Re: [PATCH 01/17] Add ida and idr helper routines.

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Tue, 2015-09-15 at 14:27 -0400, Tejun Heo wrote:
> Hello,
> 
> On Tue, Sep 15, 2015 at 09:46:01AM -0700, Lee Duncan wrote:
> > +/**
> > + * ida_get_index - allocate a ida index value
> > + * @ida		idr handle
> > + * @lock	spinlock handle protecting this index
> > + * @p_id	pointer to allocated index value
> > + *
> > + * A helper function for safely allocating an index value (id),
> > + * returning a negative errno value on failure, else 0.
> > + */
> > +static inline int ida_get_index(struct ida *ida, spinlock_t *lock, int *p_id)
> > +{
> > +	int error = -ENOMEM;
> > +
> > +	do {
> > +		if (!ida_pre_get(ida, GFP_KERNEL))
> > +			break;
> > +		spin_lock(lock);
> > +		error = ida_get_new(ida, p_id);
> > +		spin_unlock(lock);
> > +	} while (error == -EAGAIN);
> > +
> > +	return error;
> > +}
> 
> Obviously ida allocation doesn't need to be synchronized against
> anything else.  Why not just use ida_simple_get/remove()?

For most of the SCSI stuff, yes.  I'm less sure about the sd numbers.
They go up very high and get hammered a lot during system bring up and
hot plug.  I think having their own lock rather than wrapping everything
around simple_ida_lock makes more sense here just because the system is
heavily contended on getting indexes at bring up.

To continue the thought, why not move simple_ida_lock into struct ida so
we don't have to worry about the contention and can sue ida_simple_...
everywhere?

James