Merge pull request #16825 from ydb-platform/merge-libs-250407-0050

8 files changed, 35 insertions, 10 deletions

diff --git a/library/cpp/blockcodecs/codecs.h b/library/cpp/blockcodecs/codecs.h
index fd499b54b0..43a9244465 100644
--- a/library/cpp/blockcodecs/codecs.h
+++ b/library/cpp/blockcodecs/codecs.h
@@ -1,3 +1,3 @@
 #pragma once
 
-#include <library/cpp/blockcodecs/core/codecs.h>
+#include <library/cpp/blockcodecs/core/codecs.h>  // IWYU pragma: export
diff --git a/library/cpp/json/json_value.h b/library/cpp/json/json_value.h
index f70f4d2ee8..47273f6b13 100644
--- a/library/cpp/json/json_value.h
+++ b/library/cpp/json/json_value.h
@@ -1,3 +1,3 @@
 #pragma once
 
-#include <library/cpp/json/writer/json_value.h>
+#include <library/cpp/json/writer/json_value.h>  // IWYU pragma: export
diff --git a/library/cpp/svnversion/svnversion.h b/library/cpp/svnversion/svnversion.h
index b99615daa9..b1f8de5723 100644
--- a/library/cpp/svnversion/svnversion.h
+++ b/library/cpp/svnversion/svnversion.h
@@ -10,4 +10,4 @@
 #include <util/system/compiler.h>
 
 // Automatically generated functions.
-#include "build/scripts/c_templates/svnversion.h"
+#include <build/scripts/c_templates/svnversion.h>  // IWYU pragma: export
diff --git a/library/cpp/testing/gtest_extensions/gtest_extensions.h b/library/cpp/testing/gtest_extensions/gtest_extensions.h
index e20532241e..e4f336fd54 100644
--- a/library/cpp/testing/gtest_extensions/gtest_extensions.h
+++ b/library/cpp/testing/gtest_extensions/gtest_extensions.h
@@ -1,6 +1,8 @@
 #pragma once
 
+// IWYU pragma: begin_exports
 #include "assertions.h"
 #include "matchers.h"
 #include "pretty_printers.h"
 #include "probe.h"
+// IWYU pragma: end_exports
diff --git a/library/cpp/threading/future/future.h b/library/cpp/threading/future/future.h
index 35db9abbe2..91ade3f6e5 100644
--- a/library/cpp/threading/future/future.h
+++ b/library/cpp/threading/future/future.h
@@ -1,4 +1,6 @@
 #pragma once
 
+// IWYU pragma: begin_exports
 #include "core/future.h"
 #include "wait/wait.h"
+// IWYU pragma: end_exports
diff --git a/library/cpp/threading/local_executor/local_executor.h b/library/cpp/threading/local_executor/local_executor.h
index c1c824f67c..4eb67b034e 100644
--- a/library/cpp/threading/local_executor/local_executor.h
+++ b/library/cpp/threading/local_executor/local_executor.h
@@ -54,8 +54,12 @@ namespace NPar {
         // @param flags                     Same as for `Exec`.
         virtual void ExecRange(TIntrusivePtr<ILocallyExecutable> exec, int firstId, int lastId, int flags) = 0;
 
-        // 0-based ILocalExecutor worker thread identification
+        // returns:
+        //   0 for for a thread outside the internal pool
+        //      (because ILocalExecutor is allowed to use a calling thread to execute tasks as well),
+        //   1 ... GetThreadCount() for a thread inside the internal pool
         virtual int GetWorkerThreadId() const noexcept = 0;
+
         virtual int GetThreadCount() const noexcept = 0;
 
         // Describes a range of tasks with parameters from integer range [FirstId, LastId).
@@ -237,7 +241,6 @@ namespace NPar {
         int GetLPQueueSize() const noexcept;
         void ClearLPQueue();
 
-        // 0-based TLocalExecutor worker thread identification
         int GetWorkerThreadId() const noexcept override;
         int GetThreadCount() const noexcept override;
 
diff --git a/library/cpp/threading/local_executor/tbb_local_executor.cpp b/library/cpp/threading/local_executor/tbb_local_executor.cpp
index 65d6659443..91a8460b0e 100644
--- a/library/cpp/threading/local_executor/tbb_local_executor.cpp
+++ b/library/cpp/threading/local_executor/tbb_local_executor.cpp
@@ -14,9 +14,21 @@ int NPar::TTbbLocalExecutor<RespectTls>::GetThreadCount() const noexcept {
 
 template <bool RespectTls>
 int NPar::TTbbLocalExecutor<RespectTls>::GetWorkerThreadId() const noexcept {
-    return TbbArena.execute([] {
-        return tbb::this_task_arena::current_thread_index();
-    });
+    static thread_local int WorkerThreadId = -1;
+    if (WorkerThreadId == -1) {
+        // Can't rely on return value except checking that it is 'not_initialized' because of
+        //  "Since a thread may exit the arena at any time if it does not execute a task, the index of
+        //   a thread may change between any two tasks"
+        //  (https://oneapi-spec.uxlfoundation.org/specifications/oneapi/latest/elements/onetbb/source/task_scheduler/task_arena/this_task_arena_ns#_CPPv4N3tbb15this_task_arena20current_thread_indexEv)
+        const auto tbbThreadIndex = tbb::this_task_arena::current_thread_index();
+        if (tbbThreadIndex == tbb::task_arena::not_initialized) {
+            // This thread does not belong to TBB worker threads
+            WorkerThreadId = 0;
+        } else {
+            WorkerThreadId = ++RegisteredThreadCounter;
+        }
+    }
+    return WorkerThreadId;
 }
 
 template <bool RespectTls>
diff --git a/library/cpp/threading/local_executor/tbb_local_executor.h b/library/cpp/threading/local_executor/tbb_local_executor.h
index 8d790db18c..f67c07349d 100644
--- a/library/cpp/threading/local_executor/tbb_local_executor.h
+++ b/library/cpp/threading/local_executor/tbb_local_executor.h
@@ -9,6 +9,9 @@
 #include <contrib/libs/tbb/include/tbb/task_arena.h>
 #include <contrib/libs/tbb/include/tbb/task_group.h>
 
+#include <atomic>
+
+
 namespace NPar {
     template <bool RespectTls = false>
     class TTbbLocalExecutor final: public ILocalExecutor  {
@@ -16,10 +19,11 @@ namespace NPar {
         TTbbLocalExecutor(int nThreads)
             : ILocalExecutor()
             , TbbArena(nThreads)
-            , NumberOfTbbThreads(nThreads) {}
+            , NumberOfTbbThreads(nThreads)
+            , RegisteredThreadCounter(0)
+        {}
         ~TTbbLocalExecutor() noexcept override {}
 
-        // 0-based ILocalExecutor worker thread identification
         virtual int GetWorkerThreadId() const noexcept override;
         virtual int GetThreadCount() const noexcept override;
 
@@ -45,5 +49,7 @@ namespace NPar {
         mutable tbb::task_arena TbbArena;
         tbb::task_group Group;
         int NumberOfTbbThreads;
+
+        mutable std::atomic_int RegisteredThreadCounter;
     };
 }