ipath_rc.c

来自「linux 内核源代码」· C语言 代码 · 共 1,940 行 · 第 1/4 页

		if (wqe->wr.opcode == IB_WR_SEND)
			qp->s_state = OP(SEND_LAST);
		else {
			qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
			/* Immediate data comes after the BTH */
			ohdr->u.imm_data = wqe->wr.imm_data;
			hwords += 1;
		}
		if (wqe->wr.send_flags & IB_SEND_SOLICITED)
			bth0 |= 1 << 23;
		bth2 |= 1 << 31;	/* Request ACK. */
		qp->s_cur++;
		if (qp->s_cur >= qp->s_size)
			qp->s_cur = 0;
		break;

	case OP(RDMA_READ_RESPONSE_LAST):
		/*
		 * This case can only happen if a RDMA write is restarted.
		 * See ipath_restart_rc().
		 */
		ipath_init_restart(qp, wqe);
		/* FALLTHROUGH */
	case OP(RDMA_WRITE_FIRST):
		qp->s_state = OP(RDMA_WRITE_MIDDLE);
		/* FALLTHROUGH */
	case OP(RDMA_WRITE_MIDDLE):
		bth2 = qp->s_psn++ & IPATH_PSN_MASK;
		if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
			qp->s_next_psn = qp->s_psn;
		ss = &qp->s_sge;
		len = qp->s_len;
		if (len > pmtu) {
			len = pmtu;
			break;
		}
		if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
			qp->s_state = OP(RDMA_WRITE_LAST);
		else {
			qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
			/* Immediate data comes after the BTH */
			ohdr->u.imm_data = wqe->wr.imm_data;
			hwords += 1;
			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
				bth0 |= 1 << 23;
		}
		bth2 |= 1 << 31;	/* Request ACK. */
		qp->s_cur++;
		if (qp->s_cur >= qp->s_size)
			qp->s_cur = 0;
		break;

	case OP(RDMA_READ_RESPONSE_MIDDLE):
		/*
		 * This case can only happen if a RDMA read is restarted.
		 * See ipath_restart_rc().
		 */
		ipath_init_restart(qp, wqe);
		len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
		ohdr->u.rc.reth.vaddr =
			cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
		ohdr->u.rc.reth.rkey =
			cpu_to_be32(wqe->wr.wr.rdma.rkey);
		ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
		qp->s_state = OP(RDMA_READ_REQUEST);
		hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
		bth2 = qp->s_psn++ & IPATH_PSN_MASK;
		if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
			qp->s_next_psn = qp->s_psn;
		ss = NULL;
		len = 0;
		qp->s_cur++;
		if (qp->s_cur == qp->s_size)
			qp->s_cur = 0;
		break;
	}
	if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0)
		bth2 |= 1 << 31;	/* Request ACK. */
	qp->s_len -= len;
	qp->s_hdrwords = hwords;
	qp->s_cur_sge = ss;
	qp->s_cur_size = len;
	ipath_make_ruc_header(dev, qp, ohdr, bth0 | (qp->s_state << 24), bth2);
done:
	ret = 1;
bail:
	spin_unlock_irqrestore(&qp->s_lock, flags);
	return ret;
}

/**
 * send_rc_ack - Construct an ACK packet and send it
 * @qp: a pointer to the QP
 *
 * This is called from ipath_rc_rcv() and only uses the receive
 * side QP state.
 * Note that RDMA reads and atomics are handled in the
 * send side QP state and tasklet.
 */
static void send_rc_ack(struct ipath_qp *qp)
{
	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
	u16 lrh0;
	u32 bth0;
	u32 hwords;
	struct ipath_ib_header hdr;
	struct ipath_other_headers *ohdr;
	unsigned long flags;

	/* Don't send ACK or NAK if a RDMA read or atomic is pending. */
	if (qp->r_head_ack_queue != qp->s_tail_ack_queue ||
	    (qp->s_flags & IPATH_S_ACK_PENDING) ||
	    qp->s_ack_state != OP(ACKNOWLEDGE))
		goto queue_ack;

	/* Construct the header. */
	ohdr = &hdr.u.oth;
	lrh0 = IPATH_LRH_BTH;
	/* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
	hwords = 6;
	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
		hwords += ipath_make_grh(dev, &hdr.u.l.grh,
					 &qp->remote_ah_attr.grh,
					 hwords, 0);
		ohdr = &hdr.u.l.oth;
		lrh0 = IPATH_LRH_GRH;
	}
	/* read pkey_index w/o lock (its atomic) */
	bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index) |
		(OP(ACKNOWLEDGE) << 24) | (1 << 22);
	if (qp->r_nak_state)
		ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
					    (qp->r_nak_state <<
					     IPATH_AETH_CREDIT_SHIFT));
	else
		ohdr->u.aeth = ipath_compute_aeth(qp);
	lrh0 |= qp->remote_ah_attr.sl << 4;
	hdr.lrh[0] = cpu_to_be16(lrh0);
	hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
	hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
	ohdr->bth[0] = cpu_to_be32(bth0);
	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
	ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);

	/*
	 * If we can send the ACK, clear the ACK state.
	 */
	if (ipath_verbs_send(qp, &hdr, hwords, NULL, 0) == 0) {
		dev->n_unicast_xmit++;
		goto done;
	}

	/*
	 * We are out of PIO buffers at the moment.
	 * Pass responsibility for sending the ACK to the
	 * send tasklet so that when a PIO buffer becomes
	 * available, the ACK is sent ahead of other outgoing
	 * packets.
	 */
	dev->n_rc_qacks++;

queue_ack:
	spin_lock_irqsave(&qp->s_lock, flags);
	qp->s_flags |= IPATH_S_ACK_PENDING;
	qp->s_nak_state = qp->r_nak_state;
	qp->s_ack_psn = qp->r_ack_psn;
	spin_unlock_irqrestore(&qp->s_lock, flags);

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

done:
	return;
}

/**
 * reset_psn - reset the QP state to send starting from PSN
 * @qp: the QP
 * @psn: the packet sequence number to restart at
 *
 * This is called from ipath_rc_rcv() to process an incoming RC ACK
 * for the given QP.
 * Called at interrupt level with the QP s_lock held.
 */
static void reset_psn(struct ipath_qp *qp, u32 psn)
{
	u32 n = qp->s_last;
	struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
	u32 opcode;

	qp->s_cur = n;

	/*
	 * If we are starting the request from the beginning,
	 * let the normal send code handle initialization.
	 */
	if (ipath_cmp24(psn, wqe->psn) <= 0) {
		qp->s_state = OP(SEND_LAST);
		goto done;
	}

	/* Find the work request opcode corresponding to the given PSN. */
	opcode = wqe->wr.opcode;
	for (;;) {
		int diff;

		if (++n == qp->s_size)
			n = 0;
		if (n == qp->s_tail)
			break;
		wqe = get_swqe_ptr(qp, n);
		diff = ipath_cmp24(psn, wqe->psn);
		if (diff < 0)
			break;
		qp->s_cur = n;
		/*
		 * If we are starting the request from the beginning,
		 * let the normal send code handle initialization.
		 */
		if (diff == 0) {
			qp->s_state = OP(SEND_LAST);
			goto done;
		}
		opcode = wqe->wr.opcode;
	}

	/*
	 * Set the state to restart in the middle of a request.
	 * Don't change the s_sge, s_cur_sge, or s_cur_size.
	 * See ipath_do_rc_send().
	 */
	switch (opcode) {
	case IB_WR_SEND:
	case IB_WR_SEND_WITH_IMM:
		qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
		break;

	case IB_WR_RDMA_WRITE:
	case IB_WR_RDMA_WRITE_WITH_IMM:
		qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
		break;

	case IB_WR_RDMA_READ:
		qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
		break;

	default:
		/*
		 * This case shouldn't happen since its only
		 * one PSN per req.
		 */
		qp->s_state = OP(SEND_LAST);
	}
done:
	qp->s_psn = psn;
}

/**
 * ipath_restart_rc - back up requester to resend the last un-ACKed request
 * @qp: the QP to restart
 * @psn: packet sequence number for the request
 * @wc: the work completion request
 *
 * The QP s_lock should be held and interrupts disabled.
 */
void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
{
	struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
	struct ipath_ibdev *dev;

	if (qp->s_retry == 0) {
		wc->wr_id = wqe->wr.wr_id;
		wc->status = IB_WC_RETRY_EXC_ERR;
		wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
		wc->vendor_err = 0;
		wc->byte_len = 0;
		wc->qp = &qp->ibqp;
		wc->imm_data = 0;
		wc->src_qp = qp->remote_qpn;
		wc->wc_flags = 0;
		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;
		ipath_sqerror_qp(qp, wc);
		goto bail;
	}
	qp->s_retry--;

	/*
	 * Remove the QP from the timeout queue.
	 * Note: it may already have been removed by ipath_ib_timer().
	 */
	dev = to_idev(qp->ibqp.device);
	spin_lock(&dev->pending_lock);
	if (!list_empty(&qp->timerwait))
		list_del_init(&qp->timerwait);
	spin_unlock(&dev->pending_lock);

	if (wqe->wr.opcode == IB_WR_RDMA_READ)
		dev->n_rc_resends++;
	else
		dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;

	reset_psn(qp, psn);
	tasklet_hi_schedule(&qp->s_task);

bail:
	return;
}

static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
{
	if (qp->s_wait_credit) {
		qp->s_wait_credit = 0;
		tasklet_hi_schedule(&qp->s_task);
	}
	qp->s_last_psn = psn;
}

/**
 * do_rc_ack - process an incoming RC ACK
 * @qp: the QP the ACK came in on
 * @psn: the packet sequence number of the ACK
 * @opcode: the opcode of the request that resulted in the ACK
 *
 * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
 * for the given QP.
 * Called at interrupt level with the QP s_lock held and interrupts disabled.
 * Returns 1 if OK, 0 if current operation should be aborted (NAK).
 */
static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode,
		     u64 val)
{
	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
	struct ib_wc wc;
	struct ipath_swqe *wqe;
	int ret = 0;
	u32 ack_psn;
	int diff;

	/*
	 * Remove the QP from the timeout queue (or RNR timeout queue).
	 * If ipath_ib_timer() has already removed it,
	 * it's OK since we hold the QP s_lock and ipath_restart_rc()
	 * just won't find anything to restart if we ACK everything.
	 */
	spin_lock(&dev->pending_lock);
	if (!list_empty(&qp->timerwait))
		list_del_init(&qp->timerwait);
	spin_unlock(&dev->pending_lock);

	/*
	 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
	 * requests and implicitly NAK RDMA read and atomic requests issued
	 * before the NAK'ed request.  The MSN won't include the NAK'ed
	 * request but will include an ACK'ed request(s).
	 */
	ack_psn = psn;
	if (aeth >> 29)
		ack_psn--;
	wqe = get_swqe_ptr(qp, qp->s_last);

	/*
	 * The MSN might be for a later WQE than the PSN indicates so
	 * only complete WQEs that the PSN finishes.
	 */
	while ((diff = ipath_cmp24(ack_psn, wqe->lpsn)) >= 0) {
		/*
		 * RDMA_READ_RESPONSE_ONLY is a special case since
		 * we want to generate completion events for everything
		 * before the RDMA read, copy the data, then generate
		 * the completion for the read.
		 */
		if (wqe->wr.opcode == IB_WR_RDMA_READ &&
		    opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
		    diff == 0) {
			ret = 1;
			goto bail;
		}
		/*
		 * If this request is a RDMA read or atomic, and the ACK is
		 * for a later operation, this ACK NAKs the RDMA read or
		 * atomic.  In other words, only a RDMA_READ_LAST or ONLY
		 * can ACK a RDMA read and likewise for atomic ops.  Note
		 * that the NAK case can only happen if relaxed ordering is
		 * used and requests are sent after an RDMA read or atomic
		 * is sent but before the response is received.
		 */
		if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
		     (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
		    ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
		      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
		     (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
			/*
			 * The last valid PSN seen is the previous
			 * request's.
			 */
			update_last_psn(qp, wqe->psn - 1);
			/* Retry this request. */
			ipath_restart_rc(qp, wqe->psn, &wc);
			/*
			 * No need to process the ACK/NAK since we are
			 * restarting an earlier request.
			 */
			goto bail;
		}
		if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
			*(u64 *) wqe->sg_list[0].vaddr = val;
		if (qp->s_num_rd_atomic &&
		    (wqe->wr.opcode == IB_WR_RDMA_READ ||
		     wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
		     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
			qp->s_num_rd_atomic--;
			/* Restart sending task if fence is complete */
			if ((qp->s_flags & IPATH_S_FENCE_PENDING) &&
			    !qp->s_num_rd_atomic) {
				qp->s_flags &= ~IPATH_S_FENCE_PENDING;
				tasklet_hi_schedule(&qp->s_task);
			} else if (qp->s_flags & IPATH_S_RDMAR_PENDING) {
				qp->s_flags &= ~IPATH_S_RDMAR_PENDING;
				tasklet_hi_schedule(&qp->s_task);
			}
		}
		/* Post a send completion queue entry if requested. */
		if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
		    (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
			wc.wr_id = wqe->wr.wr_id;
			wc.status = IB_WC_SUCCESS;
			wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
			wc.vendor_err = 0;
			wc.byte_len = wqe->length;
			wc.imm_data = 0;
			wc.qp = &qp->ibqp;
			wc.src_qp = qp->remote_qpn;
			wc.wc_flags = 0;
			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;
			ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
		}
		qp->s_retry = qp->s_retry_cnt;
		/*
		 * If we are completing a request which is in the process of
		 * being resent, we can stop resending it since we know the
		 * responder has already seen it.
		 */
		if (qp->s_last == qp->s_cur) {
			if (++qp->s_cur >= qp->s_size)
				qp->s_cur = 0;
			qp->s_last = qp->s_cur;
			if (qp->s_last == qp->s_tail)
				break;
			wqe = get_swqe_ptr(qp, qp->s_cur);
			qp->s_state = OP(SEND_LAST);
			qp->s_psn = wqe->psn;
		} else {
			if (++qp->s_last >= qp->s_size)
				qp->s_last = 0;
			if (qp->s_last == qp->s_tail)
				break;
			wqe = get_swqe_ptr(qp, qp->s_last);
		}
	}

	switch (aeth >> 29) {
	case 0:		/* ACK */
		dev->n_rc_acks++;
		/* If this is a partial ACK, reset the retransmit timer. */
		if (qp->s_last != qp->s_tail) {
			spin_lock(&dev->pending_lock);
			if (list_empty(&qp->timerwait))
				list_add_tail(&qp->timerwait,
					&dev->pending[dev->pending_index]);
			spin_unlock(&dev->pending_lock);
			/*
			 * If we get a partial ACK for a resent operation,
			 * we can stop resending the earlier packets and
			 * continue with the next packet the receiver wants.
			 */