Re: splice: infinite busy loop lockup bug

From: Ming Lei
Date: Sat Aug 08 2020 - 22:50:28 EST


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