Koen Vandeputte
|
f3dec44520
|
cns3xxx: fix ethernet IRQ imbalance in 4.14
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>
|
2018-01-17 11:07:17 +01:00 |
|