Reposting - rediff against an updated trunk.
This patch does the following:
1. Split the QP spinlock to send and receive lock.
The only place where we have to lock both is upon modify_qp,
and that is not on data path.
2. Avoid taking any QP locks when polling CQ.
This last part is achieved by getting rid of the cur field in
mthca_wq, and calculating the number of outstanding WQEs by comparing
the last_comp and next fields. next is only updated by post,
last_comp is only updated by poll.
In a rare case where an overrun is detected, a CQ is locked
and the overrun condition is re-tested, to avoid any potential
for stale last_comp values.
Signed-off-by: Michael S. Tsirkin <[EMAIL PROTECTED]>
Index: hw/mthca/mthca_provider.h
===================================================================
--- hw/mthca/mthca_provider.h (revision 1894)
+++ hw/mthca/mthca_provider.h (working copy)
@@ -166,8 +166,8 @@ struct mthca_cq {
};
struct mthca_wq {
+ spinlock_t lock;
int max;
- int cur;
int next;
int last_comp;
void *last;
@@ -180,7 +180,6 @@ struct mthca_wq {
struct mthca_qp {
struct ib_qp ibqp;
- spinlock_t lock;
atomic_t refcount;
u32 qpn;
int is_direct;
Index: hw/mthca/mthca_cq.c
===================================================================
--- hw/mthca/mthca_cq.c (revision 1894)
+++ hw/mthca/mthca_cq.c (working copy)
@@ -423,15 +423,6 @@ static inline int mthca_poll_one(struct
is_send = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
- if (*cur_qp) {
- if (*freed) {
- wmb();
- update_cons_index(dev, cq, *freed);
- *freed = 0;
- }
- spin_unlock(&(*cur_qp)->lock);
- }
-
/*
* We do not have to take the QP table lock here,
* because CQs will be locked while QPs are removed
@@ -446,8 +437,6 @@ static inline int mthca_poll_one(struct
err = -EINVAL;
goto out;
}
-
- spin_lock(&(*cur_qp)->lock);
}
entry->qp_num = (*cur_qp)->qpn;
@@ -464,11 +453,6 @@ static inline int mthca_poll_one(struct
entry->wr_id = (*cur_qp)->wrid[wqe_index];
}
- if (wq->last_comp < wqe_index)
- wq->cur -= wqe_index - wq->last_comp;
- else
- wq->cur -= wq->max - wq->last_comp + wqe_index;
-
wq->last_comp = wqe_index;
if (0)
@@ -551,9 +535,6 @@ int mthca_poll_cq(struct ib_cq *ibcq, in
update_cons_index(dev, cq, freed);
}
- if (qp)
- spin_unlock(&qp->lock);
-
spin_unlock_irqrestore(&cq->lock, flags);
return err == 0 || err == -EAGAIN ? npolled : err;
Index: hw/mthca/mthca_qp.c
===================================================================
--- hw/mthca/mthca_qp.c (revision 1894)
+++ hw/mthca/mthca_qp.c (working copy)
@@ -577,9 +577,11 @@ int mthca_modify_qp(struct ib_qp *ibqp,
else
cur_state = attr->cur_qp_state;
} else {
- spin_lock_irq(&qp->lock);
+ spin_lock_irq(&qp->sq.lock);
+ spin_lock(&qp->rq.lock);
cur_state = qp->state;
- spin_unlock_irq(&qp->lock);
+ spin_unlock(&qp->rq.lock);
+ spin_unlock_irq(&qp->sq.lock);
}
if (attr_mask & IB_QP_STATE) {
@@ -1076,6 +1078,14 @@ static void mthca_free_memfree(struct mt
}
}
+static void mthca_wq_init(struct mthca_wq* wq)
+{
+ spin_lock_init(&wq->lock);
+ wq->next = 0;
+ wq->last_comp = wq->max - 1;
+ wq->last = NULL;
+}
+
static int mthca_alloc_qp_common(struct mthca_dev *dev,
struct mthca_pd *pd,
struct mthca_cq *send_cq,
@@ -1087,20 +1097,13 @@ static int mthca_alloc_qp_common(struct
int ret;
int i;
- spin_lock_init(&qp->lock);
atomic_set(&qp->refcount, 1);
qp->state = IB_QPS_RESET;
qp->atomic_rd_en = 0;
qp->resp_depth = 0;
qp->sq_policy = send_policy;
- qp->rq.cur = 0;
- qp->sq.cur = 0;
- qp->rq.next = 0;
- qp->sq.next = 0;
- qp->rq.last_comp = qp->rq.max - 1;
- qp->sq.last_comp = qp->sq.max - 1;
- qp->rq.last = NULL;
- qp->sq.last = NULL;
+ mthca_wq_init(&qp->sq);
+ mthca_wq_init(&qp->rq);
ret = mthca_alloc_memfree(dev, qp);
if (ret)
@@ -1394,6 +1397,24 @@ static int build_mlx_header(struct mthca
return 0;
}
+static inline int mthca_wq_overflow(struct mthca_wq* wq, int nreq,
+ struct ib_cq* ib_cq)
+{
+ int cur;
+ struct mthca_cq* cq;
+
+ cur = (wq->next - wq->last_comp - 1) & (wq->max - 1);
+ if (likely(cur + nreq < wq->max))
+ return 0;
+
+ cq = to_mcq(ib_cq);
+ spin_lock(&cq->lock);
+ cur = (wq->next - wq->last_comp - 1) & (wq->max - 1);
+ spin_unlock(&cq->lock);
+
+ return cur + nreq >= wq->max;
+}
+
int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
struct ib_send_wr **bad_wr)
{
@@ -1411,16 +1432,18 @@ int mthca_tavor_post_send(struct ib_qp *
int ind;
u8 op0 = 0;
- spin_lock_irqsave(&qp->lock, flags);
+ spin_lock_irqsave(&qp->sq.lock, flags);
/* XXX check that state is OK to post send */
ind = qp->sq.next;
for (nreq = 0; wr; ++nreq, wr = wr->next) {
- if (qp->sq.cur + nreq >= qp->sq.max) {
- mthca_err(dev, "SQ full (%d posted, %d max, %d nreq)\n",
- qp->sq.cur, qp->sq.max, nreq);
+ if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
+ mthca_err(dev, "SQ %06x full (%d next, %d last_polled,"
+ " %d max, %d nreq)\n", qp->qpn,
+ qp->sq.next, qp->sq.last_comp,
+ qp->sq.max, nreq);
err = -ENOMEM;
*bad_wr = wr;
goto out;
@@ -1591,10 +1614,9 @@ out:
MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
}
- qp->sq.cur += nreq;
qp->sq.next = ind;
- spin_unlock_irqrestore(&qp->lock, flags);
+ spin_unlock_irqrestore(&qp->sq.lock, flags);
return err;
}
@@ -1613,15 +1635,18 @@ int mthca_tavor_post_receive(struct ib_q
void *wqe;
void *prev_wqe;
- spin_lock_irqsave(&qp->lock, flags);
+ spin_lock_irqsave(&qp->rq.lock, flags);
/* XXX check that state is OK to post receive */
ind = qp->rq.next;
for (nreq = 0; wr; ++nreq, wr = wr->next) {
- if (unlikely(qp->rq.cur + nreq >= qp->rq.max)) {
- mthca_err(dev, "RQ %06x full\n", qp->qpn);
+ if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
+ mthca_err(dev, "RQ %06x full (%d next, %d last_polled,"
+ " %d max, %d nreq)\n", qp->qpn,
+ qp->rq.next, qp->rq.last_comp,
+ qp->rq.max, nreq);
err = -ENOMEM;
*bad_wr = wr;
goto out;
@@ -1688,10 +1713,9 @@ out:
MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
}
- qp->rq.cur += nreq;
qp->rq.next = ind;
- spin_unlock_irqrestore(&qp->lock, flags);
+ spin_unlock_irqrestore(&qp->rq.lock, flags);
return err;
}
@@ -1712,16 +1736,18 @@ int mthca_arbel_post_send(struct ib_qp *
int ind;
u8 op0 = 0;
- spin_lock_irqsave(&qp->lock, flags);
+ spin_lock_irqsave(&qp->sq.lock, flags);
/* XXX check that state is OK to post send */
ind = qp->sq.next & (qp->sq.max - 1);
for (nreq = 0; wr; ++nreq, wr = wr->next) {
- if (qp->sq.cur + nreq >= qp->sq.max) {
- mthca_err(dev, "SQ full (%d posted, %d max, %d nreq)\n",
- qp->sq.cur, qp->sq.max, nreq);
+ if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
+ mthca_err(dev, "SQ %06x full (%d next, %d last_polled,"
+ " %d max, %d nreq)\n", qp->qpn,
+ qp->sq.next, qp->sq.last_comp,
+ qp->sq.max, nreq);
err = -ENOMEM;
*bad_wr = wr;
goto out;
@@ -1835,7 +1861,6 @@ out:
f0 | op0);
doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
- qp->sq.cur += nreq;
qp->sq.next += nreq;
/*
@@ -1855,7 +1880,7 @@ out:
MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
}
- spin_unlock_irqrestore(&qp->lock, flags);
+ spin_unlock_irqrestore(&qp->sq.lock, flags);
return err;
}
@@ -1871,15 +1896,18 @@ int mthca_arbel_post_receive(struct ib_q
int i;
void *wqe;
- spin_lock_irqsave(&qp->lock, flags);
+ spin_lock_irqsave(&qp->rq.lock, flags);
/* XXX check that state is OK to post receive */
ind = qp->rq.next & (qp->rq.max - 1);
for (nreq = 0; wr; ++nreq, wr = wr->next) {
- if (unlikely(qp->rq.cur + nreq >= qp->rq.max)) {
- mthca_err(dev, "RQ %06x full\n", qp->qpn);
+ if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
+ mthca_err(dev, "RQ %06x full (%d next, %d last_polled,"
+ " %d max, %d nreq)\n", qp->qpn,
+ qp->rq.next, qp->rq.last_comp,
+ qp->rq.max, nreq);
err = -ENOMEM;
*bad_wr = wr;
goto out;
@@ -1921,7 +1949,6 @@ int mthca_arbel_post_receive(struct ib_q
}
out:
if (likely(nreq)) {
- qp->rq.cur += nreq;
qp->rq.next += nreq;
/*
@@ -1932,7 +1959,7 @@ out:
*qp->rq.db = cpu_to_be32(qp->rq.next & 0xffff);
}
- spin_unlock_irqrestore(&qp->lock, flags);
+ spin_unlock_irqrestore(&qp->rq.lock, flags);
return err;
}
--
MST - Michael S. Tsirkin
_______________________________________________
openib-general mailing list
[email protected]
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
