On 08/21/2017 04:23 PM, Florian Fainelli wrote:
> On 08/04/2017 10:11 AM, Eric Dumazet wrote:
>> On Fri, 2017-08-04 at 08:51 -0700, Florian Fainelli wrote:
>>> On 08/03/2017 10:36 PM, Eric Dumazet wrote:
>>>> On Thu, 2017-08-03 at 21:33 -0700, Florian Fainelli wrote:
>>>>> During testing with a background iperf pushing 1Gbit/sec worth of
>>>>> traffic and having both ifconfig and ethtool collect statistics, we
>>>>> could see quite frequent deadlocks. Convert the often accessed DSA slave
>>>>> network devices statistics to per-cpu 64-bit statistics to remove these
>>>>> deadlocks and provide fast efficient statistics updates.
>>>>>
>>>>
>>>> This seems to be a bug fix, it would be nice to get a proper tag like :
>>>>
>>>> Fixes: f613ed665bb3 ("net: dsa: Add support for 64-bit statistics")
>>>
>>> Right, should have been added, thanks!
>>>
>>>>
>>>> Problem here is that if multiple cpus can call dsa_switch_rcv() at the
>>>> same time, then u64_stats_update_begin() contract is not respected.
>>>
>>> This is really where I struggled understanding what is wrong in the
>>> non-per CPU version, my understanding is that we have:
>>>
>>> - writers for xmit executes in process context
>>> - writers for receive executes from NAPI (from the DSA's master network
>>> device through it's own NAPI doing netif_receive_skb -> netdev_uses_dsa
>>> -> netif_receive_skb)
>>>
>>> readers should all execute in process context. The test scenario that
>>> led to a deadlock involved running iperf in the background, having a
>>> while loop with both ifconfig and ethtool reading stats, and somehow
>>> when iperf exited, either reader would just be locked. So I guess this
>>> leaves us with the two writers not being mutually excluded then, right?
>>
>> You could add a debug version of u64_stats_update_begin()
>>
>> doing
>>
>> int ret = atomic_inc((atomic_t *)syncp);
>>
>> BUG_ON(ret & 1);>
>>
>> And u64_stats_update_end()
>>
>> int ret = atomic_inc((atomic_t *)syncp);
>
> so with your revised suggested patch:
>
> static inline void u64_stats_update_begin(struct u64_stats_sync *syncp)
> {
> #if BITS_PER_LONG==32 && defined(CONFIG_SMP)
> int ret = atomic_inc_return((atomic_t *)syncp);
> BUG_ON(ret & 1);
> #endif
> #if 0
> #if BITS_PER_LONG==32 && defined(CONFIG_SMP)
> write_seqcount_begin(&syncp->seq);
> #endif
> #endif
> }
>
> static inline void u64_stats_update_end(struct u64_stats_sync *syncp)
> {
> #if BITS_PER_LONG==32 && defined(CONFIG_SMP)
> int ret = atomic_inc_return((atomic_t *)syncp);
> BUG_ON(!(ret & 1));
> #endif
> #if 0
> #if BITS_PER_LONG==32 && defined(CONFIG_SMP)
> write_seqcount_end(&syncp->seq);
> #endif
> #endif
> }
>
> and this makes us choke pretty early in IRQ accounting, did I get your
> suggestion right?
Well if we return 1 from atomic_inc_return() and the previous value was
zero, of course we are going to be bugging here. The idea behind the
patch I suppose is to make sure that we always get an odd number upon
u64_stats_update_begin()/entry, and an even number upon
u64_stats_update_end()/exit, right?
>
> [ 0.015149] ------------[ cut here ]------------
> [ 0.020051] kernel BUG at ./include/linux/u64_stats_sync.h:82!
> [ 0.026221] Internal error: Oops - BUG: 0 [#1] SMP ARM
> [ 0.031661] Modules linked in:
> [ 0.034970] CPU: 0 PID: 0 Comm: swapper/0 Not tainted
> 4.13.0-rc5-01297-g7d3f0cd43fee-dirty #33
> [ 0.043990] Hardware name: Broadcom STB (Flattened Device Tree)
> [ 0.050237] task: c180a500 task.stack: c1800000
> [ 0.055065] PC is at irqtime_account_delta+0xa4/0xa8
> [ 0.060322] LR is at 0x1
> [ 0.063057] pc : [<c0250504>] lr : [<00000001>] psr: 000001d3
> [ 0.069652] sp : c1801eec ip : ee78b458 fp : c0e5ea48
> [ 0.075212] r10: c18b4b40 r9 : f0803000 r8 : ee00a800
> [ 0.080781] r7 : 00000001 r6 : c180a500 r5 : c1800000 r4 : 00000000
> [ 0.087680] r3 : 00000000 r2 : 0000ec8c r1 : ee78b3c0 r0 : ee78b440
> [ 0.094546] Flags: nzcv IRQs off FIQs off Mode SVC_32 ISA ARM
> Segment user
> [ 0.102314] Control: 30c5387d Table: 00003000 DAC: fffffffd
> [ 0.108414] Process swapper/0 (pid: 0, stack limit = 0xc1800210)
> [ 0.114791] Stack: (0xc1801eec to 0xc1802000)
> [ 0.119431] 1ee0: ee78b440 c1800000
> c180a500 00000001 c02505c8
> [ 0.128079] 1f00: 00000004 ee00a800 ffffe000 00000000 00000000
> c0227890 c17e6f20 c0278910
> [ 0.136665] 1f20: c185724c c18079a0 f080200c c1801f58 f0802000
> c0201494 c0e00c18 20000053
> [ 0.145303] 1f40: ffffffff c1801f8c ffffffff c1800000 c18b4b40
> c020d238 00000000 0000001f
> [ 0.153915] 1f60: 00040d00 00000000 efffc940 00000000 c18b4b40
> c1807440 ffffffff 00000000
> [ 0.162571] 1f80: c18b4b40 c0e5ea48 00000004 c1801fa8 c0322fb0
> c0e00c18 20000053 ffffffff
> [ 0.171226] 1fa0: c18b4b40 00000000 ffffffff ffffffff 00000000
> c0e006c0 ffffffff 00000000
> [ 0.179890] 1fc0: 00000000 c1807448 c0e5ea48 00000000 00000000
> c18b4dd4 c180745c c0e5ea44
> [ 0.188546] 1fe0: c180c0d0 00007000 420f00f3 00000000 00000000
> 00008090 00000000 00000000
> [ 0.197165] [<c0250504>] (irqtime_account_delta) from [<c02505c8>]
> (irqtime_account_irq+0xc0/0xc4)
> [ 0.206664] [<c02505c8>] (irqtime_account_irq) from [<c0227890>]
> (irq_exit+0x28/0x154)
> [ 0.215012] [<c0227890>] (irq_exit) from [<c0278910>]
> (__handle_domain_irq+0x60/0xb4)
> [ 0.223245] [<c0278910>] (__handle_domain_irq) from [<c0201494>]
> (gic_handle_irq+0x48/0x8c)
> [ 0.232035] [<c0201494>] (gic_handle_irq) from [<c020d238>]
> (__irq_svc+0x58/0x74)
> [ 0.239941] Exception stack(0xc1801f58 to 0xc1801fa0)
> [ 0.245327] 1f40:
> 00000000 0000001f
> [ 0.253948] 1f60: 00040d00 00000000 efffc940 00000000 c18b4b40
> c1807440 ffffffff 00000000
> [ 0.262534] 1f80: c18b4b40 c0e5ea48 00000004 c1801fa8 c0322fb0
> c0e00c18 20000053 ffffffff
> [ 0.271144] [<c020d238>] (__irq_svc) from [<c0e00c18>]
> (start_kernel+0x300/0x410)
> [ 0.279028] [<c0e00c18>] (start_kernel) from [<00008090>] (0x8090)
> [ 0.285547] Code: f57ff05b e3130001 18bd80f0 e7f001f2 (e7f001f2)
> [ 0.291978] ---[ end trace f68728a0d3053b52 ]---
> [ 0.296871] Kernel panic - not syncing: Fatal exception in interrupt
> [ 0.303622] ---[ end Kernel panic - not syncing: Fatal exception in
> interrupt
>
--
Florian
