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: [PATCH 1/1] mm: avoid re-using pfmemalloc page in
 page_frag_alloc()
To:     Eric Dumazet <eric.dumazet@gmail.com>,
        Matthew Wilcox <willy@infradead.org>
Cc:     Rama Nichanamatlu <rama.nichanamatlu@oracle.com>,
        linux-mm@kvack.org, netdev@vger.kernel.org,
        linux-kernel@vger.kernel.org, akpm@linux-foundation.org,
        davem@davemloft.net, kuba@kernel.org, aruna.ramakrishna@oracle.com,
        bert.barbe@oracle.com, venkat.x.venkatsubra@oracle.com,
        manjunath.b.patil@oracle.com, joe.jin@oracle.com,
        srinivas.eeda@oracle.com
References: <20201103193239.1807-1-dongli.zhang@oracle.com>
 <20201103203500.GG27442@casper.infradead.org>
 <7141038d-af06-70b2-9f50-bf9fdf252e22@oracle.com>
 <20201103211541.GH27442@casper.infradead.org>
 <20201104011640.GE2445@rnichana-ThinkPad-T480>
 <2bce996a-0a62-9d14-4310-a4c5cb1ddeae@gmail.com>
 <20201104123659.GA17076@casper.infradead.org>
 <053d1d51-430a-2fa9-fb72-fee5d2f9785c@gmail.com>
From:   Dongli Zhang <dongli.zhang@oracle.com>
Message-ID: <2bb0d1dd-5b40-e0e1-9fed-7bfcbc3de6a6@oracle.com>
Date:   Wed, 4 Nov 2020 08:41:32 -0800
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101
 Thunderbird/68.10.0
MIME-Version: 1.0
In-Reply-To: <053d1d51-430a-2fa9-fb72-fee5d2f9785c@gmail.com>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 8bit
Precedence: bulk


On 11/4/20 4:51 AM, Eric Dumazet wrote:
> 
> 
> On 11/4/20 1:36 PM, Matthew Wilcox wrote:
>> On Wed, Nov 04, 2020 at 09:50:30AM +0100, Eric Dumazet wrote:
>>> On 11/4/20 2:16 AM, Rama Nichanamatlu wrote:
>>>>> Thanks for providing the numbers.  Do you think that dropping (up to)
>>>>> 7 packets is acceptable?
>>>>
>>>> net.ipv4.tcp_syn_retries = 6
>>>>
>>>> tcp clients wouldn't even get that far leading to connect establish issues.
>>>
>>> This does not really matter. If host was under memory pressure,
>>> dropping a few packets is really not an issue.
>>>
>>> Please do not add expensive checks in fast path, just to "not drop a packet"
>>> even if the world is collapsing.
>>
>> Right, that was my first patch -- to only recheck if we're about to
>> reuse the page.  Do you think that's acceptable, or is that still too
>> close to the fast path?
> 
> I think it is totally acceptable.
> 
> The same strategy is used in NIC drivers, before recycling a page.
> 
> If page_is_pfmemalloc() returns true, they simply release the 'problematic'page
> and attempt a new allocation.
> 
> ( git grep -n page_is_pfmemalloc -- drivers/net/ethernet/ )

While the drivers may implement their own page_frag_cache to manage skb->frags ...

... the skb->data is usually allocated via __netdev_alloc_skb() or
napi_alloc_skb(), which end up to the global this_cpu_ptr(&netdev_alloc_cache)
or this_cpu_ptr(&napi_alloc_cache.page).

> 
> 
>>
>>> Also consider that NIC typically have thousands of pre-allocated page/frags
>>> for their RX ring buffers, they might all have pfmemalloc set, so we are speaking
>>> of thousands of packet drops before the RX-ring can be refilled with normal (non pfmemalloc) page/frags.
>>>
>>> If we want to solve this issue more generically, we would have to try
>>> to copy data into a non pfmemalloc frag instead of dropping skb that
>>> had frags allocated minutes ago under memory pressure.
>>
>> I don't think we need to copy anything.  We need to figure out if the
>> system is still under memory pressure, and if not, we can clear the
>> pfmemalloc bit on the frag, as in my second patch.  The 'least change'
>> way of doing that is to try to allocate a page, but the VM could export
>> a symbol that says "we're not under memory pressure any more".
>>
>> Did you want to move checking that into the networking layer, or do you
>> want to keep it in the pagefrag allocator?
> 
> I think your proposal is fine, thanks !

Hi Matthew, are you going to send out the patch to avoid pfmemalloc recycle?

Thank you very much!

Dongli Zhang