Align Windows RCU implementation to the pthread variant

Unlike the pthread variant, Windows RCU uses broadcast instead
of targeted signal calls in some places, unnecessarily increasing
the number of used cycles.

The retire_qp should wake up only one thread to proceed, not
all of them.  For update_qp, that signals the thread after
increasing writers_alloced, signalling all threads does not make
sense either.

The speedup is significant on lhash_test, running on many CPUs
(on 32 cores, a speedup from 6:20 to 1:40 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 17:25:47 2026
(Merged from https://github.com/openssl/openssl/pull/30388)

diff --git a/crypto/threads_win.c b/crypto/threads_win.c
index 8b453635720143eb938d243a3b8772448c341a0b..f96c625a3b694f117bddfc64fcdf52c98c5b78f2 100644 (file)
--- a/crypto/threads_win.c
+++ b/crypto/threads_win.c
@@ -348,7 +348,7 @@ static struct rcu_qp *update_qp(CRYPTO_RCU_LOCK *lock, uint32_t *curr_id)
 #endif
 
     /* wake up any waiters */
-    ossl_crypto_condvar_broadcast(lock->alloc_signal);
+    ossl_crypto_condvar_signal(lock->alloc_signal);
     ossl_crypto_mutex_unlock(lock->alloc_lock);
     return &lock->qp_group[current_idx];
 }
@@ -358,7 +358,7 @@ static void retire_qp(CRYPTO_RCU_LOCK *lock,
 {
     ossl_crypto_mutex_lock(lock->alloc_lock);
     lock->writers_alloced--;
-    ossl_crypto_condvar_broadcast(lock->alloc_signal);
+    ossl_crypto_condvar_signal(lock->alloc_signal);
     ossl_crypto_mutex_unlock(lock->alloc_lock);
 }