Description of problem:
Concurrently calling magic.detect_from_filename(..) by different threads, even on independent files, does often end up in segfault.

The segfault is not observed on RHEL8 (python3-magic-5.33-24.el8.noarch), it sometimes happens on RHEL9.1 (python3-file-magic-5.39-12.el9.noarch.rpm as well as python3-file-magic-5.39-14.el9.noarch.rpm), and VERY often on Gemini kernels of RHEL9.3 beta (not sure if Gemini or 9.3 affects the frequency).

Version-Release number of selected component (if applicable):
python3-file-magic-5.39-12.el9.noarch

How reproducible:
20% to 100%

Steps to Reproduce:
1. Have this script:

from concurrent.futures import ThreadPoolExecutor
import os
import magic

jobs = 4
report_paths = ['/var/tmp/dir1', '/var/tmp/dir2']

def obfuscate_report(archive):
print(archive)
for dirname, dirs, files in os.walk(archive):
for filename in files:
print(f" filename=

")
_fname = os.path.join(dirname, filename.lstrip('/'))
print(f"{_fname}:

")
return

print("===")
pool = ThreadPoolExecutor(jobs)
pool.map(obfuscate_report, report_paths, chunksize=1)
pool.shutdown(wait=True)
print("===")

2. Generate 100 text files in the two directories (each directory content will be examined by magic.detect_from_filename in separate thread):

rm -rf /var/tmp/dir1 /var/tmp/dir2
mkdir /var/tmp/dir1
for i in $(seq 1 100); do date > /var/tmp/dir1/date.${i}.txt; done
cp -r /var/tmp/dir1 /var/tmp/dir

3. Run the script:

1. python3 segfault_reproducer.py

Actual results:

1. python3 segfault_reproducer.py
   ===
   /var/tmp/dir1
   filename=date.1.txt
   /var/tmp/dir2
   filename=date.51.txt
   ===
   tcache_thread_shutdown(): unaligned tcache chunk detected
   Aborted (core dumped)
   #

(depending on RHEL/kernel version, the segfault might not always happen; it must be a race condition)

Expected results:
No segfault

Additional info:
Our real use case: sos / sosreport has a feature to obfuscate customer sensitive data in provided sosreport(s). That requires detection of file types (treat binary files differently than text files). When running the cleaner concurrently on multiple sosreports, we use ThreadPoolExecutor for the concurrency. And hit the segfaults with backtraces pointing to magic library.

Providing sosreports is very important for Red Hat support, so segfaults like these slows down investigation of support cases. Thus we treat this BZ with high priority.