Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
In-Reply-To: <a48538cf40c1645669326c92d9600fc98a13a260.camel@redhat.com>
References: <CAGXu5jL=aEXHKr5ouVdSKwG-y7xSQFLi=x1nwSjFspYiyKL1Pw@mail.gmail.com>
 <64bf81fa-0363-4b46-d8da-94285b592caa@redhat.com> <a48538cf40c1645669326c92d9600fc98a13a260.camel@redhat.com>
From:   Kees Cook <keescook@chromium.org>
Date:   Tue, 26 Jun 2018 09:54:59 -0700
Message-ID: <CAGXu5jKHz=OaU1ejYEB=t-=Gs6gVoRywFbyQw8ThHk6WYG7Qxg@mail.gmail.com>
Subject: Re: [PATCH] add param that allows bootline control of hardened usercopy
To:     Paolo Abeni <pabeni@redhat.com>
Cc:     Chris von Recklinghausen <crecklin@redhat.com>,
        LKML <linux-kernel@vger.kernel.org>, linux-mm@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Tue, Jun 26, 2018 at 2:48 AM, Paolo Abeni <pabeni@redhat.com> wrote:
> Hi,
>
> On Mon, 25 Jun 2018 11:21:38 -0700 Kees Cook <keescook@chromium.org> wrote:
>> On Mon, Jun 25, 2018 at 8:08 AM, Chris von Recklinghausen
>> <crecklin@redhat.com> wrote:
>> > Enabling HARDENED_USER_COPY causes measurable regressions in the
>> > networking performances, up to 8% under UDP flood.
>>
>> Which function is "hot"? i.e. which copy*user() is taking up the time?
>> Do you have a workload that at can be used to reproduce the problem?
>
> I'm running an a small packet UDP flood using pktgen vs. an host b2b
> connected. On the receiver side the UDP packets are processed by a
> simple user space process that just read and drop them:
>
> https://github.com/netoptimizer/network-testing/blob/master/src/udp_sink.c
>
> Not very useful from a functional PoV, it helps mostly pin-pointing
> bottle-neck in the networking stack.

Cool; thanks for the pointer!

> When running a kernel with CONFIG_HARDENED_USERCOPY=y, I see a 5-8%
> regression in the receive tput, compared to the same kernel without
> such option.
>
> With CONFIG_HARDENED_USERCOPY=y, perf shows ~6% of CPU time spent
> cumulatively in __check_object_size (~4%) and __virt_addr_valid (~2%).

Are you able to see which network functions are making the
__check_object_size() calls?

-Kees

-- 
Kees Cook
Pixel Security