- Add support for giving out a global callgraph for all threads on level 2.

- You need level 3 to get a callgraph per thread.
You normally want to see the global callgraph prior to drilling in to each thread.

4 files changed, 21 insertions, 8 deletions

diff --git a/vespamalloc/src/vespamalloc/malloc/datasegment.hpp b/vespamalloc/src/vespamalloc/malloc/datasegment.hpp
index 366747ef880..6c9c1287ff8 100644
--- a/vespamalloc/src/vespamalloc/malloc/datasegment.hpp
+++ b/vespamalloc/src/vespamalloc/malloc/datasegment.hpp
@@ -164,14 +164,15 @@ void DataSegment<MemBlockPtrT>::returnBlock(void *ptr)
 template<typename MemBlockPtrT>
 size_t DataSegment<MemBlockPtrT>::infoThread(FILE * os, int level, uint32_t thread, SizeClassT sct) const
 {
-    typedef CallGraph<typename MemBlockPtrT::Stack, 0x10000, Index> CallGraphLT;
+    using CallGraphLT = CallGraph<typename MemBlockPtrT::Stack, 0x10000, Index>;
+    bool allThreads(thread == 0);
     size_t usedCount(0);
     size_t checkedCount(0);
     size_t allocatedCount(0);
     size_t notAccounted(0);
     size_t invalidCallStacks(0);
-    std::unique_ptr<CallGraphLT> callGraph(new CallGraphLT);
-    for(size_t i=0; i <  NELEMS(_blockList); ) {
+    std::unique_ptr<CallGraphLT> callGraph = std::make_unique<CallGraphLT>();
+    for (size_t i=0; i <  NELEMS(_blockList); ) {
         const BlockT & b = _blockList[i];
         SizeClassT sc = b.sizeClass();
         if (sc == sct) {
@@ -182,7 +183,7 @@ size_t DataSegment<MemBlockPtrT>::infoThread(FILE * os, int level, uint32_t thre
                 checkedCount++;
                 if (mem.allocated()) {
                     allocatedCount++;
-                    if (mem.threadId() == thread) {
+                    if (allThreads || (mem.threadId() == thread)) {
                         usedCount++;
                         if (usedCount < _allocs2Show) {
                             mem.info(os, level);
@@ -210,8 +211,9 @@ size_t DataSegment<MemBlockPtrT>::infoThread(FILE * os, int level, uint32_t thre
             i++;
         }
     }
-    fprintf(os, "\nCallTree(Checked=%ld, GlobalAlloc=%ld(%ld%%)," "ByMeAlloc=%ld(%2.2f%%) NotAccountedDue2FullGraph=%ld InvalidCallStacks=%ld:\n",
+    fprintf(os, "\nCallTree(Checked=%ld, GlobalAlloc=%ld(%ld%%)," "By%sAlloc=%ld(%2.2f%%) NotAccountedDue2FullGraph=%ld InvalidCallStacks=%ld:\n",
             checkedCount, allocatedCount, checkedCount ? allocatedCount*100/checkedCount : 0,
+            allThreads ? "Us" : "Me",
             usedCount, checkedCount ? static_cast<double>(usedCount*100)/checkedCount : 0.0, notAccounted, invalidCallStacks);
     if ( ! callGraph->empty()) {
         Aggregator agg;
diff --git a/vespamalloc/src/vespamalloc/malloc/threadlist.hpp b/vespamalloc/src/vespamalloc/malloc/threadlist.hpp
index a9a3a351899..aa514643265 100644
--- a/vespamalloc/src/vespamalloc/malloc/threadlist.hpp
+++ b/vespamalloc/src/vespamalloc/malloc/threadlist.hpp
@@ -29,16 +29,26 @@ void ThreadListT<MemBlockPtrT, ThreadStatT>::info(FILE * os, size_t level)
         const ThreadPool & thread = _threadVector[i];
         if (thread.isActive()) {
             activeThreads++;
+            peakThreads = i;
+        }
+    }
+    fprintf(os, "#%ld active threads. Peak threads #%ld\n", activeThreads, peakThreads);
+    if ((level > 1) && ! ThreadStatT::isDummy()) {
+        for (SizeClassT sc(0); sc < NUM_SIZE_CLASSES; sc++) {
+            _allocPool.dataSegment().infoThread(os, level, 0, sc);
+        }
+    }
+    for (size_t i(0); i < getMaxNumThreads(); i++) {
+        const ThreadPool & thread = _threadVector[i];
+        if (thread.isActive()) {
             if ( ! ThreadStatT::isDummy()) {
                 fprintf(os, "Thread #%ld = pid # %d\n", i, thread.osThreadId());
                 if (thread.isUsed()) {
                     thread.info(os, level, _allocPool.dataSegment());
                 }
             }
-            peakThreads = i;
         }
     }
-    fprintf(os, "#%ld active threads. Peak threads #%ld\n", activeThreads, peakThreads);
 }
 
 template <typename MemBlockPtrT, typename ThreadStatT>
diff --git a/vespamalloc/src/vespamalloc/malloc/threadpool.h b/vespamalloc/src/vespamalloc/malloc/threadpool.h
index 4b3483c45bc..792cf2c8108 100644
--- a/vespamalloc/src/vespamalloc/malloc/threadpool.h
+++ b/vespamalloc/src/vespamalloc/malloc/threadpool.h
@@ -69,6 +69,7 @@ private:
     ThreadStatT   _stat[NUM_SIZE_CLASSES];
     uint32_t      _threadId;
     std::atomic<ssize_t> _osThreadId;
+
     static SizeClassT _alwaysReuseSCLimit __attribute__((visibility("hidden")));
     static size_t     _threadCacheLimit __attribute__((visibility("hidden")));
 };
diff --git a/vespamalloc/src/vespamalloc/malloc/threadpool.hpp b/vespamalloc/src/vespamalloc/malloc/threadpool.hpp
index 2d3c020578b..4b5277a6396 100644
--- a/vespamalloc/src/vespamalloc/malloc/threadpool.hpp
+++ b/vespamalloc/src/vespamalloc/malloc/threadpool.hpp
@@ -28,7 +28,7 @@ void ThreadPoolT<MemBlockPtrT, ThreadStatT>::info(FILE * os, size_t level, const
             }
         }
     }
-    if (level > 1) {
+    if (level > 2) {
         fprintf(os, "BlockList:%ld,%ld,%ld\n", NELEMS(_stat), sizeof(_stat), sizeof(_stat[0]));
         size_t sum(0), sumLocal(0);
         for (size_t i=0; i < NELEMS(_stat); i++) {