Messages in this thread | | | Date | Tue, 25 May 2010 11:14:05 +1000 | From | Neil Brown <> | Subject | Re: [PATCH] VFS: fix recent breakage of FS_REVAL_DOT |
| |
On Mon, 24 May 2010 12:59:03 +0100 Al Viro <viro@ZenIV.linux.org.uk> wrote:
> On Mon, May 24, 2010 at 04:57:56PM +1000, Neil Brown wrote: > > > > Commit 1f36f774b22a0ceb7dd33eca626746c81a97b6a5 broke FS_REVAL_DOT semantics. > > > > In particular, before this patch, the command > > ls -l > > in an NFS mounted directory would always check if the directory on the server > > had changed and if so would flush and refill the pagecache for the dir. > > After this patch, the same "ls -l" will repeatedly return stale date until > > the cached attributes for the directory time out. > > > > The following patch fixes this by ensuring the d_revalidate is called by > > do_last when "." is being looked-up. > > link_path_walk has already called d_revalidate, but in that case LOOKUP_OPEN > > is not set so nfs_lookup_verify_inode chooses not to do any validation. > > > > The following patch restores the original behaviour. > > > > Cc: stable@kernel.org > > Signed-off-by: NeilBrown <neilb@suse.de> > > Applied, but I really don't like the way you do it; note that e.g. foo/bar/. > gets that revalidation as well, for no good reason. If anything, shouldn't > we handle that thing in the _beginning_ of pathname resolution, not in > the end? For now it'd do, and it's a genuine regression, but... >
Thanks.
I think I see what you mean by "at the beginning" - the problem path is simply ".", and both the start and end of that are "." so we can handle at either end... But I don't think there is any other special handling of the 'start' of a path, so I imagine it would be a fairly ugly special case.
We could avoid the extra GETATTR in "foo/bar/." by allowing NFS to keep some state in the namei_data to record that it has valid attributes for a given dentry so if it sees the same dentry again it doesn't need to revalidate.
I must confess though that I don't feel I understand VFS name lookup properly any more. Since intents were added it seems to have become much more obscure and complex. I cannot help thinking that there must be a better way: distinguish between the various cases at a higher level so we don't need as many flags being passed around and interpreted by widely separate pieces of code. I don't have a concrete proposal but I would certainly be interested to work on one if there were any hope of real change. Thoughts?
Thanks, NeilBrown
> BTW, here's a question for nfs client folks: is it true that for any two > pathnames on _client_ resolving to pairs (mnt1, dentry) and (mnt2, dentry) > resp., nfs_devname(mnt1, dentry, ...) and nfs_devname(mnt2, dentry, ...) > should yield the strings that do not differ past the ':' (i.e. that the > only possible difference is going to be in spelling the server name)?
| |