diff --git a/daemon/proto.c b/daemon/proto.c
index 88e2d28c7..238ee88c6 100644
--- a/daemon/proto.c
+++ b/daemon/proto.c
@@ -101,6 +101,12 @@ main_loop (int _sock)
 	       "guestfsd: main_loop: new request, len 0x%" PRIx32 "\n",
 	       len);
 
+    /* Cancellation sent from the library and received after the
+     * previous request has finished processing.  Just ignore it.
+     */
+    if (len == GUESTFS_CANCEL_FLAG)
+      continue;
+
     if (len > GUESTFS_MESSAGE_MAX) {
       fprintf (stderr, "guestfsd: incoming message is too long (%u bytes)\n",
                len);
diff --git a/regressions/test-cancellation-download-librarycancels.sh b/regressions/test-cancellation-download-librarycancels.sh
index 20f89d233..5f6ebf1ed 100755
--- a/regressions/test-cancellation-download-librarycancels.sh
+++ b/regressions/test-cancellation-download-librarycancels.sh
@@ -18,22 +18,36 @@
 
 # Test download where the library cancels.
 #
-#
+# Download big and small files to /dev/full.  This should fail but not
+# kill the appliance.  We test various randomized file sizes because
+# there are many potential race conditions -- for example the daemon
+# may or may not send all of its data because the error condition is
+# detected.
 
 set -e
 
 rm -f test.img
 
-../fish/guestfish <<'EOF'
-add ../images/test.iso
+size=$(awk 'BEGIN{ srand(); print int(16*1024*rand()) }')
+echo "$0: test size $size (bytes)"
+
+../fish/guestfish <<EOF
+# We want the file to be fully allocated.
+alloc test.img 10M
 run
 
-mount-ro /dev/sda /
+part-disk /dev/sda mbr
+mkfs ext2 /dev/sda1
+mount-options "" /dev/sda1 /
 
-# Download a file to /dev/full.
--download /100krandom /dev/full
+fallocate64 /file $size
 
+# Download the file into /dev/full so it fails.
+-download /file /dev/full
+
+# The daemon should still be reachable after the failure.
 ping-daemon
+
 EOF
 
 rm -f test.img