diff --git a/src/common/span.h b/src/common/span.h
index d66bcbe85c96..ec5a38bcf80c 100644
--- a/src/common/span.h
+++ b/src/common/span.h
@@ -341,6 +341,29 @@ XGBOOST_DEVICE bool LexicographicalCompare(InputIt1 first1, InputIt1 last1,
  *         device_vector::data() returns a wrapped pointer.
  *       It's unclear that what kind of thrust algorithm can be used without
  *         memory error. See the test case "GPUSpan.WithTrust"
+ *
+ *    Pass iterator to kernel:
+ *       Not possible. Use subspan instead.
+ *
+ *       The underlying Span in SpanIterator is a pointer, but CUDA pass kernel
+ *       parameter by value.  If we were to hold a Span value instead of a
+ *       pointer, the following snippet will crash, violating the safety
+ *       purpose of Span:
+ *
+ *       \code{.cpp}
+ *       Span<float> span {arr_a};
+ *       auto beg = span.begin();
+ *
+ *       Span<float> span_b = arr_b;
+ *       span = span_b;
+ *
+ *       delete arr_a;
+ *       beg++;                 // crash
+ *       \endcode
+ *
+ *       While hoding a pointer or reference should avoid the problem, its a
+ *       compromise. Since we have subspan, it's acceptable not to support
+ *       passing iterator.
  */
 template <typename T,
           detail::ptrdiff_t Extent = dynamic_extent>