linux-kernel - Re: [ata] 0568e61225: stress-ng.copy-file.ops_per

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Tue, 16 Aug 2022 08:42:42 -0700
From:   Damien Le Moal <damien.lemoal@...nsource.wdc.com>
To:     John Garry <john.garry@...wei.com>,
        Oliver Sang <oliver.sang@...el.com>
Cc:     Christoph Hellwig <hch@....de>,
        "Martin K. Petersen" <martin.petersen@...cle.com>,
        LKML <linux-kernel@...r.kernel.org>,
        Linux Memory Management List <linux-mm@...ck.org>,
        linux-ide@...r.kernel.org, lkp@...ts.01.org, lkp@...el.com,
        ying.huang@...el.com, feng.tang@...el.com,
        zhengjun.xing@...ux.intel.com, fengwei.yin@...el.com
Subject: Re: [ata] 0568e61225: stress-ng.copy-file.ops_per_sec -15.0%
 regression

On 2022/08/16 3:35, John Garry wrote:
> On 16/08/2022 07:57, Oliver Sang wrote:
>>>> For me, a complete kernel log may help.
>>> and since only 1HDD, the output of the following would be helpful:
>>>
>>> /sys/block/sda/queue/max_sectors_kb
>>> /sys/block/sda/queue/max_hw_sectors_kb
>>>
>>> And for 5.19, if possible.
>> for commit
>> 0568e61225 ("ata: libata-scsi: cap ata_device->max_sectors according to shost->max_sectors")
>>
>> root@...-icl-2sp1 ~# cat /sys/block/sda/queue/max_sectors_kb
>> 512
>> root@...-icl-2sp1 ~# cat /sys/block/sda/queue/max_hw_sectors_kb
>> 512
>>
>> for both commit
>> 4cbfca5f77 ("scsi: scsi_transport_sas: cap shost opt_sectors according to DMA optimal limit")
>> and v5.19
>>
>> root@...-icl-2sp1 ~# cat /sys/block/sda/queue/max_sectors_kb
>> 1280
>> root@...-icl-2sp1 ~# cat /sys/block/sda/queue/max_hw_sectors_kb
>> 32767
>>
> 
> thanks, I appreciate this.
> 
>  From the dmesg, I see 2x SATA disks - I was under the impression that 
> the system only has 1x.
> 
> Anyway, both drives show LBA48, which means the large max hw sectors at 
> 32767KB:
> [   31.129629][ T1146] ata6.00: 1562824368 sectors, multi 1: LBA48 NCQ 
> (depth 32)
> 
> So this is what I suspected: we are capped from the default shost max 
> sectors (1024 sectors).
> 
> This seems like the simplest fix for you:
> 
> --- a/include/linux/libata.h
> +++ b/include/linux/libata.h
> @@ -1382,7 +1382,8 @@ extern const struct attribute_group 
> *ata_common_sdev_groups[];
>         .proc_name              = drv_name,                     \
>         .slave_destroy          = ata_scsi_slave_destroy,       \
>         .bios_param             = ata_std_bios_param,           \
> -       .unlock_native_capacity = ata_scsi_unlock_native_capacity
> +       .unlock_native_capacity = ata_scsi_unlock_native_capacity,\
> +       .max_sectors = ATA_MAX_SECTORS_LBA48

This is crazy large (65535 x 512 B sectors) and never result in that being
exposed as the actual max_sectors_kb since other limits will apply first
(mapping size).

The regression may come not from commands becoming tiny, but from the fact that
after the patch, max_sectors_kb is too large, causing a lot of overhead with
qemu swiotlb mapping and slowing down IO processing.

Above, it can be seen that we ed up with max_sectors_kb being 1280, which is the
default for most scsi disks (including ATA drives). That is normal. But before
that, it was 512, which likely better fits qemu swiotlb and does not generate
overhead. So the above fix will not change anything I think...

> A concern is that other drivers which use libata may have similar 
> issues, as they use default in SCSI_DEFAULT_MAX_SECTORS for max_sectors:
> hisi_sas
> pm8001
> aic9xxx
> mvsas
> isci
> 
> So they may be needlessly hobbled for some SATA disks. However I have a 
> system with hisi_sas controller and attached LBA48 disk. I tested 
> performance for v5.19 vs 6.0 and it was about the same for fio rw=read @ 
> ~120K IOPS. I can test this further.
> 
> Thanks,
> John


-- 
Damien Le Moal
Western Digital Research