BACKPORT: dm bufio: introduce a global cache replacementASB-2019-11-05_4.14

This commit introduces a global cache replacement (instead of per-client
cleanup).

If one bufio client uses the cache heavily and another client is not using
it, we want to let the first client use most of the cache. The old
algorithm would partition the cache equally betwen the clients and that is
sub-optimal.

For cache replacement, we use the clock algorithm because it doesn't
require taking any lock when the buffer is accessed.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
(cherry picked from commit 6e913b28cd279212c4580dbd8b2cf9dcd4740cfb)

Change-Id: Iad25b7058d3da32ae07959a348fa75178ff9c860
Signed-off-by: Martijn Coenen <maco@android.com>

1 files changed, 91 insertions, 7 deletions

diff --git a/drivers/md/dm-bufio.c b/drivers/md/dm-bufio.c
index 3fb6f973c99e..9440da8b0120 100644
--- a/drivers/md/dm-bufio.c
+++ b/drivers/md/dm-bufio.c
@@ -34,6 +34,7 @@
 #define DM_BUFIO_MEMORY_PERCENT		2
 #define DM_BUFIO_VMALLOC_PERCENT	25
 #define DM_BUFIO_WRITEBACK_RATIO	3
+#define DM_BUFIO_LOW_WATERMARK_RATIO	16
 
 /*
  * Check buffer ages in this interval (seconds)
@@ -151,6 +152,7 @@ struct dm_buffer {
 	void *data;
 	enum data_mode data_mode;
 	unsigned char list_mode;		/* LIST_* */
+	unsigned accessed;
 	unsigned hold_count;
 	blk_status_t read_error;
 	blk_status_t write_error;
@@ -227,6 +229,8 @@ static DEFINE_SPINLOCK(global_spinlock);
 
 static LIST_HEAD(global_queue);
 
+static unsigned long global_num = 0;
+
 /*
  * Buffers are freed after this timeout
  */
@@ -256,6 +260,11 @@ static LIST_HEAD(dm_bufio_all_clients);
  */
 static DEFINE_MUTEX(dm_bufio_clients_lock);
 
+static struct workqueue_struct *dm_bufio_wq;
+static struct delayed_work dm_bufio_cleanup_old_work;
+static struct work_struct dm_bufio_replacement_work;
+
+
 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 static void buffer_record_stack(struct dm_buffer *b)
 {
@@ -341,10 +350,16 @@ static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
 	if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
 		dm_bufio_peak_allocated = dm_bufio_current_allocated;
 
+	b->accessed = 1;
+
 	if (!unlink) {
 		list_add(&b->global_list, &global_queue);
+		global_num++;
+		if (dm_bufio_current_allocated > dm_bufio_cache_size)
+			queue_work(dm_bufio_wq, &dm_bufio_replacement_work);
 	} else {
 		list_del(&b->global_list);
+		global_num--;
 	}
 
 	spin_unlock(&global_spinlock);
@@ -530,6 +545,8 @@ static void __relink_lru(struct dm_buffer *b, int dirty)
 {
 	struct dm_bufio_client *c = b->c;
 
+	b->accessed = 1;
+
 	BUG_ON(!c->n_buffers[b->list_mode]);
 
 	c->n_buffers[b->list_mode]--;
@@ -1830,6 +1847,74 @@ static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
 	dm_bufio_unlock(c);
 }
 
+static void do_global_cleanup(struct work_struct *w)
+{
+	struct dm_bufio_client *locked_client = NULL;
+	struct dm_bufio_client *current_client;
+	struct dm_buffer *b;
+	unsigned spinlock_hold_count;
+	unsigned long threshold = dm_bufio_cache_size -
+		dm_bufio_cache_size / DM_BUFIO_LOW_WATERMARK_RATIO;
+	unsigned long loops = global_num * 2;
+
+	mutex_lock(&dm_bufio_clients_lock);
+
+	while (1) {
+		cond_resched();
+
+		spin_lock(&global_spinlock);
+		if (unlikely(dm_bufio_current_allocated <= threshold))
+			break;
+
+		spinlock_hold_count = 0;
+get_next:
+		if (!loops--)
+			break;
+		if (unlikely(list_empty(&global_queue)))
+			break;
+		b = list_entry(global_queue.prev, struct dm_buffer, global_list);
+
+		if (b->accessed) {
+			b->accessed = 0;
+			list_move(&b->global_list, &global_queue);
+			if (likely(++spinlock_hold_count < 16))
+				goto get_next;
+			spin_unlock(&global_spinlock);
+			continue;
+		}
+
+		current_client = b->c;
+		if (unlikely(current_client != locked_client)) {
+			if (locked_client)
+				dm_bufio_unlock(locked_client);
+
+			if (!dm_bufio_trylock(current_client)) {
+				spin_unlock(&global_spinlock);
+				dm_bufio_lock(current_client);
+				locked_client = current_client;
+				continue;
+			}
+
+			locked_client = current_client;
+		}
+
+		spin_unlock(&global_spinlock);
+
+		if (unlikely(!__try_evict_buffer(b, GFP_KERNEL))) {
+			spin_lock(&global_spinlock);
+			list_move(&b->global_list, &global_queue);
+			spin_unlock(&global_spinlock);
+		}
+	}
+
+	spin_unlock(&global_spinlock);
+
+	if (locked_client)
+		dm_bufio_unlock(locked_client);
+
+	mutex_unlock(&dm_bufio_clients_lock);
+}
+
 static void cleanup_old_buffers(void)
 {
 	unsigned long max_age_hz = get_max_age_hz();
@@ -1845,14 +1930,11 @@ static void cleanup_old_buffers(void)
 	mutex_unlock(&dm_bufio_clients_lock);
 }
 
-static struct workqueue_struct *dm_bufio_wq;
-static struct delayed_work dm_bufio_work;
-
 static void work_fn(struct work_struct *w)
 {
 	cleanup_old_buffers();
 
-	queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
+	queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
 			   DM_BUFIO_WORK_TIMER_SECS * HZ);
 }
 
@@ -1897,8 +1979,9 @@ static int __init dm_bufio_init(void)
 	if (!dm_bufio_wq)
 		return -ENOMEM;
 
-	INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
-	queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
+	INIT_DELAYED_WORK(&dm_bufio_cleanup_old_work, work_fn);
+	INIT_WORK(&dm_bufio_replacement_work, do_global_cleanup);
+	queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
 			   DM_BUFIO_WORK_TIMER_SECS * HZ);
 
 	return 0;
@@ -1912,7 +1995,8 @@ static void __exit dm_bufio_exit(void)
 	int bug = 0;
 	int i;
 
-	cancel_delayed_work_sync(&dm_bufio_work);
+	cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
+	flush_workqueue(dm_bufio_wq);
 	destroy_workqueue(dm_bufio_wq);
 
 	for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++)