diff --git a/src/array.jl b/src/array.jl
index c0a183b6..accb2dac 100644
--- a/src/array.jl
+++ b/src/array.jl
@@ -370,14 +370,19 @@ end
 
 ## indexing
 
-# Host-accessible arrays can be indexed from CPU, bypassing GPUArrays restrictions
+# Host-accessible arrays can be indexed from CPU, bypassing GPUArrays restrictions.
+# Wait for queued work first, e.g., the kernel computing the result of a reduction. This
+# only synchronizes the current task's stream; work submitted by other tasks, or to an
+# explicitly created queue, needs to be synchronized explicitly.
 function Base.getindex(x::oneArray{<:Any, <:Any, <:Union{oneL0.HostBuffer, oneL0.SharedBuffer}}, I::Int)
     @boundscheck checkbounds(x, I)
+    synchronize(global_stream(context(x), device()))
     return unsafe_load(pointer(x, I; type = oneL0.HostBuffer))
 end
 
 function Base.setindex!(x::oneArray{<:Any, <:Any, <:Union{oneL0.HostBuffer, oneL0.SharedBuffer}}, v, I::Int)
     @boundscheck checkbounds(x, I)
+    synchronize(global_stream(context(x), device()))
     return unsafe_store!(pointer(x, I; type = oneL0.HostBuffer), v)
 end
 
diff --git a/test/array.jl b/test/array.jl
index 5d92d759..eb21d263 100644
--- a/test/array.jl
+++ b/test/array.jl
@@ -103,6 +103,14 @@ end
   @test b == [100, 200]
 end
 
+@testset "reductions of host-accessible arrays" begin
+  for B in (oneL0.SharedBuffer, oneL0.HostBuffer)
+    a = oneArray{Float32, 1, B}(fill(1.0f0, 1024))
+    @test sum(a) == 1024
+    @test maximum(a) == 1
+  end
+end
+
 # https://github.com/JuliaGPU/CUDA.jl/issues/2191
 @testset "preserving buffer types" begin
   a = oneVector{Int,oneL0.SharedBuffer}([1])