On 2021/08/24 2:15, Bart Van Assche wrote: > On 8/23/21 12:36 AM, Niklas Cassel wrote: >> I was mainly thinking that it should be possible to do a generic fix, >> such that we eventually won't need a similar fix as yours in all the >> different I/O schedulers. > > Coming up with a generic fix would be great but I have not yet found an > elegant approach ... > > Another question is what the impact is of scheduler bypass on zoned > block devices? Is the zone locking performed by the mq-deadline > scheduler for writes to zoned block devices compatible with I/O > scheduler bypass? Without mq-deadline, in the general case, regular writes to the same zone may end up being reordered and IO errors will follow. Only zone append writes can survive the scheduler bypass as zone write locking in that case is done at the scsi disk driver level using the dispatch queue. So bypassing the scheduler for writes can work only with very special cases, namely, the user issuign small request that are never split and at QD=1 at most per zone. Any other workload (larger requests and/or higher write QD per zone) can easily trigger write errors (that is fairly easy to check). > >> However, it does not apply on top of Torvalds master or Jens's for-next >> branch because they both have reverted your cgroup support patch. >> >> If you rebase your fix and send it out, I will be happy to send out >> a Reviewed-by/Tested-by. > > I will rebase, retest and resend my patch. > > Thanks, > > Bart. > -- Damien Le Moal Western Digital Research
