Bug #121287 SIGSEGV in Shared_multi_map<Routine>::Autolocker::~Autolocker() due to unlock-before-delete race condition

Submitted: 15 Sep 11:14 Modified: 16 Oct 12:26
Reporter: nanzhou lin Email Updates:
Status: No Feedback Impact on me:
None 
Category:MySQL Server: Data Dictionary Severity:S3 (Non-critical)
Version: OS:Any
Assigned to: CPU Architecture:Any

[15 Sep 11:14] nanzhou lin
Description:
MySQL server crashes with SIGSEGV (signal 11) in the data dictionary cache infrastructure, specifically in `dd::cache::Shared_multi_map<dd::Routine>::Autolocker::~Autolocker()`. The crash occurs due to a race condition in the `Autolocker` destructor, which releases the mutex **before** executing deferred `delete` operations on evicted DD cache objects. This creates a window where another thread can access, evict, or re-delete the same objects, leading to use-after-free or double-free.

The root cause is in `sql/dd/impl/cache/shared_multi_map.h`, lines 160-173. The `Autolocker` destructor executes in this order:

1. `mysql_mutex_unlock(&m_map->m_lock)` — releases the lock
2. `delete` deferred objects — **executed without lock protection**
3. `delete` deferred elements — **executed without lock protection**

Between step 1 and step 2, another thread may acquire the lock and access the same `Shared_multi_map` instance (via `get()`, `put()`, `release()`, or `drop()`), potentially operating on the same cache elements that are pending deletion. This leads to use-after-free or double-free, causing SIGSEGV.

## Crash Backtrace

```
stack_bottom = 7fc81e617460 thread_stack 0x100000
/usr/local/engine/bin/mysqld(my_print_stacktrace(unsigned char const*, unsigned long)+0x3d) [0x557ebc5620bd]
/usr/local/engine/bin/mysqld(print_fatal_signal(int)+0x314) [0x557ebafa00a4]
/usr/local/engine/bin/mysqld(handle_fatal_signal+0x84) [0x557ebafa0134]
/lib64/libpthread.so.0(+0x13280) [0x7fce32ab8280]
/usr/local/engine/bin/mysqld(dd::cache::Shared_multi_map<dd::Routine>::Autolocker::~Autolocker()+0x66) [0x557ebc324b56]
/usr/local/engine/bin/mysqld(dd::cache::Shared_multi_map<dd::Routine>::release(dd::cache::Cache_element<dd::Routine>*)+0x6f) [0x557ebc33317f]
/usr/local/engine/bin/mysqld(dd::cache::Dictionary_client::Auto_releaser::~Auto_releaser()+0x16b8) [0x557ebac577d8]
/usr/local/engine/bin/mysqld(+0x184e254) [0x557ebb33d254]
/usr/local/engine/bin/mysqld(sp_cache_routine(THD*, enum_sp_type, sp_name const*, bool, sp_head**)+0xb4) [0x557ebb33ce84]
/usr/local/engine/bin/mysqld(sp_cache_routine(THD*, Sroutine_hash_entry*, bool, sp_head**)+0x61) [0x557ebb33cda1]
/usr/local/engine/bin/mysqld(+0x12e35ad) [0x557ebadd25ad]
/usr/local/engine/bin/mysqld(open_tables(THD*, TABLE_LIST**, unsigned int*, unsigned int, Prelocking_strategy*)+0xa96) [0x557ebab8c9f6]
/usr/local/engine/bin/mysqld(open_tables_for_query(THD*, TABLE_LIST*, unsigned int)+0x99) [0x557ebab91339]
/usr/local/engine/bin/mysqld(Sql_cmd_dml::prepare(THD*)+0x521) [0x557ebabc5b51]
/usr/local/engine/bin/mysqld(Sql_cmd_dml::execute(THD*)+0xff) [0x557ebabc4b2f]
/usr/local/engine/bin/mysqld(mysql_execute_command(THD*, bool)+0x3de) [0x557ebabb301e]
/usr/local/engine/bin/mysqld(dispatch_sql_command(THD*, Parser_state*)+0x4d0) [0x557ebabba2e0]
/usr/local/engine/bin/mysqld(dispatch_command(THD*, COM_DATA const*, enum_server_command)+0x558) [0x557ebabbb6a8]
/usr/local/engine/bin/mysqld(do_command(THD*)+0x11b) [0x557ebabbafab]
/usr/local/engine/bin/mysqld(threadpool_process_request(THD*)+0x98) [0x557ebad25718]
/usr/local/engine/bin/mysqld(+0x1237219) [0x557ebad26219]
/usr/local/engine/bin/mysqld(+0x3017be6) [0x557ebcb06be6]
/lib64/libpthread.so.0(+0x8f4b) [0x7fce32aadf4b]
/lib64/libc.so.6(clone+0x3f) [0x7fce3258f7ef]
```

How to repeat:
 High-concurrency workload with stored routines (procedures/functions)

Suggested fix:
// Unlock the multi map when being deleted (e.g. going out of scope)
// and delete the objects and elements.
~Autolocker() {
    mysql_mutex_unlock(&m_map->m_lock);     // Step 1: Lock released
    // Delete the objects.
    for (typename Object_list_type::const_iterator it =
             m_objects_to_delete.begin();
         it != m_objects_to_delete.end(); ++it)
      delete *it;                            // Step 2: Delete without lock

    // Delete the elements.
    for (typename Element_list_type::const_iterator it =
             m_elements_to_delete.begin();
         it != m_elements_to_delete.end(); ++it)
      delete *it;                            // Step 3: Delete without lock
}
[16 Sep 12:26] Shane Bester
Hello,

Can you tell us what version of mysqld you saw a crash on?

/usr/local/engine/bin/mysqld --version
sha256sum /usr/local/engine/bin/mysqld
readelf -n /usr/local/engine/bin/mysqld | grep -A1 'Build ID'
[17 Oct 1:00] Bugs System
No feedback was provided for this bug for over a month, so it is
being suspended automatically. If you are able to provide the
information that was originally requested, please do so and change
the status of the bug back to "Open".

Portions of this website are copyright © 2001, 2002 The PHP Group

Page generated in 0.050 sec. using MySQL 8.4.7-u6-cloud

Timestamp references displayed by the system are UTC.
Content reproduced on this site is the property of the respective copyright holders.
It is not reviewed in advance by Oracle and does not necessarily represent the opinion of Oracle or any other party.