vmscan: unevictable LRU scan sysctl
authorLee Schermerhorn <lee.schermerhorn@hp.com>
Sun, 19 Oct 2008 03:26:53 +0000 (20:26 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Mon, 20 Oct 2008 15:52:31 +0000 (08:52 -0700)
This patch adds a function to scan individual or all zones' unevictable
lists and move any pages that have become evictable onto the respective
zone's inactive list, where shrink_inactive_list() will deal with them.

Adds sysctl to scan all nodes, and per node attributes to individual
nodes' zones.

Kosaki: If evictable page found in unevictable lru when write
/proc/sys/vm/scan_unevictable_pages, print filename and file offset of
these pages.

[akpm@linux-foundation.org: fix one CONFIG_MMU=n build error]
[kosaki.motohiro@jp.fujitsu.com: adapt vmscan-unevictable-lru-scan-sysctl.patch to new sysfs API]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/base/node.c
include/linux/rmap.h
include/linux/swap.h
kernel/sysctl.c
mm/rmap.c
mm/vmscan.c

index fb45d88..f520709 100644 (file)
@@ -13,6 +13,7 @@
 #include <linux/nodemask.h>
 #include <linux/cpu.h>
 #include <linux/device.h>
+#include <linux/swap.h>
 
 static struct sysdev_class node_class = {
        .name = "node",
@@ -191,6 +192,8 @@ int register_node(struct node *node, int num, struct node *parent)
                sysdev_create_file(&node->sysdev, &attr_meminfo);
                sysdev_create_file(&node->sysdev, &attr_numastat);
                sysdev_create_file(&node->sysdev, &attr_distance);
+
+               scan_unevictable_register_node(node);
        }
        return error;
 }
@@ -210,6 +213,8 @@ void unregister_node(struct node *node)
        sysdev_remove_file(&node->sysdev, &attr_numastat);
        sysdev_remove_file(&node->sysdev, &attr_distance);
 
+       scan_unevictable_unregister_node(node);
+
        sysdev_unregister(&node->sysdev);
 }
 
index 955667e..1da48db 100644 (file)
@@ -75,6 +75,9 @@ void anon_vma_unlink(struct vm_area_struct *);
 void anon_vma_link(struct vm_area_struct *);
 void __anon_vma_link(struct vm_area_struct *);
 
+extern struct anon_vma *page_lock_anon_vma(struct page *page);
+extern void page_unlock_anon_vma(struct anon_vma *anon_vma);
+
 /*
  * rmap interfaces called when adding or removing pte of page
  */
index 07eda69..a3af95b 100644 (file)
@@ -7,6 +7,7 @@
 #include <linux/list.h>
 #include <linux/memcontrol.h>
 #include <linux/sched.h>
+#include <linux/node.h>
 
 #include <asm/atomic.h>
 #include <asm/page.h>
@@ -235,15 +236,29 @@ static inline int zone_reclaim(struct zone *z, gfp_t mask, unsigned int order)
 #ifdef CONFIG_UNEVICTABLE_LRU
 extern int page_evictable(struct page *page, struct vm_area_struct *vma);
 extern void scan_mapping_unevictable_pages(struct address_space *);
+
+extern unsigned long scan_unevictable_pages;
+extern int scan_unevictable_handler(struct ctl_table *, int, struct file *,
+                                       void __user *, size_t *, loff_t *);
+extern int scan_unevictable_register_node(struct node *node);
+extern void scan_unevictable_unregister_node(struct node *node);
 #else
 static inline int page_evictable(struct page *page,
                                                struct vm_area_struct *vma)
 {
        return 1;
 }
+
 static inline void scan_mapping_unevictable_pages(struct address_space *mapping)
 {
 }
+
+static inline int scan_unevictable_register_node(struct node *node)
+{
+       return 0;
+}
+
+static inline void scan_unevictable_unregister_node(struct node *node) { }
 #endif
 
 extern int kswapd_run(int nid);
index 617d41e..b3cc739 100644 (file)
@@ -833,6 +833,16 @@ static struct ctl_table kern_table[] = {
                .proc_handler   = &proc_dointvec,
        },
 #endif
+#ifdef CONFIG_UNEVICTABLE_LRU
+       {
+               .ctl_name       = CTL_UNNUMBERED,
+               .procname       = "scan_unevictable_pages",
+               .data           = &scan_unevictable_pages,
+               .maxlen         = sizeof(scan_unevictable_pages),
+               .mode           = 0644,
+               .proc_handler   = &scan_unevictable_handler,
+       },
+#endif
 /*
  * NOTE: do not add new entries to this table unless you have read
  * Documentation/sysctl/ctl_unnumbered.txt
index 7e60df9..7e90beb 100644 (file)
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -181,7 +181,7 @@ void __init anon_vma_init(void)
  * Getting a lock on a stable anon_vma from a page off the LRU is
  * tricky: page_lock_anon_vma rely on RCU to guard against the races.
  */
-static struct anon_vma *page_lock_anon_vma(struct page *page)
+struct anon_vma *page_lock_anon_vma(struct page *page)
 {
        struct anon_vma *anon_vma;
        unsigned long anon_mapping;
@@ -201,7 +201,7 @@ out:
        return NULL;
 }
 
-static void page_unlock_anon_vma(struct anon_vma *anon_vma)
+void page_unlock_anon_vma(struct anon_vma *anon_vma)
 {
        spin_unlock(&anon_vma->lock);
        rcu_read_unlock();
index e5aaaad..ca64e3e 100644 (file)
@@ -39,6 +39,7 @@
 #include <linux/freezer.h>
 #include <linux/memcontrol.h>
 #include <linux/delayacct.h>
+#include <linux/sysctl.h>
 
 #include <asm/tlbflush.h>
 #include <asm/div64.h>
@@ -2363,6 +2364,39 @@ int page_evictable(struct page *page, struct vm_area_struct *vma)
        return 1;
 }
 
+static void show_page_path(struct page *page)
+{
+       char buf[256];
+       if (page_is_file_cache(page)) {
+               struct address_space *mapping = page->mapping;
+               struct dentry *dentry;
+               pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
+
+               spin_lock(&mapping->i_mmap_lock);
+               dentry = d_find_alias(mapping->host);
+               printk(KERN_INFO "rescued: %s %lu\n",
+                      dentry_path(dentry, buf, 256), pgoff);
+               spin_unlock(&mapping->i_mmap_lock);
+       } else {
+#if defined(CONFIG_MM_OWNER) && defined(CONFIG_MMU)
+               struct anon_vma *anon_vma;
+               struct vm_area_struct *vma;
+
+               anon_vma = page_lock_anon_vma(page);
+               if (!anon_vma)
+                       return;
+
+               list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
+                       printk(KERN_INFO "rescued: anon %s\n",
+                              vma->vm_mm->owner->comm);
+                       break;
+               }
+               page_unlock_anon_vma(anon_vma);
+#endif
+       }
+}
+
+
 /**
  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
  * @page: page to check evictability and move to appropriate lru list
@@ -2382,6 +2416,9 @@ retry:
        ClearPageUnevictable(page);
        if (page_evictable(page, NULL)) {
                enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
+
+               show_page_path(page);
+
                __dec_zone_state(zone, NR_UNEVICTABLE);
                list_move(&page->lru, &zone->lru[l].list);
                __inc_zone_state(zone, NR_INACTIVE_ANON + l);
@@ -2451,4 +2488,133 @@ void scan_mapping_unevictable_pages(struct address_space *mapping)
        }
 
 }
+
+/**
+ * scan_zone_unevictable_pages - check unevictable list for evictable pages
+ * @zone - zone of which to scan the unevictable list
+ *
+ * Scan @zone's unevictable LRU lists to check for pages that have become
+ * evictable.  Move those that have to @zone's inactive list where they
+ * become candidates for reclaim, unless shrink_inactive_zone() decides
+ * to reactivate them.  Pages that are still unevictable are rotated
+ * back onto @zone's unevictable list.
+ */
+#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
+void scan_zone_unevictable_pages(struct zone *zone)
+{
+       struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
+       unsigned long scan;
+       unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
+
+       while (nr_to_scan > 0) {
+               unsigned long batch_size = min(nr_to_scan,
+                                               SCAN_UNEVICTABLE_BATCH_SIZE);
+
+               spin_lock_irq(&zone->lru_lock);
+               for (scan = 0;  scan < batch_size; scan++) {
+                       struct page *page = lru_to_page(l_unevictable);
+
+                       if (!trylock_page(page))
+                               continue;
+
+                       prefetchw_prev_lru_page(page, l_unevictable, flags);
+
+                       if (likely(PageLRU(page) && PageUnevictable(page)))
+                               check_move_unevictable_page(page, zone);
+
+                       unlock_page(page);
+               }
+               spin_unlock_irq(&zone->lru_lock);
+
+               nr_to_scan -= batch_size;
+       }
+}
+
+
+/**
+ * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
+ *
+ * A really big hammer:  scan all zones' unevictable LRU lists to check for
+ * pages that have become evictable.  Move those back to the zones'
+ * inactive list where they become candidates for reclaim.
+ * This occurs when, e.g., we have unswappable pages on the unevictable lists,
+ * and we add swap to the system.  As such, it runs in the context of a task
+ * that has possibly/probably made some previously unevictable pages
+ * evictable.
+ */
+void scan_all_zones_unevictable_pages(void)
+{
+       struct zone *zone;
+
+       for_each_zone(zone) {
+               scan_zone_unevictable_pages(zone);
+       }
+}
+
+/*
+ * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
+ * all nodes' unevictable lists for evictable pages
+ */
+unsigned long scan_unevictable_pages;
+
+int scan_unevictable_handler(struct ctl_table *table, int write,
+                          struct file *file, void __user *buffer,
+                          size_t *length, loff_t *ppos)
+{
+       proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
+
+       if (write && *(unsigned long *)table->data)
+               scan_all_zones_unevictable_pages();
+
+       scan_unevictable_pages = 0;
+       return 0;
+}
+
+/*
+ * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
+ * a specified node's per zone unevictable lists for evictable pages.
+ */
+
+static ssize_t read_scan_unevictable_node(struct sys_device *dev,
+                                         struct sysdev_attribute *attr,
+                                         char *buf)
+{
+       return sprintf(buf, "0\n");     /* always zero; should fit... */
+}
+
+static ssize_t write_scan_unevictable_node(struct sys_device *dev,
+                                          struct sysdev_attribute *attr,
+                                       const char *buf, size_t count)
+{
+       struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
+       struct zone *zone;
+       unsigned long res;
+       unsigned long req = strict_strtoul(buf, 10, &res);
+
+       if (!req)
+               return 1;       /* zero is no-op */
+
+       for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
+               if (!populated_zone(zone))
+                       continue;
+               scan_zone_unevictable_pages(zone);
+       }
+       return 1;
+}
+
+
+static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
+                       read_scan_unevictable_node,
+                       write_scan_unevictable_node);
+
+int scan_unevictable_register_node(struct node *node)
+{
+       return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
+}
+
+void scan_unevictable_unregister_node(struct node *node)
+{
+       sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
+}
+
 #endif