signal handling: User-defined interrupt handlers

Interrupt handling is a tricky problem, not just in terms of
implementation, but in terms of desired behavior: when an interrupt is
received, which code should handle it? Julia's current answer to this is
effectively to throw an `InterruptException` to the first task to hit a
safepoint. While this seems sensible (the code that's running gets
interrupted), it only really works for very basic numerical code. In the case that
multiple tasks are running concurrently, or when try-catch handlers are
registered, this system breaks down, and results in unpredictable
behavior. This unpredictable behavior includes:

- Interrupting background/runtime tasks which don't want to be
  interrupted, as they do little bits of important work (and are
  critical to library runtime functionality)
- Interrupting only one task, when multiple coordinating tasks would
  want to receive the interrupt to safely terminate a computation
- Interrupting only one library's task, when multiple libraries really
  would want to be notified about the interrupt

The above behavior makes it nearly impossible to provide reliable Ctrl-C
behavior, and results in very confused users who get stuck hitting
Ctrl-C continuously, sometimes getting caught in a hang, sometimes
triggering unrelated exception handling code they didn't mean to,
sometimes getting a segfault, and very rarely getting the behavior they
desire (with unpredictable safety of being able to continue using the
active session as intended).

This commit provides an alternative behavior for interrupts which is
more predictable: user code may now register tasks as "interrupt
handlers", which will be guaranteed to receive an `InterruptException`
whenever the session receives an interrupt signal. Additionally, when
any interrupt handlers are registered, no other tasks will receive
`InterruptException`s; only the handlers may receive them.

This behavior allows one or more libraries to register handler tasks which
will all be concurrently awoken to handle each interrupt and do whatever
is necessary to safely interrupt any running code; the extent to which
other tasks are interrupted is arbitrary and library-defined. For
example, GPU libraries like AMDGPU.jl can register a handler to safely
interrupt GPU kernels running on all GPU queues and do resource cleanup.
Concurrently, a complex runtime like the scheduler in Dagger.jl can
register a handler to interrupt running tasks on other workers when
possible, and otherwise notify the user that tasks are being shutdown.

This change is intended to be non-breaking for simple
codes: the previous behavior is maintained when no interrupt handlers
are registered. However, once some libraries start adding interrupt
handlers, other libraries will need to follow suit to ensure that users
can interrupt their computations.

base/task.jl

@@ -992,3 +992,17 @@ if Sys.iswindows()
 else
     pause() = ccall(:pause, Cvoid, ())
 end
+
+interrupt_handlers() = ccall(:jl_get_interrupt_handlers, Any, ())::Vector{Task}
+function register_interrupt_handler(t::Task)
+    handlers = interrupt_handlers()
+    if findfirst(==(t), handlers) === nothing
+        push!(handlers, t)
+    end
+    return
+end
+function unregister_interrupt_handler(t::Task)
+    handlers = interrupt_handlers()
+    deleteat!(handlers, findall(==(t), handlers))
+    return
+end

src/gc.c

@@ -2891,6 +2891,7 @@ static void gc_mark_roots(jl_gc_markqueue_t *mq)
     gc_try_claim_and_push(mq, jl_emptytuple_type, NULL);
     gc_try_claim_and_push(mq, cmpswap_names, NULL);
     gc_try_claim_and_push(mq, jl_global_roots_table, NULL);
+    gc_try_claim_and_push(mq, jl_interrupt_handlers, NULL);
 }
 
 // find unmarked objects that need to be finalized from the finalizer list "list".

src/jl_exported_data.inc

@@ -54,6 +54,7 @@
     XX(jl_int8_type) \
     XX(jl_interconditional_type) \
     XX(jl_interrupt_exception) \
+    XX(jl_interrupt_handlers) \
     XX(jl_intrinsic_type) \
     XX(jl_kwcall_func) \
     XX(jl_lineinfonode_type) \

src/jl_exported_funcs.inc

@@ -220,6 +220,7 @@
     XX(jl_get_field) \
     XX(jl_get_global) \
     XX(jl_get_image_file) \
+    XX(jl_get_interrupt_handlers) \
     XX(jl_get_JIT) \
     XX(jl_get_julia_bin) \
     XX(jl_get_julia_bindir) \

src/julia_threads.h

@@ -369,6 +369,8 @@ JL_DLLEXPORT int8_t jl_gc_is_in_finalizer(void);
 
 JL_DLLEXPORT void jl_wakeup_thread(int16_t tid);
 
+JL_DLLEXPORT void jl_schedule_task(struct _jl_task_t *task);
+
 #ifdef __cplusplus
 }
 #endif

src/signal-handling.c

@@ -304,6 +304,55 @@ static void jl_check_profile_autostop(void)
     }
 }
 
+JL_DLLEXPORT _Atomic(jl_array_t *) jl_interrupt_handlers = NULL;
+JL_DLLEXPORT jl_array_t *jl_get_interrupt_handlers(void)
+{
+    jl_array_t *handlers = jl_atomic_load_relaxed(&jl_interrupt_handlers);
+    if (!handlers) {
+        static jl_datatype_t *jl_array_task_type;
+        if (!jl_array_task_type)
+            jl_array_task_type = (jl_datatype_t *)jl_apply_type2((jl_value_t*)jl_array_type, (jl_value_t*)jl_task_type, jl_box_long(1));
+        jl_array_t *new_handlers = jl_alloc_array_1d((jl_value_t *)jl_array_task_type, 0);
+        if (jl_atomic_cmpswap(&jl_interrupt_handlers, &handlers, new_handlers)) {
+            handlers = new_handlers;
+        } else {
+            handlers = jl_atomic_load_relaxed(&jl_interrupt_handlers);
+        }
+    }
+    assert(handlers);
+    return handlers;
+}
+static _Atomic(int) handle_interrupt = 0;
+JL_DLLEXPORT void jl_schedule_interrupt_handlers(void)
+{
+    if (jl_atomic_exchange_relaxed(&handle_interrupt, 0) != 1)
+        return;
+    jl_array_t *handlers = jl_atomic_load_relaxed(&jl_interrupt_handlers);
+    if (!handlers)
+        return;
+    for (int i = 0; i < jl_array_len(handlers); i++) {
+        jl_task_t *handler = ((jl_task_t **)jl_array_data(handlers))[i];
+        assert(jl_is_task(handler));
+        if (handler->ptls)
+            continue;
+        if (jl_atomic_load_relaxed(&handler->_state) != JL_TASK_STATE_RUNNABLE)
+            continue;
+        handler->result = jl_interrupt_exception;
+        handler->_isexception = 1;
+        jl_schedule_task(handler);
+    }
+}
+static int want_interrupt_handlers(void)
+{
+    jl_array_t *handlers = jl_atomic_load_relaxed(&jl_interrupt_handlers);
+    if (handlers && (jl_array_len(handlers) > 0)) {
+        // Set flag to trigger user handlers on next task switch
+        jl_atomic_store_relaxed(&handle_interrupt, 1);
+        return 1;
+    }
+    return 0;
+}
+
 #if defined(_WIN32)
 #include "signals-win.c"
 #else

src/signals-unix.c

@@ -525,11 +525,14 @@ void usr2_handler(int sig, siginfo_t *info, void *ctx)
     jl_atomic_exchange(&ptls->signal_request, 0); // returns -1
     if (request == 2) {
         int force = jl_check_force_sigint();
+        if (!force && want_interrupt_handlers()) {
+            return;
+        }
         if (force || (!ptls->defer_signal && ptls->io_wait)) {
             jl_safepoint_consume_sigint();
+            // Force a throw
             if (force)
                 jl_safe_printf("WARNING: Force throwing a SIGINT\n");
-            // Force a throw
             jl_clear_force_sigint();
             jl_throw_in_ctx(ct, jl_interrupt_exception, sig, ctx);
         }
@@ -767,7 +770,7 @@ static void *signal_listener(void *arg)
         profile = (sig == SIGUSR1);
 #if defined(_POSIX_C_SOURCE) && _POSIX_C_SOURCE >= 199309L
         if (profile && !(info.si_code == SI_TIMER &&
-	            info.si_value.sival_ptr == &timerprof))
+            info.si_value.sival_ptr == &timerprof))
             profile = 0;
 #endif
 #endif
@@ -780,6 +783,10 @@ static void *signal_listener(void *arg)
             else if (exit_on_sigint) {
                 critical = 1;
             }
+            // FIXME: Skip this if force
+            else if (want_interrupt_handlers()) {
+                continue;
+            }
             else {
                 jl_try_deliver_sigint();
                 continue;

src/signals-win.c

@@ -221,7 +221,8 @@ static BOOL WINAPI sigint_handler(DWORD wsig) //This needs winapi types to guara
     if (!jl_ignore_sigint()) {
         if (exit_on_sigint)
             jl_exit(128 + sig); // 128 + SIGINT
-        jl_try_deliver_sigint();
+        if (!want_interrupt_handlers())
+            jl_try_deliver_sigint();
     }
     return 1;
 }

src/task.c

@@ -621,8 +621,12 @@ JL_NO_ASAN static void ctx_switch(jl_task_t *lastt)
     sanitizer_finish_switch_fiber(ptls->previous_task, jl_atomic_load_relaxed(&ptls->current_task));
 }
 
+JL_DLLIMPORT void jl_schedule_interrupt_handlers(void);
+
 JL_DLLEXPORT void jl_switch(void) JL_NOTSAFEPOINT_LEAVE JL_NOTSAFEPOINT_ENTER
 {
+    jl_schedule_interrupt_handlers();
+
     jl_task_t *ct = jl_current_task;
     jl_ptls_t ptls = ct->ptls;
     jl_task_t *t = ptls->next_task;
@@ -1164,7 +1168,7 @@ JL_DLLEXPORT void jl_task_wait()
     jl_apply(&wait_func, 1);
     ct->world_age = last_age;
 }
-
+#endif
 JL_DLLEXPORT void jl_schedule_task(jl_task_t *task)
 {
     static jl_function_t *sched_func = NULL;
@@ -1178,7 +1182,6 @@ JL_DLLEXPORT void jl_schedule_task(jl_task_t *task)
     jl_apply(args, 2);
     ct->world_age = last_age;
 }
-#endif
 
 // Do one-time initializations for task system
 void jl_init_tasks(void) JL_GC_DISABLED

src/threading.h

@@ -27,6 +27,8 @@ jl_ptls_t jl_init_threadtls(int16_t tid) JL_NOTSAFEPOINT;
 void jl_init_threadinginfra(void);
 void jl_threadfun(void *arg);
 
+extern _Atomic(jl_array_t *) jl_interrupt_handlers;
+
 #ifdef __cplusplus
 }
 #endif