mm: memory-hotplug: enable memory hotplug to handle hugepage
authorNaoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Wed, 11 Sep 2013 21:22:09 +0000 (14:22 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 11 Sep 2013 22:57:48 +0000 (15:57 -0700)
Until now we can't offline memory blocks which contain hugepages because a
hugepage is considered as an unmovable page.  But now with this patch
series, a hugepage has become movable, so by using hugepage migration we
can offline such memory blocks.

What's different from other users of hugepage migration is that we need to
decompose all the hugepages inside the target memory block into free buddy
pages after hugepage migration, because otherwise free hugepages remaining
in the memory block intervene the memory offlining.  For this reason we
introduce new functions dissolve_free_huge_page() and
dissolve_free_huge_pages().

Other than that, what this patch does is straightforwardly to add hugepage
migration code, that is, adding hugepage code to the functions which scan
over pfn and collect hugepages to be migrated, and adding a hugepage
allocation function to alloc_migrate_target().

As for larger hugepages (1GB for x86_64), it's not easy to do hotremove
over them because it's larger than memory block.  So we now simply leave
it to fail as it is.

[yongjun_wei@trendmicro.com.cn: remove duplicated include]
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Acked-by: Andi Kleen <ak@linux.intel.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Wei Yongjun <yongjun_wei@trendmicro.com.cn>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/hugetlb.h
mm/hugetlb.c
mm/memory_hotplug.c
mm/page_alloc.c
mm/page_isolation.c

index d1db00790a84dc65182cfaca7eb4673bae1d33e2..2e02c4ed1035713e7491ce837a6a62ebcf02fb18 100644 (file)
@@ -68,6 +68,7 @@ void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed);
 int dequeue_hwpoisoned_huge_page(struct page *page);
 bool isolate_huge_page(struct page *page, struct list_head *list);
 void putback_active_hugepage(struct page *page);
+bool is_hugepage_active(struct page *page);
 void copy_huge_page(struct page *dst, struct page *src);
 
 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
@@ -138,6 +139,7 @@ static inline int dequeue_hwpoisoned_huge_page(struct page *page)
 
 #define isolate_huge_page(p, l) false
 #define putback_active_hugepage(p)     do {} while (0)
+#define is_hugepage_active(x)  false
 static inline void copy_huge_page(struct page *dst, struct page *src)
 {
 }
@@ -377,6 +379,9 @@ static inline pgoff_t basepage_index(struct page *page)
        return __basepage_index(page);
 }
 
+extern void dissolve_free_huge_pages(unsigned long start_pfn,
+                                    unsigned long end_pfn);
+
 #else  /* CONFIG_HUGETLB_PAGE */
 struct hstate {};
 #define alloc_huge_page_node(h, nid) NULL
@@ -403,6 +408,7 @@ static inline pgoff_t basepage_index(struct page *page)
 {
        return page->index;
 }
+#define dissolve_free_huge_pages(s, e) do {} while (0)
 #endif /* CONFIG_HUGETLB_PAGE */
 
 #endif /* _LINUX_HUGETLB_H */
index d37b3b95c4392cb8d156ba915827dea820d3483a..fb4293b93fd0ed0ffeb4320d26fdd32c14467a88 100644 (file)
@@ -21,6 +21,7 @@
 #include <linux/rmap.h>
 #include <linux/swap.h>
 #include <linux/swapops.h>
+#include <linux/page-isolation.h>
 
 #include <asm/page.h>
 #include <asm/pgtable.h>
@@ -522,9 +523,15 @@ static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
 {
        struct page *page;
 
-       if (list_empty(&h->hugepage_freelists[nid]))
+       list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
+               if (!is_migrate_isolate_page(page))
+                       break;
+       /*
+        * if 'non-isolated free hugepage' not found on the list,
+        * the allocation fails.
+        */
+       if (&h->hugepage_freelists[nid] == &page->lru)
                return NULL;
-       page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
        list_move(&page->lru, &h->hugepage_activelist);
        set_page_refcounted(page);
        h->free_huge_pages--;
@@ -878,6 +885,44 @@ static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
        return ret;
 }
 
+/*
+ * Dissolve a given free hugepage into free buddy pages. This function does
+ * nothing for in-use (including surplus) hugepages.
+ */
+static void dissolve_free_huge_page(struct page *page)
+{
+       spin_lock(&hugetlb_lock);
+       if (PageHuge(page) && !page_count(page)) {
+               struct hstate *h = page_hstate(page);
+               int nid = page_to_nid(page);
+               list_del(&page->lru);
+               h->free_huge_pages--;
+               h->free_huge_pages_node[nid]--;
+               update_and_free_page(h, page);
+       }
+       spin_unlock(&hugetlb_lock);
+}
+
+/*
+ * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
+ * make specified memory blocks removable from the system.
+ * Note that start_pfn should aligned with (minimum) hugepage size.
+ */
+void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
+{
+       unsigned int order = 8 * sizeof(void *);
+       unsigned long pfn;
+       struct hstate *h;
+
+       /* Set scan step to minimum hugepage size */
+       for_each_hstate(h)
+               if (order > huge_page_order(h))
+                       order = huge_page_order(h);
+       VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
+       for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
+               dissolve_free_huge_page(pfn_to_page(pfn));
+}
+
 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
 {
        struct page *page;
@@ -3457,3 +3502,25 @@ void putback_active_hugepage(struct page *page)
        spin_unlock(&hugetlb_lock);
        put_page(page);
 }
+
+bool is_hugepage_active(struct page *page)
+{
+       VM_BUG_ON(!PageHuge(page));
+       /*
+        * This function can be called for a tail page because the caller,
+        * scan_movable_pages, scans through a given pfn-range which typically
+        * covers one memory block. In systems using gigantic hugepage (1GB
+        * for x86_64,) a hugepage is larger than a memory block, and we don't
+        * support migrating such large hugepages for now, so return false
+        * when called for tail pages.
+        */
+       if (PageTail(page))
+               return false;
+       /*
+        * Refcount of a hwpoisoned hugepages is 1, but they are not active,
+        * so we should return false for them.
+        */
+       if (unlikely(PageHWPoison(page)))
+               return false;
+       return page_count(page) > 0;
+}
index d595606728f9c0d90b9b91e8c49971fbd26c1dc7..0eb1a1df649d8a149b02eb5989682a5c61e73b3f 100644 (file)
@@ -30,6 +30,7 @@
 #include <linux/mm_inline.h>
 #include <linux/firmware-map.h>
 #include <linux/stop_machine.h>
+#include <linux/hugetlb.h>
 
 #include <asm/tlbflush.h>
 
@@ -1230,10 +1231,12 @@ static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
 }
 
 /*
- * Scanning pfn is much easier than scanning lru list.
- * Scan pfn from start to end and Find LRU page.
+ * Scan pfn range [start,end) to find movable/migratable pages (LRU pages
+ * and hugepages). We scan pfn because it's much easier than scanning over
+ * linked list. This function returns the pfn of the first found movable
+ * page if it's found, otherwise 0.
  */
-static unsigned long scan_lru_pages(unsigned long start, unsigned long end)
+static unsigned long scan_movable_pages(unsigned long start, unsigned long end)
 {
        unsigned long pfn;
        struct page *page;
@@ -1242,6 +1245,13 @@ static unsigned long scan_lru_pages(unsigned long start, unsigned long end)
                        page = pfn_to_page(pfn);
                        if (PageLRU(page))
                                return pfn;
+                       if (PageHuge(page)) {
+                               if (is_hugepage_active(page))
+                                       return pfn;
+                               else
+                                       pfn = round_up(pfn + 1,
+                                               1 << compound_order(page)) - 1;
+                       }
                }
        }
        return 0;
@@ -1262,6 +1272,19 @@ do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
                if (!pfn_valid(pfn))
                        continue;
                page = pfn_to_page(pfn);
+
+               if (PageHuge(page)) {
+                       struct page *head = compound_head(page);
+                       pfn = page_to_pfn(head) + (1<<compound_order(head)) - 1;
+                       if (compound_order(head) > PFN_SECTION_SHIFT) {
+                               ret = -EBUSY;
+                               break;
+                       }
+                       if (isolate_huge_page(page, &source))
+                               move_pages -= 1 << compound_order(head);
+                       continue;
+               }
+
                if (!get_page_unless_zero(page))
                        continue;
                /*
@@ -1294,7 +1317,7 @@ do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
        }
        if (!list_empty(&source)) {
                if (not_managed) {
-                       putback_lru_pages(&source);
+                       putback_movable_pages(&source);
                        goto out;
                }
 
@@ -1305,7 +1328,7 @@ do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
                ret = migrate_pages(&source, alloc_migrate_target, 0,
                                        MIGRATE_SYNC, MR_MEMORY_HOTPLUG);
                if (ret)
-                       putback_lru_pages(&source);
+                       putback_movable_pages(&source);
        }
 out:
        return ret;
@@ -1548,8 +1571,8 @@ repeat:
                drain_all_pages();
        }
 
-       pfn = scan_lru_pages(start_pfn, end_pfn);
-       if (pfn) { /* We have page on LRU */
+       pfn = scan_movable_pages(start_pfn, end_pfn);
+       if (pfn) { /* We have movable pages */
                ret = do_migrate_range(pfn, end_pfn);
                if (!ret) {
                        drain = 1;
@@ -1568,6 +1591,11 @@ repeat:
        yield();
        /* drain pcp pages, this is synchronous. */
        drain_all_pages();
+       /*
+        * dissolve free hugepages in the memory block before doing offlining
+        * actually in order to make hugetlbfs's object counting consistent.
+        */
+       dissolve_free_huge_pages(start_pfn, end_pfn);
        /* check again */
        offlined_pages = check_pages_isolated(start_pfn, end_pfn);
        if (offlined_pages < 0) {
index 7c3f8d7e2d8ea60ae44520485ff85ea62a377b72..f7cc08dad26a1b1eedacabc2dfba7bb704bc61da 100644 (file)
@@ -6008,6 +6008,17 @@ bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
                        continue;
 
                page = pfn_to_page(check);
+
+               /*
+                * Hugepages are not in LRU lists, but they're movable.
+                * We need not scan over tail pages bacause we don't
+                * handle each tail page individually in migration.
+                */
+               if (PageHuge(page)) {
+                       iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
+                       continue;
+               }
+
                /*
                 * We can't use page_count without pin a page
                 * because another CPU can free compound page.
index 0cee10ffb98d4cf8e6ad930faa1f4de925bdf3a5..d1473b2e9481731988695755a618baa0991556a7 100644 (file)
@@ -6,6 +6,7 @@
 #include <linux/page-isolation.h>
 #include <linux/pageblock-flags.h>
 #include <linux/memory.h>
+#include <linux/hugetlb.h>
 #include "internal.h"
 
 int set_migratetype_isolate(struct page *page, bool skip_hwpoisoned_pages)
@@ -252,6 +253,19 @@ struct page *alloc_migrate_target(struct page *page, unsigned long private,
 {
        gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
 
+       /*
+        * TODO: allocate a destination hugepage from a nearest neighbor node,
+        * accordance with memory policy of the user process if possible. For
+        * now as a simple work-around, we use the next node for destination.
+        */
+       if (PageHuge(page)) {
+               nodemask_t src = nodemask_of_node(page_to_nid(page));
+               nodemask_t dst;
+               nodes_complement(dst, src);
+               return alloc_huge_page_node(page_hstate(compound_head(page)),
+                                           next_node(page_to_nid(page), dst));
+       }
+
        if (PageHighMem(page))
                gfp_mask |= __GFP_HIGHMEM;