Isolate the tracemalloc Module Between Interpreters

[ericsnowcurrently]

(See #100227.)

Currently the tracemalloc module has some state in _PyRuntimeState, including objects, which is shared by all interpreters. Interpreters should be isolated from each other, for a variety of reasons (including the possibility of a per-interpreter GIL). Isolating the module will involve moving some of the state to PyInterpreterState (and, under per-interpreter GIL, guarding other state with a global lock).

---

Analysis:

(expand)

Allocators

The module installs a custom allocator (using the PEP 445 API
(docs).

(expand)

the allocator functions:

domain	func	wraps	wraps (reentrant)	actually wraps
-	tracemalloc_alloc()	<original malloc() or calloc()>	<---
-	tracemalloc_realloc()	<original realloc() or free()>	<---
-	tracemalloc_raw_alloc()	tracemalloc_alloc() *	<original malloc() or calloc()>
-	tracemalloc_alloc_gil()	tracemalloc_alloc()	<original malloc() or calloc()>
raw
	tracemalloc_raw_malloc()	tracemalloc_alloc()	<original malloc()>	tracemalloc_raw_alloc()
	tracemalloc_raw_calloc()	tracemalloc_alloc()	<original calloc()>	tracemalloc_raw_alloc()
	tracemalloc_raw_realloc()	tracemalloc_realloc() *	<original realloc()>
	tracemalloc_free()	<original free()>	<---
mem
obj
	tracemalloc_malloc_gil()	tracemalloc_alloc_gil()	<---
	tracemalloc_calloc_gil()	tracemalloc_alloc_gil()	<---
	tracemalloc_realloc_gil()	tracemalloc_realloc()	<original realloc()>
	tracemalloc_free()	<original free()>	<---

* Note that tracemalloc_raw_alloc() wraps the tracemalloc_alloc() call
with PyGILState_Ensure()/PyGILState_Release().
Likewise for tracemalloc_raw_realloc() where it calls tracemalloc_realloc().
In no other case does an allocator function use the GILState API for any calls.

State

Fields

(expand)

https://github.com/python/cpython/blob/main/Include/internal/pycore_tracemalloc.h#L57-L107
https://github.com/python/cpython/blob/main/Include/internal/pycore_runtime.h#L147

raw

cpython/Include/internal/pycore_tracemalloc.h

Lines 43 to 55 in 0675b8f

	struct
	#ifdef __GNUC__
	__attribute__((packed))
	#endif
	tracemalloc_frame {
	/* filename cannot be NULL: "<unknown>" is used if the Python frame
	filename is NULL */
	PyObject *filename;
	unsigned int lineno;
	};
	#ifdef _MSC_VER
	#pragma pack(pop)
	#endif

cpython/Include/internal/pycore_tracemalloc.h

Lines 57 to 64 in 0675b8f

	struct tracemalloc_traceback {
	Py_uhash_t hash;
	/* Number of frames stored */
	uint16_t nframe;
	/* Total number of frames the traceback had */
	uint16_t total_nframe;
	struct tracemalloc_frame frames[1];
	};

cpython/Include/internal/pycore_tracemalloc.h

Lines 57 to 107 in 0675b8f

	struct tracemalloc_traceback {
	Py_uhash_t hash;
	/* Number of frames stored */
	uint16_t nframe;
	/* Total number of frames the traceback had */
	uint16_t total_nframe;
	struct tracemalloc_frame frames[1];
	};
	struct _tracemalloc_runtime_state {
	struct _PyTraceMalloc_Config config;
	/* Protected by the GIL */
	struct {
	PyMemAllocatorEx mem;
	PyMemAllocatorEx raw;
	PyMemAllocatorEx obj;
	} allocators;
	#if defined(TRACE_RAW_MALLOC)
	PyThread_type_lock tables_lock;
	#endif
	/* Size in bytes of currently traced memory.
	Protected by TABLES_LOCK(). */
	size_t traced_memory;
	/* Peak size in bytes of traced memory.
	Protected by TABLES_LOCK(). */
	size_t peak_traced_memory;
	/* Hash table used as a set to intern filenames:
	PyObject* => PyObject*.
	Protected by the GIL */
	_Py_hashtable_t *filenames;
	/* Buffer to store a new traceback in traceback_new().
	Protected by the GIL. */
	struct tracemalloc_traceback *traceback;
	/* Hash table used as a set to intern tracebacks:
	traceback_t* => traceback_t*
	Protected by the GIL */
	_Py_hashtable_t *tracebacks;
	/* pointer (void*) => trace (trace_t*).
	Protected by TABLES_LOCK(). */
	_Py_hashtable_t *traces;
	/* domain (unsigned int) => traces (_Py_hashtable_t).
	Protected by TABLES_LOCK(). */
	_Py_hashtable_t *domains;
	struct tracemalloc_traceback empty_traceback;
	Py_tss_t reentrant_key;
	};

cpython/Modules/_tracemalloc.c

Lines 54 to 55 in 0675b8f

	typedef struct tracemalloc_frame frame_t;
	typedef struct tracemalloc_traceback traceback_t;

cpython/Modules/_tracemalloc.c

Lines 69 to 76 in 0675b8f

	/* Trace of a memory block */
	typedef struct {
	/* Size of the memory block in bytes */
	size_t size;
	/* Traceback where the memory block was allocated */
	traceback_t *traceback;
	} trace_t;

name	type	protected by	#ifdef	notes
config	struct _PyTraceMalloc_Config
. initialized	enum {}	GIL
. tracing	bool	GIL
. max_nframe	int	GIL
allocators				see PEP 445 (docs)
. mem	PyMemAllocatorEx	GIL
. raw	PyMemAllocatorEx	GIL
. obj	PyMemAllocatorEx	GIL
tables_lock	PyThread_type_lock	GIL	TRACE_RAW_MALLOC
traced_memory	size_t	tables_lock
peak_traced_memory	size_t	tables_lock
filenames	_Py_hashtable_t *	GIL		interned; effectively a set of objects
traceback	struct tracemalloc_traceback *	GIL		a temporary buffer
tracebacks	_Py_hashtable_t *	GIL		interned; effectively a set of traceback_t
traces	_Py_hashtable_t *	tables_lock		void-ptr -> trace_t
domains	_Py_hashtable_t *	tables_lock		domain -> _Py_hashtable_t * (per-domain traces)
empty_traceback	struct tracemalloc_traceback	???
reentrant_key	Py_tss_t	???

notes:

• each frame in struct tracemalloc_traceback holds a filename object
• traceback_t is a typedef for struct tracemalloc_traceback
• frame_t is a typedef for struct tracemalloc_frame

hold objects:

• filenames
• traceback (filename in each frame)
• tracebacks (filename in each frame of each traceback)
• traces (filename in each frame of each traceback)
• domains (filename in each frame of each traceback in each domain)

Usage

(expand)

simple:

name	context	get	set
config
. initialized	lifecycle	tracemalloc_init() tracemalloc_deinit()	tracemalloc_init() tracemalloc_deinit()
. tracing	module	_tracemalloc_is_tracing_impl() _tracemalloc__get_traces_impl() _tracemalloc_clear_traces_impl() _tracemalloc_get_traceback_limit_impl() _tracemalloc_get_traced_memory_impl() _tracemalloc_reset_peak_impl()
	C-API	PyTraceMalloc_Track() PyTraceMalloc_Untrack() _PyTraceMalloc_NewReference() _PyMem_DumpTraceback()
	lifecycle	tracemalloc_start() tracemalloc_stop()	tracemalloc_start() tracemalloc_stop()
	internal	tracemalloc_get_traceback()
. max_nframe	module	_tracemalloc_get_traceback_limit_impl()
	lifecycle		tracemalloc_start()
	internal	traceback_get_frames()
allocators
. mem	lifecycle	tracemalloc_start() + tracemalloc_stop()
. raw	lifecycle	tracemalloc_init() + tracemalloc_start() + tracemalloc_stop()
	internal	raw_malloc() raw_free()
. obj	lifecycle	tracemalloc_start() + tracemalloc_stop()
tables_lock	module	_tracemalloc__get_traces_impl() + _tracemalloc_get_tracemalloc_memory_impl() + _tracemalloc_get_traced_memory_impl() + _tracemalloc_reset_peak_impl() +
	C-API	PyTraceMalloc_Track() + PyTraceMalloc_Untrack() + _PyTraceMalloc_NewReference() +
	lifecycle	tracemalloc_init() tracemalloc_deinit()	tracemalloc_init() * tracemalloc_deinit() *
	allocator	tracemalloc_alloc() + tracemalloc_realloc() + tracemalloc_free() + tracemalloc_realloc_gil() + tracemalloc_raw_realloc() +
	internal	tracemalloc_get_traceback() + tracemalloc_clear_traces() +
traced_memory	module	_tracemalloc_get_traced_memory_impl() _tracemalloc_reset_peak_impl()
	internal	tracemalloc_add_trace()	tracemalloc_add_trace() tracemalloc_remove_trace() tracemalloc_clear_traces()
peak_traced_memory	module	_tracemalloc_get_traced_memory_impl()	_tracemalloc_reset_peak_impl()
	internal	tracemalloc_add_trace()	tracemalloc_add_trace() tracemalloc_clear_traces()
filenames	module	_tracemalloc_get_tracemalloc_memory_impl()
	lifecycle		tracemalloc_init() *
	internal	tracemalloc_get_frame() tracemalloc_clear_traces() +
traceback	lifecycle	tracemalloc_stop()	tracemalloc_start() * tracemalloc_stop() *
	internal	traceback_new() +
tracebacks	module	_tracemalloc_get_tracemalloc_memory_impl()
	lifecycle		tracemalloc_init() * tracemalloc_deinit() +
	internal	traceback_new() + tracemalloc_clear_traces() +
traces	module	_tracemalloc__get_traces_impl() _tracemalloc_get_tracemalloc_memory_impl()
	C-API	_PyTraceMalloc_NewReference()
	lifecycle	tracemalloc_deinit() +	tracemalloc_init() *
	internal	tracemalloc_get_traces_table() tracemalloc_add_trace() (indirect) tracemalloc_remove_trace() (indirect) tracemalloc_get_traceback() (indirect) tracemalloc_clear_traces() +
domains	module	_tracemalloc__get_traces_impl() _tracemalloc_get_tracemalloc_memory_impl()
	lifecycle	tracemalloc_deinit() +	tracemalloc_init() *
	internal	tracemalloc_get_traces_table() tracemalloc_remove_trace() (indirect) tracemalloc_get_traceback() (indirect) tracemalloc_clear_traces() + tracemalloc_add_trace() +
empty_traceback	lifecycle	tracemalloc_init() +
	internal	traceback_new()
reentrant_key	lifecycle	tracemalloc_init() + tracemalloc_deinit() +
	allocator	tracemalloc_alloc_gil() (indirect) + tracemalloc_realloc_gil() (indirect) + tracemalloc_raw_alloc() (indirect) + tracemalloc_raw_realloc() (indirect) +
	internal	get_reentrant() set_reentrant() +

* the function allocates/deallocates the value (see below)
+ the function mutates the value (see below)

simple (extraneous):

name	context	allocate/deallocate	get (assert-only)
config.tracing	internal		tracemalloc_add_trace() tracemalloc_remove_trace()
tables_lock	lifecycle	tracemalloc_init() tracemalloc_deinit()
traced_memory	internal		tracemalloc_remove_trace()
filenames	lifecycle	tracemalloc_init()
traceback	lifecycle	tracemalloc_start() tracemalloc_stop()	tracemalloc_start()
tracebacks	lifecycle	tracemalloc_init()
traces	lifecycle	tracemalloc_init()
domains	lifecycle	tracemalloc_init()

mutation of complex fields:

name	context	initialize	finalize	clear	modify
allocators.mem	lifecycle	tracemalloc_start()
allocators.raw	lifecycle	tracemalloc_init() tracemalloc_start()
allocators.obj	lifecycle	tracemalloc_start()
tables_lock	module				_tracemalloc__get_traces_impl() _tracemalloc_get_tracemalloc_memory_impl() _tracemalloc_get_traced_memory_impl() _tracemalloc_reset_peak_impl()
	C-API				PyTraceMalloc_Track() PyTraceMalloc_Untrack() _PyTraceMalloc_NewReference()
	allocator	tracemalloc_alloc()			tracemalloc_alloc() tracemalloc_realloc() tracemalloc_free() tracemalloc_realloc_gil() tracemalloc_raw_realloc()
	internal				tracemalloc_clear_traces() tracemalloc_get_traceback()
filenames	internal			tracemalloc_clear_traces()	tracemalloc_get_frame() tracemalloc_clear_traces()
	lifecycle		tracemalloc_deinit()
traceback	internal	traceback_new()
tracebacks	lifecycle		tracemalloc_deinit()
	internal			tracemalloc_clear_traces()	traceback_new()
traces	lifecycle		tracemalloc_deinit()
	internal			tracemalloc_clear_traces()
domains	lifecycle		tracemalloc_deinit()
	internal			tracemalloc_clear_traces()	tracemalloc_add_trace()
reentrant_key	lifecycle	tracemalloc_init()	tracemalloc_deinit()
	internal				set_reentrant()

indirection:

name	context	direct	indirect
allocators.raw	lifecycle	tracemalloc_start() tracemalloc_copy_trace()	raw_malloc()
		tracemalloc_stop()	raw_free()
	internal	traceback_new() tracemalloc_add_trace() tracemalloc_copy_trace()	raw_malloc()
		traceback_new() tracemalloc_add_trace() tracemalloc_remove_trace()	raw_free()
traces	internal	tracemalloc_add_trace() tracemalloc_remove_trace() tracemalloc_get_traceback()	tracemalloc_get_traces_table()
domains	internal	tracemalloc_add_trace() tracemalloc_remove_trace() tracemalloc_get_traceback()	tracemalloc_get_traces_table()
reentrant_key	allocator	tracemalloc_alloc_gil() tracemalloc_realloc_gil() tracemalloc_raw_alloc() tracemalloc_raw_realloc()	get_reentrant()
		tracemalloc_alloc_gil() tracemalloc_realloc_gil() tracemalloc_raw_alloc() tracemalloc_raw_realloc()	set_reentrant()

Linked PRs

• GH-101520: Move tracemalloc functionality into core, leaving interface in Modules. #104508

[erlend-aasland]

Are there more actionable items left here?

[encukou]

As far as I can see, there aren't.