[SYCL][L0] Fix memory leak when tracking indirect accesses (#7105)

On the plugin API boundaries extra piKernelRetain/piKernelRelease can be
called in arbitrary moments (from other threads) include *before* the
kernel was submitted to the actual GPU HW. As such, piKernelRelease has
nothing to do with submissions and indirect memory tracking updates
should be initiated from the points where the kernel actually finishes -
CleanupCompletedEvent.

Author: aelovikov-intel

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -5253,27 +5253,6 @@ pi_result piKernelRetain(pi_kernel Kernel) {
 pi_result piKernelRelease(pi_kernel Kernel) {
   PI_ASSERT(Kernel, PI_ERROR_INVALID_KERNEL);
 
-  if (IndirectAccessTrackingEnabled) {
-    // piKernelRelease is called by CleanupCompletedEvent(Event) as soon as
-    // kernel execution has finished. This is the place where we need to release
-    // memory allocations. If kernel is not in use (not submitted by some
-    // other thread) then release referenced memory allocations. As a result,
-    // memory can be deallocated and context can be removed from container in
-    // the platform. That's why we need to lock a mutex here.
-    pi_platform Plt = Kernel->Program->Context->getPlatform();
-    std::scoped_lock<pi_shared_mutex> ContextsLock(Plt->ContextsMutex);
-
-    if (--Kernel->SubmissionsCount == 0) {
-      // Kernel is not submitted for execution, release referenced memory
-      // allocations.
-      for (auto &MemAlloc : Kernel->MemAllocs) {
-        USMFreeHelper(MemAlloc->second.Context, MemAlloc->first,
-                      MemAlloc->second.OwnZeMemHandle);
-      }
-      Kernel->MemAllocs.clear();
-    }
-  }
-
   if (!Kernel->RefCount.decrementAndTest())
     return PI_SUCCESS;
 
@@ -5821,9 +5800,35 @@ static pi_result CleanupCompletedEvent(pi_event Event, bool QueueLocked) {
     Event->CleanedUp = true;
   }
 
+  auto ReleaseIndirectMem = [](pi_kernel Kernel) {
+    if (IndirectAccessTrackingEnabled) {
+      // piKernelRelease is called by CleanupCompletedEvent(Event) as soon as
+      // kernel execution has finished. This is the place where we need to
+      // release memory allocations. If kernel is not in use (not submitted by
+      // some other thread) then release referenced memory allocations. As a
+      // result, memory can be deallocated and context can be removed from
+      // container in the platform. That's why we need to lock a mutex here.
+      pi_platform Plt = Kernel->Program->Context->getPlatform();
+      std::scoped_lock<pi_shared_mutex> ContextsLock(Plt->ContextsMutex);
+
+      if (--Kernel->SubmissionsCount == 0) {
+        // Kernel is not submitted for execution, release referenced memory
+        // allocations.
+        for (auto &MemAlloc : Kernel->MemAllocs) {
+          // std::pair<void *const, MemAllocRecord> *, Hash
+          USMFreeHelper(MemAlloc->second.Context, MemAlloc->first,
+                        MemAlloc->second.OwnZeMemHandle);
+        }
+        Kernel->MemAllocs.clear();
+      }
+    }
+  };
+
   // We've reset event data members above, now cleanup resources.
-  if (AssociatedKernel)
+  if (AssociatedKernel) {
+    ReleaseIndirectMem(AssociatedKernel);
     PI_CALL(piKernelRelease(AssociatedKernel));
+  }
 
   if (AssociatedQueue) {
     {
@@ -5877,8 +5882,10 @@ static pi_result CleanupCompletedEvent(pi_event Event, bool QueueLocked) {
         }
       }
     }
-    if (DepEventKernel)
-      PI_CALL(piKernelRelease(pi_cast<pi_kernel>(DepEvent->CommandData)));
+    if (DepEventKernel) {
+      ReleaseIndirectMem(DepEventKernel);
+      PI_CALL(piKernelRelease(DepEventKernel));
+    }
     PI_CALL(piEventReleaseInternal(DepEvent));
   }