Fix linked failure with root taskgroup to kill the runtime too.

This commit is contained in:
Ben Blum 2012-07-12 02:42:56 -04:00
parent 343e9de8ef
commit 200a2ded32
6 changed files with 59 additions and 23 deletions

View file

@ -591,16 +591,21 @@ class taskgroup {
let my_pos: uint;
// let parent_group: taskgroup_arc; // TODO(bblum)
// TODO XXX bblum: add a list of empty slots to get runtime back
let mut failed: bool;
new(-tasks: taskgroup_arc, me: *rust_task, my_pos: uint) {
self.tasks = tasks; self.me = me; self.my_pos = my_pos;
self.failed = true; // This will get un-set on successful exit.
// Indicates whether this is the main (root) taskgroup. If so, failure
// here should take down the entire runtime.
let is_main: bool;
new(-tasks: taskgroup_arc, me: *rust_task, my_pos: uint, is_main: bool) {
self.tasks = tasks;
self.me = me;
self.my_pos = my_pos;
self.is_main = is_main;
}
// Runs on task exit.
drop {
if self.failed {
// If we are failing, the whole taskgroup needs to die.
if rustrt::rust_task_is_unwinding(self.me) {
// Take everybody down with us.
kill_taskgroup(self.tasks, self.me, self.my_pos);
kill_taskgroup(self.tasks, self.me, self.my_pos, self.is_main);
} else {
// Remove ourselves from the group.
leave_taskgroup(self.tasks, self.me, self.my_pos);
@ -642,7 +647,8 @@ fn leave_taskgroup(group_arc: taskgroup_arc, me: *rust_task, index: uint) {
}
// NB: Runs in destructor/post-exit context. Can't 'fail'.
fn kill_taskgroup(group_arc: taskgroup_arc, me: *rust_task, index: uint) {
fn kill_taskgroup(group_arc: taskgroup_arc, me: *rust_task, index: uint,
is_main: bool) {
// NB: We could do the killing iteration outside of the group arc, by
// having "let mut newstate" here, swapping inside, and iterating after.
// But that would let other exiting tasks fall-through and exit while we
@ -667,32 +673,40 @@ fn kill_taskgroup(group_arc: taskgroup_arc, me: *rust_task, index: uint) {
rustrt::rust_task_kill_other(task);
};
}
// Only one task should ever do this.
if is_main {
rustrt::rust_task_kill_all(me);
}
};
// (note: multiple tasks may reach this point)
};
}
fn share_parent_taskgroup() -> taskgroup_arc {
fn share_parent_taskgroup() -> (taskgroup_arc, bool) {
let me = rustrt::rust_get_task();
alt unsafe { local_get(me, taskgroup_key) } {
some(group) {
group.tasks.clone()
// Clone the shared state for the child; propagate main-ness.
(group.tasks.clone(), group.is_main)
}
none {
/* Main task, doing first spawn ever. */
// Main task, doing first spawn ever.
let tasks = arc::exclusive(some(dvec::from_elem(some(me))));
let group = @taskgroup(tasks.clone(), me, 0);
let group = @taskgroup(tasks.clone(), me, 0, true);
unsafe { local_set(me, taskgroup_key, group); }
tasks
// Tell child task it's also in the main group.
(tasks, true)
}
}
}
fn spawn_raw(opts: task_opts, +f: fn~()) {
// Decide whether the child needs to be in a new linked failure group.
let child_tg: taskgroup_arc = if opts.supervise {
let (child_tg, is_main) = if opts.supervise {
share_parent_taskgroup()
} else {
arc::exclusive(some(dvec::from_elem(none)))
// Detached from the parent group; create a new (non-main) one.
(arc::exclusive(some(dvec::from_elem(none))), false)
};
unsafe {
@ -712,7 +726,8 @@ fn spawn_raw(opts: task_opts, +f: fn~()) {
// Getting killed after here would leak the task.
let child_wrapper =
make_child_wrapper(new_task, child_tg, opts.supervise, f);
make_child_wrapper(new_task, child_tg,
opts.supervise, is_main, f);
let fptr = ptr::addr_of(child_wrapper);
let closure: *rust_closure = unsafe::reinterpret_cast(fptr);
@ -730,7 +745,8 @@ fn spawn_raw(opts: task_opts, +f: fn~()) {
}
fn make_child_wrapper(child_task: *rust_task, -child_tg: taskgroup_arc,
supervise: bool, -f: fn~()) -> fn~() {
supervise: bool, is_main: bool,
-f: fn~()) -> fn~() {
let child_tg_ptr = ~mut some(child_tg);
fn~() {
// Agh. Get move-mode items into the closure. FIXME (#2829)
@ -746,13 +762,12 @@ fn spawn_raw(opts: task_opts, +f: fn~()) {
// parent was already failing, so don't bother doing anything.
alt enlist_in_taskgroup(child_tg, child_task) {
some(my_index) {
let group = @taskgroup(child_tg, child_task, my_index);
let group =
@taskgroup(child_tg, child_task, my_index, is_main);
unsafe { local_set(child_task, taskgroup_key, group); }
// Run the child's body.
f();
// Report successful exit. (TLS cleanup code will tear
// down the group.)
group.failed = false;
// TLS cleanup code will exit the taskgroup.
}
none { }
}
@ -1006,6 +1021,7 @@ extern mod rustrt {
fn rust_task_inhibit_kill();
fn rust_task_allow_kill();
fn rust_task_kill_other(task: *rust_task);
fn rust_task_kill_all(task: *rust_task);
#[rust_stack]
fn rust_get_task_local_data(task: *rust_task) -> *libc::c_void;

View file

@ -863,6 +863,11 @@ rust_task_kill_other(rust_task *task) { /* Used for linked failure */
task->kill();
}
extern "C" void
rust_task_kill_all(rust_task *task) {
task->fail_sched_loop();
}
extern "C" rust_cond_lock*
rust_create_cond_lock() {
return new rust_cond_lock();

View file

@ -260,8 +260,8 @@ rust_task *
rust_sched_loop::create_task(rust_task *spawner, const char *name) {
rust_task *task =
new (this->kernel, "rust_task")
rust_task (this, task_state_newborn,
spawner, name, kernel->env->min_stack_size);
rust_task(this, task_state_newborn,
spawner, name, kernel->env->min_stack_size);
DLOG(this, task, "created task: " PTR ", spawner: %s, name: %s",
task, spawner ? spawner->name : "null", name);

View file

@ -129,6 +129,11 @@ cleanup_task(cleanup_args *args) {
// assert(task->task_local_data != NULL);
task->task_local_data_cleanup(task->task_local_data);
task->task_local_data = NULL;
} else if (threw_exception) {
// Edge case: If main never spawns any tasks, but fails anyway, TLS
// won't be around to take down the kernel (task.rs:kill_taskgroup,
// rust_task_kill_all). Do it here instead.
task->fail_sched_loop();
}
// FIXME (#2676): For performance we should do the annihilator
@ -282,6 +287,7 @@ rust_task::kill() {
LOG(this, task, "preparing to unwind task: 0x%" PRIxPTR, this);
}
// TODO(bblum): Move this to rust_builtin.cpp (cleanup)
extern "C" CDECL
bool rust_task_is_unwinding(rust_task *rt) {
return rt->unwinding;
@ -315,10 +321,14 @@ rust_task::begin_failure(char const *expr, char const *file, size_t line) {
#else
die();
// FIXME (#908): Need unwinding on windows. This will end up aborting
sched_loop->fail();
fail_sched_loop();
#endif
}
void rust_task::fail_sched_loop() {
sched_loop->fail();
}
void
rust_task::unsupervise()
{

View file

@ -275,6 +275,10 @@ public:
void fail();
void fail(char const *expr, char const *file, size_t line);
// Propagate failure to the entire rust runtime.
// TODO(bblum): maybe this can be done at rust-level?
void fail_sched_loop();
// Disconnect from our supervisor.
void unsupervise();

View file

@ -178,6 +178,7 @@ rust_port_task
rust_task_inhibit_kill
rust_task_allow_kill
rust_task_kill_other
rust_task_kill_all
rust_create_cond_lock
rust_destroy_cond_lock
rust_lock_cond_lock