On 09/08/2022 10:58, John Garry wrote:
>>>
>>> commit: 0568e6122574dcc1aded2979cd0245038efe22b6 ("ata: libata-scsi: 
>>> cap ata_device->max_sectors according to shost->max_sectors")
>>> https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git master
>>>
>>> in testcase: stress-ng
>>> on test machine: 96 threads 2 sockets Ice Lake with 256G memory
>>> with following parameters:
>>>
>>>     nr_threads: 10%
>>>     disk: 1HDD
>>>     testtime: 60s
>>>     fs: f2fs
>>>     class: filesystem
>>>     test: copy-file
>>>     cpufreq_governor: performance
>>>     ucode: 0xb000280
>>
>> Without knowing what the device adapter is, hard to say where the 
>> problem is. I
>> suspect that with the patch applied, we may be ending up with a small 
>> default
>> max_sectors value, causing overhead due to more commands than necessary.
>>
>> Will check what I see with my test rig.
> 
> As far as I can see, this patch should not make a difference unless the 
> ATA shost driver is setting the max_sectors value unnecessarily low.

For __ATA_BASE_SHT, we don't set max_sectors. As such, we default 
shost->max_sectors = SCSI_DEFAULT_MAX_SECTORS (=1024) in 
scsi_host_alloc(). I assume no shost dma mapping limit applied.

Then - for example - we could select dev->max_sectors = 
ATA_MAX_SECTORS_LBA48 (=65535) in ata_dev_configure().

So with commit 0568e6122574 we would have final max sectors = 1024, as 
opposed to 65535 previously. I guess that the problem is something like 
this.

If so, it seems that we would need to apply the shost dma mapping limit 
separately in ata_scsi_dev_config() and not use shost->max_sectors.

thanks,
John