Re: [Intel-wired-lan] Supermicro AOC-STGN-I1S (Intel 82599EN based 10G adapter) - poor network perfomance after moving to Debian 11.5

["Linux kernel regression tracking (Thorsten Leemhuis)" <regressions@leemhuis.info>]

On 23.01.23 20:03, Paul Menzel wrote:
> Am 23.01.23 um 19:58 schrieb Bartek Kois:
> 
>> W dniu 23.01.2023 o 19:53, Paul Menzel pisze:
> 
>>> Am 23.01.23 um 19:38 schrieb Bartek Kois:
>>>>
>>>> W dniu 22.01.2023 o 21:28, Paul Menzel pisze:
>>>>> Dear Bartek,
>>>>>
>>>>>
>>>>> Am 19.01.23 um 18:17 schrieb Bartek Kois:
>>>>>> W dniu 19.01.2023 o 18:09, Paul Menzel pisze:
>>>>>
>>>>>>> Am 19.01.23 um 17:58 schrieb Bartek Kois:
>>>>>>>> W dniu 19.01.2023 o 13:24, Bartek Kois pisze:
>>>>>>>>>
>>>>>>>>> W dniu 19.01.2023 o 11:17, Paul Menzel pisze:
>>>>>>>>>>
>>>>>>>>>> #regzbot ^introduced: 4.9.88..5.10.149
>>>>>>>
>>>>>>>>>> Am 14.01.23 um 11:23 schrieb Bartek Kois:
>>>>>>>>>>
>>>>>>>>>>> After moving from Debian 9.7 to 11.5 as soon as I perform "ip
>>>>>>>>>>> link set enp1s0 up" for my 10G adapter (AOC-STGN-I1S - Intel
>>>>>>>>>>> 82599EN based 10G adapter) I am experiencing high cpu load
>>>>>>>>>>> (even if no traffic is passing through the adapter) and
>>>>>>>>>>> network performance is low (when network is connected).
>>>>>>>>>>
>>>>>>>>>> How do you test the network performance? Please give exact
>>>>>>>>>> numbers for comparison.
>>>>>>>>>>
>>>>>>>>> I am using this server as a router for my subscribers with
>>>>>>>>> iptables (for NAT and firewall) and hfsc (for QoS). First I
>>>>>>>>> encountered this problem while migrating form Debian 9.7 to
>>>>>>>>> 11.5. Routers based  on Supermicro X11SSL-F (Intel® C232
>>>>>>>>> chipset) works with no problems after that migration, but
>>>>>>>>> routers based on Supermicro X9SCL (Intel C202 PCH) and
>>>>>>>>> Supermicro X10SLL+-F (Intel C222 Express PCH) starts behaving
>>>>>>>>> strangely with high cpu load (0.5-0.8 while before it was
>>>>>>>>> around 0.0-0.1) and subscribers not being able to utilize their
>>>>>>>>> plans. I tried to strip down the problem and ends up with clean
>>>>>>>>> system with no iptables or hfsc rules behaving the same (higher
>>>>>>>>> load) right after setting the 10G link upeven if no traffic is
>>>>>>>>> passing by.
>>>>>>>>>
>>>>>>>>>>> The cpu load is oscillating between 0.1 and 0.3 on vanilla
>>>>>>>>>>> system
>>>>>>>>>>> with no network attached. The problem can be observed on the
>>>>>>>>>>> following platforms: Supermicro X9SCL (Intel C202 PCH) and
>>>>>>>>>>> Supermicro X10SLL+-F (Intel C222 Express PCH), but for the
>>>>>>>>>>> Supermicro
>>>>>>>>>>> X11SSL-F (Intel® C232 chipset) everything is working well.
>>>>>>>>>>>
>>>>>>>>>>> Tested environments:
>>>>>>>>>>> Debian 9.7 - Linux 4.9.0-6-amd64 #1 SMP Debian
>>>>>>>>>>> 4.9.88-1+deb9u1 (2018-05-07) x86_64 GNU/Linux [all platforms
>>>>>>>>>>> working well with no problems: Supermicro X9SCL (Intel C202
>>>>>>>>>>> PCH), Supermicro X10SLL+-F (Intel C222 Express PCH),
>>>>>>>>>>> Supermicro X11SSL-F (Intel® C232 chipset)]
>>>>>>>>>>
>>>>>>>>>>> Debian 11.5 - Linux 5.10.0-19-amd64 #1 SMP Debian 5.10.149-2
>>>>>>>>>>> (2022-10-21) x86_64 GNU/Linux [older platforms: Supermicro
>>>>>>>>>>> X9SCL (Intel C202 PCH), Supermicro X10SLL+-F (Intel C222
>>>>>>>>>>> Express PCH) behave problematic as described above | newer
>>>>>>>>>>> platform: Supermicro X11SSL-F (Intel® C232 chipset) working
>>>>>>>>>>> well with no problems]
>>>>>>>>>>
>>>>>>>>>> Maybe create a bug at the Linux kernel bug tracker [1], where
>>>>>>>>>> you can attach all the logs (`dmesg`, `lspci -nnk -s …`, …).
>>>>>>>>>>
>>>>>>>>> I`ve already reported that to the Debian team
>>>>>>>>> ttps://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1024763, but
>>>>>>>>> so far nobody took care of this issue so far.
>>>>>>>>>
>>>>>>>>>>> So far to solve the problem I was trying to upgrade system to
>>>>>>>>>>> the newest stable version, upgrade kernel to version 6.x,
>>>>>>>>>>> upgrade ixgbe driver to the newest version but with no luck.
>>>>>>>>>>
>>>>>>>>>> Thank you for checking that. Too bad it’s still present. To
>>>>>>>>>> rule out some user space problem, could you test Debian 9.7
>>>>>>>>>> with a stable Linux release, currently 6.1.7?
>>>>>>>>>>
>>>>>>>>>> What does `sudo perf top --sort comm,dso` show, where the time
>>>>>>>>>> is spent?
>>>>>>>>>
>>>>>>>>> During my first test in real enviroment with subscribers I
>>>>>>>>> gether the following data through the perf:
>>>>>>>>>
>>>>>>>>>   27.83%  [kernel]                   [k] strncpy
>>>>>>>>>   14.80%  [kernel]                   [k] nft_do_chain
>>>>>>>>>    7.61%  [kernel]                   [k] memcmp
>>>>>>>>>    5.63%  [kernel]                   [k] nft_meta_get_eval
>>>>>>>>>    3.14%  [kernel]                   [k] nft_cmp_eval
>>>>>>>>>    2.79%  [kernel]                   [k] asm_exc_nmi
>>>>>>>>>    1.07%  [kernel]                   [k] module_get_kallsym
>>>>>>>>>    0.92%  [kernel]                   [k]
>>>>>>>>> kallsyms_expand_symbol.constprop.0
>>>>>>>>>    0.85%  [kernel]                   [k] ixgbe_poll
>>>>>>>>>    0.75%  [kernel]                   [k] format_decode
>>>>>>>>>    0.61%  [kernel]                   [k] number
>>>>>>>>>    0.56%  [kernel]                   [k] menu_select
>>>>>>>>>    0.54%  [kernel]                   [k] clflush_cache_range
>>>>>>>>>    0.52%  [kernel]                   [k] cpuidle_enter_state
>>>>>>>>>    0.51%  [kernel]                   [k] vsnprintf
>>>>>>>>>    0.50%  [kernel]                   [k] u32_classify
>>>>>>>>>    0.49%  [kernel]                   [k] fib_table_lookup
>>>>>>>>>    0.40%  [kernel]                   [k] dma_pte_clear_level
>>>>>>>>>    0.39%  [kernel]                   [k] domain_mapping
>>>>>>>>>    0.36%  [kernel]                   [k] ixgbe_xmit_fram
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>     PID USER      PR  NI    VIRT    RES    SHR S  %CPU %MEM
>>>>>>>>> TIME+ COMMAND
>>>>>>>>>      18 root      20   0       0      0      0 S  28.2 0.0
>>>>>>>>> 7:06.27 ksoftirqd/1
>>>>>>>>>      12 root      20   0       0      0      0 R  12.0 0.0
>>>>>>>>> 4:10.88 ksoftirqd/0
>>>>>>>
>>>>>>> […]
>>>>>>>
>>>>>>> Do you see different behavior in `/proc/interrupts`?
>>>>>>>
>>>>>> This is how it looks like for Debian 11.5 - Linux 5.10.0-19-amd64
>>>>>> #1 SMP Debian 5.10.149-2 (2022-10-21) x86_64 GNU/Linux on
>>>>>> Supermicro X10SLL+-F (Intel C222 Express PCH):
>>>>>>
>>>>>>        1 root      20   0  163948  10288   7696 S   0.0 0.1
>>>>>> 0:39.58 systemd
>>>>>
>>>>> […]
>>>>>
>>>>> The content of `/proc/interrupts` has a different format on my system.
>>>>>
>>>>> ```
>>>>> $ head -3 /proc/interrupts
>>>>>            CPU0       CPU1       CPU2       CPU3
>>>>>   1:      55560          0        113          0  IR-IO-APIC 1-edge
>>>>> i8042
>>>>>   8:          0          0          0          0  IR-IO-APIC 8-edge
>>>>> rtc0
>>>>> ```
>>>>> […]
>>>>>
>>>>>> and for Debian 9.7 - Linux 4.9.0-6-amd64 #1 SMP Debian
>>>>>> 4.9.88-1+deb9u1 on Supermicro X10SLL+-F (Intel C222 Express PCH)
>>>>>>
>>>>>> 31659 root      20   0       0      0      0 S   0.3  0.0 0:00.92
>>>>>> kworker/7:0
>>>>>>      1 root      20   0   57032   6736   5256 S   0.0  0.1 2:28.14
>>>>>> systemd
>>>>>
>>>>> […]
>>>>>>>>>>> Supermicro support suggested as follows:
>>>>>>>>>>> it might be kernel related debian 11.5 has kernel 5.10 which
>>>>>>>>>>> is a recent kernel it might not properly support the chipsets
>>>>>>>>>>> for X9 therefore i suggest to use RHEL or CentOS as they use
>>>>>>>>>>> much older kernel versions. I expect that with ubuntu 20.04
>>>>>>>>>>> you see the same problem it uses kernel 5.4
>>>>>>>>>> >>> Testing another GNU/Linux distribution for another data
>>>>>>>>>> point, might be a good idea.
>>>>>>>>>>
>>>>>>>>>> As nobody has responded yet, bisecting the issue is probably
>>>>>>>>>> the fastest way to get to the bottom of this. Luckily the
>>>>>>>>>> problem seems reproducible and you seem to be able to build a
>>>>>>>>>> Linux kernel yourself, so that should work. (For testing
>>>>>>>>>> purposes you could also test with Ubuntu, as they provide
>>>>>>>>>> Linux kernel builds for (almost) all releases in their Linux
>>>>>>>>>> kernel mainline PPA [2].)
>>>>>>>>>>
>>>>>>>>> Of course  I can try Ubuntu and report how it is working.
>>>>>>>>>
>>>>>>>> Ubuntu (5.15.0-43-generic) seems to be working in the same way
>>>>>>>> generating higher load after executing "ip link set enp1s0 up".
>>>>>>>
>>>>>>> That is good to know. (Is this Ubuntu 22.04?) What about Ubuntu
>>>>>>> 20.04 with Linux 5.4, and Ubuntu 18.04 with 4.15?
>>>>>>>
>>>>>>> Anyway, I think, you won’t come around bisecting. Another hint,
>>>>>>> make sure that you can build a 4.9 Linux kernel yourself, that
>>>>>>> does not exhibit that issue.
>>>>>>>
>>>>>> That`s right, it is 22.04. I don`t have to build it. Standard
>>>>>> kernel Linux 4.9.0-6-amd64 from Debian 9.7 worked without problems
>>>>>> for past 4 years.
>>>>>
>>>>> If nobody of the developers/maintainers is going to step up, you
>>>>> are on your own. Again, as you can reproduce this easily, the
>>>>> fastest way is to bisect the issue, which you can do on your own.
>>>>
>>>> How can I investigate that further?
>>>
>>> I repeat myself, please bisect the issue. It’s the fastest way.
>>>
>>>> I thought about trying to change some of the parameters related to
>>>> ixgbe driver and observe if anything is changing, but when I am
>>>> trying to do:
>>>>
>>>> sudo modprobe ixgbe IntMode=0
>>>>
>>>> I get the following error in the dmesg:
>>>>
>>>> [ 2137.324772] ixgbe: unknown parameter 'IntMode' ignored <<<<<<<<<
>>>
>>> […]
>>>
>>> `modinfo ixgbe` shows the supported parameters.
> 
>>> PS: If you need help bisecting, please ask. Otherwise, I am out of
>>> this thread.
>>
>> Ok, how exactly I can bisect this issue?
> 
> What have you tried so far? As written in the past, I’d first try more
> distributions, for example, older Ubuntu versions. Then, if you have
> some range, I’d use the Ubuntu PPA, and then between the release
> candidate versions, only then start doing `git bisect` as documented in
> the documentation [3].

Hmmm. I'm not an expert in that area, but if you follow Paul's advice
keep in mind that a deliberate config change by the distro might have an
impact here. Hence it might be a good idea to rule that out first by
taking a config from a working kernel and using it (with the help of
"make olddefconfig") to build your own kernel from the version that is
known to fail. But over such a wide range of versions this can be
tricky. :-/

But apart from that Paul is right afaics: nobody yet had an idea what
might cause this regression, hence we need a bisection to pin-point the
problem.

Ciao, Thorsten (wearing his 'the Linux kernel's regression tracker' hat)
--
Everything you wanna know about Linux kernel regression tracking:
https://linux-regtracking.leemhuis.info/about/#tldr
If I did something stupid, please tell me, as explained on that page.



>>>>>>>>>> [1]: https://bugzilla.kernel.org/
>>>>>>>>>> [2]: https://kernel.ubuntu.com/~kernel-ppa/mainline/
> [3]: https://www.kernel.org/doc/html/latest/admin-guide/bug-bisect.html
> 
>