Re: [PATCH] block: Make __bio_clone_fast() copy bi_vcnt

[Kent Overstreet <kent.overstreet@gmail.com>]

On Thu, Jun 28, 2018 at 04:54:44PM -0700, Bart Van Assche wrote:
> On 06/28/18 16:16, Kent Overstreet wrote:
> > On Fri, Jun 29, 2018 at 07:10:47AM +0800, Ming Lei wrote:
> > > On Thu, Jun 28, 2018 at 11:21 PM, Bart Van Assche
> > > <bart.vanassche@wdc.com> wrote:
> > > > On 06/27/18 17:30, Ming Lei wrote:
> > > > > 
> > > > > One core idea of immutable bvec is to use bio->bi_iter and the original
> > > > > bvec table to iterate over anywhere in the bio. That is why .bi_io_vec
> > > > > needs to copy, but not see any reason why .bi_vcnt needs to do.
> > > > > 
> > > > > Do you have use cases on .bi_vcnt for cloned bio?
> > > > 
> > > > So far this is only a theoretical concern. There are many functions in the
> > > > block layer that use .bi_vcnt, and it is a lot of work to figure out all the
> > > > callers of all these functions.
> > 
> > Back when I implemented immutable biovecs I thoroughly audited all the bi_vcnt
> > uses and removed all of them that weren't by the code that owns/submits the bio.
> > 
> > Grepping around I see one or two suspicious uses.. blk-merge.c in particular
> > 
> > > No, any functions using .bi_vcnt on a cloned-bio may be a bug, and we should
> > > take a close look.
> > 
> > not just cloned bios, any code using bi_vcnt on a bio it didn't create is wrong.
> > 
> > so big nack to this patch (I wasn't ccd on it though and it doesn't seem to have
> > hit lkml, so I can't find the original patch...)
> 
> Hello Kent,
> 
> Thanks for chiming in. The linux-block mailing list is archived by multiple
> websites. The entire e-mail thread is available on e.g.
> https://www.mail-archive.com/linux-block@vger.kernel.org/msg23006.html.
> 
> I have a question for you: at least in kernel v4.17 bio_clone_bioset()
> copies bi_vcnt from the source to the destination bio. However,
> __bio_clone_fast() doesn't copy bi_vcnt. Isn't that an inconsistency?

No - when you use bio_clone_bioset() you get a bio that you own and can do
whatever you want with, so it does make sense for it to initialize bi_vcnt.

e.g. you could use bio_clone_bioset() when you're going to be bouncing a bio,
iterating over each bvec and allocating a new page and copying data from the old
page to the new page.