SWDEV-296301 - Avoid deadlocks in the hostcall path

Change the scope of hostcall buffer access lock during destruction.
Make sure wait() returns the signal value after timeout. That
matches ROCr behaviour for HSA signal wait.

Change-Id: I3df34207e0c2e21972ec8052777e5742bda1dca0


[ROCm/clr commit: 9a9d10a10b]

projects/clr/rocclr/device/devhostcall.cpp

@@ -286,10 +286,12 @@ void HostcallListener::consumePackets() {
       return;
     }
 
-    amd::ScopedLock lock{listenerLock};
+    if (!idle()) {
+      amd::ScopedLock lock{listenerLock};
 
-    for (auto ii : buffers_) {
-      ii->processPackets(messages_);
+      for (auto ii : buffers_) {
+        ii->processPackets(messages_);
+      }
     }
   }
 
@@ -389,14 +391,15 @@ bool enableHostcalls(const amd::Device &dev, void* bfr, uint32_t numPackets) {
 }
 
 void disableHostcalls(void* bfr) {
-  amd::ScopedLock lock(listenerLock);
-  if (!hostcallListener) {
-    return;
+  {
+    amd::ScopedLock lock(listenerLock);
+    if (!hostcallListener) {
+      return;
+    }
+    assert(bfr && "expected a hostcall buffer");
+    auto buffer = reinterpret_cast<HostcallBuffer*>(bfr);
+    hostcallListener->removeBuffer(buffer);
   }
-  assert(bfr && "expected a hostcall buffer");
-  auto buffer = reinterpret_cast<HostcallBuffer*>(bfr);
-  hostcallListener->removeBuffer(buffer);
-
   if (hostcallListener->idle()) {
     hostcallListener->terminate();
     delete hostcallListener;

projects/clr/rocclr/device/pal/palsignal.cpp

@@ -129,7 +129,7 @@ uint64_t Signal::Wait(uint64_t value, device::Signal::Condition c, uint64_t time
     float timeoutInSec = timeout / (1000 * 1000);
     result = event_.Wait(timeoutInSec);
 
-    if (result != Pal::Result::Success) {
+    if ((result != Pal::Result::Success) && (result != Pal::Result::Timeout)) {
       return -1;
     }
 
@@ -142,7 +142,7 @@ uint64_t Signal::Wait(uint64_t value, device::Signal::Condition c, uint64_t time
       auto end = amd::Os::timeNanos();
       auto duration = 1000 * (end - start); // convert to us
       if (duration >= timeout) {
-        return -1;
+        return amdSignal_->value;
       }
 
       if (!cmp(amdSignal_->value, value)) {