judge: implement instant aborts

The way this works is:

- Worker creates a tempdir, and sets `tempfile.tempdir` to this directory.
- Worker sends back the tempdir. The parent process is responsible for cleaning
  it up when the worker exits.

Abortions are then implemented as sending `SIGKILL` to the worker.

As a side benefit of this implementation, we also get to drop the hacky
`CompiledExecutor` cache deletion.

Author: Xyene

dmoj/graders/base.py

@@ -21,14 +21,6 @@ def grade(self, case):
     def _generate_binary(self):
         raise NotImplementedError
 
-    def abort_grading(self):
-        self._abort_requested = True
-        if self._current_proc:
-            try:
-                self._current_proc.kill()
-            except OSError:
-                pass
-
     def _resolve_testcases(self, cfg, batch_no=0):
         cases = []
         for case_config in cfg:

dmoj/judge.py

@@ -2,8 +2,10 @@
 import logging
 import multiprocessing
 import os
+import shutil
 import signal
 import sys
+import tempfile
 import threading
 import traceback
 from enum import Enum
@@ -41,14 +43,13 @@ class IPC(Enum):
     BATCH_END = 'BATCH-END'
     GRADING_BEGIN = 'GRADING-BEGIN'
     GRADING_END = 'GRADING-END'
-    GRADING_ABORTED = 'GRADING-ABORTED'
     UNHANDLED_EXCEPTION = 'UNHANDLED-EXCEPTION'
-    REQUEST_ABORT = 'REQUEST-ABORT'
 
 
+class JudgeWorkerAborted(Exception):
+    pass
+
 # This needs to be at least as large as the timeout for the largest compiler time limit, but we don't enforce that here.
-# (Otherwise, aborting during a compilation that exceeds this time limit would result in a `TimeoutError` IE instead of
-# a `CompileError`.)
 IPC_TIMEOUT = 60  # seconds
 
 
@@ -128,8 +129,6 @@ def begin_grading(self, submission: Submission, report=logger.info, blocking=Fal
             )
         )
 
-        # FIXME(tbrindus): what if we receive an abort from the judge before IPC handshake completes? We'll send
-        # an abort request down the pipe, possibly messing up the handshake.
         self.current_judge_worker = JudgeWorker(submission)
 
         ipc_ready_signal = threading.Event()
@@ -147,13 +146,19 @@ def _grading_thread_main(self, ipc_ready_signal: threading.Event, report) -> Non
         assert self.current_judge_worker is not None
 
         try:
+            worker_tempdir = None
+
+            def _ipc_hello(_report, tempdir: str):
+                nonlocal worker_tempdir
+                ipc_ready_signal.set()
+                worker_tempdir = tempdir
+
             ipc_handler_dispatch: Dict[IPC, Callable] = {
-                IPC.HELLO: lambda _report: ipc_ready_signal.set(),
+                IPC.HELLO: _ipc_hello,
                 IPC.COMPILE_ERROR: self._ipc_compile_error,
                 IPC.COMPILE_MESSAGE: self._ipc_compile_message,
                 IPC.GRADING_BEGIN: self._ipc_grading_begin,
                 IPC.GRADING_END: self._ipc_grading_end,
-                IPC.GRADING_ABORTED: self._ipc_grading_aborted,
                 IPC.BATCH_BEGIN: self._ipc_batch_begin,
                 IPC.BATCH_END: self._ipc_batch_end,
                 IPC.RESULT: self._ipc_result,
@@ -176,12 +181,22 @@ def _grading_thread_main(self, ipc_ready_signal: threading.Event, report) -> Non
                     % (self.current_submission.problem_id, self.current_submission.id)
                 )
             )
+        except JudgeWorkerAborted:
+            self.packet_manager.submission_aborted_packet()
         except Exception:  # noqa: E722, we want to catch everything
             self.log_internal_error()
         finally:
             self.current_judge_worker.wait_with_timeout()
             self.current_judge_worker = None
 
+            print('cleaning up', worker_tempdir)
+            os.system('ls -al %s' % worker_tempdir)
+            if worker_tempdir:
+                try:
+                    shutil.rmtree(worker_tempdir)
+                except:  # noqa: E722
+                    pass
+
             # Might not have been set if an exception was encountered before HELLO message, so signal here to keep the
             # other side from waiting forever.
             ipc_ready_signal.set()
@@ -232,10 +247,6 @@ def _ipc_batch_begin(self, report, batch_number: int) -> None:
     def _ipc_batch_end(self, _report, _batch_number: int) -> None:
         self.packet_manager.batch_end_packet()
 
-    def _ipc_grading_aborted(self, report) -> None:
-        self.packet_manager.submission_aborted_packet()
-        report(ansi_style('#ansi[Forcefully terminating grading. Temporary files may not be deleted.](red|bold)'))
-
     def _ipc_unhandled_exception(self, _report, message: str) -> None:
         logger.error('Unhandled exception in worker process')
         self.log_internal_error(message=message)
@@ -254,10 +265,9 @@ def abort_grading(self, submission_id: Optional[int] = None) -> None:
                 'Received abortion request for %d, but %d is currently running', submission_id, worker.submission.id
             )
         else:
-            logger.info('Received abortion request for %d', worker.submission.id)
-            # These calls are idempotent, so it doesn't matter if we raced and the worker has exited already.
-            worker.request_abort_grading()
-            worker.wait_with_timeout()
+            logger.info('Received abortion request for %d, killing worker', worker.submission.id)
+            # This call is idempotent, so it doesn't matter if we raced and the worker has exited already.
+            worker.abort_grading__kill_worker()
 
     def listen(self) -> None:
         """
@@ -270,7 +280,8 @@ def murder(self) -> None:
         """
         End any submission currently executing, and exit the judge.
         """
-        self.abort_grading()
+        if self.current_judge_worker:
+            self.current_judge_worker.abort_grading__kill_worker()
         self.updater_exit = True
         self.updater_signal.set()
         if self.packet_manager:
@@ -304,8 +315,8 @@ def log_internal_error(self, exc: Optional[BaseException] = None, message: Optio
 class JudgeWorker:
     def __init__(self, submission: Submission) -> None:
         self.submission = submission
-        self._abort_requested = False
-        self._sent_sigkill_to_worker_process = False
+        self._aborted = False
+        self._timed_out = False
         # FIXME(tbrindus): marked Any pending grader cleanups.
         self.grader: Any = None
 
@@ -331,8 +342,12 @@ def communicate(self) -> Generator[Tuple[IPC, tuple], None, None]:
                 self.worker_process.kill()
                 raise
             except EOFError:
-                if self._sent_sigkill_to_worker_process:
-                    raise TimeoutError('worker did not exit in %d seconds, so it was killed' % IPC_TIMEOUT)
+                if self._aborted:
+                    raise JudgeWorkerAborted() from None
+
+                if self._timed_out:
+                    raise TimeoutError('worker did not exit in %d seconds, so it was killed' % IPC_TIMEOUT) from None
+
                 raise
             except Exception:
                 logger.error('Failed to read IPC message from worker!')
@@ -354,16 +369,14 @@ def wait_with_timeout(self) -> None:
             finally:
                 if self.worker_process.is_alive():
                     logger.error('Worker is still alive, sending SIGKILL!')
-                    self._sent_sigkill_to_worker_process = True
+                    self._timed_out = True
                     self.worker_process.kill()
 
-    def request_abort_grading(self) -> None:
-        assert self.worker_process_conn
-
-        try:
-            self.worker_process_conn.send((IPC.REQUEST_ABORT, ()))
-        except Exception:
-            logger.exception('Failed to send abort request to worker, did it race?')
+    def abort_grading__kill_worker(self) -> None:
+        if self.worker_process and self.worker_process.is_alive():
+            self._aborted = True
+            self.worker_process.kill()
+            self.worker_process.join(timeout=1)
 
     def _worker_process_main(
         self,
@@ -384,15 +397,12 @@ def _ipc_recv_thread_main() -> None:
             while True:
                 try:
                     ipc_type, data = judge_process_conn.recv()
-                except:  # noqa: E722, whatever happened, we have to abort now.
+                except:  # noqa: E722, whatever happened, we have to exit now.
                     logger.exception('Judge unexpectedly hung up!')
-                    self._do_abort()
                     return
 
                 if ipc_type == IPC.BYE:
                     return
-                elif ipc_type == IPC.REQUEST_ABORT:
-                    self._do_abort()
                 else:
                     raise RuntimeError('worker got unexpected IPC message from judge: %s' % ((ipc_type, data),))
 
@@ -402,9 +412,12 @@ def _report_unhandled_exception() -> None:
             judge_process_conn.send((IPC.UNHANDLED_EXCEPTION, (message,)))
             judge_process_conn.send((IPC.BYE, ()))
 
+        tempdir = tempfile.mkdtemp('dmoj-judge-worker')
+        tempfile.tempdir = tempdir
+
         ipc_recv_thread = None
         try:
-            judge_process_conn.send((IPC.HELLO, ()))
+            judge_process_conn.send((IPC.HELLO, (tempdir,)))
 
             ipc_recv_thread = threading.Thread(target=_ipc_recv_thread_main, daemon=True)
             ipc_recv_thread.start()
@@ -439,15 +452,6 @@ def _report_unhandled_exception() -> None:
                 if ipc_recv_thread.is_alive():
                     logger.error('Judge IPC recv thread is still alive after timeout, shutting worker down anyway!')
 
-            # FIXME(tbrindus): we need to do this because cleaning up temporary directories happens on __del__, which
-            # won't get called if we exit the process right now (so we'd leak all files created by the grader). This
-            # should be refactored to have an explicit `cleanup()` or similar, rather than relying on refcounting
-            # working out.
-            from dmoj.executors.compiled_executor import _CompiledExecutorMeta
-
-            for cached_executor in _CompiledExecutorMeta.compiled_binary_cache.values():
-                cached_executor.is_cached = False
-                cached_executor.cleanup()
             self.grader = None
 
     def _grade_cases(self) -> Generator[Tuple[IPC, tuple], None, None]:
@@ -503,11 +507,6 @@ def _grade_cases(self) -> Generator[Tuple[IPC, tuple], None, None]:
                 else:
                     result = self.grader.grade(case)
 
-                    # If the submission was killed due to a user-initiated abort, any result is meaningless.
-                    if self._abort_requested:
-                        yield IPC.GRADING_ABORTED, ()
-                        return
-
                     if result.result_flag & Result.WA:
                         # If we failed a 0-point case, we will short-circuit every case after this.
                         is_short_circuiting_enabled |= not case.points
@@ -532,11 +531,6 @@ def _grade_cases(self) -> Generator[Tuple[IPC, tuple], None, None]:
 
         yield IPC.GRADING_END, ()
 
-    def _do_abort(self) -> None:
-        self._abort_requested = True
-        if self.grader:
-            self.grader.abort_grading()
-
 
 class ClassicJudge(Judge):
     def __init__(self, host, port, **kwargs) -> None: