Skip to content

[gevent] improving support for the new tracing API#188

Merged
palazzem merged 9 commits into
palazzem/async-pythonfrom
palazzem/gevent-support
Mar 1, 2017
Merged

[gevent] improving support for the new tracing API#188
palazzem merged 9 commits into
palazzem/async-pythonfrom
palazzem/gevent-support

Conversation

@palazzem

@palazzem palazzem commented Feb 17, 2017

Copy link
Copy Markdown

What it does

Improves the gevent support for the new tracing API, through:

  • GeventContextProvider that provides the right Context to the tracer.trace() using the current Greenlet as a carrier
  • the gevent.Greenlet is overridden with the TracedGreenlet instance. It behaves the same way of the original Greenlet class, but it moves the Context among greenlets.

Flow

  • the main greenlet doesn't have a Context
  • when a greenlet is created, a new Context is attached only if the tracer is used
  • if a new greenlet is created from another greenlet, they share the same Context only if available in the invoker function

Usage

# even if not strictly required, it's better that the
# patch happens before importing gevent
from ddtrace import patch, tracer
patch(gevent=True)

import gevent
# do whatever you need and use tracer as usual

@palazzem palazzem force-pushed the palazzem/gevent-support branch 2 times, most recently from 6d812a5 to 8dc1941 Compare February 17, 2017 18:40
Comment thread ddtrace/contrib/gevent/__init__.py Outdated

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is there a reason why tracer.configure cannot happen by default in the ddtrace.contrib.gevent.patch? Is there an event where we would want to patch the Greenlet class but not the tracer's context provider

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

actually we don't have a way to patch things using the configure() method. I don't like having these two lines too because they're error prone (both are required). What I can do is:

  • having the class for this context provider that patches automatically the code like:
# even if not strictly required, it's better that the
# patch happens before importing gevent
from ddtrace import tracer
from ddtrace.contrib.gevent import GeventContextProvider
tracer.configure(context_provider=GeventContextProvider())

import gevent
# do whatever you need

In that case, under the hood, the constructor will call the gevent patch()

@talwai talwai Feb 17, 2017

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

i'm leaning towards:

from ddtrace import patch_all; patch_all(gevent=True)

calls tracer.configure() under the hood. Still I'm fine with either, i'll let you be the judge

@palazzem palazzem Feb 17, 2017

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

oh gotcha.... yes that could be another option! I may check if it applies properly to other async modules, because I want to keep the API consistent among them.

Comment thread ddtrace/contrib/gevent/greenlet.py Outdated

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

reuse CONTEXT_ATTR ?

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

👍

Comment thread ddtrace/contrib/gevent/patch.py Outdated

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

does testing suggest we need the latter two patches? i would think these shortcuts end up initializing a new gevent.Greenlet anyway which is aliased to a TracedGreenlet?

@palazzem palazzem Feb 17, 2017

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

These shortcuts are replaced because gevent module keeps the original reference for the Greenlet class. Because both are class methods, it means that when they're used, the original class will be initialized.

(If I remove them, tests will fail)

@palazzem palazzem force-pushed the palazzem/gevent-support branch from 6ff5410 to 12ddefa Compare February 17, 2017 19:53
@palazzem palazzem requested a review from LotharSee February 17, 2017 19:54
@palazzem palazzem force-pushed the palazzem/async-python branch from 0477cab to 18a90b3 Compare February 20, 2017 10:43
@palazzem palazzem force-pushed the palazzem/gevent-support branch from 6bbb11f to 76987da Compare February 20, 2017 10:44
@palazzem palazzem force-pushed the palazzem/gevent-support branch from 76987da to 6192bb2 Compare February 20, 2017 10:46
@palazzem palazzem added this to the 0.6.0 milestone Feb 20, 2017

@LeoCavaille LeoCavaille left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Some minor docs nitpick, the code lgtm, it's really easy to read through with the context notion.

Comment thread ddtrace/contrib/gevent/__init__.py Outdated
"""
To trace a request in a gevent-ed environment, configure the tracer to use greenlet-local
storage, rather than the default thread-local storage.
To trace a request in a gevent-ed environment, configure the tracer to use the Greenlet

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

gevent-ed sounds weird? gevent?

Comment thread ddtrace/contrib/gevent/__init__.py Outdated
To trace a request in a gevent-ed environment, configure the tracer to use greenlet-local
storage, rather than the default thread-local storage.
To trace a request in a gevent-ed environment, configure the tracer to use the Greenlet
context provider, rather than the default one that relies in thread-local storaging.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

relies in -> relies on

Comment thread ddtrace/contrib/gevent/greenlet.py Outdated
``Greenlet`` class that is used to replace the original ``gevent``
class. This class is supposed to do ``Context`` replacing operation, so
that any greenlet inherits the context from the parent Greenlet.
When a new greenlet is spawn from the main greenlet, a new instance

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

spawned

@LotharSee LotharSee left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Very good work, looking great! It is all good to me.

@palazzem palazzem merged commit 52b9ba7 into palazzem/async-python Mar 1, 2017
@palazzem palazzem deleted the palazzem/gevent-support branch March 1, 2017 12:25
gh-worker-dd-mergequeue-cf854d Bot pushed a commit that referenced this pull request Apr 23, 2026
…llocator (#17664)

## Description

This PR fixes a segmentation fault in the memory allocation profiler that occurs when a hook call races with `memalloc` start/stop operations. The issue arises from concurrent access to the saved allocator struct, which could be partially written while being read, resulting in`NULL` function pointers being dereferenced.  The key indicator in that case is that `#1 0x0000000000000000` frame -- we are trying to execute a null function pointer.

````
Error UnixSignal: Process terminated with SEGV_MAPERR (SIGSEGV)
#0   0x00007ff3c303a8d4  
#1   0x0000000000000000 memalloc_alloc (/go/src/github.com/DataDog/apm-reliability/dd-trace-py/ddtrace/profiling/collector/_memalloc.cpp:68)
#2   0x00007ff39dcb3b20 memalloc_alloc (/go/src/github.com/DataDog/apm-reliability/dd-trace-py/ddtrace/profiling/collector/_memalloc.cpp:68)
#3   0x00007ff39dcb3b20 memalloc_malloc(void*, unsigned long) (/go/src/github.com/DataDog/apm-reliability/dd-trace-py/ddtrace/profiling/collector/_memalloc.cpp:80)
#4   0x00007ff3c3087e1b PyUnicode_New 
#5   0x00007ff3c30889f4  
#6   0x00007ff3c3170c84  
#7   0x00007ff3c316b931  
#8   0x00007ff3c31aaac8  
#9   0x00007ff3c31033ac  
#10  0x00007ff3c310e2a6 PyObject_CallMethodObjArgs 
#11  0x00007ff3c310e46d  
#12  0x00007ff3c31a96c2  
#13  0x00007ff3c3102fd7 PyObject_Vectorcall 
#14  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#15  0x00007ff3c323c094  
#16  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#17  0x00007ff3c323c094  
#18  0x00007ff3c30e997d PyObject_CallOneArg 
#19  0x00007ff3c306a480 _PyObject_GenericGetAttrWithDict 
#20  0x00007ff3c30c620d PyObject_GetAttr 
#21  0x00007ff3c32309e7 _PyEval_EvalFrameDefault 
#22  0x00007ff3c323c094  
#23  0x00007ff3c312880e  
#24  0x00007ff3c30e917c _PyObject_MakeTpCall 
#25  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#26  0x00007ff3c323c094  
#27  0x00007ff3c312880e  
#28  0x00007ff3c30e917c _PyObject_MakeTpCall 
#29  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#30  0x00007ff3c323c094  
#31  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#32  0x00007ff3c323c094  
#33  0x00007ff3c317d0fd  
#34  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#35  0x00007ff3c323c094  
#36  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#37  0x00007ff3c323c094  
#38  0x00007ff3c317d1b5  
#39  0x00007ff3c3102fd7 PyObject_Vectorcall 
#40  0x00007ff3c3232f4a _PyEval_EvalFrameDefault 
#41  0x00007ff3c3240da5  
#42  0x00007ff3c324112d  
#43  0x00007ff3c3233be1 _PyEval_EvalFrameDefault 
#44  0x00007ff3c323c094  
#45  0x00007ff3c317d1b5  
#46  0x00007ff3c3102fd7 PyObject_Vectorcall 
#47  0x00007ff3c3232f4a _PyEval_EvalFrameDefault 
#48  0x00007ff3c323c094  
#49  0x00007ff3c31033ac  
#50  0x00007ff3c310358d PyObject_CallFunctionObjArgs 
#51  0x00007ff3bf7eb91d WraptBoundFunctionWrapper_call (/project/src/wrapt/_wrappers.c:3750)
#52  0x00007ff3c3104055 _PyObject_Call 
#53  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#54  0x00007ff3c323c094  
#55  0x00007ff3c317d23c  
#56  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#57  0x00007ff3c323c094  
#58  0x00007ff3c310416f _PyObject_Call 
#59  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#60  0x00007ff3c3240da5  
#61  0x00007ff3c324112d  
#62  0x00007ff3c3233be1 _PyEval_EvalFrameDefault 
#63  0x00007ff3c323c094  
#64  0x00007ff3c317d1b5  
#65  0x00007ff3c3102fd7 PyObject_Vectorcall 
#66  0x00007ff3c3232f4a _PyEval_EvalFrameDefault 
#67  0x00007ff3c323c094  
#68  0x00007ff3c317d1b5  
#69  0x00007ff3c310416f _PyObject_Call 
#70  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#71  0x00007ff3c323c094  
#72  0x00007ff3c317d1b5  
#73  0x00007ff3c310416f _PyObject_Call 
#74  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#75  0x00007ff3c323c094  
#76  0x00007ff3c31033ac  
#77  0x00007ff3c310358d PyObject_CallFunctionObjArgs 
#78  0x00007ff3bf7eb91d WraptBoundFunctionWrapper_call (/project/src/wrapt/_wrappers.c:3750)
#79  0x00007ff3c30e917c _PyObject_MakeTpCall 
#80  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#81  0x00007ff3c323c094  
#82  0x00007ff3c317d518  
#83  0x00007ff3c3155963  
#84  0x00007ff3c315393d  
#85  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#86  0x00007ff3c323c094  
#87  0x00007ff3c317d0fd  
#88  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#89  0x00007ff3c323c094  
#90  0x00007ff3c317d0fd  
#91  0x00007ff3c317d518  
#92  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#93  0x00007ff3c323c094  
#94  0x00007ff3c30e9371 _PyObject_FastCallDictTstate 
#95  0x00007ff3c30e958d _PyObject_Call_Prepend 
#96  0x00007ff3c3109150  
#97  0x00007ff3c3104055 _PyObject_Call 
#98  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#99  0x00007ff3c323c094  
#100 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#101 0x00007ff3c30e958d _PyObject_Call_Prepend 
#102 0x00007ff3c3109150  
#103 0x00007ff3c30e917c _PyObject_MakeTpCall 
#104 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#105 0x00007ff3c323c094  
#106 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#107 0x00007ff3c30e958d _PyObject_Call_Prepend 
#108 0x00007ff3c3109150  
#109 0x00007ff3c30e917c _PyObject_MakeTpCall 
#110 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#111 0x00007ff3c323c094  
#112 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#113 0x00007ff3c30e958d _PyObject_Call_Prepend 
#114 0x00007ff3c3109150  
#115 0x00007ff3c30e917c _PyObject_MakeTpCall 
#116 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#117 0x00007ff3c323c094  
#118 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#119 0x00007ff3c30e958d _PyObject_Call_Prepend 
#120 0x00007ff3c3109150  
#121 0x00007ff3c30e917c _PyObject_MakeTpCall 
#122 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#123 0x00007ff3c323c094  
#124 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#125 0x00007ff3c30e958d _PyObject_Call_Prepend 
#126 0x00007ff3c3109150  
#127 0x00007ff3c30e917c _PyObject_MakeTpCall 
#128 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#129 0x00007ff3c323c094  
#130 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#131 0x00007ff3c30e958d _PyObject_Call_Prepend 
#132 0x00007ff3c3109150  
#133 0x00007ff3c30e917c _PyObject_MakeTpCall 
#134 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#135 0x00007ff3c323c094  
#136 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#137 0x00007ff3c30e958d _PyObject_Call_Prepend 
#138 0x00007ff3c3109150  
#139 0x00007ff3c30e917c _PyObject_MakeTpCall 
#140 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#141 0x00007ff3c323c094  
#142 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#143 0x00007ff3c30e958d _PyObject_Call_Prepend 
#144 0x00007ff3c3109150  
#145 0x00007ff3c30e917c _PyObject_MakeTpCall 
#146 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#147 0x00007ff3c323c094  
#148 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#149 0x00007ff3c30e958d _PyObject_Call_Prepend 
#150 0x00007ff3c3109150  
#151 0x00007ff3c30e917c _PyObject_MakeTpCall 
#152 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#153 0x00007ff3c323c094  
#154 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#155 0x00007ff3c30e958d _PyObject_Call_Prepend 
#156 0x00007ff3c3109150  
#157 0x00007ff3c30e917c _PyObject_MakeTpCall 
#158 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#159 0x00007ff3c323c094  
#160 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#161 0x00007ff3c30e958d _PyObject_Call_Prepend 
#162 0x00007ff3c3109150  
#163 0x00007ff3c30e917c _PyObject_MakeTpCall 
#164 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#165 0x00007ff3c323c094  
#166 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#167 0x00007ff3c30e958d _PyObject_Call_Prepend 
#168 0x00007ff3c3109150  
#169 0x00007ff3c30e917c _PyObject_MakeTpCall 
#170 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#171 0x00007ff3c323c094  
#172 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#173 0x00007ff3c30e958d _PyObject_Call_Prepend 
#174 0x00007ff3c3109150  
#175 0x00007ff3c30e917c _PyObject_MakeTpCall 
#176 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#177 0x00007ff3c323c094  
#178 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#179 0x00007ff3c30e958d _PyObject_Call_Prepend 
#180 0x00007ff3c3109150  
#181 0x00007ff3c30e917c _PyObject_MakeTpCall 
#182 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#183 0x00007ff3c323c094  
#184 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#185 0x00007ff3c30e958d _PyObject_Call_Prepend 
#186 0x00007ff3c3109150  
#187 0x00007ff3c30e917c _PyObject_MakeTpCall 
#188 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#189 0x00007ff3c323c094  
#190 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#191 0x00007ff3c323c094  
#192 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#193 0x00007ff3c30e958d _PyObject_Call_Prepend 
#194 0x00007ff3c3109150  
#195 0x00007ff3c30e917c _PyObject_MakeTpCall 
#196 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#197 0x00007ff3c323c094  
#198 0x00007ff3c317d0fd  
#199 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#200 0x00007ff3c323c094  
#201 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#202 0x00007ff3c323c094  
#203 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#204 0x00007ff3c323c094  
#205 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#206 0x00007ff3c323c094  
#207 0x00007ff3c317d23c  
#208 0x00007ff3c31a7ec5  
#209 0x00007ff3c301ac77  
#210 0x00007ff3c357c573  
````

The fix implements two key changes.

1. **Hook functions (`memalloc_alloc`, `memalloc_realloc`)**: Snapshot the allocator struct locally before use and guard indirect function calls with `NULL` checks. This prevents crashes if a partially-written struct is observed during a start/stop race.

2. **Start/stop operations (`memalloc_start`, `memalloc_stop`)**: Use local variables and single assignments when publishing the allocator struct to `global_memalloc_ctx.pymem_allocator_obj`. This ensures concurrent hook calls observe either the old or new struct, never a partially-written intermediate state.

The real root cause is that `PyMem_GetAllocator` is not documented as atomic, and the struct could be read field-by-field while being written to concurrently.  By using local copies and single assignments, we ensure atomicity at the C level and prevent observation of inconsistent state.

Co-authored-by: thomas.kowalski <thomas.kowalski@datadoghq.com>
emmettbutler pushed a commit that referenced this pull request Apr 24, 2026
…llocator (#17664)

## Description

This PR fixes a segmentation fault in the memory allocation profiler that occurs when a hook call races with `memalloc` start/stop operations. The issue arises from concurrent access to the saved allocator struct, which could be partially written while being read, resulting in`NULL` function pointers being dereferenced.  The key indicator in that case is that `#1 0x0000000000000000` frame -- we are trying to execute a null function pointer.

````
Error UnixSignal: Process terminated with SEGV_MAPERR (SIGSEGV)
#0   0x00007ff3c303a8d4  
#1   0x0000000000000000 memalloc_alloc (/go/src/github.com/DataDog/apm-reliability/dd-trace-py/ddtrace/profiling/collector/_memalloc.cpp:68)
#2   0x00007ff39dcb3b20 memalloc_alloc (/go/src/github.com/DataDog/apm-reliability/dd-trace-py/ddtrace/profiling/collector/_memalloc.cpp:68)
#3   0x00007ff39dcb3b20 memalloc_malloc(void*, unsigned long) (/go/src/github.com/DataDog/apm-reliability/dd-trace-py/ddtrace/profiling/collector/_memalloc.cpp:80)
#4   0x00007ff3c3087e1b PyUnicode_New 
#5   0x00007ff3c30889f4  
#6   0x00007ff3c3170c84  
#7   0x00007ff3c316b931  
#8   0x00007ff3c31aaac8  
#9   0x00007ff3c31033ac  
#10  0x00007ff3c310e2a6 PyObject_CallMethodObjArgs 
#11  0x00007ff3c310e46d  
#12  0x00007ff3c31a96c2  
#13  0x00007ff3c3102fd7 PyObject_Vectorcall 
#14  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#15  0x00007ff3c323c094  
#16  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#17  0x00007ff3c323c094  
#18  0x00007ff3c30e997d PyObject_CallOneArg 
#19  0x00007ff3c306a480 _PyObject_GenericGetAttrWithDict 
#20  0x00007ff3c30c620d PyObject_GetAttr 
#21  0x00007ff3c32309e7 _PyEval_EvalFrameDefault 
#22  0x00007ff3c323c094  
#23  0x00007ff3c312880e  
#24  0x00007ff3c30e917c _PyObject_MakeTpCall 
#25  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#26  0x00007ff3c323c094  
#27  0x00007ff3c312880e  
#28  0x00007ff3c30e917c _PyObject_MakeTpCall 
#29  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#30  0x00007ff3c323c094  
#31  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#32  0x00007ff3c323c094  
#33  0x00007ff3c317d0fd  
#34  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#35  0x00007ff3c323c094  
#36  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#37  0x00007ff3c323c094  
#38  0x00007ff3c317d1b5  
#39  0x00007ff3c3102fd7 PyObject_Vectorcall 
#40  0x00007ff3c3232f4a _PyEval_EvalFrameDefault 
#41  0x00007ff3c3240da5  
#42  0x00007ff3c324112d  
#43  0x00007ff3c3233be1 _PyEval_EvalFrameDefault 
#44  0x00007ff3c323c094  
#45  0x00007ff3c317d1b5  
#46  0x00007ff3c3102fd7 PyObject_Vectorcall 
#47  0x00007ff3c3232f4a _PyEval_EvalFrameDefault 
#48  0x00007ff3c323c094  
#49  0x00007ff3c31033ac  
#50  0x00007ff3c310358d PyObject_CallFunctionObjArgs 
#51  0x00007ff3bf7eb91d WraptBoundFunctionWrapper_call (/project/src/wrapt/_wrappers.c:3750)
#52  0x00007ff3c3104055 _PyObject_Call 
#53  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#54  0x00007ff3c323c094  
#55  0x00007ff3c317d23c  
#56  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#57  0x00007ff3c323c094  
#58  0x00007ff3c310416f _PyObject_Call 
#59  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#60  0x00007ff3c3240da5  
#61  0x00007ff3c324112d  
#62  0x00007ff3c3233be1 _PyEval_EvalFrameDefault 
#63  0x00007ff3c323c094  
#64  0x00007ff3c317d1b5  
#65  0x00007ff3c3102fd7 PyObject_Vectorcall 
#66  0x00007ff3c3232f4a _PyEval_EvalFrameDefault 
#67  0x00007ff3c323c094  
#68  0x00007ff3c317d1b5  
#69  0x00007ff3c310416f _PyObject_Call 
#70  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#71  0x00007ff3c323c094  
#72  0x00007ff3c317d1b5  
#73  0x00007ff3c310416f _PyObject_Call 
#74  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#75  0x00007ff3c323c094  
#76  0x00007ff3c31033ac  
#77  0x00007ff3c310358d PyObject_CallFunctionObjArgs 
#78  0x00007ff3bf7eb91d WraptBoundFunctionWrapper_call (/project/src/wrapt/_wrappers.c:3750)
#79  0x00007ff3c30e917c _PyObject_MakeTpCall 
#80  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#81  0x00007ff3c323c094  
#82  0x00007ff3c317d518  
#83  0x00007ff3c3155963  
#84  0x00007ff3c315393d  
#85  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#86  0x00007ff3c323c094  
#87  0x00007ff3c317d0fd  
#88  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#89  0x00007ff3c323c094  
#90  0x00007ff3c317d0fd  
#91  0x00007ff3c317d518  
#92  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#93  0x00007ff3c323c094  
#94  0x00007ff3c30e9371 _PyObject_FastCallDictTstate 
#95  0x00007ff3c30e958d _PyObject_Call_Prepend 
#96  0x00007ff3c3109150  
#97  0x00007ff3c3104055 _PyObject_Call 
#98  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#99  0x00007ff3c323c094  
#100 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#101 0x00007ff3c30e958d _PyObject_Call_Prepend 
#102 0x00007ff3c3109150  
#103 0x00007ff3c30e917c _PyObject_MakeTpCall 
#104 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#105 0x00007ff3c323c094  
#106 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#107 0x00007ff3c30e958d _PyObject_Call_Prepend 
#108 0x00007ff3c3109150  
#109 0x00007ff3c30e917c _PyObject_MakeTpCall 
#110 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#111 0x00007ff3c323c094  
#112 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#113 0x00007ff3c30e958d _PyObject_Call_Prepend 
#114 0x00007ff3c3109150  
#115 0x00007ff3c30e917c _PyObject_MakeTpCall 
#116 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#117 0x00007ff3c323c094  
#118 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#119 0x00007ff3c30e958d _PyObject_Call_Prepend 
#120 0x00007ff3c3109150  
#121 0x00007ff3c30e917c _PyObject_MakeTpCall 
#122 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#123 0x00007ff3c323c094  
#124 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#125 0x00007ff3c30e958d _PyObject_Call_Prepend 
#126 0x00007ff3c3109150  
#127 0x00007ff3c30e917c _PyObject_MakeTpCall 
#128 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#129 0x00007ff3c323c094  
#130 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#131 0x00007ff3c30e958d _PyObject_Call_Prepend 
#132 0x00007ff3c3109150  
#133 0x00007ff3c30e917c _PyObject_MakeTpCall 
#134 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#135 0x00007ff3c323c094  
#136 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#137 0x00007ff3c30e958d _PyObject_Call_Prepend 
#138 0x00007ff3c3109150  
#139 0x00007ff3c30e917c _PyObject_MakeTpCall 
#140 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#141 0x00007ff3c323c094  
#142 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#143 0x00007ff3c30e958d _PyObject_Call_Prepend 
#144 0x00007ff3c3109150  
#145 0x00007ff3c30e917c _PyObject_MakeTpCall 
#146 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#147 0x00007ff3c323c094  
#148 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#149 0x00007ff3c30e958d _PyObject_Call_Prepend 
#150 0x00007ff3c3109150  
#151 0x00007ff3c30e917c _PyObject_MakeTpCall 
#152 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#153 0x00007ff3c323c094  
#154 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#155 0x00007ff3c30e958d _PyObject_Call_Prepend 
#156 0x00007ff3c3109150  
#157 0x00007ff3c30e917c _PyObject_MakeTpCall 
#158 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#159 0x00007ff3c323c094  
#160 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#161 0x00007ff3c30e958d _PyObject_Call_Prepend 
#162 0x00007ff3c3109150  
#163 0x00007ff3c30e917c _PyObject_MakeTpCall 
#164 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#165 0x00007ff3c323c094  
#166 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#167 0x00007ff3c30e958d _PyObject_Call_Prepend 
#168 0x00007ff3c3109150  
#169 0x00007ff3c30e917c _PyObject_MakeTpCall 
#170 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#171 0x00007ff3c323c094  
#172 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#173 0x00007ff3c30e958d _PyObject_Call_Prepend 
#174 0x00007ff3c3109150  
#175 0x00007ff3c30e917c _PyObject_MakeTpCall 
#176 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#177 0x00007ff3c323c094  
#178 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#179 0x00007ff3c30e958d _PyObject_Call_Prepend 
#180 0x00007ff3c3109150  
#181 0x00007ff3c30e917c _PyObject_MakeTpCall 
#182 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#183 0x00007ff3c323c094  
#184 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#185 0x00007ff3c30e958d _PyObject_Call_Prepend 
#186 0x00007ff3c3109150  
#187 0x00007ff3c30e917c _PyObject_MakeTpCall 
#188 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#189 0x00007ff3c323c094  
#190 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#191 0x00007ff3c323c094  
#192 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#193 0x00007ff3c30e958d _PyObject_Call_Prepend 
#194 0x00007ff3c3109150  
#195 0x00007ff3c30e917c _PyObject_MakeTpCall 
#196 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#197 0x00007ff3c323c094  
#198 0x00007ff3c317d0fd  
#199 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#200 0x00007ff3c323c094  
#201 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#202 0x00007ff3c323c094  
#203 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#204 0x00007ff3c323c094  
#205 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#206 0x00007ff3c323c094  
#207 0x00007ff3c317d23c  
#208 0x00007ff3c31a7ec5  
#209 0x00007ff3c301ac77  
#210 0x00007ff3c357c573  
````

The fix implements two key changes.

1. **Hook functions (`memalloc_alloc`, `memalloc_realloc`)**: Snapshot the allocator struct locally before use and guard indirect function calls with `NULL` checks. This prevents crashes if a partially-written struct is observed during a start/stop race.

2. **Start/stop operations (`memalloc_start`, `memalloc_stop`)**: Use local variables and single assignments when publishing the allocator struct to `global_memalloc_ctx.pymem_allocator_obj`. This ensures concurrent hook calls observe either the old or new struct, never a partially-written intermediate state.

The real root cause is that `PyMem_GetAllocator` is not documented as atomic, and the struct could be read field-by-field while being written to concurrently.  By using local copies and single assignments, we ensure atomicity at the C level and prevent observation of inconsistent state.

Co-authored-by: thomas.kowalski <thomas.kowalski@datadoghq.com>
emmettbutler pushed a commit that referenced this pull request May 6, 2026
…llocator (#17664)

## Description

This PR fixes a segmentation fault in the memory allocation profiler that occurs when a hook call races with `memalloc` start/stop operations. The issue arises from concurrent access to the saved allocator struct, which could be partially written while being read, resulting in`NULL` function pointers being dereferenced.  The key indicator in that case is that `#1 0x0000000000000000` frame -- we are trying to execute a null function pointer.

````
Error UnixSignal: Process terminated with SEGV_MAPERR (SIGSEGV)
#0   0x00007ff3c303a8d4  
#1   0x0000000000000000 memalloc_alloc (/go/src/github.com/DataDog/apm-reliability/dd-trace-py/ddtrace/profiling/collector/_memalloc.cpp:68)
#2   0x00007ff39dcb3b20 memalloc_alloc (/go/src/github.com/DataDog/apm-reliability/dd-trace-py/ddtrace/profiling/collector/_memalloc.cpp:68)
#3   0x00007ff39dcb3b20 memalloc_malloc(void*, unsigned long) (/go/src/github.com/DataDog/apm-reliability/dd-trace-py/ddtrace/profiling/collector/_memalloc.cpp:80)
#4   0x00007ff3c3087e1b PyUnicode_New 
#5   0x00007ff3c30889f4  
#6   0x00007ff3c3170c84  
#7   0x00007ff3c316b931  
#8   0x00007ff3c31aaac8  
#9   0x00007ff3c31033ac  
#10  0x00007ff3c310e2a6 PyObject_CallMethodObjArgs 
#11  0x00007ff3c310e46d  
#12  0x00007ff3c31a96c2  
#13  0x00007ff3c3102fd7 PyObject_Vectorcall 
#14  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#15  0x00007ff3c323c094  
#16  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#17  0x00007ff3c323c094  
#18  0x00007ff3c30e997d PyObject_CallOneArg 
#19  0x00007ff3c306a480 _PyObject_GenericGetAttrWithDict 
#20  0x00007ff3c30c620d PyObject_GetAttr 
#21  0x00007ff3c32309e7 _PyEval_EvalFrameDefault 
#22  0x00007ff3c323c094  
#23  0x00007ff3c312880e  
#24  0x00007ff3c30e917c _PyObject_MakeTpCall 
#25  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#26  0x00007ff3c323c094  
#27  0x00007ff3c312880e  
#28  0x00007ff3c30e917c _PyObject_MakeTpCall 
#29  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#30  0x00007ff3c323c094  
#31  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#32  0x00007ff3c323c094  
#33  0x00007ff3c317d0fd  
#34  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#35  0x00007ff3c323c094  
#36  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#37  0x00007ff3c323c094  
#38  0x00007ff3c317d1b5  
#39  0x00007ff3c3102fd7 PyObject_Vectorcall 
#40  0x00007ff3c3232f4a _PyEval_EvalFrameDefault 
#41  0x00007ff3c3240da5  
#42  0x00007ff3c324112d  
#43  0x00007ff3c3233be1 _PyEval_EvalFrameDefault 
#44  0x00007ff3c323c094  
#45  0x00007ff3c317d1b5  
#46  0x00007ff3c3102fd7 PyObject_Vectorcall 
#47  0x00007ff3c3232f4a _PyEval_EvalFrameDefault 
#48  0x00007ff3c323c094  
#49  0x00007ff3c31033ac  
#50  0x00007ff3c310358d PyObject_CallFunctionObjArgs 
#51  0x00007ff3bf7eb91d WraptBoundFunctionWrapper_call (/project/src/wrapt/_wrappers.c:3750)
#52  0x00007ff3c3104055 _PyObject_Call 
#53  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#54  0x00007ff3c323c094  
#55  0x00007ff3c317d23c  
#56  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#57  0x00007ff3c323c094  
#58  0x00007ff3c310416f _PyObject_Call 
#59  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#60  0x00007ff3c3240da5  
#61  0x00007ff3c324112d  
#62  0x00007ff3c3233be1 _PyEval_EvalFrameDefault 
#63  0x00007ff3c323c094  
#64  0x00007ff3c317d1b5  
#65  0x00007ff3c3102fd7 PyObject_Vectorcall 
#66  0x00007ff3c3232f4a _PyEval_EvalFrameDefault 
#67  0x00007ff3c323c094  
#68  0x00007ff3c317d1b5  
#69  0x00007ff3c310416f _PyObject_Call 
#70  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#71  0x00007ff3c323c094  
#72  0x00007ff3c317d1b5  
#73  0x00007ff3c310416f _PyObject_Call 
#74  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#75  0x00007ff3c323c094  
#76  0x00007ff3c31033ac  
#77  0x00007ff3c310358d PyObject_CallFunctionObjArgs 
#78  0x00007ff3bf7eb91d WraptBoundFunctionWrapper_call (/project/src/wrapt/_wrappers.c:3750)
#79  0x00007ff3c30e917c _PyObject_MakeTpCall 
#80  0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#81  0x00007ff3c323c094  
#82  0x00007ff3c317d518  
#83  0x00007ff3c3155963  
#84  0x00007ff3c315393d  
#85  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#86  0x00007ff3c323c094  
#87  0x00007ff3c317d0fd  
#88  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#89  0x00007ff3c323c094  
#90  0x00007ff3c317d0fd  
#91  0x00007ff3c317d518  
#92  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#93  0x00007ff3c323c094  
#94  0x00007ff3c30e9371 _PyObject_FastCallDictTstate 
#95  0x00007ff3c30e958d _PyObject_Call_Prepend 
#96  0x00007ff3c3109150  
#97  0x00007ff3c3104055 _PyObject_Call 
#98  0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#99  0x00007ff3c323c094  
#100 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#101 0x00007ff3c30e958d _PyObject_Call_Prepend 
#102 0x00007ff3c3109150  
#103 0x00007ff3c30e917c _PyObject_MakeTpCall 
#104 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#105 0x00007ff3c323c094  
#106 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#107 0x00007ff3c30e958d _PyObject_Call_Prepend 
#108 0x00007ff3c3109150  
#109 0x00007ff3c30e917c _PyObject_MakeTpCall 
#110 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#111 0x00007ff3c323c094  
#112 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#113 0x00007ff3c30e958d _PyObject_Call_Prepend 
#114 0x00007ff3c3109150  
#115 0x00007ff3c30e917c _PyObject_MakeTpCall 
#116 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#117 0x00007ff3c323c094  
#118 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#119 0x00007ff3c30e958d _PyObject_Call_Prepend 
#120 0x00007ff3c3109150  
#121 0x00007ff3c30e917c _PyObject_MakeTpCall 
#122 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#123 0x00007ff3c323c094  
#124 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#125 0x00007ff3c30e958d _PyObject_Call_Prepend 
#126 0x00007ff3c3109150  
#127 0x00007ff3c30e917c _PyObject_MakeTpCall 
#128 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#129 0x00007ff3c323c094  
#130 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#131 0x00007ff3c30e958d _PyObject_Call_Prepend 
#132 0x00007ff3c3109150  
#133 0x00007ff3c30e917c _PyObject_MakeTpCall 
#134 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#135 0x00007ff3c323c094  
#136 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#137 0x00007ff3c30e958d _PyObject_Call_Prepend 
#138 0x00007ff3c3109150  
#139 0x00007ff3c30e917c _PyObject_MakeTpCall 
#140 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#141 0x00007ff3c323c094  
#142 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#143 0x00007ff3c30e958d _PyObject_Call_Prepend 
#144 0x00007ff3c3109150  
#145 0x00007ff3c30e917c _PyObject_MakeTpCall 
#146 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#147 0x00007ff3c323c094  
#148 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#149 0x00007ff3c30e958d _PyObject_Call_Prepend 
#150 0x00007ff3c3109150  
#151 0x00007ff3c30e917c _PyObject_MakeTpCall 
#152 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#153 0x00007ff3c323c094  
#154 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#155 0x00007ff3c30e958d _PyObject_Call_Prepend 
#156 0x00007ff3c3109150  
#157 0x00007ff3c30e917c _PyObject_MakeTpCall 
#158 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#159 0x00007ff3c323c094  
#160 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#161 0x00007ff3c30e958d _PyObject_Call_Prepend 
#162 0x00007ff3c3109150  
#163 0x00007ff3c30e917c _PyObject_MakeTpCall 
#164 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#165 0x00007ff3c323c094  
#166 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#167 0x00007ff3c30e958d _PyObject_Call_Prepend 
#168 0x00007ff3c3109150  
#169 0x00007ff3c30e917c _PyObject_MakeTpCall 
#170 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#171 0x00007ff3c323c094  
#172 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#173 0x00007ff3c30e958d _PyObject_Call_Prepend 
#174 0x00007ff3c3109150  
#175 0x00007ff3c30e917c _PyObject_MakeTpCall 
#176 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#177 0x00007ff3c323c094  
#178 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#179 0x00007ff3c30e958d _PyObject_Call_Prepend 
#180 0x00007ff3c3109150  
#181 0x00007ff3c30e917c _PyObject_MakeTpCall 
#182 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#183 0x00007ff3c323c094  
#184 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#185 0x00007ff3c30e958d _PyObject_Call_Prepend 
#186 0x00007ff3c3109150  
#187 0x00007ff3c30e917c _PyObject_MakeTpCall 
#188 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#189 0x00007ff3c323c094  
#190 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#191 0x00007ff3c323c094  
#192 0x00007ff3c30e92f1 _PyObject_FastCallDictTstate 
#193 0x00007ff3c30e958d _PyObject_Call_Prepend 
#194 0x00007ff3c3109150  
#195 0x00007ff3c30e917c _PyObject_MakeTpCall 
#196 0x00007ff3c32335a2 _PyEval_EvalFrameDefault 
#197 0x00007ff3c323c094  
#198 0x00007ff3c317d0fd  
#199 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#200 0x00007ff3c323c094  
#201 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#202 0x00007ff3c323c094  
#203 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#204 0x00007ff3c323c094  
#205 0x00007ff3c3233dd3 _PyEval_EvalFrameDefault 
#206 0x00007ff3c323c094  
#207 0x00007ff3c317d23c  
#208 0x00007ff3c31a7ec5  
#209 0x00007ff3c301ac77  
#210 0x00007ff3c357c573  
````

The fix implements two key changes.

1. **Hook functions (`memalloc_alloc`, `memalloc_realloc`)**: Snapshot the allocator struct locally before use and guard indirect function calls with `NULL` checks. This prevents crashes if a partially-written struct is observed during a start/stop race.

2. **Start/stop operations (`memalloc_start`, `memalloc_stop`)**: Use local variables and single assignments when publishing the allocator struct to `global_memalloc_ctx.pymem_allocator_obj`. This ensures concurrent hook calls observe either the old or new struct, never a partially-written intermediate state.

The real root cause is that `PyMem_GetAllocator` is not documented as atomic, and the struct could be read field-by-field while being written to concurrently.  By using local copies and single assignments, we ensure atomicity at the C level and prevent observation of inconsistent state.

Co-authored-by: thomas.kowalski <thomas.kowalski@datadoghq.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants