Also inject a failure for createWritableDirectory when testing that A…

…ctionOutputDirectoryHelper propagates exceptions.

Failure to do causes a flaky test failure when simulating an IOException while the `knownDirectories` cache is disabled. The reason is that `ActionOutputDirectoryHelper#forceCreateDirectoryAndParents` calls either `createDirectory` or `createWritableDirectory`, depending on whether the parent directory is in the cache or not. However, "disabling" the cache sets its size to zero, which doesn't prevent insertion; rather, it causes inserted entries to be almost immediately deleted by a background thread. Thus, if only one of `createDirectory` and `createWritableDirectory` throws, the test outcome depends on how fast the background thread runs.

(We could make the zero-sized cache an actual no-op, but it's unlikely that anyone would want to set it to zero outside of test code, so it's not worth the trouble.)

Fixes #21471.

PiperOrigin-RevId: 615513608
Change-Id: Id2247596c6af0e5d5142072de8309227a1d1cbd1

src/test/java/com/google/devtools/build/lib/actions/ActionOutputDirectoryHelperTest.java

@@ -264,6 +264,14 @@ public boolean createDirectory(PathFragment path) throws IOException {
         return super.createDirectory(path);
       }
 
+      @Override
+      public boolean createWritableDirectory(PathFragment path) throws IOException {
+        if (path.equals(failingPath)) {
+          throw injectedException;
+        }
+        return super.createWritableDirectory(path);
+      }
+
       @Override
       public void createDirectoryAndParents(PathFragment path) throws IOException {
         if (path.equals(failingPath)) {