On Tue, Nov 07, 2006 at 11:16:13AM +0000, Keir Fraser wrote: > Applied, except for one chunk which added a barrier before update of
> req_cons. This was not mentioned in the changeset comment and I do not
> believe there is any race that needs fixing (certainly not between memcpy of
> request data and read of req_cons in make_response()). I applied the rest of
> the patch as-is.
Thanks you for applying.
Hmm.. since you don't believe any race, then please revert
this mb patch except NULL check chunk.
I certainly observed that tapdisk failed to get new request on IA64.
If it is really caused by the race, I will sent a patch again.
(or I'll sent another patch which fixes another issues)
why is wmb() in write_rsp_to_ring() of tools/blktap/drivers/tapdisk.c
RING_PUSH_RESPONSES() of kick_responses() issues wmb() so that
wmb() in write_rsp_to_ring() isn't needed, I think.
Xen-devel mailing list