avoid deadlock when using compile cache with blocking executors

- avoid taking cache lock in compile task destructor
- avoid holding cache lock while posting compile task
- add unit test trying to provoke deadlock
- bonus: fix blocking executor implementation
  (did not work with threads <= task limit)

2 files changed, 3 insertions, 3 deletions

diff --git a/searchcore/src/vespa/searchcore/proton/server/proton.cpp b/searchcore/src/vespa/searchcore/proton/server/proton.cpp
index 719ba359ccf..962ee65c10d 100644
--- a/searchcore/src/vespa/searchcore/proton/server/proton.cpp
+++ b/searchcore/src/vespa/searchcore/proton/server/proton.cpp
@@ -302,8 +302,8 @@ Proton::init(const BootstrapConfig::SP & configSnapshot)
     _warmupExecutor = std::make_unique<vespalib::ThreadStackExecutor>(4, 128*1024, index_warmup_executor);
 
     const size_t sharedThreads = deriveCompactionCompressionThreads(protonConfig, hwInfo.cpu());
-    _sharedExecutor = std::make_unique<vespalib::BlockingThreadStackExecutor>(sharedThreads, 128*1024, sharedThreads*16, proton_shared_executor);
-    _compile_cache_executor_binding = vespalib::eval::CompileCache::bind(*_sharedExecutor);
+    _sharedExecutor = std::make_shared<vespalib::BlockingThreadStackExecutor>(sharedThreads, 128*1024, sharedThreads*16, proton_shared_executor);
+    _compile_cache_executor_binding = vespalib::eval::CompileCache::bind(_sharedExecutor);
     InitializeThreads initializeThreads;
     if (protonConfig.initialize.threads > 0) {
         initializeThreads = std::make_shared<vespalib::ThreadStackExecutor>(protonConfig.initialize.threads, 128 * 1024, initialize_executor);
diff --git a/searchcore/src/vespa/searchcore/proton/server/proton.h b/searchcore/src/vespa/searchcore/proton/server/proton.h
index d0b76bd5804..d5c1a8b7b78 100644
--- a/searchcore/src/vespa/searchcore/proton/server/proton.h
+++ b/searchcore/src/vespa/searchcore/proton/server/proton.h
@@ -112,7 +112,7 @@ private:
     ProtonConfigurer                _protonConfigurer;
     ProtonConfigFetcher             _protonConfigFetcher;
     std::unique_ptr<vespalib::ThreadStackExecutorBase> _warmupExecutor;
-    std::unique_ptr<vespalib::ThreadStackExecutorBase> _sharedExecutor;
+    std::shared_ptr<vespalib::ThreadStackExecutorBase> _sharedExecutor;
     vespalib::eval::CompileCache::ExecutorBinding::UP _compile_cache_executor_binding;
     matching::QueryLimiter          _queryLimiter;
     vespalib::Clock                 _clock;