Merge tag 'for-4.15/dm' of git://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm

Pull device mapper updates from Mike Snitzer:

 - a few conversions from atomic_t to ref_count_t

 - a DM core fix for a race during device destruction that could result
   in a BUG_ON

 - a stable@ fix for a DM cache race condition that could lead to data
   corruption when operating in writeback mode (writethrough is default)

 - various DM cache cleanups and improvements

 - add DAX support to the DM log-writes target

 - a fix for the DM zoned target's ability to deal with the last zone of
   the drive being smaller than all others

 - a stable@ DM crypt and DM integrity fix for a negative check that was
   to restrictive (prevented slab debug with XFS ontop of DM crypt from
   working)

 - a DM raid target fix for a panic that can occur when forcing a raid
   to sync

* tag 'for-4.15/dm' of git://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm: (25 commits)
  dm cache: lift common migration preparation code to alloc_migration()
  dm cache: remove usused deferred_cells member from struct cache
  dm cache policy smq: allocate cache blocks in order
  dm cache policy smq: change max background work from 10240 to 4096 blocks
  dm cache background tracker: limit amount of background work that may be issued at once
  dm cache policy smq: take origin idle status into account when queuing writebacks
  dm cache policy smq: handle races with queuing background_work
  dm raid: fix panic when attempting to force a raid to sync
  dm integrity: allow unaligned bv_offset
  dm crypt: allow unaligned bv_offset
  dm: small cleanup in dm_get_md()
  dm: fix race between dm_get_from_kobject() and __dm_destroy()
  dm: allocate struct mapped_device with kvzalloc
  dm zoned: ignore last smaller runt zone
  dm space map metadata: use ARRAY_SIZE
  dm log writes: add support for DAX
  dm log writes: add support for inline data buffers
  dm cache: simplify get_per_bio_data() by removing data_size argument
  dm cache: remove all obsolete writethrough-specific code
  dm cache: submit writethrough writes in parallel to origin and cache
  ...

1 files changed, 33 insertions, 9 deletions

diff --git a/drivers/md/dm-cache-policy-smq.c b/drivers/md/dm-cache-policy-smq.c
index e5eb9c9b4bc8..4ab23d0075f6 100644
--- a/drivers/md/dm-cache-policy-smq.c
+++ b/drivers/md/dm-cache-policy-smq.c
@@ -213,6 +213,19 @@ static void l_del(struct entry_space *es, struct ilist *l, struct entry *e)
 		l->nr_elts--;
 }
 
+static struct entry *l_pop_head(struct entry_space *es, struct ilist *l)
+{
+	struct entry *e;
+
+	for (e = l_head(es, l); e; e = l_next(es, e))
+		if (!e->sentinel) {
+			l_del(es, l, e);
+			return e;
+		}
+
+	return NULL;
+}
+
 static struct entry *l_pop_tail(struct entry_space *es, struct ilist *l)
 {
 	struct entry *e;
@@ -719,7 +732,7 @@ static struct entry *alloc_entry(struct entry_alloc *ea)
 	if (l_empty(&ea->free))
 		return NULL;
 
-	e = l_pop_tail(ea->es, &ea->free);
+	e = l_pop_head(ea->es, &ea->free);
 	init_entry(e);
 	ea->nr_allocated++;
 
@@ -1158,13 +1171,13 @@ static void clear_pending(struct smq_policy *mq, struct entry *e)
 	e->pending_work = false;
 }
 
-static void queue_writeback(struct smq_policy *mq)
+static void queue_writeback(struct smq_policy *mq, bool idle)
 {
 	int r;
 	struct policy_work work;
 	struct entry *e;
 
-	e = q_peek(&mq->dirty, mq->dirty.nr_levels, !mq->migrations_allowed);
+	e = q_peek(&mq->dirty, mq->dirty.nr_levels, idle);
 	if (e) {
 		mark_pending(mq, e);
 		q_del(&mq->dirty, e);
@@ -1174,12 +1187,16 @@ static void queue_writeback(struct smq_policy *mq)
 		work.cblock = infer_cblock(mq, e);
 
 		r = btracker_queue(mq->bg_work, &work, NULL);
-		WARN_ON_ONCE(r); // FIXME: finish, I think we have to get rid of this race.
+		if (r) {
+			clear_pending(mq, e);
+			q_push_front(&mq->dirty, e);
+		}
 	}
 }
 
 static void queue_demotion(struct smq_policy *mq)
 {
+	int r;
 	struct policy_work work;
 	struct entry *e;
 
@@ -1189,7 +1206,7 @@ static void queue_demotion(struct smq_policy *mq)
 	e = q_peek(&mq->clean, mq->clean.nr_levels / 2, true);
 	if (!e) {
 		if (!clean_target_met(mq, true))
-			queue_writeback(mq);
+			queue_writeback(mq, false);
 		return;
 	}
 
@@ -1199,12 +1216,17 @@ static void queue_demotion(struct smq_policy *mq)
 	work.op = POLICY_DEMOTE;
 	work.oblock = e->oblock;
 	work.cblock = infer_cblock(mq, e);
-	btracker_queue(mq->bg_work, &work, NULL);
+	r = btracker_queue(mq->bg_work, &work, NULL);
+	if (r) {
+		clear_pending(mq, e);
+		q_push_front(&mq->clean, e);
+	}
 }
 
 static void queue_promotion(struct smq_policy *mq, dm_oblock_t oblock,
 			    struct policy_work **workp)
 {
+	int r;
 	struct entry *e;
 	struct policy_work work;
 
@@ -1234,7 +1256,9 @@ static void queue_promotion(struct smq_policy *mq, dm_oblock_t oblock,
 	work.op = POLICY_PROMOTE;
 	work.oblock = oblock;
 	work.cblock = infer_cblock(mq, e);
-	btracker_queue(mq->bg_work, &work, workp);
+	r = btracker_queue(mq->bg_work, &work, workp);
+	if (r)
+		free_entry(&mq->cache_alloc, e);
 }
 
 /*----------------------------------------------------------------*/
@@ -1418,7 +1442,7 @@ static int smq_get_background_work(struct dm_cache_policy *p, bool idle,
 	r = btracker_issue(mq->bg_work, result);
 	if (r == -ENODATA) {
 		if (!clean_target_met(mq, idle)) {
-			queue_writeback(mq);
+			queue_writeback(mq, idle);
 			r = btracker_issue(mq->bg_work, result);
 		}
 	}
@@ -1778,7 +1802,7 @@ static struct dm_cache_policy *__smq_create(dm_cblock_t cache_size,
 	mq->next_hotspot_period = jiffies;
 	mq->next_cache_period = jiffies;
 
-	mq->bg_work = btracker_create(10240); /* FIXME: hard coded value */
+	mq->bg_work = btracker_create(4096); /* FIXME: hard coded value */
 	if (!mq->bg_work)
 		goto bad_btracker;