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 */
> + if (ret) {
> + trace_qemu_rdma_advise_mr(name, len, addr, strerror(errno));
> + } else {
> + trace_qemu_rdma_advise_mr(name, len, addr, "successed");
> + }
> +}
> +
> static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma)
> {
> int i;
> @@ -1140,6 +1164,17 @@ on_demand:
> perror("Failed to register local dest ram block!\n");
> break;
> }
> +
> + if (access & IBV_ACCESS_ON_DEMAND) {
> + qemu_rdma_advise_prefetch_write_mr(rdma->pd,
> + (uintptr_t)
> +
> local->block[i].local_host_addr,
> + local->block[i].length,
> + local->block[i].mr->lkey,
> + local->block[i].block_name,
> + true);
> + }
> +
> rdma->total_registrations++;
> }
>
> @@ -1244,6 +1279,11 @@ on_demand:
> rdma->total_registrations);
> return -1;
> }
> + if (access & IBV_ACCESS_ON_DEMAND) {
> + qemu_rdma_advise_prefetch_write_mr(rdma->pd,
> (uintptr_t)chunk_start,
> + len, block->pmr[chunk]->lkey,
> + block->block_name, rkey);
> + }
> rdma->total_registrations++;
> }
>
> diff --git a/migration/trace-events b/migration/trace-events
> index 5f6aa580def..901c1d54c12 100644
> --- a/migration/trace-events
> +++ b/migration/trace-events
> @@ -213,6 +213,7 @@ qemu_rdma_poll_other(const char *compstr, int64_t comp,
> int left) "other complet
> qemu_rdma_post_send_control(const char *desc) "CONTROL: sending %s.."
> qemu_rdma_register_and_get_keys(uint64_t len, void *start) "Registering %"
> PRIu64 " bytes @ %p"
> qemu_rdma_register_odp_mr(const char *name) "Try to register On-Demand
> Paging memory region: %s"
> +qemu_rdma_advise_mr(const char *name, uint32_t len, uint64_t addr, const
> char *res) "Try to advise block %s prefetch write at %" PRIu32 "@0x%" PRIx64
> ": %s"
> qemu_rdma_registration_handle_compress(int64_t length, int index, int64_t
> offset) "Zapping zero chunk: %" PRId64 " bytes, index %d, offset %" PRId64
> qemu_rdma_registration_handle_finished(void) ""
> qemu_rdma_registration_handle_ram_blocks(void) ""
> --
> 2.31.1
>
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.
Anyway
Reviewed-by: Marcel Apfelbaum <marcel.apfelb...@gmail.com>
Thanks,
Marcel
