pycares has a Use-After-Free Vulnerability

Summary

pycares is vulnerable to a use-after-free condition that occurs when a Channel object is garbage collected while DNS queries are still pending. This results in a fatal Python error and interpreter crash.

Details

Root Cause

The vulnerability stems from improper handling of callback references when the Channel object is destroyed:

1. When a DNS query is initiated, pycares stores a callback reference using ffi.new_handle()
2. If the Channel object is garbage collected while queries are pending, the callback references become invalid
3. When c-ares attempts to invoke the callback, it accesses freed memory, causing a fatal error

This issue was much more likely to occur when using event_thread=True but could happen without it under the right circumstances.

Technical Details

The core issue is a race condition between Python's garbage collector and c-ares's callback execution:

1. When __del__ is called from within a c-ares callback context, we cannot immediately call ares_destroy() because c-ares is still executing code after the callback returns
2. c-ares needs to execute cleanup code after our Python callback returns (specifically at lines 1422-1429 in ares_process.c)
3. If we destroy the channel too quickly, c-ares accesses freed memory

Impact

Applications using pycares can be crashed remotely by triggering DNS queries that result in Channel objects being garbage collected before query completion. This is particularly problematic in scenarios where:

• Channel objects are created per-request
• Multiple failed DNS queries are processed rapidly
• The application doesn't properly manage Channel lifecycle

The error manifests as:

Fatal Python error: b_from_handle: ffi.from_handle() detected that the address passed points to garbage

Fix

The vulnerability has been fixed in pycares 4.9.0 by implementing a safe channel destruction mechanism

Mitigation

For Application Developers

1. Upgrade to pycares >= 4.9.0 - This version includes the fix and requires no code changes
2. Best practices (optional but recommended):
   python

   # Explicit cleanup
   channel.close()
   
   # Or use context manager
   with pycares.Channel() as channel:
       # ... use channel ...
   # Automatically closed

3. Avoid creating Channel objects per-request - Prefer long-lived instances for better performance and safety

The fix is completely transparent - no API changes or code modifications are required.

Credit

This vulnerability was reported by @vEpiphyte through the aio-libs security program.