Merge tag 'dm-3.9-changes' of git://git.kernel.org/pub/scm/linux/kernel/git/agk/linux-dm

Pull device-mapper update from Alasdair G Kergon:
 "The main addition here is a long-desired target framework to allow an
  SSD to be used as a cache in front of a slower device.  Cache tuning
  is delegated to interchangeable policy modules so these can be
  developed independently of the mechanics needed to shuffle the data
  around.

  Other than that, kcopyd users acquire a throttling parameter, ioctl
  buffer usage gets streamlined, more mempool reliance is reduced and
  there are a few other bug fixes and tidy-ups."

* tag 'dm-3.9-changes' of git://git.kernel.org/pub/scm/linux/kernel/git/agk/linux-dm: (30 commits)
  dm cache: add cleaner policy
  dm cache: add mq policy
  dm: add cache target
  dm persistent data: add bitset
  dm persistent data: add transactional array
  dm thin: remove cells from stack
  dm bio prison: pass cell memory in
  dm persistent data: add btree_walk
  dm: add target num_write_bios fn
  dm kcopyd: introduce configurable throttling
  dm ioctl: allow message to return data
  dm ioctl: optimize functions without variable params
  dm ioctl: introduce ioctl_flags
  dm: merge io_pool and tio_pool
  dm: remove unused _rq_bio_info_cache
  dm: fix limits initialization when there are no data devices
  dm snapshot: add missing module aliases
  dm persistent data: set some btree fn parms const
  dm: refactor bio cloning
  dm: rename bio cloning functions
  ...

1 files changed, 120 insertions, 1 deletions

diff --git a/drivers/md/dm-kcopyd.c b/drivers/md/dm-kcopyd.c
index 68c02673263..d581fe5d2fa 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)