IB/ipath: Support larger IB_QP_MAX_DEST_RD_ATOMIC and IB_QP_MAX_QP_RD_ATOMIC
This patch adds support for multiple RDMA reads and atomics to be sent
before an ACK is required to be seen by the requester.
Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
diff --git a/drivers/infiniband/hw/ipath/ipath_ruc.c b/drivers/infiniband/hw/ipath/ipath_ruc.c
index cda8493..d9c2a9b 100644
--- a/drivers/infiniband/hw/ipath/ipath_ruc.c
+++ b/drivers/infiniband/hw/ipath/ipath_ruc.c
@@ -255,6 +255,7 @@
unsigned long flags;
struct ib_wc wc;
u64 sdata;
+ atomic64_t *maddr;
qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
if (!qp) {
@@ -311,7 +312,7 @@
sqp->s_rnr_retry--;
dev->n_rnr_naks++;
sqp->s_rnr_timeout =
- ib_ipath_rnr_table[sqp->r_min_rnr_timer];
+ ib_ipath_rnr_table[qp->r_min_rnr_timer];
ipath_insert_rnr_queue(sqp);
goto done;
}
@@ -344,20 +345,22 @@
wc.sl = sqp->remote_ah_attr.sl;
wc.dlid_path_bits = 0;
wc.port_num = 0;
+ spin_lock_irqsave(&sqp->s_lock, flags);
ipath_sqerror_qp(sqp, &wc);
+ spin_unlock_irqrestore(&sqp->s_lock, flags);
goto done;
}
break;
case IB_WR_RDMA_READ:
+ if (unlikely(!(qp->qp_access_flags &
+ IB_ACCESS_REMOTE_READ)))
+ goto acc_err;
if (unlikely(!ipath_rkey_ok(qp, &sqp->s_sge, wqe->length,
wqe->wr.wr.rdma.remote_addr,
wqe->wr.wr.rdma.rkey,
IB_ACCESS_REMOTE_READ)))
goto acc_err;
- if (unlikely(!(qp->qp_access_flags &
- IB_ACCESS_REMOTE_READ)))
- goto acc_err;
qp->r_sge.sge = wqe->sg_list[0];
qp->r_sge.sg_list = wqe->sg_list + 1;
qp->r_sge.num_sge = wqe->wr.num_sge;
@@ -365,22 +368,22 @@
case IB_WR_ATOMIC_CMP_AND_SWP:
case IB_WR_ATOMIC_FETCH_AND_ADD:
+ if (unlikely(!(qp->qp_access_flags &
+ IB_ACCESS_REMOTE_ATOMIC)))
+ goto acc_err;
if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, sizeof(u64),
- wqe->wr.wr.rdma.remote_addr,
- wqe->wr.wr.rdma.rkey,
+ wqe->wr.wr.atomic.remote_addr,
+ wqe->wr.wr.atomic.rkey,
IB_ACCESS_REMOTE_ATOMIC)))
goto acc_err;
/* Perform atomic OP and save result. */
- sdata = wqe->wr.wr.atomic.swap;
- spin_lock_irqsave(&dev->pending_lock, flags);
- qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
- if (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
- *(u64 *) qp->r_sge.sge.vaddr =
- qp->r_atomic_data + sdata;
- else if (qp->r_atomic_data == wqe->wr.wr.atomic.compare_add)
- *(u64 *) qp->r_sge.sge.vaddr = sdata;
- spin_unlock_irqrestore(&dev->pending_lock, flags);
- *(u64 *) sqp->s_sge.sge.vaddr = qp->r_atomic_data;
+ maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
+ sdata = wqe->wr.wr.atomic.compare_add;
+ *(u64 *) sqp->s_sge.sge.vaddr =
+ (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
+ (u64) atomic64_add_return(sdata, maddr) - sdata :
+ (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
+ sdata, wqe->wr.wr.atomic.swap);
goto send_comp;
default:
@@ -441,7 +444,7 @@
send_comp:
sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
- if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &sqp->s_flags) ||
+ if (!(sqp->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;
@@ -503,7 +506,7 @@
* We clear the tasklet flag now since we are committing to return
* from the tasklet function.
*/
- clear_bit(IPATH_S_BUSY, &qp->s_flags);
+ clear_bit(IPATH_S_BUSY, &qp->s_busy);
tasklet_unlock(&qp->s_task);
want_buffer(dev->dd);
dev->n_piowait++;
@@ -542,6 +545,9 @@
wr->sg_list[0].addr & (sizeof(u64) - 1))) {
ret = -EINVAL;
goto bail;
+ } else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) {
+ ret = -EINVAL;
+ goto bail;
}
/* IB spec says that num_sge == 0 is OK. */
if (wr->num_sge > qp->s_max_sge) {
@@ -648,7 +654,7 @@
u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
struct ipath_other_headers *ohdr;
- if (test_and_set_bit(IPATH_S_BUSY, &qp->s_flags))
+ if (test_and_set_bit(IPATH_S_BUSY, &qp->s_busy))
goto bail;
if (unlikely(qp->remote_ah_attr.dlid == dev->dd->ipath_lid)) {
@@ -684,19 +690,15 @@
*/
spin_lock_irqsave(&qp->s_lock, flags);
- /* Sending responses has higher priority over sending requests. */
- if (qp->s_ack_state != IB_OPCODE_RC_ACKNOWLEDGE &&
- (bth0 = ipath_make_rc_ack(qp, ohdr, pmtu)) != 0)
- bth2 = qp->s_ack_psn++ & IPATH_PSN_MASK;
- else if (!((qp->ibqp.qp_type == IB_QPT_RC) ?
- ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2) :
- ipath_make_uc_req(qp, ohdr, pmtu, &bth0, &bth2))) {
+ if (!((qp->ibqp.qp_type == IB_QPT_RC) ?
+ ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2) :
+ ipath_make_uc_req(qp, ohdr, pmtu, &bth0, &bth2))) {
/*
* Clear the busy bit before unlocking to avoid races with
* adding new work queue items and then failing to process
* them.
*/
- clear_bit(IPATH_S_BUSY, &qp->s_flags);
+ clear_bit(IPATH_S_BUSY, &qp->s_busy);
spin_unlock_irqrestore(&qp->s_lock, flags);
goto bail;
}
@@ -729,7 +731,7 @@
goto again;
clear:
- clear_bit(IPATH_S_BUSY, &qp->s_flags);
+ clear_bit(IPATH_S_BUSY, &qp->s_busy);
bail:
return;
}