kernel26-magellan/patches-2.6.20-r1/0022-2.6.20-mm-filesize_dependant_lru_cache_add.patch

When reading from large files through the generic file read functions into
page cache we can detect when a file is so large that it is unlikely to be
fully cached in ram.

Add a tunable /proc/sys/vm/tail_largefiles that puts them at the tail of the
inactive list to minimise their harm on present mapped pages and pagecache
and enable it by default.

Signed-off-by: Con Kolivas <kernel@kolivas.org>

---
 Documentation/filesystems/proc.txt |    8 +++++
 Documentation/sysctl/vm.txt        |    2 -
 include/linux/sysctl.h             |    1 
 kernel/sysctl.c                    |    9 ++++++
 mm/filemap.c                       |   53 +++++++++++++++++++++++++++++++++++--
 mm/swap.c                          |    3 --
 6 files changed, 71 insertions(+), 5 deletions(-)

Index: linux-2.6.20-ck1/mm/filemap.c
===================================================================
--- linux-2.6.20-ck1.orig/mm/filemap.c	2007-02-05 22:52:04.000000000 +1100
+++ linux-2.6.20-ck1/mm/filemap.c	2007-02-16 19:01:34.000000000 +1100
@@ -466,6 +466,16 @@ int add_to_page_cache_lru(struct page *p
 	return ret;
 }
 
+int add_to_page_cache_lru_tail(struct page *page,
+	struct address_space *mapping, pgoff_t offset, gfp_t gfp_mask)
+{
+	int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
+
+	if (ret == 0)
+		lru_cache_add_tail(page);
+	return ret;
+}
+
 #ifdef CONFIG_NUMA
 struct page *__page_cache_alloc(gfp_t gfp)
 {
@@ -856,6 +866,34 @@ static void shrink_readahead_size_eio(st
 	ra->ra_pages /= 4;
 }
 
+/*
+ * Sysctl which determines whether we should read from large files to the
+ * tail of the inactive lru list.
+ */
+int vm_tail_largefiles __read_mostly = 1;
+
+static inline int nr_mapped(void)
+{
+	return global_page_state(NR_FILE_MAPPED) +
+		global_page_state(NR_ANON_PAGES);
+}
+
+/*
+ * This examines how large in pages a file size is and returns 1 if it is
+ * more than half the unmapped ram. Avoid doing read_page_state which is
+ * expensive unless we already know it is likely to be large enough.
+ */
+static int large_isize(unsigned long nr_pages)
+{
+	if (nr_pages * 6 > vm_total_pages) {
+		 unsigned long unmapped_ram = vm_total_pages - nr_mapped();
+
+		if (nr_pages * 2 > unmapped_ram)
+			return 1;
+	}
+	return 0;
+}
+
 /**
  * do_generic_mapping_read - generic file read routine
  * @mapping:	address_space to be read
@@ -1064,8 +1102,19 @@ no_cached_page:
 				goto out;
 			}
 		}
-		error = add_to_page_cache_lru(cached_page, mapping,
-						index, GFP_KERNEL);
+
+		/*
+		 * If we know the file is large we add the pages read to the
+		 * end of the lru as we're unlikely to be able to cache the
+		 * whole file in ram so make those pages the first to be
+		 * dropped if not referenced soon.
+		 */
+		if (vm_tail_largefiles && large_isize(end_index))
+			error = add_to_page_cache_lru_tail(cached_page,
+						mapping, index, GFP_KERNEL);
+		else
+			error = add_to_page_cache_lru(cached_page, mapping,
+							index, GFP_KERNEL);
 		if (error) {
 			if (error == -EEXIST)
 				goto find_page;
Index: linux-2.6.20-ck1/mm/swap.c
===================================================================
--- linux-2.6.20-ck1.orig/mm/swap.c	2007-02-16 19:01:33.000000000 +1100
+++ linux-2.6.20-ck1/mm/swap.c	2007-02-16 19:01:34.000000000 +1100
@@ -434,8 +434,7 @@ void __pagevec_lru_add_active(struct pag
 
 /*
  * Function used uniquely to put pages back to the lru at the end of the
- * inactive list to preserve the lru order. Currently only used by swap
- * prefetch.
+ * inactive list to preserve the lru order.
  */
 void fastcall lru_cache_add_tail(struct page *page)
 {
Index: linux-2.6.20-ck1/kernel/sysctl.c
===================================================================
--- linux-2.6.20-ck1.orig/kernel/sysctl.c	2007-02-16 19:01:33.000000000 +1100
+++ linux-2.6.20-ck1/kernel/sysctl.c	2007-02-16 19:01:34.000000000 +1100
@@ -77,6 +77,7 @@ extern int pid_max_min, pid_max_max;
 extern int sysctl_drop_caches;
 extern int percpu_pagelist_fraction;
 extern int compat_log;
+extern int vm_tail_largefiles;
 
 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 static int maxolduid = 65535;
@@ -918,6 +919,14 @@ static ctl_table vm_table[] = {
 		.mode		= 0644,
 		.proc_handler	= &proc_dointvec,
 	},
+	{
+		.ctl_name	= VM_TAIL_LARGEFILES,
+		.procname	= "tail_largefiles",
+		.data		= &vm_tail_largefiles,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+	},
 #ifdef CONFIG_HUGETLB_PAGE
 	 {
 		.ctl_name	= VM_HUGETLB_PAGES,
Index: linux-2.6.20-ck1/Documentation/filesystems/proc.txt
===================================================================
--- linux-2.6.20-ck1.orig/Documentation/filesystems/proc.txt	2007-02-05 22:51:59.000000000 +1100
+++ linux-2.6.20-ck1/Documentation/filesystems/proc.txt	2007-02-16 19:01:34.000000000 +1100
@@ -1324,6 +1324,14 @@ To free pagecache, dentries and inodes:
 As this is a non-destructive operation and dirty objects are not freeable, the
 user should run `sync' first.
 
+tail_largefiles
+---------------
+
+When enabled reads from large files to the tail end of the inactive lru list.
+This means that any cache from reading large files is dropped very quickly,
+preventing loss of mapped ram and useful pagecache when large files are read.
+This does, however, make caching less effective when working with large files.
+
 
 2.5 /proc/sys/dev - Device specific parameters
 ----------------------------------------------
Index: linux-2.6.20-ck1/Documentation/sysctl/vm.txt
===================================================================
--- linux-2.6.20-ck1.orig/Documentation/sysctl/vm.txt	2007-02-16 19:01:33.000000000 +1100
+++ linux-2.6.20-ck1/Documentation/sysctl/vm.txt	2007-02-16 19:01:34.000000000 +1100
@@ -39,7 +39,7 @@ Currently, these files are in /proc/sys/
 
 dirty_ratio, dirty_background_ratio, dirty_expire_centisecs,
 dirty_writeback_centisecs, vfs_cache_pressure, laptop_mode,
-block_dump, swap_token_timeout, drop-caches:
+block_dump, swap_token_timeout, drop-caches, tail_largefiles:
 
 See Documentation/filesystems/proc.txt
 
Index: linux-2.6.20-ck1/include/linux/sysctl.h
===================================================================
--- linux-2.6.20-ck1.orig/include/linux/sysctl.h	2007-02-16 19:01:33.000000000 +1100
+++ linux-2.6.20-ck1/include/linux/sysctl.h	2007-02-16 19:01:34.000000000 +1100
@@ -204,6 +204,7 @@ enum
 	VM_MIN_SLAB=35,		 /* Percent pages ignored by zone reclaim */
 	VM_SWAP_PREFETCH=36,	/* swap prefetch */
 	VM_HARDMAPLIMIT=37,	/* Make mapped a hard limit */
+	VM_TAIL_LARGEFILES=38,	/* Read large files to lru tail */
 };
 
 
1	niro	106	When reading from large files through the generic file read functions into
2			page cache we can detect when a file is so large that it is unlikely to be
3			fully cached in ram.
4
5			Add a tunable /proc/sys/vm/tail_largefiles that puts them at the tail of the
6			inactive list to minimise their harm on present mapped pages and pagecache
7			and enable it by default.
8
9			Signed-off-by: Con Kolivas <kernel@kolivas.org>
10
11			---
12			Documentation/filesystems/proc.txt \| 8 +++++
13			Documentation/sysctl/vm.txt \| 2 -
14			include/linux/sysctl.h \| 1
15			kernel/sysctl.c \| 9 ++++++
16			mm/filemap.c \| 53 +++++++++++++++++++++++++++++++++++--
17			mm/swap.c \| 3 --
18			6 files changed, 71 insertions(+), 5 deletions(-)
19
20			Index: linux-2.6.20-ck1/mm/filemap.c
21			===================================================================
22			--- linux-2.6.20-ck1.orig/mm/filemap.c 2007-02-05 22:52:04.000000000 +1100
23			+++ linux-2.6.20-ck1/mm/filemap.c 2007-02-16 19:01:34.000000000 +1100
24			@@ -466,6 +466,16 @@ int add_to_page_cache_lru(struct page *p
25			return ret;
26			}
27
28			+int add_to_page_cache_lru_tail(struct page *page,
29			+ struct address_space *mapping, pgoff_t offset, gfp_t gfp_mask)
30			+{
31			+ int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
32			+
33			+ if (ret == 0)
34			+ lru_cache_add_tail(page);
35			+ return ret;
36			+}
37			+
38			#ifdef CONFIG_NUMA
39			struct page *__page_cache_alloc(gfp_t gfp)
40			{
41			@@ -856,6 +866,34 @@ static void shrink_readahead_size_eio(st
42			ra->ra_pages /= 4;
43			}
44
45			+/*
46			+ * Sysctl which determines whether we should read from large files to the
47			+ * tail of the inactive lru list.
48			+ */
49			+int vm_tail_largefiles __read_mostly = 1;
50			+
51			+static inline int nr_mapped(void)
52			+{
53			+ return global_page_state(NR_FILE_MAPPED) +
54			+ global_page_state(NR_ANON_PAGES);
55			+}
56			+
57			+/*
58			+ * This examines how large in pages a file size is and returns 1 if it is
59			+ * more than half the unmapped ram. Avoid doing read_page_state which is
60			+ * expensive unless we already know it is likely to be large enough.
61			+ */
62			+static int large_isize(unsigned long nr_pages)
63			+{
64			+ if (nr_pages * 6 > vm_total_pages) {
65			+ unsigned long unmapped_ram = vm_total_pages - nr_mapped();
66			+
67			+ if (nr_pages * 2 > unmapped_ram)
68			+ return 1;
69			+ }
70			+ return 0;
71			+}
72			+
73			/**
74			* do_generic_mapping_read - generic file read routine
75			* @mapping: address_space to be read
76			@@ -1064,8 +1102,19 @@ no_cached_page:
77			goto out;
78			}
79			}
80			- error = add_to_page_cache_lru(cached_page, mapping,
81			- index, GFP_KERNEL);
82			+
83			+ /*
84			+ * If we know the file is large we add the pages read to the
85			+ * end of the lru as we're unlikely to be able to cache the
86			+ * whole file in ram so make those pages the first to be
87			+ * dropped if not referenced soon.
88			+ */
89			+ if (vm_tail_largefiles && large_isize(end_index))
90			+ error = add_to_page_cache_lru_tail(cached_page,
91			+ mapping, index, GFP_KERNEL);
92			+ else
93			+ error = add_to_page_cache_lru(cached_page, mapping,
94			+ index, GFP_KERNEL);
95			if (error) {
96			if (error == -EEXIST)
97			goto find_page;
98			Index: linux-2.6.20-ck1/mm/swap.c
99			===================================================================
100			--- linux-2.6.20-ck1.orig/mm/swap.c 2007-02-16 19:01:33.000000000 +1100
101			+++ linux-2.6.20-ck1/mm/swap.c 2007-02-16 19:01:34.000000000 +1100
102			@@ -434,8 +434,7 @@ void __pagevec_lru_add_active(struct pag
103
104			/*
105			* Function used uniquely to put pages back to the lru at the end of the
106			- * inactive list to preserve the lru order. Currently only used by swap
107			- * prefetch.
108			+ * inactive list to preserve the lru order.
109			*/
110			void fastcall lru_cache_add_tail(struct page *page)
111			{
112			Index: linux-2.6.20-ck1/kernel/sysctl.c
113			===================================================================
114			--- linux-2.6.20-ck1.orig/kernel/sysctl.c 2007-02-16 19:01:33.000000000 +1100
115			+++ linux-2.6.20-ck1/kernel/sysctl.c 2007-02-16 19:01:34.000000000 +1100
116			@@ -77,6 +77,7 @@ extern int pid_max_min, pid_max_max;
117			extern int sysctl_drop_caches;
118			extern int percpu_pagelist_fraction;
119			extern int compat_log;
120			+extern int vm_tail_largefiles;
121
122			/* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
123			static int maxolduid = 65535;
124			@@ -918,6 +919,14 @@ static ctl_table vm_table[] = {
125			.mode = 0644,
126			.proc_handler = &proc_dointvec,
127			},
128			+ {
129			+ .ctl_name = VM_TAIL_LARGEFILES,
130			+ .procname = "tail_largefiles",
131			+ .data = &vm_tail_largefiles,
132			+ .maxlen = sizeof(int),
133			+ .mode = 0644,
134			+ .proc_handler = &proc_dointvec,
135			+ },
136			#ifdef CONFIG_HUGETLB_PAGE
137			{
138			.ctl_name = VM_HUGETLB_PAGES,
139			Index: linux-2.6.20-ck1/Documentation/filesystems/proc.txt
140			===================================================================
141			--- linux-2.6.20-ck1.orig/Documentation/filesystems/proc.txt 2007-02-05 22:51:59.000000000 +1100
142			+++ linux-2.6.20-ck1/Documentation/filesystems/proc.txt 2007-02-16 19:01:34.000000000 +1100
143			@@ -1324,6 +1324,14 @@ To free pagecache, dentries and inodes:
144			As this is a non-destructive operation and dirty objects are not freeable, the
145			user should run `sync' first.
146
147			+tail_largefiles
148			+---------------
149			+
150			+When enabled reads from large files to the tail end of the inactive lru list.
151			+This means that any cache from reading large files is dropped very quickly,
152			+preventing loss of mapped ram and useful pagecache when large files are read.
153			+This does, however, make caching less effective when working with large files.
154			+
155
156			2.5 /proc/sys/dev - Device specific parameters
157			----------------------------------------------
158			Index: linux-2.6.20-ck1/Documentation/sysctl/vm.txt
159			===================================================================
160			--- linux-2.6.20-ck1.orig/Documentation/sysctl/vm.txt 2007-02-16 19:01:33.000000000 +1100
161			+++ linux-2.6.20-ck1/Documentation/sysctl/vm.txt 2007-02-16 19:01:34.000000000 +1100
162			@@ -39,7 +39,7 @@ Currently, these files are in /proc/sys/
163
164			dirty_ratio, dirty_background_ratio, dirty_expire_centisecs,
165			dirty_writeback_centisecs, vfs_cache_pressure, laptop_mode,
166			-block_dump, swap_token_timeout, drop-caches:
167			+block_dump, swap_token_timeout, drop-caches, tail_largefiles:
168
169			See Documentation/filesystems/proc.txt
170
171			Index: linux-2.6.20-ck1/include/linux/sysctl.h
172			===================================================================
173			--- linux-2.6.20-ck1.orig/include/linux/sysctl.h 2007-02-16 19:01:33.000000000 +1100
174			+++ linux-2.6.20-ck1/include/linux/sysctl.h 2007-02-16 19:01:34.000000000 +1100
175			@@ -204,6 +204,7 @@ enum
176			VM_MIN_SLAB=35, /* Percent pages ignored by zone reclaim */
177			VM_SWAP_PREFETCH=36, /* swap prefetch */
178			VM_HARDMAPLIMIT=37, /* Make mapped a hard limit */
179			+ VM_TAIL_LARGEFILES=38, /* Read large files to lru tail */
180			};
181
182