dm kcopyd: introduce configurable throttling

This patch allows the administrator to reduce the rate at which kcopyd
issues I/O.

Each module that uses kcopyd acquires a throttle parameter that can be
set in /sys/module/*/parameters.

We maintain a history of kcopyd usage by each module in the variables
io_period and total_period in struct dm_kcopyd_throttle. The actual
kcopyd activity is calculated as a percentage of time equal to
"(100 * io_period / total_period)".  This is compared with the user-defined
throttle percentage threshold and if it is exceeded, we sleep.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>

4 files changed, 132 insertions, 4 deletions

diff --git a/drivers/md/dm-kcopyd.c b/drivers/md/dm-kcopyd.c
index 68c0267..d581fe5 100644
--- a/drivers/md/dm-kcopyd.c
+++ b/drivers/md/dm-kcopyd.c
@@ -22,6 +22,7 @@
 #include <linux/vmalloc.h>
 #include <linux/workqueue.h>
 #include <linux/mutex.h>
+#include <linux/delay.h>
 #include <linux/device-mapper.h>
 #include <linux/dm-kcopyd.h>
 
@@ -51,6 +52,8 @@ struct dm_kcopyd_client {
 	struct workqueue_struct *kcopyd_wq;
 	struct work_struct kcopyd_work;
 
+	struct dm_kcopyd_throttle *throttle;
+
 /*
  * We maintain three lists of jobs:
  *
@@ -68,6 +71,117 @@ struct dm_kcopyd_client {
 
 static struct page_list zero_page_list;
 
+static DEFINE_SPINLOCK(throttle_spinlock);
+
+/*
+ * IO/IDLE accounting slowly decays after (1 << ACCOUNT_INTERVAL_SHIFT) period.
+ * When total_period >= (1 << ACCOUNT_INTERVAL_SHIFT) the counters are divided
+ * by 2.
+ */
+#define ACCOUNT_INTERVAL_SHIFT		SHIFT_HZ
+
+/*
+ * Sleep this number of milliseconds.
+ *
+ * The value was decided experimentally.
+ * Smaller values seem to cause an increased copy rate above the limit.
+ * The reason for this is unknown but possibly due to jiffies rounding errors
+ * or read/write cache inside the disk.
+ */
+#define SLEEP_MSEC			100
+
+/*
+ * Maximum number of sleep events. There is a theoretical livelock if more
+ * kcopyd clients do work simultaneously which this limit avoids.
+ */
+#define MAX_SLEEPS			10
+
+static void io_job_start(struct dm_kcopyd_throttle *t)
+{
+	unsigned throttle, now, difference;
+	int slept = 0, skew;
+
+	if (unlikely(!t))
+		return;
+
+try_again:
+	spin_lock_irq(&throttle_spinlock);
+
+	throttle = ACCESS_ONCE(t->throttle);
+
+	if (likely(throttle >= 100))
+		goto skip_limit;
+
+	now = jiffies;
+	difference = now - t->last_jiffies;
+	t->last_jiffies = now;
+	if (t->num_io_jobs)
+		t->io_period += difference;
+	t->total_period += difference;
+
+	/*
+	 * Maintain sane values if we got a temporary overflow.
+	 */
+	if (unlikely(t->io_period > t->total_period))
+		t->io_period = t->total_period;
+
+	if (unlikely(t->total_period >= (1 << ACCOUNT_INTERVAL_SHIFT))) {
+		int shift = fls(t->total_period >> ACCOUNT_INTERVAL_SHIFT);
+		t->total_period >>= shift;
+		t->io_period >>= shift;
+	}
+
+	skew = t->io_period - throttle * t->total_period / 100;
+
+	if (unlikely(skew > 0) && slept < MAX_SLEEPS) {
+		slept++;
+		spin_unlock_irq(&throttle_spinlock);
+		msleep(SLEEP_MSEC);
+		goto try_again;
+	}
+
+skip_limit:
+	t->num_io_jobs++;
+
+	spin_unlock_irq(&throttle_spinlock);
+}
+
+static void io_job_finish(struct dm_kcopyd_throttle *t)
+{
+	unsigned long flags;
+
+	if (unlikely(!t))
+		return;
+
+	spin_lock_irqsave(&throttle_spinlock, flags);
+
+	t->num_io_jobs--;
+
+	if (likely(ACCESS_ONCE(t->throttle) >= 100))
+		goto skip_limit;
+
+	if (!t->num_io_jobs) {
+		unsigned now, difference;
+
+		now = jiffies;
+		difference = now - t->last_jiffies;
+		t->last_jiffies = now;
+
+		t->io_period += difference;
+		t->total_period += difference;
+
+		/*
+		 * Maintain sane values if we got a temporary overflow.
+		 */
+		if (unlikely(t->io_period > t->total_period))
+			t->io_period = t->total_period;
+	}
+
+skip_limit:
+	spin_unlock_irqrestore(&throttle_spinlock, flags);
+}
+
+
 static void wake(struct dm_kcopyd_client *kc)
 {
 	queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
@@ -348,6 +462,8 @@ static void complete_io(unsigned long error, void *context)
 	struct kcopyd_job *job = (struct kcopyd_job *) context;
 	struct dm_kcopyd_client *kc = job->kc;
 
+	io_job_finish(kc->throttle);
+
 	if (error) {
 		if (job->rw & WRITE)
 			job->write_err |= error;
@@ -389,6 +505,8 @@ static int run_io_job(struct kcopyd_job *job)
 		.client = job->kc->io_client,
 	};
 
+	io_job_start(job->kc->throttle);
+
 	if (job->rw == READ)
 		r = dm_io(&io_req, 1, &job->source, NULL);
 	else
@@ -695,7 +813,7 @@ int kcopyd_cancel(struct kcopyd_job *job, int block)
 /*-----------------------------------------------------------------
  * Client setup
  *---------------------------------------------------------------*/
-struct dm_kcopyd_client *dm_kcopyd_client_create(void)
+struct dm_kcopyd_client *dm_kcopyd_client_create(struct dm_kcopyd_throttle *throttle)
 {
 	int r = -ENOMEM;
 	struct dm_kcopyd_client *kc;
@@ -708,6 +826,7 @@ struct dm_kcopyd_client *dm_kcopyd_client_create(void)
 	INIT_LIST_HEAD(&kc->complete_jobs);
 	INIT_LIST_HEAD(&kc->io_jobs);
 	INIT_LIST_HEAD(&kc->pages_jobs);
+	kc->throttle = throttle;
 
 	kc->job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
 	if (!kc->job_pool)
diff --git a/drivers/md/dm-raid1.c b/drivers/md/dm-raid1.c
index e2ea977..d053098 100644
--- a/drivers/md/dm-raid1.c
+++ b/drivers/md/dm-raid1.c
@@ -82,6 +82,9 @@ struct mirror_set {
 	struct mirror mirror[0];
 };
 
+DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(raid1_resync_throttle,
+		"A percentage of time allocated for raid resynchronization");
+
 static void wakeup_mirrord(void *context)
 {
 	struct mirror_set *ms = context;
@@ -1111,7 +1114,7 @@ static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 		goto err_destroy_wq;
 	}
 
-	ms->kcopyd_client = dm_kcopyd_client_create();
+	ms->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle);
 	if (IS_ERR(ms->kcopyd_client)) {
 		r = PTR_ERR(ms->kcopyd_client);
 		goto err_destroy_wq;
diff --git a/drivers/md/dm-snap.c b/drivers/md/dm-snap.c
index 58c2b58..c0e0702 100644
--- a/drivers/md/dm-snap.c
+++ b/drivers/md/dm-snap.c
@@ -124,6 +124,9 @@ struct dm_snapshot {
 #define RUNNING_MERGE          0
 #define SHUTDOWN_MERGE         1
 
+DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
+		"A percentage of time allocated for copy on write");
+
 struct dm_dev *dm_snap_origin(struct dm_snapshot *s)
 {
 	return s->origin;
@@ -1108,7 +1111,7 @@ static int snapshot_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 		goto bad_hash_tables;
 	}
 
-	s->kcopyd_client = dm_kcopyd_client_create();
+	s->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle);
 	if (IS_ERR(s->kcopyd_client)) {
 		r = PTR_ERR(s->kcopyd_client);
 		ti->error = "Could not create kcopyd client";
diff --git a/drivers/md/dm-thin.c b/drivers/md/dm-thin.c
index 303e11d..35d9d03 100644
--- a/drivers/md/dm-thin.c
+++ b/drivers/md/dm-thin.c
@@ -26,6 +26,9 @@
 #define PRISON_CELLS 1024
 #define COMMIT_PERIOD HZ
 
+DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
+		"A percentage of time allocated for copy on write");
+
 /*
  * The block size of the device holding pool data must be
  * between 64KB and 1GB.
@@ -1642,7 +1645,7 @@ static struct pool *pool_create(struct mapped_device *pool_md,
 		goto bad_prison;
 	}
 
-	pool->copier = dm_kcopyd_client_create();
+	pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
 	if (IS_ERR(pool->copier)) {
 		r = PTR_ERR(pool->copier);
 		*error = "Error creating pool's kcopyd client";