SIGSEGV in libmsodbcsql during process teardown (iconv_close) on macOS arm64 with concurrent encrypted connections #685

Summary

On macOS arm64, the process crashes with SIGSEGV during interpreter/process teardown (exit() → __cxa_finalize_ranges), inside the driver's static destructor calling iconv_close. The crash surfaces at shutdown when multiple worker threads have been using encrypted connections (Encrypt=yes) concurrently. The process terminates with exit code 139 and no Python traceback.

This appears related to, but distinct from, #668 (TLS stream corruption on concurrent encrypted reads — data corruption, no crash) and #679 (SIGSEGV in SQLFetchScroll during concurrent fetchmany). Here the faulting frame is not SQLFetchScroll but the driver's teardown path via libiconv.

Environment

• macOS 26.5.2 (build 25F84), Apple Silicon (arm64)
• Python 3.12.11
• mssql-python 1.11.0 (bundled libmsodbcsql.18.dylib, mssql_python/libs/macos/arm64)
• SQLAlchemy 2.1.0b3
• Connection options: Encrypt=yes, TrustServerCertificate=yes, SNAPSHOT isolation
• Concurrency: multiple worker threads. mssql-python's native connection pooling is enabled (enable_pooling=True), which uses a process-global pooling manager, so encrypted connections are pooled and reused across the process rather than being one-per-thread.

Crash

Signal: SIGSEGV / EXC_BAD_ACCESS (KERN_INVALID_ADDRESS), subtype "possible pointer authentication failure" at a wild address.

Faulting thread (from the macOS .ips crash report):

libiconv.2.dylib       _citrus_iconv_close
libiconv.2.dylib       __bsd_iconv_close
libmsodbcsql.18.dylib  <driver frame>
libsystem_c.dylib      __cxa_finalize_ranges
libsystem_c.dylib      exit
libdyld.dylib          dyld4::LibSystemHelpers::exit(int) const
dyld                   dyld4::LibSystemHelpersWrapper::exit(int) const
dyld                   start

So the crash happens while the process is exiting: exit() runs C++ static destructors (__cxa_finalize_ranges), one of which belongs to libmsodbcsql.18.dylib and calls iconv_close on what looks like an already-freed or corrupted iconv handle, producing a pointer-authentication failure.

Scenario

1. Several worker threads read from different tables concurrently under SNAPSHOT isolation over pooled encrypted connections (Encrypt=yes).
2. The application begins shutting down (an unrelated application-level error triggers a fail-fast exit()) while those threads/connections are still being torn down.
3. During exit(), the driver's static-destructor cleanup segfaults in iconv_close.

The iconv involvement is notable: iconv is the character-encoding conversion layer, and #668 reports TLS/stream corruption specifically under concurrent encrypted reads on arm64. A corrupted iconv handle that only faults at driver finalization is consistent with encoding/stream state being left in a bad state under that same concurrency.

Relation to existing issues

• TLS stream corruption under concurrent threaded reads on macOS arm64 #668 — TLS stream corruption on concurrent encrypted reads (arm64): data corruption / protocol errors, no hard crash.
• SIGSEGV in SQLFetchScroll (libmsodbcsql.18) under concurrent threaded fetch on macOS arm64 #679 — SIGSEGV in SQLFetchScroll during concurrent fetchmany (arm64).

This report is a third manifestation on the same platform and connection profile (macOS arm64, Encrypt=yes, concurrent threaded reads), but the crash occurs at process teardown via libiconv, not during an active fetch. Filing separately so the teardown/iconv stack is tracked; it may share the root cause reported in #668.

Notes

• No Python-level exception is raised for the crash itself; only the exit code (139) and the macOS crash report indicate it.