On Mon, 2025-02-10 at 09:27 +0100, Vlastimil Babka wrote:
> On 2/8/25 16:46, Joanne Koong wrote:
> > On Sat, Feb 8, 2025 at 2:11 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > >
> > > On Fri, Feb 07, 2025 at 04:22:56PM -0800, Joanne Koong wrote:
> > > > > Thanks, Josef. I guess we can at least try to confirm we're on the right track.
> > > > > Can anyone affected see if this (only compile tested) patch fixes the issue?
> > > > > Created on top of 6.13.1.
> > > >
> > > > This fixes the crash for me on 6.14.0-rc1. I ran the repro using
> > > > Mantas's instructions for Obfuscate. I was able to trigger the crash
> > > > on a clean build and then with this patch, I'm not seeing the crash
> > > > anymore.
> > >
> > > Since this patch fixes the bug, we're looking for one call to folio_put()
> > > too many. Is it possibly in fuse_try_move_page()? In particular, this
> > > one:
> > >
> > > /* Drop ref for ap->pages[] array */
> > > folio_put(oldfolio);
> > >
> > > I don't know fuse very well. Maybe this isn't it.
> >
> > Yeah, this looks it to me. We don't grab a folio reference for the
> > ap->pages[] array for readahead and it tracks with Mantas's
> > fuse_dev_splice_write() dmesg. this patch fixed the crash for me when
> > I tested it yesterday:
> >
> > diff --git a/fs/fuse/file.c b/fs/fuse/file.c
> > index 7d92a5479998..172cab8e2caf 100644
> > --- a/fs/fuse/file.c
> > +++ b/fs/fuse/file.c
> > @@ -955,8 +955,10 @@ static void fuse_readpages_end(struct fuse_mount
> > *fm, struct fuse_args *args,
> > fuse_invalidate_atime(inode);
> > }
> >
> > - for (i = 0; i < ap->num_folios; i++)
> > + for (i = 0; i < ap->num_folios; i++) {
> > folio_end_read(ap->folios[i], !err);
> > + folio_put(ap->folios[i]);
> > + }
> > if (ia->ff)
> > fuse_file_put(ia->ff, false);
> >
> > @@ -1049,6 +1051,7 @@ static void fuse_readahead(struct readahead_control *rac)
> >
> > while (ap->num_folios < cur_pages) {
> > folio = readahead_folio(rac);
> > + folio_get(folio);
>
> This is almost the same as my patch, but balances the folio_put() in
> readahead_folio() with another folio_get(), while mine uses
> __readahead_folio() that does not do folio_put() in the first place.
>
> But I think neither patch proves the extraneous folio_put() comes from
> fuse_try_move_page().
>
> > ap->folios[ap->num_folios] = folio;
> > ap->descs[ap->num_folios].length = folio_size(folio);
> > ap->num_folios++;
> >
> >
> > I reran it just now with a printk by that ref drop in
> > fuse_try_move_page() and I'm indeed seeing that path get hit.
>
> It might get hit, but is it hit in the readahead paths? One way to test
> would be to instead of yours above or mine change, to stop doing the
> foio_put() in fuse_try_move_page(). But maybe it's called also from other
> contexts that do expect it, and will leak memory otherwise.
>
I think you're right that there is a double put in
fuse_try_move_page(). Let's assume that we enter that function and the
refcount on "oldpage" is 1:
1/ We take a reference to "oldfolio" when we enter the function, now
refcount is 2.
2/ We drop a reference on "oldfolio" with the call to
replace_page_cache_folio. Now refcount is 1.
3/ Now there are 2 folio_put(oldfolio) calls on the way out of the
function, refcount goes to -1.
Maybe it's expected that this function consumes an extra folio
reference, but it's certainly not evident why that is if so. I don't
see why the callers would expect that either.
> > Not sure why fstests didn't pick this up though since splice is
> > enabled by default in passthrough_hp, i'll look into this next week.
> >
>
>
--
Jeff Layton <jlayton@xxxxxxxxxx>
