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.
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.
Core static types will continue to use the global value. All other types
will use the per-interpreter value. They all share the same range, where
the global types use values < 2^16 and each interpreter uses values
higher than that.
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 function is like Py_AtExit() but for a single interpreter. This is a companion to the atexit module's register() function, taking a C callback instead of a Python one.
We also update the _xxinterpchannels module to use _Py_AtExit(), which is the motivating case. (This is inspired by pain points felt while working on gh-101660.)
We can revisit the options for keeping it global later, if desired. For now the approach seems quite complex, so we've gone with the simpler isolation solution in the meantime.
https://github.com/python/cpython/issues/100227
The essentially eliminates the global variable, with the associated benefits. This is also a precursor to isolating this bit of state to PyInterpreterState.
Folks that currently read _Py_RefTotal directly would have to start using _Py_GetGlobalRefTotal() instead.
https://github.com/python/cpython/issues/102304
We've factored out a struct from the two PyThreadState fields. This accomplishes two things:
* make it clear that the trashcan-related code doesn't need any other parts of PyThreadState
* allows us to use the trashcan mechanism even when there isn't a "current" thread state
We still expect the caller to hold the GIL.
https://github.com/python/cpython/issues/59956
The objective of this change is to help make the GILState-related code easier to understand. This mostly involves moving code around and some semantically equivalent refactors. However, there are a also a small number of slight changes in structure and behavior:
* tstate_current is moved out of _PyRuntimeState.gilstate
* autoTSSkey is moved out of _PyRuntimeState.gilstate
* autoTSSkey is initialized earlier
* autoTSSkey is re-initialized (after fork) earlier
https://github.com/python/cpython/issues/59956
* move _PyRuntime.global_objects.interned to _PyRuntime.cached_objects.interned_strings (and use _Py_CACHED_OBJECT())
* rename _PyRuntime.global_objects to _PyRuntime.static_objects
(This also relates to gh-96075.)
https://github.com/python/cpython/issues/90111
The global allocators were stored in 3 static global variables: _PyMem_Raw, _PyMem, and _PyObject. State for the "small block" allocator was stored in another 13. That makes a total of 16 global variables. We are moving all 16 to the _PyRuntimeState struct as part of the work for gh-81057. (If PEP 684 is accepted then we will follow up by moving them all to PyInterpreterState.)
https://github.com/python/cpython/issues/81057
This was added for bpo-40514 (gh-84694) to test out a per-interpreter GIL. However, it has since proven unnecessary to keep the experiment in the repo. (It can be done as a branch in a fork like normal.) So here we are removing:
* the configure option
* the macro
* the code enabled by the macro
Previously, the main interpreter was allocated on the heap during runtime initialization. Here we instead embed it into _PyRuntimeState, which means it is statically allocated as part of the _PyRuntime global. The same goes for the initial thread state (of each interpreter, including the main one). Consequently there are fewer allocations during runtime/interpreter init, fewer possible failures, and better memory locality.
FYI, this also helps efforts to consolidate globals, which in turns helps work on subinterpreter isolation.
https://bugs.python.org/issue45953
The array of small PyLong objects has been statically declared. Here I also statically initialize them. Consequently they are no longer initialized dynamically during runtime init.
I've also moved them under a new sub-struct in _PyRuntimeState, in preparation for static allocation and initialization of other global objects.
https://bugs.python.org/issue45953
This change is strictly renames and moving code around. It helps in the following ways:
* ensures type-related init functions focus strictly on one of the three aspects (state, objects, types)
* passes in PyInterpreterState * to all those functions, simplifying work on moving types/objects/state to the interpreter
* consistent naming conventions help make what's going on more clear
* keeping API related to a type in the corresponding header file makes it more obvious where to look for it
https://bugs.python.org/issue46008
Previously, basic initialization of PyInterprterState happened in PyInterpreterState_New() (along with allocation and adding the new interpreter to the runtime state). This prevented us from initializing interpreter states that were allocated separately (e.g. statically or in a free list). We've addressed that here by factoring out a separate function just for initialization. We've done the same for PyThreadState. _PyRuntimeState was sorted out when we added it since _PyRuntime is statically allocated. However, here we update the existing init code to line up with the functions for PyInterpreterState and PyThreadState.
https://bugs.python.org/issue46008
Fix the _PyUnicode_FromId() function (_Py_IDENTIFIER(var) API) when
Py_Initialize() / Py_Finalize() is called multiple times:
preserve _PyRuntime.unicode_ids.next_index value.
Use _PyRuntimeState_INIT macro instead memset(0) to reset
_PyRuntimeState members to zero.
Make _PyUnicode_FromId() function compatible with subinterpreters.
Each interpreter now has an array of identifier objects (interned
strings decoded from UTF-8).
* Add PyInterpreterState.unicode.identifiers: array of identifiers
objects.
* Add _PyRuntimeState.unicode_ids used to allocate unique indexes
to _Py_Identifier.
* Rewrite the _Py_Identifier structure.
Microbenchmark on _PyUnicode_FromId(&PyId_a) with _Py_IDENTIFIER(a):
[ref] 2.42 ns +- 0.00 ns -> [atomic] 3.39 ns +- 0.00 ns: 1.40x slower
This change adds 1 ns per _PyUnicode_FromId() call in average.
PyOS_AfterFork_Child() helper functions now return a PyStatus:
PyOS_AfterFork_Child() is now responsible to handle errors.
* Move _PySignal_AfterFork() to the internal C API
* Add #ifdef HAVE_FORK on _PyGILState_Reinit(), _PySignal_AfterFork()
and _PyInterpreterState_DeleteExceptMain().
In the experimental isolated subinterpreters build mode, the GIL is
now per-interpreter.
Move gil from _PyRuntimeState.ceval to PyInterpreterState.ceval.
new_interpreter() always get the config from the main interpreter.
Move recursion_limit member from _PyRuntimeState.ceval to
PyInterpreterState.ceval.
* Py_SetRecursionLimit() now only sets _Py_CheckRecursionLimit
of ceval.c if the current Python thread is part of the main
interpreter.
* Inline _Py_MakeEndRecCheck() into _Py_LeaveRecursiveCall().
* Convert _Py_RecursionLimitLowerWaterMark() macro into a static
inline function.
Victor Stinner
Pablo Galindo Salgado
Eric Snow
Gabriele N. Tornetta
Виталий Дмитриев
Kumar Aditya
Mark Shannon