Re: [linus:master] [io_uring] caec5ebe77: stress-ng.io-uring.ops_per_sec -33.1% regression

[Yin, Fengwei]

Hi Jens,

On 8/17/2023 3:27 AM, Jens Axboe wrote:
>>  static const struct file_operations shmem_file_operations = {
>>         .mmap           = shmem_mmap,
>> -       .open           = generic_file_open,
>> +       .open           = shmem_file_open,
>>         .get_unmapped_area = shmem_get_unmapped_area,
>>  #ifdef CONFIG_TMPFS
>>         .llseek         = shmem_file_llseek,
>>
>> The performance change when running stress-ng.io_uring with testing file
>> in tmpfs is gone.
>>
>> This is just the information FYI. I may miss something obviously here. Thanks.
> This actually highlighted a problem with the old nowait logic, in that
> it assumed !bdev would mean that nowait was fine. Looking at shmem, we
> definitely need IOCB_NOWAIT handling in there to make that safe. So the
> old code was buggy, and conversely, we can't also just add the
> FMODE_NOWAIT without first making those improvements to shmem first.
Thanks a lot for the explanation.

> 
> Basically you'd want to ensure that the read_iter and write_iter paths
> honor IOCB_NOWAIT. Once that's done, then FMODE_NOWAIT can indeed be set
> in the open helper.
> 
> I might take a stab at this, but I'm out sick right now so don't think
> it'd be cohesive if I did it right now.
Take care.


Regards
Yin, Fengwei