Merge branch 'linus/master' into rdma.git for-next

rdma.git merge resolution for the 4.19 merge window

Conflicts:
 drivers/infiniband/core/rdma_core.c
   - Use the rdma code and revise with the new spelling for
     atomic_fetch_add_unless
 drivers/nvme/host/rdma.c
   - Replace max_sge with max_send_sge in new blk code
 drivers/nvme/target/rdma.c
   - Use the blk code and revise to use NULL for ib_post_recv when
     appropriate
   - Replace max_sge with max_recv_sge in new blk code
 net/rds/ib_send.c
   - Use the net code and revise to use NULL for ib_post_recv when
     appropriate

Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

1 files changed, 91 insertions, 34 deletions

diff --git a/drivers/md/bcache/writeback.c b/drivers/md/bcache/writeback.c
index ad45ebe1a74b..481d4cf38ac0 100644
--- a/drivers/md/bcache/writeback.c
+++ b/drivers/md/bcache/writeback.c
@@ -27,7 +27,7 @@ static uint64_t __calc_target_rate(struct cached_dev *dc)
 	 * flash-only devices
 	 */
 	uint64_t cache_sectors = c->nbuckets * c->sb.bucket_size -
-				bcache_flash_devs_sectors_dirty(c);
+				atomic_long_read(&c->flash_dev_dirty_sectors);
 
 	/*
 	 * Unfortunately there is no control of global dirty data.  If the
@@ -104,11 +104,56 @@ static void __update_writeback_rate(struct cached_dev *dc)
 
 	dc->writeback_rate_proportional = proportional_scaled;
 	dc->writeback_rate_integral_scaled = integral_scaled;
-	dc->writeback_rate_change = new_rate - dc->writeback_rate.rate;
-	dc->writeback_rate.rate = new_rate;
+	dc->writeback_rate_change = new_rate -
+			atomic_long_read(&dc->writeback_rate.rate);
+	atomic_long_set(&dc->writeback_rate.rate, new_rate);
 	dc->writeback_rate_target = target;
 }
 
+static bool set_at_max_writeback_rate(struct cache_set *c,
+				       struct cached_dev *dc)
+{
+	/*
+	 * Idle_counter is increased everytime when update_writeback_rate() is
+	 * called. If all backing devices attached to the same cache set have
+	 * identical dc->writeback_rate_update_seconds values, it is about 6
+	 * rounds of update_writeback_rate() on each backing device before
+	 * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
+	 * to each dc->writeback_rate.rate.
+	 * In order to avoid extra locking cost for counting exact dirty cached
+	 * devices number, c->attached_dev_nr is used to calculate the idle
+	 * throushold. It might be bigger if not all cached device are in write-
+	 * back mode, but it still works well with limited extra rounds of
+	 * update_writeback_rate().
+	 */
+	if (atomic_inc_return(&c->idle_counter) <
+	    atomic_read(&c->attached_dev_nr) * 6)
+		return false;
+
+	if (atomic_read(&c->at_max_writeback_rate) != 1)
+		atomic_set(&c->at_max_writeback_rate, 1);
+
+	atomic_long_set(&dc->writeback_rate.rate, INT_MAX);
+
+	/* keep writeback_rate_target as existing value */
+	dc->writeback_rate_proportional = 0;
+	dc->writeback_rate_integral_scaled = 0;
+	dc->writeback_rate_change = 0;
+
+	/*
+	 * Check c->idle_counter and c->at_max_writeback_rate agagain in case
+	 * new I/O arrives during before set_at_max_writeback_rate() returns.
+	 * Then the writeback rate is set to 1, and its new value should be
+	 * decided via __update_writeback_rate().
+	 */
+	if ((atomic_read(&c->idle_counter) <
+	     atomic_read(&c->attached_dev_nr) * 6) ||
+	    !atomic_read(&c->at_max_writeback_rate))
+		return false;
+
+	return true;
+}
+
 static void update_writeback_rate(struct work_struct *work)
 {
 	struct cached_dev *dc = container_of(to_delayed_work(work),
@@ -136,13 +181,20 @@ static void update_writeback_rate(struct work_struct *work)
 		return;
 	}
 
-	down_read(&dc->writeback_lock);
-
-	if (atomic_read(&dc->has_dirty) &&
-	    dc->writeback_percent)
-		__update_writeback_rate(dc);
+	if (atomic_read(&dc->has_dirty) && dc->writeback_percent) {
+		/*
+		 * If the whole cache set is idle, set_at_max_writeback_rate()
+		 * will set writeback rate to a max number. Then it is
+		 * unncessary to update writeback rate for an idle cache set
+		 * in maximum writeback rate number(s).
+		 */
+		if (!set_at_max_writeback_rate(c, dc)) {
+			down_read(&dc->writeback_lock);
+			__update_writeback_rate(dc);
+			up_read(&dc->writeback_lock);
+		}
+	}
 
-	up_read(&dc->writeback_lock);
 
 	/*
 	 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
@@ -422,27 +474,6 @@ static void read_dirty(struct cached_dev *dc)
 
 		delay = writeback_delay(dc, size);
 
-		/* If the control system would wait for at least half a
-		 * second, and there's been no reqs hitting the backing disk
-		 * for awhile: use an alternate mode where we have at most
-		 * one contiguous set of writebacks in flight at a time.  If
-		 * someone wants to do IO it will be quick, as it will only
-		 * have to contend with one operation in flight, and we'll
-		 * be round-tripping data to the backing disk as quickly as
-		 * it can accept it.
-		 */
-		if (delay >= HZ / 2) {
-			/* 3 means at least 1.5 seconds, up to 7.5 if we
-			 * have slowed way down.
-			 */
-			if (atomic_inc_return(&dc->backing_idle) >= 3) {
-				/* Wait for current I/Os to finish */
-				closure_sync(&cl);
-				/* And immediately launch a new set. */
-				delay = 0;
-			}
-		}
-
 		while (!kthread_should_stop() &&
 		       !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
 		       delay) {
@@ -476,6 +507,9 @@ void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned inode,
 	if (!d)
 		return;
 
+	if (UUID_FLASH_ONLY(&c->uuids[inode]))
+		atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
+
 	stripe = offset_to_stripe(d, offset);
 	stripe_offset = offset & (d->stripe_size - 1);
 
@@ -673,10 +707,14 @@ static int bch_writeback_thread(void *arg)
 }
 
 /* Init */
+#define INIT_KEYS_EACH_TIME	500000
+#define INIT_KEYS_SLEEP_MS	100
 
 struct sectors_dirty_init {
 	struct btree_op	op;
 	unsigned	inode;
+	size_t		count;
+	struct bkey	start;
 };
 
 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
@@ -691,18 +729,37 @@ static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
 		bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
 					     KEY_START(k), KEY_SIZE(k));
 
+	op->count++;
+	if (atomic_read(&b->c->search_inflight) &&
+	    !(op->count % INIT_KEYS_EACH_TIME)) {
+		bkey_copy_key(&op->start, k);
+		return -EAGAIN;
+	}
+
 	return MAP_CONTINUE;
 }
 
 void bch_sectors_dirty_init(struct bcache_device *d)
 {
 	struct sectors_dirty_init op;
+	int ret;
 
 	bch_btree_op_init(&op.op, -1);
 	op.inode = d->id;
-
-	bch_btree_map_keys(&op.op, d->c, &KEY(op.inode, 0, 0),
-			   sectors_dirty_init_fn, 0);
+	op.count = 0;
+	op.start = KEY(op.inode, 0, 0);
+
+	do {
+		ret = bch_btree_map_keys(&op.op, d->c, &op.start,
+					 sectors_dirty_init_fn, 0);
+		if (ret == -EAGAIN)
+			schedule_timeout_interruptible(
+				msecs_to_jiffies(INIT_KEYS_SLEEP_MS));
+		else if (ret < 0) {
+			pr_warn("sectors dirty init failed, ret=%d!", ret);
+			break;
+		}
+	} while (ret == -EAGAIN);
 }
 
 void bch_cached_dev_writeback_init(struct cached_dev *dc)
@@ -715,7 +772,7 @@ void bch_cached_dev_writeback_init(struct cached_dev *dc)
 	dc->writeback_running		= true;
 	dc->writeback_percent		= 10;
 	dc->writeback_delay		= 30;
-	dc->writeback_rate.rate		= 1024;
+	atomic_long_set(&dc->writeback_rate.rate, 1024);
 	dc->writeback_rate_minimum	= 8;
 
 	dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;