Re: [PATCH] cifs: add comments explaining cifs_new_fileinfo behavior

On Tue, 11 May 2010 10:12:07 -0500
Shirish Pargaonkar <[email protected]> wrote:

> On Tue, May 11, 2010 at 10:06 AM, Jeff Layton <[email protected]> wrote:
> > On Tue, 11 May 2010 09:45:54 -0500
> > Shirish Pargaonkar <[email protected]> wrote:
> >
> >> On Tue, May 11, 2010 at 9:12 AM, Jeff Layton <[email protected]> wrote:
> >> > On Tue, 11 May 2010 18:18:08 +0530
> >> > Suresh Jayaraman <[email protected]> wrote:
> >> >
> >> >> On 05/11/2010 05:03 PM, Jeff Layton wrote:
> >> >> > On Tue, 11 May 2010 16:33:44 +0530
> >> >> > Suresh Jayaraman <[email protected]> wrote:
> >> >> >
> >> >> >>>>
> >> >> >>>> +/*
> >> >> >>>> + * When called with struct file pointer set to NULL, there is no way we could
> >> >> >>>> + * update file->private_data, but getting it stuck on openFileList provides a
> >> >> >>>> + * way to access it from cifs_fill_filedata and thereby set file->private_data
> >> >> >>>> + * from cifs_open.
> >> >> >>>> + */
> >> >> >>
> >> >> >>>
> >> >> >>> The comments help, but this code has bothered me for some time. Is it
> >> >> >>> possible for the create to return success and for something else to
> >> >> >>> happen such that the cifs_open is never called? I'd imagine that it is,
> >> >> >>> and if so, then this this open file will be "leaked".
> >> >> >>
> >> >> >> I asked Shirish exactly the same question while discussing in the
> >> >> >> #samba-technical irc channel. He does not see a leak, but thought you or
> >> >> >> Steve will have a better idea..
> >> >> >>
> >> >> >> Steve:  is such a situation not possible at all?
> >> >> >>
> >> >> >
> >> >> > I'm pretty sure it is possible and may even be somewhat likely.
> >> >> > Consider this situation. vfs_create gets called from
> >> >> > __open_namei_create. Something like this:
> >> >> >
> >> >> > do_last
> >> >> >   __open_namei_create
> >> >> >     vfs_create
> >> >> >        inode create operation
> >> >> >
> >> >> > ...after this, __open_namei_create calls may_open to check permissions
> >> >> > and can return an error. If that occurs, then I don't think the open op
> >> >> > will ever be called.
> >> >> >
> >> >> > I think you can probably reproduce this by doing something like this:
> >> >> >
> >> >> > Have samba export a world-writable directory. Mount up the share with a
> >> >> > user's credentials. Make sure that unix extensions are enabled. Have a
> >> >> > different user do something like this into a file on the mount:
> >> >> >
> >> >> >     echo foo > /path/to/share/testfile
> >> >> >
> >> >> > It's probable that the file will be created, but the open-for-write
> >> >> > permission check will fail and the open file will be left dangling.
> >> >> >
> >> >>
> >> >> A quick test shows it is not leaking atleast in this case. What happens is:
> >> >>   cifs_lookup() to the file returns NULL
> >> >>   cifs_posix_open()
> >> >>     CIFSPOSIXCreate() (file gets created)
> >> >>     cifs_new_fileinfo() (updated the openFileList)
> >> >>   lookup_instantiate_filp (gets the filep, calls cifs_open)
> >> >>   followed by a cifs_close
> >> >>
> >> >> On the wire, I see SET_PATH_INFO with Level Of Interest set to Set File
> >> >> Posix Open call followed by a Close.
> >> >>
> >> >>
> >> >> Thanks,
> >> >>
> >> >
> >> > Hmm ok...sounds like the create on lookup stuff might be getting in the
> >> > way of reproducing this (even though you said that cifs_lookup returned
> >> > NULL). Maybe could do better reproducing this with a program that does
> >> > an open(...O_EXCL|O_CREAT|O_WRONLY....) or something? The O_EXCL makes
> >> > it fall through to cifs_create which doesn't do lookup_instantiate_filp.
> >> >
> >> > In any case, I think the problem is valid. Clearly nothing will clean
> >> > these up if cifs_open is never called after cifs_create is...right?
> >> >
> >>
> >> It is possible, during unmount, cifs goes through openFileList, freeing up
> >> structures on the list, have to delve into code to affirm this either way.
> >> If cifs is not doing it, it should start doing it.
> >>
> >
> > When I say "leaked" I don't mean that they are never cleaned up (though
> > I'm not certain that they are), but rather that they outlive the open
> > syscall even though it failed.
> >
> > We should be cleaning these up when the open fails. It's incorrect to
> > assume that cifs_open will always be called after cifs_create on an
> > open(...O_CREAT...)
> 
> And along with cleaning them up, we should close the file at the server also?
> 

Yes, I believe so.


-- 
Jeff Layton <[email protected]>