1.0.10.36: support for Darwin versions that support __DARWIN_UNIX03
[sbcl.git] / src / runtime / x86-darwin-os.c
index fa924d5..97d3999 100644 (file)
@@ -7,7 +7,24 @@
 #endif
 
 #include "thread.h"
+#include "validate.h"
+#include "runtime.h"
+#include "interrupt.h"
 #include "x86-darwin-os.h"
+#include "genesis/fdefn.h"
+
+#include <mach/mach.h>
+#include <mach/mach_error.h>
+#include <mach/mach_types.h>
+#include <mach/sync_policy.h>
+#include <mach/vm_region.h>
+#include <mach/machine/thread_state.h>
+#include <mach/machine/thread_status.h>
+#include <sys/_types.h>
+#include <sys/ucontext.h>
+#include <pthread.h>
+#include <assert.h>
+#include <stdlib.h>
 
 #ifdef LISP_FEATURE_SB_THREAD
 
@@ -28,6 +45,10 @@ void set_data_desc_addr(data_desc_t* desc, void* addr)
 
 #endif
 
+#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
+kern_return_t mach_thread_init(mach_port_t thread_exception_port);
+#endif
+
 int arch_os_thread_init(struct thread *thread) {
 #ifdef LISP_FEATURE_SB_THREAD
     int n;
@@ -36,7 +57,7 @@ int arch_os_thread_init(struct thread *thread) {
     data_desc_t ldt_entry = { 0, 0, 0, DESC_DATA_WRITE,
                               3, 1, 0, DESC_DATA_32B, DESC_GRAN_BYTE, 0 };
 
-    set_data_desc_addr(&ldt_entry, (unsigned long) thread);
+    set_data_desc_addr(&ldt_entry, thread);
     set_data_desc_size(&ldt_entry, dynamic_values_bytes);
 
     thread_mutex_lock(&modify_ldt_lock);
@@ -58,6 +79,9 @@ int arch_os_thread_init(struct thread *thread) {
     thread->tls_cookie=n;
     pthread_setspecific(specials,thread);
 #endif
+#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
+    mach_thread_init(THREAD_STRUCT_TO_EXCEPTION_PORT(thread));
+#endif
 
 #ifdef LISP_FEATURE_C_STACK_IS_CONTROL_STACK
     stack_t sigstack;
@@ -77,8 +101,8 @@ int arch_os_thread_cleanup(struct thread *thread) {
 #if defined(LISP_FEATURE_SB_THREAD)
     int n = thread->tls_cookie;
 
-    /* Set the %%fs register back to 0 and free the the ldt
-     * by setting it to NULL.
+    /* Set the %%fs register back to 0 and free the ldt by setting it
+     * to NULL.
      */
     FSHOW_SIGNAL((stderr, "/ TLS: Freeing LDT %x\n", n));
 
@@ -90,3 +114,540 @@ int arch_os_thread_cleanup(struct thread *thread) {
     return 1;                  /* success */
 }
 
+#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
+
+void sigill_handler(int signal, siginfo_t *siginfo, void *void_context);
+void sigtrap_handler(int signal, siginfo_t *siginfo, void *void_context);
+void memory_fault_handler(int signal, siginfo_t *siginfo, void *void_context);
+
+/* exc_server handles mach exception messages from the kernel and
+ * calls catch exception raise. We use the system-provided
+ * mach_msg_server, which, I assume, calls exc_server in a loop.
+ *
+ */
+extern boolean_t exc_server();
+
+/* This executes in the faulting thread as part of the signal
+ * emulation.  It is passed a context with the uc_mcontext field
+ * pointing to a valid block of memory. */
+void build_fake_signal_context(os_context_t *context,
+                               x86_thread_state32_t *thread_state,
+                               x86_float_state32_t *float_state) {
+    pthread_sigmask(0, NULL, &context->uc_sigmask);
+    context->uc_mcontext->SS = *thread_state;
+    context->uc_mcontext->FS = *float_state;
+}
+
+/* This executes in the faulting thread as part of the signal
+ * emulation.  It is effectively the inverse operation from above. */
+void update_thread_state_from_context(x86_thread_state32_t *thread_state,
+                                      x86_float_state32_t *float_state,
+                                      os_context_t *context) {
+    *thread_state = context->uc_mcontext->SS;
+    *float_state = context->uc_mcontext->FS;
+    pthread_sigmask(SIG_SETMASK, &context->uc_sigmask, NULL);
+}
+
+/* Modify a context to push new data on its stack. */
+void push_context(u32 data, x86_thread_state32_t *thread_state)
+{
+    u32 *stack_pointer;
+
+    stack_pointer = (u32*) thread_state->ESP;
+    *(--stack_pointer) = data;
+    thread_state->ESP = (unsigned int) stack_pointer;
+}
+
+void align_context_stack(x86_thread_state32_t *thread_state)
+{
+    /* 16byte align the stack (provided that the stack is, as it
+     * should be, 4byte aligned. */
+    while (thread_state->ESP & 15) push_context(0, thread_state);
+}
+
+/* Stack allocation starts with a context that has a mod-4 ESP value
+ * and needs to leave a context with a mod-16 ESP that will restore
+ * the old ESP value and other register state when activated.  The
+ * first part of this is the recovery trampoline, which loads ESP from
+ * EBP, pops EBP, and returns. */
+asm("_stack_allocation_recover: movl %ebp, %esp; popl %ebp; ret;");
+
+void open_stack_allocation(x86_thread_state32_t *thread_state)
+{
+    void stack_allocation_recover(void);
+
+    push_context(thread_state->EIP, thread_state);
+    push_context(thread_state->EBP, thread_state);
+    thread_state->EBP = thread_state->ESP;
+    thread_state->EIP = (unsigned int) stack_allocation_recover;
+
+    align_context_stack(thread_state);
+}
+
+/* Stack allocation of data starts with a context with a mod-16 ESP
+ * value and reserves some space on it by manipulating the ESP
+ * register. */
+void *stack_allocate(x86_thread_state32_t *thread_state, size_t size)
+{
+    /* round up size to 16byte multiple */
+    size = (size + 15) & -16;
+
+    thread_state->ESP = ((u32)thread_state->ESP) - size;
+
+    return (void *)thread_state->ESP;
+}
+
+/* Arranging to invoke a C function is tricky, as we have to assume
+ * cdecl calling conventions (caller removes args) and x86/darwin
+ * alignment requirements.  The simplest way to arrange this,
+ * actually, is to open a new stack allocation.
+ * WARNING!!! THIS DOES NOT PRESERVE REGISTERS! */
+void call_c_function_in_context(x86_thread_state32_t *thread_state,
+                                void *function,
+                                int nargs,
+                                ...)
+{
+    va_list ap;
+    int i;
+    u32 *stack_pointer;
+
+    /* Set up to restore stack on exit. */
+    open_stack_allocation(thread_state);
+
+    /* Have to keep stack 16byte aligned on x86/darwin. */
+    for (i = (3 & -nargs); i; i--) {
+        push_context(0, thread_state);
+    }
+
+    thread_state->ESP = ((u32)thread_state->ESP) - nargs * 4;
+    stack_pointer = (u32 *)thread_state->ESP;
+
+    va_start(ap, nargs);
+    for (i = 0; i < nargs; i++) {
+        //push_context(va_arg(ap, u32), thread_state);
+        stack_pointer[i] = va_arg(ap, u32);
+    }
+    va_end(ap);
+
+    push_context(thread_state->EIP, thread_state);
+    thread_state->EIP = (unsigned int) function;
+}
+
+void signal_emulation_wrapper(x86_thread_state32_t *thread_state,
+                              x86_float_state32_t *float_state,
+                              int signal,
+                              siginfo_t *siginfo,
+                              void (*handler)(int, siginfo_t *, void *))
+{
+
+    /* CLH: FIXME **NOTE: HACK ALERT!** Ideally, we would allocate
+     * context and regs on the stack as local variables, but this
+     * causes problems for the lisp debugger. When it walks the stack
+     * for a back trace, it sees the 1) address of the local variable
+     * on the stack and thinks that is a frame pointer to a lisp
+     * frame, and, 2) the address of the sap that we alloc'ed in
+     * dynamic space and thinks that is a return address, so it,
+     * heuristicly (and wrongly), chooses that this should be
+     * interpreted as a lisp frame instead of as a C frame.
+     * We can work around this in this case by os_validating the
+     * context (and regs just for symmetry).
+     */
+
+    os_context_t *context;
+#if __DARWIN_UNIX03
+    struct __darwin_mcontext32 *regs;
+#else
+    struct mcontext *regs;
+#endif
+
+    context = (os_context_t*) os_validate(0, sizeof(os_context_t));
+#if __DARWIN_UNIX03
+    regs = (struct __darwin_mcontext32*) os_validate(0, sizeof(struct __darwin_mcontext32));
+#else
+    regs = (struct mcontext*) os_validate(0, sizeof(struct mcontext));
+#endif
+    context->uc_mcontext = regs;
+
+    /* when BSD signals are fired, they mask they signals in sa_mask
+       which always seem to be the blockable_sigset, for us, so we
+       need to:
+       1) save the current sigmask
+       2) block blockable signals
+       3) call the signal handler
+       4) restore the sigmask */
+
+    build_fake_signal_context(context, thread_state, float_state);
+
+    block_blockable_signals();
+
+    handler(signal, siginfo, context);
+
+    update_thread_state_from_context(thread_state, float_state, context);
+
+    os_invalidate((os_vm_address_t)context, sizeof(os_context_t));
+#if __DARWIN_UNIX03
+    os_invalidate((os_vm_address_t)regs, sizeof(struct __darwin_mcontext32));
+#else
+    os_invalidate((os_vm_address_t)regs, sizeof(struct mcontext));
+#endif
+
+    /* Trap to restore the signal context. */
+    asm volatile ("movl %0, %%eax; movl %1, %%ebx; .long 0xffff0b0f"
+                  : : "r" (thread_state), "r" (float_state));
+}
+
+/* Convenience wrapper for the above */
+void call_handler_on_thread(mach_port_t thread,
+                            x86_thread_state32_t *thread_state,
+                            int signal,
+                            siginfo_t *siginfo,
+                            void (*handler)(int, siginfo_t *, void *))
+{
+    x86_thread_state32_t new_state;
+    x86_thread_state32_t *save_thread_state;
+    x86_float_state32_t *save_float_state;
+    mach_msg_type_number_t state_count;
+    siginfo_t *save_siginfo;
+    kern_return_t ret;
+    /* Initialize the new state */
+    new_state = *thread_state;
+    open_stack_allocation(&new_state);
+    /* Save old state */
+    save_thread_state = (x86_thread_state32_t *)stack_allocate(&new_state, sizeof(*save_thread_state));
+    *save_thread_state = *thread_state;
+    /* Save float state */
+    save_float_state = (x86_float_state32_t *)stack_allocate(&new_state, sizeof(*save_float_state));
+    state_count = x86_FLOAT_STATE32_COUNT;
+    if ((ret = thread_get_state(thread,
+                                x86_FLOAT_STATE32,
+                                (thread_state_t)save_float_state,
+                                &state_count)) != KERN_SUCCESS)
+        lose("thread_get_state (x86_THREAD_STATE32) failed %d\n", ret);
+    /* Set up siginfo */
+    save_siginfo = stack_allocate(&new_state, sizeof(*siginfo));
+    if (siginfo == NULL)
+        save_siginfo = siginfo;
+    else
+        *save_siginfo = *siginfo;
+    /* Prepare to call */
+    call_c_function_in_context(&new_state,
+                               signal_emulation_wrapper,
+                               5,
+                               save_thread_state,
+                               save_float_state,
+                               signal,
+                               save_siginfo,
+                               handler);
+    /* Update the thread state */
+    state_count = x86_THREAD_STATE32_COUNT;
+    if ((ret = thread_set_state(thread,
+                                x86_THREAD_STATE32,
+                                (thread_state_t)&new_state,
+                                state_count)) != KERN_SUCCESS)
+        lose("thread_set_state (x86_FLOAT_STATE32) failed %d\n", ret);
+
+}
+
+#if defined DUMP_CONTEXT
+void dump_context(x86_thread_state32_t *thread_state)
+{
+    int i;
+    u32 *stack_pointer;
+
+    printf("eax: %08lx  ecx: %08lx  edx: %08lx  ebx: %08lx\n",
+           thread_state->EAX, thread_state->ECX, thread_state->EDX, thread_state->EAX);
+    printf("esp: %08lx  ebp: %08lx  esi: %08lx  edi: %08lx\n",
+           thread_state->ESP, thread_state->EBP, thread_state->ESI, thread_state->EDI);
+    printf("eip: %08lx  eflags: %08lx\n",
+           thread_state->EIP, thread_state->EFLAGS);
+    printf("cs: %04hx  ds: %04hx  es: %04hx  "
+           "ss: %04hx  fs: %04hx  gs: %04hx\n",
+           thread_state->CS,
+           thread_state->DS,
+           thread_state->ES,
+           thread_state->SS,
+           thread_state->FS,
+           thread_state->GS);
+
+    stack_pointer = (u32 *)thread_state->ESP;
+    for (i = 0; i < 48; i+=4) {
+        printf("%08x:  %08x %08x %08x %08x\n",
+               thread_state->ESP + (i * 4),
+               stack_pointer[i],
+               stack_pointer[i+1],
+               stack_pointer[i+2],
+               stack_pointer[i+3]);
+    }
+}
+#endif
+
+void
+control_stack_exhausted_handler(int signal, siginfo_t *siginfo, void *void_context) {
+    os_context_t *context = arch_os_get_context(&void_context);
+
+    arrange_return_to_lisp_function
+        (context, SymbolFunction(CONTROL_STACK_EXHAUSTED_ERROR));
+}
+
+void
+undefined_alien_handler(int signal, siginfo_t *siginfo, void *void_context) {
+    os_context_t *context = arch_os_get_context(&void_context);
+
+    arrange_return_to_lisp_function
+        (context, SymbolFunction(UNDEFINED_ALIEN_VARIABLE_ERROR));
+}
+
+kern_return_t
+catch_exception_raise(mach_port_t exception_port,
+                      mach_port_t thread,
+                      mach_port_t task,
+                      exception_type_t exception,
+                      exception_data_t code_vector,
+                      mach_msg_type_number_t code_count)
+{
+    struct thread *th = (struct thread*) exception_port;
+    x86_thread_state32_t thread_state;
+    mach_msg_type_number_t state_count;
+    vm_address_t region_addr;
+    vm_size_t region_size;
+    vm_region_basic_info_data_t region_info;
+    mach_msg_type_number_t info_count;
+    mach_port_t region_name;
+    void *addr = NULL;
+    int signal = 0;
+    void (*handler)(int, siginfo_t *, void *) = NULL;
+    siginfo_t siginfo;
+    kern_return_t ret;
+
+    /* Get state and info */
+    state_count = x86_THREAD_STATE32_COUNT;
+    if ((ret = thread_get_state(thread,
+                                x86_THREAD_STATE32,
+                                (thread_state_t)&thread_state,
+                                &state_count)) != KERN_SUCCESS)
+        lose("thread_get_state (x86_THREAD_STATE32) failed %d\n", ret);
+    switch (exception) {
+    case EXC_BAD_ACCESS:
+        signal = SIGBUS;
+        /* Check if write protection fault */
+        if ((code_vector[0] & OS_VM_PROT_ALL) == 0) {
+            ret = KERN_INVALID_RIGHT;
+            break;
+        }
+        /* Get vm_region info */
+        region_addr = (vm_address_t)code_vector[1];
+        info_count = VM_REGION_BASIC_INFO_COUNT;
+        if ((ret = vm_region(mach_task_self(),
+                             &region_addr,
+                             &region_size,
+                             VM_REGION_BASIC_INFO,
+                             (vm_region_info_t)&region_info,
+                             &info_count,
+                             &region_name)))
+            lose("vm_region (VM_REGION_BASIC_INFO) failed failed %d\n", ret);
+        /* Check if still protected */
+        if ((region_info.protection & OS_VM_PROT_ALL) == 0) {
+          /* KLUDGE:
+           * If two threads fault on the same page, the protection
+           * is cleared as the first thread runs memory_fault_handler.
+           * Grep for "not marked as write-protected" in gencgc.c
+           */
+            ret = KERN_SUCCESS;
+            break;
+        }
+        addr = (void*)code_vector[1];
+        /* At stack guard */
+        if (os_trunc_to_page(addr) == CONTROL_STACK_GUARD_PAGE(th)) {
+            protect_control_stack_guard_page_thread(0, th);
+            protect_control_stack_return_guard_page_thread(1, th);
+            handler = control_stack_exhausted_handler;
+            break;
+        }
+        /* Return from stack guard */
+        if (os_trunc_to_page(addr) == CONTROL_STACK_RETURN_GUARD_PAGE(th)) {
+            protect_control_stack_guard_page_thread(1, th);
+            protect_control_stack_return_guard_page_thread(0, th);
+            break;
+        }
+        /* Undefined alien */
+        if (os_trunc_to_page(addr) == undefined_alien_address) {
+            handler = undefined_alien_handler;
+            break;
+        }
+        /* Regular memory fault */
+        handler = memory_fault_handler;
+        break;
+    case EXC_BAD_INSTRUCTION:
+        signal = SIGTRAP;
+        /* Check if illegal instruction trap */
+        if (code_vector[0] != EXC_I386_INVOP) {
+            ret = KERN_INVALID_RIGHT;
+            break;
+        }
+        /* Check if UD2 instruction */
+        if (*(unsigned short *)thread_state.EIP != 0x0b0f) {
+            /* KLUDGE: There are two ways we could get here:
+             * 1) We're executing data and we've hit some truly
+             *    illegal opcode, of which there are a few, see
+             *    Intel 64 and IA-32 Architectures
+             *    Sofware Developer's Manual
+             *    Volume 3A page 5-34)
+             * 2) The kernel started an unrelated signal handler
+             *    before we got a chance to run. The context that
+             *    caused the exception is saved in a stack frame
+             *    somewhere down below.
+             * In either case we rely on the exception to retrigger,
+             * eventually bailing out if we're spinning on case 2).
+             */
+            static mach_port_t last_thread;
+            static unsigned int last_eip;
+            if (last_thread == thread && last_eip == thread_state.EIP)
+                ret = KERN_INVALID_RIGHT;
+            else
+                ret = KERN_SUCCESS;
+            last_thread = thread;
+            last_eip = thread_state.EIP;
+            break;
+        }
+        /* Skip the trap code */
+        thread_state.EIP += 2;
+        /* Return from handler? */
+        if (*(unsigned short *)thread_state.EIP == 0xffff) {
+            if ((ret = thread_set_state(thread,
+                                        x86_THREAD_STATE32,
+                                        (thread_state_t)thread_state.EAX,
+                                        x86_THREAD_STATE32_COUNT)) != KERN_SUCCESS)
+                lose("thread_set_state (x86_THREAD_STATE32) failed %d\n", ret);
+            if ((ret = thread_set_state(thread,
+                                        x86_FLOAT_STATE32,
+                                        (thread_state_t)thread_state.EBX,
+                                        x86_FLOAT_STATE32_COUNT)) != KERN_SUCCESS)
+                lose("thread_set_state (x86_FLOAT_STATE32) failed %d\n", ret);
+            break;
+        }
+        /* Trap call */
+        handler = sigtrap_handler;
+        break;
+    default:
+        ret = KERN_INVALID_RIGHT;
+    }
+    /* Call handler */
+    if (handler != 0) {
+      siginfo.si_signo = signal;
+      siginfo.si_addr = addr;
+      call_handler_on_thread(thread, &thread_state, signal, &siginfo, handler);
+    }
+    return ret;
+}
+
+void *
+mach_exception_handler(void *port)
+{
+  mach_msg_server(exc_server, 2048, (mach_port_t) port, 0);
+  /* mach_msg_server should never return, but it should dispatch mach
+   * exceptions to our catch_exception_raise function
+   */
+  abort();
+}
+
+#endif
+
+#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
+
+/* Sets up the thread that will listen for mach exceptions. note that
+   the exception handlers will be run on this thread. This is
+   different from the BSD-style signal handling situation in which the
+   signal handlers run in the relevant thread directly. */
+
+mach_port_t mach_exception_handler_port_set = MACH_PORT_NULL;
+
+pthread_t
+setup_mach_exception_handling_thread()
+{
+    kern_return_t ret;
+    pthread_t mach_exception_handling_thread = NULL;
+    pthread_attr_t attr;
+
+    /* allocate a mach_port for this process */
+    ret = mach_port_allocate(mach_task_self(),
+                             MACH_PORT_RIGHT_PORT_SET,
+                             &mach_exception_handler_port_set);
+
+    /* create the thread that will receive the mach exceptions */
+
+    FSHOW((stderr, "Creating mach_exception_handler thread!\n"));
+
+    pthread_attr_init(&attr);
+    pthread_create(&mach_exception_handling_thread,
+                   &attr,
+                   mach_exception_handler,
+                   (void*) mach_exception_handler_port_set);
+    pthread_attr_destroy(&attr);
+
+    return mach_exception_handling_thread;
+}
+
+/* tell the kernel that we want EXC_BAD_ACCESS exceptions sent to the
+   exception port (which is being listened to do by the mach
+   exception handling thread). */
+kern_return_t
+mach_thread_init(mach_port_t thread_exception_port)
+{
+    kern_return_t ret;
+    /* allocate a named port for the thread */
+
+    FSHOW((stderr, "Allocating mach port %x\n", thread_exception_port));
+
+    ret = mach_port_allocate_name(mach_task_self(),
+                                  MACH_PORT_RIGHT_RECEIVE,
+                                  thread_exception_port);
+    if (ret) {
+        lose("mach_port_allocate_name failed with return_code %d\n", ret);
+    }
+
+    /* establish the right for the thread_exception_port to send messages */
+    ret = mach_port_insert_right(mach_task_self(),
+                                 thread_exception_port,
+                                 thread_exception_port,
+                                 MACH_MSG_TYPE_MAKE_SEND);
+    if (ret) {
+        lose("mach_port_insert_right failed with return_code %d\n", ret);
+    }
+
+    ret = thread_set_exception_ports(mach_thread_self(),
+                                     EXC_MASK_BAD_ACCESS | EXC_MASK_BAD_INSTRUCTION,
+                                     thread_exception_port,
+                                     EXCEPTION_DEFAULT,
+                                     THREAD_STATE_NONE);
+    if (ret) {
+        lose("thread_set_exception_port failed with return_code %d\n", ret);
+    }
+
+    ret = mach_port_move_member(mach_task_self(),
+                                thread_exception_port,
+                                mach_exception_handler_port_set);
+    if (ret) {
+        lose("mach_port_ failed with return_code %d\n", ret);
+    }
+
+    return ret;
+}
+
+void
+setup_mach_exceptions() {
+    setup_mach_exception_handling_thread();
+    mach_thread_init(THREAD_STRUCT_TO_EXCEPTION_PORT(all_threads));
+}
+
+pid_t
+mach_fork() {
+    pid_t pid = fork();
+    if (pid == 0) {
+        setup_mach_exceptions();
+        return pid;
+    } else {
+        return pid;
+    }
+}
+
+#endif