Re: [PATCH v3 20/21] mm: split zone->lru_lock

From: Konstantin Khlebnikov
Date: Tue Feb 28 2012 - 01:39:48 EST

KAMEZAWA Hiroyuki wrote:
On Thu, 23 Feb 2012 17:53:23 +0400
Konstantin Khlebnikov<khlebnikov@xxxxxxxxxx> wrote:

Looks like all ready for splitting zone->lru_lock into per-lruvec pieces.

lruvec locking loop protected with rcu, actually there is irq-disabling instead
of rcu_read_lock(). Memory controller already releases its lru-vectors after
syncronize_rcu() in cgroup_diput(). Probably it should be replaced with synchronize_sched()

Signed-off-by: Konstantin Khlebnikov<khlebnikov@xxxxxxxxxx>

Do we need rcu_read_lock() even if we check isolated pages at pre_destroy() ?
If pre_destroy() ends, pages under a memcg being destroyed were moved to other
cgroup while it's usolated.

So,
- PageLRU(page) guarantees lruvec is valid.
- if !PageLRU(page), the caller of lru_lock should know what it does.
Once isolated, pre_destroy() never ends and page_lruvec(page) is always stable.

This will be racy, because lruvec can be changed and released between checking PageLRU() and page_lruvec() or
between page_lruvec() and spin_lock(), so we need to protect whole operation with single locking interval.


Thanks,
-Kame

---
include/linux/mmzone.h | 3 +-
mm/compaction.c | 2 +
mm/internal.h | 66 +++++++++++++++++++++++++-----------------------
mm/page_alloc.c | 2 +
mm/swap.c | 2 +
5 files changed, 40 insertions(+), 35 deletions(-)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 2e3a298..9880150 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -304,6 +304,8 @@ struct zone_reclaim_stat {
};

struct lruvec {
+ spinlock_t lru_lock;
+
struct list_head pages_lru[NR_LRU_LISTS];
unsigned long pages_count[NR_LRU_COUNTERS];

@@ -386,7 +388,6 @@ struct zone {
ZONE_PADDING(_pad1_)

/* Fields commonly accessed by the page reclaim scanner */
- spinlock_t lru_lock;
struct lruvec lruvec;

unsigned long pages_scanned; /* since last reclaim */
diff --git a/mm/compaction.c b/mm/compaction.c
index fa74cbe..8661bb58 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -306,7 +306,7 @@ static isolate_migrate_t isolate_migratepages(struct zone *zone,
lruvec = NULL;
}
if (need_resched() ||
- (lruvec&& spin_is_contended(&zone->lru_lock))) {
+ (lruvec&& spin_is_contended(&lruvec->lru_lock))) {
if (lruvec)
unlock_lruvec_irq(lruvec);
lruvec = NULL;
diff --git a/mm/internal.h b/mm/internal.h
index 6dd2e70..9a9fd53 100644
--- a/mm/internal.h
+++ b/mm/internal.h
@@ -15,27 +15,27 @@

static inline void lock_lruvec(struct lruvec *lruvec, unsigned long *flags)
{
- spin_lock_irqsave(&lruvec_zone(lruvec)->lru_lock, *flags);
+ spin_lock_irqsave(&lruvec->lru_lock, *flags);
}

static inline void lock_lruvec_irq(struct lruvec *lruvec)
{
- spin_lock_irq(&lruvec_zone(lruvec)->lru_lock);
+ spin_lock_irq(&lruvec->lru_lock);
}

static inline void unlock_lruvec(struct lruvec *lruvec, unsigned long *flags)
{
- spin_unlock_irqrestore(&lruvec_zone(lruvec)->lru_lock, *flags);
+ spin_unlock_irqrestore(&lruvec->lru_lock, *flags);
}

static inline void unlock_lruvec_irq(struct lruvec *lruvec)
{
- spin_unlock_irq(&lruvec_zone(lruvec)->lru_lock);
+ spin_unlock_irq(&lruvec->lru_lock);
}

static inline void wait_lruvec_unlock(struct lruvec *lruvec)
{
- spin_unlock_wait(&lruvec_zone(lruvec)->lru_lock);
+ spin_unlock_wait(&lruvec->lru_lock);
}

#ifdef CONFIG_CGROUP_MEM_RES_CTLR
@@ -46,37 +46,39 @@ static inline void wait_lruvec_unlock(struct lruvec *lruvec)
static inline struct lruvec *__relock_page_lruvec(struct lruvec *locked_lruvec,
struct page *page)
{
- /* Currenyly only one lru_lock per-zone */
- return page_lruvec(page);
+ struct lruvec *lruvec;
+
+ do {
+ lruvec = page_lruvec(page);
+ if (likely(lruvec == locked_lruvec))
+ return lruvec;
+ spin_unlock(&locked_lruvec->lru_lock);
+ spin_lock(&lruvec->lru_lock);
+ locked_lruvec = lruvec;
+ } while (1);
}

static inline struct lruvec *relock_page_lruvec_irq(struct lruvec *lruvec,
struct page *page)
{
- struct zone *zone = page_zone(page);
-
if (!lruvec) {
- spin_lock_irq(&zone->lru_lock);
- } else if (zone != lruvec_zone(lruvec)) {
- unlock_lruvec_irq(lruvec);
- spin_lock_irq(&zone->lru_lock);
+ local_irq_disable();
+ lruvec = page_lruvec(page);
+ spin_lock(&lruvec->lru_lock);
}
- return page_lruvec(page);
+ return __relock_page_lruvec(lruvec, page);
}

static inline struct lruvec *relock_page_lruvec(struct lruvec *lruvec,
struct page *page,
unsigned long *flags)
{
- struct zone *zone = page_zone(page);
-
if (!lruvec) {
- spin_lock_irqsave(&zone->lru_lock, *flags);
- } else if (zone != lruvec_zone(lruvec)) {
- unlock_lruvec(lruvec, flags);
- spin_lock_irqsave(&zone->lru_lock, *flags);
+ local_irq_save(*flags);
+ lruvec = page_lruvec(page);
+ spin_lock(&lruvec->lru_lock);
}
- return page_lruvec(page);
+ return __relock_page_lruvec(lruvec, page);
}

/*
@@ -87,22 +89,24 @@ static inline struct lruvec *relock_page_lruvec(struct lruvec *lruvec,
static inline bool __lock_page_lruvec_irq(struct lruvec **lruvec,
struct page *page)
{
- struct zone *zone;
bool ret = false;

+ rcu_read_lock();
+ /*
+ * If we see there PageLRU(), it means page has valid lruvec link.
+ * We need protect whole operation with single rcu-interval, otherwise
+ * lruvec which hold this LRU sign can run out before we secure it.
+ */
if (PageLRU(page)) {
if (!*lruvec) {
- zone = page_zone(page);
- spin_lock_irq(&zone->lru_lock);
- } else
- zone = lruvec_zone(*lruvec);
-
- if (PageLRU(page)) {
*lruvec = page_lruvec(page);
+ lock_lruvec_irq(*lruvec);
+ }
+ *lruvec = __relock_page_lruvec(*lruvec, page);
+ if (PageLRU(page))
ret = true;
- } else
- *lruvec =&zone->lruvec;
}
+ rcu_read_unlock();

return ret;
}
@@ -110,7 +114,7 @@ static inline bool __lock_page_lruvec_irq(struct lruvec **lruvec,
/* Wait for lruvec unlock before locking other lruvec for the same page */
static inline void __wait_lruvec_unlock(struct lruvec *lruvec)
{
- /* Currently only one lru_lock per-zone */
+ wait_lruvec_unlock(lruvec);
}

#else /* CONFIG_CGROUP_MEM_RES_CTLR */
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index ab42446..beadcc9 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -4294,6 +4294,7 @@ void init_zone_lruvec(struct zone *zone, struct lruvec *lruvec)
enum lru_list lru;

memset(lruvec, 0, sizeof(struct lruvec));
+ spin_lock_init(&lruvec->lru_lock);
for_each_lru(lru)
INIT_LIST_HEAD(&lruvec->pages_lru[lru]);
#ifdef CONFIG_CGROUP_MEM_RES_CTLR
@@ -4369,7 +4370,6 @@ static void __paginginit free_area_init_core(struct pglist_data *pgdat,
#endif
zone->name = zone_names[j];
spin_lock_init(&zone->lock);
- spin_lock_init(&zone->lru_lock);
zone_seqlock_init(zone);
zone->zone_pgdat = pgdat;

diff --git a/mm/swap.c b/mm/swap.c
index 998c71c..8156181 100644
--- a/mm/swap.c
+++ b/mm/swap.c
@@ -700,7 +700,7 @@ void lru_add_page_tail(struct lruvec *lruvec,
VM_BUG_ON(!PageHead(page));
VM_BUG_ON(PageCompound(page_tail));
VM_BUG_ON(PageLRU(page_tail));
- VM_BUG_ON(NR_CPUS != 1&& !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
+ VM_BUG_ON(NR_CPUS != 1&& !spin_is_locked(&lruvec->lru_lock));

SetPageLRU(page_tail);





--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/