Optimize Windows RCU thread signalling.

With the pthread variant, a thread truly wakes up after
the pthread_mutex_unlock call, even if pthread_cond_signal
is called before.

This is not true for the Windows variant. The thread is
woken up in WakeConditionVariable but goes back to sleep.
Reordering (signalling thread after unlocking) should save
some time during transitions and should be safe in this context.

The speedup is visible on lhash_test, running on many CPUs
(on 32 cores, a speedup from 1:40 to 1:05 minutes on test hw).

Co-Authored-By: Claude Opus 4.6 Extended <noreply@anthropic.com>
Signed-off-by: Milan Broz <gmazyland@gmail.com>
Reviewed-by: Saša Nedvědický <sashan@openssl.org>
Reviewed-by: Nikola Pajkovsky <nikolap@openssl.org>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
MergeDate: Fri Mar 13 17:25:49 2026
(Merged from https://github.com/openssl/openssl/pull/30388)

(cherry picked from commit cefa8bb0ac6b2a83b4e57de2b5c40ac939cf1f95)

diff --git a/crypto/threads_win.c b/crypto/threads_win.c
index f96c625a3b694f117bddfc64fcdf52c98c5b78f2..db24796009846a13647d5ee06a586a6a83e640a8 100644 (file)
--- a/crypto/threads_win.c
+++ b/crypto/threads_win.c
@@ -347,9 +347,9 @@ static struct rcu_qp *update_qp(CRYPTO_RCU_LOCK *lock, uint32_t *curr_id)
     InterlockedExchange((LONG volatile *)&lock->reader_idx, tmp);
 #endif
 
+    ossl_crypto_mutex_unlock(lock->alloc_lock);
     /* wake up any waiters */
     ossl_crypto_condvar_signal(lock->alloc_signal);
-    ossl_crypto_mutex_unlock(lock->alloc_lock);
     return &lock->qp_group[current_idx];
 }
 
@@ -358,8 +358,8 @@ static void retire_qp(CRYPTO_RCU_LOCK *lock,
 {
     ossl_crypto_mutex_lock(lock->alloc_lock);
     lock->writers_alloced--;
-    ossl_crypto_condvar_signal(lock->alloc_signal);
     ossl_crypto_mutex_unlock(lock->alloc_lock);
+    ossl_crypto_condvar_signal(lock->alloc_signal);
 }
 
 void ossl_synchronize_rcu(CRYPTO_RCU_LOCK *lock)
@@ -388,8 +388,8 @@ void ossl_synchronize_rcu(CRYPTO_RCU_LOCK *lock)
     } while (count != (uint64_t)0);
 
     lock->next_to_retire++;
-    ossl_crypto_condvar_broadcast(lock->prior_signal);
     ossl_crypto_mutex_unlock(lock->prior_lock);
+    ossl_crypto_condvar_broadcast(lock->prior_signal);
 
     retire_qp(lock, qp);