cns3xxx: fix ethernet IRQ imbalance in 4.14
authorKoen Vandeputte <koen.vandeputte@ncentric.com>
Thu, 11 Jan 2018 15:04:37 +0000 (16:04 +0100)
committerJohn Crispin <john@phrozen.org>
Wed, 17 Jan 2018 10:07:17 +0000 (11:07 +0100)
Kernel 4.14 contains more warnings regarding NAPI usage, which showed
following warning:

[ 4016.420000] ------------[ cut here ]------------
[ 4016.430000] WARNING: CPU: 1 PID: 14 at kernel/irq/manage.c:525
__enable_irq+0x3c/0x78
[ 4016.440000] Unbalanced enable for IRQ 51
[ 4016.440000] Modules linked in: ath9k ath9k_common pppoe ppp_async
ath9k_hw ath pppox ppp_generic nf_conntrack_ipv6 mac80211 iptable_nat
ipt_REJECT ipt_MASQUERADE cfg80211 xt_time xt_tcpudp xt_state xt_nat
xt_multiport xt_mark xt_mac xt_limit xt_conntrack xt_comment xt_TCPMSS
xt_REDIRECT xt_LOG slhc nf_reject_ipv4 nf_nat_redirs
[ 4016.500000] CPU: 1 PID: 14 Comm: ksoftirqd/1 Tainted: G        W
       4.14.11 #0
[ 4016.500000] Hardware name: Gateworks Corporation Laguna Platform
[ 4016.510000] Backtrace:
[ 4016.510000] [<c010a28c>] (dump_backtrace) from [<c010a56c>]
(show_stack+0x18/0x1c)
[ 4016.520000]  r7:c059df4b r6:00000000 r5:60000093 r4:00000000
[ 4016.530000] [<c010a554>] (show_stack) from [<c048b04c>]
(dump_stack+0x94/0xb4)
[ 4016.530000] [<c048afb8>] (dump_stack) from [<c011760c>]
(__warn+0xc8/0xf8)
[ 4016.540000]  r7:c059df4b r6:c015144c r5:00000000 r4:cf863e08
[ 4016.550000] [<c0117544>] (__warn) from [<c011767c>]
(warn_slowpath_fmt+0x40/0x48)
[ 4016.550000]  r9:cf863eb0 r8:00000036 r7:00000000 r6:c0636360
r5:00000033 r4:cf80a500
[ 4016.560000] [<c0117640>] (warn_slowpath_fmt) from [<c015144c>]
(__enable_irq+0x3c/0x78)
[ 4016.570000]  r3:00000033 r2:c059e0c5
[ 4016.570000] [<c0151410>] (__enable_irq) from [<c01514ec>]
(enable_irq+0x64/0x7c)
[ 4016.580000] [<c0151488>] (enable_irq) from [<c0335c14>]
(eth_poll+0x28c/0x558)
[ 4016.590000]  r5:d08926c0 r4:cf894488
[ 4016.590000] [<c0335988>] (eth_poll) from [<c03b1718>]
(net_rx_action+0xfc/0x2e8)
[ 4016.600000]  r10:cf863eb8 r9:cf863eb0 r8:00000001 r7:0f7b5000
r6:c0636360 r5:cfdeb360
[ 4016.610000]  r4:cf894488
[ 4016.610000] [<c03b161c>] (net_rx_action) from [<c0101510>]
(__do_softirq+0xe0/0x228)
[ 4016.620000]  r10:00000100 r9:c070204c r8:c0702040 r7:00000003
r6:00000008 r5:40000003
[ 4016.620000]  r4:cf862000
[ 4016.630000] [<c0101430>] (__do_softirq) from [<c011ae28>]
(run_ksoftirqd+0x34/0x58)
[ 4016.630000]  r10:c0133964 r9:00000000 r8:00000001 r7:00000000
r6:c070b44c r5:cf862000
[ 4016.640000]  r4:cf8037e0
[ 4016.640000] [<c011adf4>] (run_ksoftirqd) from [<c0133adc>]
(smpboot_thread_fn+0x178/0x190)
[ 4016.650000] [<c0133964>] (smpboot_thread_fn) from [<c013099c>]
(kthread+0x11c/0x138)
[ 4016.660000]  r9:cf8037e0 r8:cf802cdc r7:cf839df8 r6:cf803800
r5:00000000 r4:cf802cc0
[ 4016.670000] [<c0130880>] (kthread) from [<c0107218>]
(ret_from_fork+0x14/0x3c)
[ 4016.680000]  r10:00000000 r9:00000000 r8:00000000 r7:00000000
r6:00000000 r5:c0130880
[ 4016.680000]  r4:cf803800 r3:cf862000
[ 4016.690000] ---[ end trace 51114df08f429115 ]---

This is triggered by calling eth_schedule_poll(sw) after the IRQ has
been re-enabled.

Rework the network code to only enable IRQ's again if NAPI agrees it's
safe to do so.

Also only re-enable IRQ *after* cleaning up the RX ring and re-enabling
DMA, which otherwise resulted in ugly warnings regarding dirty page
fragments.
These popped up nearly immediately when building the kernel with O2 iso
Os.

--> Note that this change fixes stability issues, at the cost of ~8%
throughput performance.

While at it, also change the iface name being used in warning prints,
making it more obvious: "switch%d" --> "cns3xxx_eth"

The changes have been tested on 4 boards, each moving ~30TB of data

Signed-off-by: Koen Vandeputte <koen.vandeputte@ncentric.com>
target/linux/cns3xxx/patches-4.14/140-fix-ethernet-irq-imbalance.patch [new file with mode: 0644]

diff --git a/target/linux/cns3xxx/patches-4.14/140-fix-ethernet-irq-imbalance.patch b/target/linux/cns3xxx/patches-4.14/140-fix-ethernet-irq-imbalance.patch
new file mode 100644 (file)
index 0000000..3538586
--- /dev/null
@@ -0,0 +1,77 @@
+--- a/drivers/net/ethernet/cavium/cns3xxx_eth.c
++++ b/drivers/net/ethernet/cavium/cns3xxx_eth.c
+@@ -712,26 +712,20 @@ static int eth_poll(struct napi_struct *
+       }
+       rx_ring->cur_index = i;
+-      if (!received) {
+-              napi_complete(napi);
+-              enable_irq(sw->rx_irq);
+-              budget = 0;
+-
+-              /* If 1 or more frames came in during IRQ enable, re-schedule */
+-              if (rx_ring->desc[i].cown)
+-                      eth_schedule_poll(sw);
+-      }
+-
+-      spin_lock_bh(&tx_lock);
+-      eth_complete_tx(sw);
+-      spin_unlock_bh(&tx_lock);
+       cns3xxx_alloc_rx_buf(sw, received);
+-
+       wmb();
+       enable_rx_dma(sw);
+-      return budget;
++      if (received < budget && napi_complete_done(napi, received)) {
++              enable_irq(sw->rx_irq);
++      }
++      
++      spin_lock_bh(&tx_lock);
++      eth_complete_tx(sw);
++      spin_unlock_bh(&tx_lock);
++      
++      return received;
+ }
+ static void eth_set_desc(struct sw *sw, struct _tx_ring *tx_ring, int index,
+@@ -856,18 +850,6 @@ static void cns3xxx_get_drvinfo(struct n
+       strcpy(info->bus_info, "internal");
+ }
+-static int cns3xxx_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
+-{
+-      struct port *port = netdev_priv(dev);
+-      return phy_ethtool_gset(port->phydev, cmd);
+-}
+-
+-static int cns3xxx_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
+-{
+-      struct port *port = netdev_priv(dev);
+-      return phy_ethtool_sset(port->phydev, cmd);
+-}
+-
+ static int cns3xxx_nway_reset(struct net_device *dev)
+ {
+       struct port *port = netdev_priv(dev);
+@@ -876,8 +858,8 @@ static int cns3xxx_nway_reset(struct net
+ static struct ethtool_ops cns3xxx_ethtool_ops = {
+       .get_drvinfo = cns3xxx_get_drvinfo,
+-      .get_settings = cns3xxx_get_settings,
+-      .set_settings = cns3xxx_set_settings,
++      .get_link_ksettings = phy_ethtool_get_link_ksettings,
++      .set_link_ksettings = phy_ethtool_set_link_ksettings,
+       .nway_reset = cns3xxx_nway_reset,
+       .get_link = ethtool_op_get_link,
+ };
+@@ -1177,7 +1159,7 @@ static int eth_init_one(struct platform_
+               goto err_remove_mdio;
+       }
+-      strcpy(napi_dev->name, "switch%d");
++      strcpy(napi_dev->name, "cns3xxx_eth");
+       napi_dev->features = NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_FRAGLIST;
+       SET_NETDEV_DEV(napi_dev, &pdev->dev);