In both clang-10 and clang-12, a WITH_MSAN build of the server would crash with SIGSEGV if the ENCRYPT() function is evaluated. The problem seems to be related to the MSAN interceptor of the crypt() function. I was unable to reproduce this with a simple test program:

#include <crypt.h>

#include <stdio.h>

int main(int argc, char **argv)

{

  printf("crypt: %s\n", crypt(argv[1], "123"));

  return 0;

}

clang-10 -fsanitize=undefined crypt.c -lcrypt

./a.out foo

I see that re-entrant versions of the C function crypt() exist, but we are using the original version with LOCK_crypt around it. Perhaps one of the other variants of the function would be intercepted better?

Note: No memory appears to be uninitialized. The following patch did not trigger anything; the very first sign of trouble was the SIGSEGV.

diff --git a/sql/item_strfunc.cc b/sql/item_strfunc.cc

index 0635556be40..3f9e023c793 100644

--- a/sql/item_strfunc.cc

+++ b/sql/item_strfunc.cc

@@ -2280,6 +2280,8 @@ String *Item_func_encrypt::val_str(String *str)

     salt_ptr= salt_str->c_ptr_safe();

   }

   mysql_mutex_lock(&LOCK_crypt);

+  MEM_CHECK_DEFINED(res->c_ptr_safe(), res->length);

+  MEM_CHECK_DEFINED(salt_ptr, 2);

   char *tmp= crypt(res->c_ptr_safe(),salt_ptr);

   if (!tmp)

   {

To work around this problem, I will move some encrypt() test cases from the files main.view and main.func_encrypt to the test main.func_crypt, which will be disabled in cmake -DWITH_MSAN=ON builds for now.