Thrashing notification

This patch provides a way to notify user applications when the system is about to thrash. It checks the scanning priority of the inactive lru list and notifies user applications via sysfs when the priority reaches a threshold. In comparison to Marcelo Tosatti's oom notification patch, this patch also works on systems without swap.

Applications can poll() on this sysfs file and can then free memory in one way or another to prevent an oom situation.

Using a test application http://spng.se/oomtest/ that uses multiple allocator threads and a single release thread one can see that this works fairly well. See http://spng.se/oomtest/ for more details and graphs.

Signed-off-by Daniel Sp??ng <[email protected]>

diff -purN linux-2.6.23.1-mm1/include/linux/thrashing_notify.h linux-2.6.23.1-mm1_thrashing/include/linux/thrashing_notify.h
--- linux-2.6.23.1-mm1/include/linux/thrashing_notify.h 1970-01-01
01:00:00.000000000 +0100
+++
linux-2.6.23.1-mm1_thrashing/include/linux/thrashing_notify.h 2007-11-05 14:23:26.000000000 +0100

(quoted diff) +8 -0

@@ -0,0 +1,8 @@
+#ifndef _LINUX_THRASHING_NOTIFY_H
+#define _LINUX_THRASHING_NOTIFY_H
+
+void thrashing_notify(int priority);
+
+extern int thrashing_notifier_threshold;
+
+#endif /* _LINUX_THRASHING_NOTIFY_H */

diff -purN linux-2.6.23.1-mm1/kernel/sysctl.c linux-2.6.23.1-mm1_thrashing/kernel/sysctl.c
--- linux-2.6.23.1-mm1/kernel/sysctl.c 2007-11-01 14:59:16.000000000
+0100
+++ linux-2.6.23.1-mm1_thrashing/kernel/sysctl.c 2007-11-05
14:22:29.000000000 +0100

(quoted diff) +12 -0

@@ -46,6 +46,7 @@
 #include <linux/nfs_fs.h>
 #include <linux/acpi.h>
 #include <linux/reboot.h>
+#include <linux/thrashing_notify.h>
 #include <asm/uaccess.h>
 #include <asm/processor.h>

@@ -102,6 +103,7 @@ static int minolduid;
 static int min_percpu_pagelist_fract = 8;
 static int ngroups_max = NGROUPS_MAX;
+static int def_priority = DEF_PRIORITY;
 #ifdef CONFIG_KMOD
 extern char modprobe_path[];

@@ -1071,6 +1073,16 @@ static struct ctl_table vm_table[] = {
 		.extra1		= &zero,
 	},
 #endif
+	{
+		.ctl_name	= CTL_UNNUMBERED,
+		.procname	= "thrashing_notifier_threshold",
+		.data		= &thrashing_notifier_threshold,
+		.maxlen		= sizeof thrashing_notifier_threshold,
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec_minmax,
+		.extra1		= &zero,
+		.extra2		= &def_priority,
+	},
 /*
  * NOTE: do not add new entries to this table unless you have read
  * Documentation/sysctl/ctl_unnumbered.txt

diff -purN linux-2.6.23.1-mm1/mm/Makefile linux-2.6.23.1-mm1_thrashing/mm/Makefile

mm/Makefile +0 -0

14:22:11.000000000 +0100

mm/Makefile +1 -1

@@ -11,7 +11,7 @@ obj-y			:= bootmem.o filemap.o mempool.o
 			   page_alloc.o page-writeback.o pdflush.o \
 			   readahead.o swap.o truncate.o vmscan.o \
 			   prio_tree.o util.o mmzone.o vmstat.o backing-dev.o \
-			   page_isolation.o $(mmu-y)
+			   page_isolation.o thrashing_notify.o $(mmu-y)

obj-$(CONFIG_BOUNCE) += bounce.o
obj-$(CONFIG_SWAP) += page_io.o swap_state.o swapfile.o thrash.o
diff -purN linux-2.6.23.1-mm1/mm/thrashing_notify.c linux-2.6.23.1-mm1_thrashing/mm/thrashing_notify.c
--- linux-2.6.23.1-mm1/mm/thrashing_notify.c 1970-01-01
01:00:00.000000000 +0100
+++ linux-2.6.23.1-mm1_thrashing/mm/thrashing_notify.c 2007-11-05
14:22:46.000000000 +0100

mm/Makefile +32 -0

@@ -0,0 +1,56 @@
+/*
+ * mm/thrashing_notify.c
+ *
+ * Copyright (C) 2007 Daniel Sp??ng <[email protected]>
+ *
+ * Released under the GPL, see the file COPYING for details.
+ */
+
+#include <linux/thrashing_notify.h>
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/mman.h>
+#include <linux/init.h>
+#include <linux/types.h>
+#include <linux/kobject.h>
+#include <linux/sysfs.h>
+
+/*
+ * The count of thrashing occasions.
+ *
+ * Published to userspace at /sys/kernel/nr_thrashing
+ */
+int nr_thrashing = 0;
+
+int thrashing_notifier_threshold = 4;
+
+static ssize_t nr_thrashing_show(struct kset *kset, char *page)
+{
+	return sprintf(page, "%u\n", nr_thrashing);
+}
+
+static struct subsys_attribute nr_thrashing_attr =  

__ATTR_RO(nr_thrashing);

mm/Makefile +24 -0

+
+static struct attribute *nr_thrashing_attrs[] = {
+	&nr_thrashing_attr.attr,
+	NULL,
+};
+
+static struct attribute_group nr_thrashing_attr_group = {
+	.attrs  = nr_thrashing_attrs,
+};
+
+void thrashing_notify(int priority)
+{
+	nr_thrashing++;
+	sysfs_notify(&kernel_subsys.kobj, NULL, "nr_thrashing");
+}
+
+static int __init thrashing_init(void)
+{
+	return sysfs_create_group(&kernel_subsys.kobj,
+			       &nr_thrashing_attr_group);
+}
+
+module_init(thrashing_init)
+

diff -purN linux-2.6.23.1-mm1/mm/vmscan.c linux-2.6.23.1-mm1_thrashing/mm/vmscan.c

mm/vmscan.c +0 -0

14:21:55.000000000 +0100

mm/vmscan.c +3 -0

@@ -39,6 +39,7 @@
 #include <linux/kthread.h>
 #include <linux/freezer.h>
 #include <linux/memcontrol.h>
+#include <linux/thrashing_notify.h>
 #include <asm/tlbflush.h>
 #include <asm/div64.h>

@@ -1285,6 +1286,9 @@ static unsigned long do_try_to_free_page
 		sc->nr_io_pages = 0;
 		if (!priority)
 			disable_swap_token();
+		if (priority == thrashing_notifier_threshold)
+			thrashing_notify(priority);
 		nr_reclaimed += shrink_zones(priority, zones, sc);
 		/*
 		 * Don't shrink slabs when reclaiming memory from

This patch provides a way to notify user applications when the system is about to thrash. It checks the scanning priority of the inactive lru list and notifies user applications via sysfs when the priority reaches a threshold. In comparison to Marcelo Tosatti's oom notification patch, this patch also works on systems without swap.

Applications can poll() on this sysfs file and can then free memory in one way or another to prevent an oom situation.

Using a test application http://spng.se/oomtest/ that uses multiple allocator threads and a single release thread one can see that this works fairly well. See http://spng.se/oomtest/ for more details and graphs.

Signed-off-by Daniel Spang <[email protected]>

diff -purN linux-2.6.23.1-mm1/include/linux/thrashing_notify.h linux-2.6.23.1-mm1_thrashing/include/linux/thrashing_notify.h

include/linux/thrashing_notify.h +8 -0

@@ -0,0 +1,8 @@
+#ifndef _LINUX_THRASHING_NOTIFY_H
+#define _LINUX_THRASHING_NOTIFY_H
+
+void thrashing_notify(int priority);
+
+extern int thrashing_notifier_threshold;
+
+#endif /* _LINUX_THRASHING_NOTIFY_H */

diff -purN linux-2.6.23.1-mm1/kernel/sysctl.c linux-2.6.23.1-mm1_thrashing/kernel/sysctl.c

kernel/sysctl.c +12 -0

@@ -46,6 +46,7 @@
 #include <linux/nfs_fs.h>
 #include <linux/acpi.h>
 #include <linux/reboot.h>
+#include <linux/thrashing_notify.h>
 
 #include <asm/uaccess.h>
 #include <asm/processor.h>

@@ -102,6 +103,7 @@ static int minolduid;
 static int min_percpu_pagelist_fract = 8;
 
 static int ngroups_max = NGROUPS_MAX;
+static int def_priority = DEF_PRIORITY;
 
 #ifdef CONFIG_KMOD
 extern char modprobe_path[];

@@ -1071,6 +1073,16 @@ static struct ctl_table vm_table[] = {
 		.extra1		= &zero,
 	},
 #endif
+	{
+		.ctl_name	= CTL_UNNUMBERED,
+		.procname	= "thrashing_notifier_threshold",
+		.data		= &thrashing_notifier_threshold,
+		.maxlen		= sizeof thrashing_notifier_threshold,
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec_minmax,
+		.extra1		= &zero,
+		.extra2		= &def_priority,
+	},
 /*
  * NOTE: do not add new entries to this table unless you have read
  * Documentation/sysctl/ctl_unnumbered.txt

diff -purN linux-2.6.23.1-mm1/mm/Makefile linux-2.6.23.1-mm1_thrashing/mm/Makefile

mm/Makefile +1 -1

@@ -11,7 +11,7 @@ obj-y			:= bootmem.o filemap.o mempool.o
 			   page_alloc.o page-writeback.o pdflush.o \
 			   readahead.o swap.o truncate.o vmscan.o \
 			   prio_tree.o util.o mmzone.o vmstat.o backing-dev.o \
-			   page_isolation.o $(mmu-y)
+			   page_isolation.o thrashing_notify.o $(mmu-y)
 
 obj-$(CONFIG_BOUNCE)	+= bounce.o
 obj-$(CONFIG_SWAP)	+= page_io.o swap_state.o swapfile.o thrash.o

diff -purN linux-2.6.23.1-mm1/mm/thrashing_notify.c linux-2.6.23.1-mm1_thrashing/mm/thrashing_notify.c

mm/thrashing_notify.c +56 -0

@@ -0,0 +1,56 @@
+/*
+ * mm/thrashing_notify.c
+ *
+ * Copyright (C) 2007 Daniel Spang <[email protected]>
+ *
+ * Released under the GPL, see the file COPYING for details.
+ */
+
+#include <linux/thrashing_notify.h>
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/mman.h>
+#include <linux/init.h>
+#include <linux/types.h>
+#include <linux/kobject.h>
+#include <linux/sysfs.h>
+
+/*
+ * The count of thrashing occasions.
+ *
+ * Published to userspace at /sys/kernel/nr_thrashing
+ */
+int nr_thrashing = 0;
+
+int thrashing_notifier_threshold = 4;
+
+static ssize_t nr_thrashing_show(struct kset *kset, char *page)
+{
+	return sprintf(page, "%u\n", nr_thrashing);
+}
+
+static struct subsys_attribute nr_thrashing_attr = __ATTR_RO(nr_thrashing);
+
+static struct attribute *nr_thrashing_attrs[] = {
+	&nr_thrashing_attr.attr,
+	NULL,
+};
+
+static struct attribute_group nr_thrashing_attr_group = {
+	.attrs  = nr_thrashing_attrs,
+};
+
+void thrashing_notify(int priority)
+{
+	nr_thrashing++;
+	sysfs_notify(&kernel_subsys.kobj, NULL, "nr_thrashing");
+}
+
+static int __init thrashing_init(void)
+{
+	return sysfs_create_group(&kernel_subsys.kobj,
+			       &nr_thrashing_attr_group);
+}
+
+module_init(thrashing_init)
+

diff -purN linux-2.6.23.1-mm1/mm/vmscan.c linux-2.6.23.1-mm1_thrashing/mm/vmscan.c

mm/vmscan.c +3 -0

@@ -39,6 +39,7 @@
 #include <linux/kthread.h>
 #include <linux/freezer.h>
 #include <linux/memcontrol.h>
+#include <linux/thrashing_notify.h>
 
 #include <asm/tlbflush.h>
 #include <asm/div64.h>

@@ -1285,6 +1286,9 @@ static unsigned long do_try_to_free_page
 		sc->nr_io_pages = 0;
 		if (!priority)
 			disable_swap_token();
+		if (priority == thrashing_notifier_threshold)
+			thrashing_notify(priority);
 		nr_reclaimed += shrink_zones(priority, zones, sc);
 		/*
 		 * Don't shrink slabs when reclaiming memory from

Hooking into try_to_free_pages() makes the scheme suspectible to specifics such as:

Remember that notifications are sent to applications which can allocate globally...

Quoted 2 lines · On Mon, 5 Nov 2007 13:30:25 -0500 Marcelo Tosatti <[email protected]> wrote:

Hooking into try_to_free_pages() makes the scheme suspectible to specifics such as:

The specific of where the hook is can be changed. I am sure the two of you can come up with the best way to do things. Just keep shooting holes in each other's ideas until one idea remains which neither of you can find a problem with[1] :)

Quoted 2 lines

Remember that notifications are sent to applications which can allocate globally...

This is the bigger problem with the sysfs code: every task that watches the sysfs node will get woken up. That could be a big problem when there are hundreds of processes watching that file.

Marcelo's code, which only wakes up one task at a time, has the potential to work much better. That code can also be enhanced to wake up tasks that use a lot of memory on the specific NUMA node that has a memory shortage.

[1] Yes, that is how I usually come up with VM ideas :)

I have actually no problem at all using a device to get the message to userspace. My patch was more like a demonstration of when to trigger the notification. I still (obviously) think that we need a notification for systems without swap too.

A concern, or feature =), with the notify-on-swap method is that with responsive user applications, it will never use swap at all. There are for sure systems where this behavior is desirable, but for example desktop systems, the memory occupied by inactive processes might be better used by active ones.

To get out of the "my patch is better" line of conversation, I guess you and Marcelo should probably try to figure out some threshold that you both agree on.

Quoted 5 lines

A concern, or feature =), with the notify-on-swap method is that with responsive user applications, it will never use swap at all. There are for sure systems where this behavior is desirable, but for example desktop systems, the memory occupied by inactive processes might be better used by active ones.

Well, if the inactive processes get woken up by the low memory notification and free some of their memory, the active processes will use the memory from the inactive ones :)

Hooking into try_to_free_pages() makes the scheme suspectible to specifics such as:

- can the task writeout pages?
- is the allocation a higher order one?
- in what zones is it operating on?

Remember that notifications are sent to applications which can allocate globally... It is not very useful to send notifications for a userspace which has a large percentage of its memory in highmem if the system is having a lowmem zone shortage (granted that the notify-on-swap heuristic has that problem, but you can then argue that swap affects system performance globally, and it generally does in desktop systems).

Other than that tuning "priority" from try_to_free_pages() is rather difficult for users/admins.

My previous patches had the zone limitation, but the following way of asking "are we low on memory?" gets rid of it:

(quoted diff) +14 -0

+static unsigned int mem_notify_poll(struct file *file, poll_table *wait)
+{
+       unsigned int val = 0;
+       struct zone *zone;
+       int tpages_low, tpages_free, tpages_reserve;
+
+       tpages_low = tpages_free = tpages_reserve = 0;
+
+       poll_wait(file, &mem_wait, wait);
+
+       for_each_zone(zone) {
+               if (!populated_zone(zone))
+                       continue;
+               tpages_low += zone->pages_low;