rocr: GPU core file location support (#1732)

* rocr: WIP Support dump of GPU core file

* WIP new core dump tests compile

* WIP: anony namespaces, test updates, progress

Added disabled Fault test. Other non-disabled coredump tests don't work.

* WIP: address code review feedback

* WIP: gpu core dump rocrtst works; combined

* WIP: remove rocrtst changes for this commit

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_aql_queue.cpp

@@ -1323,8 +1323,8 @@ bool AqlQueue::ExceptionHandler(hsa_signal_value_t error_code, void* arg) {
     return exceptionHandlerDone();
   }
 
-  // Fallback if KFD does not support GPU core dump. In this case, there core dump is
-  // generated by hsa-runtime.
+  // Fallback if KFD does not support GPU core dump. In this case, the core
+  // dump is generated by hsa-runtime.
   if (!core::Runtime::runtime_singleton_->KfdVersion().supports_core_dump &&
                 queue->agent_->supported_isas()[0]->GetMajorVersion() != 11) {
 

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/runtime.cpp

@@ -2236,7 +2236,7 @@ bool Runtime::VMFaultHandler(hsa_signal_value_t val, void* arg) {
       PrintMemoryMapNear(reinterpret_cast<void*>(fault.VirtualAddress));
 #endif
     }
-    // Fallback if KFD does not support GPU core dump. In this case, there core dump is
+    // Fallback if KFD does not support GPU core dump. In this case, the core dump is
     // generated by hsa-runtime.
     if (faulty_agent &&
         faulty_agent->supported_isas()[0]->GetMajorVersion() != 11 &&

projects/rocr-runtime/runtime/hsa-runtime/core/util/flag.h

@@ -298,6 +298,14 @@ class Flag {
 
     var = os::GetEnvVar("HSA_CO_DMACOPY_SIZE");
     co_dmacopy_size_ = var.empty() ? 1024*1024 : atoi(var.c_str());
+
+    var = os::GetEnvVar("HSA_COREDUMP_SHOW_PROGRESS");
+    enable_core_dump_progress_ = (var == "1");
+
+    var = os::GetEnvVar("HSA_DISABLE_COREDUMP_ON_EXCEPTION");
+    core_dump_disable_ = (var == "1");
+
+    core_dump_pattern_ = os::GetEnvVar("HSA_COREDUMP_PATTERN");
   }
 
   void parse_masks(uint32_t maxGpu, uint32_t maxCU) {
@@ -430,6 +438,17 @@ class Flag {
 
   bool enable_dxg_detection() const { return enable_dxg_detection_; }
 
+  [[nodiscard]]
+  bool core_dump_disable() const { return core_dump_disable_; }
+
+  [[nodiscard]]
+  bool enable_core_dump_progress() const {
+                                       return enable_core_dump_progress_; }
+
+  [[nodiscard]]
+  const std::string& core_dump_pattern() const {
+                                         return core_dump_pattern_; }
+
   void set_sdma(bool peer_sdma, bool sdma_gang) {
     enable_peer_sdma_ = peer_sdma ? SDMA_ENABLE : SDMA_DISABLE;
     enable_sdma_gang_ = sdma_gang ? SDMA_ENABLE : SDMA_DISABLE;
@@ -522,6 +541,10 @@ class Flag {
 
   size_t co_dmacopy_size_;
 
+  bool core_dump_disable_ = false;
+  bool enable_core_dump_progress_ = false;
+  std::string core_dump_pattern_;
+
   // Map GPU index post RVD to its default cu mask.
   std::map<uint32_t, std::vector<uint32_t>> cu_mask_;