RDS/IB: create a work queue for FMR flushing

This patch moves the FMR flushing work in to its own mult-threaded work queue.
This is to maintain performance in preparation for returning the main krdsd
work queue back to a single threaded work queue to avoid deep-rooted
concurrency bugs.

This is also good because it further separates FMRs, which might be removed
some day, from the rest of the code base.

Signed-off-by: Zach Brown <zach.brown@oracle.com>

3 files changed, 33 insertions, 3 deletions

diff --git a/net/rds/ib.c b/net/rds/ib.c
index af1ef18b6ff..d2007b93161 100644
--- a/net/rds/ib.c
+++ b/net/rds/ib.c
@@ -351,6 +351,7 @@ void rds_ib_exit(void)
         rds_ib_sysctl_exit();
         rds_ib_recv_exit();
         rds_trans_unregister(&rds_ib_transport);
+        rds_ib_fmr_exit();
 }
 
 struct rds_transport rds_ib_transport = {
@@ -386,10 +387,14 @@ int __init rds_ib_init(void)
 
         INIT_LIST_HEAD(&rds_ib_devices);
 
-        ret = ib_register_client(&rds_ib_client);
+        ret = rds_ib_fmr_init();
         if (ret)
                 goto out;
 
+        ret = ib_register_client(&rds_ib_client);
+        if (ret)
+                goto out_fmr_exit;
+
         ret = rds_ib_sysctl_init();
         if (ret)
                 goto out_ibreg;
@@ -412,6 +417,8 @@ out_sysctl:
         rds_ib_sysctl_exit();
 out_ibreg:
         rds_ib_unregister_client();
+out_fmr_exit:
+        rds_ib_fmr_exit();
 out:
         return ret;
 }
diff --git a/net/rds/ib.h b/net/rds/ib.h
index e9f9ddf440c..fd4ea69d244 100644
--- a/net/rds/ib.h
+++ b/net/rds/ib.h
@@ -308,6 +308,8 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 void rds_ib_sync_mr(void *trans_private, int dir);
 void rds_ib_free_mr(void *trans_private, int invalidate);
 void rds_ib_flush_mrs(void);
+int __init rds_ib_fmr_init(void);
+void __exit rds_ib_fmr_exit(void);
 
 /* ib_recv.c */
 int __init rds_ib_recv_init(void);
diff --git a/net/rds/ib_rdma.c b/net/rds/ib_rdma.c
index 00f3995351c..0eb597670c5 100644
--- a/net/rds/ib_rdma.c
+++ b/net/rds/ib_rdma.c
@@ -691,6 +691,26 @@ out_nolock:
         return ret;
 }
 
+struct workqueue_struct *rds_ib_fmr_wq;
+
+int __init rds_ib_fmr_init(void)
+{
+        rds_ib_fmr_wq = create_workqueue("rds_fmr_flushd");
+        if (!rds_ib_fmr_wq)
+                return -ENOMEM;
+        return 0;
+}
+
+/*
+ * By the time this is called all the IB devices should have been torn down and
+ * had their pools freed.  As each pool is freed its work struct is waited on,
+ * so the pool flushing work queue should be idle by the time we get here.
+ */
+void __exit rds_ib_fmr_exit(void)
+{
+        destroy_workqueue(rds_ib_fmr_wq);
+}
+
 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 {
         struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
@@ -718,7 +738,7 @@ void rds_ib_free_mr(void *trans_private, int invalidate)
         /* If we've pinned too many pages, request a flush */
         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
             atomic_read(&pool->dirty_count) >= pool->max_items / 10)
-                queue_delayed_work(rds_wq, &pool->flush_worker, 10);
+                queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
 
         if (invalidate) {
                 if (likely(!in_interrupt())) {
@@ -726,7 +746,8 @@ void rds_ib_free_mr(void *trans_private, int invalidate)
                 } else {
                         /* We get here if the user created a MR marked
                          * as use_once and invalidate at the same time. */
-                        queue_delayed_work(rds_wq, &pool->flush_worker, 10);
+                        queue_delayed_work(rds_ib_fmr_wq,
+                                           &pool->flush_worker, 10);
                 }
         }