RDS: Refill recv ring directly from tasklet
authorAndy Grover <andy.grover@oracle.com>
Fri, 19 Mar 2010 00:19:52 +0000 (17:19 -0700)
committerAndy Grover <andy.grover@oracle.com>
Thu, 9 Sep 2010 01:12:11 +0000 (18:12 -0700)
Performance is better if we use allocations that don't block
to refill the receive ring. Since the whole reason we were
kicking out to the worker thread was so we could do blocking
allocs, we no longer need to do this.

Remove gfp params from rds_ib_recv_refill(); we always use
GFP_NOWAIT.

Signed-off-by: Andy Grover <andy.grover@oracle.com>
net/rds/ib.h
net/rds/ib_cm.c
net/rds/ib_recv.c

index 202140a84f0c2836615e6e15275e232607919bfd..4c2ee2401979fa7a5894779962a07b9f87072009 100644 (file)
@@ -297,8 +297,7 @@ void rds_ib_flush_mrs(void);
 int __init rds_ib_recv_init(void);
 void rds_ib_recv_exit(void);
 int rds_ib_recv(struct rds_connection *conn);
-int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
-                      gfp_t page_gfp, int prefill);
+int rds_ib_recv_refill(struct rds_connection *conn, int prefill);
 void rds_ib_inc_free(struct rds_incoming *inc);
 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
                             size_t size);
index 1a91af75f4c74b4bb15c6123e26302a1775ce706..75eda9c82135fea944b394d27090ae88ca92face 100644 (file)
@@ -135,7 +135,7 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
        rds_ib_recv_init_ring(ic);
        /* Post receive buffers - as a side effect, this will update
         * the posted credit count. */
-       rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1);
+       rds_ib_recv_refill(conn, 1);
 
        /* Tune RNR behavior */
        rds_ib_tune_rnr(ic, &qp_attr);
index e9fe08a86c406762c071b782f72742658b97a59e..8f041f7954a28d11528037d3a79feba64b4a703f 100644 (file)
@@ -135,8 +135,7 @@ void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
 }
 
 static int rds_ib_recv_refill_one(struct rds_connection *conn,
-                                 struct rds_ib_recv_work *recv,
-                                 gfp_t kptr_gfp, gfp_t page_gfp)
+                                 struct rds_ib_recv_work *recv)
 {
        struct rds_ib_connection *ic = conn->c_transport_data;
        dma_addr_t dma_addr;
@@ -148,8 +147,7 @@ static int rds_ib_recv_refill_one(struct rds_connection *conn,
                        rds_ib_stats_inc(s_ib_rx_alloc_limit);
                        goto out;
                }
-               recv->r_ibinc = kmem_cache_alloc(rds_ib_incoming_slab,
-                                                kptr_gfp);
+               recv->r_ibinc = kmem_cache_alloc(rds_ib_incoming_slab, GFP_NOWAIT);
                if (!recv->r_ibinc) {
                        atomic_dec(&rds_ib_allocation);
                        goto out;
@@ -159,7 +157,7 @@ static int rds_ib_recv_refill_one(struct rds_connection *conn,
        }
 
        if (!recv->r_frag) {
-               recv->r_frag = kmem_cache_alloc(rds_ib_frag_slab, kptr_gfp);
+               recv->r_frag = kmem_cache_alloc(rds_ib_frag_slab, GFP_NOWAIT);
                if (!recv->r_frag)
                        goto out;
                INIT_LIST_HEAD(&recv->r_frag->f_item);
@@ -167,7 +165,7 @@ static int rds_ib_recv_refill_one(struct rds_connection *conn,
        }
 
        if (!ic->i_frag.f_page) {
-               ic->i_frag.f_page = alloc_page(page_gfp);
+               ic->i_frag.f_page = alloc_page(GFP_NOWAIT);
                if (!ic->i_frag.f_page)
                        goto out;
                ic->i_frag.f_offset = 0;
@@ -221,8 +219,7 @@ out:
  *
  * -1 is returned if posting fails due to temporary resource exhaustion.
  */
-int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
-                      gfp_t page_gfp, int prefill)
+int rds_ib_recv_refill(struct rds_connection *conn, int prefill)
 {
        struct rds_ib_connection *ic = conn->c_transport_data;
        struct rds_ib_recv_work *recv;
@@ -241,7 +238,7 @@ int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
                }
 
                recv = &ic->i_recvs[pos];
-               ret = rds_ib_recv_refill_one(conn, recv, kptr_gfp, page_gfp);
+               ret = rds_ib_recv_refill_one(conn, recv);
                if (ret) {
                        ret = -1;
                        break;
@@ -856,11 +853,8 @@ void rds_ib_recv_tasklet_fn(unsigned long data)
        if (rds_ib_ring_empty(&ic->i_recv_ring))
                rds_ib_stats_inc(s_ib_rx_ring_empty);
 
-       /*
-        * If the ring is running low, then schedule the thread to refill.
-        */
        if (rds_ib_ring_low(&ic->i_recv_ring))
-               queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
+               rds_ib_recv_refill(conn, 0);
 }
 
 int rds_ib_recv(struct rds_connection *conn)
@@ -875,7 +869,7 @@ int rds_ib_recv(struct rds_connection *conn)
         * we're really low and we want the caller to back off for a bit.
         */
        mutex_lock(&ic->i_recv_mutex);
-       if (rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 0))
+       if (rds_ib_recv_refill(conn, 0))
                ret = -ENOMEM;
        else
                rds_ib_stats_inc(s_ib_rx_refill_from_thread);