Re: [PATCH 1/1] [virt] virtio-blk: Use ida to allocate disk index

[Rusty Russell <rusty@rustcorp.com.au>]

On Wed, 15 Jun 2011 09:06:38 +0200, Tejun Heo <tj@kernel.org> wrote:
> It's inherited from idr which was designed to have separate
> prepare/allocation stages so that allocation can happen inside an
> outer spinlock.  It doesn't have too much to do with optimization.
> It's mostly to be able to use sleepable context for memory allocation
> while allowing atomic id[ra] allocation.

It might have made sense for a few callers, but as a general mechanism
it stinks.  It's a lot of dancing to avoid GFP_ATOMIC allocations; we'd
be better making idr_get_new() take a gfp_t, and have an idr_pre_alloc()
for those who care.

*Sure* there's a chance of racing and we will need to do an atomic
allocation.  But can anyone justify the current complexity for all
callers?

> > + * ida_simple_get - get a new id.
> > + * @ida: the (initialized) ida.
> > + * @min_id: the minimum id (inclusive)
> > + * @max_id: the maximum id (inclusive)
> > + *
> > + * Allocates an id in the range min_id <= id <= max_id, or returns -ENOSPC.
> > + * On allocation failure, returns -ENOMEM.  This function can sleep.
> > + *
> > + * Use ida_simple_remove() to get rid of an id.
> > + */
> > +int ida_simple_get(struct ida *ida, int min_id, int max_id)
> 
> Hmmm... new interface different from existing id[ra] style, but yeah
> something like the above would have made more sense from the
> beginning.  The only thing is that isn't (begin <= range < end) more
> conventional form to express ranges?

Yes, but how to express an unlimited range then?  I could used unsigned
and 0x80000000, but that seemed crude.

Cheers,
Rusty.