Hi Marcel
On 22/08/2021 16:39, Marcel Apfelbaum wrote: > Hi, > > On Sat, Jul 31, 2021 at 5:03 PM Li Zhijian <lizhij...@cn.fujitsu.com> wrote: >> The responder mr registering with ODP will sent RNR NAK back to >> the requester in the face of the page fault. >> --------- >> ibv_poll_cq wc.status=13 RNR retry counter exceeded! >> ibv_poll_cq wrid=WRITE RDMA! >> --------- >> ibv_advise_mr(3) helps to make pages present before the actual IO is >> conducted so that the responder does page fault as little as possible. >> >> Signed-off-by: Li Zhijian <lizhij...@cn.fujitsu.com> >> --- >> migration/rdma.c | 40 ++++++++++++++++++++++++++++++++++++++++ >> migration/trace-events | 1 + >> 2 files changed, 41 insertions(+) >> >> diff --git a/migration/rdma.c b/migration/rdma.c >> index 8784b5f22a6..a2ad00d665f 100644 >> --- a/migration/rdma.c >> +++ b/migration/rdma.c >> @@ -1117,6 +1117,30 @@ static int qemu_rdma_alloc_qp(RDMAContext *rdma) >> return 0; >> } >> >> +/* >> + * ibv_advise_mr to avoid RNR NAK error as far as possible. >> + * The responder mr registering with ODP will sent RNR NAK back to >> + * the requester in the face of the page fault. >> + */ >> +static void qemu_rdma_advise_prefetch_write_mr(struct ibv_pd *pd, uint64_t >> addr, >> + uint32_t len, uint32_t lkey, >> + const char *name, bool wr) >> +{ >> + int ret; >> + int advice = wr ? IBV_ADVISE_MR_ADVICE_PREFETCH_WRITE : >> + IBV_ADVISE_MR_ADVICE_PREFETCH; >> + struct ibv_sge sg_list = {.lkey = lkey, .addr = addr, .length = len}; >> + >> + ret = ibv_advise_mr(pd, advice, >> + IB_UVERBS_ADVISE_MR_FLAG_FLUSH, &sg_list, 1); >> + /* ignore the error */ > Following > https://github.com/linux-rdma/rdma-core/blob/master/libibverbs/man/ibv_advise_mr.3.md > it looks like it is a best-effort optimization, > I don't see any down-sides to it. > However it seems like it is recommended to use > IBV_ADVISE_MR_FLAG_FLUSH in order to > increase the optimization chances. Good catch, i will update it soon. Thanks > > Anyway > > Reviewed-by: Marcel Apfelbaum <marcel.apfelb...@gmail.com> > > Thanks, > Marcel > >