On Fri 21-05-21 08:42:16, Ming Lei wrote: > On Fri, May 21, 2021 at 12:33:52AM +0200, Jan Kara wrote: > > Most of the merging happens at bio level. There should not be much > > merging happening at request level anymore. Furthermore if we backmerged > > a request to the previous one, the chances to be able to merge the > > result to even previous request are slim - that could succeed only if > > requests were inserted in 2 1 3 order. Merging more requests in > > Right, but some workload has this kind of pattern. > > For example of qemu IO emulation, it often can be thought as single job, > native aio, direct io with high queue depth. IOs is originated from one VM, but > may be from multiple jobs in the VM, so bio merge may not hit much because of IO > emulation timing(virtio-scsi/blk's MQ, or IO can be interleaved from multiple > jobs via the SQ transport), but request merge can really make a difference, see > recent patch in the following link: > > https://lore.kernel.org/linux-block/3f61e939-d95a-1dd1-6870-e66795cfc1b1@xxxxxxx/T/#t Oh, request merging definitely does make a difference. But the elevator hash & merge logic I'm modifying here is used only by BFQ and MQ-DEADLINE AFAICT. And these IO schedulers will already call blk_mq_sched_try_merge() from their \.bio_merge handler which gets called from blk_mq_submit_bio(). So all the merging that can happen in the code I remove should have already happened. Or am I missing something? Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
