Fixes for high CPU counts

Allow any number of build threads.

Use _SC_NPROCESSORS_ONLN to find online CPUs only:

Some Linux kernels are configured to allow "CPU hotswap" which means
they pretend to have some huge array (128) of CPUs that might come
online. This is weird. If we ask for the number of online CPUs that
should be a saner value.

std::thread::hardware_concurrency() has the same issue, I think it uses the total
count, which seems like a pretty bad bug in itself but outside the scope
of this project.

src/BuildQueue.cpp

@@ -735,15 +735,10 @@ namespace t2
     queue->m_ExpensiveWaitCount = 0;
     queue->m_ExpensiveWaitList  = HeapAllocateArray<NodeState*>(heap, capacity);
 
-    CHECK(queue->m_Queue);
-
-    if (queue->m_Config.m_ThreadCount > kMaxBuildThreads)
-    {
-      Log(kWarning, "too many build threads (%d) - clamping to %d",
-          queue->m_Config.m_ThreadCount, kMaxBuildThreads);
+    queue->m_Threads = HeapAllocateArrayZeroed<ThreadId>(config->m_Heap, config->m_ThreadCount);
+    queue->m_ThreadState = HeapAllocateArrayZeroed<ThreadState>(config->m_Heap, config->m_ThreadCount);
 
-      queue->m_Config.m_ThreadCount = kMaxBuildThreads;
-    }
+    CHECK(queue->m_Queue);
 
     Log(kDebug, "build queue initialized; ring buffer capacity = %u", queue->m_QueueCapacity);
 
@@ -797,6 +792,9 @@ namespace t2
     CondDestroy(&queue->m_WorkAvailable);
     MutexDestroy(&queue->m_Lock);
 
+    HeapFree(config->m_Heap, queue->m_ThreadState);
+    HeapFree(config->m_Heap, queue->m_Threads);
+
     // Unblock all signals on the main thread.
     SignalHandlerSetCondition(nullptr);
     SignalBlockThread(false);

src/BuildQueue.hpp

@@ -17,11 +17,6 @@ namespace t2
   struct StatCache;
   struct DigestCache;
 
-  enum
-  {
-    kMaxBuildThreads = 64
-  };
-
   struct BuildQueueConfig
   {
     enum
@@ -73,8 +68,8 @@ namespace t2
     int32_t            m_PendingNodeCount;
     int32_t            m_FailedNodeCount;
     int32_t            m_CurrentPassIndex;
-    ThreadId           m_Threads[kMaxBuildThreads];
-    ThreadState        m_ThreadState[kMaxBuildThreads];
+    ThreadId          *m_Threads;
+    ThreadState       *m_ThreadState;
     int32_t            m_ExpensiveRunning;
     int32_t            m_ExpensiveWaitCount;
     NodeState        **m_ExpensiveWaitList;

src/Common.cpp

@@ -24,10 +24,6 @@
 #include <sys/sysctl.h>
 #endif
 
-#if defined(TUNDRA_LINUX)
-#include <thread>
-#endif
-
 #if defined(TUNDRA_WIN32)
 #include <windows.h>
 #include <ctype.h>
@@ -346,9 +342,7 @@ int GetCpuCount()
   GetSystemInfo(&si);
   return (int) si.dwNumberOfProcessors;
 #elif defined(TUNDRA_LINUX)
-  return (int)std::thread::hardware_concurrency();
-#else
-  long nprocs_max = sysconf(_SC_NPROCESSORS_CONF);
+  long nprocs_max = sysconf(_SC_NPROCESSORS_ONLN);
   if (nprocs_max < 0)
     CroakErrno("couldn't get CPU count");
   return (int) nprocs_max;

src/Exec.hpp

@@ -15,7 +15,7 @@ namespace t2
     bool    m_WasSignalled;
   };
 
-  void ExecInit(void);
+  void ExecInit(int thread_count);
 
   ExecResult ExecuteProcess(
         const char*         cmd_line,

src/ExecUnix.cpp

@@ -37,7 +37,7 @@ static void SetFdNonBlocking(int fd)
 		CroakErrno("couldn't unblock fd %d", fd);
 }
 
-void ExecInit(void)
+void ExecInit(int thread_count)
 {
 	TerminalIoInit();
 }

src/ExecWin32.cpp

@@ -45,7 +45,7 @@ static char              s_TemporaryDir[MAX_PATH];
 static DWORD             s_TundraPid;
 static Mutex             s_FdMutex;
 
-static HANDLE s_TempFiles[kMaxBuildThreads];
+static HANDLE           *s_TempFiles;
 
 static HANDLE AllocFd(int job_id)
 {
@@ -176,8 +176,9 @@ static char UTF8_WindowsEnvironment[128*1024];
 
 static size_t g_Win32EnvCount;
 
-void ExecInit(void)
+void ExecInit(int thread_count)
 {
+  s_TempFiles = (HANDLE*) calloc(thread_count, sizeof s_TempFiles[0]);
   s_TundraPid = GetCurrentProcessId();
 
   if (0 == GetTempPathA(sizeof(s_TemporaryDir), s_TemporaryDir))

src/Hash.cpp

@@ -65,8 +65,8 @@ void HashAddStringFoldCase(HashState* self, const char* path)
     char c = *path++;
     if (c == 0)
       return;
-      c = FoldCase(c);
-      HashUpdate(self, &c, 1);
+    c = FoldCase(c);
+    HashUpdate(self, &c, 1);
   }
 }
 

src/Main.cpp

@@ -334,7 +334,7 @@ int main(int argc, char* argv[])
 
   uint64_t start_time = TimerGet();
 
-  ExecInit();
+  ExecInit(options.m_ThreadCount);
 
   if (options.m_WorkingDir)
   {