From mboxrd@z Thu Jan 1 00:00:00 1970 From: huhaiyan@huawei.com (Haiyan Hu) Date: Wed, 4 Sep 2013 17:15:51 +0800 Subject: [PATCH] NVMe: Avoid caculate cq head doorbel in nvme_process_cq() In-Reply-To: <20130903155316.GY4707@linux.intel.com> References: <20130903155316.GY4707@linux.intel.com> Message-ID: <5226FA47.10205@huawei.com> On 2013/9/3 23:53, Matthew Wilcox wrote: > On Mon, Sep 02, 2013@06:49:34AM +0000, Huhaiyan (haiyan) wrote: >> We use 2 pointers, one for sq tail doorbell, the other for cq head doorbell in struct nvme_queue, to avoid calculate cq head doorbell in nvme_process_cq(). >> This change can reduce latency for the admin/io command. > > Have you been able to measure a difference? If so, can you share > relative numbers? > Sorry for my inaccurate description. Our test data shows that this patch does not change cmd process latency. But I think it can improve code efficiency, maybe a little. In addition, the format of the patch I sent is corrupted. I will resend it later. Thank you.