On Fri, Mar 04, 2022 at 04:26:21PM -0500, Mike Snitzer wrote: > Hi, > > I've rebased Ming's latest [1] ontop of dm-5.18 [2] (which is based on > for-5.18/block). End result available in dm-5.18-biopoll branch [3] > > These changes add bio polling support to DM. Tested with linear and > striped DM targets. > > IOPS improvement was ~5% on my baremetal system with a single Intel > Optane NVMe device (555K hipri=1 vs 525K hipri=0). > > Ming has seen better improvement while testing within a VM: > dm-linear: hipri=1 vs hipri=0 15~20% iops improvement > dm-stripe: hipri=1 vs hipri=0 ~30% iops improvement > > I'd like to merge these changes via the DM tree when the 5.18 merge > window opens. The first block patch that adds ->poll_bio to > block_device_operations will need review so that I can take it > through the DM tree. Reason for going through the DM tree is there > have been some fairly extensive changes queued in dm-5.18 that build > on for-5.18/block. So I think it easiest to just add the block > depenency via DM tree since DM is first consumer of ->poll_bio > > FYI, Ming does have another DM patch [4] that looks to avoid using > hlist but I only just saw it. bio_split() _is_ involved (see > dm_split_and_process_bio) so I'm not exactly sure where he is going > with that change. io_uring(polling) workloads often cares latency, so big IO request isn't involved usually, I guess. Then bio_split() is seldom called in dm_split_and_process_bio(), such as if 4k random IO is run on dm-linear or dm-stripe via io_uring, bio_split() won't be run into. Single list is enough here, and efficient than hlist, just need a little care to delete element from the list since linux kernel doesn't have generic single list implementation. > But that is DM-implementation detail that we'll > sort out. Yeah, that patch also needs more test. Thanks, Ming
