iwch_qp.c

linux内核源码

		else
			*ecode = RDMAP_STAG_NOT_ASSOC;
		break;
	case TPT_ERR_QPID:
		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
		*ecode = RDMAP_STAG_NOT_ASSOC;
		break;
	case TPT_ERR_ACCESS:
		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
		*ecode = RDMAP_ACC_VIOL;
		break;
	case TPT_ERR_WRAP:
		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
		*ecode = RDMAP_TO_WRAP;
		break;
	case TPT_ERR_BOUND:
		if (tagged) {
			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
			*ecode = DDPT_BASE_BOUNDS;
		} else {
			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
			*ecode = RDMAP_BASE_BOUNDS;
		}
		break;
	case TPT_ERR_INVALIDATE_SHARED_MR:
	case TPT_ERR_INVALIDATE_MR_WITH_MW_BOUND:
		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
		*ecode = RDMAP_CANT_INV_STAG;
		break;
	case TPT_ERR_ECC:
	case TPT_ERR_ECC_PSTAG:
	case TPT_ERR_INTERNAL_ERR:
		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
		*ecode = 0;
		break;
	case TPT_ERR_OUT_OF_RQE:
		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
		*ecode = DDPU_INV_MSN_NOBUF;
		break;
	case TPT_ERR_PBL_ADDR_BOUND:
		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
		*ecode = DDPT_BASE_BOUNDS;
		break;
	case TPT_ERR_CRC:
		*layer_type = LAYER_MPA|DDP_LLP;
		*ecode = MPA_CRC_ERR;
		break;
	case TPT_ERR_MARKER:
		*layer_type = LAYER_MPA|DDP_LLP;
		*ecode = MPA_MARKER_ERR;
		break;
	case TPT_ERR_PDU_LEN_ERR:
		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
		*ecode = DDPU_MSG_TOOBIG;
		break;
	case TPT_ERR_DDP_VERSION:
		if (tagged) {
			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
			*ecode = DDPT_INV_VERS;
		} else {
			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
			*ecode = DDPU_INV_VERS;
		}
		break;
	case TPT_ERR_RDMA_VERSION:
		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
		*ecode = RDMAP_INV_VERS;
		break;
	case TPT_ERR_OPCODE:
		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
		*ecode = RDMAP_INV_OPCODE;
		break;
	case TPT_ERR_DDP_QUEUE_NUM:
		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
		*ecode = DDPU_INV_QN;
		break;
	case TPT_ERR_MSN:
	case TPT_ERR_MSN_GAP:
	case TPT_ERR_MSN_RANGE:
	case TPT_ERR_IRD_OVERFLOW:
		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
		*ecode = DDPU_INV_MSN_RANGE;
		break;
	case TPT_ERR_TBIT:
		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
		*ecode = 0;
		break;
	case TPT_ERR_MO:
		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
		*ecode = DDPU_INV_MO;
		break;
	default:
		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
		*ecode = 0;
		break;
	}
}

/*
 * This posts a TERMINATE with layer=RDMA, type=catastrophic.
 */
int iwch_post_terminate(struct iwch_qp *qhp, struct respQ_msg_t *rsp_msg)
{
	union t3_wr *wqe;
	struct terminate_message *term;
	struct sk_buff *skb;

	PDBG("%s %d\n", __FUNCTION__, __LINE__);
	skb = alloc_skb(40, GFP_ATOMIC);
	if (!skb) {
		printk(KERN_ERR "%s cannot send TERMINATE!\n", __FUNCTION__);
		return -ENOMEM;
	}
	wqe = (union t3_wr *)skb_put(skb, 40);
	memset(wqe, 0, 40);
	wqe->send.rdmaop = T3_TERMINATE;

	/* immediate data length */
	wqe->send.plen = htonl(4);

	/* immediate data starts here. */
	term = (struct terminate_message *)wqe->send.sgl;
	build_term_codes(rsp_msg, &term->layer_etype, &term->ecode);
	wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_SEND) |
			 V_FW_RIWR_FLAGS(T3_COMPLETION_FLAG | T3_NOTIFY_FLAG));
	wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(qhp->ep->hwtid));
	skb->priority = CPL_PRIORITY_DATA;
	return cxgb3_ofld_send(qhp->rhp->rdev.t3cdev_p, skb);
}

/*
 * Assumes qhp lock is held.
 */
static void __flush_qp(struct iwch_qp *qhp, unsigned long *flag)
{
	struct iwch_cq *rchp, *schp;
	int count;

	rchp = get_chp(qhp->rhp, qhp->attr.rcq);
	schp = get_chp(qhp->rhp, qhp->attr.scq);

	PDBG("%s qhp %p rchp %p schp %p\n", __FUNCTION__, qhp, rchp, schp);
	/* take a ref on the qhp since we must release the lock */
	atomic_inc(&qhp->refcnt);
	spin_unlock_irqrestore(&qhp->lock, *flag);

	/* locking heirarchy: cq lock first, then qp lock. */
	spin_lock_irqsave(&rchp->lock, *flag);
	spin_lock(&qhp->lock);
	cxio_flush_hw_cq(&rchp->cq);
	cxio_count_rcqes(&rchp->cq, &qhp->wq, &count);
	cxio_flush_rq(&qhp->wq, &rchp->cq, count);
	spin_unlock(&qhp->lock);
	spin_unlock_irqrestore(&rchp->lock, *flag);

	/* locking heirarchy: cq lock first, then qp lock. */
	spin_lock_irqsave(&schp->lock, *flag);
	spin_lock(&qhp->lock);
	cxio_flush_hw_cq(&schp->cq);
	cxio_count_scqes(&schp->cq, &qhp->wq, &count);
	cxio_flush_sq(&qhp->wq, &schp->cq, count);
	spin_unlock(&qhp->lock);
	spin_unlock_irqrestore(&schp->lock, *flag);

	/* deref */
	if (atomic_dec_and_test(&qhp->refcnt))
	        wake_up(&qhp->wait);

	spin_lock_irqsave(&qhp->lock, *flag);
}

static void flush_qp(struct iwch_qp *qhp, unsigned long *flag)
{
	if (t3b_device(qhp->rhp))
		cxio_set_wq_in_error(&qhp->wq);
	else
		__flush_qp(qhp, flag);
}


/*
 * Return non zero if at least one RECV was pre-posted.
 */
static int rqes_posted(struct iwch_qp *qhp)
{
	return fw_riwrh_opcode((struct fw_riwrh *)qhp->wq.queue) == T3_WR_RCV;
}

static int rdma_init(struct iwch_dev *rhp, struct iwch_qp *qhp,
				enum iwch_qp_attr_mask mask,
				struct iwch_qp_attributes *attrs)
{
	struct t3_rdma_init_attr init_attr;
	int ret;

	init_attr.tid = qhp->ep->hwtid;
	init_attr.qpid = qhp->wq.qpid;
	init_attr.pdid = qhp->attr.pd;
	init_attr.scqid = qhp->attr.scq;
	init_attr.rcqid = qhp->attr.rcq;
	init_attr.rq_addr = qhp->wq.rq_addr;
	init_attr.rq_size = 1 << qhp->wq.rq_size_log2;
	init_attr.mpaattrs = uP_RI_MPA_IETF_ENABLE |
		qhp->attr.mpa_attr.recv_marker_enabled |
		(qhp->attr.mpa_attr.xmit_marker_enabled << 1) |
		(qhp->attr.mpa_attr.crc_enabled << 2);

	/*
	 * XXX - The IWCM doesn't quite handle getting these
	 * attrs set before going into RTS.  For now, just turn
	 * them on always...
	 */
#if 0
	init_attr.qpcaps = qhp->attr.enableRdmaRead |
		(qhp->attr.enableRdmaWrite << 1) |
		(qhp->attr.enableBind << 2) |
		(qhp->attr.enable_stag0_fastreg << 3) |
		(qhp->attr.enable_stag0_fastreg << 4);
#else
	init_attr.qpcaps = 0x1f;
#endif
	init_attr.tcp_emss = qhp->ep->emss;
	init_attr.ord = qhp->attr.max_ord;
	init_attr.ird = qhp->attr.max_ird;
	init_attr.qp_dma_addr = qhp->wq.dma_addr;
	init_attr.qp_dma_size = (1UL << qhp->wq.size_log2);
	init_attr.flags = rqes_posted(qhp) ? RECVS_POSTED : 0;
	init_attr.irs = qhp->ep->rcv_seq;
	PDBG("%s init_attr.rq_addr 0x%x init_attr.rq_size = %d "
	     "flags 0x%x qpcaps 0x%x\n", __FUNCTION__,
	     init_attr.rq_addr, init_attr.rq_size,
	     init_attr.flags, init_attr.qpcaps);
	ret = cxio_rdma_init(&rhp->rdev, &init_attr);
	PDBG("%s ret %d\n", __FUNCTION__, ret);
	return ret;
}

int iwch_modify_qp(struct iwch_dev *rhp, struct iwch_qp *qhp,
				enum iwch_qp_attr_mask mask,
				struct iwch_qp_attributes *attrs,
				int internal)
{
	int ret = 0;
	struct iwch_qp_attributes newattr = qhp->attr;
	unsigned long flag;
	int disconnect = 0;
	int terminate = 0;
	int abort = 0;
	int free = 0;
	struct iwch_ep *ep = NULL;

	PDBG("%s qhp %p qpid 0x%x ep %p state %d -> %d\n", __FUNCTION__,
	     qhp, qhp->wq.qpid, qhp->ep, qhp->attr.state,
	     (mask & IWCH_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);

	spin_lock_irqsave(&qhp->lock, flag);

	/* Process attr changes if in IDLE */
	if (mask & IWCH_QP_ATTR_VALID_MODIFY) {
		if (qhp->attr.state != IWCH_QP_STATE_IDLE) {
			ret = -EIO;
			goto out;
		}
		if (mask & IWCH_QP_ATTR_ENABLE_RDMA_READ)
			newattr.enable_rdma_read = attrs->enable_rdma_read;
		if (mask & IWCH_QP_ATTR_ENABLE_RDMA_WRITE)
			newattr.enable_rdma_write = attrs->enable_rdma_write;
		if (mask & IWCH_QP_ATTR_ENABLE_RDMA_BIND)
			newattr.enable_bind = attrs->enable_bind;
		if (mask & IWCH_QP_ATTR_MAX_ORD) {
			if (attrs->max_ord >
			    rhp->attr.max_rdma_read_qp_depth) {
				ret = -EINVAL;
				goto out;
			}
			newattr.max_ord = attrs->max_ord;
		}
		if (mask & IWCH_QP_ATTR_MAX_IRD) {
			if (attrs->max_ird >
			    rhp->attr.max_rdma_reads_per_qp) {
				ret = -EINVAL;
				goto out;
			}
			newattr.max_ird = attrs->max_ird;
		}
		qhp->attr = newattr;
	}

	if (!(mask & IWCH_QP_ATTR_NEXT_STATE))
		goto out;
	if (qhp->attr.state == attrs->next_state)
		goto out;

	switch (qhp->attr.state) {
	case IWCH_QP_STATE_IDLE:
		switch (attrs->next_state) {
		case IWCH_QP_STATE_RTS:
			if (!(mask & IWCH_QP_ATTR_LLP_STREAM_HANDLE)) {
				ret = -EINVAL;
				goto out;
			}
			if (!(mask & IWCH_QP_ATTR_MPA_ATTR)) {
				ret = -EINVAL;
				goto out;
			}
			qhp->attr.mpa_attr = attrs->mpa_attr;
			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
			qhp->ep = qhp->attr.llp_stream_handle;
			qhp->attr.state = IWCH_QP_STATE_RTS;

			/*
			 * Ref the endpoint here and deref when we
			 * disassociate the endpoint from the QP.  This
			 * happens in CLOSING->IDLE transition or *->ERROR
			 * transition.
			 */
			get_ep(&qhp->ep->com);
			spin_unlock_irqrestore(&qhp->lock, flag);
			ret = rdma_init(rhp, qhp, mask, attrs);
			spin_lock_irqsave(&qhp->lock, flag);
			if (ret)
				goto err;
			break;
		case IWCH_QP_STATE_ERROR:
			qhp->attr.state = IWCH_QP_STATE_ERROR;
			flush_qp(qhp, &flag);
			break;
		default:
			ret = -EINVAL;
			goto out;
		}
		break;
	case IWCH_QP_STATE_RTS:
		switch (attrs->next_state) {
		case IWCH_QP_STATE_CLOSING:
			BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
			qhp->attr.state = IWCH_QP_STATE_CLOSING;
			if (!internal) {
				abort=0;
				disconnect = 1;
				ep = qhp->ep;
			}
			break;
		case IWCH_QP_STATE_TERMINATE:
			qhp->attr.state = IWCH_QP_STATE_TERMINATE;
			if (t3b_device(qhp->rhp))
				cxio_set_wq_in_error(&qhp->wq);
			if (!internal)
				terminate = 1;
			break;
		case IWCH_QP_STATE_ERROR:
			qhp->attr.state = IWCH_QP_STATE_ERROR;
			if (!internal) {
				abort=1;
				disconnect = 1;
				ep = qhp->ep;
			}
			goto err;
			break;
		default:
			ret = -EINVAL;
			goto out;
		}
		break;
	case IWCH_QP_STATE_CLOSING:
		if (!internal) {
			ret = -EINVAL;
			goto out;
		}
		switch (attrs->next_state) {
			case IWCH_QP_STATE_IDLE:
				qhp->attr.state = IWCH_QP_STATE_IDLE;
				qhp->attr.llp_stream_handle = NULL;
				put_ep(&qhp->ep->com);
				qhp->ep = NULL;
				wake_up(&qhp->wait);
				break;
			case IWCH_QP_STATE_ERROR:
				goto err;
			default:
				ret = -EINVAL;
				goto err;
		}
		break;
	case IWCH_QP_STATE_ERROR:
		if (attrs->next_state != IWCH_QP_STATE_IDLE) {
			ret = -EINVAL;
			goto out;
		}

		if (!Q_EMPTY(qhp->wq.sq_rptr, qhp->wq.sq_wptr) ||
		    !Q_EMPTY(qhp->wq.rq_rptr, qhp->wq.rq_wptr)) {
			ret = -EINVAL;
			goto out;
		}
		qhp->attr.state = IWCH_QP_STATE_IDLE;
		memset(&qhp->attr, 0, sizeof(qhp->attr));
		break;
	case IWCH_QP_STATE_TERMINATE:
		if (!internal) {
			ret = -EINVAL;
			goto out;
		}
		goto err;
		break;
	default:
		printk(KERN_ERR "%s in a bad state %d\n",
		       __FUNCTION__, qhp->attr.state);
		ret = -EINVAL;
		goto err;
		break;
	}
	goto out;
err:
	PDBG("%s disassociating ep %p qpid 0x%x\n", __FUNCTION__, qhp->ep,
	     qhp->wq.qpid);

	/* disassociate the LLP connection */
	qhp->attr.llp_stream_handle = NULL;
	ep = qhp->ep;
	qhp->ep = NULL;
	qhp->attr.state = IWCH_QP_STATE_ERROR;
	free=1;
	wake_up(&qhp->wait);
	BUG_ON(!ep);
	flush_qp(qhp, &flag);
out:
	spin_unlock_irqrestore(&qhp->lock, flag);

	if (terminate)
		iwch_post_terminate(qhp, NULL);

	/*
	 * If disconnect is 1, then we need to initiate a disconnect
	 * on the EP.  This can be a normal close (RTS->CLOSING) or
	 * an abnormal close (RTS/CLOSING->ERROR).
	 */
	if (disconnect)
		iwch_ep_disconnect(ep, abort, GFP_KERNEL);

	/*
	 * If free is 1, then we've disassociated the EP from the QP
	 * and we need to dereference the EP.
	 */
	if (free)
		put_ep(&ep->com);

	PDBG("%s exit state %d\n", __FUNCTION__, qhp->attr.state);
	return ret;
}

static int quiesce_qp(struct iwch_qp *qhp)
{
	spin_lock_irq(&qhp->lock);
	iwch_quiesce_tid(qhp->ep);
	qhp->flags |= QP_QUIESCED;
	spin_unlock_irq(&qhp->lock);
	return 0;
}

static int resume_qp(struct iwch_qp *qhp)
{
	spin_lock_irq(&qhp->lock);
	iwch_resume_tid(qhp->ep);
	qhp->flags &= ~QP_QUIESCED;
	spin_unlock_irq(&qhp->lock);
	return 0;
}

int iwch_quiesce_qps(struct iwch_cq *chp)
{
	int i;
	struct iwch_qp *qhp;

	for (i=0; i < T3_MAX_NUM_QP; i++) {
		qhp = get_qhp(chp->rhp, i);
		if (!qhp)
			continue;
		if ((qhp->attr.rcq == chp->cq.cqid) && !qp_quiesced(qhp)) {
			quiesce_qp(qhp);
			continue;
		}
		if ((qhp->attr.scq == chp->cq.cqid) && !qp_quiesced(qhp))
			quiesce_qp(qhp);
	}
	return 0;
}

int iwch_resume_qps(struct iwch_cq *chp)
{
	int i;
	struct iwch_qp *qhp;

	for (i=0; i < T3_MAX_NUM_QP; i++) {
		qhp = get_qhp(chp->rhp, i);
		if (!qhp)
			continue;
		if ((qhp->attr.rcq == chp->cq.cqid) && qp_quiesced(qhp)) {
			resume_qp(qhp);
			continue;
		}
		if ((qhp->attr.scq == chp->cq.cqid) && qp_quiesced(qhp))
			resume_qp(qhp);
	}
	return 0;
}