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;
Message-ID: <2b900d3b-9b52-5bc3-4ba8-24249f3c2e42@huawei.com>
Date:   Sat, 13 Aug 2022 08:23:36 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:91.0) Gecko/20100101
 Thunderbird/91.6.1
Subject: Re: [ata] 0568e61225: stress-ng.copy-file.ops_per_sec -15.0%
 regression
To:     Damien Le Moal <damien.lemoal@opensource.wdc.com>,
        Oliver Sang <oliver.sang@intel.com>
CC:     Christoph Hellwig <hch@lst.de>,
        "Martin K. Petersen" <martin.petersen@oracle.com>,
        LKML <linux-kernel@vger.kernel.org>,
        "Linux Memory Management List" <linux-mm@kvack.org>,
        <linux-ide@vger.kernel.org>, <lkp@lists.01.org>, <lkp@intel.com>,
        <ying.huang@intel.com>, <feng.tang@intel.com>,
        <zhengjun.xing@linux.intel.com>, <fengwei.yin@intel.com>
References: <YuzPMMnnY739Tnit@xsang-OptiPlex-9020>
 <1f498d4a-f93f-ceb4-b713-753196e5e08d@opensource.wdc.com>
 <3451fa5a-6229-073f-ae18-0c232cd48ed5@huawei.com>
 <e4106ffa-3842-45c0-9756-5226cfcfa17d@opensource.wdc.com>
 <YvXeuCAK780OuJPz@xsang-OptiPlex-9020>
 <2e9cf5a6-c043-5ccf-e363-097c6c941891@huawei.com>
 <4642848c-a386-d6a0-6255-8b16800e0548@opensource.wdc.com>
 <fd2b1dbe-8482-cb89-2568-4909db1239b0@huawei.com>
 <840a5f98-a53c-ce08-2833-f41d8c9a015b@opensource.wdc.com>
From:   John Garry <john.garry@huawei.com>
In-Reply-To: <840a5f98-a53c-ce08-2833-f41d8c9a015b@opensource.wdc.com>
Content-Type: text/plain; charset="UTF-8"; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 12/08/2022 19:27, Damien Le Moal wrote:
>> Interestingly ata dev max_sectors kb also gets capped from 32MB (LBA48)
>> -> 256KB due to swiotlb max mapping size. (It would be capped by shost
>> default max sectors 512KB without that swiotlb limit). I assume capping
>> due to swiotlb limit is not occuring on Oliver's machine.
> Yes, I was suspecting that we may be seeing a difference for anything that is
> not AHCI, e.g. with other drivers.
> 
> But that seems to be the correct thing to do, no ? 
Yes, this should be the correct thing to do.

 > How was this working before
 > without applying the swiotlb limit ?
 >
Not sure. I would need to check the libata code for how it handles DMA 
mapping errors, which I assume would occur for when we exceed the 
swiotlb limit.

Having said this, from limited testing, whenever I check megaraid sas or 
mpt3sas for performance, the length of request data never/rarely comes 
close to max sectors. That why I am surprised with the regression which 
Oliver reports.

Thanks,
John