On Thu 20-02-20 19:07:01, Yufen Yu wrote: > Hi, > > On 2020/2/19 20:55, Jan Kara wrote: > > Hi! > > > > On Sat 15-02-20 21:54:08, Yufen Yu wrote: > > > > > I've now noticed there's commit 68f23b8906 "memcg: fix a crash in wb_workfn > > when a device disappears" from end of January which tries to address the > > issue you're looking into. Now AFAIU the code is till somewhat racy after > > that commit so I wanted to mention this mostly so that you fixup also the > > new bdi_dev_name() while you're fixing blkg_dev_name(). > > > > Also I was wondering about one thing: If we really care about bdi->dev only > > for the name, won't we be much better off with just copying the name to > > bdi->name on registration? Sure it would consume a bit of memory for the > > name copy but I don't think we really care and things would be IMO *much* > > simpler that way... Yufen, Tejun, what do you think? > > > > I think copying the name to bdi->name is also need protected by lock. > Otherwise, the reader of bdi->name may read incorrect value when > re-registion have not copy the name completely. Right? So, I also think > using RCU to protect object lifetimes may be a better way. OK, fair enough. :) Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
