[PATCH]swap-zero-page-bug

[PATCH]swap-zero-page-bug

[Li Xiaodong]

[-- Attachment #1: Type: text/plain, Size: 916 bytes --]

According to Documentation/filesystems/proc.txt, kernel should swap one page
as vm.page-cluster is 0, while it seems not like so in mm/swapfile.c:
     ...
	int our_page_cluster = page_cluster;
	...
	if (!our_page_cluster)  /* no readahead */
                return 0;
	...

It is better to remove that checkpoint.

Signed-off-by: Li Xiaodong<lixd@cn.fujitsu.com>
--- swapfile.c.orig	2008-10-29 21:29:36.000000000 +0800
+++ swapfile.c	2008-10-29 21:59:49.000000000 +0800
@@ -1810,9 +1810,6 @@
 	pgoff_t base, end;
 	int nr_pages = 0;
 
-	if (!our_page_cluster)	/* no readahead */
-		return 0;
-
 	si = &swap_info[swp_type(entry)];
 	target = swp_offset(entry);
 	base = (target >> our_page_cluster) << our_page_cluster; @@ -1847,5
+1844,5 @@
 	 * if only 1, say 0, since there's then no readahead to be done.
 	 */
 	*offset = ++toff;
-	return nr_pages? ++nr_pages: 0;
+	return nr_pages ? ++nr_pages : nr_pages;
 }


[-- Attachment #2: [PATCH]swap-zero-page-bug.patch --]
[-- Type: application/octet-stream, Size: 619 bytes --]

Signed-off-by: Li Xiaodong<lixd@cn.fujitsu.com>
--- swapfile.c.orig	2008-10-29 21:29:36.000000000 +0800
+++ swapfile.c	2008-10-29 21:59:49.000000000 +0800
@@ -1810,9 +1810,6 @@
 	pgoff_t base, end;
 	int nr_pages = 0;
 
-	if (!our_page_cluster)	/* no readahead */
-		return 0;
-
 	si = &swap_info[swp_type(entry)];
 	target = swp_offset(entry);
 	base = (target >> our_page_cluster) << our_page_cluster;
@@ -1847,5 +1844,5 @@
 	 * if only 1, say 0, since there's then no readahead to be done.
 	 */
 	*offset = ++toff;
-	return nr_pages? ++nr_pages: 0;
+	return nr_pages ? ++nr_pages : nr_pages;
 }

Re: [PATCH]swap-zero-page-bug

[Hugh Dickins]

On Wed, 29 Oct 2008, Li Xiaodong wrote:

> According to Documentation/filesystems/proc.txt, kernel should swap one page
> as vm.page-cluster is 0, while it seems not like so in mm/swapfile.c:
>      ...
> 	int our_page_cluster = page_cluster;
> 	...
> 	if (!our_page_cluster)  /* no readahead */
>                 return 0;
> 	...
> 
> It is better to remove that checkpoint.

No, page_cluster 0 allows swapin one page at a time, but disallows
swapin readahead (better named readaround) i.e. disallows reading
more than the one page actually requested.

Compare with how valid_swaphandles() is used from swapin_readahead():
there's actually one extra call to read_swap_cache_async() - because
we must ignore failure on the readaround but not on the page requested.

That "return 0" avoids the overhead of a pointlessly doubled
read_swap_cache_async() when there's only one page to be read.

Your patch might be needed if you'd found that swapin of one page
was not working at all - but you'd be making a stronger case if
you'd found that, I think!

Hugh