Messages in this thread | ![/](/images/icornerl.gif) | | Date | Tue, 29 Apr 2014 20:03:24 +0200 | Subject | Re: dcache shrink list corruption? | From | Miklos Szeredi <> |
| |
On Tue, Apr 29, 2014 at 7:43 PM, Linus Torvalds <torvalds@linux-foundation.org> wrote: > On Tue, Apr 29, 2014 at 9:01 AM, Miklos Szeredi <miklos@szeredi.hu> wrote: >> This was reported by IBM for 3.12, but if my analysis is right, it affects >> current kernel as well as older ones. >> >> So the question is: does anything protect the shrink list from concurrent >> modification by one or more dput() instances? > > Ugh. I don't see anything. The shrinking list is a private list, so > adding on its own would be entirely safe, and I think that's where the > "we don't need no steenking locking" comes from. > > But yes, the dentries are then visible on the hash chains, and there > can be concurrent removals from the list. > > That new global lock smells, though - and if we want to use a global > lock, we should simply use the existing per-superblock LRU lock, not > make up some new global one. The moving case already holds it, can't > we just take it in the add/del case? Was there some reason you didn't > do that?
Because we no longer have that. It now uses the list_lru thing, with a "per-node" lock, whatever that one is.
Introducing a new per-sb lock should be OK.
Another idea, which could have subtler effects, is simply not to kill a dentry that is on the shrink list (indicated by DCACHE_SHRINK_LIST), since it's bound to get killed anyway. But that's a change in behaviour...
> > Let me think about it, maybe there's some trick we can do by virtue of > the list head being private and us holding the dentry lock. So at > least on addition, we have *two* of the tree involved nodes locked. > Does that perhaps allow for any lockless games to be played?
I don't understand how that could be made to work.
| ![\](/images/icornerr.gif) |