rbx.cp 0.5.46__py3-none-any.whl → 0.5.48__py3-none-any.whl

rbx/box/solutions.py

@@ -616,7 +616,7 @@ def _get_solution_repr(sol: Solution) -> List[Tuple[str, str]]:
     ]
 
 
-def pick_solutions(tracked_solutions: Optional[Set[str]]) -> List[str]:
+async def pick_solutions(tracked_solutions: Optional[Set[str]]) -> List[str]:
     pkg = package.find_problem_package_or_die()
     if tracked_solutions is None:
         tracked_solutions = set(str(sol.path) for sol in pkg.solutions)
@@ -630,7 +630,7 @@ def pick_solutions(tracked_solutions: Optional[Set[str]]) -> List[str]:
         if str(sol.path) in tracked_solutions
     ]
 
-    picked = questionary.checkbox('Select solutions', choices=choices).ask()
+    picked = await questionary.checkbox('Select solutions', choices=choices).ask_async()
     if picked is None:
         raise typer.Abort()
     return picked

rbx/box/statements/build_statements.py

@@ -7,7 +7,7 @@ import syncer
 import typer
 
 from rbx import annotations, console
-from rbx.box import environment, package
+from rbx.box import environment, naming, package
 from rbx.box.schema import Package
 from rbx.box.statements.builders import (
     BUILDER_LIST,
@@ -291,6 +291,7 @@ def build_statement(
         output_type=output_type,
         use_samples=use_samples,
         is_editorial=is_editorial,
+        short_name=naming.get_problem_shortname(),
     )
     statement_path = (
         package.get_build_path()

rbx/box/stresses.py

@@ -50,7 +50,6 @@ def _compile_finder(finder: CodeItem) -> str:
     return digest
 
 
-@syncer.sync
 async def run_stress(
     name: str,
     timeoutInSeconds: int,

rbx/box/tasks.py

@@ -194,7 +194,8 @@ async def _run_communication_solution_on_testcase(
             output_path = testcase.outputPath
         else:
             output_path = output_dir / testcase.inputPath.with_suffix('.out').name
-        error_path = output_path.with_suffix('.err')
+        solution_error_path = output_path.with_suffix('.sol.err')
+        interactor_error_path = output_path.with_suffix('.int.err')
         log_path = output_path.with_suffix('.log')
         output_path.parent.mkdir(parents=True, exist_ok=True)
 
@@ -204,7 +205,7 @@ async def _run_communication_solution_on_testcase(
         interactor_item = CommunicationItem(
             code=package.get_interactor(),
             executable=DigestOrSource.create(interactor_digest),
-            stderr=DigestOrDest.create(error_path),
+            stderr=DigestOrDest.create(interactor_error_path),
             extra_config=interactor_extra_config,
             extra_args='interactor.in interactor.out',
             inputs=[
@@ -230,6 +231,7 @@ async def _run_communication_solution_on_testcase(
         solution_item = CommunicationItem(
             code=solution,
             executable=DigestOrSource.create(compiled_digest),
+            stderr=DigestOrDest.create(solution_error_path),
             extra_config=extra_config,
             capture=DigestOrDest.create(solution_capture_path)
             if solution_capture_path
@@ -248,7 +250,7 @@ async def _run_communication_solution_on_testcase(
             checker_digest,
             run_log,
             interactor_run_log,
-            error_path,
+            interactor_error_path,
             testcase,
             output_path,
         )
@@ -263,7 +265,7 @@ async def _run_communication_solution_on_testcase(
             log=TestcaseLog(
                 **(run_log.model_dump() if run_log is not None else {}),
                 stdout_absolute_path=output_path.absolute(),
-                stderr_absolute_path=error_path.absolute(),
+                stderr_absolute_path=solution_error_path.absolute(),
                 log_absolute_path=log_path.absolute(),
             ),
         )

rbx/grading/judge/sandbox.py

@@ -6,6 +6,7 @@ import logging
 import os
 import pathlib
 import select
+import signal
 import stat
 import subprocess
 import sys
@@ -14,6 +15,7 @@ from typing import IO, Any, Dict, List, Optional
 
 import pydantic
 
+from rbx.grading import processing_context
 from rbx.grading.judge import cacher, storage
 
 logger = logging.getLogger(__name__)
@@ -187,6 +189,7 @@ class SandboxBase(abc.ABC):
 
     EXIT_SANDBOX_ERROR = 'sandbox error'
     EXIT_OK = 'ok'
+    EXIT_TERMINATED = 'terminated'
     EXIT_SIGNAL = 'signal'
     EXIT_TIMEOUT = 'timeout'
     EXIT_TIMEOUT_WALL = 'wall timeout'
@@ -225,6 +228,7 @@ class SandboxBase(abc.ABC):
         self.cmd_file = pathlib.PosixPath('commands.log')
 
         self.params = params or SandboxParams()
+        self.pid = None
 
         # Set common environment variables.
         # Specifically needed by Python, that searches the home for
@@ -324,6 +328,28 @@ class SandboxBase(abc.ABC):
         """
         pass
 
+    def set_pid(self, pid: int):
+        processing_context.add_to_processing_context(pid)
+        self.pid = pid
+
+    def get_pid(self) -> Optional[int]:
+        """Return the PID of the sandboxed process.
+
+        return (int|None): the PID of the sandboxed process, or None if
+            the sandboxed process is not running.
+
+        """
+        return self.pid
+
+    @abc.abstractmethod
+    def get_detailed_logs(self) -> str:
+        """Return the detailed logs of the sandbox.
+
+        return (string): the detailed logs of the sandbox.
+
+        """
+        pass
+
     @abc.abstractmethod
     def get_human_exit_description(self) -> str:
         """Get the status of the sandbox and return a human-readable
@@ -667,7 +693,9 @@ class SandboxBase(abc.ABC):
             did).
 
         """
-        return exitcode == 0
+        # SIGTERM can be safely ignored, just in case it leaks away from
+        # the sandbox.
+        return exitcode == 0 or exitcode == -signal.SIGTERM
 
     @abc.abstractmethod
     def initialize(self):

rbx/grading/judge/sandboxes/isolate.py

@@ -420,6 +420,7 @@ class IsolateSandbox(SandboxBase):
         return (string): the main reason why the sandbox terminated.
 
         """
+        # TODO: figure out EXIT_TERMINATED
         assert self.log is not None
         status_list = self.get_status_list()
         if 'XX' in status_list:
@@ -462,6 +463,14 @@ class IsolateSandbox(SandboxBase):
             return 'Execution failed because the return code was nonzero'
         return ''
 
+    def get_detailed_logs(self) -> str:
+        """Return the detailed logs of the sandbox.
+
+        return (string): the detailed logs of the sandbox.
+
+        """
+        return str(self.log)
+
     def inner_absolute_path(self, path: pathlib.Path) -> pathlib.Path:
         """Translate from a relative path inside the sandbox to an
         absolute path inside the sandbox.
@@ -602,6 +611,7 @@ class IsolateSandbox(SandboxBase):
         # std*** to interfere with command. Otherwise we let the
         # caller handle these issues.
         with popen as p:
+            self.set_pid(p.pid)
             exitcode = self.translate_box_exitcode(
                 wait_without_std([p], actually_pipe_to_stdout=self.debug)[0]
             )

rbx/grading/judge/sandboxes/stupid_sandbox.py

@@ -57,7 +57,6 @@ class StupidSandbox(SandboxBase):
         self.initialize()
 
         self.exec_num = -1
-        self.popen = None
         self.log = None
         self.returncode = None
 
@@ -192,9 +191,13 @@ class StupidSandbox(SandboxBase):
         return (string): the main reason why the sandbox terminated.
 
         """
-        if self.returncode != 0:
+        if self.returncode is not None and not self.translate_box_exitcode(
+            self.returncode
+        ):
             return self.EXIT_SANDBOX_ERROR
         status_list = self.get_status_list()
+        if 'TE' in status_list:
+            return self.EXIT_TERMINATED
         if 'WT' in status_list:
             return self.EXIT_TIMEOUT_WALL
         if 'TO' in status_list:
@@ -218,6 +221,9 @@ class StupidSandbox(SandboxBase):
         assert self.log is not None
         return int(self.log['exit-code'])
 
+    def get_detailed_logs(self) -> str:
+        return str(self.log)
+
     def get_human_exit_description(self) -> str:
         """Get the status of the sandbox and return a human-readable
         string describing it.
@@ -299,13 +305,19 @@ class StupidSandbox(SandboxBase):
             + self.get_timeit_args()
             + command
         )
-        self.returncode = subprocess.call(
+        with subprocess.Popen(
             real_command,
             stdin=subprocess.PIPE,
             stdout=subprocess.PIPE,
             stderr=subprocess.STDOUT,
             env={**os.environ, **self.params.set_env},
-        )
+        ) as p:
+            self.set_pid(p.pid)
+            try:
+                self.returncode = p.wait()
+            except Exception:
+                p.kill()
+                raise
         self.hydrate_logs()
         return self.translate_box_exitcode(self.returncode)
 

rbx/grading/judge/sandboxes/timeit.py

@@ -223,6 +223,7 @@ def wait_and_finish(
     pid: int,
     options: Options,
     start_time: float,
+    status_holder: Set[str],
     alarm_msg: Optional[List[Optional[str]]] = None,
 ):
     _, exitstatus, ru = os.wait4(pid, 0)
@@ -237,7 +238,7 @@ def wait_and_finish(
     if exitcode < 0:
         entries.append(f'exit-sig: {-exitcode}')
 
-    status = set()
+    status = status_holder
     if exitcode > 0:
         status.add('RE')
     if exitcode < 0:
@@ -284,9 +285,11 @@ def main():
             os.chdir(options.chdir)
         set_rlimits(options)
         redirect_fds(options)
+        signal.signal(signal.SIGPIPE, signal.SIG_DFL)
         os.execvp(options.argv[0], options.argv)
 
     alarm_msg: List[Optional[str]] = [None]
+    status_holder: Set[str] = set()
 
     def handle_alarm(*args, **kwargs):
         nonlocal alarm_msg
@@ -311,15 +314,21 @@ def main():
 
         signal.setitimer(signal.ITIMER_REAL, 0.3)
 
+    def handle_sub_term(*args, **kwargs):
+        nonlocal status_holder
+        status_holder.add('TE')
+        os.kill(sub_pid, 9)
+
     signal.setitimer(signal.ITIMER_REAL, 0.3)
     signal.signal(signal.SIGALRM, handle_alarm)
-    wait_and_finish(sub_pid, options, start_time, alarm_msg=alarm_msg)
+    signal.signal(signal.SIGTERM, handle_sub_term)
 
+    wait_and_finish(sub_pid, options, start_time, status_holder, alarm_msg=alarm_msg)
     # Cancel alarm before exiting to avoid surprises.
     signal.setitimer(signal.ITIMER_REAL, 0)
 
     # Exit gracefully.
-    sys.exit()
+    sys.exit(0)
 
 
 if __name__ == '__main__':

rbx/grading/processing_context.py

@@ -0,0 +1,48 @@
+import contextlib
+import os
+import signal
+import threading
+from typing import Optional, Set
+
+_processing_context_pids: Optional[Set[int]] = None
+_lock = threading.Lock()
+
+# Creating a processing context is not thread-safe, but adding to it is.
+
+
+@contextlib.contextmanager
+def new_processing_context():
+    global _processing_context_pids
+    with _lock:
+        old_processing_context_pids = _processing_context_pids
+        _processing_context_pids = set()
+    try:
+        yield
+    finally:
+        with _lock:
+            _processing_context_pids = old_processing_context_pids
+
+
+def get_processing_context() -> Set[int]:
+    with _lock:
+        return _processing_context_pids or set()
+
+
+def add_to_processing_context(pid: int):
+    global _processing_context_pids
+    with _lock:
+        if _processing_context_pids is None:
+            return
+        _processing_context_pids.add(pid)
+
+
+def terminate_all_processes_in_context():
+    with _lock:
+        if _processing_context_pids is None:
+            return
+        for pid in _processing_context_pids:
+            try:
+                os.kill(pid, signal.SIGTERM)
+            except OSError:
+                pass
+        _processing_context_pids.clear()

rbx/grading/steps.py

@@ -21,6 +21,7 @@ from rich.text import Text
 from rbx import utils
 from rbx.config import get_bits_stdcpp, get_jngen, get_testlib
 from rbx.console import console
+from rbx.grading import processing_context
 from rbx.grading.judge.sandbox import SandboxBase, SandboxParams
 from rbx.grading.judge.storage import Storage, copyfileobj
 
@@ -197,6 +198,7 @@ class RunLog(BaseModel):
     exitstatus: str = SandboxBase.EXIT_SANDBOX_ERROR
     time: Optional[float] = 0.0
     memory: Optional[int] = 0
+    sandbox: str = ''
     warnings: bool = False
     metadata: Optional[RunLogMetadata] = None
 
@@ -210,7 +212,7 @@ class RunLog(BaseModel):
             return 'OK'
         time = self.time or 0.0
         memory = self.memory or 0
-        return f'FAILED with exit code {self.exitcode} and sandbox status {self.exitstatus} (time: {time}s, memory: {memory//(1024*1024)}MB)'
+        return f'FAILED with exit code {self.exitcode} and sandbox status {self.exitstatus} (time: {time}s, memory: {memory // (1024 * 1024)}MB)'
 
 
 class PreprocessLog(RunLog):
@@ -330,7 +332,7 @@ def _expand_part(part: str, sandbox: SandboxBase) -> List[str]:
 
 def _split_and_expand(command: str, sandbox: SandboxBase) -> List[str]:
     res = []
-    parts = shlex.split(command.format(memory=sandbox.params.address_space))
+    parts = shlex.split(command.format(memory=sandbox.params.address_space or 2048))
     for part in parts:
         res.extend(_expand_part(part, sandbox))
     return res
@@ -567,6 +569,7 @@ def compile(
             memory=sandbox.get_memory_used(),
             warnings=_check_for_compilation_warnings(sandbox, stderr_file),
             log='\n'.join(std_outputs),
+            sandbox=sandbox.get_detailed_logs(),
         )
         logs.append(log)
 
@@ -594,6 +597,7 @@ async def run(
     sandbox: SandboxBase,
     artifacts: GradingArtifacts,
     metadata: Optional[RunLogMetadata] = None,
+    kill_on_processing_context_exit: bool = False,
 ) -> Optional[RunLog]:
     _process_input_artifacts(artifacts, sandbox)
     _process_fifos(artifacts, sandbox)
@@ -609,6 +613,9 @@ async def run(
         )
         return None
 
+    if sandbox.get_exit_code() != 0 and kill_on_processing_context_exit:
+        processing_context.terminate_all_processes_in_context()
+
     if not _process_output_artifacts(artifacts, sandbox):
         return None
 
@@ -625,6 +632,7 @@ async def run(
         time=sandbox.get_execution_time(),
         memory=sandbox.get_memory_used(),
         metadata=metadata,
+        sandbox=sandbox.get_detailed_logs(),
     )
     if metadata is not None and metadata.is_sanitized:
         run_log.warnings = _check_for_sanitizer_warnings(
@@ -649,19 +657,22 @@ async def run_coordinated(
     interactor: CoordinatedRunParams,
     solution: CoordinatedRunParams,
 ) -> Tuple[Optional[RunLog], Optional[RunLog]]:
-    runs = tuple(
-        run(
-            params.command,
-            params.params,
-            params.sandbox,
-            params.artifacts,
-            params.metadata,
+    with processing_context.new_processing_context():
+        runs = tuple(
+            run(
+                params.command,
+                params.params,
+                params.sandbox,
+                params.artifacts,
+                params.metadata,
+                kill_on_processing_context_exit=True,
+            )
+            for params in [interactor, solution]
+        )
+        return typing.cast(
+            Tuple[Optional[RunLog], Optional[RunLog]],
+            tuple(await asyncio.gather(*runs)),
         )
-        for params in [interactor, solution]
-    )
-    return typing.cast(
-        Tuple[Optional[RunLog], Optional[RunLog]], tuple(await asyncio.gather(*runs))
-    )
 
 
 def _normalize_checked_words(s: str) -> Tuple[str, ...]:

rbx/resources/packagers/boca/checker.sh

@@ -1,9 +1,10 @@
 ### START OF CHECKER COMPILATION
+CDIR=$(pwd)
 CHECKER_PATH="checker.cpp"
-CHECKER_OUT="../checker.exe"
+CHECKER_OUT="checker.exe"
 
 # find compiler
-cc=`which g++`
+cc=$(which g++)
 [ -x "$cc" ] || cc=/usr/bin/g++
 if [ ! -x "$cc" ]; then
     echo "$cc not found or it's not executable"
@@ -12,25 +13,26 @@ fi
 read -r -d '' TestlibContent <<"EOF"
 {{testlib_content}}
 EOF
-  
+
 read -r -d '' CheckerContent <<"EOF"
 {{checker_content}}
 EOF
 
-printf "%s" "${TestlibContent}" > testlib.h
-printf "%s" "${CheckerContent}" > $CHECKER_PATH
+printf "%s" "${TestlibContent}" >testlib.h
+printf "%s" "${CheckerContent}" >$CHECKER_PATH
 
 checker_hash=($(md5sum $CHECKER_PATH))
 checker_cache="/tmp/boca-chk-${checker_hash}"
 
 echo "Polygon checker hash: $checker_hash"
+echo "Copying polygon checker to $CDIR/$CHECKER_OUT"
 if [ -f "$checker_cache" ]; then
     echo "Recovering polygon checker from cache: $checker_cache"
     cp "$checker_cache" $CHECKER_OUT -f
 else
     echo "Compiling polygon checker: $CHECKER_PATH"
     $cc {{rbxFlags}} $CHECKER_PATH -o $CHECKER_OUT
-    
+
     if [ $? -ne 0 ]; then
         echo "Checker could not be compiled"
         exit 47

rbx/resources/packagers/boca/compare.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+if [ ! -r "$1" -o ! -r "$2" ]; then
+  echo "Parameter problem"
+  exit 43
+fi
+
+INTERACTOREXITCODE=$(grep '^testlib exitcode' "$1" | awk '{print $NF}')
+if [[ -n $INTERACTOREXITCODE ]]; then
+  echo "interactor exitcode = $INTERACTOREXITCODE"
+  if [[ $INTERACTOREXITCODE -eq 1 ]]; then
+    echo "interactor return wrong answer"
+    exit 6
+  elif [[ $INTERACTOREXITCODE -eq 2 ]]; then
+    echo "interactor invalid input"
+    exit 6
+  elif [[ $INTERACTOREXITCODE -eq 3 ]]; then
+    echo "interactor failed with exit code 3"
+    exit 43
+  else
+    echo "interactor failed with exit code $INTERACTOREXITCODE"
+    exit 47
+  fi
+fi
+
+# Next lines of this script just compares team_output and sol_output,
+# although it is possible to change them to more complex evaluations.
+output=$(./checker.exe $3 $1 $2 2>&1 >/dev/null)
+EC=$?
+
+echo "checker exitcode = $EC"
+echo "$output"
+
+if [ $EC -eq 0 ]; then
+  echo "checker found no differences"
+  exit 4
+elif [ $EC -eq 1 ]; then
+  echo "checker found differences"
+  exit 6
+elif [ $EC -eq 2 ]; then
+  echo "checker found invalid output"
+  exit 6
+elif [ $EC -eq 3 ]; then
+  echo "judge failed with $EC"
+  exit 43
+else
+  echo "judge failed with $EC"
+  exit 47
+fi