Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next

Pull networking updates from David Miller:
 "Another merge window, another set of networking changes.  I've heard
  rumblings that the lightweight tunnels infrastructure has been voted
  networking change of the year.  But what do I know?

   1) Add conntrack support to openvswitch, from Joe Stringer.

   2) Initial support for VRF (Virtual Routing and Forwarding), which
      allows the segmentation of routing paths without using multiple
      devices.  There are some semantic kinks to work out still, but
      this is a reasonably strong foundation.  From David Ahern.

   3) Remove spinlock fro act_bpf fast path, from Alexei Starovoitov.

   4) Ignore route nexthops with a link down state in ipv6, just like
      ipv4.  From Andy Gospodarek.

   5) Remove spinlock from fast path of act_gact and act_mirred, from
      Eric Dumazet.

   6) Document the DSA layer, from Florian Fainelli.

   7) Add netconsole support to bcmgenet, systemport, and DSA.  Also
      from Florian Fainelli.

   8) Add Mellanox Switch Driver and core infrastructure, from Jiri
      Pirko.

   9) Add support for "light weight tunnels", which allow for
      encapsulation and decapsulation without bearing the overhead of a
      full blown netdevice.  From Thomas Graf, Jiri Benc, and a cast of
      others.

  10) Add Identifier Locator Addressing support for ipv6, from Tom
      Herbert.

  11) Support fragmented SKBs in iwlwifi, from Johannes Berg.

  12) Allow perf PMUs to be accessed from eBPF programs, from Kaixu Xia.

  13) Add BQL support to 3c59x driver, from Loganaden Velvindron.

  14) Stop using a zero TX queue length to mean that a device shouldn't
      have a qdisc attached, use an explicit flag instead.  From Phil
      Sutter.

  15) Use generic geneve netdevice infrastructure in openvswitch, from
      Pravin B Shelar.

  16) Add infrastructure to avoid re-forwarding a packet in software
      that was already forwarded by a hardware switch.  From Scott
      Feldman.

  17) Allow AF_PACKET fanout function to be implemented in a bpf
      program, from Willem de Bruijn"

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1458 commits)
  netfilter: nf_conntrack: make nf_ct_zone_dflt built-in
  netfilter: nf_dup{4, 6}: fix build error when nf_conntrack disabled
  net: fec: clear receive interrupts before processing a packet
  ipv6: fix exthdrs offload registration in out_rt path
  xen-netback: add support for multicast control
  bgmac: Update fixed_phy_register()
  sock, diag: fix panic in sock_diag_put_filterinfo
  flow_dissector: Use 'const' where possible.
  flow_dissector: Fix function argument ordering dependency
  ixgbe: Resolve "initialized field overwritten" warnings
  ixgbe: Remove bimodal SR-IOV disabling
  ixgbe: Add support for reporting 2.5G link speed
  ixgbe: fix bounds checking in ixgbe_setup_tc for 82598
  ixgbe: support for ethtool set_rxfh
  ixgbe: Avoid needless PHY access on copper phys
  ixgbe: cleanup to use cached mask value
  ixgbe: Remove second instance of lan_id variable
  ixgbe: use kzalloc for allocating one thing
  flow: Move __get_hash_from_flowi{4,6} into flow_dissector.c
  ixgbe: Remove unused PCI bus types
  ...

1 files changed, 65 insertions, 11 deletions

diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
index cac5b45..6bbe620 100644
--- a/net/rds/ib_recv.c
+++ b/net/rds/ib_recv.c
@@ -297,7 +297,7 @@ static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic
 }
 
 static int rds_ib_recv_refill_one(struct rds_connection *conn,
-				  struct rds_ib_recv_work *recv, int prefill)
+				  struct rds_ib_recv_work *recv, gfp_t gfp)
 {
 	struct rds_ib_connection *ic = conn->c_transport_data;
 	struct ib_sge *sge;
@@ -305,7 +305,7 @@ static int rds_ib_recv_refill_one(struct rds_connection *conn,
 	gfp_t slab_mask = GFP_NOWAIT;
 	gfp_t page_mask = GFP_NOWAIT;
 
-	if (prefill) {
+	if (gfp & __GFP_WAIT) {
 		slab_mask = GFP_KERNEL;
 		page_mask = GFP_HIGHUSER;
 	}
@@ -347,6 +347,24 @@ out:
 	return ret;
 }
 
+static int acquire_refill(struct rds_connection *conn)
+{
+	return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0;
+}
+
+static void release_refill(struct rds_connection *conn)
+{
+	clear_bit(RDS_RECV_REFILL, &conn->c_flags);
+
+	/* We don't use wait_on_bit()/wake_up_bit() because our waking is in a
+	 * hot path and finding waiters is very rare.  We don't want to walk
+	 * the system-wide hashed waitqueue buckets in the fast path only to
+	 * almost never find waiters.
+	 */
+	if (waitqueue_active(&conn->c_waitq))
+		wake_up_all(&conn->c_waitq);
+}
+
 /*
  * This tries to allocate and post unused work requests after making sure that
  * they have all the allocations they need to queue received fragments into
@@ -354,15 +372,23 @@ out:
  *
  * -1 is returned if posting fails due to temporary resource exhaustion.
  */
-void rds_ib_recv_refill(struct rds_connection *conn, int prefill)
+void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
 {
 	struct rds_ib_connection *ic = conn->c_transport_data;
 	struct rds_ib_recv_work *recv;
 	struct ib_recv_wr *failed_wr;
 	unsigned int posted = 0;
 	int ret = 0;
+	bool can_wait = !!(gfp & __GFP_WAIT);
 	u32 pos;
 
+	/* the goal here is to just make sure that someone, somewhere
+	 * is posting buffers.  If we can't get the refill lock,
+	 * let them do their thing
+	 */
+	if (!acquire_refill(conn))
+		return;
+
 	while ((prefill || rds_conn_up(conn)) &&
 	       rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
 		if (pos >= ic->i_recv_ring.w_nr) {
@@ -372,7 +398,7 @@ void rds_ib_recv_refill(struct rds_connection *conn, int prefill)
 		}
 
 		recv = &ic->i_recvs[pos];
-		ret = rds_ib_recv_refill_one(conn, recv, prefill);
+		ret = rds_ib_recv_refill_one(conn, recv, gfp);
 		if (ret) {
 			break;
 		}
@@ -402,6 +428,24 @@ void rds_ib_recv_refill(struct rds_connection *conn, int prefill)
 
 	if (ret)
 		rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
+
+	release_refill(conn);
+
+	/* if we're called from the softirq handler, we'll be GFP_NOWAIT.
+	 * in this case the ring being low is going to lead to more interrupts
+	 * and we can safely let the softirq code take care of it unless the
+	 * ring is completely empty.
+	 *
+	 * if we're called from krdsd, we'll be GFP_KERNEL.  In this case
+	 * we might have raced with the softirq code while we had the refill
+	 * lock held.  Use rds_ib_ring_low() instead of ring_empty to decide
+	 * if we should requeue.
+	 */
+	if (rds_conn_up(conn) &&
+	    ((can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
+	    rds_ib_ring_empty(&ic->i_recv_ring))) {
+		queue_delayed_work(rds_wq, &conn->c_recv_w, 1);
+	}
 }
 
 /*
@@ -982,10 +1026,17 @@ static inline void rds_poll_cq(struct rds_ib_connection *ic,
 		}
 
 		/*
-		 * It's very important that we only free this ring entry if we've truly
-		 * freed the resources allocated to the entry.  The refilling path can
-		 * leak if we don't.
+		 * rds_ib_process_recv() doesn't always consume the frag, and
+		 * we might not have called it at all if the wc didn't indicate
+		 * success. We already unmapped the frag's pages, though, and
+		 * the following rds_ib_ring_free() call tells the refill path
+		 * that it will not find an allocated frag here. Make sure we
+		 * keep that promise by freeing a frag that's still on the ring.
 		 */
+		if (recv->r_frag) {
+			rds_ib_frag_free(ic, recv->r_frag);
+			recv->r_frag = NULL;
+		}
 		rds_ib_ring_free(&ic->i_recv_ring, 1);
 	}
 }
@@ -1016,7 +1067,7 @@ void rds_ib_recv_tasklet_fn(unsigned long data)
 		rds_ib_stats_inc(s_ib_rx_ring_empty);
 
 	if (rds_ib_ring_low(&ic->i_recv_ring))
-		rds_ib_recv_refill(conn, 0);
+		rds_ib_recv_refill(conn, 0, GFP_NOWAIT);
 }
 
 int rds_ib_recv(struct rds_connection *conn)
@@ -1025,8 +1076,10 @@ int rds_ib_recv(struct rds_connection *conn)
 	int ret = 0;
 
 	rdsdebug("conn %p\n", conn);
-	if (rds_conn_up(conn))
+	if (rds_conn_up(conn)) {
 		rds_ib_attempt_ack(ic);
+		rds_ib_recv_refill(conn, 0, GFP_KERNEL);
+	}
 
 	return ret;
 }
@@ -1049,9 +1102,10 @@ int rds_ib_recv_init(void)
 	rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
 					sizeof(struct rds_page_frag),
 					0, SLAB_HWCACHE_ALIGN, NULL);
-	if (!rds_ib_frag_slab)
+	if (!rds_ib_frag_slab) {
 		kmem_cache_destroy(rds_ib_incoming_slab);
-	else
+		rds_ib_incoming_slab = NULL;
+	} else
 		ret = 0;
 out:
 	return ret;