better signal handling

Instead of relying on creating a fake stack frame, and having no signals
delivered, kernel bugs, accidentally gc_collect, or other issues occur
during the delivery and execution of these calls, use the ability we
added recently to emulate a longjmp into a unw_context to eliminate any
time where there would exist any invalid states.

Secondly, when calling jl_exit_thread0_cb, we used to end up completely
smashing the unwind info (with CFI_NOUNWIND), but this makes core files
from SIGQUIT much less helpful, so we now have a `fake_stack_pop`
function with contains the necessary CFI directives such that a minimal
unwind from the debugger will likely still succeed up into the frames
that were removed. We cannot do this perfectly on AArch64 since that
platform's DWARF spec lacks the ability to do so. On other platforms,
this should be possible to implement exactly (subject to libunwind
implementation quality).

Author: vtjnash

src/jl_exported_funcs.inc

@@ -420,7 +420,6 @@
     XX(jl_set_zero_subnormals) \
     XX(jl_sigatomic_begin) \
     XX(jl_sigatomic_end) \
-    XX(jl_sig_throw) \
     XX(jl_spawn) \
     XX(jl_specializations_get_linfo) \
     XX(jl_specializations_lookup) \

src/julia.h

@@ -2310,7 +2310,6 @@ JL_DLLEXPORT int jl_set_task_tid(jl_task_t *task, int16_t tid) JL_NOTSAFEPOINT;
 JL_DLLEXPORT int jl_set_task_threadpoolid(jl_task_t *task, int8_t tpid) JL_NOTSAFEPOINT;
 JL_DLLEXPORT void JL_NORETURN jl_throw(jl_value_t *e JL_MAYBE_UNROOTED);
 JL_DLLEXPORT void JL_NORETURN jl_rethrow(void);
-JL_DLLEXPORT void JL_NORETURN jl_sig_throw(void);
 JL_DLLEXPORT void JL_NORETURN jl_rethrow_other(jl_value_t *e JL_MAYBE_UNROOTED);
 JL_DLLEXPORT void JL_NORETURN jl_no_exc_handler(jl_value_t *e, jl_task_t *ct);
 JL_DLLEXPORT JL_CONST_FUNC jl_gcframe_t **(jl_get_pgcstack)(void) JL_GLOBALLY_ROOTED JL_NOTSAFEPOINT;

src/julia_threads.h

@@ -187,6 +187,9 @@ typedef struct _jl_tls_states_t {
     // Saved exception for previous *external* API call or NULL if cleared.
     // Access via jl_exception_occurred().
     struct _jl_value_t *previous_exception;
+#ifdef _OS_DARWIN_
+    jl_jmp_buf *volatile safe_restore;
+#endif
 
     // currently-held locks, to be released when an exception is thrown
     small_arraylist_t locks;

src/rtutils.c

@@ -269,25 +269,38 @@ JL_DLLEXPORT void jl_eh_restore_state(jl_task_t *ct, jl_handler_t *eh)
     // `eh` may be not equal to `ct->eh`. See `jl_pop_handler`
     // This function should **NOT** have any safepoint before the ones at the
     // end.
-    sig_atomic_t old_defer_signal = ct->ptls->defer_signal;
+    jl_ptls_t ptls = ct->ptls;
+    sig_atomic_t old_defer_signal = ptls->defer_signal;
     ct->eh = eh->prev;
     ct->gcstack = eh->gcstack;
-    small_arraylist_t *locks = &ct->ptls->locks;
+    small_arraylist_t *locks = &ptls->locks;
     int unlocks = locks->len > eh->locks_len;
     if (unlocks) {
         for (size_t i = locks->len; i > eh->locks_len; i--)
             jl_mutex_unlock_nogc((jl_mutex_t*)locks->items[i - 1]);
         locks->len = eh->locks_len;
     }
     ct->world_age = eh->world_age;
-    ct->ptls->defer_signal = eh->defer_signal;
-    int8_t old_gc_state = jl_atomic_load_relaxed(&ct->ptls->gc_state);
+    ptls->defer_signal = eh->defer_signal;
+    int8_t old_gc_state = jl_atomic_load_relaxed(&ptls->gc_state);
     if (old_gc_state != eh->gc_state)
-        jl_atomic_store_release(&ct->ptls->gc_state, eh->gc_state);
+        jl_atomic_store_release(&ptls->gc_state, eh->gc_state);
     if (!old_gc_state || !eh->gc_state) // it was or is unsafe now
-        jl_gc_safepoint_(ct->ptls);
+        jl_gc_safepoint_(ptls);
+    jl_value_t *exception = ptls->sig_exception;
+    if (exception) {
+        int8_t oldstate = jl_gc_unsafe_enter(ptls);
+        /* The temporary ptls->bt_data is rooted by special purpose code in the
+        GC. This exists only for the purpose of preserving bt_data until we
+        set ptls->bt_size=0 below. */
+        jl_push_excstack(ct, &ct->excstack, exception,
+                         ptls->bt_data, ptls->bt_size);
+        ptls->bt_size = 0;
+        ptls->sig_exception = NULL;
+        jl_gc_unsafe_leave(ptls, oldstate);
+    }
     if (old_defer_signal && !eh->defer_signal)
-        jl_sigint_safepoint(ct->ptls);
+        jl_sigint_safepoint(ptls);
     if (jl_atomic_load_relaxed(&jl_gc_have_pending_finalizers) &&
             unlocks && eh->locks_len == 0) {
         jl_gc_run_pending_finalizers(ct);

src/signals-mach.c

@@ -222,38 +222,81 @@ typedef arm_exception_state64_t host_exception_state_t;
 #define HOST_EXCEPTION_STATE_COUNT ARM_EXCEPTION_STATE64_COUNT
 #endif
 
-static void jl_call_in_state(jl_ptls_t ptls2, host_thread_state_t *state,
-                             void (*fptr)(void))
+// create a fake function that describes the variable manipulations in jl_call_in_state
+__attribute__((naked)) static void fake_stack_pop(void)
 {
 #ifdef _CPU_X86_64_
-    uintptr_t rsp = state->__rsp;
+    __asm__ volatile (
+        "  .cfi_signal_frame\n"
+        "  .cfi_def_cfa %rsp, 0\n" // CFA here uses %rsp directly
+        "  .cfi_offset %rip, 0\n" // previous value of %rip at CFA
+        "  .cfi_offset %rsp, 8\n" // previous value of %rsp at CFA
+        "  nop\n"
+    );
 #elif defined(_CPU_AARCH64_)
-    uintptr_t rsp = state->__sp;
+    __asm__ volatile (
+        "  .cfi_signal_frame\n"
+        "  .cfi_def_cfa sp, 0\n" // use sp as fp here
+        "  .cfi_offset lr, 0\n"
+        "  .cfi_offset sp, 8\n"
+        // Anything else got smashed, since we didn't explictly copy all of the
+        // state object to the stack (to build a real sigreturn frame).
+        // This is also not quite valid, since the AArch64 DWARF spec lacks the ability to define how to restore the LR reigster correctly,
+        // so normally libunwind implementations on linux detect this function specially and hack around the invalid info:
+        // https://github.com/llvm/llvm-project/commit/c82deed6764cbc63966374baf9721331901ca958
+        " nop\n"
+    );
 #else
-#error "julia: throw-in-context not supported on this platform"
+CFI_NORETURN
 #endif
-    if (ptls2 == NULL || is_addr_on_sigstack(ptls2, (void*)rsp)) {
-        rsp = (rsp - 256) & ~(uintptr_t)15; // redzone and re-alignment
-    }
-    else {
-        rsp = (uintptr_t)ptls2->signal_stack + (ptls2->signal_stack_size ? ptls2->signal_stack_size : sig_stack_size);
-    }
-    assert(rsp % 16 == 0);
-    rsp -= 16;
+}
 
+static void jl_call_in_state(jl_ptls_t ptls2, host_thread_state_t *state,
+                             void (*fptr)(void))
+{
+#ifdef _CPU_X86_64_
+    uintptr_t sp = state->__rsp;
+#elif defined(_CPU_AARCH64_)
+    uintptr_t sp = state->__sp;
+#endif
+    sp = (sp - 256) & ~(uintptr_t)15; // redzone and re-alignment
+    assert(sp % 16 == 0);
+    sp -= 16;
 #ifdef _CPU_X86_64_
-    rsp -= sizeof(void*);
-    state->__rsp = rsp; // set stack pointer
+    // set return address to NULL
+    *(uintptr_t*)sp = 0;
+    // pushq %sp
+    sp -= sizeof(void*);
+    *(uintptr_t*)sp = state->__rsp;
+    // pushq %rip
+    sp -= sizeof(void*);
+    *(uintptr_t*)sp = state->__rip;
+    // pushq .fake_stack_pop + 1; aka call from fake_stack_pop
+    sp -= sizeof(void*);
+    *(uintptr_t*)sp = (uintptr_t)&fake_stack_pop + 1;
+    state->__rsp = sp; // set stack pointer
     state->__rip = (uint64_t)fptr; // "call" the function
 #elif defined(_CPU_AARCH64_)
-    state->__sp = rsp;
-    state->__pc = (uint64_t)fptr;
-    state->__lr = 0;
+    // push {%sp, %pc + 4}
+    sp -= sizeof(void*);
+    *(uintptr_t*)sp = state->__sp;
+    sp -= sizeof(void*);
+    *(uintptr_t*)sp = (uintptr_t)state->__pc;
+    state->__sp = sp; // x31
+    state->__pc = (uint64_t)fptr; // pc
+    state->__lr = (uintptr_t)&fake_stack_pop + 4; // x30
 #else
 #error "julia: throw-in-context not supported on this platform"
 #endif
 }
 
+static void jl_longjmp_in_state(host_thread_state_t *state, jl_jmp_buf jmpbuf)
+{
+
+    if (!jl_simulate_longjmp(jmpbuf, (bt_context_t*)state))
+        abort(); // unreachable
+}
+
 #ifdef _CPU_X86_64_
 int is_write_fault(host_exception_state_t exc_state) {
     return exc_reg_is_write_fault(exc_state.__err);
@@ -275,25 +318,36 @@ static void jl_throw_in_thread(jl_ptls_t ptls2, mach_port_t thread, jl_value_t *
     host_thread_state_t state;
     kern_return_t ret = thread_get_state(thread, MACH_THREAD_STATE, (thread_state_t)&state, &count);
     HANDLE_MACH_ERROR("thread_get_state", ret);
-    if (1) { // XXX: !jl_has_safe_restore(ptls2)
+    if (ptls2->safe_restore) {
+        jl_longjmp_in_state(&state, *ptls2->safe_restore);
+    }
+    else {
         assert(exception);
         ptls2->bt_size =
             rec_backtrace_ctx(ptls2->bt_data, JL_MAX_BT_SIZE, (bt_context_t *)&state,
-                              NULL /*current_task?*/);
+                            NULL /*current_task?*/);
         ptls2->sig_exception = exception;
+        ptls2->io_wait = 0;
+        jl_task_t *ct = ptls2->current_task;
+        jl_handler_t *eh = ct->eh;
+        if (eh != NULL) {
+            asan_unpoison_task_stack(ct, &eh->eh_ctx);
+            jl_longjmp_in_state(&state, eh->eh_ctx);
+        }
+        else {
+            jl_no_exc_handler(exception, ct);
+        }
     }
-    jl_call_in_state(ptls2, &state, &jl_sig_throw);
     ret = thread_set_state(thread, MACH_THREAD_STATE, (thread_state_t)&state, count);
     HANDLE_MACH_ERROR("thread_set_state", ret);
 }
 
 static void segv_handler(int sig, siginfo_t *info, void *context)
 {
     assert(sig == SIGSEGV || sig == SIGBUS);
-    if (jl_get_safe_restore()) { // restarting jl_ or jl_unwind_stepn
-        jl_task_t *ct = jl_get_current_task();
-        jl_ptls_t ptls = ct == NULL ? NULL : ct->ptls;
-        jl_call_in_state(ptls, (host_thread_state_t*)jl_to_bt_context(context), &jl_sig_throw);
+    jl_jmp_buf *saferestore = jl_get_safe_restore();
+    if (saferestore) { // restarting jl_ or jl_unwind_stepn
+        jl_longjmp_in_state((host_thread_state_t*)jl_to_bt_context(context), *saferestore);
         return;
     }
     jl_task_t *ct = jl_get_current_task();
@@ -352,12 +406,10 @@ kern_return_t catch_mach_exception_raise(
         jl_safe_printf("ERROR: Exception handler triggered on unmanaged thread.\n");
         return KERN_INVALID_ARGUMENT;
     }
-    // XXX: jl_throw_in_thread or segv_handler will eventually check this, but
-    //      we would like to avoid some of this work if we could detect this earlier
-    // if (jl_has_safe_restore(ptls2)) {
-    //     jl_throw_in_thread(ptls2, thread, NULL);
-    //     return KERN_SUCCESS;
-    // }
+    if (ptls2->safe_restore) {
+        jl_throw_in_thread(ptls2, thread, NULL);
+        return KERN_SUCCESS;
+    }
     if (jl_atomic_load_acquire(&ptls2->gc_state) == JL_GC_STATE_WAITING)
         return KERN_FAILURE;
     if (exception == EXC_ARITHMETIC) {
@@ -516,7 +568,6 @@ static void jl_try_deliver_sigint(void)
 
 static void JL_NORETURN jl_exit_thread0_cb(int signo)
 {
-CFI_NORETURN
     jl_critical_error(signo, 0, NULL, jl_current_task);
     jl_atexit_hook(128);
     jl_raise(signo);