Re: [RFC/PATCH 0/5] read/write on attribute w/o show/store should return -ENOSYS

[Dmitry Torokhov <dmitry.torokhov@gmail.com>]

On 4/28/05, Chris Wright <chrisw@osdl.org> wrote:
> * Greg KH (gregkh@suse.de) wrote:
> > On Thu, Apr 28, 2005 at 12:30:09AM -0500, Dmitry Torokhov wrote:
> > > Hi,
> > >
> > > Jean Delvare has noticed that if a driver happens to declare its
> > > attribute as RW but doesn't provide store() method attempt to write
> > > into such attribute will cause spinning process as most of the
> > > attribute implementations return 0 in case of missing store causing
> > > endless retries. In some cases missing show/store will return -EPERM,
> > > -EACCESS or -EINVAL.
> > >
> > > I think we should unify implementations and have them all return -ENOSYS
> > > (function not implemented) when corresponding method (show/store) is
> > > missing.
> >
> > What is the POSIX standard for this?  ENOSYS or EACCESS?
> 
> SuSv3 suggests EBADF, however we already do EINVAL at VFS for no write
> op.  Although, returning 0 (i.e. wrote zero bytes) is still meaningful
> too.
> 

Returning 0 causes caller to immediately repeat operation causing the
loop and 100% CPU utiluization as was reported. If ENOSYS is not
acceptable (after all the problem is sysfs-specific, other fs-es
either support write or they don't for entire volume) I'd say (looking
at the descripptions) we should use ENXIO instead of EBADF:

[ENXIO] 
A request was made of a nonexistent device, or the request was outside
the capabilities of the device.

[EBADF] 
The fildes argument is not a valid file descriptor open for writing.

We have a valid FD and it was opened for wrinting, so EBADF is not
really applicable. But then I may be talking complete gibberish...

-- 
Dmitry