memcg: move charges of anonymous page
authorDaisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Wed, 10 Mar 2010 23:22:14 +0000 (15:22 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 12 Mar 2010 23:52:36 +0000 (15:52 -0800)
This patch is the core part of this move-charge-at-task-migration feature.
 It implements functions to move charges of anonymous pages mapped only by
the target task.

Implementation:
- define struct move_charge_struct and a valuable of it(mc) to remember the
  count of pre-charges and other information.
- At can_attach(), get anon_rss of the target mm, call __mem_cgroup_try_charge()
  repeatedly and count up mc.precharge.
- At attach(), parse the page table, find a target page to be move, and call
  mem_cgroup_move_account() about the page.
- Cancel all precharges if mc.precharge > 0 on failure or at the end of
  task move.

[akpm@linux-foundation.org: a little simplification]
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Balbir Singh <balbir@linux.vnet.ibm.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Paul Menage <menage@google.com>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/memcontrol.c

index 59ffaf5..22f088f 100644 (file)
@@ -21,6 +21,7 @@
 #include <linux/memcontrol.h>
 #include <linux/cgroup.h>
 #include <linux/mm.h>
+#include <linux/hugetlb.h>
 #include <linux/pagemap.h>
 #include <linux/smp.h>
 #include <linux/page-flags.h>
@@ -243,9 +244,17 @@ struct mem_cgroup {
  * left-shifted bitmap of these types.
  */
 enum move_type {
+       MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
        NR_MOVE_TYPE,
 };
 
+/* "mc" and its members are protected by cgroup_mutex */
+static struct move_charge_struct {
+       struct mem_cgroup *from;
+       struct mem_cgroup *to;
+       unsigned long precharge;
+} mc;
+
 /*
  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
  * limit reclaim to prevent infinite loops, if they ever occur.
@@ -1513,7 +1522,7 @@ charged:
         * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
         * if they exceeds softlimit.
         */
-       if (mem_cgroup_soft_limit_check(mem))
+       if (page && mem_cgroup_soft_limit_check(mem))
                mem_cgroup_update_tree(mem, page);
 done:
        return 0;
@@ -1690,8 +1699,9 @@ static void __mem_cgroup_move_account(struct page_cgroup *pc,
        /*
         * We charges against "to" which may not have any tasks. Then, "to"
         * can be under rmdir(). But in current implementation, caller of
-        * this function is just force_empty() and it's garanteed that
-        * "to" is never removed. So, we don't check rmdir status here.
+        * this function is just force_empty() and move charge, so it's
+        * garanteed that "to" is never removed. So, we don't check rmdir
+        * status here.
         */
 }
 
@@ -3428,11 +3438,171 @@ static int mem_cgroup_populate(struct cgroup_subsys *ss,
 }
 
 /* Handlers for move charge at task migration. */
-static int mem_cgroup_can_move_charge(void)
+static int mem_cgroup_do_precharge(void)
 {
+       int ret = -ENOMEM;
+       struct mem_cgroup *mem = mc.to;
+
+       ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false, NULL);
+       if (ret || !mem)
+               return -ENOMEM;
+
+       mc.precharge++;
+       return ret;
+}
+
+/**
+ * is_target_pte_for_mc - check a pte whether it is valid for move charge
+ * @vma: the vma the pte to be checked belongs
+ * @addr: the address corresponding to the pte to be checked
+ * @ptent: the pte to be checked
+ * @target: the pointer the target page will be stored(can be NULL)
+ *
+ * Returns
+ *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
+ *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
+ *     move charge. if @target is not NULL, the page is stored in target->page
+ *     with extra refcnt got(Callers should handle it).
+ *
+ * Called with pte lock held.
+ */
+/* We add a new member later. */
+union mc_target {
+       struct page     *page;
+};
+
+/* We add a new type later. */
+enum mc_target_type {
+       MC_TARGET_NONE, /* not used */
+       MC_TARGET_PAGE,
+};
+
+static int is_target_pte_for_mc(struct vm_area_struct *vma,
+               unsigned long addr, pte_t ptent, union mc_target *target)
+{
+       struct page *page;
+       struct page_cgroup *pc;
+       int ret = 0;
+       bool move_anon = test_bit(MOVE_CHARGE_TYPE_ANON,
+                                       &mc.to->move_charge_at_immigrate);
+
+       if (!pte_present(ptent))
+               return 0;
+
+       page = vm_normal_page(vma, addr, ptent);
+       if (!page || !page_mapped(page))
+               return 0;
+       /*
+        * TODO: We don't move charges of file(including shmem/tmpfs) pages for
+        * now.
+        */
+       if (!move_anon || !PageAnon(page))
+               return 0;
+       /*
+        * TODO: We don't move charges of shared(used by multiple processes)
+        * pages for now.
+        */
+       if (page_mapcount(page) > 1)
+               return 0;
+       if (!get_page_unless_zero(page))
+               return 0;
+
+       pc = lookup_page_cgroup(page);
+       /*
+        * Do only loose check w/o page_cgroup lock. mem_cgroup_move_account()
+        * checks the pc is valid or not under the lock.
+        */
+       if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
+               ret = MC_TARGET_PAGE;
+               if (target)
+                       target->page = page;
+       }
+
+       if (!ret || !target)
+               put_page(page);
+
+       return ret;
+}
+
+static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
+                                       unsigned long addr, unsigned long end,
+                                       struct mm_walk *walk)
+{
+       struct vm_area_struct *vma = walk->private;
+       pte_t *pte;
+       spinlock_t *ptl;
+
+       pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
+       for (; addr != end; pte++, addr += PAGE_SIZE)
+               if (is_target_pte_for_mc(vma, addr, *pte, NULL))
+                       mc.precharge++; /* increment precharge temporarily */
+       pte_unmap_unlock(pte - 1, ptl);
+       cond_resched();
+
        return 0;
 }
 
+static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
+{
+       unsigned long precharge;
+       struct vm_area_struct *vma;
+
+       down_read(&mm->mmap_sem);
+       for (vma = mm->mmap; vma; vma = vma->vm_next) {
+               struct mm_walk mem_cgroup_count_precharge_walk = {
+                       .pmd_entry = mem_cgroup_count_precharge_pte_range,
+                       .mm = mm,
+                       .private = vma,
+               };
+               if (is_vm_hugetlb_page(vma))
+                       continue;
+               /* TODO: We don't move charges of shmem/tmpfs pages for now. */
+               if (vma->vm_flags & VM_SHARED)
+                       continue;
+               walk_page_range(vma->vm_start, vma->vm_end,
+                                       &mem_cgroup_count_precharge_walk);
+       }
+       up_read(&mm->mmap_sem);
+
+       precharge = mc.precharge;
+       mc.precharge = 0;
+
+       return precharge;
+}
+
+#define PRECHARGE_AT_ONCE      256
+static int mem_cgroup_precharge_mc(struct mm_struct *mm)
+{
+       int ret = 0;
+       int count = PRECHARGE_AT_ONCE;
+       unsigned long precharge = mem_cgroup_count_precharge(mm);
+
+       while (!ret && precharge--) {
+               if (signal_pending(current)) {
+                       ret = -EINTR;
+                       break;
+               }
+               if (!count--) {
+                       count = PRECHARGE_AT_ONCE;
+                       cond_resched();
+               }
+               ret = mem_cgroup_do_precharge();
+       }
+
+       return ret;
+}
+
+static void mem_cgroup_clear_mc(void)
+{
+       /* we must uncharge all the leftover precharges from mc.to */
+       while (mc.precharge) {
+               mem_cgroup_cancel_charge(mc.to);
+               mc.precharge--;
+       }
+       mc.from = NULL;
+       mc.to = NULL;
+}
+
 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
                                struct cgroup *cgroup,
                                struct task_struct *p,
@@ -3450,11 +3620,19 @@ static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
                mm = get_task_mm(p);
                if (!mm)
                        return 0;
-
                /* We move charges only when we move a owner of the mm */
-               if (mm->owner == p)
-                       ret = mem_cgroup_can_move_charge();
-
+               if (mm->owner == p) {
+                       VM_BUG_ON(mc.from);
+                       VM_BUG_ON(mc.to);
+                       VM_BUG_ON(mc.precharge);
+                       mc.from = from;
+                       mc.to = mem;
+                       mc.precharge = 0;
+
+                       ret = mem_cgroup_precharge_mc(mm);
+                       if (ret)
+                               mem_cgroup_clear_mc();
+               }
                mmput(mm);
        }
        return ret;
@@ -3465,10 +3643,95 @@ static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
                                struct task_struct *p,
                                bool threadgroup)
 {
+       mem_cgroup_clear_mc();
 }
 
-static void mem_cgroup_move_charge(void)
+static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
+                               unsigned long addr, unsigned long end,
+                               struct mm_walk *walk)
 {
+       int ret = 0;
+       struct vm_area_struct *vma = walk->private;
+       pte_t *pte;
+       spinlock_t *ptl;
+
+retry:
+       pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
+       for (; addr != end; addr += PAGE_SIZE) {
+               pte_t ptent = *(pte++);
+               union mc_target target;
+               int type;
+               struct page *page;
+               struct page_cgroup *pc;
+
+               if (!mc.precharge)
+                       break;
+
+               type = is_target_pte_for_mc(vma, addr, ptent, &target);
+               switch (type) {
+               case MC_TARGET_PAGE:
+                       page = target.page;
+                       if (isolate_lru_page(page))
+                               goto put;
+                       pc = lookup_page_cgroup(page);
+                       if (!mem_cgroup_move_account(pc, mc.from, mc.to)) {
+                               css_put(&mc.to->css);
+                               mc.precharge--;
+                       }
+                       putback_lru_page(page);
+put:                   /* is_target_pte_for_mc() gets the page */
+                       put_page(page);
+                       break;
+               default:
+                       break;
+               }
+       }
+       pte_unmap_unlock(pte - 1, ptl);
+       cond_resched();
+
+       if (addr != end) {
+               /*
+                * We have consumed all precharges we got in can_attach().
+                * We try charge one by one, but don't do any additional
+                * charges to mc.to if we have failed in charge once in attach()
+                * phase.
+                */
+               ret = mem_cgroup_do_precharge();
+               if (!ret)
+                       goto retry;
+       }
+
+       return ret;
+}
+
+static void mem_cgroup_move_charge(struct mm_struct *mm)
+{
+       struct vm_area_struct *vma;
+
+       lru_add_drain_all();
+       down_read(&mm->mmap_sem);
+       for (vma = mm->mmap; vma; vma = vma->vm_next) {
+               int ret;
+               struct mm_walk mem_cgroup_move_charge_walk = {
+                       .pmd_entry = mem_cgroup_move_charge_pte_range,
+                       .mm = mm,
+                       .private = vma,
+               };
+               if (is_vm_hugetlb_page(vma))
+                       continue;
+               /* TODO: We don't move charges of shmem/tmpfs pages for now. */
+               if (vma->vm_flags & VM_SHARED)
+                       continue;
+               ret = walk_page_range(vma->vm_start, vma->vm_end,
+                                               &mem_cgroup_move_charge_walk);
+               if (ret)
+                       /*
+                        * means we have consumed all precharges and failed in
+                        * doing additional charge. Just abandon here.
+                        */
+                       break;
+       }
+       up_read(&mm->mmap_sem);
 }
 
 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
@@ -3477,7 +3740,18 @@ static void mem_cgroup_move_task(struct cgroup_subsys *ss,
                                struct task_struct *p,
                                bool threadgroup)
 {
-       mem_cgroup_move_charge();
+       struct mm_struct *mm;
+
+       if (!mc.to)
+               /* no need to move charge */
+               return;
+
+       mm = get_task_mm(p);
+       if (mm) {
+               mem_cgroup_move_charge(mm);
+               mmput(mm);
+       }
+       mem_cgroup_clear_mc();
 }
 
 struct cgroup_subsys mem_cgroup_subsys = {