[dm-devel] [PATCH 2/2] block: Avoid deadlocks with bio allocation by stacking drivers

[Kent Overstreet]

On Mon, Sep 10, 2012 at 04:01:01PM -0700, Tejun Heo wrote:
> Hello,
> 
> On Mon, Sep 10, 2012 at 3:50 PM, Alasdair G Kergon <agk at redhat.com> wrote:
> >> >     Note that this doesn't do anything for allocation from other mempools.
> >
> > Note that dm has several cases of this, so this patch should not be used with
> > dm yet.  Mikulas is studying those cases to see whether anything like this
> > might be feasible/sensible or not.
> 
> IIUC, Kent posted a patch which converts all of them to use front-pad
> (there's no reason not to, really). This better come after that but
> it's not like this is gonna break something which isn't broken now.

Not all, I only did the easy one - you know how dm has all those crazy
abstraction layers? They've got multiple per bio allocations because of
that; the core dm code does one, and then some other code takes that
struct dm_io* and allocates its own state pointing to that (which then
points to the original bio...)

So front_pad should still work, but you need to have say dm_crypt pass
the amount of front pad it needs to the core dm code when it creates the
bio_set, and then dm crypt can use container_of(struct dm_io) and embed
like everything does that use the bio_set front pad.

*I'm probably misremembering all the names.