[Drbd-dev] [PATCH v5 11/12] block: Add bio_clone_bioset()

[Kent Overstreet]

On Wed, Aug 08, 2012 at 11:52:51PM -0700, Tejun Heo wrote:
> On Wed, Aug 08, 2012 at 07:56:10PM -0700, Kent Overstreet wrote:
> > > So, bio_clone() loses its function comment.  Also, does it even make
> > > sense to call bio_clone() from fs_bio_set?
> > 
> > I'll re add the function comment if you want, just for a single line
> > wrapper I don't know if it's worth the cost - comments get out of date,
> > and they're more stuff to wade through.
> 
> People actually look at docbook generated docs.  I don't know why but
> they do.  It's a utility function at block layer.  Please just add the
> comment.

Will do then.

> > > Let's say it's so, then
> > > what's the difference from using _kmalloc variant?
> > 
> > bio_kmalloc() fails if nr_iovecs > 1024, bio_alloc_bioset() fails if
> > nr_iovecs > 256
> > 
> > and bio_alloc_bioset() is mempool backed, bio_kmalloc() is not.
> > 
> > AFAICT that's it.
> 
> So, the thing is being mempool backed doesn't mean anything if
> multiple layers use the pool. 

It's worse than just using kmalloc, because then you've introduced the
possibility of deadlock.

> I *suspect* fs_bio_set is supposed to
> be used by fs layer - ie. where bios originate.  The reason why I
> wondered about bio_clone() is that bio_clone() is almost always used
> from stacking drivers and stacking driver tapping into fs reserve is
> buggy.  So, I'm wondering whether cloning from fs_bio_set should be
> supported at all.

That's actually a really good point.

I just grepped and there's actually only 3 callers - I thought there'd
be more.  That should be easy to fix, at least.