On 11/10/21 12:08 PM, Jaegeuk Kim wrote: > On 11/09, Jens Axboe wrote: >> On 11/9/21 9:39 AM, Christoph Hellwig wrote: >>> On Mon, Nov 08, 2021 at 06:13:36PM -0800, Jaegeuk Kim wrote: >>>> This patch adds a way to attach HIPRI by expanding the existing sysfs's >>>> data_io_flag. User can measure IO performance by enabling it. >>> >>> NAK. This flag should only be used when explicitly specified by >>> the submitter of the I/O. >> >> Yes, this cannot be set in the middle for a multitude of reasons. I wonder >> if we should add a comment to that effect near the definition of it. > > Not surprising. I was wondering we can add this for testing purpose only. > Btw, is there a reasonable way that filesystem can use IO polling? Whether an IO is polled or not belongs to the issuer of the IO, as it comes with certain obligations like "I will actively poll for the completion of this request", and it incurs certain restrictions in the block layer in terms of whether or not you can ever sleep for requests. You could certainly use in in an fs, but only IFF you are the original issuer of the request, which then also means that you are the one that needs to poll for completion of it. It's not a drive-by "let's set this flag to speed things up" kind of flag, there are a lot more moving parts than that. I don't think it will be useful for you. -- Jens Axboe
