Isolate the tracemalloc Module Between Interpreters #101520

ericsnowcurrently · 2023-02-02T21:41:41Z

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

The text was updated successfully, but these errors were encountered:

…e in Modules. (#104508)

* main: (26 commits) pythonGH-101520: Move tracemalloc functionality into core, leaving interface in Modules. (python#104508) typing: Add more tests for TypeVar (python#104571) pythongh-104572: Improve error messages for invalid constructs in PEP 695 contexts (python#104573) typing: Use PEP 695 syntax in typing.py (python#104553) pythongh-102153: Start stripping C0 control and space chars in `urlsplit` (python#102508) pythongh-104469: Update README.txt for _testcapi (pythongh-104529) pythonGH-103092: isolate `_elementtree` (python#104561) pythongh-104050: Add typing to Argument Clinic converters (python#104547) pythonGH-103906: Remove immortal refcounting in the interpreter (pythonGH-103909) pythongh-87474: Fix file descriptor leaks in subprocess.Popen (python#96351) pythonGH-103092: isolate `pyexpat` (python#104506) pythongh-75367: Fix data descriptor detection in inspect.getattr_static (python#104517) pythongh-104050: Add more annotations to `Tools/clinic.py` (python#104544) pythongh-104555: Fix isinstance() and issubclass() for runtime-checkable protocols that use PEP 695 (python#104556) pythongh-103865: add monitoring support to LOAD_SUPER_ATTR (python#103866) CODEOWNERS: Assign new PEP 695 files to myself (python#104551) pythonGH-104510: Fix refleaks in `_io` base types (python#104516) pythongh-104539: Fix indentation error in logging.config.rst (python#104545) pythongh-104050: Don't star-import 'types' in Argument Clinic (python#104543) pythongh-104050: Add basic typing to CConverter in clinic.py (python#104538) ...

…terface in Modules. (python#104508)

ericsnowcurrently added type-feature A feature request or enhancement interpreter-core (Objects, Python, Grammar, and Parser dirs) topic-subinterpreters 3.12 new features, bugs and security fixes labels Feb 2, 2023

ericsnowcurrently mentioned this issue Feb 2, 2023

Improve Interpreter Isolation #100227

Open

ericsnowcurrently mentioned this issue Mar 28, 2023

Isolate Stdlib Extension Modules #103092

Open

26 tasks

bedevere-bot mentioned this issue May 15, 2023

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

Merged

markshannon added a commit that referenced this issue May 17, 2023

GH-101520: Move tracemalloc functionality into core, leaving interfac…

f7df173

…e in Modules. (#104508)

JelleZijlstra pushed a commit to JelleZijlstra/cpython that referenced this issue May 18, 2023

pythonGH-101520: Move tracemalloc functionality into core, leaving in…

4fea66d

…terface in Modules. (python#104508)

Isolate the tracemalloc Module Between Interpreters #101520

Isolate the tracemalloc Module Between Interpreters #101520

Comments

ericsnowcurrently commented Feb 2, 2023 • edited by bedevere-bot

Allocators

State

Fields

Usage

Linked PRs

ericsnowcurrently commented Feb 2, 2023 •

edited by bedevere-bot