On Tue, Feb 19, 2019 at 11:08:56AM +0100, Bernard Metzler wrote:
> +/*
> + * siw_create_qp()
> + *
> + * Create QP of requested size on given device.
> + *
> + * @base_pd: Base PD contained in siw PD
> + * @attrs: Initial QP attributes.
> + * @udata: used to provide QP ID, SQ and RQ size back to user.
> + */
> +
> +struct ib_qp *siw_create_qp(struct ib_pd *base_pd,
> + struct ib_qp_init_attr *attrs,
> + struct ib_udata *udata)
> +{
> + struct siw_qp *qp = NULL;
> + struct siw_pd *pd = to_siw_pd(base_pd);
> + struct ib_device *base_dev = base_pd->device;
> + struct siw_device *sdev = to_siw_dev(base_dev);
> + struct siw_cq *scq = NULL, *rcq = NULL;
> +
> + unsigned long flags;
> + int num_sqe, num_rqe, rv = 0;
> +
> + siw_dbg(sdev, "create new qp\n");
> +
> + if (atomic_inc_return(&sdev->num_qp) > SIW_MAX_QP) {
> + siw_dbg(sdev, "too many qp's\n");
> + rv = -ENOMEM;
> + goto err_out;
> + }
> + if (attrs->qp_type != IB_QPT_RC) {
> + siw_dbg(sdev, "only rc qp's supported\n");
> + rv = -EINVAL;
> + goto err_out;
> + }
> + if ((attrs->cap.max_send_wr > SIW_MAX_QP_WR) ||
> + (attrs->cap.max_recv_wr > SIW_MAX_QP_WR) ||
> + (attrs->cap.max_send_sge > SIW_MAX_SGE) ||
> + (attrs->cap.max_recv_sge > SIW_MAX_SGE)) {
> + siw_dbg(sdev, "qp size error\n");
> + rv = -EINVAL;
> + goto err_out;
> + }
> + if (attrs->cap.max_inline_data > SIW_MAX_INLINE) {
> + siw_dbg(sdev, "max inline send: %d > %d\n",
> + attrs->cap.max_inline_data, (int)SIW_MAX_INLINE);
> + rv = -EINVAL;
> + goto err_out;
> + }
> + /*
> + * NOTE: we allow for zero element SQ and RQ WQE's SGL's
> + * but not for a QP unable to hold any WQE (SQ + RQ)
> + */
> + if (attrs->cap.max_send_wr + attrs->cap.max_recv_wr == 0) {
> + siw_dbg(sdev, "qp must have send or receive queue\n");
> + rv = -EINVAL;
> + goto err_out;
> + }
> +
> + scq = siw_cq_id2obj(sdev, ((struct siw_cq *)attrs->send_cq)->hdr.id);
> + rcq = siw_cq_id2obj(sdev, ((struct siw_cq *)attrs->recv_cq)->hdr.id);
> +
> + if (!scq || (!rcq && !attrs->srq)) {
> + siw_dbg(sdev, "send cq or receive cq invalid\n");
> + rv = -EINVAL;
> + goto err_out;
> + }
> + qp = kzalloc(sizeof(*qp), GFP_KERNEL);
> + if (!qp) {
> + rv = -ENOMEM;
> + goto err_out;
> + }
> +
> + init_rwsem(&qp->state_lock);
> + spin_lock_init(&qp->sq_lock);
> + spin_lock_init(&qp->rq_lock);
> + spin_lock_init(&qp->orq_lock);
> +
> + if (!base_pd->uobject)
> + qp->kernel_verbs = 1;
New drivers should not have the word 'uobject' in them. This is '!udata'
> + rv = siw_qp_add(sdev, qp);
> + if (rv)
> + goto err_out;
> +
> + num_sqe = roundup_pow_of_two(attrs->cap.max_send_wr);
> + num_rqe = roundup_pow_of_two(attrs->cap.max_recv_wr);
> +
> + if (qp->kernel_verbs)
> + qp->sendq = vzalloc(num_sqe * sizeof(struct siw_sqe));
> + else
> + qp->sendq = vmalloc_user(num_sqe * sizeof(struct siw_sqe));
> +
> + if (qp->sendq == NULL) {
> + siw_dbg_qp(qp, "send queue size %d alloc failed\n", num_sqe);
> + rv = -ENOMEM;
> + goto err_out_idr;
> + }
> + if (attrs->sq_sig_type != IB_SIGNAL_REQ_WR) {
> + if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
> + qp->attrs.flags |= SIW_SIGNAL_ALL_WR;
> + else {
> + rv = -EINVAL;
> + goto err_out_idr;
> + }
> + }
> + qp->pd = pd;
> + qp->scq = scq;
> + qp->rcq = rcq;
> +
> + if (attrs->srq) {
> + /*
> + * SRQ support.
> + * Verbs 6.3.7: ignore RQ size, if SRQ present
> + * Verbs 6.3.5: do not check PD of SRQ against PD of QP
> + */
> + qp->srq = to_siw_srq(attrs->srq);
> + qp->attrs.rq_size = 0;
> + siw_dbg_qp(qp, "[SRQ 0x%p] attached\n", qp->srq);
> + } else if (num_rqe) {
> + if (qp->kernel_verbs)
> + qp->recvq = vzalloc(num_rqe * sizeof(struct siw_rqe));
> + else
> + qp->recvq = vmalloc_user(num_rqe *
> + sizeof(struct siw_rqe));
> +
> + if (qp->recvq == NULL) {
> + siw_dbg_qp(qp, "recv queue size %d alloc failed\n",
> + num_rqe);
> + rv = -ENOMEM;
> + goto err_out_idr;
> + }
> +
> + qp->attrs.rq_size = num_rqe;
> + }
> + qp->attrs.sq_size = num_sqe;
> + qp->attrs.sq_max_sges = attrs->cap.max_send_sge;
> + qp->attrs.rq_max_sges = attrs->cap.max_recv_sge;
> +
> + /* Make those two tunables fixed for now. */
> + qp->tx_ctx.gso_seg_limit = gso_seg_limit;
> + qp->tx_ctx.zcopy_tx = zcopy_tx;
> +
> + qp->attrs.state = SIW_QP_STATE_IDLE;
> +
> + if (udata) {
> + struct siw_uresp_create_qp uresp;
> + struct siw_ucontext *ctx;
> +
> + memset(&uresp, 0, sizeof(uresp));
> + ctx = to_siw_ctx(base_pd->uobject->context);
Here too, this is rdma_udata_to_drv_context()
> + uresp.sq_key = uresp.rq_key = SIW_INVAL_UOBJ_KEY;
> + uresp.num_sqe = num_sqe;
> + uresp.num_rqe = num_rqe;
> + uresp.qp_id = QP_ID(qp);
> +
> + if (qp->sendq) {
> + uresp.sq_key = siw_insert_uobj(ctx, qp->sendq,
> + num_sqe * sizeof(struct siw_sqe));
> + if (uresp.sq_key > SIW_MAX_UOBJ_KEY)
> + siw_dbg_qp(qp, "preparing mmap sq failed\n");
> + }
> + if (qp->recvq) {
> + uresp.rq_key = siw_insert_uobj(ctx, qp->recvq,
> + num_rqe * sizeof(struct siw_rqe));
> + if (uresp.rq_key > SIW_MAX_UOBJ_KEY)
> + siw_dbg_qp(qp, "preparing mmap rq failed\n");
> + }
> + rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
> + if (rv)
> + goto err_out_idr;
> + }
> + qp->tx_cpu = siw_get_tx_cpu(sdev);
> + if (qp->tx_cpu < 0) {
> + rv = -EINVAL;
> + goto err_out_idr;
> + }
> + qp->base_qp.qp_num = QP_ID(qp);
> +
> + siw_pd_get(pd);
> +
> + INIT_LIST_HEAD(&qp->devq);
> + spin_lock_irqsave(&sdev->lock, flags);
> + list_add_tail(&qp->devq, &sdev->qp_list);
> + spin_unlock_irqrestore(&sdev->lock, flags);
> +
> + return &qp->base_qp;
> +
> +err_out_idr:
> + siw_remove_obj(&sdev->lock, &sdev->qp_idr, &qp->hdr);
> +err_out:
> + if (scq)
> + siw_cq_put(scq);
> + if (rcq)
> + siw_cq_put(rcq);
> +
> + if (qp) {
> + if (qp->sendq)
> + vfree(qp->sendq);
> + if (qp->recvq)
> + vfree(qp->recvq);
> + kfree(qp);
> + }
> + atomic_dec(&sdev->num_qp);
> +
> + return ERR_PTR(rv);
> +}
> +
> +/*
> + * Minimum siw_query_qp() verb interface.
> + *
> + * @qp_attr_mask is not used but all available information is provided
> + */
> +int siw_query_qp(struct ib_qp *base_qp, struct ib_qp_attr *qp_attr,
> + int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
> +{
> + struct siw_qp *qp;
> + struct siw_device *sdev;
> +
> + if (base_qp && qp_attr && qp_init_attr) {
> + qp = to_siw_qp(base_qp);
> + sdev = to_siw_dev(base_qp->device);
> + } else
> + return -EINVAL;
> +
> + qp_attr->cap.max_inline_data = SIW_MAX_INLINE;
> + qp_attr->cap.max_send_wr = qp->attrs.sq_size;
> + qp_attr->cap.max_send_sge = qp->attrs.sq_max_sges;
> + qp_attr->cap.max_recv_wr = qp->attrs.rq_size;
> + qp_attr->cap.max_recv_sge = qp->attrs.rq_max_sges;
> + qp_attr->path_mtu = siw_mtu_net2base(sdev->netdev->mtu);
> + qp_attr->max_rd_atomic = qp->attrs.irq_size;
> + qp_attr->max_dest_rd_atomic = qp->attrs.orq_size;
> +
> + qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE |
> + IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ;
> +
> + qp_init_attr->qp_type = base_qp->qp_type;
> + qp_init_attr->send_cq = base_qp->send_cq;
> + qp_init_attr->recv_cq = base_qp->recv_cq;
> + qp_init_attr->srq = base_qp->srq;
> +
> + qp_init_attr->cap = qp_attr->cap;
> +
> + return 0;
> +}
> +
> +int siw_verbs_modify_qp(struct ib_qp *base_qp, struct ib_qp_attr *attr,
> + int attr_mask, struct ib_udata *udata)
> +{
> + struct siw_qp_attrs new_attrs;
> + enum siw_qp_attr_mask siw_attr_mask = 0;
> + struct siw_qp *qp = to_siw_qp(base_qp);
> + int rv = 0;
> +
> + if (!attr_mask)
> + return 0;
> +
> + memset(&new_attrs, 0, sizeof(new_attrs));
> +
> + if (attr_mask & IB_QP_ACCESS_FLAGS) {
> +
> + siw_attr_mask = SIW_QP_ATTR_ACCESS_FLAGS;
> +
> + if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
> + new_attrs.flags |= SIW_RDMA_READ_ENABLED;
> + if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
> + new_attrs.flags |= SIW_RDMA_WRITE_ENABLED;
> + if (attr->qp_access_flags & IB_ACCESS_MW_BIND)
> + new_attrs.flags |= SIW_RDMA_BIND_ENABLED;
> + }
> + if (attr_mask & IB_QP_STATE) {
> + siw_dbg_qp(qp, "desired ib qp state: %s\n",
> + ib_qp_state_to_string[attr->qp_state]);
> +
> + new_attrs.state = ib_qp_state_to_siw_qp_state[attr->qp_state];
> +
> + if (new_attrs.state > SIW_QP_STATE_RTS)
> + qp->tx_ctx.tx_suspend = 1;
> +
> + siw_attr_mask |= SIW_QP_ATTR_STATE;
> + }
> + if (!attr_mask)
> + goto out;
> +
> + down_write(&qp->state_lock);
> +
> + rv = siw_qp_modify(qp, &new_attrs, siw_attr_mask);
> +
> + up_write(&qp->state_lock);
> +out:
> + return rv;
> +}
> +
> +int siw_destroy_qp(struct ib_qp *base_qp)
> +{
> + struct siw_qp *qp = to_siw_qp(base_qp);
> + struct siw_qp_attrs qp_attrs;
> +
> + siw_dbg_qp(qp, "state %d, cep 0x%p\n", qp->attrs.state, qp->cep);
> +
> + /*
> + * Mark QP as in process of destruction to prevent from
> + * any async callbacks to RDMA core
> + */
> + qp->attrs.flags |= SIW_QP_IN_DESTROY;
> + qp->rx_ctx.rx_suspend = 1;
> +
> + down_write(&qp->state_lock);
> +
> + qp_attrs.state = SIW_QP_STATE_ERROR;
> + (void)siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE);
> +
> + if (qp->cep) {
> + siw_cep_put(qp->cep);
> + qp->cep = NULL;
> + }
> +
> + up_write(&qp->state_lock);
> +
> + kfree(qp->rx_ctx.mpa_crc_hd);
> + kfree(qp->tx_ctx.mpa_crc_hd);
> +
> + /* Drop references */
> + siw_cq_put(qp->scq);
> + siw_cq_put(qp->rcq);
> + siw_pd_put(qp->pd);
> + qp->scq = qp->rcq = NULL;
> +
> + siw_qp_put(qp);
> +
> + return 0;
> +}
> +
> +/*
> + * siw_copy_sgl()
> + *
> + * Copy SGL from RDMA core representation to local
> + * representation.
> + */
> +static inline void siw_copy_sgl(struct ib_sge *sge, struct siw_sge *siw_sge,
> + int num_sge)
> +{
> + while (num_sge--) {
> + siw_sge->laddr = sge->addr;
> + siw_sge->length = sge->length;
> + siw_sge->lkey = sge->lkey;
> +
> + siw_sge++; sge++;
> + }
> +}
> +
> +/*
> + * siw_copy_inline_sgl()
> + *
> + * Prepare sgl of inlined data for sending. For userland callers
> + * function checks if given buffer addresses and len's are within
> + * process context bounds.
> + * Data from all provided sge's are copied together into the wqe,
> + * referenced by a single sge.
> + */
> +static int siw_copy_inline_sgl(const struct ib_send_wr *core_wr,
> + struct siw_sqe *sqe)
> +{
> + struct ib_sge *core_sge = core_wr->sg_list;
> + void *kbuf = &sqe->sge[1];
> + int num_sge = core_wr->num_sge,
> + bytes = 0;
> +
> + sqe->sge[0].laddr = (u64)kbuf;
> + sqe->sge[0].lkey = 0;
> +
> + while (num_sge--) {
> + if (!core_sge->length) {
> + core_sge++;
> + continue;
> + }
> + bytes += core_sge->length;
> + if (bytes > SIW_MAX_INLINE) {
> + bytes = -EINVAL;
> + break;
> + }
> + memcpy(kbuf, (void *)(uintptr_t)core_sge->addr,
> + core_sge->length);
> +
> + kbuf += core_sge->length;
> + core_sge++;
> + }
> + sqe->sge[0].length = bytes > 0 ? bytes : 0;
> + sqe->num_sge = bytes > 0 ? 1 : 0;
> +
> + return bytes;
> +}
> +
> +/*
> + * siw_post_send()
> + *
> + * Post a list of S-WR's to a SQ.
> + *
> + * @base_qp: Base QP contained in siw QP
> + * @wr: Null terminated list of user WR's
> + * @bad_wr: Points to failing WR in case of synchronous failure.
> + */
> +int siw_post_send(struct ib_qp *base_qp, const struct ib_send_wr *wr,
> + const struct ib_send_wr **bad_wr)
> +{
> + struct siw_qp *qp = to_siw_qp(base_qp);
> + struct siw_wqe *wqe = tx_wqe(qp);
> +
> + unsigned long flags;
> + int rv = 0;
> +
> + siw_dbg_qp(qp, "state %d\n", qp->attrs.state);
> +
> + /*
> + * Try to acquire QP state lock. Must be non-blocking
> + * to accommodate kernel clients needs.
> + */
> + if (!down_read_trylock(&qp->state_lock)) {
> + *bad_wr = wr;
> + return -ENOTCONN;
> + }
> +
> + if (unlikely(qp->attrs.state != SIW_QP_STATE_RTS)) {
> + up_read(&qp->state_lock);
> + *bad_wr = wr;
> + return -ENOTCONN;
> + }
> + if (wr && !qp->kernel_verbs) {
> + siw_dbg_qp(qp, "wr must be empty for user mapped sq\n");
> + up_read(&qp->state_lock);
> + *bad_wr = wr;
> + return -EINVAL;
> + }
> +
> + spin_lock_irqsave(&qp->sq_lock, flags);
> +
> + while (wr) {
> + u32 idx = qp->sq_put % qp->attrs.sq_size;
> + struct siw_sqe *sqe = &qp->sendq[idx];
> +
> + if (sqe->flags) {
> + siw_dbg_qp(qp, "sq full\n");
> + rv = -ENOMEM;
> + break;
> + }
> + if (wr->num_sge > qp->attrs.sq_max_sges) {
> + siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
> + rv = -EINVAL;
> + break;
> + }
> + sqe->id = wr->wr_id;
> +
> + if ((wr->send_flags & IB_SEND_SIGNALED) ||
> + (qp->attrs.flags & SIW_SIGNAL_ALL_WR))
> + sqe->flags |= SIW_WQE_SIGNALLED;
> +
> + if (wr->send_flags & IB_SEND_FENCE)
> + sqe->flags |= SIW_WQE_READ_FENCE;
> +
> + switch (wr->opcode) {
> +
> + case IB_WR_SEND:
> + case IB_WR_SEND_WITH_INV:
> + if (wr->send_flags & IB_SEND_SOLICITED)
> + sqe->flags |= SIW_WQE_SOLICITED;
> +
> + if (!(wr->send_flags & IB_SEND_INLINE)) {
> + siw_copy_sgl(wr->sg_list, sqe->sge,
> + wr->num_sge);
> + sqe->num_sge = wr->num_sge;
> + } else {
> + rv = siw_copy_inline_sgl(wr, sqe);
> + if (rv <= 0) {
> + rv = -EINVAL;
> + break;
> + }
> + sqe->flags |= SIW_WQE_INLINE;
> + sqe->num_sge = 1;
> + }
> + if (wr->opcode == IB_WR_SEND)
> + sqe->opcode = SIW_OP_SEND;
> + else {
> + sqe->opcode = SIW_OP_SEND_REMOTE_INV;
> + sqe->rkey = wr->ex.invalidate_rkey;
> + }
> + break;
> +
> + case IB_WR_RDMA_READ_WITH_INV:
> + case IB_WR_RDMA_READ:
> + /*
> + * OFED WR restricts RREAD sink to SGL containing
> + * 1 SGE only. we could relax to SGL with multiple
> + * elements referring the SAME ltag or even sending
> + * a private per-rreq tag referring to a checked
> + * local sgl with MULTIPLE ltag's. would be easy
> + * to do...
> + */
> + if (unlikely(wr->num_sge != 1)) {
> + rv = -EINVAL;
> + break;
> + }
> + siw_copy_sgl(wr->sg_list, &sqe->sge[0], 1);
> + /*
> + * NOTE: zero length RREAD is allowed!
> + */
> + sqe->raddr = rdma_wr(wr)->remote_addr;
> + sqe->rkey = rdma_wr(wr)->rkey;
> + sqe->num_sge = 1;
> +
> + if (wr->opcode == IB_WR_RDMA_READ)
> + sqe->opcode = SIW_OP_READ;
> + else
> + sqe->opcode = SIW_OP_READ_LOCAL_INV;
> + break;
> +
> + case IB_WR_RDMA_WRITE:
> + if (!(wr->send_flags & IB_SEND_INLINE)) {
> + siw_copy_sgl(wr->sg_list, &sqe->sge[0],
> + wr->num_sge);
> + sqe->num_sge = wr->num_sge;
> + } else {
> + rv = siw_copy_inline_sgl(wr, sqe);
> + if (unlikely(rv < 0)) {
> + rv = -EINVAL;
> + break;
> + }
> + sqe->flags |= SIW_WQE_INLINE;
> + sqe->num_sge = 1;
> + }
> + sqe->raddr = rdma_wr(wr)->remote_addr;
> + sqe->rkey = rdma_wr(wr)->rkey;
> + sqe->opcode = SIW_OP_WRITE;
> +
> + break;
> +
> + case IB_WR_REG_MR:
> + sqe->base_mr = (uint64_t)reg_wr(wr)->mr;
> + sqe->rkey = reg_wr(wr)->key;
> + sqe->access = SIW_MEM_LREAD;
> + if (reg_wr(wr)->access & IB_ACCESS_LOCAL_WRITE)
> + sqe->access |= SIW_MEM_LWRITE;
> + if (reg_wr(wr)->access & IB_ACCESS_REMOTE_WRITE)
> + sqe->access |= SIW_MEM_RWRITE;
> + if (reg_wr(wr)->access & IB_ACCESS_REMOTE_READ)
> + sqe->access |= SIW_MEM_RREAD;
> + sqe->opcode = SIW_OP_REG_MR;
> +
> + break;
> +
> + case IB_WR_LOCAL_INV:
> + sqe->rkey = wr->ex.invalidate_rkey;
> + sqe->opcode = SIW_OP_INVAL_STAG;
> +
> + break;
> +
> + default:
> + siw_dbg_qp(qp, "ib wr type %d unsupported\n",
> + wr->opcode);
> + rv = -EINVAL;
> + break;
> + }
> + siw_dbg_qp(qp, "opcode %d, flags 0x%x\n",
> + sqe->opcode, sqe->flags);
> +
> + if (unlikely(rv < 0))
> + break;
> +
> + /* make SQE only vaild after completely written */
> + smp_wmb();
> + sqe->flags |= SIW_WQE_VALID;
> +
> + qp->sq_put++;
> + wr = wr->next;
> + }
> +
> + /*
> + * Send directly if SQ processing is not in progress.
> + * Eventual immediate errors (rv < 0) do not affect the involved
> + * RI resources (Verbs, 8.3.1) and thus do not prevent from SQ
> + * processing, if new work is already pending. But rv must be passed
> + * to caller.
> + */
> + if (wqe->wr_status != SIW_WR_IDLE) {
> + spin_unlock_irqrestore(&qp->sq_lock, flags);
> + goto skip_direct_sending;
> + }
> + rv = siw_activate_tx(qp);
> + spin_unlock_irqrestore(&qp->sq_lock, flags);
> +
> + if (rv <= 0)
> + goto skip_direct_sending;
> +
> + if (qp->kernel_verbs) {
> + rv = siw_sq_start(qp);
> + } else {
> + qp->tx_ctx.in_syscall = 1;
> +
> + if (siw_qp_sq_process(qp) != 0 && !(qp->tx_ctx.tx_suspend))
> + siw_qp_cm_drop(qp, 0);
> +
> + qp->tx_ctx.in_syscall = 0;
> + }
> +skip_direct_sending:
> +
> + up_read(&qp->state_lock);
> +
> + if (rv >= 0)
> + return 0;
> + /*
> + * Immediate error
> + */
> + siw_dbg_qp(qp, "error %d\n", rv);
> +
> + *bad_wr = wr;
> + return rv;
> +}
> +
> +/*
> + * siw_post_receive()
> + *
> + * Post a list of R-WR's to a RQ.
> + *
> + * @base_qp: Base QP contained in siw QP
> + * @wr: Null terminated list of user WR's
> + * @bad_wr: Points to failing WR in case of synchronous failure.
> + */
> +int siw_post_receive(struct ib_qp *base_qp, const struct ib_recv_wr *wr,
> + const struct ib_recv_wr **bad_wr)
> +{
> + struct siw_qp *qp = to_siw_qp(base_qp);
> + unsigned long flags;
> + int rv = 0;
> +
> + if (qp->srq) {
> + *bad_wr = wr;
> + return -EOPNOTSUPP; /* what else from errno.h? */
> + }
> + /*
> + * Try to acquire QP state lock. Must be non-blocking
> + * to accommodate kernel clients needs.
> + */
> + if (!down_read_trylock(&qp->state_lock)) {
> + *bad_wr = wr;
> + return -ENOTCONN;
> + }
> + if (!qp->kernel_verbs) {
> + siw_dbg_qp(qp, "no kernel post_recv for user mapped sq\n");
> + up_read(&qp->state_lock);
> + *bad_wr = wr;
> + return -EINVAL;
> + }
> + if (qp->attrs.state > SIW_QP_STATE_RTS) {
> + up_read(&qp->state_lock);
> + *bad_wr = wr;
> + return -EINVAL;
> + }
> + /*
> + * Serialize potentially multiple producers.
> + * Not needed for single threaded consumer side.
> + */
> + spin_lock_irqsave(&qp->rq_lock, flags);
> +
> + while (wr) {
> + u32 idx = qp->rq_put % qp->attrs.rq_size;
> + struct siw_rqe *rqe = &qp->recvq[idx];
> +
> + if (rqe->flags) {
> + siw_dbg_qp(qp, "Receive Queue full\n");
> + rv = -ENOMEM;
> + break;
> + }
> + if (wr->num_sge > qp->attrs.rq_max_sges) {
> + siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
> + rv = -EINVAL;
> + break;
> + }
> + rqe->id = wr->wr_id;
> + rqe->num_sge = wr->num_sge;
> + siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
> +
> + /* make sure RQE is completely written before valid */
> + smp_wmb();
> +
> + rqe->flags = SIW_WQE_VALID;
> +
> + qp->rq_put++;
> + wr = wr->next;
> + }
> + spin_unlock_irqrestore(&qp->rq_lock, flags);
> +
> + up_read(&qp->state_lock);
> +
> + if (rv < 0) {
> + siw_dbg_qp(qp, "error %d\n", rv);
> + *bad_wr = wr;
> + }
> + return rv > 0 ? 0 : rv;
> +}
> +
> +int siw_destroy_cq(struct ib_cq *base_cq)
> +{
> + struct siw_cq *cq = to_siw_cq(base_cq);
> + struct ib_device *base_dev = base_cq->device;
> + struct siw_device *sdev = to_siw_dev(base_dev);
> +
> + siw_cq_flush(cq);
> +
> + siw_remove_obj(&sdev->lock, &sdev->cq_idr, &cq->hdr);
> + siw_cq_put(cq);
> +
> + return 0;
> +}
> +
> +/*
> + * siw_create_cq()
> + *
> + * Create CQ of requested size on given device.
> + *
> + * @base_dev: RDMA device contained in siw device
> + * @size: maximum number of CQE's allowed.
> + * @ib_context: user context.
> + * @udata: used to provide CQ ID back to user.
> + */
> +
> +struct ib_cq *siw_create_cq(struct ib_device *base_dev,
> + const struct ib_cq_init_attr *attr,
> + struct ib_ucontext *ib_context,
> + struct ib_udata *udata)
> +{
> + struct siw_cq *cq = NULL;
> + struct siw_device *sdev = to_siw_dev(base_dev);
> + struct siw_uresp_create_cq uresp;
> + int rv, size = attr->cqe;
> +
> + if (!base_dev) {
> + rv = -ENODEV;
> + goto err_out;
> + }
> + if (atomic_inc_return(&sdev->num_cq) > SIW_MAX_CQ) {
> + siw_dbg(sdev, "too many cq's\n");
> + rv = -ENOMEM;
> + goto err_out;
> + }
> + if (size < 1 || size > sdev->attrs.max_cqe) {
> + siw_dbg(sdev, "cq size error: %d\n", size);
> + rv = -EINVAL;
> + goto err_out;
> + }
> + cq = kzalloc(sizeof(*cq), GFP_KERNEL);
> + if (!cq) {
> + rv = -ENOMEM;
> + goto err_out;
> + }
> + size = roundup_pow_of_two(size);
> + cq->base_cq.cqe = size;
> + cq->num_cqe = size;
> +
> + if (!ib_context) {
> + cq->kernel_verbs = 1;
> + cq->queue = vzalloc(size * sizeof(struct siw_cqe)
> + + sizeof(struct siw_cq_ctrl));
> + } else {
> + cq->queue = vmalloc_user(size * sizeof(struct siw_cqe)
> + + sizeof(struct siw_cq_ctrl));
> + }
> + if (cq->queue == NULL) {
> + rv = -ENOMEM;
> + goto err_out;
> + }
> +
> + rv = siw_cq_add(sdev, cq);
> + if (rv)
> + goto err_out;
> +
> + spin_lock_init(&cq->lock);
> +
> + cq->notify = &((struct siw_cq_ctrl *)&cq->queue[size])->notify;
> +
> + if (!cq->kernel_verbs) {
> + struct siw_ucontext *ctx = to_siw_ctx(ib_context);
> +
> + uresp.cq_key = siw_insert_uobj(ctx, cq->queue,
> + size * sizeof(struct siw_cqe) +
> + sizeof(struct siw_cq_ctrl));
> +
> + if (uresp.cq_key > SIW_MAX_UOBJ_KEY)
> + siw_dbg(sdev, "[CQ %d]: preparing mmap failed\n",
> + OBJ_ID(cq));
> +
> + uresp.cq_id = OBJ_ID(cq);
> + uresp.num_cqe = size;
> +
> + rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
> + if (rv)
> + goto err_out_idr;
> + }
> + return &cq->base_cq;
> +
> +err_out_idr:
> + siw_remove_obj(&sdev->lock, &sdev->cq_idr, &cq->hdr);
> +err_out:
> + siw_dbg(sdev, "cq creation failed: %d", rv);
> +
> + if (cq && cq->queue)
> + vfree(cq->queue);
> +
> + kfree(cq);
> + atomic_dec(&sdev->num_cq);
> +
> + return ERR_PTR(rv);
> +}
> +
> +/*
> + * siw_poll_cq()
> + *
> + * Reap CQ entries if available and copy work completion status into
> + * array of WC's provided by caller. Returns number of reaped CQE's.
> + *
> + * @base_cq: Base CQ contained in siw CQ.
> + * @num_cqe: Maximum number of CQE's to reap.
> + * @wc: Array of work completions to be filled by siw.
> + */
> +int siw_poll_cq(struct ib_cq *base_cq, int num_cqe, struct ib_wc *wc)
> +{
> + struct siw_cq *cq = to_siw_cq(base_cq);
> + int i;
> +
> + for (i = 0; i < num_cqe; i++) {
> + if (!(siw_reap_cqe(cq, wc)))
> + break;
> + wc++;
> + }
> + return i;
> +}
> +
> +/*
> + * siw_req_notify_cq()
> + *
> + * Request notification for new CQE's added to that CQ.
> + * Defined flags:
> + * o SIW_CQ_NOTIFY_SOLICITED lets siw trigger a notification
> + * event if a WQE with notification flag set enters the CQ
> + * o SIW_CQ_NOTIFY_NEXT_COMP lets siw trigger a notification
> + * event if a WQE enters the CQ.
> + * o IB_CQ_REPORT_MISSED_EVENTS: return value will provide the
> + * number of not reaped CQE's regardless of its notification
> + * type and current or new CQ notification settings.
> + *
> + * @base_cq: Base CQ contained in siw CQ.
> + * @flags: Requested notification flags.
> + */
> +int siw_req_notify_cq(struct ib_cq *base_cq, enum ib_cq_notify_flags flags)
> +{
> + struct siw_cq *cq = to_siw_cq(base_cq);
> +
> + siw_dbg(cq->hdr.sdev, "[CQ %d]: flags: 0x%8x\n", OBJ_ID(cq), flags);
> +
> + if ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED)
> + /* CQ event for next solicited completion */
> + smp_store_mb(*cq->notify, SIW_NOTIFY_SOLICITED);
> + else
> + /* CQ event for any signalled completion */
> + smp_store_mb(*cq->notify, SIW_NOTIFY_ALL);
> +
> + if (flags & IB_CQ_REPORT_MISSED_EVENTS)
> + return cq->cq_put - cq->cq_get;
> +
> + return 0;
> +}
> +
> +/*
> + * siw_dereg_mr()
> + *
> + * Release Memory Region.
> + *
> + * TODO: Update function if Memory Windows are supported by siw:
> + * Is OFED core checking for MW dependencies for current
> + * MR before calling MR deregistration?.
> + *
> + * @base_mr: Base MR contained in siw MR.
> + */
> +int siw_dereg_mr(struct ib_mr *base_mr)
> +{
> + struct siw_mr *mr;
> + struct siw_device *sdev = to_siw_dev(base_mr->device);
> +
> + mr = to_siw_mr(base_mr);
> +
> + siw_dbg(sdev, "[MEM %d]: deregister mr, #ref's %d\n",
> + mr->mem.hdr.id, kref_read(&mr->mem.hdr.ref));
> +
> + mr->mem.stag_valid = 0;
> +
> + siw_remove_obj(&sdev->lock, &sdev->mem_idr, &mr->mem.hdr);
> + siw_mem_put(&mr->mem);
> +
> + return 0;
> +}
> +
> +static struct siw_mr *siw_create_mr(struct siw_device *sdev, void *mem_obj,
> + u64 start, u64 len, int rights)
> +{
> + struct siw_mr *mr = kzalloc(sizeof(*mr), GFP_KERNEL);
> + unsigned long flags;
> +
> + if (!mr)
> + return NULL;
> +
> + mr->mem.stag_valid = 0;
> +
> + if (siw_mem_add(sdev, &mr->mem) < 0) {
> + kfree(mr);
> + return NULL;
> + }
> + siw_dbg(sdev, "[MEM %d]: new mr, object 0x%p\n",
> + mr->mem.hdr.id, mem_obj);
> +
> + mr->base_mr.lkey = mr->base_mr.rkey = mr->mem.hdr.id << 8;
> +
> + mr->mem.va = start;
> + mr->mem.len = len;
> + mr->mem.mr = NULL;
> + mr->mem.perms = SIW_MEM_LREAD | /* not selectable in RDMA core */
> + (rights & IB_ACCESS_REMOTE_READ ? SIW_MEM_RREAD : 0) |
> + (rights & IB_ACCESS_LOCAL_WRITE ? SIW_MEM_LWRITE : 0) |
> + (rights & IB_ACCESS_REMOTE_WRITE ? SIW_MEM_RWRITE : 0);
> +
> + mr->mem_obj = mem_obj;
> +
> + INIT_LIST_HEAD(&mr->devq);
> + spin_lock_irqsave(&sdev->lock, flags);
> + list_add_tail(&mr->devq, &sdev->mr_list);
> + spin_unlock_irqrestore(&sdev->lock, flags);
> +
> + return mr;
> +}
> +
> +/*
> + * siw_reg_user_mr()
> + *
> + * Register Memory Region.
> + *
> + * @base_pd: Base PD contained in siw PD.
> + * @start: starting address of MR (virtual address)
> + * @len: len of MR
> + * @rnic_va: not used by siw
> + * @rights: MR access rights
> + * @udata: user buffer to communicate STag and Key.
> + */
> +struct ib_mr *siw_reg_user_mr(struct ib_pd *base_pd, u64 start, u64 len,
> + u64 rnic_va, int rights, struct ib_udata *udata)
> +{
> + struct siw_mr *mr = NULL;
> + struct siw_pd *pd = to_siw_pd(base_pd);
> + struct siw_umem *umem = NULL;
> + struct siw_ureq_reg_mr ureq;
> + struct siw_uresp_reg_mr uresp;
> + struct siw_device *sdev = pd->hdr.sdev;
> +
> + unsigned long mem_limit = rlimit(RLIMIT_MEMLOCK);
> + int rv;
> +
> + siw_dbg(sdev, "[PD %d]: start: 0x%016llx, va: 0x%016llx, len: %llu\n",
> + OBJ_ID(pd), (unsigned long long)start,
> + (unsigned long long)rnic_va, (unsigned long long)len);
> +
> + if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
> + siw_dbg(sdev, "[PD %d]: too many mr's\n", OBJ_ID(pd));
> + rv = -ENOMEM;
> + goto err_out;
> + }
> + if (!len) {
> + rv = -EINVAL;
> + goto err_out;
> + }
> + if (mem_limit != RLIM_INFINITY) {
> + unsigned long num_pages =
> + (PAGE_ALIGN(len + (start & ~PAGE_MASK))) >> PAGE_SHIFT;
> + mem_limit >>= PAGE_SHIFT;
> +
> + if (num_pages > mem_limit - current->mm->locked_vm) {
> + siw_dbg(sdev,
> + "[PD %d]: pages req %lu, max %lu, lock %lu\n",
> + OBJ_ID(pd), num_pages, mem_limit,
> + current->mm->locked_vm);
> + rv = -ENOMEM;
> + goto err_out;
> + }
> + }
> + umem = siw_umem_get(start, len);
> + if (IS_ERR(umem)) {
> + rv = PTR_ERR(umem);
> + siw_dbg(sdev, "[PD %d]: getting user memory failed: %d\n",
> + OBJ_ID(pd), rv);
> + umem = NULL;
> + goto err_out;
> + }
> + mr = siw_create_mr(sdev, umem, start, len, rights);
> + if (!mr) {
> + rv = -ENOMEM;
> + goto err_out;
> + }
> + if (udata) {
> + rv = ib_copy_from_udata(&ureq, udata, sizeof(ureq));
> + if (rv)
> + goto err_out;
> +
> + mr->base_mr.lkey |= ureq.stag_key;
> + mr->base_mr.rkey |= ureq.stag_key;
> + uresp.stag = mr->base_mr.lkey;
> +
> + rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
> + if (rv)
> + goto err_out;
> + }
> + mr->pd = pd;
> + siw_pd_get(pd);
> +
> + mr->mem.stag_valid = 1;
> +
> + return &mr->base_mr;
> +
> +err_out:
> + if (mr) {
> + siw_remove_obj(&sdev->lock, &sdev->mem_idr, &mr->mem.hdr);
> + siw_mem_put(&mr->mem);
> + umem = NULL;
> + } else
> + atomic_dec(&sdev->num_mr);
> +
> + if (umem)
> + siw_umem_release(umem);
> +
> + return ERR_PTR(rv);
> +}
> +
> +struct ib_mr *siw_alloc_mr(struct ib_pd *base_pd, enum ib_mr_type mr_type,
> + u32 max_sge)
> +{
> + struct siw_mr *mr;
> + struct siw_pd *pd = to_siw_pd(base_pd);
> + struct siw_device *sdev = pd->hdr.sdev;
> + struct siw_pbl *pbl = NULL;
> + int rv;
> +
> + if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
> + siw_dbg(sdev, "[PD %d]: too many mr's\n", OBJ_ID(pd));
> + rv = -ENOMEM;
> + goto err_out;
> + }
> + if (mr_type != IB_MR_TYPE_MEM_REG) {
> + siw_dbg(sdev, "[PD %d]: mr type %d unsupported\n",
> + OBJ_ID(pd), mr_type);
> + rv = -EOPNOTSUPP;
> + goto err_out;
> + }
> + if (max_sge > SIW_MAX_SGE_PBL) {
> + siw_dbg(sdev, "[PD %d]: too many sge's: %d\n",
> + OBJ_ID(pd), max_sge);
> + rv = -ENOMEM;
> + goto err_out;
> + }
> + pbl = siw_pbl_alloc(max_sge);
> + if (IS_ERR(pbl)) {
> + rv = PTR_ERR(pbl);
> + siw_dbg(sdev, "[PD %d]: pbl allocation failed: %d\n",
> + OBJ_ID(pd), rv);
> + pbl = NULL;
> + goto err_out;
> + }
> + mr = siw_create_mr(sdev, pbl, 0, max_sge * PAGE_SIZE, 0);
> + if (!mr) {
> + rv = -ENOMEM;
> + goto err_out;
> + }
> + mr->mem.is_pbl = 1;
> + mr->pd = pd;
> + siw_pd_get(pd);
> +
> + siw_dbg(sdev, "[PD %d], [MEM %d]: success\n",
> + OBJ_ID(pd), OBJ_ID(&mr->mem));
> +
> + return &mr->base_mr;
> +
> +err_out:
> + if (pbl)
> + siw_pbl_free(pbl);
> +
> + siw_dbg(sdev, "[PD %d]: failed: %d\n", OBJ_ID(pd), rv);
> +
> + atomic_dec(&sdev->num_mr);
> +
> + return ERR_PTR(rv);
> +}
> +
> +/* Just used to count number of pages being mapped */
> +static int siw_set_pbl_page(struct ib_mr *base_mr, u64 buf_addr)
> +{
> + return 0;
> +}
> +
> +int siw_map_mr_sg(struct ib_mr *base_mr, struct scatterlist *sl, int num_sle,
> + unsigned int *sg_off)
> +{
> + struct scatterlist *slp;
> + struct siw_mr *mr = to_siw_mr(base_mr);
> + struct siw_pbl *pbl = mr->pbl;
> + struct siw_pble *pble = pbl->pbe;
> + u64 pbl_size;
> + int i, rv;
> +
> + if (!pbl) {
> + siw_dbg(mr->mem.hdr.sdev, "[MEM %d]: no pbl allocated\n",
> + OBJ_ID(&mr->mem));
> + return -EINVAL;
> + }
> + if (pbl->max_buf < num_sle) {
> + siw_dbg(mr->mem.hdr.sdev, "[MEM %d]: too many sge's: %d>%d\n",
> + OBJ_ID(&mr->mem), mr->pbl->max_buf, num_sle);
> + return -ENOMEM;
> + }
> +
> + for_each_sg(sl, slp, num_sle, i) {
> + if (sg_dma_len(slp) == 0) {
> + siw_dbg(mr->mem.hdr.sdev, "[MEM %d]: empty sge\n",
> + OBJ_ID(&mr->mem));
> + return -EINVAL;
> + }
> + if (i == 0) {
> + pble->addr = sg_dma_address(slp);
> + pble->size = sg_dma_len(slp);
> + pble->pbl_off = 0;
> + pbl_size = pble->size;
> + pbl->num_buf = 1;
> +
> + continue;
> + }
> + /* Merge PBL entries if adjacent */
> + if (pble->addr + pble->size == sg_dma_address(slp))
> + pble->size += sg_dma_len(slp);
> + else {
> + pble++;
> + pbl->num_buf++;
> + pble->addr = sg_dma_address(slp);
> + pble->size = sg_dma_len(slp);
> + pble->pbl_off = pbl_size;
> + }
> + pbl_size += sg_dma_len(slp);
> +
> + siw_dbg(mr->mem.hdr.sdev,
> + "[MEM %d]: sge[%d], size %llu, addr %p, total %llu\n",
> + OBJ_ID(&mr->mem), i, pble->size, (void *)pble->addr,
> + pbl_size);
> + }
> + rv = ib_sg_to_pages(base_mr, sl, num_sle, sg_off, siw_set_pbl_page);
> + if (rv > 0) {
> + mr->mem.len = base_mr->length;
> + mr->mem.va = base_mr->iova;
> + siw_dbg(mr->mem.hdr.sdev,
> + "[MEM %d]: %llu byte, %u SLE into %u entries\n",
> + OBJ_ID(&mr->mem), mr->mem.len, num_sle, pbl->num_buf);
> + }
> + return rv;
> +}
> +
> +/*
> + * siw_get_dma_mr()
> + *
> + * Create a (empty) DMA memory region, where no umem is attached.
> + */
> +struct ib_mr *siw_get_dma_mr(struct ib_pd *base_pd, int rights)
> +{
> + struct siw_mr *mr;
> + struct siw_pd *pd = to_siw_pd(base_pd);
> + struct siw_device *sdev = pd->hdr.sdev;
> + int rv;
> +
> + if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
> + siw_dbg(sdev, "[PD %d]: too many mr's\n", OBJ_ID(pd));
> + rv = -ENOMEM;
> + goto err_out;
> + }
> + mr = siw_create_mr(sdev, NULL, 0, ULONG_MAX, rights);
> + if (!mr) {
> + rv = -ENOMEM;
> + goto err_out;
> + }
> + mr->mem.stag_valid = 1;
> +
> + mr->pd = pd;
> + siw_pd_get(pd);
> +
> + siw_dbg(sdev, "[PD %d], [MEM %d]: success\n",
> + OBJ_ID(pd), OBJ_ID(&mr->mem));
> +
> + return &mr->base_mr;
> +
> +err_out:
> + atomic_dec(&sdev->num_mr);
> +
> + return ERR_PTR(rv);
> +}
> +
> +/*
> + * siw_create_srq()
> + *
> + * Create Shared Receive Queue of attributes @init_attrs
> + * within protection domain given by @base_pd.
> + *
> + * @base_pd: Base PD contained in siw PD.
> + * @init_attrs: SRQ init attributes.
> + * @udata: not used by siw.
> + */
> +struct ib_srq *siw_create_srq(struct ib_pd *base_pd,
> + struct ib_srq_init_attr *init_attrs,
> + struct ib_udata *udata)
> +{
> + struct siw_srq *srq = NULL;
> + struct ib_srq_attr *attrs = &init_attrs->attr;
> + struct siw_pd *pd = to_siw_pd(base_pd);
> + struct siw_device *sdev = pd->hdr.sdev;
> +
> + int kernel_verbs = base_pd->uobject ? 0 : 1;
bool kernel_verbs = !udata;
> + memset(&uresp, 0, sizeof(uresp));
> + ctx = to_siw_ctx(base_pd->uobject->context);
.. and so on throughout.
Jason
