mm: introduce new field "managed_pages" to struct zone
authorJiang Liu <liuj97@gmail.com>
Wed, 12 Dec 2012 21:52:12 +0000 (13:52 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 13 Dec 2012 01:38:34 +0000 (17:38 -0800)
Currently a zone's present_pages is calcuated as below, which is
inaccurate and may cause trouble to memory hotplug.

spanned_pages - absent_pages - memmap_pages - dma_reserve.

During fixing bugs caused by inaccurate zone->present_pages, we found
zone->present_pages has been abused.  The field zone->present_pages may
have different meanings in different contexts:

1) pages existing in a zone.
2) pages managed by the buddy system.

For more discussions about the issue, please refer to:
  http://lkml.org/lkml/2012/11/5/866
  https://patchwork.kernel.org/patch/1346751/

This patchset tries to introduce a new field named "managed_pages" to
struct zone, which counts "pages managed by the buddy system".  And revert
zone->present_pages to count "physical pages existing in a zone", which
also keep in consistence with pgdat->node_present_pages.

We will set an initial value for zone->managed_pages in function
free_area_init_core() and will adjust it later if the initial value is
inaccurate.

For DMA/normal zones, the initial value is set to:

(spanned_pages - absent_pages - memmap_pages - dma_reserve)

Later zone->managed_pages will be adjusted to the accurate value when the
bootmem allocator frees all free pages to the buddy system in function
free_all_bootmem_node() and free_all_bootmem().

The bootmem allocator doesn't touch highmem pages, so highmem zones'
managed_pages is set to the accurate value "spanned_pages - absent_pages"
in function free_area_init_core() and won't be updated anymore.

This patch also adds a new field "managed_pages" to /proc/zoneinfo
and sysrq showmem.

[akpm@linux-foundation.org: small comment tweaks]
Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Maciej Rutecki <maciej.rutecki@gmail.com>
Tested-by: Chris Clayton <chris2553@googlemail.com>
Cc: "Rafael J . Wysocki" <rjw@sisk.pl>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/mmzone.h
mm/bootmem.c
mm/memory_hotplug.c
mm/nobootmem.c
mm/page_alloc.c
mm/vmstat.c

index 0c0b1d6..cd55dad 100644 (file)
@@ -460,17 +460,44 @@ struct zone {
        unsigned long           zone_start_pfn;
 
        /*
-        * zone_start_pfn, spanned_pages and present_pages are all
-        * protected by span_seqlock.  It is a seqlock because it has
-        * to be read outside of zone->lock, and it is done in the main
-        * allocator path.  But, it is written quite infrequently.
+        * spanned_pages is the total pages spanned by the zone, including
+        * holes, which is calculated as:
+        *      spanned_pages = zone_end_pfn - zone_start_pfn;
         *
-        * The lock is declared along with zone->lock because it is
+        * present_pages is physical pages existing within the zone, which
+        * is calculated as:
+        *      present_pages = spanned_pages - absent_pages(pags in holes);
+        *
+        * managed_pages is present pages managed by the buddy system, which
+        * is calculated as (reserved_pages includes pages allocated by the
+        * bootmem allocator):
+        *      managed_pages = present_pages - reserved_pages;
+        *
+        * So present_pages may be used by memory hotplug or memory power
+        * management logic to figure out unmanaged pages by checking
+        * (present_pages - managed_pages). And managed_pages should be used
+        * by page allocator and vm scanner to calculate all kinds of watermarks
+        * and thresholds.
+        *
+        * Locking rules:
+        *
+        * zone_start_pfn and spanned_pages are protected by span_seqlock.
+        * It is a seqlock because it has to be read outside of zone->lock,
+        * and it is done in the main allocator path.  But, it is written
+        * quite infrequently.
+        *
+        * The span_seq lock is declared along with zone->lock because it is
         * frequently read in proximity to zone->lock.  It's good to
         * give them a chance of being in the same cacheline.
+        *
+        * Write access to present_pages and managed_pages at runtime should
+        * be protected by lock_memory_hotplug()/unlock_memory_hotplug().
+        * Any reader who can't tolerant drift of present_pages and
+        * managed_pages should hold memory hotplug lock to get a stable value.
         */
-       unsigned long           spanned_pages;  /* total size, including holes */
-       unsigned long           present_pages;  /* amount of memory (excluding holes) */
+       unsigned long           spanned_pages;
+       unsigned long           present_pages;
+       unsigned long           managed_pages;
 
        /*
         * rarely used fields:
index 26d057a..19262ac 100644 (file)
@@ -229,6 +229,22 @@ static unsigned long __init free_all_bootmem_core(bootmem_data_t *bdata)
        return count;
 }
 
+static void reset_node_lowmem_managed_pages(pg_data_t *pgdat)
+{
+       struct zone *z;
+
+       /*
+        * In free_area_init_core(), highmem zone's managed_pages is set to
+        * present_pages, and bootmem allocator doesn't allocate from highmem
+        * zones. So there's no need to recalculate managed_pages because all
+        * highmem pages will be managed by the buddy system. Here highmem
+        * zone also includes highmem movable zone.
+        */
+       for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
+               if (!is_highmem(z))
+                       z->managed_pages = 0;
+}
+
 /**
  * free_all_bootmem_node - release a node's free pages to the buddy allocator
  * @pgdat: node to be released
@@ -238,6 +254,7 @@ static unsigned long __init free_all_bootmem_core(bootmem_data_t *bdata)
 unsigned long __init free_all_bootmem_node(pg_data_t *pgdat)
 {
        register_page_bootmem_info_node(pgdat);
+       reset_node_lowmem_managed_pages(pgdat);
        return free_all_bootmem_core(pgdat->bdata);
 }
 
@@ -250,6 +267,10 @@ unsigned long __init free_all_bootmem(void)
 {
        unsigned long total_pages = 0;
        bootmem_data_t *bdata;
+       struct pglist_data *pgdat;
+
+       for_each_online_pgdat(pgdat)
+               reset_node_lowmem_managed_pages(pgdat);
 
        list_for_each_entry(bdata, &bdata_list, list)
                total_pages += free_all_bootmem_core(bdata);
index c6cd8b5..b7c93ca 100644 (file)
@@ -106,6 +106,7 @@ static void get_page_bootmem(unsigned long info,  struct page *page,
 void __ref put_page_bootmem(struct page *page)
 {
        unsigned long type;
+       static DEFINE_MUTEX(ppb_lock);
 
        type = (unsigned long) page->lru.next;
        BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE ||
@@ -115,7 +116,14 @@ void __ref put_page_bootmem(struct page *page)
                ClearPagePrivate(page);
                set_page_private(page, 0);
                INIT_LIST_HEAD(&page->lru);
+
+               /*
+                * Please refer to comment for __free_pages_bootmem()
+                * for why we serialize here.
+                */
+               mutex_lock(&ppb_lock);
                __free_pages_bootmem(page, 0);
+               mutex_unlock(&ppb_lock);
        }
 
 }
@@ -748,6 +756,7 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages, int online_typ
                return ret;
        }
 
+       zone->managed_pages += onlined_pages;
        zone->present_pages += onlined_pages;
        zone->zone_pgdat->node_present_pages += onlined_pages;
        if (onlined_pages) {
@@ -1321,6 +1330,7 @@ repeat:
        /* reset pagetype flags and makes migrate type to be MOVABLE */
        undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
        /* removal success */
+       zone->managed_pages -= offlined_pages;
        zone->present_pages -= offlined_pages;
        zone->zone_pgdat->node_present_pages -= offlined_pages;
        totalram_pages -= offlined_pages;
index bd82f6b..b8294fc 100644 (file)
@@ -137,6 +137,22 @@ unsigned long __init free_low_memory_core_early(int nodeid)
        return count;
 }
 
+static void reset_node_lowmem_managed_pages(pg_data_t *pgdat)
+{
+       struct zone *z;
+
+       /*
+        * In free_area_init_core(), highmem zone's managed_pages is set to
+        * present_pages, and bootmem allocator doesn't allocate from highmem
+        * zones. So there's no need to recalculate managed_pages because all
+        * highmem pages will be managed by the buddy system. Here highmem
+        * zone also includes highmem movable zone.
+        */
+       for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
+               if (!is_highmem(z))
+                       z->managed_pages = 0;
+}
+
 /**
  * free_all_bootmem_node - release a node's free pages to the buddy allocator
  * @pgdat: node to be released
@@ -146,6 +162,7 @@ unsigned long __init free_low_memory_core_early(int nodeid)
 unsigned long __init free_all_bootmem_node(pg_data_t *pgdat)
 {
        register_page_bootmem_info_node(pgdat);
+       reset_node_lowmem_managed_pages(pgdat);
 
        /* free_low_memory_core_early(MAX_NUMNODES) will be called later */
        return 0;
@@ -158,6 +175,11 @@ unsigned long __init free_all_bootmem_node(pg_data_t *pgdat)
  */
 unsigned long __init free_all_bootmem(void)
 {
+       struct pglist_data *pgdat;
+
+       for_each_online_pgdat(pgdat)
+               reset_node_lowmem_managed_pages(pgdat);
+
        /*
         * We need to use MAX_NUMNODES instead of NODE_DATA(0)->node_id
         *  because in some case like Node0 doesn't have RAM installed
index 2bf0d43..0b6a6d0 100644 (file)
@@ -735,6 +735,13 @@ static void __free_pages_ok(struct page *page, unsigned int order)
        local_irq_restore(flags);
 }
 
+/*
+ * Read access to zone->managed_pages is safe because it's unsigned long,
+ * but we still need to serialize writers. Currently all callers of
+ * __free_pages_bootmem() except put_page_bootmem() should only be used
+ * at boot time. So for shorter boot time, we shift the burden to
+ * put_page_bootmem() to serialize writers.
+ */
 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
 {
        unsigned int nr_pages = 1 << order;
@@ -750,6 +757,7 @@ void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
                set_page_count(p, 0);
        }
 
+       page_zone(page)->managed_pages += 1 << order;
        set_page_refcounted(page);
        __free_pages(page, order);
 }
@@ -2984,6 +2992,7 @@ void show_free_areas(unsigned int filter)
                        " isolated(anon):%lukB"
                        " isolated(file):%lukB"
                        " present:%lukB"
+                       " managed:%lukB"
                        " mlocked:%lukB"
                        " dirty:%lukB"
                        " writeback:%lukB"
@@ -3013,6 +3022,7 @@ void show_free_areas(unsigned int filter)
                        K(zone_page_state(zone, NR_ISOLATED_ANON)),
                        K(zone_page_state(zone, NR_ISOLATED_FILE)),
                        K(zone->present_pages),
+                       K(zone->managed_pages),
                        K(zone_page_state(zone, NR_MLOCK)),
                        K(zone_page_state(zone, NR_FILE_DIRTY)),
                        K(zone_page_state(zone, NR_WRITEBACK)),
@@ -4502,48 +4512,54 @@ static void __paginginit free_area_init_core(struct pglist_data *pgdat,
 
        for (j = 0; j < MAX_NR_ZONES; j++) {
                struct zone *zone = pgdat->node_zones + j;
-               unsigned long size, realsize, memmap_pages;
+               unsigned long size, realsize, freesize, memmap_pages;
 
                size = zone_spanned_pages_in_node(nid, j, zones_size);
-               realsize = size - zone_absent_pages_in_node(nid, j,
+               realsize = freesize = size - zone_absent_pages_in_node(nid, j,
                                                                zholes_size);
 
                /*
-                * Adjust realsize so that it accounts for how much memory
+                * Adjust freesize so that it accounts for how much memory
                 * is used by this zone for memmap. This affects the watermark
                 * and per-cpu initialisations
                 */
                memmap_pages =
                        PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
-               if (realsize >= memmap_pages) {
-                       realsize -= memmap_pages;
+               if (freesize >= memmap_pages) {
+                       freesize -= memmap_pages;
                        if (memmap_pages)
                                printk(KERN_DEBUG
                                       "  %s zone: %lu pages used for memmap\n",
                                       zone_names[j], memmap_pages);
                } else
                        printk(KERN_WARNING
-                               "  %s zone: %lu pages exceeds realsize %lu\n",
-                               zone_names[j], memmap_pages, realsize);
+                               "  %s zone: %lu pages exceeds freesize %lu\n",
+                               zone_names[j], memmap_pages, freesize);
 
                /* Account for reserved pages */
-               if (j == 0 && realsize > dma_reserve) {
-                       realsize -= dma_reserve;
+               if (j == 0 && freesize > dma_reserve) {
+                       freesize -= dma_reserve;
                        printk(KERN_DEBUG "  %s zone: %lu pages reserved\n",
                                        zone_names[0], dma_reserve);
                }
 
                if (!is_highmem_idx(j))
-                       nr_kernel_pages += realsize;
-               nr_all_pages += realsize;
+                       nr_kernel_pages += freesize;
+               nr_all_pages += freesize;
 
                zone->spanned_pages = size;
-               zone->present_pages = realsize;
+               zone->present_pages = freesize;
+               /*
+                * Set an approximate value for lowmem here, it will be adjusted
+                * when the bootmem allocator frees pages into the buddy system.
+                * And all highmem pages will be managed by the buddy system.
+                */
+               zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
 #ifdef CONFIG_NUMA
                zone->node = nid;
-               zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
+               zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
                                                / 100;
-               zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
+               zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
 #endif
                zone->name = zone_names[j];
                spin_lock_init(&zone->lock);
index 9a4a522..df14808 100644 (file)
@@ -994,14 +994,16 @@ static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
                   "\n        high     %lu"
                   "\n        scanned  %lu"
                   "\n        spanned  %lu"
-                  "\n        present  %lu",
+                  "\n        present  %lu"
+                  "\n        managed  %lu",
                   zone_page_state(zone, NR_FREE_PAGES),
                   min_wmark_pages(zone),
                   low_wmark_pages(zone),
                   high_wmark_pages(zone),
                   zone->pages_scanned,
                   zone->spanned_pages,
-                  zone->present_pages);
+                  zone->present_pages,
+                  zone->managed_pages);
 
        for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
                seq_printf(m, "\n    %-12s %lu", vmstat_text[i],