git/list[1] front-page[2] threads[3] people[4] search[5] about
 

[JGIT PATCH 06/13] Better handle concurrent reads during a WindowCache reconfiguration

From
Shawn O. Pearce <spearce@spearce.org>
Date
Apr 28, 2009, 21:12 UTC
Message-ID
<1240953146-12878-7-git-send-email-spearce@spearce.org>
In-Reply-To
<1240953146-12878-6-git-send-email-spearce@spearce.org>

If the WindowCache is reconfigured during a read we may wind up with the newly read window referenced by the old cache. This could cause skew on the PackFile's reference counts, potentially never allowing the pack's reference count to drop to 0, and pegging the file open.

One way to deal with this is to use a multi-reader/single-writer lock around the cache variable, and only permit reconfiguration when there are no concurrent accesses to the cache. This is also quite costly for an operation that occurs very infrequently, if ever.

Another approach is a double checked idiom. If the cache was replaced while we were accessing the window, there is a good chance that we put the window back into the old cache. If that occurs, calling removeAll again on the old cache will ensure the window isn't referenced anymore by that old cache.

Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
---
 .../src/org/spearce/jgit/lib/WindowCache.java      |   11 ++++++++++-
 1 files changed, 10 insertions(+), 1 deletions(-)
diff --git a/org.spearce.jgit/src/org/spearce/jgit/lib/WindowCache.java b/org.spearce.jgit/src/org/spearce/jgit/lib/WindowCache.java
index aaad033..91c7cf3 100644
--- a/org.spearce.jgit/src/org/spearce/jgit/lib/WindowCache.java
+++ b/org.spearce.jgit/src/org/spearce/jgit/lib/WindowCache.java
@@ -123,7 +123,16 @@ public static void reconfigure(final WindowCacheConfig cfg) {
 	static final ByteWindow get(final PackFile pack, final long offset)
 			throws IOException {
 		final WindowCache c = cache;
-		return c.getOrLoad(pack, c.toStart(offset));
+		final ByteWindow r = c.getOrLoad(pack, c.toStart(offset));
+		if (c != cache) {
+			// The cache was reconfigured while we were using the old one
+			// to load this window. The window is still valid, but our
+			// cache may think its still live. Ensure the window is removed
+			// from the old cache so resources can be released.
+			//
+			c.removeAll();
+		}
+		return r;
 	}
 
 	static final void purge(final PackFile pack) {
-- 
1.6.3.rc1.205.g37f8
Previous: Shawn O. PearceNext: Shawn O. Pearce
Message 7 of 18 in “Misc. bug fixes and cleanups”
  1. 00/13 Misc. bug fixes and cleanupsShawn O. Pearce, Apr 28, 2009
  2. 01/13 Fix performance problem recently introduced to DeltaPackedObjectLoaderShawn O. Pearce, Apr 28, 2009
  3. 02/13 Don't use ByteWindows when checking pack file headers/footersShawn O. Pearce, Apr 28, 2009
  4. 03/13 Rewrite WindowCache to be easier to follow and maintainShawn O. Pearce, Apr 28, 2009
  5. 04/13 Document the IllegalArgumentException thrown by WindowCache.reconfigureShawn O. Pearce, Apr 28, 2009
  6. 05/13 Create the new WindowCache before clearing the old oneShawn O. Pearce, Apr 28, 2009
  7. 06/13 Better handle concurrent reads during a WindowCache reconfigurationShawn O. Pearce, Apr 28, 2009
  8. 07/13 Clear dead OffsetCache cells when clearing a referenceShawn O. Pearce, Apr 28, 2009
  9. 08/13 Work around Sun JVM bug "Cleared SoftReference not added to queue"Shawn O. Pearce, Apr 28, 2009
  10. 09/13 Replace inefficient new String(String) constructor to silence FindBugsShawn O. Pearce, Apr 28, 2009
  11. 10/13 Replace inefficient new Long(long) constructor to silence FindBugsShawn O. Pearce, Apr 28, 2009
  12. 11/13 Change IndexPack to use ObjectIdSubclassMap instead of ObjectIdMapShawn O. Pearce, Apr 28, 2009
  13. 12/13 Use getCachedBytes in IndexPack to avoid an unnecessary copyShawn O. Pearce, Apr 28, 2009
  14. 13/13 Remove ObjectIdMap from the JGit libraryShawn O. Pearce, Apr 28, 2009
  15. Robin RosenbergApr 29, 2009
  16. 12/11 Add a test for LongMap, IndexPack's helper classShawn O. Pearce, Apr 29, 2009
  17. Robin RosenbergApr 29, 2009
  18. Shawn O. PearceApr 29, 2009

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.