Eric Biggers <ebiggers3@xxxxxxxxx> wrote: > Fix the bug by moving ->reject_error out of the union with ->payload, > then using nonzero ->reject_error to mean that the key is negative. > > This eliminates the need for KEY_FLAG_NEGATIVE, which we remove as well > so that we don't have to handle memory ordering between > KEY_FLAG_NEGATIVE and ->reject_error. We *do* still need to handle > memory ordering between KEY_FLAG_INSTANTIATED and ->reject_error, but > that was needed before (and for KEY_FLAG_NEGATIVE as well --- though it > wasn't done correctly, which was another bug). Hmmm... My only objection is that it extends struct key still further, but you do have a point. If we're going to do this, can we eliminate KEY_FLAG_INSTANTIATED also and make ->reject_error into ->status? 0 -> uninstantiated 1 -> instantiated <0 -> negatively instantiated/rejected Also, can we make it a short int please? Also, can setting this field be done with a release barrier rather than a write barrier? David