Details
-
Bug
-
Status: Closed (View Workflow)
-
Major
-
Resolution: Fixed
-
10.5, 10.3.0, 10.4.2, 10.1(EOL), 10.2(EOL)
Description
Several tests fail on 10.4 kvm-asan due to stack overrun:
10.4 af912664989e0c3ee9cdb6caf8ec439029e7405c |
Failing test(s): sys_vars.max_sp_recursion_depth_func main.signal_demo3 compat/oracle.sp-package rpl.rpl_row_sp011 main.sp
|
The main contributor to this stack overrun is mysql_execute_command().
Rebuilding with cmake -DWITH_WSREP=OFF shrunk the size of the mysql_execute_command() stack frame in the ASAN debug build from 29312 to 17568 bytes. A quick and dirty workaround made the tests pass:
diff --git a/sql/wsrep_mysqld.h b/sql/wsrep_mysqld.h
|
--- a/sql/wsrep_mysqld.h
|
+++ b/sql/wsrep_mysqld.h
|
@@ -245,7 +245,7 @@ extern wsrep_seqno_t wsrep_locked_seqno;
|
// prefix all messages with "WSREP"
|
#define WSREP_LOG(fun, ...) \
|
do { \
|
- char msg[1024]= {'\0'}; \
|
+ char msg[1]= {'\0'}; \
|
snprintf(msg, sizeof(msg) - 1, ## __VA_ARGS__); \
|
fun("WSREP: %s", msg); \
|
} while(0) |
It turns out that this bug had been fixed in MariaDB 10.3.1 but the fix was destroyed in the Galera 4 merge (MDEV-16405).
This bug should affect all builds, especially affecting the execution of complex stored procedures, which involves recursive calls to mysql_execute_command(). The WSREP_DEBUG macro invocations will expand to WSREP_LOG also in non-debug builds, and obviously, the stack memory will be allocated even with the default settings (WSREP or Galera are disabled).