core dump: Generates a core dump from a fault event

Extracts and creates a core dump ELF file from a fault event, using
core dump front end.

Signed-off-by: Alex Sierra <Alex.Sierra@amd.com>
Change-Id: Ibbbe41b3d13dd3fcb90161e927d48c329cf513a9


[ROCm/ROCR-Runtime commit: 803e37ded5]

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

@@ -67,6 +67,7 @@
 #include "core/inc/default_signal.h"
 #include "core/inc/hsa_ext_amd_impl.h"
 #include "core/inc/amd_gpu_pm4.h"
+#include "core/inc/amd_core_dump.hpp"
 
 namespace rocr {
 namespace AMD {
@@ -1231,6 +1232,17 @@ bool AqlQueue::ExceptionHandler(hsa_signal_value_t error_code, void* arg) {
     return false;
   }
 
+  // Fallback if KFD does not support GPU core dump. In this case, there core dump is
+  // generated by hsa-runtime.
+  if (!core::Runtime::runtime_singleton_->KfdVersion().supports_core_dump) {
+    if (amd::coredump::dump_gpu_core())
+      debug_print("GPU core dump failed\n");
+    // supports_core_dump flag is overwritten to avoid generate core dump file again
+    // caught by a different exception handler. Such as VMFaultHandler.
+    core::Runtime::runtime_singleton_->KfdVersion(
+      core::Runtime::runtime_singleton_->KfdVersion().supports_exception_debugging, true);
+  }
+
   for (auto& error : QueueErrors) {
     if (error_code & (1 << (error.code - 1))) {
       errorCode = error.status;

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

@@ -70,6 +70,7 @@
 #include "core/util/os.h"
 #include "core/inc/exceptions.h"
 #include "inc/hsa_ven_amd_aqlprofile.h"
+#include "core/inc/amd_core_dump.hpp"
 
 #ifndef HSA_VERSION_MAJOR
 #define HSA_VERSION_MAJOR 1
@@ -1358,6 +1359,13 @@ bool Runtime::VMFaultHandler(hsa_signal_value_t val, void* arg) {
   HsaMemoryAccessFault& fault =
       vm_fault_event->EventData.EventData.MemoryAccessFault;
 
+  // Fallback if KFD does not support GPU core dump. In this case, there core dump is
+  // generated by hsa-runtime.
+  if (!runtime_singleton_->KfdVersion().supports_core_dump) {
+    if (amd::coredump::dump_gpu_core())
+      debug_print("GPU core dump failed\n");
+  }
+
   hsa_status_t custom_handler_status = HSA_STATUS_ERROR;
   auto system_event_handlers = runtime_singleton_->GetSystemEventHandlers();
   // If custom handler is registered, pack the fault info and call the handler