[Libomptarget] Improve next-gen AMDGPU plugin error messages

[jhuber6]

Summary

The next-gen plugin properly prints errors. This patch improves the
error messages by including the Node-ID of the GPU that failed as well
as a textual representation of the enumeration values.

[kevinsala]

LGTM

openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp

2460

We could print the fatal message even if hsa_agent_get_info fails, right? We can print the GPU id or unknown/unspecified if the call failed.

[jhuber6]

openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp

2460

Yeah, we could probably just initialize it to -1 and continue like nothing happened.

[kevinsala]

openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp

2460

That's fine. We have to reinterpret the value as signed, or it will print the uint32_t's maximum value.

[arsenm]

openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp

2435

Is it really possible to have multiple of these hit at a time?

[jhuber6]

openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp

2435

Yes, but I think the only one I'm aware of is a typical segfault will present HSA_AMD_MEMORY_FAULT_PAGE_NOT_PRESENT and HSA_AMD_MEMORY_FAULT_READ_ONLY at the same time.

openmp/libomptarget/include/Debug.h

#define FATAL_MESSAGE0(_num, _str)                                             \
   do {                                                                         \
     fprintf(stderr, GETNAME(TARGET_NAME) " fatal error %d: %s\n", _num, _str); \
     abort();                                                                   \
   } while (0)
 
 /// Print fatal error message with a printf string and error identifier
 #define FATAL_MESSAGE(_num, _str, ...)                                         \
   do {                                                                         \
-    fprintf(stderr, GETNAME(TARGET_NAME) " fatal error %d:" _str "\n", _num,   \
+    fprintf(stderr, GETNAME(TARGET_NAME) " fatal error %d: " _str "\n", _num,  \
             __VA_ARGS__);                                                      \
     abort();                                                                   \
   } while (0)
 
 /// Print a generic error string from libomptarget or a plugin RTL
 #define FAILURE_MESSAGE(...)                                                   \
   do {                                                                         \
     fprintf(stderr, GETNAME(TARGET_NAME) " error: ");                          \

openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp

Error initImpl(GenericDeviceTy &Device, DeviceImageTy &Image) override {
     // TODO: Read the kernel descriptor for the max threads per block. May be
     // read from the image.
 
     return Plugin::success();
   }
 
   /// Launch the AMDGPU kernel function.
   Error launchImpl(GenericDeviceTy &GenericDevice, uint32_t NumThreads,
-                   uint64_t NumBlocks, 
-                   KernelArgsTy &KernelArgs, void *Args,
+                   uint64_t NumBlocks, KernelArgsTy &KernelArgs, void *Args,
                    AsyncInfoWrapperTy &AsyncInfoWrapper) const override;
 
   /// The default number of blocks is common to the whole device.
   uint32_t getDefaultNumBlocks(GenericDeviceTy &GenericDevice) const override {
     return GenericDevice.getDefaultNumBlocks();
   }
 
   /// The default number of threads is common to the whole device.
struct AMDGPUPluginTy final : public GenericPluginTy {
   }
 
 private:
   /// Event handler that will be called by ROCr if an event is detected.
   static hsa_status_t eventHandler(const hsa_amd_event_t *Event, void *) {
     if (Event->event_type != HSA_AMD_GPU_MEMORY_FAULT_EVENT)
       return HSA_STATUS_SUCCESS;
 
-    std::string Reasons;
+    SmallVector<std::string> Reasons;
     uint32_t ReasonsMask = Event->memory_fault.fault_reason_mask;
     if (ReasonsMask & HSA_AMD_MEMORY_FAULT_PAGE_NOT_PRESENT)
-      Reasons += "HSA_AMD_MEMORY_FAULT_PAGE_NOT_PRESENT, ";
+      Reasons.emplace_back("Page not present or supervisor privilege");
     if (ReasonsMask & HSA_AMD_MEMORY_FAULT_READ_ONLY)
-      Reasons += " HSA_AMD_MEMORY_FAULT_READ_ONLY, ";
+      Reasons.emplace_back("Write access to a read-only page");
     if (ReasonsMask & HSA_AMD_MEMORY_FAULT_NX)
-      Reasons += " HSA_AMD_MEMORY_FAULT_NX, ";
+      Reasons.emplace_back("Execute access to a page marked NX");
     if (ReasonsMask & HSA_AMD_MEMORY_FAULT_HOST_ONLY)
-      Reasons += " HSA_AMD_MEMORY_FAULT_HOST_ONLY, ";
+      Reasons.emplace_back("GPU attempted access to a host only page");
     if (ReasonsMask & HSA_AMD_MEMORY_FAULT_DRAMECC)
-      Reasons += " HSA_AMD_MEMORY_FAULT_DRAMECC, ";
+      Reasons.emplace_back("DRAM ECC failure");
     if (ReasonsMask & HSA_AMD_MEMORY_FAULT_IMPRECISE)
-      Reasons += " HSA_AMD_MEMORY_FAULT_IMPRECISE, ";
+      Reasons.emplace_back("Can't determine the exact fault address");
     if (ReasonsMask & HSA_AMD_MEMORY_FAULT_SRAMECC)
-      Reasons += " HSA_AMD_MEMORY_FAULT_SRAMECC, ";
+      Reasons.emplace_back("SRAM ECC failure (ie registers, no fault address)");
     if (ReasonsMask & HSA_AMD_MEMORY_FAULT_HANG)
-      Reasons += " HSA_AMD_MEMORY_FAULT_HANG, ";
+      Reasons.emplace_back("GPU reset following unspecified hang");
 
     // If we do not know the reason, say so, otherwise remove the trailing comma
     // and space.
     if (Reasons.empty())
-      Reasons = "Unknown (Mask: " + std::to_string(ReasonsMask) + ")";
-    else
-      Reasons.resize(Reasons.size() - /* ', ' */ 2);
+      Reasons.emplace_back("Unknown (" + std::to_string(ReasonsMask) + ")");
+
+    uint32_t Node = -1;
+    hsa_agent_get_info(Event->memory_fault.agent, HSA_AGENT_INFO_NODE, &Node);
 
     // Abort the execution since we do not recover from this error.
     FATAL_MESSAGE(1,
-                  "Found HSA_AMD_GPU_MEMORY_FAULT_EVENT in agent %" PRIu64
-                  " at virtual address %p and reasons: %s",
-                  Event->memory_fault.agent.handle,
-                  (void *)Event->memory_fault.virtual_address, Reasons.data());
+                  "Memory access fault by GPU %" PRIu32 " (agent 0x%" PRIx64
+                  ") at virtual address %p. Reasons: %s",
+                  Node, Event->memory_fault.agent.handle,
+                  (void *)Event->memory_fault.virtual_address,
+                  llvm::join(Reasons, ", ").c_str());
 
     return HSA_STATUS_ERROR;
   }
 
   /// Indicate whether the HSA runtime was correctly initialized. Even if there
   /// is no available devices this boolean will be true. It indicates whether
   /// we can safely call HSA functions (e.g., hsa_shut_down).
   bool Initialized;