On 7/17/24 13:40, tj@xxxxxxxxxx wrote:
Hello, Waiman.
On Wed, Jul 17, 2024 at 01:37:56PM -0400, Waiman Long wrote:
bis->sync is still being used in blk_cgroup_bio_start(). Replacing it with a
global lock may kill performance. We may have to use a per-cpu lock if we
want to go this route of eliminating bis->sync.
So, the idea is to keep using u64_sync for blkg->iostat_cpu and use
blkg_stat_lock for blkg->iostat. The former is the only one which is updated
in hot path, right?
Well, it can be confusing whether we are dealing with blkg->iostat or
blkg->iostat_cpu. In many cases, we are dealing with iostat_cpu instead
of iostat like __blkcg_rstat_flush() and blkg_clear_stat(). So we can't
eliminate the use of u64_stats_update_begin_irqsave() in those cases.
Cheers,
Longman
