On Wed, Feb 03, 2021 at 03:38:54PM +0100, Miklos Szeredi wrote: > On Wed, Feb 3, 2021 at 3:28 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote: > > > > On Wed, Feb 03, 2021 at 03:13:29PM +0100, Miklos Szeredi wrote: > > > On Wed, Feb 3, 2021 at 2:58 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote: > > > > > > > Network filesystems frequently need to use the credentials attached to > > > > a struct file in order to communicate with the server. There's no point > > > > fighting this reality. > > > > > > IDGI. Credentials can be taken from the file and from the task. In > > > this case every filesystem except cifs looks at task creds. Why are > > > network filesystem special in this respect? > > > > I don't necessarily mean 'struct cred'. I mean "the authentication > > that the client has performed to the server". Which is not a per-task > > thing, it's stored in the struct file, which is why we have things like > > > > int (*write_begin)(struct file *, struct address_space *mapping, > > loff_t pos, unsigned len, unsigned flags, > > struct page **pagep, void **fsdata); > > > > disk filesystems ignore the struct file argument, but network filesystems > > very much use it. > > Fine for file I/O. That's authorized at open time for all > filesystems, not just network ones. > > Not fine for metadata operations (IMO). I.e. ->[gs]etattr() don't > take a file argument either, even though on the uAPI there are plenty > of open file based variants. That's a fine statement of principle, but if the filesystem needs to contact the server, then your principle must accede to reality. But let's talk specifics. What does CIFS need to contact the server for? Could it be cached earlier?
