net: systemport: Fix 64-bit stats deadlock
authorFlorian Fainelli <f.fainelli@gmail.com>
Tue, 12 Sep 2017 20:14:26 +0000 (13:14 -0700)
committerDavid S. Miller <davem@davemloft.net>
Fri, 15 Sep 2017 21:24:10 +0000 (14:24 -0700)
We can enter a deadlock situation because there is no sufficient protection
when ndo_get_stats64() runs in process context to guard against RX or TX NAPI
contexts running in softirq, this can lead to the following lockdep splat and
actual deadlock was experienced as well with an iperf session in the background
and a while loop doing ifconfig + ethtool.

[    5.780350] ================================
[    5.784679] WARNING: inconsistent lock state
[    5.789011] 4.13.0-rc7-02179-g32fae27c725d #70 Not tainted
[    5.794561] --------------------------------
[    5.798890] inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} usage.
[    5.804971] swapper/0/0 [HC0[0]:SC1[1]:HE0:SE0] takes:
[    5.810175]  (&syncp->seq#2){+.?...}, at: [<c0768a28>] bcm_sysport_tx_reclaim+0x30/0x54
[    5.818327] {SOFTIRQ-ON-W} state was registered at:
[    5.823278]   bcm_sysport_get_stats64+0x17c/0x258
[    5.828053]   dev_get_stats+0x38/0xac
[    5.831776]   rtnl_fill_stats+0x30/0x118
[    5.835761]   rtnl_fill_ifinfo+0x538/0xe24
[    5.839921]   rtmsg_ifinfo_build_skb+0x6c/0xd8
[    5.844430]   rtmsg_ifinfo_event.part.5+0x14/0x44
[    5.849201]   rtmsg_ifinfo+0x20/0x28
[    5.852837]   register_netdevice+0x628/0x6b8
[    5.857171]   register_netdev+0x14/0x24
[    5.861051]   bcm_sysport_probe+0x30c/0x438
[    5.865280]   platform_drv_probe+0x50/0xb0
[    5.869418]   driver_probe_device+0x2e8/0x450
[    5.873817]   __driver_attach+0x104/0x120
[    5.877871]   bus_for_each_dev+0x7c/0xc0
[    5.881834]   bus_add_driver+0x1b0/0x270
[    5.885797]   driver_register+0x78/0xf4
[    5.889675]   do_one_initcall+0x54/0x190
[    5.893646]   kernel_init_freeable+0x144/0x1d0
[    5.898135]   kernel_init+0x8/0x110
[    5.901665]   ret_from_fork+0x14/0x2c
[    5.905363] irq event stamp: 24263
[    5.908804] hardirqs last  enabled at (24262): [<c08eecf0>] net_rx_action+0xc4/0x4e4
[    5.916624] hardirqs last disabled at (24263): [<c0a7da00>] _raw_spin_lock_irqsave+0x1c/0x98
[    5.925143] softirqs last  enabled at (24258): [<c022a7fc>] irq_enter+0x84/0x98
[    5.932524] softirqs last disabled at (24259): [<c022a918>] irq_exit+0x108/0x16c
[    5.939985]
[    5.939985] other info that might help us debug this:
[    5.946576]  Possible unsafe locking scenario:
[    5.946576]
[    5.952556]        CPU0
[    5.955031]        ----
[    5.957506]   lock(&syncp->seq#2);
[    5.960955]   <Interrupt>
[    5.963604]     lock(&syncp->seq#2);
[    5.967227]
[    5.967227]  *** DEADLOCK ***
[    5.967227]
[    5.973222] 1 lock held by swapper/0/0:
[    5.977092]  #0:  (&(&ring->lock)->rlock){..-...}, at: [<c0768a18>] bcm_sysport_tx_reclaim+0x20/0x54

So just remove the u64_stats_update_begin()/end() pair in ndo_get_stats64()
since it does not appear to be useful for anything. No inconsistency was
observed with either ifconfig or ethtool, global TX counts equal the sum of
per-queue TX counts on a 32-bit architecture.

Fixes: 10377ba7673d ("net: systemport: Support 64bit statistics")
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/broadcom/bcmsysport.c

index a6572b51435aef4fd630a3868b546a35b7e500cf..c3c53f6cd9e62cb6551533153f455a2cd7285954 100644 (file)
@@ -1735,11 +1735,8 @@ static void bcm_sysport_get_stats64(struct net_device *dev,
                stats->tx_packets += tx_packets;
        }
 
-       /* lockless update tx_bytes and tx_packets */
-       u64_stats_update_begin(&priv->syncp);
        stats64->tx_bytes = stats->tx_bytes;
        stats64->tx_packets = stats->tx_packets;
-       u64_stats_update_end(&priv->syncp);
 
        do {
                start = u64_stats_fetch_begin_irq(&priv->syncp);