*/
#include "ipath_verbs.h"
-#include "ipath_common.h"
+#include "ipath_kernel.h"
/* cut down ridiculously long IB macro names */
#define OP(x) IB_OPCODE_RC_##x
qp->s_rnr_timeout)
goto done;
+ /* Limit the number of packets sent without an ACK. */
+ if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT) > 0) {
+ qp->s_wait_credit = 1;
+ dev->n_rc_stalls++;
+ 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);
+ goto done;
+ }
+
/* header size in 32-bit words LRH+BTH = (8+12)/4. */
hwords = 5;
bth0 = 0;
/* Check if send work queue is empty. */
if (qp->s_tail == qp->s_head)
goto done;
- qp->s_psn = wqe->psn = qp->s_next_psn;
+ wqe->psn = qp->s_next_psn;
newreq = 1;
}
/*
* original work request since we may need to resend
* it.
*/
- qp->s_sge.sge = wqe->sg_list[0];
- qp->s_sge.sg_list = wqe->sg_list + 1;
- qp->s_sge.num_sge = wqe->wr.num_sge;
- qp->s_len = len = wqe->length;
+ len = wqe->length;
ss = &qp->s_sge;
bth2 = 0;
switch (wqe->wr.opcode) {
default:
goto done;
}
+ qp->s_sge.sge = wqe->sg_list[0];
+ qp->s_sge.sg_list = wqe->sg_list + 1;
+ qp->s_sge.num_sge = wqe->wr.num_sge;
+ qp->s_len = wqe->length;
if (newreq) {
qp->s_tail++;
if (qp->s_tail >= qp->s_size)
qp->s_tail = 0;
}
- bth2 |= qp->s_psn++ & IPATH_PSN_MASK;
- if ((int)(qp->s_psn - qp->s_next_psn) > 0)
- qp->s_next_psn = qp->s_psn;
+ bth2 |= qp->s_psn & IPATH_PSN_MASK;
+ if (wqe->wr.opcode == IB_WR_RDMA_READ)
+ qp->s_psn = wqe->lpsn + 1;
+ else {
+ qp->s_psn++;
+ if ((int)(qp->s_psn - qp->s_next_psn) > 0)
+ qp->s_next_psn = qp->s_psn;
+ }
/*
* Put the QP on the pending list so lost ACKs will cause
* a retry. More than one request can be pending so the
ss = &qp->s_sge;
len = qp->s_len;
if (len > pmtu) {
- /*
- * Request an ACK every 1/2 MB to avoid retransmit
- * timeouts.
- */
- if (((wqe->length - len) % (512 * 1024)) == 0)
- bth2 |= 1 << 31;
len = pmtu;
break;
}
ss = &qp->s_sge;
len = qp->s_len;
if (len > pmtu) {
- /*
- * Request an ACK every 1/2 MB to avoid retransmit
- * timeouts.
- */
- if (((wqe->length - len) % (512 * 1024)) == 0)
- bth2 |= 1 << 31;
len = pmtu;
break;
}
*/
goto done;
}
+ 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;
lrh0 = IPATH_LRH_GRH;
}
/* read pkey_index w/o lock (its atomic) */
- bth0 = ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
+ bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index);
if (qp->r_nak_state)
ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
(qp->r_nak_state <<
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(ipath_layer_get_lid(dev->dd));
+ 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);
struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
struct ipath_ibdev *dev;
- /*
- * If there are no requests pending, we are done.
- */
- if (ipath_cmp24(psn, qp->s_next_psn) >= 0 ||
- qp->s_last == qp->s_tail)
- goto done;
-
if (qp->s_retry == 0) {
wc->wr_id = wqe->wr.wr_id;
wc->status = IB_WC_RETRY_EXC_ERR;
dev->n_rc_resends += (int)qp->s_psn - (int)psn;
reset_psn(qp, psn);
-
-done:
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
struct ib_wc wc;
struct ipath_swqe *wqe;
int ret = 0;
+ u32 ack_psn;
/*
* Remove the QP from the timeout queue (or RNR timeout queue).
list_del_init(&qp->timerwait);
spin_unlock(&dev->pending_lock);
+ /* Nothing is pending to ACK/NAK. */
+ if (unlikely(qp->s_last == qp->s_tail))
+ goto bail;
+
/*
* 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);
- /* Nothing is pending to ACK/NAK. */
- if (qp->s_last == qp->s_tail)
- goto bail;
-
/*
* The MSN might be for a later WQE than the PSN indicates so
* only complete WQEs that the PSN finishes.
*/
- while (ipath_cmp24(psn, wqe->lpsn) >= 0) {
- /* If we are ACKing a WQE, the MSN should be >= the SSN. */
- if (ipath_cmp24(aeth, wqe->ssn) < 0)
- break;
+ while (ipath_cmp24(ack_psn, wqe->lpsn) >= 0) {
/*
* If this request is a RDMA read or atomic, and the ACK is
* for a later operation, this ACK NAKs the RDMA read or
* is sent but before the response is received.
*/
if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
- opcode != OP(RDMA_READ_RESPONSE_LAST)) ||
+ (opcode != OP(RDMA_READ_RESPONSE_LAST) ||
+ ipath_cmp24(ack_psn, wqe->lpsn) != 0)) ||
((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
(opcode != OP(ATOMIC_ACKNOWLEDGE) ||
* The last valid PSN seen is the previous
* request's.
*/
- qp->s_last_psn = wqe->psn - 1;
+ update_last_psn(qp, wqe->psn - 1);
/* Retry this request. */
ipath_restart_rc(qp, wqe->psn, &wc);
/*
*/
goto bail;
}
+ if (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)
+ tasklet_hi_schedule(&qp->s_task);
/* Post a send completion queue entry if requested. */
if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &qp->s_flags) ||
(wqe->wr.send_flags & IB_SEND_SIGNALED)) {
ipath_get_credit(qp, aeth);
qp->s_rnr_retry = qp->s_rnr_retry_cnt;
qp->s_retry = qp->s_retry_cnt;
- qp->s_last_psn = psn;
+ update_last_psn(qp, psn);
ret = 1;
goto bail;
goto bail;
/* The last valid PSN is the previous PSN. */
- qp->s_last_psn = psn - 1;
+ update_last_psn(qp, psn - 1);
dev->n_rc_resends += (int)qp->s_psn - (int)psn;
case 3: /* NAK */
/* The last valid PSN seen is the previous request's. */
if (qp->s_last != qp->s_tail)
- qp->s_last_psn = wqe->psn - 1;
+ update_last_psn(qp, wqe->psn - 1);
switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
IPATH_AETH_CREDIT_MASK) {
case 0: /* PSN sequence error */
/* no AETH, no ACK */
if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
dev->n_rdma_seq++;
- ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
+ if (qp->s_last != qp->s_tail)
+ ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
goto ack_done;
}
rdma_read:
* since we don't want s_sge modified.
*/
qp->s_len -= pmtu;
- qp->s_last_psn = psn;
+ update_last_psn(qp, psn);
spin_unlock_irqrestore(&qp->s_lock, flags);
ipath_copy_sge(&qp->s_sge, data, pmtu);
goto bail;
/* ACKs READ req. */
if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
dev->n_rdma_seq++;
- ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
+ if (qp->s_last != qp->s_tail)
+ ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
goto ack_done;
}
/* FALLTHROUGH */
* Address range must be a subset of the original
* request and start on pmtu boundaries.
*/
- ok = ipath_rkey_ok(dev, &qp->s_rdma_sge,
+ ok = ipath_rkey_ok(qp, &qp->s_rdma_sge,
qp->s_rdma_len, vaddr, rkey,
IB_ACCESS_REMOTE_READ);
if (unlikely(!ok)) {
return 1;
}
+static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
+{
+ spin_lock_irq(&qp->s_lock);
+ qp->state = IB_QPS_ERR;
+ ipath_error_qp(qp, err);
+ spin_unlock_irq(&qp->s_lock);
+}
+
/**
* ipath_rc_rcv - process an incoming RC packet
* @dev: the device this packet came in on
struct ib_reth *reth;
int header_in_data;
+ /* Validate the SLID. See Ch. 9.6.1.5 */
+ if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
+ goto done;
+
/* Check for GRH */
if (!has_grh) {
ohdr = &hdr->u.oth;
* 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;
+ header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
if (header_in_data) {
psn = be32_to_cpu(((__be32 *) data)[0]);
data += sizeof(__be32);
*/
if (qp->r_ack_state >= OP(COMPARE_SWAP))
goto send_ack;
- /* XXX Flush WQEs */
- qp->state = IB_QPS_ERR;
+ ipath_rc_error(qp, IB_WC_REM_INV_REQ_ERR);
qp->r_ack_state = OP(SEND_ONLY);
qp->r_nak_state = IB_NAK_INVALID_REQUEST;
qp->r_ack_psn = qp->r_psn;
goto nack_inv;
ipath_copy_sge(&qp->r_sge, data, tlen);
qp->r_msn++;
- if (opcode == OP(RDMA_WRITE_LAST) ||
- opcode == OP(RDMA_WRITE_ONLY))
+ if (!qp->r_wrid_valid)
break;
+ qp->r_wrid_valid = 0;
wc.wr_id = qp->r_wr_id;
wc.status = IB_WC_SUCCESS;
wc.opcode = IB_WC_RECV;
int ok;
/* Check rkey & NAK */
- ok = ipath_rkey_ok(dev, &qp->r_sge,
+ ok = ipath_rkey_ok(qp, &qp->r_sge,
qp->r_len, vaddr, rkey,
IB_ACCESS_REMOTE_WRITE);
if (unlikely(!ok))
int ok;
/* Check rkey & NAK */
- ok = ipath_rkey_ok(dev, &qp->s_rdma_sge,
+ ok = ipath_rkey_ok(qp, &qp->s_rdma_sge,
qp->s_rdma_len, vaddr, rkey,
IB_ACCESS_REMOTE_READ);
if (unlikely(!ok)) {
goto nack_inv;
rkey = be32_to_cpu(ateth->rkey);
/* Check rkey & NAK */
- if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge,
+ if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
sizeof(u64), vaddr, rkey,
IB_ACCESS_REMOTE_ATOMIC)))
goto nack_acc;
* is pending though.
*/
if (qp->r_ack_state < OP(COMPARE_SWAP)) {
- /* XXX Flush WQEs */
- qp->state = IB_QPS_ERR;
+ ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
qp->r_ack_state = OP(RDMA_WRITE_ONLY);
qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
qp->r_ack_psn = qp->r_psn;