On Thu, Sep 20, 2012 at 05:25:55PM -0700, Tejun Heo wrote:
> Hello, Kent.
>
> On Mon, Sep 10, 2012 at 05:22:30PM -0700, Kent Overstreet wrote:
> > A bunch of what __blk_queue_bounce() was doing was problematic for the
> > immutable bvec work; this cleans that up and the code is quite a bit
> > smaller, too.
> >
> > The __bio_for_each_segment() in copy_to_high_bio_irq() was changed
> > because that one's looping over the original bio, not the bounce bio -
> > since the bounce code doesn't own that bio the __ version wasn't
> > correct.
>
> I do like the new implementation. I think the function is broken
> before and after tho. Allocating from fs_bio_set from block layer is
> never safe and nothing seems to prevent multiple allocators compete in
> the bounce page mempool. This will need a separate bioset and the
> multiple mempool allocation would have to be put inside a mutex.

Yeah, I should've at least made a note of that.

I should really add "audit all uses of fs_bio_set" to my todo list.

> Also, how was this tested?

Changed queue_bounce_pfn() to return 0, forcing all io to be bounced.

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel

On Thu, Sep 20, 2012 at 05:27:06PM -0700, Tejun Heo wrote:
> On Mon, Sep 10, 2012 at 05:22:30PM -0700, Kent Overstreet wrote:
> > A bunch of what __blk_queue_bounce() was doing was problematic for the
> > immutable bvec work; this cleans that up and the code is quite a bit
> > smaller, too.
> >
> > The __bio_for_each_segment() in copy_to_high_bio_irq() was changed
> > because that one's looping over the original bio, not the bounce bio -
> > since the bounce code doesn't own that bio the __ version wasn't
> > correct.
>
> Also, I can't understand the above at all. I can think why it
> wouldn't be necessary but why is it wrong because bounce code doesn't
> own it?

Another prep work thing - in current code, it isn't really wrong
(slightly inconsistent though).

But the idea is that anything that doesn't own the bio shouldn't assume
anything about bi_idx; the bounce code should loop over the bio starting
from wherever it was when the bio got to the bounce code, not the start
of the bio.

A later patch makes this clearer - __bio_for_each_segment() gets removed
in favor of bio_for_each_segment_all(), and it documents that
bio_for_each_segment_all() is only for code that owns the bio.

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel