semaphore plus:

(1) add an option for in-order resume of threads (by default not);
(2) avoid thundering herd of threads whose needs are not met;

thread/thread.cpp

@@ -1749,44 +1749,47 @@ R"(
     {
         if (count == 0) return 0;
         splock.lock();
+        DEFER(splock.unlock());
         CURRENT->semaphore_count = count;
-        int ret = 0;
         while (!try_subtract(count)) {
-            ret = waitq::wait_defer(timeout, spinlock_unlock, &splock);
-            ERRNO err;
-            splock.lock();
-            if (ret < 0) {
+            int ret = waitq::wait_defer(timeout, spinlock_unlock, &splock);
+            splock.lock();  // assuming errno NOT changed
+            if (unlikely(ret < 0)) {    // got interrupted
                 CURRENT->semaphore_count = 0;
-                // when timeout, we need to try to resume next thread(s) in q
-                if (err.no == ETIMEDOUT) try_resume();
-                splock.unlock();
-                errno = err.no;
+                if (!m_ooo_resume)
+                    try_resume(m_count.load());
                 return ret;
             }
         }
-        try_resume();
-        splock.unlock();
         return 0;
     }
-    void semaphore::try_resume()
-    {
-        auto cnt = m_count.load();
-        while(true)
-        {
+    void semaphore::try_resume(uint64_t cnt) {
+        if (cnt == 0) return;
+        while(true) {
             ScopedLockHead h(this);
-            if (!h) return;
+            if (!h) break;
             auto th = (thread*)h;
             auto& qfcount = th->semaphore_count;
             if (qfcount > cnt) break;
             cnt -= qfcount;
             qfcount = 0;
             prelocked_thread_interrupt(th, -1);
         }
+        if (!q.th || !cnt || !m_ooo_resume)
+            return;
+        for (auto th = q.th->next();
+                  th!= q.th && cnt;
+                  th = th->next()) {
+            auto& qfcount = th->semaphore_count;
+            if (qfcount <= cnt) {
+                cnt -= qfcount;
+                qfcount = 0;
+                prelocked_thread_interrupt(th, -1);
+            }
+        }
     }
-    bool semaphore::try_subtract(uint64_t count)
-    {
-        while(true)
-        {
+    inline bool semaphore::try_subtract(uint64_t count) {
+        while(true) {
             auto mc = m_count.load();
             if (mc < count)
                 return false;

thread/thread.h

@@ -432,7 +432,8 @@ namespace photon
     class semaphore : protected waitq
     {
     public:
-        explicit semaphore(uint64_t count = 0) : m_count(count) { }
+        explicit semaphore(uint64_t count = 0, bool in_order_resume = true)
+            : m_count(count), m_ooo_resume(!in_order_resume) { }
         int wait(uint64_t count, Timeout timeout = {}) {
             int ret = 0;
             do {
@@ -441,12 +442,11 @@ namespace photon
             return ret;
         }
         int wait_interruptible(uint64_t count, Timeout timeout = {});
-        int signal(uint64_t count)
-        {
+        int signal(uint64_t count) {
             if (count == 0) return 0;
             SCOPED_LOCK(splock);
-            m_count.fetch_add(count);
-            resume_one();
+            auto cnt = m_count.fetch_add(count) + count;
+            try_resume(cnt);
             return 0;
         }
         uint64_t count() const {
@@ -455,9 +455,10 @@ namespace photon
 
     protected:
         std::atomic<uint64_t> m_count;
+        bool m_ooo_resume;
         spinlock splock;
         bool try_subtract(uint64_t count);
-        void try_resume();
+        void try_resume(uint64_t count);
     };
 
     // to be different to timer flags