Christoph Hellwig <hch@xxxxxxxxxxxxx> writes: > On Tue, Mar 13, 2012 at 12:00:05PM +0100, Miklos Szeredi wrote: >> > Do we really need the opendata structure? >> > >> > It seems like we could just pass a struct path instead of the dentry >> > passed directly and the vfsmount in it. There should be no need to >> > preallocate the file before calling into ->atomic_open, as it's only >> > used to pass around f_flags - but we already pass that one to >> > ->atomic_open directly and might as well pass it on to finish_open and >> > allocate the file there. >> >> We really don't want to get into the situation where the open fails >> after a successful create(*). Which means the file needs to be allocated >> prior to calling ->atomic_open and needs to be passed to finish_open() >> toghether with the vfsmount and dentry. >> >> In the first version of the patch I set filp->f_path.mnt to nd->path.mnt >> and passed the half initialized filp to ->atomic_open. But then decided >> that it's confusing for the filesystem code to deal with a half baked >> filp (does it need to be fput on error? etc...) >> >> Doing it with an opaque opendata makes this cleaner I think. > > Make sense. Can you throw in another cleanup patch to really just make > it a pass-through and not also use it as a boolean flag if open_flags > should be obeyed? This probably will change sematincs for the various > filesystems, but given that they should behave the same way that's a > good thing. It's not just that. The filesystems will create some state if od is non-NULL, which is released in f_op->release. If od is always non-NULL then the VFS has to call ->release on a dummy file, that file has to be allocated, which might fail... So this brings with it a couple of issues that I didn't want to deal with. But yes, it would probably be a good cleanup... Thanks, Miklos -- 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