On Tue, Mar 08, 2016 at 07:16:40PM +0100, Christoph Hellwig wrote:
> Replace the homegrown RDMA READ/WRITE code in srpt with the generic API.
> The only real twist here is that we need to allocate one Linux scatterlist
> per direct buffer in the SRP command, and chain them before handing them
> off to the target core.
>
> As a side-effect of the conversion the driver will also chain the SEND
> of the SRP response to the RDMA WRITE WRs for a DATA OUT command, and
> properly account for RDMA WRITE WRs instead of just for RDMA READ WRs
> like the driver previously did.
>
> We now allocate half of the SQ size to RDMA READ/WRITE contexts, assuming
> by default one RDMA READ or WRITE operation per command. If a command
> has multiple operations it will eat into the budget but will still succeed,
> possible after waiting for WQEs to be available.
>
> Also ensure the QPs request the maximum allowed SGEs so that RDMA R/W API
> works correctly.
>
> Signed-off-by: Christoph Hellwig <hch@xxxxxx>
> ---
> drivers/infiniband/ulp/srpt/ib_srpt.c | 693 +++++++++++-----------------------
> drivers/infiniband/ulp/srpt/ib_srpt.h | 30 +-
> 2 files changed, 242 insertions(+), 481 deletions(-)
>
> diff --git a/drivers/infiniband/ulp/srpt/ib_srpt.c b/drivers/infiniband/ulp/srpt/ib_srpt.c
> index 25bdaee..05bdec8 100644
> --- a/drivers/infiniband/ulp/srpt/ib_srpt.c
> +++ b/drivers/infiniband/ulp/srpt/ib_srpt.c
> @@ -765,52 +765,6 @@ static int srpt_post_recv(struct srpt_device *sdev,
> }
>
> /**
> - * srpt_post_send() - Post an IB send request.
> - *
> - * Returns zero upon success and a non-zero value upon failure.
> - */
> -static int srpt_post_send(struct srpt_rdma_ch *ch,
> - struct srpt_send_ioctx *ioctx, int len)
> -{
> - struct ib_sge list;
> - struct ib_send_wr wr, *bad_wr;
> - struct srpt_device *sdev = ch->sport->sdev;
> - int ret;
> -
> - atomic_inc(&ch->req_lim);
> -
> - ret = -ENOMEM;
> - if (unlikely(atomic_dec_return(&ch->sq_wr_avail) < 0)) {
> - pr_warn("IB send queue full (needed 1)\n");
> - goto out;
> - }
> -
> - ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, len,
> - DMA_TO_DEVICE);
> -
> - list.addr = ioctx->ioctx.dma;
> - list.length = len;
> - list.lkey = sdev->pd->local_dma_lkey;
> -
> - ioctx->ioctx.cqe.done = srpt_send_done;
> - wr.next = NULL;
> - wr.wr_cqe = &ioctx->ioctx.cqe;
> - wr.sg_list = &list;
> - wr.num_sge = 1;
> - wr.opcode = IB_WR_SEND;
> - wr.send_flags = IB_SEND_SIGNALED;
> -
> - ret = ib_post_send(ch->qp, &wr, &bad_wr);
> -
> -out:
> - if (ret < 0) {
> - atomic_inc(&ch->sq_wr_avail);
> - atomic_dec(&ch->req_lim);
> - }
> - return ret;
> -}
> -
> -/**
> * srpt_zerolength_write() - Perform a zero-length RDMA write.
> *
> * A quote from the InfiniBand specification: C9-88: For an HCA responder
> @@ -843,6 +797,110 @@ static void srpt_zerolength_write_done(struct ib_cq *cq, struct ib_wc *wc)
> }
> }
>
> +static int srpt_alloc_rw_ctxs(struct srpt_send_ioctx *ioctx,
> + struct srp_direct_buf *db, int nbufs, struct scatterlist **sg,
> + unsigned *sg_cnt)
> +{
> + enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd);
> + struct srpt_rdma_ch *ch = ioctx->ch;
> + struct scatterlist *prev = NULL;
> + unsigned prev_nents;
> + int ret, i;
> +
> + if (nbufs == 1) {
> + ioctx->rw_ctxs = &ioctx->s_rw_ctx;
> + } else {
> + ioctx->rw_ctxs = kmalloc_array(nbufs, sizeof(*ioctx->rw_ctxs),
> + GFP_KERNEL);
> + if (!ioctx->rw_ctxs)
> + return -ENOMEM;
> + }
> +
> + for (i = 0; i < nbufs; i++, db++) {
> + struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
> + u64 remote_addr = be64_to_cpu(db->va);
> + u32 size = be32_to_cpu(db->len);
> + u32 rkey = be32_to_cpu(db->key);
> +
> + ret = target_alloc_sgl(&ctx->sg, &ctx->nents, size, false,
> + i < nbufs - 1);
> + if (ret)
> + goto unwind;
> +
> + ret = rdma_rw_ctx_init(&ctx->rw, ch->qp, ch->sport->port,
> + ctx->sg, ctx->nents, 0, remote_addr, rkey, dir);
> + if (ret < 0) {
> + target_free_sgl(ctx->sg, ctx->nents);
> + goto unwind;
> + }
> +
> + ioctx->n_rdma += ret;
> +
> + if (prev) {
> + sg_unmark_end(&prev[prev_nents - 1]);
> + sg_chain(prev, prev_nents + 1, ctx->sg);
> + } else {
> + *sg = ctx->sg;
> + }
> +
> + prev = ctx->sg;
> + prev_nents = ctx->nents;
> +
> + *sg_cnt += ctx->nents;
> + }
> +
> + ioctx->n_rw_ctx = nbufs;
> + return 0;
> +
> +unwind:
> + while (--i >= 0) {
> + struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
> +
> + rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port,
> + ctx->sg, ctx->nents, dir);
> + target_free_sgl(ctx->sg, ctx->nents);
> + }
> + if (ioctx->rw_ctxs && ioctx->rw_ctxs != &ioctx->s_rw_ctx)
You can jump to unwind label from one place only. In that stage
"ioctx->rw_ctxs != NULL" will always be true.
> + kfree(ioctx->rw_ctxs);
> + return ret;
> +}
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
