Re: [PATCH RERESEND 10/11] splice: file->pipe: -EINVAL for non-regular files w/o FMODE_NOWAIT

[Jens Axboe]

On 12/14/23 11:45 AM, Ahelenia Ziemia?ska wrote:
> We request non-blocking I/O in the generic implementation, but some
> files ? ttys ? only check O_NONBLOCK. Refuse them here, lest we
> risk sleeping with the pipe locked for indeterminate lengths of
> time.

A worthy goal here is ensuring that _everybody_ honors IOCB_NOWAIT,
rather than just rely on O_NONBLOCK. This does involve converting to
->read_iter/->write_iter if the driver isn't already using it, but some
of them already have that, yet don't check IOCB_NOWAIT or treat it the
same as O_NONBLOCK.

Adding special checks like this is not a good idea, imho.

> This also masks inconsistent wake-ups (usually every second line)
> when splicing from ttys in icanon mode.
> 
> Regular files don't /have/ a distinct O_NONBLOCK mode,
> because they always behave non-blockingly, and for them FMODE_NOWAIT is
> used in the purest sense of
>   /* File is capable of returning -EAGAIN if I/O will block */
> which is not set by the vast majority of filesystems,
> and it's not the semantic we want here.

The main file systems do very much set it, like btrfs, ext4, and xfs. If
you look at total_file_systems / ones_flagging_it the ratio may be high,
but in terms of installed userbase, the majority definitely will have
it. Also see comment on cover letter for addressing this IOCB_NOWAIT
confusion.

-- 
Jens Axboe