Re: [Resend] 3.2-rc6+: Reported regressions from 3.0 and 3.1

[Hugh Dickins <hughd@google.com>]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 2927 bytes --]

On Tue, 20 Dec 2011, Linus Torvalds wrote:
> On Tue, Dec 20, 2011 at 8:23 PM, Dave Kleikamp <dave.kleikamp@oracle.com> wrote:
> >
> > I don't think this is a regression.  It's been seen before, but the
> > patch never got submitted, or was lost somewhere. I believe this
> > will fix it.
> 
> Hmm. This patch looks obviously correct. But it looks *so* obviously
> correct that it just makes me suspicious - this is not new or seldom
> used code, it's been this way for ages and used all the time. That
> line literally goes back to 2007, commit eb2be189317d0. And it looks
> like even before that we had a GFP_KERNEL for the add_to_page_cache()
> case and that goes back to before the git history. So this is
> *ancient*.
> 
> Maybe almost nobody uses __read_cache_page() with a non-GFP_KERNEL gfp
> and as a result we've not noticed.
> 
> Or maybe there is some crazy reason why it calls "add_to_page_cache()"
> with GFP_KERNEL.
> 
> Adding the usual suspects for mm/filemap.c to the cc line (Andrew is
> already cc'd, but Al and Hugh should comment).
> 
> Ack's, people? Is it really as obvious as it looks, and we've just had
> this bug forever?

Certainly

Acked-by: Hugh Dickins <hughd@google.com>

from me (and add_to_page_cache_locked does the masking of inappropriate
bits when passing on down, so no need to worry about that aspect).

I agree that it's odd that we've never noticed it before, but I don't
think the GFP_KERNEL there has any more significance than oversight.
Nick cleaned up some similar instances in filemap.c a few years back,
I guess ones he hit in testing, but this just got left over.

page_cache_read()'s GFP_KERNEL looks similarly worrying, but as it's
only called by filemap_fault(), I suppose it's actually okay.

Ooh, maybe you should also update that comment on GFP_KERNEL above
read_cache_page_gfp()...

Hugh

> 
>             Linus
> 
> --- snip snip ---
> > vfs: __read_cache_page should use gfp argument rather than GFP_KERNEL
> >
> > lockdep reports a deadlock in jfs because a special inode's rw semaphore
> > is taken recursively. The mapping's gfp mask is GFP_NOFS, but is not used
> > when __read_cache_page() calls add_to_page_cache_lru().
> >
> > Signed-off-by: Dave Kleikamp <dave.kleikamp@oracle.com>
> >
> > diff --git a/mm/filemap.c b/mm/filemap.c
> > index c106d3b..c9ea3df 100644
> > --- a/mm/filemap.c
> > +++ b/mm/filemap.c
> > @@ -1828,7 +1828,7 @@ repeat:
> >                page = __page_cache_alloc(gfp | __GFP_COLD);
> >                if (!page)
> >                        return ERR_PTR(-ENOMEM);
> > -               err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
> > +               err = add_to_page_cache_lru(page, mapping, index, gfp);
> >                if (unlikely(err)) {
> >                        page_cache_release(page);
> >                        if (err == -EEXIST)