> > > - anonymous objects (shmem) are not accounted.
> > IMHO, shmem should be accounted.
> > I agree it's difficult in your implementation,
> > but are you going to support it?
> 
> it should be trivial to track how much swap an anonymous object is using.
> i'm not sure how it should be associated with cgroups, tho.
> 
> YAMAMOTO Takashi

i implemented shmem swap accounting.  see below.

YAMAMOTO Takashi


the following is another swap controller, which was designed and
implemented independently from nishimura-san's one.

some random differences from nishimura-san's one:
- counts and limits the number of ptes with swap entries instead of
  on-disk swap slots.  (except swap slots used by anonymous objects,
  which are counted differently.  see below.)
- no swapon-time memory allocation.
- precise wrt moving tasks between cgroups.
- [NEW] anonymous objects (shmem) are associated to a cgroup when they are
  created and the number of on-disk swap slots used by them are counted for
  the cgroup.  the association is persistent except cgroup removal,
  in which case, its associated objects are moved to init_mm's cgroup.


Signed-off-by: YAMAMOTO Takashi <yamamoto@valinux.co.jp>
---

--- linux-2.6.25-rc8-mm2/mm/swapcontrol.c.BACKUP	2008-04-21 12:06:44.000000000 +0900
+++ linux-2.6.25-rc8-mm2/mm/swapcontrol.c	2008-04-28 14:16:03.000000000 +0900
@@ -0,0 +1,447 @@
+
+/*
+ * swapcontrol.c COPYRIGHT FUJITSU LIMITED 2008
+ *
+ * Author: yamamoto@valinux.co.jp
+ */
+
+#include <linux/err.h>
+#include <linux/cgroup.h>
+#include <linux/hugetlb.h>
+#include <linux/res_counter.h>
+#include <linux/pagemap.h>
+#include <linux/slab.h>
+#include <linux/swap.h>
+#include <linux/swapcontrol.h>
+#include <linux/swapops.h>
+#include <linux/shmem_fs.h>
+
+struct swap_cgroup {
+	struct cgroup_subsys_state scg_css;
+	struct res_counter scg_counter;
+	struct list_head scg_shmem_list;
+};
+
+#define	task_to_css(task) task_subsys_state((task), swap_cgroup_subsys_id)
+#define	css_to_scg(css)	container_of((css), struct swap_cgroup, scg_css)
+#define	cg_to_css(cg)	cgroup_subsys_state((cg), swap_cgroup_subsys_id)
+#define	cg_to_scg(cg)	css_to_scg(cg_to_css(cg))
+
+static DEFINE_MUTEX(swap_cgroup_shmem_lock);
+
+static struct swap_cgroup *
+swap_cgroup_prepare_ptp(struct page *ptp, struct mm_struct *mm)
+{
+	struct swap_cgroup *scg = ptp->ptp_swap_cgroup;
+
+	BUG_ON(mm == NULL);
+	BUG_ON(mm->swap_cgroup == NULL);
+	if (scg == NULL) {
+		/*
+		 * see swap_cgroup_attach.
+		 */
+		smp_rmb();
+		scg = mm->swap_cgroup;
+		BUG_ON(scg == NULL);
+		ptp->ptp_swap_cgroup = scg;
+	}
+
+	return scg;
+}
+
+/*
+ * called with page table locked.
+ */
+int
+swap_cgroup_charge(struct page *ptp, struct mm_struct *mm)
+{
+	struct swap_cgroup * const scg = swap_cgroup_prepare_ptp(ptp, mm);
+
+	return res_counter_charge(&scg->scg_counter, PAGE_CACHE_SIZE);
+}
+
+/*
+ * called with page table locked.
+ */
+void
+swap_cgroup_uncharge(struct page *ptp)
+{
+	struct swap_cgroup * const scg = ptp->ptp_swap_cgroup;
+
+	BUG_ON(scg == NULL);
+	res_counter_uncharge(&scg->scg_counter, PAGE_CACHE_SIZE);
+}
+
+/*
+ * called with both page tables locked.
+ */
+void
+swap_cgroup_remap_charge(struct page *oldptp, struct page *newptp,
+    struct mm_struct *mm)
+{
+	struct swap_cgroup * const oldscg = oldptp->ptp_swap_cgroup;
+	struct swap_cgroup * const newscg = swap_cgroup_prepare_ptp(newptp, mm);
+
+	BUG_ON(oldscg == NULL);
+	BUG_ON(newscg == NULL);
+	if (oldscg == newscg) {
+		return;
+	}
+
+	/*
+	 * swap_cgroup_attach is in progress.
+	 */
+
+	res_counter_charge_force(&newscg->scg_counter, PAGE_CACHE_SIZE);
+	res_counter_uncharge(&oldscg->scg_counter, PAGE_CACHE_SIZE);
+}
+
+void
+swap_cgroup_init_mm(struct mm_struct *mm, struct task_struct *t)
+{
+	struct swap_cgroup *scg;
+	struct cgroup *cg;
+
+	/* mm->swap_cgroup is not NULL in the case of dup_mm */
+	cg = task_cgroup(t, swap_cgroup_subsys_id);
+	BUG_ON(cg == NULL);
+	scg = cg_to_scg(cg);
+	BUG_ON(scg == NULL);
+	css_get(&scg->scg_css);
+	mm->swap_cgroup = scg;
+}
+
+void
+swap_cgroup_exit_mm(struct mm_struct *mm)
+{
+	struct swap_cgroup * const scg = mm->swap_cgroup;
+
+	/*
+	 * note that swap_cgroup_attach does get_task_mm which
+	 * prevents us from being called.  we can assume mm->swap_cgroup
+	 * is stable.
+	 */
+
+	BUG_ON(scg == NULL);
+	mm->swap_cgroup = NULL; /* it isn't necessary.  just being tidy. */
+	css_put(&scg->scg_css);
+}
+
+static u64
+swap_cgroup_read_u64(struct cgroup *cg, struct cftype *cft)
+{
+
+	return res_counter_read_u64(&cg_to_scg(cg)->scg_counter, cft->private);
+}
+
+static int
+swap_cgroup_write_u64(struct cgroup *cg, struct cftype *cft, u64 val)
+{
+	struct res_counter *counter = &cg_to_scg(cg)->scg_counter;
+	unsigned long flags;
+
+	/* XXX res_counter_write_u64 */
+	BUG_ON(cft->private != RES_LIMIT);
+	spin_lock_irqsave(&counter->lock, flags);
+	counter->limit = val;
+	spin_unlock_irqrestore(&counter->lock, flags);
+	return 0;
+}
+
+static const struct cftype files[] = {
+	{
+		.name = "usage_in_bytes",
+		.private = RES_USAGE,
+		.read_u64 = swap_cgroup_read_u64,
+	},
+	{
+		.name = "failcnt",
+		.private = RES_FAILCNT,
+		.read_u64 = swap_cgroup_read_u64,
+	},
+	{
+		.name = "limit_in_bytes",
+		.private = RES_LIMIT,
+		.read_u64 = swap_cgroup_read_u64,
+		.write_u64 = swap_cgroup_write_u64,
+	},
+};
+
+static struct cgroup_subsys_state *
+swap_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cg)
+{
+	struct swap_cgroup *scg;
+
+	scg = kzalloc(sizeof(*scg), GFP_KERNEL);
+	if (scg == NULL) {
+		return ERR_PTR(-ENOMEM);
+	}
+	res_counter_init(&scg->scg_counter);
+	if (unlikely(cg->parent == NULL)) {
+		init_mm.swap_cgroup = scg;
+	}
+	INIT_LIST_HEAD(&scg->scg_shmem_list);
+	return &scg->scg_css;
+}
+
+static void
+swap_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cg)
+{
+	struct swap_cgroup *oldscg = cg_to_scg(cg);
+	struct swap_cgroup *newscg;
+	struct list_head *pos;
+	struct list_head *next;
+
+	/*
+	 * move our anonymous objects to init_mm's group.
+	 */
+	newscg = init_mm.swap_cgroup;
+	BUG_ON(newscg == NULL);
+	BUG_ON(newscg == oldscg);
+
+	mutex_lock(&swap_cgroup_shmem_lock);
+	list_for_each_safe(pos, next, &oldscg->scg_shmem_list) {
+		struct shmem_inode_info * const info =
+		    list_entry(pos, struct shmem_inode_info, swap_cgroup_list);
+		long bytes;
+
+		spin_lock(&info->lock);
+		BUG_ON(info->swap_cgroup != oldscg);
+		bytes = info->swapped << PAGE_SHIFT;
+		info->swap_cgroup = newscg;
+		res_counter_uncharge(&oldscg->scg_counter, bytes);
+		res_counter_charge_force(&newscg->scg_counter, bytes);
+		spin_unlock(&info->lock);
+		list_del_init(&info->swap_cgroup_list);
+		list_add(&info->swap_cgroup_list, &newscg->scg_shmem_list);
+	}
+	mutex_unlock(&swap_cgroup_shmem_lock);
+
+	BUG_ON(res_counter_read_u64(&oldscg->scg_counter, RES_USAGE) != 0);
+	kfree(oldscg);
+}
+
+static int
+swap_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cg)
+{
+
+	return cgroup_add_files(cg, ss, files, ARRAY_SIZE(files));
+}
+
+struct swap_cgroup_attach_mm_cb_args {
+	struct vm_area_struct *vma;
+	struct swap_cgroup *oldscg;
+	struct swap_cgroup *newscg;
+};
+
+static int
+swap_cgroup_attach_mm_cb(pmd_t *pmd, unsigned long startva, unsigned long endva,
+    void *private)
+{
+	const struct swap_cgroup_attach_mm_cb_args * const args = private;
+	struct vm_area_struct * const vma = args->vma;
+	struct swap_cgroup * const oldscg = args->oldscg;
+	struct swap_cgroup * const newscg = args->newscg;
+	struct page *ptp;
+	spinlock_t *ptl;
+	const pte_t *startpte;
+	const pte_t *pte;
+	unsigned long va;
+	int swslots;
+	int bytes;
+
+	BUG_ON((startva & ~PMD_MASK) != 0);
+	BUG_ON((endva & ~PMD_MASK) != 0);
+
+	startpte = pte_offset_map_lock(vma->vm_mm, pmd, startva, &ptl);
+	ptp = pmd_page(*pmd);
+	if (ptp->ptp_swap_cgroup == NULL || ptp->ptp_swap_cgroup == newscg) {
+		goto out;
+	}
+	BUG_ON(ptp->ptp_swap_cgroup != oldscg);
+
+	/*
+	 * count the number of swap entries in this page table page.
+	 */
+	swslots = 0;
+	for (va = startva, pte = startpte; va != endva;
+	    pte++, va += PAGE_SIZE) {
+		const pte_t pt_entry = *pte;
+
+		if (pte_present(pt_entry)) {
+			continue;
+		}
+		if (pte_none(pt_entry)) {
+			continue;
+		}
+		if (pte_file(pt_entry)) {
+			continue;
+		}
+		if (is_migration_entry(pte_to_swp_entry(pt_entry))) {
+			continue;
+		}
+		swslots++;
+	}
+
+	bytes = swslots * PAGE_CACHE_SIZE;
+	res_counter_uncharge(&oldscg->scg_counter, bytes);
+	/*
+	 * XXX ignore newscg's limit because cgroup ->attach method can't fail.
+	 */
+	res_counter_charge_force(&newscg->scg_counter, bytes);
+	ptp->ptp_swap_cgroup = newscg;
+out:
+	pte_unmap_unlock(startpte, ptl);
+
+	return 0;
+}
+
+static const struct mm_walk swap_cgroup_attach_mm_walk = {
+	.pmd_entry = swap_cgroup_attach_mm_cb,
+};
+
+static void
+swap_cgroup_attach_mm(struct mm_struct *mm, struct swap_cgroup *oldscg,
+    struct swap_cgroup *newscg)
+{
+	struct swap_cgroup_attach_mm_cb_args args;
+	struct vm_area_struct *vma;
+
+	args.oldscg = oldscg;
+	args.newscg = newscg;
+	down_read(&mm->mmap_sem);
+	for (vma = mm->mmap; vma; vma = vma->vm_next) {
+		if (is_vm_hugetlb_page(vma)) {
+			continue;
+		}
+		args.vma = vma;
+		walk_page_range(mm, vma->vm_start & PMD_MASK,
+		    (vma->vm_end + PMD_SIZE - 1) & PMD_MASK,
+		    &swap_cgroup_attach_mm_walk, &args);
+	}
+	up_read(&mm->mmap_sem);
+}
+
+static void
+swap_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *newcg,
+    struct cgroup *oldcg, struct task_struct *t)
+{
+	struct swap_cgroup *oldmmscg;
+	struct swap_cgroup *ol