Re: [PATCH net-next v3 0/5] page_pool: recycle buffers

From: Shay Agroskin <shayagr@amazon.com>
To: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Yunsheng Lin <linyunsheng@huawei.com>,
	Ilias Apalodimas <ilias.apalodimas@linaro.org>,
	Matteo Croce <mcroce@linux.microsoft.com>,
	<netdev@vger.kernel.org>, <linux-mm@kvack.org>,
	Ayush Sawal <ayush.sawal@chelsio.com>,
	Vinay Kumar Yadav <vinay.yadav@chelsio.com>,
	Rohit Maheshwari <rohitm@chelsio.com>,
	"David S. Miller" <davem@davemloft.net>,
	Jakub Kicinski <kuba@kernel.org>,
	Thomas Petazzoni <thomas.petazzoni@bootlin.com>,
	Marcin Wojtas <mw@semihalf.com>,
	"Russell King" <linux@armlinux.org.uk>,
	Mirko Lindner <mlindner@marvell.com>,
	Stephen Hemminger <stephen@networkplumber.org>,
	Tariq Toukan <tariqt@nvidia.com>,
	Jesper Dangaard Brouer <hawk@kernel.org>,
	"Alexei Starovoitov" <ast@kernel.org>,
	Daniel Borkmann <daniel@iogearbox.net>,
	"John Fastabend" <john.fastabend@gmail.com>,
	Boris Pismenny <borisp@nvidia.com>, Arnd Bergmann <arnd@arndb.de>,
	Andrew Morton <akpm@linux-foundation.org>,
	"Peter Zijlstra (Intel)" <peterz@infradead.org>,
	Vlastimil Babka <vbabka@suse.cz>, Yu Zhao <yuzhao@google.com>,
	Will Deacon <will@kernel.org>,
	Michel Lespinasse <walken@google.com>,
	Fenghua Yu <fenghua.yu@intel.com>, Roman Gushchin <guro@fb.com>,
	Hugh Dickins <hughd@google.com>, Peter Xu <peterx@redhat.com>,
	Jason Gunthorpe <jgg@ziepe.ca>,
	Guoqing Jiang <guoqing.jiang@cloud.ionos.com>,
	"Jonathan Lemon" <jonathan.lemon@gmail.com>,
	Alexander Lobakin <alobakin@pm.me>,
	"Cong Wang" <cong.wang@bytedance.com>, wenxu <wenxu@ucloud.cn>,
	Kevin Hao <haokexin@gmail.com>,
	Aleksandr Nogikh <nogikh@google.com>,
	Jakub Sitnicki <jakub@cloudflare.com>,
	Marco Elver <elver@google.com>,
	Willem de Bruijn <willemb@google.com>,
	Miaohe Lin <linmiaohe@huawei.com>,
	Guillaume Nault <gnault@redhat.com>,
	<linux-kernel@vger.kernel.org>, <linux-rdma@vger.kernel.org>,
	<bpf@vger.kernel.org>, Matthew Wilcox <willy@infradead.org>,
	Eric Dumazet <edumazet@google.com>,
	David Ahern <dsahern@gmail.com>,
	Lorenzo Bianconi <lorenzo@kernel.org>,
	Saeed Mahameed <saeedm@nvidia.com>, Andrew Lunn <andrew@lunn.ch>,
	Paolo Abeni <pabeni@redhat.com>
Subject: Re: [PATCH net-next v3 0/5] page_pool: recycle buffers
Date: Sun, 9 May 2021 08:11:35 +0300	[thread overview]
Message-ID: <pj41zl4kfclce0.fsf@u570694869fb251.ant.amazon.com> (raw)
In-Reply-To: <20210507121953.59e22aa8@carbon>

Jesper Dangaard Brouer <brouer@redhat.com> writes:

> On Fri, 7 May 2021 16:28:30 +0800
> Yunsheng Lin <linyunsheng@huawei.com> wrote:
>
>> On 2021/5/7 15:06, Ilias Apalodimas wrote:
>> > On Fri, May 07, 2021 at 11:23:28AM +0800, Yunsheng Lin wrote:  
>> >> On 2021/5/6 20:58, Ilias Apalodimas wrote:  
>> >>>>>>  
>> >>>>>
> ...
>> > 
>> > 
>> > I think both choices are sane.  What I am trying to explain 
>> > here, is
>> > regardless of what we choose now, we can change it in the 
>> > future without
>> > affecting the API consumers at all.  What will change 
>> > internally is the way we
>> > lookup the page pool pointer we are trying to recycle.  
>> 
>> It seems the below API need changing?
>> +static inline void skb_mark_for_recycle(struct sk_buff *skb, 
>> struct page *page,
>> +					struct xdp_mem_info *mem)
>
> I don't think we need to change this API, to support future 
> memory
> models.  Notice that xdp_mem_info have a 'type' member.

Hi,
Providing that we will (possibly as a future optimization) store 
the pointer to the page pool in struct page instead of strcut 
xdp_mem_info, passing
xdp_mem_info * instead of struct page_pool * would mean that for 
every packet we'll need to call
             xa = rhashtable_lookup(mem_id_ht, &mem->id, 
             mem_id_rht_params);
             xa->page_pool;

which might pressure the Dcache to fetch a pointer that might be 
present already in cache as part of driver's data-structures.

I tend to agree with Yunsheng that it makes more sense to adjust 
the API for the clear use-case now rather than using xdp_mem_info 
indirection. It seems to me like
the page signature provides the same information anyway and allows 
to support different memory types.

Shay

>
> Naming in Computer Science is a hard problem ;-). Something that 
> seems
> to confuse a lot of people is the naming of the struct 
> "xdp_mem_info".  
> Maybe we should have named it "mem_info" instead or 
> "net_mem_info", as
> it doesn't indicate that the device is running XDP.
>
> I see XDP as the RX-layer before the network stack, that helps 
> drivers
> to support different memory models, also for handling normal 
> packets
> that doesn't get process by XDP, and the drivers doesn't even 
> need to
> support XDP to use the "xdp_mem_info" type.