[LUCENE-8663] NRTCachingDirectory.slowFileExists may open a file while it's inaccessible

• Type: Bug
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Fix Version/s: 9.0
• Lucene Fields: New
• Assignee: Dawid Weiss
• Reporter: Dawid Weiss
• Created: 29/Jan/19 09:40
• Updated: 28/Aug/22 15:41
• Resolved: 30/Jan/19 08:44

Description

This build:
https://builds.apache.org/job/Lucene-Solr-NightlyTests-master/1761/consoleText

failed with the following stack trace:

FAILED:  org.apache.lucene.util.TestOfflineSorter.testThreadSafety

Error Message:
Captured an uncaught exception in thread: Thread[id=2665, name=Thread-2493, state=RUNNABLE, group=TGRP-TestOfflineSorter]

Stack Trace:
com.carrotsearch.randomizedtesting.UncaughtExceptionError: Captured an uncaught exception in thread: Thread[id=2665, name=Thread-2493, state=RUNNABLE, group=TGRP-TestOfflineSorter]
Caused by: java.lang.RuntimeException: java.nio.file.AccessDeniedException: Can't open a file still open for writing: unsorted_tmp_0.tmp
        at __randomizedtesting.SeedInfo.seed([972EEC94E272C842]:0)
        at org.apache.lucene.util.TestOfflineSorter$2.run(TestOfflineSorter.java:271)
Caused by: java.nio.file.AccessDeniedException: Can't open a file still open for writing: unsorted_tmp_0.tmp
        at org.apache.lucene.store.ByteBuffersDirectory$FileEntry.openInput(ByteBuffersDirectory.java:250)
        at org.apache.lucene.store.ByteBuffersDirectory.openInput(ByteBuffersDirectory.java:222)
        at org.apache.lucene.store.NRTCachingDirectory.slowFileExists(NRTCachingDirectory.java:292)
        at org.apache.lucene.store.NRTCachingDirectory.createTempOutput(NRTCachingDirectory.java:266)
        at org.apache.lucene.store.MockDirectoryWrapper.createTempOutput(MockDirectoryWrapper.java:697)
        at org.apache.lucene.util.TestOfflineSorter.checkSort(TestOfflineSorter.java:183)
        at org.apache.lucene.util.TestOfflineSorter.access$100(TestOfflineSorter.java:49)
        at org.apache.lucene.util.TestOfflineSorter$2.run(TestOfflineSorter.java:267)

The problem in slowFileExists is that it probes for file existence by trying to open it in a directory. We don't allow opening files that are still open by another thread (for writing), hence the exception thrown is AccessDeniedException.

I couldn't reproduce the problem with the same seed. Takes some luck to hit the right scenario in a multi-threaded run.

The contract for Directory.fileLength is less strict so a straightforward patch is to use that instead.