Date:   Tue, 14 Jun 2022 08:42:34 +0800
From:   Feng Tang <feng.tang@...el.com>
To:     Eric Dumazet <edumazet@...gle.com>
Cc:     Willy Tarreau <w@....eu>, Moshe Kol <moshe.kol@...l.huji.ac.il>,
        fengwei.yin@...el.com, kernel test robot <oliver.sang@...el.com>,
        Jakub Kicinski <kuba@...nel.org>,
        Yossi Gilad <yossi.gilad@...l.huji.ac.il>,
        Amit Klein <aksecurity@...il.com>,
        LKML <linux-kernel@...r.kernel.org>,
        netdev <netdev@...r.kernel.org>, lkp@...ts.01.org,
        kbuild test robot <lkp@...el.com>,
        "Huang, Ying" <ying.huang@...el.com>, zhengjun.xing@...ux.intel.com
Subject: Re: [tcp] e926147618: stress-ng.icmp-flood.ops_per_sec -8.7%
 regression

On Mon, Jun 13, 2022 at 10:10:57AM -0700, Eric Dumazet wrote:
> On Sun, Jun 12, 2022 at 7:09 PM Feng Tang <feng.tang@...el.com> wrote:
> >
> > Hi,
> >
> > On Wed, Jun 08, 2022 at 09:34:41AM +0200, Willy Tarreau wrote:
> > > On Wed, Jun 08, 2022 at 10:26:12AM +0300, Moshe Kol wrote:
> > > > Hmm, How is the ICMP flood stress test related to TCP connections?
> > >
> > > To me it's not directly related, unless the test pre-establishes many
> > > connections, or is affected in a way or another by a larger memory
> > > allocation of this part.
> >
> > Fengwei and I discussed and thought this could be a data alignment
> > related case, that one module's data alignment change affects other
> > modules' alignment, and we had a patch for detecting similar cases [1]
> >
> > After some debugging, this could be related with the bss section
> > alignment changes, that if we forced all module's bss section to be
> > 4KB aligned, then the stress-ng icmp-flood case will have almost no
> > performance difference for the 2 commits:
> >
> > 10025135            +0.8%   10105711 ±  2%  stress-ng.icmp-flood.ops_per_sec
> >
> > The debug patch is:
> >
> > diff --git a/arch/x86/kernel/vmlinux.lds.S b/arch/x86/kernel/vmlinux.lds.S
> > index 7fda7f27e7620..7eb626b98620c 100644
> > --- a/arch/x86/kernel/vmlinux.lds.S
> > +++ b/arch/x86/kernel/vmlinux.lds.S
> > @@ -378,7 +378,9 @@ SECTIONS
> >
> >         /* BSS */
> >         . = ALIGN(PAGE_SIZE);
> > -       .bss : AT(ADDR(.bss) - LOAD_OFFSET) {
> > +       .bss : AT(ADDR(.bss) - LOAD_OFFSET)
> > +       SUBALIGN(PAGE_SIZE)
> > +       {
> >                 __bss_start = .;
> >                 *(.bss..page_aligned)
> >                 . = ALIGN(PAGE_SIZE);
> >
> > The 'table_perturb[]' used to be in bss section, and with the commit
> > of moving it to runtime allocation, other data structures following it
> > in the .bss section will get affected accordingly.
> >
> 
> As the 'regression' is seen with ICMP workload, can you please share with us
> the symbols close to icmp_global (without your align patch)

Here are the info for the 2 commits (kernels for the original reprot):

$ egrep -8 "icmp_global$" System.map-5.18.0-rc4-00242-gca7af0402550
ffffffff837adc00 b tcp_cong_list_lock
ffffffff837adc08 b tcpmhash_entries
ffffffff837adc10 b fastopen_seqlock
ffffffff837adc18 b tcp_metrics_lock
ffffffff837adc1c b tcp_ulp_list_lock
ffffffff837adc20 B raw_v4_hashinfo
ffffffff837ae440 B udp_encap_needed_key
ffffffff837ae480 B udp_memory_allocated
ffffffff837ae488 b icmp_global
ffffffff837ae4a0 b inet_addr_lst
ffffffff837aeca0 b inetsw_lock
ffffffff837aecc0 b inetsw
ffffffff837aed80 b fib_info_devhash
ffffffff837af580 b fib_info_cnt
ffffffff837af584 b fib_info_hash_bits
ffffffff837af588 b fib_info_hash_size
ffffffff837af590 b fib_info_laddrhash

$ egrep -8 "icmp_global$" System.map-5.18.0-rc4-00243-ge9261476184b
ffffffff837ad800 b tcp_cong_list_lock
ffffffff837ad808 b tcpmhash_entries
ffffffff837ad810 b fastopen_seqlock
ffffffff837ad818 b tcp_metrics_lock
ffffffff837ad81c b tcp_ulp_list_lock
ffffffff837ad820 B raw_v4_hashinfo
ffffffff837ae040 B udp_encap_needed_key
ffffffff837ae080 B udp_memory_allocated
ffffffff837ae088 b icmp_global
ffffffff837ae0a0 b inet_addr_lst
ffffffff837ae8a0 b inetsw_lock
ffffffff837ae8c0 b inetsw
ffffffff837ae980 b fib_info_devhash
ffffffff837af180 b fib_info_cnt
ffffffff837af184 b fib_info_hash_bits
ffffffff837af188 b fib_info_hash_size
ffffffff837af190 b fib_info_laddrhash

>From above we can see some symbols are offseted of 0x400 (1KB),
like icmp_global or fib_xxx, as the 'table_perturb' used to be
1KB long in bss section.

> I suspect we should move icmp_global to a dedicated cache line.
> 
> $ nm -v vmlinux|egrep -8 "icmp_global$"
> ffffffff835bc490 b tcp_cong_list_lock
> ffffffff835bc494 b fastopen_seqlock
> ffffffff835bc49c b tcp_metrics_lock
> ffffffff835bc4a0 b tcpmhash_entries
> ffffffff835bc4a4 b tcp_ulp_list_lock
> ffffffff835bc4a8 B raw_v4_hashinfo
> ffffffff835bccc0 B udp_memory_allocated      << Note sure why it is
> not already in a dedicated cache line>>

IIUC, 0x....C0 means it is already cacheline aligned?


> ffffffff835bccc8 B udp_encap_needed_key
> ffffffff835bccd8 b icmp_global                               <<<HERE>>
> ffffffff835bccf0 b inet_addr_lst
> ffffffff835bd4f0 b inetsw_lock
> ffffffff835bd500 b inetsw
> ffffffff835bd5b0 b fib_info_lock
> ffffffff835bd5b4 b fib_info_cnt
> ffffffff835bd5b8 b fib_info_hash_size
> ffffffff835bd5c0 b fib_info_hash
> ffffffff835bd5c8 b fib_info_laddrhash

Fengwei found that from the perf-profile info of report, there are
quite difference around ping table handling.

     43.72            +4.1       47.81 ±  2% icmp_rcv.ip_protocol_deliver_rcu.ip_local_deliver_finish.__netif_receive_skb_one_core.process_backlog
     17.10 ±  7%      +6.0       23.10 ±  6%ping_rcv.icmp_rcv.ip_protocol_deliver_rcu.ip_local_deliver_finish.__netif_receive_skb_one_core
     17.05 ±  7%      +6.0       23.06 ±  6%ping_lookup.ping_rcv.icmp_rcv.ip_protocol_deliver_rcu.ip_local_deliver_finish

And the "ping_table[]" also sits in bss section and got affected:

  Before: 
  ffffffff837af600 b ping_table

  After:
  ffffffff837af200 b ping_table

The cacheline alignment is not changed, and I suspect it cause some
TLB changes.

Thanks,
Feng

Powered by blists - more mailing lists