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.
This avoids filling the memory occupied by ob_tid, ob_ref_local, and
ob_ref_shared with debug bytes (e.g., 0xDD) in mimalloc in the
free-threaded build.
This change adds an `eval_breaker` field to `PyThreadState`. The primary
motivation is for performance in free-threaded builds: with thread-local eval
breakers, we can stop a specific thread (e.g., for an async exception) without
interrupting other threads.
The source of truth for the global instrumentation version is stored in the
`instrumentation_version` field in PyInterpreterState. Threads usually read the
version from their local `eval_breaker`, where it continues to be colocated
with the eval breaker bits.
This adds a safe memory reclamation scheme based on FreeBSD's "GUS" and
quiescent state based reclamation (QSBR). The API provides a mechanism
for callers to detect when it is safe to free memory that may be
concurrently accessed by readers.
Makes _PyType_Lookup thread safe, including:
Thread safety of the underlying cache.
Make mutation of mro and type members thread safe
Also _PyType_GetMRO and _PyType_GetBases are currently returning borrowed references which aren't safe.
This is a temporary fix to unblock embedders that do not call Py_Main().
_PyInterpreterState_IsRunningMain() will always return true for the main interpreter, even in corner cases where it technically should not. The (future) full solution will do the right thing in those corner cases.
Biased reference counting maintains two refcount fields in each object:
`ob_ref_local` and `ob_ref_shared`. The true refcount is the sum of these two
fields. In some cases, when refcounting operations are split across threads,
the ob_ref_shared field can be negative (although the total refcount must be
at least zero). In this case, the thread that decremented the refcount
requests that the owning thread give up ownership and merge the refcount
fields.
This marks dead ThreadHandles as non-joinable earlier in
`PyOS_AfterFork_Child()` before we execute any Python code. The handles
are stored in a global linked list in `_PyRuntimeState` because `fork()`
affects the entire process.
For interpreters that share state with the main interpreter, this points
to the same static memory structure. For interpreters with their own
obmalloc state, it is heap allocated. Add free_obmalloc_arenas() which
will free the obmalloc arenas and radix tree structures for interpreters
with their own obmalloc state.
Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
* gh-112529: Implement GC for free-threaded builds
This implements a mark and sweep GC for the free-threaded builds of
CPython. The implementation relies on mimalloc to find GC tracked
objects (i.e., "containers").
The `--disable-gil` builds occasionally need to pause all but one thread. Some
examples include:
* Cyclic garbage collection, where this is often called a "stop the world event"
* Before calling `fork()`, to ensure a consistent state for internal data structures
* During interpreter shutdown, to ensure that daemon threads aren't accessing Python objects
This adds the following functions to implement global and per-interpreter pauses:
* `_PyEval_StopTheWorldAll()` and `_PyEval_StartTheWorldAll()` (for the global runtime)
* `_PyEval_StopTheWorld()` and `_PyEval_StartTheWorld()` (per-interpreter)
(The function names may change.)
These functions are no-ops outside of the `--disable-gil` build.
* gh-112532: Tag mimalloc heaps and pages
Mimalloc pages are data structures that contain contiguous allocations
of the same block size. Note that they are distinct from operating
system pages. Mimalloc pages are contained in segments.
When a thread exits, it abandons any segments and contained pages that
have live allocations. These segments and pages may be later reclaimed
by another thread. To support GC and certain thread-safety guarantees in
free-threaded builds, we want pages to only be reclaimed by the
corresponding heap in the claimant thread. For example, we want pages
containing GC objects to only be claimed by GC heaps.
This allows heaps and pages to be tagged with an integer tag that is
used to ensure that abandoned pages are only claimed by heaps with the
same tag. Heaps can be initialized with a tag (0-15); any page allocated
by that heap copies the corresponding tag.
* Fix conversion warning
* gh-112532: Isolate abandoned segments by interpreter
Mimalloc segments are data structures that contain memory allocations along
with metadata. Each segment is "owned" by a thread. When a thread exits,
it abandons its segments to a global pool to be later reclaimed by other
threads. This changes the pool to be per-interpreter instead of process-wide.
This will be important for when we use mimalloc to find GC objects in the
`--disable-gil` builds. We want heaps to only store Python objects from a
single interpreter. Absent this change, the abandoning and reclaiming process
could break this isolation.
* Add missing '&_mi_abandoned_default' to 'tld_empty'
* gh-112532: Use separate mimalloc heaps for GC objects
In `--disable-gil` builds, we now use four separate heaps in
anticipation of using mimalloc to find GC objects when the GIL is
disabled. To support this, we also make a few changes to mimalloc:
* `mi_heap_t` and `mi_tld_t` initialization is split from allocation.
This allows us to have a `mi_tld_t` per-`PyThreadState`, which is
important to keep interpreter isolation, since the same OS thread may
run in multiple interpreters (using different PyThreadStates.)
* Heap abandoning (mi_heap_collect_ex) can now be called from a
different thread than the one that created the heap. This is necessary
because we may clear and delete the containing PyThreadStates from a
different thread during finalization and after fork().
* Use enum instead of defines and guard mimalloc includes.
* The enum typedef will be convenient for future PRs that use the type.
* Guarding the mimalloc includes allows us to unconditionally include
pycore_mimalloc.h from other header files that rely on things like
`struct _mimalloc_thread_state`.
* Only define _mimalloc_thread_state in Py_GIL_DISABLED builds
The `PyThreadState_Clear()` function must only be called with the GIL
held and must be called from the same interpreter as the passed in
thread state. Otherwise, any Python objects on the thread state may be
destroyed using the wrong interpreter, leading to memory corruption.
This is also important for `Py_GIL_DISABLED` builds because free lists
will be associated with PyThreadStates and cleared in
`PyThreadState_Clear()`.
This fixes two places that called `PyThreadState_Clear()` from the wrong
interpreter and adds an assertion to `PyThreadState_Clear()`.
This replaces some usages of PyThread_type_lock with PyMutex, which does not require memory allocation to initialize.
This simplifies some of the runtime initialization and is also one step towards avoiding changing the default raw memory allocator during initialize/finalization, which can be non-thread-safe in some circumstances.
Every PyThreadState instance is now actually a _PyThreadStateImpl.
It is safe to cast from `PyThreadState*` to `_PyThreadStateImpl*` and back.
The _PyThreadStateImpl will contain fields that we do not want to expose
in the public C API.
Critical sections are helpers to replace the global interpreter lock
with finer grained locking. They provide similar guarantees to the GIL
and avoid the deadlock risk that plain locking involves. Critical
sections are implicitly ended whenever the GIL would be released. They
are resumed when the GIL would be acquired. Nested critical sections
behave as if the sections were interleaved.
This moves several general internal APIs out of _xxsubinterpretersmodule.c and into the new Python/crossinterp.c (and the corresponding internal headers).
Specifically:
* _Py_excinfo, etc.: the initial implementation for non-object exception snapshots (in pycore_pyerrors.h and Python/errors.c)
* _PyXI_exception_info, etc.: helpers for passing an exception beween interpreters (wraps _Py_excinfo)
* _PyXI_namespace, etc.: helpers for copying a dict of attrs between interpreters
* _PyXI_Enter(), _PyXI_Exit(): functions that abstract out the transitions between one interpreter and a second that will do some work temporarily
Again, these were all abstracted out of _xxsubinterpretersmodule.c as generalizations. I plan on proposing these as public API at some point.
This is partly to clear this stuff out of pystate.c, but also in preparation for moving some code out of _xxsubinterpretersmodule.c. This change also moves this stuff to the internal API (new: Include/internal/pycore_crossinterp.h). @vstinner did this previously and I undid it. Now I'm re-doing it. :/
This adds a new field 'state' to PyThreadState that can take on one of three values: _Py_THREAD_ATTACHED, _Py_THREAD_DETACHED, or _Py_THREAD_GC. The "attached" and "detached" states correspond closely to acquiring and releasing the GIL. The "gc" state is current unused, but will be used to implement stop-the-world GC for --disable-gil builds in the near future.
We do the following:
* add a per-interpreter XID registry (PyInterpreterState.xidregistry)
* put heap types there (keep static types in _PyRuntimeState.xidregistry)
* clear the registries during interpreter/runtime finalization
* avoid duplicate entries in the registry (when _PyCrossInterpreterData_RegisterClass() is called more than once for a type)
* use Py_TYPE() instead of PyObject_Type() in _PyCrossInterpreterData_Lookup()
The per-interpreter registry helps preserve isolation between interpreters. This is important when heap types are registered, which is something we haven't been doing yet but I will likely do soon.
Add PyThreadState_GetUnchecked() function: similar to
PyThreadState_Get(), but don't issue a fatal error if it is NULL. The
caller is responsible to check if the result is NULL. Previously,
this function was private and known as _PyThreadState_UncheckedGet().
In a few places we switch to another interpreter without knowing if it has a thread state associated with the current thread. For the main interpreter there wasn't much of a problem, but for subinterpreters we were *mostly* okay re-using the tstate created with the interpreter (located via PyInterpreterState_ThreadHead()). There was a good chance that tstate wasn't actually in use by another thread.
However, there are no guarantees of that. Furthermore, re-using an already used tstate is currently fragile. To address this, now we create a new thread state in each of those places and use it.
One consequence of this change is that PyInterpreterState_ThreadHead() may not return NULL (though that won't happen for the main interpreter).
The existence of background threads running on a subinterpreter was preventing interpreters from getting properly destroyed, as well as impacting the ability to run the interpreter again. It also affected how we wait for non-daemon threads to finish.
We add PyInterpreterState.threads.main, with some internal C-API functions.
PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock.
PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events.
This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock.
Uncontended acquisition + release:
* Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns
* macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns
* Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns
PR Overview:
The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below).
* PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now.
* _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex.
* _PyRawMutex: A word sized lock used to implement _PyParking_Lot.
* PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR.
* pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue)
---------
Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
There is a WIP proposal to enable webassembly stack switching which have been
implemented in v8:
https://github.com/WebAssembly/js-promise-integration
It is not possible to switch stacks that contain JS frames so the Emscripten JS
trampolines that allow calling functions with the wrong number of arguments
don't work in this case. However, the js-promise-integration proposal requires
the [type reflection for Wasm/JS API](https://github.com/WebAssembly/js-types)
proposal, which allows us to actually count the number of arguments a function
expects.
For better compatibility with stack switching, this PR checks if type reflection
is available, and if so we use a switch block to decide the appropriate
signature. If type reflection is unavailable, we should use the current EMJS
trampoline.
We cache the function argument counts since when I didn't cache them performance
was negatively affected.
Co-authored-by: T. Wouters <thomas@python.org>
Co-authored-by: Brett Cannon <brett@python.org>
Fix _thread.start_new_thread() race condition. If a thread is created
during Python finalization, the newly spawned thread now exits
immediately instead of trying to access freed memory and lead to a
crash.
thread_run() calls PyEval_AcquireThread() which checks if the thread
must exit. The problem was that tstate was dereferenced earlier in
_PyThreadState_Bind() which leads to a crash most of the time.
Move _PyThreadState_CheckConsistency() from thread_run() to
_PyThreadState_Bind().
thread_run() of _threadmodule.c now calls
_PyThreadState_CheckConsistency() to check if tstate is a dangling
pointer when Python is built in debug mode.
Rename ceval_gil.c is_tstate_valid() to
_PyThreadState_CheckConsistency() to reuse it in _threadmodule.c.
Symbols of the C API should be prefixed by "Py_" to avoid conflict
with existing names in 3rd party C extensions on "#include <Python.h>".
test.pythoninfo now logs Py_C_RECURSION_LIMIT constant and other
_testcapi and _testinternalcapi constants.
pycore_create_interpreter() now returns a status, rather than
calling Py_FatalError().
* PyInterpreterState_New() now calls Py_ExitStatusException() instead
of calling Py_FatalError() directly.
* Replace Py_FatalError() with PyStatus in init_interpreter() and
_PyObject_InitState().
* _PyErr_SetFromPyStatus() now raises RuntimeError, instead of
ValueError. It can now call PyErr_NoMemory(), raise MemoryError,
if it detects _PyStatus_NO_MEMORY() error message.
Python built with "configure --with-trace-refs" (tracing references)
is now ABI compatible with Python release build and debug build.
Moreover, it now also supports the Limited API.
Change Py_TRACE_REFS build:
* Remove _PyObject_EXTRA_INIT macro.
* The PyObject structure no longer has two extra members (_ob_prev
and _ob_next).
* Use a hash table (_Py_hashtable_t) to trace references (all
objects): PyInterpreterState.object_state.refchain.
* Py_TRACE_REFS build is now ABI compatible with release build and
debug build.
* Limited C API extensions can now be built with Py_TRACE_REFS:
xxlimited, xxlimited_35, _testclinic_limited.
* No longer rename PyModule_Create2() and PyModule_FromDefAndSpec2()
functions to PyModule_Create2TraceRefs() and
PyModule_FromDefAndSpec2TraceRefs().
* _Py_PrintReferenceAddresses() is now called before
finalize_interp_delete() which deletes the refchain hash table.
* test_tracemalloc find_trace() now also filters by size to ignore
the memory allocated by _PyRefchain_Trace().
Test changes for Py_TRACE_REFS:
* Add test.support.Py_TRACE_REFS constant.
* Add test_sys.test_getobjects() to test sys.getobjects() function.
* test_exceptions skips test_recursion_normalizing_with_no_memory()
and test_memory_error_in_PyErr_PrintEx() if Python is built with
Py_TRACE_REFS.
* test_repl skips test_no_memory().
* test_capi skisp test_set_nomemory().
This fixes a crasher due to a race condition, triggered infrequently when two isolated (own GIL) subinterpreters simultaneously initialize their sys or builtins modules. The crash happened due the combination of the "detached" thread state we were using and the "last holder" logic we use for the GIL. It turns out it's tricky to use the same thread state for different threads. Who could have guessed?
We solve the problem by eliminating the one object we were still sharing between interpreters. We replace it with a low-level hashtable, using the "raw" allocator to avoid tying it to the main interpreter.
We also remove the accommodations for "detached" thread states, which were a dubious idea to start with.
The _xxsubinterpreters module should not rely on internal API. Some of the functions it uses were recently moved there however. Here we move them back (and expose them properly).
Rename private C API constants:
* Rename PY_MONITORING_UNGROUPED_EVENTS to _PY_MONITORING_UNGROUPED_EVENTS
* Rename PY_MONITORING_EVENTS to _PY_MONITORING_EVENTS
Remove private _PyThreadState and _PyInterpreterState C API
functions: move them to the internal C API (pycore_pystate.h and
pycore_interp.h). Don't export most of these functions anymore, but
still export functions used by tests.
Remove _PyThreadState_Prealloc() and _PyThreadState_Init() from the C
API, but keep it in the stable API.
* Replace PyWeakref_GET_OBJECT() with _PyWeakref_GET_REF().
* _sqlite/blob.c now holds a strong reference to the blob object
while calling close_blob().
* _xidregistry_find_type() now holds a strong reference to registered
while using it.
* Add table describing possible executable classes for out-of-process debuggers.
* Remove shim code object creation code as it is no longer needed.
* Make lltrace a bit more robust w.r.t. non-standard frames.
For a while now, pending calls only run in the main thread (in the main interpreter). This PR changes things to allow any thread run a pending call, unless the pending call was explicitly added for the main thread to run.
The risk of a race with this state is relatively low, but we play it safe anyway. We do avoid using the lock in performance-sensitive cases where the risk of a race is very, very low.
This avoids the problematic race in drop_gil() by skipping the FORCE_SWITCHING code there for finalizing threads.
(The idea for this approach came out of discussions with @markshannon.)
This is the culmination of PEP 684 (and of my 8-year long multi-core Python project)!
Each subinterpreter may now be created with its own GIL (via Py_NewInterpreterFromConfig()). If not so configured then the interpreter will share with the main interpreter--the status quo since subinterpreters were added decades ago. The main interpreter always has its own GIL and subinterpreters from Py_NewInterpreter() will always share with the main interpreter.
In preparation for a per-interpreter GIL, we add PyInterpreterState.ceval.gil, set it to the shared GIL for each interpreter, and use that rather than using _PyRuntime.ceval.gil directly. Note that _PyRuntime.ceval.gil is still the actual GIL.
This function no longer makes sense, since its runtime parameter is
no longer used. Use directly _PyThreadState_GET() and
_PyInterpreterState_GET() instead.
This is strictly about moving the "obmalloc" runtime state from
`_PyRuntimeState` to `PyInterpreterState`. Doing so improves isolation
between interpreters, specifically most of the memory (incl. objects)
allocated for each interpreter's use. This is important for a
per-interpreter GIL, but such isolation is valuable even without it.
FWIW, a per-interpreter obmalloc is the proverbial
canary-in-the-coalmine when it comes to the isolation of objects between
interpreters. Any object that leaks (unintentionally) to another
interpreter is highly likely to cause a crash (on debug builds at
least). That's a useful thing to know, relative to interpreter
isolation.
We replace _PyRuntime.tstate_current with a thread-local variable. As part of this change, we add a _Py_thread_local macro in pyport.h (only for the core runtime) to smooth out the compiler differences. The main motivation here is in support of a per-interpreter GIL, but this change also provides some performance improvement opportunities.
Note that we do not provide a fallback to the thread-local, either falling back to the old tstate_current or to thread-specific storage (PyThread_tss_*()). If that proves problematic then we can circle back. I consider it unlikely, but will run the buildbots to double-check.
Also note that this does not change any of the code related to the GILState API, where it uses a thread state stored in thread-specific storage. I suspect we can combine that with _Py_tss_tstate (from here). However, that can be addressed separately and is not urgent (nor critical).
(While this change was mostly done independently, I did take some inspiration from earlier (~2020) work by @markshannon (main...markshannon:threadstate_in_tls) and @vstinner (#23976).)
* The majority of the monitoring code is in instrumentation.c
* The new instrumentation bytecodes are in bytecodes.c
* legacy_tracing.c adapts the new API to the old sys.setrace and sys.setprofile APIs
Sam Gross
Guido van Rossum
mpage
Brett Simmers
Steve Dower
Michael Droettboom
Victor Stinner
Mark Shannon
Dino Viehland
Eric Snow
Donghee Na
Andrew Rogers
Neil Schemenauer
Hugo van Kemenade
Tian Gao
Brett Cannon
Hood Chatham
Irit Katriel