The patch introduces a new path of posting buffers from user-space to kernel post_send/post_recv using shared page.
The shared page is allocated and defined by vendor driver. The vendor driver during create QP decides when the new path will be used setting a field use_shpage_for_rxtx field in connection context. When the connection context indicates that the shared page should be used the uverbs_post_send()/uverbs_post_recv() calls directly the vendor driver private post_send()/post_recv() functions improving the kernel path. Signed-off-by: Mirek Walukiewicz <[email protected]> --- drivers/infiniband/core/uverbs_cmd.c | 55 +++++++++++++++++++++++++++------- include/rdma/ib_verbs.h | 1 + 2 files changed, 45 insertions(+), 11 deletions(-) diff --git a/drivers/infiniband/core/uverbs_cmd.c b/drivers/infiniband/core/uverbs_cmd.c index c426992..ba6655c 100644 --- a/drivers/infiniband/core/uverbs_cmd.c +++ b/drivers/infiniband/core/uverbs_cmd.c @@ -299,6 +299,7 @@ ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file, INIT_LIST_HEAD(&ucontext->srq_list); INIT_LIST_HEAD(&ucontext->ah_list); ucontext->closing = 0; + ucontext->use_shpage_for_rxtx = 0; resp.num_comp_vectors = file->device->num_comp_vectors; @@ -1448,15 +1449,31 @@ ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file, if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr)) return -EINVAL; + qp = idr_read_qp(cmd.qp_handle, file->ucontext); + if (!qp) + goto out_raw_qp; + + if (file->ucontext->use_shpage_for_rxtx) { + /* pass NULL pointers as the information about */ + /* buffers is passed using endor defined shared page */ + resp.bad_wr = 0; + ret = qp->device->post_send(qp, NULL, NULL); + if (ret) + resp.bad_wr = cmd.wr_count; + + if (copy_to_user((void __user *) (unsigned long) + cmd.response, + &resp, + sizeof resp)) + ret = -EFAULT; + put_qp_read(qp); + goto out_raw_qp; + } user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL); if (!user_wr) return -ENOMEM; - qp = idr_read_qp(cmd.qp_handle, file->ucontext); - if (!qp) - goto out; - is_ud = qp->qp_type == IB_QPT_UD; sg_ind = 0; last = NULL; @@ -1576,9 +1593,8 @@ out_put: wr = next; } -out: kfree(user_wr); - +out_raw_qp: return ret ? ret : in_len; } @@ -1680,16 +1696,33 @@ ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file, if (copy_from_user(&cmd, buf, sizeof cmd)) return -EFAULT; + qp = idr_read_qp(cmd.qp_handle, file->ucontext); + if (!qp) + goto out_raw_qp; + + if (file->ucontext->use_shpage_for_rxtx) { + resp.bad_wr = 0; + /* pass NULL pointers as the information about */ + /* buffers is passed using endor defined shared page */ + ret = qp->device->post_recv(qp, NULL, NULL); + if (ret) + resp.bad_wr = cmd.wr_count; + + if (copy_to_user((void __user *) (unsigned long) + cmd.response, + &resp, + sizeof resp)) + ret = -EFAULT; + put_qp_read(qp); + goto out_raw_qp; + } + wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd, in_len - sizeof cmd, cmd.wr_count, cmd.sge_count, cmd.wqe_size); if (IS_ERR(wr)) return PTR_ERR(wr); - qp = idr_read_qp(cmd.qp_handle, file->ucontext); - if (!qp) - goto out; - resp.bad_wr = 0; ret = qp->device->post_recv(qp, wr, &bad_wr); @@ -1706,13 +1739,13 @@ ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file, &resp, sizeof resp)) ret = -EFAULT; -out: while (wr) { next = wr->next; kfree(wr); wr = next; } +out_raw_qp: return ret ? ret : in_len; } diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h index 55cd0a0..ab31c03 100644 --- a/include/rdma/ib_verbs.h +++ b/include/rdma/ib_verbs.h @@ -831,6 +831,7 @@ struct ib_ucontext { struct list_head srq_list; struct list_head ah_list; int closing; + int use_shpage_for_rxtx; }; struct ib_uobject { -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
