linux-kernel - Re: splice: infinite busy loop lockup bug

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20200809024957.GD2134904@T590>
Date:   Sun, 9 Aug 2020 10:49:57 +0800
From:   Ming Lei <ming.lei@...hat.com>
To:     Matthew Wilcox <willy@...radead.org>
Cc:     Al Viro <viro@...iv.linux.org.uk>,
        Tetsuo Handa <penguin-kernel@...ove.sakura.ne.jp>,
        syzkaller-bugs@...glegroups.com,
        syzbot <syzbot+61acc40a49a3e46e25ea@...kaller.appspotmail.com>,
        linux-fsdevel@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: splice: infinite busy loop lockup bug

On Sun, Aug 09, 2020 at 10:31:23AM +0800, Ming Lei wrote:
> On Fri, Aug 07, 2020 at 03:11:48PM +0100, Matthew Wilcox wrote:
> > On Fri, Aug 07, 2020 at 09:41:14PM +0800, Ming Lei wrote:
> > > On Fri, Aug 07, 2020 at 01:38:54PM +0100, Al Viro wrote:
> > > > FWIW, my preference would be to have for_each_bvec() advance past zero-length
> > > > segments; I'll need to go through its uses elsewhere in the tree first, though
> > > > (after I grab some sleep),
> > > 
> > > Usually block layer doesn't allow/support zero bvec, however we can make
> > > for_each_bvec() to support it only.
> > > 
> > > Tetsuo, can you try the following patch?
> > > 
> > > diff --git a/include/linux/bvec.h b/include/linux/bvec.h
> > > index ac0c7299d5b8..b03c793dd28d 100644
> > > --- a/include/linux/bvec.h
> > > +++ b/include/linux/bvec.h
> > > @@ -117,11 +117,19 @@ static inline bool bvec_iter_advance(const struct bio_vec *bv,
> > >  	return true;
> > >  }
> > >  
> > > +static inline void bvec_iter_skip_zero_vec(const struct bio_vec *bv,
> > > +		struct bvec_iter *iter)
> > > +{
> > > +	iter->bi_idx++;
> > > +	iter->bi_bvec_done = 0;
> > > +}
> > > +
> > >  #define for_each_bvec(bvl, bio_vec, iter, start)			\
> > >  	for (iter = (start);						\
> > >  	     (iter).bi_size &&						\
> > > -		((bvl = bvec_iter_bvec((bio_vec), (iter))), 1);	\
> > > -	     bvec_iter_advance((bio_vec), &(iter), (bvl).bv_len))
> > > +		((bvl = bvec_iter_bvec((bio_vec), (iter))), 1);		\
> > > +	  (bvl).bv_len ? bvec_iter_advance((bio_vec), &(iter), (bvl).bv_len) : \
> > > +			bvec_iter_skip_zero_vec((bio_vec), &(iter)))
> > 
> > Uhm, bvec_iter_advance() already skips over zero length bio_vecs.
> > 
> >         while (bytes && bytes >= bv[idx].bv_len) {
> >                 bytes -= bv[idx].bv_len;
> >                 idx++;
> >         }
> 
> The issue is that zero (bvl).bv_len passed to bvec_iter_advance(), so
> the iterator can't move on.
> 
> And I tried to avoid change to bvec_iter_advance() since this exact
> issue only exists on for_each_bvec, and block layer won't support/allow
> zero-length bvec.
> 
> > 
> > The problem is when the _first_ bio_vec is zero length.
> 
> It can be any zero-length bvec during the iterating. 
> 
> > Maybe something more
> > like this (which doesn't even compile, but hopefully makes my point):
> > 
> > @@ -86,12 +86,24 @@ struct bvec_iter_all {
> >         (mp_bvec_iter_page((bvec), (iter)) +                    \
> >          mp_bvec_iter_page_idx((bvec), (iter)))
> >  
> > -#define bvec_iter_bvec(bvec, iter)                             \
> > -((struct bio_vec) {                                            \
> > -       .bv_page        = bvec_iter_page((bvec), (iter)),       \
> > -       .bv_len         = bvec_iter_len((bvec), (iter)),        \
> > -       .bv_offset      = bvec_iter_offset((bvec), (iter)),     \
> > -})
> > +static inline bool bvec_iter_bvec(struct bio_vec *bv, struct bio_vec *bvec,
> > +               struct bvec_iter *iter)
> > +{
> > +       unsigned int idx = iter->bi_idx;
> > +
> > +       if (!iter->bi_size)
> > +               return false;
> > +
> > +       while (!bv[idx].bv_len)
> > +               idx++;
> > +       iter->bi_idx = idx;
> > +
> > +       bv->bv_page = bvec_iter_page(bvec, *iter);
> > +       bv->bv_len = bvec_iter_len(bvec, *iter);
> > +       bv->bv_offset = bvec_iter_offset(bvec, *iter);
> > +
> > +       return true;
> > +}
> >  
> >  static inline bool bvec_iter_advance(const struct bio_vec *bv,
> >                 struct bvec_iter *iter, unsigned bytes)
> > @@ -119,8 +131,7 @@ static inline bool bvec_iter_advance(const struct bio_vec *bv,
> >  
> >  #define for_each_bvec(bvl, bio_vec, iter, start)                       \
> >         for (iter = (start);                                            \
> > -            (iter).bi_size &&                                          \
> > -               ((bvl = bvec_iter_bvec((bio_vec), (iter))), 1); \
> > +            bvec_iter_bvec(&(bvl), (bio_vec), &(iter));                \
> >              bvec_iter_advance((bio_vec), &(iter), (bvl).bv_len))
> >  
> >  /* for iterating one bio from start to end */
> > 
> > (I find the whole bvec handling a mess of confusing macros and would
> > welcome more of it being inline functions, in general).
> 
> The above change may bring more code duplication. Meantime, it can't
> work because (bvl).bv_len isn't taken into account into bvec_iter_bvec(),
> then how can the iterator advance?

oops, looks you change bvec_iter_bvec() only for skipping zero length bvec,
and this way might work(still ->bi_bvec_done isn't reset). However the
change is ugly, cause the iterator is supposed to not be updated in
bvec_iter_bvec(). Also block layer code doesn't require such change.

BTW, I agree on switching to inline if performance isn't affected.


Thanks,
Ming