Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Jesper Dangaard Brouer <jbrouer@redhat.com>
Message-ID: <3e6c191a-3be3-d6ff-92a2-2685bade2e66@redhat.com>
Date:   Wed, 21 Jun 2023 13:55:00 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.10.0
Cc:     brouer@redhat.com, Jesper Dangaard Brouer <jbrouer@redhat.com>,
        Yunsheng Lin <linyunsheng@huawei.com>, davem@davemloft.net,
        kuba@kernel.org, pabeni@redhat.com, netdev@vger.kernel.org,
        linux-kernel@vger.kernel.org,
        Jesper Dangaard Brouer <hawk@kernel.org>,
        Ilias Apalodimas <ilias.apalodimas@linaro.org>,
        Eric Dumazet <edumazet@google.com>,
        Maryam Tahhan <mtahhan@redhat.com>, bpf <bpf@vger.kernel.org>
Subject: Re: [PATCH net-next v3 3/4] page_pool: introduce page_pool_alloc()
 API
Content-Language: en-US
To:     Lorenzo Bianconi <lorenzo@kernel.org>,
        Alexander Duyck <alexander.duyck@gmail.com>
References: <c06f6f59-6c35-4944-8f7a-7f6f0e076649@huawei.com>
 <CAKgT0UccmDe+CE6=zDYQHi1=3vXf5MptzDo+BsPrKdmP5j9kgQ@mail.gmail.com>
 <0ba1bf9c-2e45-cd44-60d3-66feeb3268f3@redhat.com>
 <dcc9db4c-207b-e118-3d84-641677cd3d80@huawei.com>
 <f8ce176f-f975-af11-641c-b56c53a8066a@redhat.com>
 <CAKgT0UfzP30OiBQu+YKefLD+=32t+oA6KGzkvsW6k7CMTXU8KA@mail.gmail.com>
 <699563f5-c4fa-0246-5e79-61a29e1a8db3@redhat.com>
 <CAKgT0UcNOYwxRP_zkaBaZh-VBL-CriL8dFG-VY7-FUyzxfHDWw@mail.gmail.com>
 <ZI8dP5+guKdR7IFE@lore-desk>
 <CAKgT0UfFVFa4zT2DnPZEGaHp0uh5V1u1aGymgdL4Vu8Q1VV8hQ@mail.gmail.com>
 <ZJIXSyjxPf7FQQKo@lore-rh-laptop>
In-Reply-To: <ZJIXSyjxPf7FQQKo@lore-rh-laptop>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk


On 20/06/2023 23.16, Lorenzo Bianconi wrote:
> [...]
> 
>>> I did some experiments using page_frag_cache/page_frag_alloc() instead of
>>> page_pools in a simple environment I used to test XDP for veth driver.
>>> In particular, I allocate a new buffer in veth_convert_skb_to_xdp_buff() from
>>> the page_frag_cache in order to copy the full skb in the new one, actually
>>> "linearizing" the packet (since we know the original skb length).
>>> I run an iperf TCP connection over a veth pair where the
>>> remote device runs the xdp_rxq_info sample (available in the kernel source
>>> tree, with action XDP_PASS):
>>>
>>> TCP clietn -- v0 === v1 (xdp_rxq_info) -- TCP server
>>>
>>> net-next (page_pool):
>>> - MTU 1500B: ~  7.5 Gbps
>>> - MTU 8000B: ~ 15.3 Gbps
>>>
>>> net-next + page_frag_alloc:
>>> - MTU 1500B: ~  8.4 Gbps
>>> - MTU 8000B: ~ 14.7 Gbps
>>>
>>> It seems there is no a clear "win" situation here (at least in this environment
>>> and we this simple approach). Moreover:
>>
>> For the 1500B packets it is a win, but for 8000B it looks like there
>> is a regression. Any idea what is causing it?
> 
> nope, I have not looked into it yet.
> 

I think I can explain via using micro-benchmark numbers.
(Lorenzo and I have discussed this over IRC, so this is our summary)

*** MTU 1500***

* The MTU 1500 case, where page_frag_alloc is faster than PP (page_pool):

The PP alloc a 4K page for MTU 1500. The cost of alloc + recycle via
ptr_ring cost 48 cycles (page_pool02_ptr_ring Per elem: 48 cycles(tsc)).

The page_frag_alloc API allocates a 32KB order-3 page, and chops it up
for packets.  The order-3 alloc + free cost 514 cycles (page_bench01:
alloc_pages order:3(32768B) 514 cycles). The MTU 1500 needs alloc size
1514+320+256 = 2090 bytes.  In 32KB we can fit 15 packets.  Thus, the
amortized cost per packet is only 34.3 cycles (514/15).

Thus, this explains why page_frag_alloc API have an advantage here, as
amortized cost per packet is lower (for page_frag_alloc).


*** MTU 8000 ***

* The MTU 8000 case, where PP is faster than page_frag_alloc.

The page_frag_alloc API cannot slice the same 32KB into as many packets.
The MTU 8000 needs alloc size 8000+14+256+320 = 8590 bytes.  This is can
only store 3 full packets (32768/8590 = 3.81).
Thus, cost is 514/3 = 171 cycles.

The PP is actually challenged at MTU 8000, because it unfortunately
leads to allocating 3 full pages (12KiB), due to needed alloc size 8590
bytes. Thus cost is 3x 48 cycles = 144 cycles.
(There is also a chance of Jakubs "allow_direct" optimization in 
page_pool_return_skb_page to increase performance for PP).

Thus, this explains why PP is fastest in this case.


*** Surprising insights ***

My (maybe) surprising conclusion is that we should combine the two
approaches.  Which is basically what Lin's patchset is doing!
Thus, I'm actually suddenly become a fan of this patchset...

The insight is that PP can also work with higher-order pages and the
cost of PP recycles via ptr_ring will be the same, regardless of page
order size.  Thus, we can reduced the order-3 cost 514 cycles to
basically 48 cycles, and fit 15 packets (MTU 1500) resulting is
amortized allocator cost 48/15 = 3.2 cycles.

On the PP alloc-side this will be amazingly fast. When PP recycles frags
side, see page_pool_defrag_page() there is an atomic_sub operation.
I've measured atomic_inc to cost 17 cycles (for optimal non-contended
case), thus 3+17 = 20 cycles, it should still be a win.


--Jesper