netdev.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Jesper Dangaard Brouer <brouer@redhat.com>
To: Brenden Blanco <bblanco@plumgrid.com>
Cc: davem@davemloft.net, netdev@vger.kernel.org, tom@herbertland.com,
	alexei.starovoitov@gmail.com, Or Gerlitz <ogerlitz@mellanox.com>,
	daniel@iogearbox.net, john.fastabend@gmail.com,
	brouer@redhat.com
Subject: Re: [RFC PATCH 5/5] Add sample for adding simple drop program to link
Date: Wed, 6 Apr 2016 22:01:00 +0200	[thread overview]
Message-ID: <20160406220100.0df04925@redhat.com> (raw)
In-Reply-To: <20160406214848.7568235b@redhat.com>

On Wed, 6 Apr 2016 21:48:48 +0200
Jesper Dangaard Brouer <brouer@redhat.com> wrote:

> I'm testing with this program and these patches, after getting past the
> challenge of compiling the samples/bpf files ;-)
> 
> 
> On Fri,  1 Apr 2016 18:21:58 -0700 Brenden Blanco <bblanco@plumgrid.com> wrote:
> 
> > Add a sample program that only drops packets at the
> > BPF_PROG_TYPE_PHYS_DEV hook of a link. With the drop-only program,
> > observed single core rate is ~14.6Mpps.  
> 
> On my i7-4790K CPU @ 4.00GHz I'm seeing 9.7Mpps (single flow/cpu).
> (generator: pktgen_sample03_burst_single_flow.sh)
> 
>  # ./netdrvx1 $(</sys/class/net/mlx4p1/ifindex)
>  sh: /sys/kernel/debug/tracing/kprobe_events: No such file or directory
>  Success: Loaded file ./netdrvx1_kern.o
>  proto 17:    9776320 drops/s
> 
> These numbers are quite impressive. Compared to: sending it to local
> socket that drop packets 1.7Mpps. Compared to: dropping with iptables
> in "raw" table 3.7Mpps.
> 
> If I do multiple flows, via ./pktgen_sample05_flow_per_thread.sh
> then I hit this strange 14.5Mpps limit (proto 17:   14505558 drops/s).
> And the RX 4x CPUs are starting to NOT use 100% in softirq, they have
> some cycles attributed to %idle. (I verified generator is sending at
> 24Mpps).
> 
> 
> > Other tests were run, for instance without the dropcnt increment or
> > without reading from the packet header, the packet rate was mostly
> > unchanged.  
> 
> If I change the program to not touch packet data (don't call
> load_byte()) then the performance increase to 14.6Mpps (single
> flow/cpu).  And the RX CPU is mostly idle... mlx4_en_process_rx_cq()
> and page alloc/free functions taking the time.
> 
> > $ perf record -a samples/bpf/netdrvx1 $(</sys/class/net/eth0/ifindex)
> > proto 17:   14597724 drops/s
> > 
> > ./pktgen_sample03_burst_single_flow.sh -i $DEV -d $IP -m $MAC -t 4
> > Running... ctrl^C to stop
> > Device: eth4@0
> > Result: OK: 6486875(c6485849+d1026) usec, 23689465 (60byte,0frags)
> >   3651906pps 1752Mb/sec (1752914880bps) errors: 0
> > Device: eth4@1
> > Result: OK: 6486874(c6485656+d1217) usec, 23689489 (60byte,0frags)
> >   3651911pps 1752Mb/sec (1752917280bps) errors: 0
> > Device: eth4@2
> > Result: OK: 6486851(c6485730+d1120) usec, 23687853 (60byte,0frags)
> >   3651672pps 1752Mb/sec (1752802560bps) errors: 0
> > Device: eth4@3
> > Result: OK: 6486879(c6485807+d1071) usec, 23688954 (60byte,0frags)
> >   3651825pps 1752Mb/sec (1752876000bps) errors: 0
> > 
> > perf report --no-children:
> >   18.36%  ksoftirqd/1    [mlx4_en]         [k] mlx4_en_process_rx_cq
> >   15.98%  swapper        [kernel.vmlinux]  [k] poll_idle
> >   12.71%  ksoftirqd/1    [mlx4_en]         [k] mlx4_en_alloc_frags
> >    6.87%  ksoftirqd/1    [mlx4_en]         [k] mlx4_en_free_frag
> >    4.20%  ksoftirqd/1    [kernel.vmlinux]  [k] get_page_from_freelist
> >    4.09%  swapper        [mlx4_en]         [k] mlx4_en_process_rx_cq
> >    3.32%  ksoftirqd/1    [kernel.vmlinux]  [k] sk_load_byte_positive_offset
> >    2.39%  ksoftirqd/1    [mdio]            [k] 0x00000000000074cd
> >    2.23%  swapper        [mlx4_en]         [k] mlx4_en_alloc_frags
> >    2.20%  ksoftirqd/1    [kernel.vmlinux]  [k] free_pages_prepare
> >    2.08%  ksoftirqd/1    [mlx4_en]         [k] mlx4_call_bpf
> >    1.57%  ksoftirqd/1    [kernel.vmlinux]  [k] percpu_array_map_lookup_elem
> >    1.35%  ksoftirqd/1    [mdio]            [k] 0x00000000000074fa
> >    1.09%  ksoftirqd/1    [kernel.vmlinux]  [k] free_one_page
> >    1.02%  ksoftirqd/1    [kernel.vmlinux]  [k] bpf_map_lookup_elem
> >    0.90%  ksoftirqd/1    [kernel.vmlinux]  [k] __alloc_pages_nodemask
> >    0.88%  swapper        [kernel.vmlinux]  [k] intel_idle
> >    0.82%  ksoftirqd/1    [mdio]            [k] 0x00000000000074be
> >    0.80%  swapper        [mlx4_en]         [k] mlx4_en_free_frag  
> 
> My picture (single flow/cpu) looks a little bit different:
> 
>  +   64.33%  ksoftirqd/7    [kernel.vmlinux]  [k] __bpf_prog_run
>  +    9.60%  ksoftirqd/7    [mlx4_en]         [k] mlx4_en_alloc_frags
>  +    7.71%  ksoftirqd/7    [mlx4_en]         [k] mlx4_en_process_rx_cq
>  +    5.47%  ksoftirqd/7    [mlx4_en]         [k] mlx4_en_free_frag
>  +    1.68%  ksoftirqd/7    [kernel.vmlinux]  [k] get_page_from_freelist
>  +    1.52%  ksoftirqd/7    [mlx4_en]         [k] mlx4_call_bpf
>  +    1.02%  ksoftirqd/7    [kernel.vmlinux]  [k] free_pages_prepare
>  +    0.72%  ksoftirqd/7    [mlx4_en]         [k] mlx4_alloc_pages.isra.20
>  +    0.70%  ksoftirqd/7    [kernel.vmlinux]  [k] __rcu_read_unlock
>  +    0.65%  ksoftirqd/7    [kernel.vmlinux]  [k] percpu_array_map_lookup_elem
> 
> On my i7-4790K CPU, I don't have DDIO, thus I assume this high cost in
> __bpf_prog_run is due to a cache-miss on the packet data.

Before someone else point out the obvious... I forgot to enable JIT.
Enable it::

 # echo 1 > /proc/sys/net/core/bpf_jit_enable

Performance increased to: 10.8Mpps (proto 17:   10819446 drops/s)

 Samples: 51K of event 'cycles', Event count (approx.): 56775706510
   Overhead  Command      Shared Object     Symbol
 +   55.90%  ksoftirqd/7  [kernel.vmlinux]  [k] sk_load_byte_positive_offset
 +   10.71%  ksoftirqd/7  [mlx4_en]         [k] mlx4_en_alloc_frags
 +    8.26%  ksoftirqd/7  [mlx4_en]         [k] mlx4_en_process_rx_cq
 +    5.94%  ksoftirqd/7  [mlx4_en]         [k] mlx4_en_free_frag
 +    2.04%  ksoftirqd/7  [kernel.vmlinux]  [k] get_page_from_freelist
 +    2.03%  ksoftirqd/7  [kernel.vmlinux]  [k] percpu_array_map_lookup_elem
 +    1.42%  ksoftirqd/7  [mlx4_en]         [k] mlx4_call_bpf
 +    1.04%  ksoftirqd/7  [kernel.vmlinux]  [k] free_pages_prepare
 +    1.03%  ksoftirqd/7  [kernel.vmlinux]  [k] __rcu_read_unlock
 +    0.97%  ksoftirqd/7  [mlx4_en]         [k] mlx4_alloc_pages.isra.20
 +    0.95%  ksoftirqd/7  [devlink]         [k] 0x0000000000005f87
 +    0.58%  ksoftirqd/7  [devlink]         [k] 0x0000000000005f8f
 +    0.49%  ksoftirqd/7  [kernel.vmlinux]  [k] __free_pages_ok
 +    0.47%  ksoftirqd/7  [kernel.vmlinux]  [k] __rcu_read_lock
 +    0.46%  ksoftirqd/7  [kernel.vmlinux]  [k] free_one_page
 +    0.38%  ksoftirqd/7  [kernel.vmlinux]  [k] net_rx_action
 +    0.36%  ksoftirqd/7  [kernel.vmlinux]  [k] bpf_map_lookup_elem
 +    0.36%  ksoftirqd/7  [kernel.vmlinux]  [k] __mod_zone_page_state
 +    0.34%  ksoftirqd/7  [kernel.vmlinux]  [k] __alloc_pages_nodemask
 +    0.32%  ksoftirqd/7  [kernel.vmlinux]  [k] _raw_spin_lock
 +    0.31%  ksoftirqd/7  [devlink]         [k] 0x0000000000005f0a
 +    0.29%  ksoftirqd/7  [kernel.vmlinux]  [k] next_zones_zonelist

It is a very likely cache-miss in sk_load_byte_positive_offset().

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  Author of http://www.iptv-analyzer.org
  LinkedIn: http://www.linkedin.com/in/brouer

  reply	other threads:[~2016-04-06 20:01 UTC|newest]

Thread overview: 62+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2016-04-02  1:21 [RFC PATCH 0/5] Add driver bpf hook for early packet drop Brenden Blanco
2016-04-02  1:21 ` [RFC PATCH 1/5] bpf: add PHYS_DEV prog type for early driver filter Brenden Blanco
2016-04-02 16:39   ` Tom Herbert
2016-04-03  7:02     ` Brenden Blanco
2016-04-04 22:07       ` Thomas Graf
2016-04-05  8:19         ` Jesper Dangaard Brouer
2016-04-04  8:49   ` Daniel Borkmann
2016-04-04 13:07     ` Jesper Dangaard Brouer
2016-04-04 13:36       ` Daniel Borkmann
2016-04-04 14:09         ` Tom Herbert
2016-04-04 15:12           ` Jesper Dangaard Brouer
2016-04-04 15:29             ` Brenden Blanco
2016-04-04 16:07               ` John Fastabend
2016-04-04 16:17                 ` Brenden Blanco
2016-04-04 20:00                   ` Alexei Starovoitov
2016-04-04 22:04                     ` Thomas Graf
2016-04-05  2:25                       ` Alexei Starovoitov
2016-04-05  8:11                         ` Jesper Dangaard Brouer
2016-04-05  9:29                     ` Jesper Dangaard Brouer
2016-04-05 22:06                       ` Alexei Starovoitov
2016-04-04 14:33       ` Eric Dumazet
2016-04-04 15:18         ` Edward Cree
2016-04-02  1:21 ` [RFC PATCH 2/5] net: add ndo to set bpf prog in adapter rx Brenden Blanco
2016-04-02  1:21 ` [RFC PATCH 3/5] rtnl: add option for setting link bpf prog Brenden Blanco
2016-04-02  1:21 ` [RFC PATCH 4/5] mlx4: add support for fast rx drop bpf program Brenden Blanco
2016-04-02  2:08   ` Eric Dumazet
2016-04-02  2:47     ` Alexei Starovoitov
2016-04-04 14:57       ` Jesper Dangaard Brouer
2016-04-04 15:22         ` Eric Dumazet
2016-04-04 18:50           ` Alexei Starovoitov
2016-04-05 14:15             ` Or Gerlitz
2016-04-06  4:05               ` Brenden Blanco
2016-04-03  6:15     ` Brenden Blanco
2016-04-05  2:20       ` Brenden Blanco
2016-04-05  2:44         ` Eric Dumazet
2016-04-05 18:59         ` Eran Ben Elisha
2016-04-02  8:23   ` Jesper Dangaard Brouer
2016-04-03  6:11     ` Brenden Blanco
2016-04-04 18:27       ` Alexei Starovoitov
2016-04-05  6:04         ` Jesper Dangaard Brouer
2016-04-02 18:40   ` Johannes Berg
2016-04-03  6:38     ` Brenden Blanco
2016-04-04  7:35       ` Johannes Berg
2016-04-04  9:57         ` Daniel Borkmann
2016-04-04 18:46           ` Alexei Starovoitov
2016-04-04 21:01             ` Daniel Borkmann
2016-04-05  1:17               ` Alexei Starovoitov
2016-04-04  8:33   ` Jesper Dangaard Brouer
2016-04-04  9:22   ` Daniel Borkmann
2016-04-02  1:21 ` [RFC PATCH 5/5] Add sample for adding simple drop program to link Brenden Blanco
2016-04-06 19:48   ` Jesper Dangaard Brouer
2016-04-06 20:01     ` Jesper Dangaard Brouer [this message]
2016-04-06 23:11       ` Alexei Starovoitov
2016-04-06 20:03     ` Daniel Borkmann
2016-04-02 16:47 ` [RFC PATCH 0/5] Add driver bpf hook for early packet drop Tom Herbert
2016-04-03  5:41   ` Brenden Blanco
2016-04-04  7:48     ` Jesper Dangaard Brouer
2016-04-04 18:10       ` Alexei Starovoitov
2016-04-02 18:41 ` Johannes Berg
2016-04-02 22:57   ` Tom Herbert
2016-04-03  2:28     ` Lorenzo Colitti
2016-04-04  7:37       ` Johannes Berg

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20160406220100.0df04925@redhat.com \
    --to=brouer@redhat.com \
    --cc=alexei.starovoitov@gmail.com \
    --cc=bblanco@plumgrid.com \
    --cc=daniel@iogearbox.net \
    --cc=davem@davemloft.net \
    --cc=john.fastabend@gmail.com \
    --cc=netdev@vger.kernel.org \
    --cc=ogerlitz@mellanox.com \
    --cc=tom@herbertland.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).