Repository navigation
Memory sanitizer generates false positives on data returned by os.getrandom() #148850
Description
Bug report
Bug description:
When python is built with memory sanitizer enabled, reading any data returned by os.getrandom() will trigger a memory sanitizer false positive, because the buffer is filled using a syscall, and the memory isn't marked as having been written to.
Currently,
$ cat test.py
import os
data = os.getrandom(32)
h = data.hex()
print(f'h={h}')
$ ./python ./test.py
Uninitialized bytes in memchr at offset 18 inside [0x707000006458, 66)
==242471==WARNING: MemorySanitizer: use-of-uninitialized-value
...
SUMMARY: MemorySanitizer: use-of-uninitialized-value /home/peter/dev/pvanvugt-cpython/Objects/stringlib/fastsearch.h:58:13 in ucs1lib_find_char
Exiting
I have verified that applying __masan_unpoison() to the data fixes this.
CPython versions tested on:
CPython main branch
Operating systems tested on:
Linux
Activity
picnixz commentedon Apr 22, 2026
Is
__masan_unpoisonavailable everywhere? isn't it possible to suppress this without this builtin call? because it may not be portable enough on all PEP-11 platforms and compilers. Is it also a possible issue in MSAN itself? can we fix it using some build options instead?encukou commentedon Apr 22, 2026
Yes,
__msan_unpoisonis available on any platform with (clang-style) MSAN support. That's not all PEP-11 platforms, but that's OK.We already use
__msan_unpoisonin a bunch of places.See also #11375. 2018 was a different time though :)
I'd prefer adding a wrapper macro to
pyport.h, in the spirit of_Py_MEMORY_SANITIZERor_Py_NO_SANITIZE_THREAD: e.g._Py_MSAN_UNPOISON(p, sz)should call__msan_unpoisonif it's available, and do nothing otherwise.petervanvugt commentedon Apr 23, 2026
Thank you, @encukou , #148851 has been updated per your feedback.
pythongh-148850: Fix msan false positive in os.getrandom
gh-148850: Fix memory sanitizer false positive in os.getrandom (GH-14…