269                       mach_msg_type_number_t code_count)
270 {
271     kern_return_t ret;
272     int signal;
273     siginfo_t* siginfo;
274
275 #ifdef LISP_FEATURE_SB_THREAD
276     thread_mutex_lock(&mach_exception_lock);
277 #endif
278
279     x86_thread_state64_t thread_state;
280     mach_msg_type_number_t thread_state_count = x86_THREAD_STATE64_COUNT;
281
282     x86_float_state64_t float_state;
283     mach_msg_type_number_t float_state_count = x86_FLOAT_STATE64_COUNT;
284
285     x86_exception_state64_t exception_state;
286     mach_msg_type_number_t exception_state_count = x86_EXCEPTION_STATE64_COUNT;
287
288     x86_thread_state64_t backup_thread_state;
289     x86_thread_state64_t *target_thread_state;
290     x86_float_state64_t *target_float_state;
291
292     os_vm_address_t addr;
293
294     struct thread *th = (struct thread*) exception_port;
295
296     FSHOW((stderr,"/entering catch_exception_raise with exception: %d\n", exception));
297
298     switch (exception) {
299
300     case EXC_BAD_ACCESS:
301         signal = SIGBUS;
302         ret = thread_get_state(thread,
303                                x86_THREAD_STATE64,
304                                (thread_state_t)&thread_state,
305                                &thread_state_count);
306         ret = thread_get_state(thread,
307                                x86_FLOAT_STATE64,
308                                (thread_state_t)&float_state,
309                                &float_state_count);
310         ret = thread_get_state(thread,
311                                x86_EXCEPTION_STATE64,
312                                (thread_state_t)&exception_state,
313                                &exception_state_count);
314         addr = (void*)exception_state.faultvaddr;
315
316
317         /* note the os_context hackery here.  When the signal handler returns,
318          * it won't go back to what it was doing ... */
319         if(addr >= CONTROL_STACK_GUARD_PAGE(th) &&
320            addr < CONTROL_STACK_GUARD_PAGE(th) + os_vm_page_size) {
321             /* We hit the end of the control stack: disable guard page
322              * protection so the error handler has some headroom, protect the
323              * previous page so that we can catch returns from the guard page
324              * and restore it. */
325             protect_control_stack_guard_page_thread(0, th);
326             protect_control_stack_return_guard_page_thread(1, th);
327
328             backup_thread_state = thread_state;
329             open_stack_allocation(&thread_state);
330
331             /* Save thread state */
332             target_thread_state =
333                 stack_allocate(&thread_state, sizeof(*target_thread_state));
334             (*target_thread_state) = backup_thread_state;
335
336             /* Save float state */
337             target_float_state =
338                 stack_allocate(&thread_state, sizeof(*target_float_state));
339             (*target_float_state) = float_state;
340
341             /* Set up siginfo */
342             siginfo = stack_allocate(&thread_state, sizeof(*siginfo));
343             /* what do we need to put in our fake siginfo?  It looks like
344              * the x86 code only uses si_signo and si_adrr. */
345             siginfo->si_signo = signal;
346             siginfo->si_addr = (void*)exception_state.faultvaddr;
347
348             call_c_function_in_context(&thread_state,
349                                        signal_emulation_wrapper,
350                                        5,
351                                        target_thread_state,
352                                        target_float_state,
353                                        signal,
354                                        siginfo,
355                                        control_stack_exhausted_handler);
356         }
357         else if(addr >= CONTROL_STACK_RETURN_GUARD_PAGE(th) &&
358                 addr < CONTROL_STACK_RETURN_GUARD_PAGE(th) + os_vm_page_size) {
359             /* We're returning from the guard page: reprotect it, and
360              * unprotect this one. This works even if we somehow missed
361              * the return-guard-page, and hit it on our way to new
362              * exhaustion instead. */
363             protect_control_stack_guard_page_thread(1, th);
364             protect_control_stack_return_guard_page_thread(0, th);
365         }
366         else if (addr >= undefined_alien_address &&
367                  addr < undefined_alien_address + os_vm_page_size) {
368             backup_thread_state = thread_state;
369             open_stack_allocation(&thread_state);
370
371             /* Save thread state */
372             target_thread_state =
373                 stack_allocate(&thread_state, sizeof(*target_thread_state));
374             (*target_thread_state) = backup_thread_state;
375
376             target_float_state =
377                 stack_allocate(&thread_state, sizeof(*target_float_state));
378             (*target_float_state) = float_state;
379
380             /* Set up siginfo */
381             siginfo = stack_allocate(&thread_state, sizeof(*siginfo));
382             /* what do we need to put in our fake siginfo?  It looks like
383              * the x86 code only uses si_signo and si_adrr. */
384             siginfo->si_signo = signal;
385             siginfo->si_addr = (void*)exception_state.faultvaddr;
386
387             call_c_function_in_context(&thread_state,
388                                        signal_emulation_wrapper,
389                                        5,
390                                        target_thread_state,
391                                        target_float_state,
392                                        signal,
393                                        siginfo,
394                                        undefined_alien_handler);
395         } else {
396
397             backup_thread_state = thread_state;
398             open_stack_allocation(&thread_state);
399
400             /* Save thread state */
401             target_thread_state =
402                 stack_allocate(&thread_state, sizeof(*target_thread_state));
403             (*target_thread_state) = backup_thread_state;
404
405             target_float_state =
406                 stack_allocate(&thread_state, sizeof(*target_float_state));
407             (*target_float_state) = float_state;
408
409             /* Set up siginfo */
410             siginfo = stack_allocate(&thread_state, sizeof(*siginfo));
411             /* what do we need to put in our fake siginfo?  It looks like
412              * the x86 code only uses si_signo and si_adrr. */
413             siginfo->si_signo = signal;
414             siginfo->si_addr = (void*)exception_state.faultvaddr;
415
416             call_c_function_in_context(&thread_state,
417                                        signal_emulation_wrapper,
418                                        5,
419                                        target_thread_state,
420                                        target_float_state,
421                                        signal,
422                                        siginfo,
423                                        memory_fault_handler);
424         }
425         ret = thread_set_state(thread,
426                                x86_THREAD_STATE64,
427                                (thread_state_t)&thread_state,
428                                thread_state_count);
429
430         ret = thread_set_state(thread,
431                                x86_FLOAT_STATE64,
432                                (thread_state_t)&float_state,
433                                float_state_count);
434 #ifdef LISP_FEATURE_SB_THREAD
435         thread_mutex_unlock(&mach_exception_lock);
436 #endif
437         return KERN_SUCCESS;
438
439     case EXC_BAD_INSTRUCTION:
440
441         ret = thread_get_state(thread,
442                                x86_THREAD_STATE64,
443                                (thread_state_t)&thread_state,
444                                &thread_state_count);
445         ret = thread_get_state(thread,
446                                x86_FLOAT_STATE64,
447                                (thread_state_t)&float_state,
448                                &float_state_count);
449         ret = thread_get_state(thread,
450                                x86_EXCEPTION_STATE64,
451                                (thread_state_t)&exception_state,
452                                &exception_state_count);
453         if (0xffffffffffff0b0f == *((u64 *)thread_state.rip)) {
454             /* fake sigreturn. */
455
456             /* When we get here, thread_state.rax is a pointer to a
457              * thread_state to restore. */
458             /* thread_state = *((thread_state_t *)thread_state.rax); */
459
460             ret = thread_set_state(thread,
461                                    x86_THREAD_STATE64,
462                                    (thread_state_t) thread_state.rax,
463                                    /* &thread_state, */
464                                    thread_state_count);
465
466             ret = thread_set_state(thread,
467                                    x86_FLOAT_STATE64,
468                                    (thread_state_t) thread_state.rbx,
469                                    /* &thread_state, */
470                                    float_state_count);
471         } else {
472
473             backup_thread_state = thread_state;
474             open_stack_allocation(&thread_state);
475
476             /* Save thread state */
477             target_thread_state =
478                 stack_allocate(&thread_state, sizeof(*target_thread_state));
479             (*target_thread_state) = backup_thread_state;
480
481             target_float_state =
482                 stack_allocate(&thread_state, sizeof(*target_float_state));
483             (*target_float_state) = float_state;
484
485             /* Set up siginfo */
486             siginfo = stack_allocate(&thread_state, sizeof(*siginfo));
487             /* what do we need to put in our fake siginfo?  It looks like
488              * the x86 code only uses si_signo and si_adrr. */
489             if (*((unsigned short *)target_thread_state->rip) == 0x0b0f) {
490                 signal = SIGTRAP;
491                 siginfo->si_signo = signal;
492                 siginfo->si_addr = (void*)exception_state.faultvaddr;
493                 target_thread_state->rip += 2;
494                 call_c_function_in_context(&thread_state,
495                                            signal_emulation_wrapper,
496                                            5,
497                                            target_thread_state,
498                                            target_float_state,
499                                            signal,
500                                            siginfo,
501                                            sigtrap_handler);
502             } else {
503                 signal = SIGILL;
504                 siginfo->si_signo = signal;
505                 siginfo->si_addr = (void*)exception_state.faultvaddr;
506
507                 call_c_function_in_context(&thread_state,
508                                            signal_emulation_wrapper,
509                                            5,
510                                            target_thread_state,
511                                            target_float_state,
512                                            signal,
513                                            siginfo,
514                                            sigill_handler);
515             }
516             ret = thread_set_state(thread,
517                                    x86_THREAD_STATE64,
518                                    (thread_state_t)&thread_state,
519                                    thread_state_count);
520             ret = thread_set_state(thread,
521                                    x86_FLOAT_STATE64,
522                                    (thread_state_t)&float_state,
523                                    float_state_count);
524         }
525 #ifdef LISP_FEATURE_SB_THREAD
526         thread_mutex_unlock(&mach_exception_lock);
527 #endif
528         return KERN_SUCCESS;
529
530     default:
531 #ifdef LISP_FEATURE_SB_THREAD
532         thread_mutex_unlock(&mach_exception_lock);
533 #endif
534         return KERN_INVALID_RIGHT;
535     }
536 }
537
538 void *
539 mach_exception_handler(void *port)
540 {
541   mach_msg_server(exc_server, 2048, (mach_port_t) port, 0);
542   /* mach_msg_server should never return, but it should dispatch mach
543    * exceptions to our catch_exception_raise function
544    */
545   abort();
546 }
547
548 /* Sets up the thread that will listen for mach exceptions. note that
549    the exception handlers will be run on this thread. This is
550    different from the BSD-style signal handling situation in which the
551    signal handlers run in the relevant thread directly. */
552
553 mach_port_t mach_exception_handler_port_set = MACH_PORT_NULL;
554
555 pthread_t
556 setup_mach_exception_handling_thread()
557 {
558     kern_return_t ret;
559     pthread_t mach_exception_handling_thread = NULL;
560     pthread_attr_t attr;
561
562     /* allocate a mach_port for this process */
563     ret = mach_port_allocate(mach_task_self(),
564                              MACH_PORT_RIGHT_PORT_SET,
565                              &mach_exception_handler_port_set);
566
567     /* create the thread that will receive the mach exceptions */
568
569     FSHOW((stderr, "Creating mach_exception_handler thread!\n"));
570
571     pthread_attr_init(&attr);
572     pthread_create(&mach_exception_handling_thread,
573                    &attr,
574                    mach_exception_handler,
575                    (void*) mach_exception_handler_port_set);
576     pthread_attr_destroy(&attr);
577
578     return mach_exception_handling_thread;
579 }
580
581 /* tell the kernel that we want EXC_BAD_ACCESS exceptions sent to the
582    exception port (which is being listened to do by the mach
583    exception handling thread). */
584 kern_return_t
585 mach_thread_init(mach_port_t thread_exception_port)
586 {
587     kern_return_t ret;
588     /* allocate a named port for the thread */
589
590     FSHOW((stderr, "Allocating mach port %x\n", thread_exception_port));
591
592     ret = mach_port_allocate_name(mach_task_self(),
593                                   MACH_PORT_RIGHT_RECEIVE,
594                                   thread_exception_port);
595     if (ret) {
596         lose("mach_port_allocate_name failed with return_code %d\n", ret);
597     }
598
599     /* establish the right for the thread_exception_port to send messages */
600     ret = mach_port_insert_right(mach_task_self(),
601                                  thread_exception_port,
602                                  thread_exception_port,
603                                  MACH_MSG_TYPE_MAKE_SEND);
604     if (ret) {
605         lose("mach_port_insert_right failed with return_code %d\n", ret);
606     }
607
608     ret = thread_set_exception_ports(mach_thread_self(),
609                                      EXC_MASK_BAD_ACCESS | EXC_MASK_BAD_INSTRUCTION,
610                                      thread_exception_port,
611                                      EXCEPTION_DEFAULT,
612                                      THREAD_STATE_NONE);
613     if (ret) {
614         lose("thread_set_exception_port failed with return_code %d\n", ret);
615     }
616
617     ret = mach_port_move_member(mach_task_self(),
618                                 thread_exception_port,
619                                 mach_exception_handler_port_set);
620     if (ret) {
621         lose("mach_port_ failed with return_code %d\n", ret);
622     }
623
624     return ret;
625 }
626
627 void
628 setup_mach_exceptions() {
629     setup_mach_exception_handling_thread();
630     mach_thread_init(THREAD_STRUCT_TO_EXCEPTION_PORT(all_threads));
631 }
632
633 pid_t
634 mach_fork() {
635     pid_t pid = fork();
636     if (pid == 0) {
637         setup_mach_exceptions();
638         return pid;
639     } else {
640         return pid;
641     }
642 }
643
644 #endif