Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Subject: Re: The value of FB_MTU eats two pages
To:     Menglong Dong <menglong8.dong@gmail.com>
Cc:     ying.xue@windriver.com, David Miller <davem@davemloft.net>,
        Jakub Kicinski <kuba@kernel.org>, etdev@vger.kernel.org,
        tipc-discussion@lists.sourceforge.net,
        LKML <linux-kernel@vger.kernel.org>
References: <CADxym3baupJJ7Q9otxtoQ-DH5e-J2isg-LZj2CsOqRPo70AL4A@mail.gmail.com>
From:   Jon Maloy <jmaloy@redhat.com>
Message-ID: <e91baaba-e00a-4b16-0787-e9460dacfbb9@redhat.com>
Date:   Wed, 2 Jun 2021 15:50:27 -0400
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101
 Thunderbird/78.10.1
MIME-Version: 1.0
In-Reply-To: <CADxym3baupJJ7Q9otxtoQ-DH5e-J2isg-LZj2CsOqRPo70AL4A@mail.gmail.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Transfer-Encoding: 7bit
Content-Language: en-US
Precedence: bulk


On 6/1/21 10:18 AM, Menglong Dong wrote:
> Hello!
>
> I have a question about the value of FB_MTU in tipc, how does the '3744' form?
> I notice that it is used in 'tipc_msg_build()' when memory allocation
> fails, and it
> tries to fall back to a smaller MTU to avoid unnecessary sending failures.
>
> However, the size of the data allocated will be more than 4096 when FB_MTU
> is 3744. I did a rough calculation, the size of data will more than 4200:
>
> (FB_MTU + TIPCHDR + BUF_HEADROOM + sizeof(struct skb_shared_info))
>
> Therefore, 8192 will be allocated from slab, and about 4000 of it will
> not be used.
>
> FB_MTU is used for low memory, and I think eating two pages will make it worse.
> Do I miss something?
>
> Thanks!
> Menglong Dong
>
Hi Dong,
The value is based on empiric knowledge.
When I determined it I made a small loop in a kernel driver where I 
allocated skbs (using tipc_buf_acquire) with an increasing size 
(incremented with 1 each iteration), and then printed out the 
corresponding truesize.

That gave the value we are using now.

Now, when re-running the test I get a different value, so something has 
obviously changed since then.

[ 1622.158586] skb(513) =>> truesize 2304, prev skb(512) => prev 
truesize 1280
[ 1622.162074] skb(1537) =>> truesize 4352, prev skb(1536) => prev 
truesize 2304
[ 1622.165984] skb(3585) =>> truesize 8448, prev skb(3584) => prev 
truesize 4352

As you can see, the optimal value now, for an x86_64 machine compiled 
with gcc, is 3584 bytes, not 3744.

Feel free to post a patch for this if you want to.

Thanks
///Jon Maloy