On 7/11/21 8:18 PM, Yu Kuai wrote:

If there are multiple active queues while sharing a tag set, it's not
necessary to limit the available tags as same share for each active queue
if no one ever failed to get driver tag. And fall back to same share if
someone do failed to get driver tag.

This modification will be beneficial if total queue_depth of disks
on the same host is less than total tags.

This patch adds new atomic operations in the hot path and hence probably
has a negative performance impact. What is the performance impact of
this patch for e.g. null_blk when submitting I/O from all CPU cores?

Thanks,

Bart.