Re: [PATCH v2 3/8] fs: clear or set FMODE_LSEEK based on llseek function

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

On Sat, Jun 25, 2022 at 06:39:52AM -0700, Christoph Hellwig wrote:
> On Sat, Jun 25, 2022 at 02:29:09PM +0100, Al Viro wrote:
> > I wouldn't bet on that - as it is, an ->open() instance can decide
> > in some cases to clear FMODE_LSEEK, despite having file_operations
> > with non-NULL ->llseek.
> 
> The interesting cases here are nonseekable_open and stream_open,
> and I don't see why we could not fix this up in the file_operations.

What's the point, really?  We can easily enforce "no FMODE_LSEEK ever
observed on files with NULL ->llseek" (this series does that), so we
can use that check alone in e.g. vfs_llseek() or dump_skip().

Sure, we are tight on bits in ->f_mode, but there's a better way to
relieve that problem - split the field into "stuff that needs to
be preserved all the way until __fput()" and the rest; the latter
could sit next to ->f_iocb_flags, with no increase of struct file
size.

So if you are worried about FMODE_... space getting exhausted, that's
better dealt with in a different way, IMO.