[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <d1b06606-f01c-918e-0921-5d6c697f9c89@linux.ibm.com>
Date: Tue, 28 Feb 2023 13:52:18 +0100
From: Wenjia Zhang <wenjia@...ux.ibm.com>
To: Guangguan Wang <guangguan.wang@...ux.alibaba.com>,
Kai <KaiShen@...ux.alibaba.com>, kgraul@...ux.ibm.com,
jaka@...ux.ibm.com
Cc: kuba@...nel.org, davem@...emloft.net, netdev@...r.kernel.org,
linux-s390@...r.kernel.org, linux-rdma@...r.kernel.org
Subject: Re: [PATCH net-next v2] net/smc: Use percpu ref for wr tx reference
On 28.02.23 13:15, Guangguan Wang wrote:
>
> On 2023/2/28 19:34, Kai wrote:
>>
>>
>> On 2023/2/28 6:55 下午, Wenjia Zhang wrote:
>>
>>> @Kai, the performance improvement seems not so giant, but the method looks good, indeed. However, to keep the consistency of the code, I'm wondering why you only use the perf_ref for wr_tx_wait, but not for wr_reg_refcnt?
>> Didn't check the similar refcnt, my bad.
>> On the other hand, Our work is inspired by performance analysis, it seems wr_reg_refcnt is not on the IO path. It may not contribute to performance improvement.
>> And inspired by your comment, it seems we can also make the refcnt cdc_pend_tx_wr a perfcpu one. I will look into this.
>>
>> Thanks
>
> cdc_pend_tx_wr needs to be zero value tested every time it decreases in smc_cdc_tx_handler.
> I don't think this is the right scenario for percpu_ref.
I agree, that's why I didn't mention it;)
But could you please check about wr_reg_refcnt? Because we do need to
find the right balance between the code consistency and improvement
Powered by blists - more mailing lists