From de3912dae4d8da6d1cf3ff1938bf6bd9299512be Mon Sep 17 00:00:00 2001
From: Sean Keely <Sean.Keely@amd.com>
Date: Wed, 25 Oct 2017 20:08:36 -0500
Subject: [PATCH] Emit fragment map and thunk ptr info with VM faults.

Change-Id: If1302f674df7a636529c64bf66dfdda755a70c32


[ROCm/ROCR-Runtime commit: 9212e7a09f211d51b047c82774984ac4a2c4c3c8]
---
 .../hsa-runtime/core/runtime/runtime.cpp      | 68 ++++++++++++++++---
 1 file changed, 60 insertions(+), 8 deletions(-)

diff --git a/projects/rocr-runtime/runtime/hsa-runtime/core/runtime/runtime.cpp b/projects/rocr-runtime/runtime/hsa-runtime/core/runtime/runtime.cpp
index 497d9b85cd..553300265d 100644
--- a/projects/rocr-runtime/runtime/hsa-runtime/core/runtime/runtime.cpp
+++ b/projects/rocr-runtime/runtime/hsa-runtime/core/runtime/runtime.cpp
@@ -1101,15 +1101,67 @@ bool Runtime::VMFaultHandler(hsa_signal_value_t val, void* arg) {
         reason += "Unknown";
       }
 
-      fprintf(stderr,
-              "Memory access fault by GPU node-%u on address %p%s. Reason: %s.\n",
-              fault.NodeId, reinterpret_cast<const void*>(fault.VirtualAddress),
-              (fault.Failure.Imprecise == 1) ? "(may not be exact address)" : "",
-              reason.c_str());
-    } else {
-      assert(false && "GPU memory access fault.");
-    }
+      core::Agent* faultingAgent = runtime_singleton_->agents_by_node_[fault.NodeId][0];
 
+      fprintf(
+          stderr,
+          "Memory access fault by GPU node-%u (Agent handle: %p) on address %p%s. Reason: %s.\n",
+          fault.NodeId, reinterpret_cast<void*>(faultingAgent->public_handle().handle),
+          reinterpret_cast<const void*>(fault.VirtualAddress),
+          (fault.Failure.Imprecise == 1) ? "(may not be exact address)" : "", reason.c_str());
+
+#ifndef NDEBUG
+      runtime_singleton_->memory_lock_.Acquire();
+      auto it = runtime_singleton_->allocation_map_.upper_bound(
+          reinterpret_cast<void*>(fault.VirtualAddress));
+      for (int i = 0; i < 2; i++) {
+        if (it != runtime_singleton_->allocation_map_.begin()) it--;
+      }
+      fprintf(stderr, "Nearby memory map:\n");
+      auto start = it;
+      for (int i = 0; i < 3; i++) {
+        if (it == runtime_singleton_->allocation_map_.end()) break;
+        std::string kind = "Non-HSA";
+        if (it->second.region != nullptr) {
+          const amd::MemoryRegion* region =
+              static_cast<const amd::MemoryRegion*>(it->second.region);
+          if (region->IsSystem())
+            kind = "System";
+          else if (region->IsLocalMemory())
+            kind = "VRAM";
+          else if (region->IsScratch())
+            kind = "Scratch";
+          else if (region->IsLDS())
+            kind = "LDS";
+        }
+        fprintf(stderr, "%p, 0x%lx, %s\n", it->first, it->second.size, kind.c_str());
+        it++;
+      }
+      fprintf(stderr, "\n");
+      it = start;
+      runtime_singleton_->memory_lock_.Release();
+      hsa_amd_pointer_info_t info;
+      PtrInfoBlockData block;
+      uint32_t count;
+      hsa_agent_t* canAccess;
+      info.size = sizeof(info);
+      for (int i = 0; i < 3; i++) {
+        if (it == runtime_singleton_->allocation_map_.end()) break;
+        runtime_singleton_->PtrInfo(const_cast<void*>(it->first), &info, malloc, &count, &canAccess,
+                                    &block);
+        fprintf(stderr, "PtrInfo:\n\tAddress: %p/%p\n\tSize: 0x%lx\n\tType: %u\n\tOwner: %p\n",
+                info.agentBaseAddress, info.hostBaseAddress, info.sizeInBytes, info.type,
+                reinterpret_cast<void*>(info.agentOwner.handle));
+        fprintf(stderr, "\tCanAccess: %u\n", count);
+        for (int t = 0; t < count; t++)
+          fprintf(stderr, "\t\t%p\n", reinterpret_cast<void*>(canAccess[t].handle));
+        fprintf(stderr, "\tIn block: %p, 0x%lx\n", block.base, block.length);
+        free(canAccess);
+        it++;
+      }
+#endif  //! NDEBUG
+    }
+    assert(false && "GPU memory access fault.");
     std::abort();
   }
   // No need to keep the signal because we are done.