Regression in 1.1.1 against 1.1.0 in SSL_CTX_new

[Harald Koch]

> Am 16.04.2020 um 17:07 schrieb Tomas Mraz <tmraz at redhat.com>:
> 
> On Thu, 2020-04-16 at 15:42 +0200, Harald Koch wrote:
>> Hello list,
>> 
>> I have a TLS server which is started on demand in a multithreaded
>> (pthread) application. The TLS server is one thread which is being
>> started and stopped. At first start, the TLS server initialized with
>> SSL_CTX_new with TLS_server_method works as expected, after cleaning
>> up, eliminating the thread and starting it again at a later time in
>> the same process, SSL_CTX_new returns NULL. I’ve been digging deeper
>> into the initialization code, and found out that in
>> crypto/threads_pthread.c, function CRYPTO_THREAD_set_local the call
>> to pthread_setspecific returns a value != 0 (in my case: 22). The
>> error queue of openSSL stays empty. The same code works with openSSL
>> 1.1.0 in all versions.
>> Some posts googled state that before usage, be sure to run
>> OPENSSL_init_ssl (which I do, even if not required to my analysis
>> since it’s already called in one of the called functions deeper in
>> the library).
>> Am I missing something in a multithreaded environment?
> 
> Is this pure old 1.1.1 version or a current release from the 1.1.1
> branch (i.e. 1.1.1f)?
It’s 1.1.1f, also tested 1.1.1c. In 1.1.0t it works. I can test against other versions if you want to. As a speciality, I compile openSSL with gzip support („./config enable-zlib ...“, for support of compressed SMIME contents in other application).

> Do you call the OPENSSL_init_ssl from the main thread or from the TLS
> server thread?

I call it from the TLS server thread (created by pthread_create):

if (!OPENSSL_init_ssl(OPENSSL_INIT_LOAD_SSL_STRINGS, NULL))
	return;
I tried to do it only once (instead of every started thread): no difference.

In addition, I load random data via /dev/urandom (also tested only once or every time the server thread starts):
	RAND_load_file("/dev/urandom", 256);