Miklos Szeredi wrote: > On Mon, 20 Oct 2008, Christoph Lameter wrote: >>>> kick_inodes() only works on inodes that first have undergone >>>> get_inodes() where we establish a refcount under inode_lock(). The final >>>> cleanup in kick_inodes() is done under iprune_mutex. You are looking at >>>> the loop that does writeback and invalidates attached dentries. This can >>>> fail for various reasons. >>> Yes, but I'm not at all sure that calling remove_inode_buffers() or >>> invalidate_mapping_pages() is OK on a live inode. They should be done >>> after checking the refcount, just like prune_icache() does. >> Dont we do the same on a truncate? > > Yes, with i_mutex and i_alloc_sem held. There is another call to invalidate_mapping_pages() in prune_icache (that is where this code originates). No i_mutex and i_alloc. Only iprune_mutex held and that seems to be for the protection of the list. So just checking inode->i_count would do the trick? >>> Also, while d_invalidate() is not actually wrong here, because you >>> check S_ISDIR(), but it's still the wrong function to use. You really >>> just want to shrink the children. Invalidation means: the filesystem >>> found out that the cached inode is invalid, so we want to throw it >>> away. In the future it might actually be able to do it for >>> directories as well, but currently it cannot because of possible >>> mounts on the dentry. >> Thats the same issue as with the dentries. The new function could deal with >> both situations? > > Sure. > > The big issue is dealing with umount. You could do something like > grab_super() on sb before getting a ref on the inode/dentry. But I'm > not sure this is a good idea. There must be a simpler way to achieve > this... Taking a lock on vfsmount_lock? But that would make dentry reclaim a pain. We are only interested in the reclaim a dentry if its currently unused. If so then why does unmount matter? Both unmount and reclaim will attempt to remove the dentry. Have a look at get_dentries(). It takes the dcache_lock and checks the dentry state. Either the entry is ignored or dget_locked() removes it from the lru. If its off the LRU then it can no longer be reclaimed by umount. -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
