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: <b28b0e3e-87e4-5a02-c172-2d1424405a5a@redhat.com>
Date:   Thu, 15 Jun 2023 16:00:13 +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, hawk@kernel.org, ilias.apalodimas@linaro.org,
        netdev@vger.kernel.org, linux-kernel@vger.kernel.org,
        davem@davemloft.net, edumazet@google.com, pabeni@redhat.com
Subject: Re: [PATCH net-next] page pool: not return page to alloc cache during
 pool destruction
Content-Language: en-US
To:     Jakub Kicinski <kuba@kernel.org>,
        Liang Chen <liangchen.linux@gmail.com>
References: <20230615013645.7297-1-liangchen.linux@gmail.com>
 <20230614212031.7e1b6893@kernel.org>
In-Reply-To: <20230614212031.7e1b6893@kernel.org>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk


On 15/06/2023 06.20, Jakub Kicinski wrote:
> On Thu, 15 Jun 2023 09:36:45 +0800 Liang Chen wrote:
>> When destroying a page pool, the alloc cache and recycle ring are emptied.
>> If there are inflight pages, the retry process will periodically check the
>> recycle ring for recently returned pages, but not the alloc cache (alloc
>> cache is only emptied once). As a result, any pages returned to the alloc
>> cache after the page pool destruction will be stuck there and cause the
>> retry process to continuously look for inflight pages and report warnings.
>>
>> To safeguard against this situation, any pages returning to the alloc cache
>> after pool destruction should be prevented.
> 
> Let's hear from the page pool maintainers but I think the driver
> is supposed to prevent allocations while pool is getting destroyed.

Yes, this is a driver API violation. Direct returns (allow_direct) can
only happen from drivers RX path, e.g while driver is active processing
packets (in NAPI).  When driver is shutting down a page_pool, it MUST
have stopped RX path and NAPI (napi_disable()) before calling
page_pool_destroy()  Thus, this situation cannot happen and if it does
it is a driver bug.

> Perhaps we can add DEBUG_NET_WARN_ON_ONCE() for this condition to
> prevent wasting cycles in production builds?
> 

For this page_pool code path ("allow_direct") it is extremely important
we avoid wasting cycles in production.  As this is used for XDP_DROP
use-cases for 100Gbit/s NICs.

At 100Gbit/s with 64 bytes Ethernet frames (84 on wire), the wirespeed
is 148.8Mpps which gives CPU 6.72 nanosec to process each packet.
The microbench[1] shows (below signature) that page_pool_alloc_pages() +
page_pool_recycle_direct() cost 4.041 ns (or 14 cycles(tsc)).
Thus, for this code fast-path every cycle counts.

In practice PCIe transactions/sec seems limit total system to 108Mpps
(with multiple RX-queues + descriptor compression) thus 9.26 nanosec to
process each packet. Individual hardware RX queues seems be limited to
around 36Mpps thus 27.77 nanosec to process each packet.

Adding a DEBUG_NET_WARN_ON_ONCE will be annoying as I like to run my
testlab kernels with CONFIG_DEBUG_NET, which will change this extreme
fash-path slightly (adding some unlikely's affecting code layout to the
mix).

Question to Liang Chen: Did you hit this bug in practice?

--Jesper

CPU E5-1650 v4 @ 3.60GHz
  tasklet_page_pool01_fast_path Per elem:  14 cycles(tsc)  4.041 ns
  tasklet_page_pool02_ptr_ring  Per elem:  49 cycles(tsc) 13.622 ns
  tasklet_page_pool03_slow      Per elem: 162 cycles(tsc) 45.198 ns

[1] 
https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/lib/bench_page_pool_simple.c