Re: Possible regression with buffered writes + NOWAIT behavior, under memory pressure

[Dave Chinner <david@fromorbit.com>]

On Mon, Feb 10, 2025 at 09:18:04PM +0000, Matthew Wilcox wrote:
> On Tue, Feb 11, 2025 at 08:09:31AM +1100, Dave Chinner wrote:
> > Better to only do the FGP_NOWAIT check when a failure occurs; that
> > puts it in the slow path rather than having to evaluate it
> > unnecessarily every time through the function/loop. i.e.
> > 
> >  		folio = filemap_alloc_folio(gfp, order);
> > -		if (!folio)
> > -			return ERR_PTR(-ENOMEM);
> > +		if (!folio) {
> > +			if (fgp_flags & FGP_NOWAIT)
> > +				err = -EAGAIN;
> > +			else
> > +				err = -ENOMEM;
> > +			continue;
> > +		}
> 
> Or would we be better off handling ENOMEM the same way we handle EAGAIN?
> eg something like:
> 
> +++ b/io_uring/io_uring.c
> @@ -1842,7 +1842,7 @@ void io_wq_submit_work(struct io_wq_work *work)
> 
>         do {
>                 ret = io_issue_sqe(req, issue_flags);
> -               if (ret != -EAGAIN)
> +               if (ret != -EAGAIN || ret != -ENOMEM)
>                         break;

This still allows -ENOMEM to escape to userspace instead of -EAGAIN
via pwritev2(RWF_NOWAIT) and AIO write interfaces.

-Dave.
-- 
Dave Chinner
david@fromorbit.com