Quoting Dan Smith (danms@xxxxxxxxxx):
> SH> I don't think this part of the comment is quite right here. The
> SH> "on failure this cleans up the object itself" really is for the
> SH> ref_drop under IS_ERR() check below.
>
> SH> The ref_drop here is for the ref taken by obj_new(), which is only
> SH> done in this path of course.
>
> As just discussed on IRC, I couldn't really correct the comment
> because the logic it was describing was incorrect. Below is a version
> with the correct comment *and* logic :)
>
> --
> Dan Smith
> IBM Linux Technology Center
> email: danms@xxxxxxxxxx
>
> Make restore_obj() tolerate a preexisting object in the hash (v3)
>
> ... as long as the pointer is the same as that returned from the restore
> function. Also move the compulsory ref_drop() so that it only gets
> done if we created the new object.
>
> The existing object tolerance is important for netdev restore because it
> means that I can refer to a peer by its objref instead of needing the
> (previously-rejected) veth_peer() function. If this is not acceptable,
> then I'll need to keep a separate list of pairs.
>
> Changes in v3:
> - Fix the logic in the case where we need to do an obj_new() and fail,
> so that we don't do ref_drop() twice
>
> Changes in v2:
> - Check that the type of the object already in the hash matches that
> of the objref header we're reading.
> - Add a comment about why and how we might get into this sort of
> situation.
>
> Signed-off-by: Dan Smith <danms@xxxxxxxxxx>
Thanks, Dan, for also fixing pre-existing bug :)
Acked-by: Serge Hallyn <serue@xxxxxxxxxx>
-serge
>
> diff --git a/checkpoint/objhash.c b/checkpoint/objhash.c
> index 7208382..3b360cb 100644
> --- a/checkpoint/objhash.c
> +++ b/checkpoint/objhash.c
> @@ -1064,17 +1064,32 @@ int restore_obj(struct ckpt_ctx *ctx, struct ckpt_hdr_objref *h)
> if (IS_ERR(ptr))
> return PTR_ERR(ptr);
>
> - if (obj_find_by_objref(ctx, h->objref))
> - obj = ERR_PTR(-EINVAL);
> - else
> + obj = obj_find_by_objref(ctx, h->objref);
> + if (!obj) {
> obj = obj_new(ctx, ptr, h->objref, h->objtype);
> - /*
> - * Drop an extra reference to the object returned by ops->restore:
> - * On success, this clears the extra reference taken by obj_new(),
> - * and on failure, this cleans up the object itself.
> - */
> - ops->ref_drop(ptr, 0);
> + /*
> + * Drop an extra reference to the object returned by
> + * ops->restore to balance the one taken by obj_new()
> + */
> + if (!IS_ERR(obj))
> + ops->ref_drop(ptr, 0);
> + } else if ((obj->ptr != ptr) || (obj->ops->obj_type != h->objtype)) {
> + /* Normally, we expect an object to not already exist
> + * in the hash. However, for some special scenarios
> + * where we're restoring sets of objects that must be
> + * co-allocated (such, as veth netdev pairs) we need
> + * to tolerate this case if the second restore returns
> + * the correct type and pointer, as specified in the
> + * existing object. If either of those don't match,
> + * we fail.
> + */
> + obj = ERR_PTR(-EINVAL);
> + }
> +
> if (IS_ERR(obj)) {
> + /* This releases our final reference on the object
> + * returned by ops->restore()
> + */
> ops->ref_drop(ptr, 1);
> return PTR_ERR(obj);
> }
_______________________________________________
Containers mailing list
Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linux-foundation.org/mailman/listinfo/containers
