make spinlocks degrade better when under high contention

dbtests/spin_lock_test.cpp

@@ -20,6 +20,7 @@
 #include <boost/thread/thread.hpp>
 #include "dbtests.h"
 #include "../util/concurrency/spin_lock.h"
+#include "../util/timer.h"
 
 namespace {
 
@@ -73,8 +74,10 @@ namespace {
             int counter = 0;
 
             const int threads = 64;
-            const int incs = 10000;
+            const int incs = 50000;
             LockTester* testers[threads];
+
+            Timer timer;
 
             for ( int i = 0; i < threads; i++ ) {
                 testers[i] = new LockTester( &spin, &counter );
@@ -87,7 +90,10 @@ namespace {
                 ASSERT_EQUALS( testers[i]->requests(), incs );
                 delete testers[i];
             }
-
+
+            int ms = timer.millis();
+            log() << "spinlock ConcurrentIncs time: " << ms << endl;
+
             ASSERT_EQUALS( counter, threads*incs );
 #if defined(__linux__)
             ASSERT( SpinLock::isfast() );

util/concurrency/spin_lock.cpp

@@ -45,7 +45,36 @@ namespace mongo {
 #if defined(_WIN32)
         EnterCriticalSection(&_cs);
 #elif defined(__USE_XOPEN2K)
-        pthread_spin_lock( &_lock );
+
+        /**
+         * this is designed to perform close to the default spin lock
+         * the reason for the mild insanity is to prevent horrible performance
+         * when contention spikes 
+         * it allows spinlocks to be used in many more places
+         * which is good because even with this change they are about 8x faster on linux
+         */
+
+        if ( pthread_spin_trylock( &_lock ) == 0 )
+            return;
+
+        for ( int i=0; i<1000; i++ ) 
+            if ( pthread_spin_trylock( &_lock ) == 0 )
+                return;
+
+        for ( int i=0; i<1000; i++ ) {
+            if ( pthread_spin_trylock( &_lock ) == 0 )
+                return;
+            pthread_yield();
+        }
+
+        struct timespec t;
+        t.tv_sec = 0;
+        t.tv_nsec = 5000000;
+
+        while ( pthread_spin_trylock( &_lock ) != 0 ) {
+            nanosleep(&t, NULL);
+        }
+
 #elif defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4)
         // fast path
         if (!_locked && !__sync_lock_test_and_set(&_locked, true)) {