Hi Tejun, On 18/2/8 23:23, Tejun Heo wrote: > Hello, Joseph. > > On Thu, Feb 08, 2018 at 10:29:43AM +0800, Joseph Qi wrote: >> So you mean checking css->refcnt to prevent the further use of >> blkg_get? I think it makes sense. > > Yes. > >> IMO, we should use css_tryget_online instead, and rightly after taking > > Not really. An offline css still can have a vast amount of IO to > drain and write out. > IIUC, we have to identify it is in blkcg_css_offline now which will blkg_put. Since percpu_ref_kill_and_confirm in kill_css will set flag __PERCPU_REF_DEAD, so we can use this to avoid the race. IOW, if __PERCPU_REF_DEAD is set now, we know blkcg css is in offline and continue access blkg may risk double free. Thus we choose to skip these ios. I don't get how css_tryget works since it doesn't care the flag __PERCPU_REF_DEAD. Also css_tryget can't prevent blkcg_css from offlining since the race happens blkcg_css_offline is in progress. Am I missing something here? Thanks, Joseph >> queue_lock. Because there may be more use of blkg_get in blk_throtl_bio >> in the futher. Actually it already has two now. One is in >> blk_throtl_assoc_bio, and the other is in throtl_qnode_add_bio. >> What do you think of this? > > As long as we avoid making the bypass paths expensive, whatever makes > the code easier to read and maintain would be better. css ref ops are > extremely cheap anyway. > > Thanks. >
