dm table: set flush capability based on underlying devices

DM has always advertised both REQ_FLUSH and REQ_FUA flush capabilities
regardless of whether or not a given DM device's underlying devices
also advertised a need for them.

Block's flush-merge changes from 2.6.39 have proven to be more costly
for DM devices.  Performance regressions have been reported even when
DM's underlying devices do not advertise that they have a write cache.

Fix the performance regressions by configuring a DM device's flushing
capabilities based on those of the underlying devices' capabilities.

Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>

2 files changed, 43 insertions, 1 deletions

diff --git a/drivers/md/dm-table.c b/drivers/md/dm-table.c
index 259ce99302fc..986b8754bb08 100644
--- a/drivers/md/dm-table.c
+++ b/drivers/md/dm-table.c
@@ -1248,9 +1248,45 @@ static void dm_table_set_integrity(struct dm_table *t)
                                blk_get_integrity(template_disk));
 }
 
+static int device_flush_capable(struct dm_target *ti, struct dm_dev *dev,
+                                sector_t start, sector_t len, void *data)
+{
+        unsigned flush = (*(unsigned *)data);
+        struct request_queue *q = bdev_get_queue(dev->bdev);
+
+        return q && (q->flush_flags & flush);
+}
+
+static bool dm_table_supports_flush(struct dm_table *t, unsigned flush)
+{
+        struct dm_target *ti;
+        unsigned i = 0;
+
+        /*
+         * Require at least one underlying device to support flushes.
+         * t->devices includes internal dm devices such as mirror logs
+         * so we need to use iterate_devices here, which targets
+         * supporting flushes must provide.
+         */
+        while (i < dm_table_get_num_targets(t)) {
+                ti = dm_table_get_target(t, i++);
+
+                if (!ti->num_flush_requests)
+                        continue;
+
+                if (ti->type->iterate_devices &&
+                    ti->type->iterate_devices(ti, device_flush_capable, &flush))
+                        return 1;
+        }
+
+        return 0;
+}
+
 void dm_table_set_restrictions(struct dm_table *t, struct request_queue *q,
                                struct queue_limits *limits)
 {
+        unsigned flush = 0;
+
         /*
          * Copy table's limits to the DM device's request_queue
          */
@@ -1261,6 +1297,13 @@ void dm_table_set_restrictions(struct dm_table *t, struct request_queue *q,
         else
                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
 
+        if (dm_table_supports_flush(t, REQ_FLUSH)) {
+                flush |= REQ_FLUSH;
+                if (dm_table_supports_flush(t, REQ_FUA))
+                        flush |= REQ_FUA;
+        }
+        blk_queue_flush(q, flush);
+
         dm_table_set_integrity(t);
 
         /*
diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index 1000eaf984ef..52b39f335bb3 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c
@@ -1808,7 +1808,6 @@ static void dm_init_md_queue(struct mapped_device *md)
         blk_queue_make_request(md->queue, dm_request);
         blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
         blk_queue_merge_bvec(md->queue, dm_merge_bvec);
-        blk_queue_flush(md->queue, REQ_FLUSH | REQ_FUA);
 }
 
 /*