ipath_rc.c

来自「LINUX 2.6.17.4的源码」· C语言 代码 · 共 1,857 行 · 第 1/4 页

	qp->s_ack_state = opcode;
	qp->s_nak_state = 0;
	qp->s_ack_psn = psn;
resched:
	return 0;

done:
	return 1;
}

/**
 * ipath_rc_rcv - process an incoming RC packet
 * @dev: the device this packet came in on
 * @hdr: the header of this packet
 * @has_grh: true if the header has a GRH
 * @data: the packet data
 * @tlen: the packet length
 * @qp: the QP for this packet
 *
 * This is called from ipath_qp_rcv() to process an incoming RC packet
 * for the given QP.
 * Called at interrupt level.
 */
void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
		  int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
{
	struct ipath_other_headers *ohdr;
	u32 opcode;
	u32 hdrsize;
	u32 psn;
	u32 pad;
	unsigned long flags;
	struct ib_wc wc;
	u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
	int diff;
	struct ib_reth *reth;
	int header_in_data;

	/* Check for GRH */
	if (!has_grh) {
		ohdr = &hdr->u.oth;
		hdrsize = 8 + 12;	/* LRH + BTH */
		psn = be32_to_cpu(ohdr->bth[2]);
		header_in_data = 0;
	} else {
		ohdr = &hdr->u.l.oth;
		hdrsize = 8 + 40 + 12;	/* LRH + GRH + BTH */
		/*
		 * The header with GRH is 60 bytes and the core driver sets
		 * the eager header buffer size to 56 bytes so the last 4
		 * bytes of the BTH header (PSN) is in the data buffer.
		 */
		header_in_data =
			ipath_layer_get_rcvhdrentsize(dev->dd) == 16;
		if (header_in_data) {
			psn = be32_to_cpu(((__be32 *) data)[0]);
			data += sizeof(__be32);
		} else
			psn = be32_to_cpu(ohdr->bth[2]);
	}
	/*
	 * The opcode is in the low byte when its in network order
	 * (top byte when in host order).
	 */
	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;

	/*
	 * Process responses (ACKs) before anything else.  Note that the
	 * packet sequence number will be for something in the send work
	 * queue rather than the expected receive packet sequence number.
	 * In other words, this QP is the requester.
	 */
	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
		ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
				  hdrsize, pmtu, header_in_data);
		goto bail;
	}

	spin_lock_irqsave(&qp->r_rq.lock, flags);

	/* Compute 24 bits worth of difference. */
	diff = ipath_cmp24(psn, qp->r_psn);
	if (unlikely(diff)) {
		if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
				       psn, diff, header_in_data))
			goto done;
		goto resched;
	}

	/* Check for opcode sequence errors. */
	switch (qp->r_state) {
	case OP(SEND_FIRST):
	case OP(SEND_MIDDLE):
		if (opcode == OP(SEND_MIDDLE) ||
		    opcode == OP(SEND_LAST) ||
		    opcode == OP(SEND_LAST_WITH_IMMEDIATE))
			break;
	nack_inv:
	/*
	 * A NAK will ACK earlier sends and RDMA writes.  Don't queue the
	 * NAK if a RDMA read, atomic, or NAK is pending though.
	 */
	spin_lock(&qp->s_lock);
	if (qp->s_ack_state >= OP(RDMA_READ_REQUEST) &&
	    qp->s_ack_state != IB_OPCODE_ACKNOWLEDGE) {
		spin_unlock(&qp->s_lock);
		goto done;
	}
	/* XXX Flush WQEs */
	qp->state = IB_QPS_ERR;
	qp->s_ack_state = OP(SEND_ONLY);
	qp->s_nak_state = IB_NAK_INVALID_REQUEST;
	qp->s_ack_psn = qp->r_psn;
	goto resched;

	case OP(RDMA_WRITE_FIRST):
	case OP(RDMA_WRITE_MIDDLE):
		if (opcode == OP(RDMA_WRITE_MIDDLE) ||
		    opcode == OP(RDMA_WRITE_LAST) ||
		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
			break;
		goto nack_inv;

	case OP(RDMA_READ_REQUEST):
	case OP(COMPARE_SWAP):
	case OP(FETCH_ADD):
		/*
		 * Drop all new requests until a response has been sent.  A
		 * new request then ACKs the RDMA response we sent.  Relaxed
		 * ordering would allow new requests to be processed but we
		 * would need to keep a queue of rwqe's for all that are in
		 * progress.  Note that we can't RNR NAK this request since
		 * the RDMA READ or atomic response is already queued to be
		 * sent (unless we implement a response send queue).
		 */
		goto done;

	default:
		if (opcode == OP(SEND_MIDDLE) ||
		    opcode == OP(SEND_LAST) ||
		    opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
		    opcode == OP(RDMA_WRITE_MIDDLE) ||
		    opcode == OP(RDMA_WRITE_LAST) ||
		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
			goto nack_inv;
		break;
	}

	wc.imm_data = 0;
	wc.wc_flags = 0;

	/* OK, process the packet. */
	switch (opcode) {
	case OP(SEND_FIRST):
		if (!ipath_get_rwqe(qp, 0)) {
		rnr_nak:
			/*
			 * A RNR NAK will ACK earlier sends and RDMA writes.
			 * Don't queue the NAK if a RDMA read or atomic
			 * is pending though.
			 */
			spin_lock(&qp->s_lock);
			if (qp->s_ack_state >=
			    OP(RDMA_READ_REQUEST) &&
			    qp->s_ack_state != IB_OPCODE_ACKNOWLEDGE) {
				spin_unlock(&qp->s_lock);
				goto done;
			}
			qp->s_ack_state = OP(SEND_ONLY);
			qp->s_nak_state = IB_RNR_NAK | qp->s_min_rnr_timer;
			qp->s_ack_psn = qp->r_psn;
			goto resched;
		}
		qp->r_rcv_len = 0;
		/* FALLTHROUGH */
	case OP(SEND_MIDDLE):
	case OP(RDMA_WRITE_MIDDLE):
	send_middle:
		/* Check for invalid length PMTU or posted rwqe len. */
		if (unlikely(tlen != (hdrsize + pmtu + 4)))
			goto nack_inv;
		qp->r_rcv_len += pmtu;
		if (unlikely(qp->r_rcv_len > qp->r_len))
			goto nack_inv;
		ipath_copy_sge(&qp->r_sge, data, pmtu);
		break;

	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
		/* consume RWQE */
		if (!ipath_get_rwqe(qp, 1))
			goto rnr_nak;
		goto send_last_imm;

	case OP(SEND_ONLY):
	case OP(SEND_ONLY_WITH_IMMEDIATE):
		if (!ipath_get_rwqe(qp, 0))
			goto rnr_nak;
		qp->r_rcv_len = 0;
		if (opcode == OP(SEND_ONLY))
			goto send_last;
		/* FALLTHROUGH */
	case OP(SEND_LAST_WITH_IMMEDIATE):
	send_last_imm:
		if (header_in_data) {
			wc.imm_data = *(__be32 *) data;
			data += sizeof(__be32);
		} else {
			/* Immediate data comes after BTH */
			wc.imm_data = ohdr->u.imm_data;
		}
		hdrsize += 4;
		wc.wc_flags = IB_WC_WITH_IMM;
		/* FALLTHROUGH */
	case OP(SEND_LAST):
	case OP(RDMA_WRITE_LAST):
	send_last:
		/* Get the number of bytes the message was padded by. */
		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
		/* Check for invalid length. */
		/* XXX LAST len should be >= 1 */
		if (unlikely(tlen < (hdrsize + pad + 4)))
			goto nack_inv;
		/* Don't count the CRC. */
		tlen -= (hdrsize + pad + 4);
		wc.byte_len = tlen + qp->r_rcv_len;
		if (unlikely(wc.byte_len > qp->r_len))
			goto nack_inv;
		ipath_copy_sge(&qp->r_sge, data, tlen);
		atomic_inc(&qp->msn);
		if (opcode == OP(RDMA_WRITE_LAST) ||
		    opcode == OP(RDMA_WRITE_ONLY))
			break;
		wc.wr_id = qp->r_wr_id;
		wc.status = IB_WC_SUCCESS;
		wc.opcode = IB_WC_RECV;
		wc.vendor_err = 0;
		wc.qp_num = qp->ibqp.qp_num;
		wc.src_qp = qp->remote_qpn;
		wc.pkey_index = 0;
		wc.slid = qp->remote_ah_attr.dlid;
		wc.sl = qp->remote_ah_attr.sl;
		wc.dlid_path_bits = 0;
		wc.port_num = 0;
		/* Signal completion event if the solicited bit is set. */
		ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
			       (ohdr->bth[0] &
				__constant_cpu_to_be32(1 << 23)) != 0);
		break;

	case OP(RDMA_WRITE_FIRST):
	case OP(RDMA_WRITE_ONLY):
	case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
		/* consume RWQE */
		/* RETH comes after BTH */
		if (!header_in_data)
			reth = &ohdr->u.rc.reth;
		else {
			reth = (struct ib_reth *)data;
			data += sizeof(*reth);
		}
		hdrsize += sizeof(*reth);
		qp->r_len = be32_to_cpu(reth->length);
		qp->r_rcv_len = 0;
		if (qp->r_len != 0) {
			u32 rkey = be32_to_cpu(reth->rkey);
			u64 vaddr = be64_to_cpu(reth->vaddr);
			int ok;

			/* Check rkey & NAK */
			ok = ipath_rkey_ok(dev, &qp->r_sge,
					   qp->r_len, vaddr, rkey,
					   IB_ACCESS_REMOTE_WRITE);
			if (unlikely(!ok)) {
			nack_acc:
				/*
				 * A NAK will ACK earlier sends and RDMA
				 * writes.  Don't queue the NAK if a RDMA
				 * read, atomic, or NAK is pending though.
				 */
				spin_lock(&qp->s_lock);
				if (qp->s_ack_state >=
				    OP(RDMA_READ_REQUEST) &&
				    qp->s_ack_state !=
				    IB_OPCODE_ACKNOWLEDGE) {
					spin_unlock(&qp->s_lock);
					goto done;
				}
				/* XXX Flush WQEs */
				qp->state = IB_QPS_ERR;
				qp->s_ack_state = OP(RDMA_WRITE_ONLY);
				qp->s_nak_state =
					IB_NAK_REMOTE_ACCESS_ERROR;
				qp->s_ack_psn = qp->r_psn;
				goto resched;
			}
		} else {
			qp->r_sge.sg_list = NULL;
			qp->r_sge.sge.mr = NULL;
			qp->r_sge.sge.vaddr = NULL;
			qp->r_sge.sge.length = 0;
			qp->r_sge.sge.sge_length = 0;
		}
		if (unlikely(!(qp->qp_access_flags &
			       IB_ACCESS_REMOTE_WRITE)))
			goto nack_acc;
		if (opcode == OP(RDMA_WRITE_FIRST))
			goto send_middle;
		else if (opcode == OP(RDMA_WRITE_ONLY))
			goto send_last;
		if (!ipath_get_rwqe(qp, 1))
			goto rnr_nak;
		goto send_last_imm;

	case OP(RDMA_READ_REQUEST):
		/* RETH comes after BTH */
		if (!header_in_data)
			reth = &ohdr->u.rc.reth;
		else {
			reth = (struct ib_reth *)data;
			data += sizeof(*reth);
		}
		spin_lock(&qp->s_lock);
		if (qp->s_ack_state != OP(ACKNOWLEDGE) &&
		    qp->s_ack_state >= IB_OPCODE_RDMA_READ_REQUEST) {
			spin_unlock(&qp->s_lock);
			goto done;
		}
		qp->s_rdma_len = be32_to_cpu(reth->length);
		if (qp->s_rdma_len != 0) {
			u32 rkey = be32_to_cpu(reth->rkey);
			u64 vaddr = be64_to_cpu(reth->vaddr);
			int ok;

			/* Check rkey & NAK */
			ok = ipath_rkey_ok(dev, &qp->s_rdma_sge,
					   qp->s_rdma_len, vaddr, rkey,
					   IB_ACCESS_REMOTE_READ);
			if (unlikely(!ok)) {
				spin_unlock(&qp->s_lock);
				goto nack_acc;
			}
			/*
			 * Update the next expected PSN.  We add 1 later
			 * below, so only add the remainder here.
			 */
			if (qp->s_rdma_len > pmtu)
				qp->r_psn += (qp->s_rdma_len - 1) / pmtu;
		} else {
			qp->s_rdma_sge.sg_list = NULL;
			qp->s_rdma_sge.num_sge = 0;
			qp->s_rdma_sge.sge.mr = NULL;
			qp->s_rdma_sge.sge.vaddr = NULL;
			qp->s_rdma_sge.sge.length = 0;
			qp->s_rdma_sge.sge.sge_length = 0;
		}
		if (unlikely(!(qp->qp_access_flags &
			       IB_ACCESS_REMOTE_READ)))
			goto nack_acc;
		/*
		 * We need to increment the MSN here instead of when we
		 * finish sending the result since a duplicate request would
		 * increment it more than once.
		 */
		atomic_inc(&qp->msn);
		qp->s_ack_state = opcode;
		qp->s_nak_state = 0;
		qp->s_ack_psn = psn;
		qp->r_psn++;
		qp->r_state = opcode;
		goto rdmadone;

	case OP(COMPARE_SWAP):
	case OP(FETCH_ADD): {
		struct ib_atomic_eth *ateth;
		u64 vaddr;
		u64 sdata;
		u32 rkey;

		if (!header_in_data)
			ateth = &ohdr->u.atomic_eth;
		else {
			ateth = (struct ib_atomic_eth *)data;
			data += sizeof(*ateth);
		}
		vaddr = be64_to_cpu(ateth->vaddr);
		if (unlikely(vaddr & (sizeof(u64) - 1)))
			goto nack_inv;
		rkey = be32_to_cpu(ateth->rkey);
		/* Check rkey & NAK */
		if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge,
					    sizeof(u64), vaddr, rkey,
					    IB_ACCESS_REMOTE_ATOMIC)))
			goto nack_acc;
		if (unlikely(!(qp->qp_access_flags &
			       IB_ACCESS_REMOTE_ATOMIC)))
			goto nack_acc;
		/* Perform atomic OP and save result. */
		sdata = be64_to_cpu(ateth->swap_data);
		spin_lock(&dev->pending_lock);
		qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
		if (opcode == OP(FETCH_ADD))
			*(u64 *) qp->r_sge.sge.vaddr =
				qp->r_atomic_data + sdata;
		else if (qp->r_atomic_data ==
			 be64_to_cpu(ateth->compare_data))
			*(u64 *) qp->r_sge.sge.vaddr = sdata;
		spin_unlock(&dev->pending_lock);
		atomic_inc(&qp->msn);
		qp->r_atomic_psn = psn & IPS_PSN_MASK;
		psn |= 1 << 31;
		break;
	}

	default:
		/* Drop packet for unknown opcodes. */
		goto done;
	}
	qp->r_psn++;
	qp->r_state = opcode;
	/* Send an ACK if requested or required. */
	if (psn & (1 << 31)) {
		/*
		 * Coalesce ACKs unless there is a RDMA READ or
		 * ATOMIC pending.
		 */
		spin_lock(&qp->s_lock);
		if (qp->s_ack_state == OP(ACKNOWLEDGE) ||
		    qp->s_ack_state < IB_OPCODE_RDMA_READ_REQUEST) {
			qp->s_ack_state = opcode;
			qp->s_nak_state = 0;
			qp->s_ack_psn = psn;
			qp->s_ack_atomic = qp->r_atomic_data;
			goto resched;
		}
		spin_unlock(&qp->s_lock);
	}
done:
	spin_unlock_irqrestore(&qp->r_rq.lock, flags);
	goto bail;

resched:
	/*
	 * Try to send ACK right away but not if ipath_do_rc_send() is
	 * active.
	 */
	if (qp->s_hdrwords == 0 &&
	    (qp->s_ack_state < IB_OPCODE_RDMA_READ_REQUEST ||
	     qp->s_ack_state >= IB_OPCODE_COMPARE_SWAP))
		send_rc_ack(qp);

rdmadone:
	spin_unlock(&qp->s_lock);
	spin_unlock_irqrestore(&qp->r_rq.lock, flags);

	/* Call ipath_do_rc_send() in another thread. */
	tasklet_hi_schedule(&qp->s_task);

bail:
	return;
}