netdev - Re: [RFC PATCH 5/5] Add sample for adding simple drop program to link

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <20160406231147.GA11421@ast-mbp.thefacebook.com>
Date:	Wed, 6 Apr 2016 16:11:49 -0700
From:	Alexei Starovoitov <alexei.starovoitov@...il.com>
To:	Jesper Dangaard Brouer <brouer@...hat.com>
Cc:	Brenden Blanco <bblanco@...mgrid.com>, davem@...emloft.net,
	netdev@...r.kernel.org, tom@...bertland.com,
	Or Gerlitz <ogerlitz@...lanox.com>, daniel@...earbox.net,
	john.fastabend@...il.com
Subject: Re: [RFC PATCH 5/5] Add sample for adding simple drop program to link

On Wed, Apr 06, 2016 at 10:01:00PM +0200, Jesper Dangaard Brouer wrote:
> On Wed, 6 Apr 2016 21:48:48 +0200
> Jesper Dangaard Brouer <brouer@...hat.com> wrote:
> > 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).
...
> > 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.

Please try it with module param log_num_mgm_entry_size=-1
It should get to 20Mpps when bpf doesn't touch the packet.

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

yes, likely due to missing ddio as you said.