Re: crush multiweight implementation details

[Sage Weil <sweil@redhat.com>]

On Mon, 10 Apr 2017, Loic Dachary wrote:
> Hi Sage,
> 
> We could have:
> 
> struct crush_choose_arg {
>   __u32 bucket_id;
>   __u32 num_items;
>   __u32 *ids; // override the bucket items for placement
>   __u32 num_positions;
>   __u32 *weights; // size is num_positions*num_items
> };
>  
> struct crush_choose_arg_map {
>   struct crush_choose_arg *args;
>   __u32 size;
> };
> 
> and
> 
> void crush_init_workspace(const struct crush_map *m, struct crush_choose_arg_map *arg_map, void *v) {
> ...
> if (m->buckets[b]->id == arg_map[b]->bucket_id)
>    w->work[b]->arg = arg_map[b];
> ...
> }
> 
> with
> 
> struct crush_work_bucket {
>     __u32 perm_x; /* @x for which *perm is defined */
>     __u32 perm_n; /* num elements of *perm that are permuted/defined */
>     __u32 *perm;  /* Permutation of the bucket's items */
>     struct crush_choose_arg *arg;
> };
> 
> There would be no need to change the code path since crush_bucket_choose 
> already is given the crush_work_bucket. And crush_init_workspace already 
> has logic that is algorithm dependent. And all the sanity checks could 
> be done in crush_init_workspace so that the choose function only does 
> what's absolutely necessary.

Allowing overrides of the bucket items too makes me nervous (do we have a 
use for that yet?), but otherwise this looks reasonable!

sage