Fix Arena MR to support simultaneous access by PTDS and other streams (…

…#1395)

Replaces #1394, this is targeted for 24.02.

fixes #1393

In Spark with the Spark Rapids accelerator using cudf 23.12 snapshot we have an application that is reading ORC files, doing some light processing and then writing ORC files. It consistently fails while doing the ORC write with:

```
terminate called after throwing an instance of 'rmm::logic_error'
  what():  RMM failure at:/home/jenkins/agent/workspace/jenkins-spark-rapids-jni_nightly-dev-594-cuda11/thirdparty/cudf/cpp/build/_deps/rmm-src/include/rmm/mr/device/arena_memory_resource.hpp:238: allocation not found
```

The underlying issue is brought about because Spark with the Rapids accelerate is using ARENA allocator with per default streams enabled.  CUDF recently added its own stream pool that is used in addition to when per default streams are used.  
It's now possible to use per thread default streams along with another pool of streams. This means that it's possible for an arena to move from a thread or stream arena back  into the global arena during a defragmentation and then move down into another arena type. For instance, thread arena -> global arena -> stream arena. If this happens and  there was an allocation from it while it was a thread arena, we now have to check to see if the allocation is part of a stream arena.

I added a test here. I was trying to make sure that all the allocations were now in stream arenas, if there is a better way to do this please let me know.

Authors:
  - Thomas Graves (https://github.com/tgravescs)

Approvers:
  - Lawrence Mitchell (https://github.com/wence-)
  - Bradley Dice (https://github.com/bdice)
  - Rong Ou (https://github.com/rongou)
  - Mark Harris (https://github.com/harrism)

URL: #1395

include/rmm/mr/device/arena_memory_resource.hpp

@@ -235,7 +235,26 @@ class arena_memory_resource final : public device_memory_resource {
       }
     }
 
-    if (!global_arena_.deallocate(ptr, bytes)) { RMM_FAIL("allocation not found"); }
+    if (!global_arena_.deallocate(ptr, bytes)) {
+      // It's possible to use per thread default streams along with another pool of streams.
+      // This means that it's possible for an allocation to move from a thread or stream arena
+      // back into the global arena during a defragmentation and then move down into another arena
+      // type. For instance, thread arena -> global arena -> stream arena. If this happens and
+      // there was an allocation from it while it was a thread arena, we now have to check to
+      // see if the allocation is part of a stream arena, and vice versa.
+      // Only do this in exceptional cases to not affect performance and have to check all
+      // arenas all the time.
+      if (use_per_thread_arena(stream)) {
+        for (auto& stream_arena : stream_arenas_) {
+          if (stream_arena.second.deallocate(ptr, bytes)) { return; }
+        }
+      } else {
+        for (auto const& thread_arena : thread_arenas_) {
+          if (thread_arena.second->deallocate(ptr, bytes)) { return; }
+        }
+      }
+      RMM_FAIL("allocation not found");
+    }
   }
 
   /**

tests/mr/device/arena_mr_tests.cpp

@@ -533,6 +533,37 @@ TEST_F(ArenaTest, Defragment)  // NOLINT
   }());
 }
 
+TEST_F(ArenaTest, PerThreadToStreamDealloc)  // NOLINT
+{
+  // This is testing that deallocation of a ptr still works when
+  // it was originally allocated in a superblock that was in a thread
+  // arena that then moved to global arena during a defragmentation
+  // and then moved to a stream arena.
+  auto const arena_size = superblock::minimum_size * 2;
+  arena_mr mr(rmm::mr::get_current_device_resource(), arena_size);
+  // Create an allocation from a per thread arena
+  void* thread_ptr = mr.allocate(256, rmm::cuda_stream_per_thread);
+  // Create an allocation in a stream arena to force global arena
+  // to be empty
+  cuda_stream stream{};
+  void* ptr = mr.allocate(32_KiB, stream);
+  mr.deallocate(ptr, 32_KiB, stream);
+  // at this point the global arena doesn't have any superblocks so
+  // the next allocation causes defrag. Defrag causes all superblocks
+  // from the thread and stream arena allocated above to go back to
+  // global arena and it allocates one superblock to the stream arena.
+  auto* ptr1 = mr.allocate(superblock::minimum_size, rmm::cuda_stream_view{});
+  // Allocate again to make sure all superblocks from
+  // global arena are owned by a stream arena instead of a thread arena
+  // or the global arena.
+  auto* ptr2 = mr.allocate(32_KiB, rmm::cuda_stream_view{});
+  // The original thread ptr is now owned by a stream arena so make
+  // sure deallocation works.
+  mr.deallocate(thread_ptr, 256, rmm::cuda_stream_per_thread);
+  mr.deallocate(ptr1, superblock::minimum_size, rmm::cuda_stream_view{});
+  mr.deallocate(ptr2, 32_KiB, rmm::cuda_stream_view{});
+}
+
 TEST_F(ArenaTest, DumpLogOnFailure)  // NOLINT
 {
   arena_mr mr{rmm::mr::get_current_device_resource(), 1_MiB, true};