Re: Yet another pipe related oops.

[Al Viro <viro@ZenIV.linux.org.uk>]

On Wed, Mar 27, 2013 at 09:33:35AM -0700, Linus Torvalds wrote:

> Applied.
> 
> Do we actually have files with NULL f_ops pointers? Should we? What
> could we possibly do with a file descriptor that doesn't have any
> fops?

We shouldn't, at least not for something that has been successfully
opened.  I've a patch series cleaning that up a bit in the local
queue; will check for bitrot and throw into for-next.

Another thing that is a definite for-next fodder - we really have no
reason to put anything non-regular or opened not for write into ->s_files.
And since read-only opens outnumber write-only/read-write ones by far
(two orders of magnitude for something like kernel build), that gives
a nice reduction of files_lglock accesses.  OTOH, the only remaining
user of those lists is forced remount to read-only, and I'm not at all
sure we wouldn't be better off by leaving those opened files alone and
just teaching file_start_write() to fail with EROFS on such fs.  Then
we could get rid of files_lglock and ->s_files completely...

> Also, perhaps we should do something more akin to what we do for
> dentry functions where we validate them on registration, and we could
> fix up or validate read/write pointers, with semantics something like
> 
>     if (!fop->write)
>         fop->write = fop->aio_write ? do_sync_write : EINVAL_write;
>     if (!fop->read)
>         fop->read = fop->aio_read ? do_sync_read : EINVAL_read;
> 
> kind of things?

As it is, file_operations instances are const, and it's a good idea, IMO...