We were not properly accounting for interpreter memory leaks at
shutdown and had two sources of leaks:
* Objects that use deferred reference counting and were reachable via
static types outlive the final GC. We now disable deferred reference
counting on all objects if we are calling the GC due to interpreter
shutdown.
* `_PyMem_FreeDelayed` did not properly check for interpreter shutdown
so we had some memory blocks that were enqueued to be freed, but
never actually freed.
* `_PyType_FinalizeIdPool` wasn't called at interpreter shutdown.
The free-threaded build partially stores heap type reference counts in
distributed manner in per-thread arrays. This avoids reference count
contention when creating or destroying instances.
Co-authored-by: Ken Jin <kenjin@python.org>
This combines and updates our freelist handling to use a consistent
implementation. Objects in the freelist are linked together using the
first word of memory block.
If configured with freelists disabled, these operations are essentially
no-ops.
The `_PyThreadState_Bind()` function is called before the first
`PyEval_AcquireThread()` so it's not synchronized with the stop the
world GC. We had a race where `gc_visit_heaps()` might visit a thread's
heap while it's being initialized.
Use a simple atomic int to avoid visiting heaps for threads that are not
yet fully initialized (i.e., before `tstate_mimalloc_bind()` is called).
The race was reproducible by running:
`python Lib/test/test_importlib/partial/pool_in_threads.py`.
The free-threaded build currently immortalizes objects that use deferred
reference counting (see gh-117783). This typically happens once the
first non-main thread is created, but the behavior can be suppressed for
tests, in subinterpreters, or during a compile() call.
This fixes a race condition involving the tracking of whether the
behavior is suppressed.
Release the GIL before calling `_Py_qsbr_unregister`.
The deadlock could occur when the GIL was enabled at runtime. The
`_Py_qsbr_unregister` call might block while holding the GIL because the
thread state was not active, but the GIL was still held.
Make sure that `gilstate_counter` is not zero in when calling
`PyThreadState_Clear()`. A destructor called from `PyThreadState_Clear()` may
call back into `PyGILState_Ensure()` and `PyGILState_Release()`. If
`gilstate_counter` is zero, it will try to create a new thread state before
the current active thread state is destroyed, leading to an assertion failure
or crash.
`drop_gil()` assumes that its caller is attached, which means that the current
thread holds the GIL if and only if the GIL is enabled, and the enabled-state
of the GIL won't change. This isn't true, though, because `detach_thread()`
calls `_PyEval_ReleaseLock()` after detaching and
`_PyThreadState_DeleteCurrent()` calls it after removing the current thread
from consideration for stop-the-world requests (effectively detaching it).
Fix this by remembering whether or not a thread acquired the GIL when it last
attached, in `PyThreadState._status.holds_gil`, and check this in `drop_gil()`
instead of `gil->enabled`.
This fixes a crash in `test_multiprocessing_pool_circular_import()`, so I've
reenabled it.
`_Py_qsbr_unregister` is called when the PyThreadState is already
detached, so the access to `tstate->qsbr` isn't safe without locking the
shared mutex. Grab the `struct _qsbr_shared` from the interpreter
instead.
Add the ability to enable/disable the GIL at runtime, and use that in
the C module loading code.
We can't know before running a module init function if it supports
free-threading, so the GIL is temporarily enabled before doing so. If
the module declares support for running without the GIL, the GIL is
later disabled. Otherwise, the GIL is permanently enabled, and will
never be disabled again for the life of the current interpreter.
We already intern and immortalize most string constants. In the
free-threaded build, other constants can be a source of reference count
contention because they are shared by all threads running the same code
objects.
The module itself is a thin wrapper around calls to functions in
`Python/codecs.c`, so that's where the meaningful changes happened:
- Move codecs-related state that lives on `PyInterpreterState` to a
struct declared in `pycore_codecs.h`.
- In free-threaded builds, add a mutex to `codecs_state` to synchronize
operations on `search_path`. Because `search_path_mutex` is used as a
normal mutex and not a critical section, we must be extremely careful
with operations called while holding it.
- The codec registry is explicitly initialized as part of
`_PyUnicode_InitEncodings` to simplify thread-safety.
The code for Tier 2 is now only compiled when configured
with `--enable-experimental-jit[=yes|interpreter]`.
We drop support for `PYTHON_UOPS` and -`Xuops`,
but you can disable the interpreter or JIT
at runtime by setting `PYTHON_JIT=0`.
You can also build it without enabling it by default
using `--enable-experimental-jit=yes-off`;
enable with `PYTHON_JIT=1`.
On Windows, the `build.bat` script supports
`--experimental-jit`, `--experimental-jit-off`,
`--experimental-interpreter`.
In the C code, `_Py_JIT` is defined as before
when the JIT is enabled; the new variable
`_Py_TIER2` is defined when the JIT *or* the
interpreter is enabled. It is actually a bitmask:
1: JIT; 2: default-off; 4: interpreter.
Avoid detaching thread state when stopping the world. When re-attaching
the thread state, the thread would attempt to resume the top-most
critical section, which might now be held by a thread paused for our
stop-the-world request.
Deferred reference counting is not fully implemented yet. As a temporary
measure, we immortalize objects that would use deferred reference
counting to avoid multi-threaded scaling bottlenecks.
This is only performed in the free-threaded build once the first
non-main thread is started. Additionally, some tests, including refleak
tests, suppress this behavior.
Makes sys.settrace, sys.setprofile, and monitoring generally thread-safe.
Mostly uses a stop-the-world approach and synchronization around the code object's _co_instrumentation_version. There may be a little bit of extra synchronization around the monitoring data that's required to be TSAN clean.
TSAN erroneously reports a data race between the `_Py_atomic_compare_exchange_int`
on `tstate->state` in `tstate_try_attach()` and the non-atomic load of
`tstate->state` in `start_the_world`. The `_Py_atomic_compare_exchange_int` fails,
but TSAN erroneously treats it as a store.
This is similar to the situation with threading._DummyThread. The methods (incl. __del__()) of interpreters.Interpreter objects must be careful with interpreters not created by interpreters.create(). The simplest thing to start with is to disable any method that modifies or runs in the interpreter. As part of this, the runtime keeps track of where an interpreter was created. We also handle interpreter "refcounts" properly.
This keeps track of the per-thread total reference count operations in
PyThreadState in the free-threaded builds. The count is merged into the
interpreter's total when the thread exits.
Most mutable data is protected by a striped lock that is keyed on the
referenced object's address. The weakref's hash is protected using the
weakref's per-object lock.
Note that this only affects free-threaded builds. Apart from some minor
refactoring, the added code is all either gated by `ifdef`s or is a no-op
(e.g. `Py_BEGIN_CRITICAL_SECTION`).
I had meant to switch everything to InterpreterError when I added it a while back. At the time I missed a few key spots.
As part of this, I've added print-the-exception to _PyXI_InitTypes() and fixed an error case in `_PyStaticType_InitBuiltin().
This fixes a crash in `test_threading.test_reinit_tls_after_fork()` when
running with the GIL disabled. We already properly handle the case where
the thread state is `_Py_THREAD_ATTACHED` in `tstate_delete_common()` --
we just need to remove an assertion.
Keeping the thread attached means that a stop-the-world pause, such as
for a `fork()`, won't commence until we remove our thread state from the
interpreter's linked list. This prevents a crash when the child process
tries to clean up the dead thread states.
This adds a stop the world pause to make the two functions thread-safe
when the GIL is disabled in the free-threaded build.
Additionally, the main test thread may call `sys._current_exceptions()` as
soon as `g_raised.set()` is called. The background thread may not yet reach
the `leave_g.wait()` line.
When I added _PyInterpreterState_IsRunningMain() and friends last year, I tried to accommodate applications that embed Python but don't call _PyInterpreterState_SetRunningMain() (not that they're expected to). That mostly worked fine until my recent changes in gh-117049, where the subtleties with the fallback code led to failures; the change ended up breaking test_tools.test_freeze, which exercises a basic embedding situation.
The simplest fix is to drop the fallback code I originally added to _PyInterpreterState_IsRunningMain() (and later to _PyThreadState_IsRunningMain()). I've kept the fallback in the _xxsubinterpreters module though. I've also updated Py_FrozenMain() to call _PyInterpreterState_SetRunningMain().
Split `_PyThreadState_DeleteExcept` into two functions:
- `_PyThreadState_RemoveExcept` removes all thread states other than one
passed as an argument. It returns the removed thread states as a
linked list.
- `_PyThreadState_DeleteList` deletes those dead thread states. It may
call destructors, so we want to "start the world" before calling
`_PyThreadState_DeleteList` to avoid potential deadlocks.
I added it quite a while ago as a strategy for managing interpreter lifetimes relative to the PEP 554 (now 734) implementation. Relatively recently I refactored that implementation to no longer rely on InterpreterID objects. Thus now I'm removing it.
Mostly we unify the two different implementations of the conversion code (from PyObject * to int64_t. We also drop the PyArg_ParseTuple()-style converter function, as well as rename and move PyInterpreterID_LookUp().
This changes the free-threaded build to perform a stop-the-world pause
before deleting other thread states when forking and during shutdown.
This fixes some crashes when using multiprocessing and during shutdown
when running with `PYTHON_GIL=0`.
This also changes `PyOS_BeforeFork` to acquire the runtime lock
(i.e., `HEAD_LOCK(&_PyRuntime)`) before forking to ensure that data
protected by the runtime lock (and not just the GIL or stop-the-world)
is in a consistent state before forking.
Somehow we ended up with two separate counter variables tracking "the next function version".
Most likely this was a historical accident where an old branch was updated incorrectly.
This PR merges the two counters into a single one: `interp->func_state.next_version`.
There is a race between when `Thread._tstate_lock` is released[^1] in `Thread._wait_for_tstate_lock()`
and when `Thread._stop()` asserts[^2] that it is unlocked. Consider the following execution
involving threads A, B, and C:
1. A starts.
2. B joins A, blocking on its `_tstate_lock`.
3. C joins A, blocking on its `_tstate_lock`.
4. A finishes and releases its `_tstate_lock`.
5. B acquires A's `_tstate_lock` in `_wait_for_tstate_lock()`, releases it, but is swapped
out before calling `_stop()`.
6. C is scheduled, acquires A's `_tstate_lock` in `_wait_for_tstate_lock()` but is swapped
out before releasing it.
7. B is scheduled, calls `_stop()`, which asserts that A's `_tstate_lock` is not held.
However, C holds it, so the assertion fails.
The race can be reproduced[^3] by inserting sleeps at the appropriate points in
the threading code. To do so, run the `repro_join_race.py` from the linked repo.
There are two main parts to this PR:
1. `_tstate_lock` is replaced with an event that is attached to `PyThreadState`.
The event is set by the runtime prior to the thread being cleared (in the same
place that `_tstate_lock` was released). `Thread.join()` blocks waiting for the
event to be set.
2. `_PyInterpreterState_WaitForThreads()` provides the ability to wait for all
non-daemon threads to exit. To do so, an `is_daemon` predicate was added to
`PyThreadState`. This field is set each time a thread is created. `threading._shutdown()`
now calls into `_PyInterpreterState_WaitForThreads()` instead of waiting on
`_tstate_lock`s.
[^1]: 441affc9e7/Lib/threading.py (L1201)
[^2]: 441affc9e7/Lib/threading.py (L1115)
[^3]: 8194653279
---------
Co-authored-by: blurb-it[bot] <43283697+blurb-it[bot]@users.noreply.github.com>
Co-authored-by: Antoine Pitrou <antoine@python.org>
This moves `current_fast_clear()` up so that the current thread state is
`NULL` while running `tstate_delete_common()`.
This doesn't fix any bugs, but it means that we are more consistent that
`_PyThreadState_GET() != NULL` means that the thread is "attached".
This implements the delayed reuse of mimalloc pages that contain Python
objects in the free-threaded build.
Allocations of the same size class are grouped in data structures called
pages. These are different from operating system pages. For thread-safety, we
want to ensure that memory used to store PyObjects remains valid as long as
there may be concurrent lock-free readers; we want to delay using it for
other size classes, in other heaps, or returning it to the operating system.
When a mimalloc page becomes empty, instead of immediately freeing it, we tag
it with a QSBR goal and insert it into a per-thread state linked list of
pages to be freed. When mimalloc needs a fresh page, we process the queue and
free any still empty pages that are now deemed safe to be freed. Pages
waiting to be freed are still available for allocations of the same size
class and allocating from a page prevent it from being freed. There is
additional logic to handle abandoned pages when threads exit.
A previous commit introduced a bug to `interpreter_clear()`: it set
`interp->ceval.instrumentation_version` to 0, without making the corresponding
change to `tstate->eval_breaker` (which holds a thread-local copy of the
version). After this happens, Python code can still run due to object finalizers
during a GC, and the version check in bytecodes.c will see a different result
than the one in instrumentation.c causing an infinite loop.
The fix itself is straightforward: clear `tstate->eval_breaker` when clearing
`interp->ceval.instrumentation_version`.
This adds `_PyMem_FreeDelayed()` and supporting functions. The
`_PyMem_FreeDelayed()` function frees memory with the same allocator as
`PyMem_Free()`, but after some delay to ensure that concurrent lock-free
readers have finished.
Sam Gross
mpage
Ken Jin
Eric Snow
Kirill Podoprigora
Brett Simmers
Alex Turner
Dino Viehland
Guido van Rossum
Mark Shannon
Steve Dower
Michael Droettboom