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

["Raphael S. Carvalho" <raphaelsc@scylladb.com>]

> > A possible way to fix it is this one-liner, but I am not well versed
> > in this area, so someone may end up suggesting a better fix:
> > diff --git a/mm/filemap.c b/mm/filemap.c
> > index 804d7365680c..9e698a619545 100644
> > --- a/mm/filemap.c
> > +++ b/mm/filemap.c
> > @@ -1964,7 +1964,7 @@ struct folio *__filemap_get_folio(struct
> > address_space *mapping, pgoff_t index,
> >                 do {
> >                         gfp_t alloc_gfp = gfp;
> >
> > -                       err = -ENOMEM;
> > +                       err = (fgp_flags & FGP_NOWAIT) ? -ENOMEM : -EAGAIN;
>
> Sorry, I actually meant this:
> +                       err = (fgp_flags & FGP_NOWAIT) ? -EAGAIN : -ENOMEM;

Digging a bit more, I realized a better patch (assuming regression
indeed exists) is this one, since it accounts for ENOMEM coming from
filemap_add_folio, which might allocate in xas_split_alloc() under
same fgp flags:

diff --git a/mm/filemap.c b/mm/filemap.c
index 804d7365680c..dcf1f57e0a9a 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1984,6 +1984,8 @@ struct folio *__filemap_get_folio(struct
address_space *mapping, pgoff_t index,
                        folio = NULL;
                } while (order-- > min_order);

+               if ((fgp_flags & FGP_NOWAIT) && err == -ENOMEM)
+                       return ERR_PTR(-EAGAIN);
                if (err == -EEXIST)
                        goto repeat;
                if (err)