Add BlockHostUntilDoneWithStatus, which returns Status rather than bool.

Also fixed a deadlock in Stream::BlockHostUntilDone.  The problem with the
original code was that it grabbed mu_ before looping over substreams, and would
call CheckError with mu_ still held.  But CheckError will attempt to lock mu_ in
the failure case, which would deadlock.

PiperOrigin-RevId: 178191634

1 files changed, 1 insertions, 0 deletions

diff --git a/tensorflow/stream_executor/stream_executor_internal.h b/tensorflow/stream_executor/stream_executor_internal.h
index 14445a7657..d2426f46e2 100644
--- a/tensorflow/stream_executor/stream_executor_internal.h
+++ b/tensorflow/stream_executor/stream_executor_internal.h
@@ -220,6 +220,7 @@ class StreamExecutorInterface {
   virtual bool StartTimer(Stream *stream, Timer *timer) = 0;
   virtual bool StopTimer(Stream *stream, Timer *timer) = 0;
   virtual bool BlockHostUntilDone(Stream *stream) = 0;
+  virtual port::Status BlockHostUntilDoneWithStatus(Stream *stream);
   virtual int PlatformDeviceCount() = 0;
   virtual port::Status EnablePeerAccessTo(StreamExecutorInterface *other) = 0;
   virtual bool CanEnablePeerAccessTo(StreamExecutorInterface *other) = 0;