[Patch 3/7] 2.4.20-pre4/ext3: evict truncated buffers more easily.

* [Patch 3/7] 2.4.20-pre4/ext3: evict truncated buffers more easily.
@ 2002-09-26 13:55 Stephen Tweedie
  2002-09-30 19:20 ` Marcelo Tosatti
  0 siblings, 1 reply; 2+ messages in thread
From: Stephen Tweedie @ 2002-09-26 13:55 UTC (permalink / raw)
  To: Marcelo Tosatti, linux-kernel; +Cc: Stephen Tweedie

Performance tweak: when we truncate a file but encounter pages which are still
being journaled from a previous transaction, we can't evict the page from
memory immediately.

This patch just makes it a little easier for the VM to evict the page later on:
the page is marked unreferenced while we're committing the transaction which
pins it, and the commit logic tries to free the page completely once the
transaction has committed.

--- linux-2.4-ext3merge/fs/jbd/commit.c.=K0002=.orig	Thu Sep 26 12:19:14 2002
+++ linux-2.4-ext3merge/fs/jbd/commit.c	Thu Sep 26 12:25:37 2002
@@ -683,13 +683,25 @@
 			JBUFFER_TRACE(jh, "refile for checkpoint writeback");
 			__journal_refile_buffer(jh);
 		} else {
+			struct page *page = bh->b_page;
+			
 			J_ASSERT_BH(bh, !buffer_dirty(bh));
 			J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 			__journal_unfile_buffer(jh);
 			jh->b_transaction = 0;
 			__journal_remove_journal_head(bh);
-			__brelse(bh);
+
+			if (TryLockPage(page)) {	
+				__brelse(bh);
+			} else {
+				__brelse(bh);
+				page_cache_get(page);
+				try_to_free_buffers(page, 0);
+				unlock_page(page);
+				page_cache_release(page);
+			}
 		}
+		
 		spin_unlock(&journal_datalist_lock);
 	}
 
--- linux-2.4-ext3merge/fs/jbd/transaction.c.=K0002=.orig	Thu Sep 26 12:19:14 2002
+++ linux-2.4-ext3merge/fs/jbd/transaction.c	Thu Sep 26 12:25:37 2002
@@ -1903,8 +1903,17 @@
 	unlock_journal(journal);
 
 	if (!offset) {
-		if (!may_free || !try_to_free_buffers(page, 0))
+		if (!may_free || !try_to_free_buffers(page, 0)) {
+			if (!offset) {
+				/* We are still using the page, but only
+                                   because a transaction is pinning the
+                                   page.  Once it commits, we want to
+                                   encourage the page to be reaped as
+                                   quickly as possible. */
+				ClearPageReferenced(page);
+			}
 			return 0;
+		}
 		J_ASSERT(page->buffers == NULL);
 	}
 	return 1;

^ permalink raw reply	[flat|nested] 2+ messages in thread