Deadlock Fix for HSA and Serialization Disable/Enabling support (#582)

* Initial barrier

* Working on profiler serializer extraction

* Current progress

* Serializtion Support

* source formatting (clang-format v11) (#583)

Co-authored-by: bwelton <1683479+bwelton@users.noreply.github.com>

* cmake formatting (cmake-format) (#584)

Co-authored-by: bwelton <1683479+bwelton@users.noreply.github.com>

* Minor fix

* Current Progress

* Current progress

* More fixes

* Serialization Fixes

* Bug fix

* source formatting (clang-format v11) (#600)

Co-authored-by: bwelton <1683479+bwelton@users.noreply.github.com>

* More fixes

* More minor fixes

* source formatting (clang-format v11) (#603)

Co-authored-by: bwelton <1683479+bwelton@users.noreply.github.com>

* source formatting (clang-format v11) (#604)

Co-authored-by: bwelton <1683479+bwelton@users.noreply.github.com>

* Lock order inversion false positive

* order fix

* More changes

* source formatting (clang-format v11) (#607)

Co-authored-by: bwelton <1683479+bwelton@users.noreply.github.com>

* minor test fix

* Minor test changes

---------

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: bwelton <1683479+bwelton@users.noreply.github.com>

source/lib/rocprofiler-sdk/hsa/queue_controller.hpp

@@ -24,6 +24,7 @@
 
 #include <rocprofiler-sdk/rocprofiler.h>
 
+#include "lib/rocprofiler-sdk/hsa/profile_serializer.hpp"
 #include "lib/rocprofiler-sdk/hsa/queue.hpp"
 
 #include <cstdint>
@@ -36,27 +37,6 @@ namespace rocprofiler
 {
 namespace hsa
 {
-/*This is a profiler serializer. It should be instantiated
-only once for the profiler. The following is the
-description of each field.
-1. dispatch_queue - The queue to which the currently dispatched kernel
-        belongs to.
-        At any given time, in serialization only one kernel
-        can be executing.
-2. dispatch_ready- It is a software data structure which holds
-        the queues which have a kernel ready to be dispatched.
-        This stores the queues in FIFO order.
-3. serializer_mutex - The mutex is used for thread synchronization
-        while accessing the singleton instance of this structure.
-Currently, in case of profiling kernels are serialized by default.
-*/
-struct profiler_serializer_t
-{
-    const Queue*             dispatch_queue{nullptr};
-    std::deque<const Queue*> dispatch_ready;
-    bool                     enabled{true};
-};
-
 // Tracks and manages HSA queues
 class QueueController
 {
@@ -65,6 +45,7 @@ public:
         std::tuple<rocprofiler_agent_t, Queue::queue_cb_t, Queue::completed_cb_t>;
     using queue_iterator_cb_t    = std::function<void(const Queue*)>;
     using callback_iterator_cb_t = std::function<void(ClientID, const agent_callback_tuple_t&)>;
+    using queue_map_t            = std::unordered_map<hsa_queue_t*, std::unique_ptr<Queue>>;
 
     QueueController() = default;
     // Initializes the QueueInterceptor. This must be delayed until
@@ -86,7 +67,7 @@ public:
     const CoreApiTable& get_core_table() const { return _core_table; }
     const AmdExtTable&  get_ext_table() const { return _ext_table; }
 
-    // Gets the list of supported HSA agents that can be intercepted
+    // Gets the list of supported HSA agents that can be Pintercepted
     const auto& get_supported_agents() const { return _supported_agents; }
     auto&       get_supported_agents() { return _supported_agents; }
 
@@ -94,32 +75,41 @@ public:
 
     void iterate_queues(const queue_iterator_cb_t&) const;
     void set_queue_state(queue_state state, hsa_queue_t* hsa_queue);
-    void profiler_serializer_register_ready_signal_handler(const hsa_signal_t& signal,
-                                                           void*               data) const;
+
     void add_dispatch_ready(const Queue* queue);
-    template <typename FuncT>
-    void profiler_serializer(FuncT&& lambda);
 
     void iterate_callbacks(const callback_iterator_cb_t&) const;
 
+    common::Synchronized<hsa::profiler_serializer>& serializer() { return _profiler_serializer; }
+
     /**
      * Disable serialization for QueueController, has no effect if counter collection
      * is not in use (which defaults to no serialization mechanism). Should only be used for
      * testing.
      */
+    void enable_serialization();
     void disable_serialization();
 
+    // Prints current state of signals for queues, used for debugging. Only prints
+    // serialization related signals if not compiled in debug mode.
+    void print_debug_signals() const;
+
+#if !defined(NDEBUG)
+    // Tracks the creation of all signals in queues, used for debugging and disabled
+    // in release mode (adds locking around signal creation).
+    common::Synchronized<std::unordered_map<uint64_t, hsa_signal_t>> _debug_signals;
+#endif
+
 private:
-    using queue_map_t       = std::unordered_map<hsa_queue_t*, std::unique_ptr<Queue>>;
     using client_id_map_t   = std::unordered_map<ClientID, agent_callback_tuple_t>;
     using agent_cache_map_t = std::unordered_map<uint32_t, AgentCache>;
 
-    CoreApiTable                                _core_table       = {};
-    AmdExtTable                                 _ext_table        = {};
-    common::Synchronized<queue_map_t>           _queues           = {};
-    common::Synchronized<client_id_map_t>       _callback_cache   = {};
-    agent_cache_map_t                           _supported_agents = {};
-    common::Synchronized<profiler_serializer_t> _profiler_serializer;
+    CoreApiTable                                   _core_table       = {};
+    AmdExtTable                                    _ext_table        = {};
+    common::Synchronized<queue_map_t>              _queues           = {};
+    common::Synchronized<client_id_map_t>          _callback_cache   = {};
+    agent_cache_map_t                              _supported_agents = {};
+    common::Synchronized<hsa::profiler_serializer> _profiler_serializer;
 };
 
 QueueController&