kiobuf setup overhead

[Eric Barton]

I've noticed that alloc_kiovec() calls alloc_kiobuf_bhs(), which...

> int alloc_kiobuf_bhs(struct kiobuf * kiobuf)
> {
> 	int i;
> 
> 	for (i = 0; i < KIO_MAX_SECTORS; i++)
> 		if (!(kiobuf->bh[i] = kmem_cache_alloc(bh_cachep, SLAB_KERNEL))) {
> 			while (i--) {
> 				kmem_cache_free(bh_cachep, kiobuf->bh[i]);
> 				kiobuf->bh[i] = NULL;
> 			}
> 			return -ENOMEM;
> 		}
> 	return 0;
> }

...where KIO_MAX_SECTORS is currently 1024.  This makes a kiobuf quite a
heavy-weight object to allocate and free.  

Was this high-overhead initialisation left in because kiobufs shouldn't be
allocated and freed at the drop of a hat in any case?

If not, allocating the kiobuf->bh[i], could be delayed until the number
required is known.

This would be beneficial for direct I/O in the scsi generic driver, which
currently allocates and frees a kiobuf on each read/write.

-- 

                Cheers,
                        Eric