From c8286944b802c5ce4063ec3c334b38c6757a9434 Mon Sep 17 00:00:00 2001 From: Steve Wise Date: Fri, 2 May 2008 11:17:41 -0500 Subject: RDMA/cxgb3: QP flush fixes - Flush the QP only after the HW disables the connection. Currently we flush the QP when transitioning to CLOSING. This exposes a race condition where the HW can complete a RECV WR, for instance, -and- the SW can flush that same WR. - Only call CQ event handlers on flush IFF we actually flushed something. Signed-off-by: Steve Wise Signed-off-by: Roland Dreier --- drivers/infiniband/hw/cxgb3/cxio_hal.c | 13 ++++++++++--- drivers/infiniband/hw/cxgb3/cxio_hal.h | 4 ++-- drivers/infiniband/hw/cxgb3/iwch_qp.c | 13 ++++++++----- 3 files changed, 20 insertions(+), 10 deletions(-) (limited to 'drivers/infiniband/hw/cxgb3') diff --git a/drivers/infiniband/hw/cxgb3/cxio_hal.c b/drivers/infiniband/hw/cxgb3/cxio_hal.c index ed2ee4ba4b7..5fd8506a865 100644 --- a/drivers/infiniband/hw/cxgb3/cxio_hal.c +++ b/drivers/infiniband/hw/cxgb3/cxio_hal.c @@ -359,9 +359,10 @@ static void insert_recv_cqe(struct t3_wq *wq, struct t3_cq *cq) cq->sw_wptr++; } -void cxio_flush_rq(struct t3_wq *wq, struct t3_cq *cq, int count) +int cxio_flush_rq(struct t3_wq *wq, struct t3_cq *cq, int count) { u32 ptr; + int flushed = 0; PDBG("%s wq %p cq %p\n", __func__, wq, cq); @@ -369,8 +370,11 @@ void cxio_flush_rq(struct t3_wq *wq, struct t3_cq *cq, int count) PDBG("%s rq_rptr %u rq_wptr %u skip count %u\n", __func__, wq->rq_rptr, wq->rq_wptr, count); ptr = wq->rq_rptr + count; - while (ptr++ != wq->rq_wptr) + while (ptr++ != wq->rq_wptr) { insert_recv_cqe(wq, cq); + flushed++; + } + return flushed; } static void insert_sq_cqe(struct t3_wq *wq, struct t3_cq *cq, @@ -394,9 +398,10 @@ static void insert_sq_cqe(struct t3_wq *wq, struct t3_cq *cq, cq->sw_wptr++; } -void cxio_flush_sq(struct t3_wq *wq, struct t3_cq *cq, int count) +int cxio_flush_sq(struct t3_wq *wq, struct t3_cq *cq, int count) { __u32 ptr; + int flushed = 0; struct t3_swsq *sqp = wq->sq + Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2); ptr = wq->sq_rptr + count; @@ -405,7 +410,9 @@ void cxio_flush_sq(struct t3_wq *wq, struct t3_cq *cq, int count) insert_sq_cqe(wq, cq, sqp); sqp++; ptr++; + flushed++; } + return flushed; } /* diff --git a/drivers/infiniband/hw/cxgb3/cxio_hal.h b/drivers/infiniband/hw/cxgb3/cxio_hal.h index 2bcff7f5046..69ab08ebc68 100644 --- a/drivers/infiniband/hw/cxgb3/cxio_hal.h +++ b/drivers/infiniband/hw/cxgb3/cxio_hal.h @@ -173,8 +173,8 @@ u32 cxio_hal_get_pdid(struct cxio_hal_resource *rscp); void cxio_hal_put_pdid(struct cxio_hal_resource *rscp, u32 pdid); int __init cxio_hal_init(void); void __exit cxio_hal_exit(void); -void cxio_flush_rq(struct t3_wq *wq, struct t3_cq *cq, int count); -void cxio_flush_sq(struct t3_wq *wq, struct t3_cq *cq, int count); +int cxio_flush_rq(struct t3_wq *wq, struct t3_cq *cq, int count); +int cxio_flush_sq(struct t3_wq *wq, struct t3_cq *cq, int count); void cxio_count_rcqes(struct t3_cq *cq, struct t3_wq *wq, int *count); void cxio_count_scqes(struct t3_cq *cq, struct t3_wq *wq, int *count); void cxio_flush_hw_cq(struct t3_cq *cq); diff --git a/drivers/infiniband/hw/cxgb3/iwch_qp.c b/drivers/infiniband/hw/cxgb3/iwch_qp.c index 9b4be889c58..79dbe5beae5 100644 --- a/drivers/infiniband/hw/cxgb3/iwch_qp.c +++ b/drivers/infiniband/hw/cxgb3/iwch_qp.c @@ -655,6 +655,7 @@ static void __flush_qp(struct iwch_qp *qhp, unsigned long *flag) { struct iwch_cq *rchp, *schp; int count; + int flushed; rchp = get_chp(qhp->rhp, qhp->attr.rcq); schp = get_chp(qhp->rhp, qhp->attr.scq); @@ -669,20 +670,22 @@ static void __flush_qp(struct iwch_qp *qhp, unsigned long *flag) spin_lock(&qhp->lock); cxio_flush_hw_cq(&rchp->cq); cxio_count_rcqes(&rchp->cq, &qhp->wq, &count); - cxio_flush_rq(&qhp->wq, &rchp->cq, count); + flushed = cxio_flush_rq(&qhp->wq, &rchp->cq, count); spin_unlock(&qhp->lock); spin_unlock_irqrestore(&rchp->lock, *flag); - (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context); + if (flushed) + (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context); /* locking heirarchy: cq lock first, then qp lock. */ spin_lock_irqsave(&schp->lock, *flag); spin_lock(&qhp->lock); cxio_flush_hw_cq(&schp->cq); cxio_count_scqes(&schp->cq, &qhp->wq, &count); - cxio_flush_sq(&qhp->wq, &schp->cq, count); + flushed = cxio_flush_sq(&qhp->wq, &schp->cq, count); spin_unlock(&qhp->lock); spin_unlock_irqrestore(&schp->lock, *flag); - (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context); + if (flushed) + (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context); /* deref */ if (atomic_dec_and_test(&qhp->refcnt)) @@ -880,7 +883,6 @@ int iwch_modify_qp(struct iwch_dev *rhp, struct iwch_qp *qhp, ep = qhp->ep; get_ep(&ep->com); } - flush_qp(qhp, &flag); break; case IWCH_QP_STATE_TERMINATE: qhp->attr.state = IWCH_QP_STATE_TERMINATE; @@ -911,6 +913,7 @@ int iwch_modify_qp(struct iwch_dev *rhp, struct iwch_qp *qhp, } switch (attrs->next_state) { case IWCH_QP_STATE_IDLE: + flush_qp(qhp, &flag); qhp->attr.state = IWCH_QP_STATE_IDLE; qhp->attr.llp_stream_handle = NULL; put_ep(&qhp->ep->com); -- cgit v1.2.3 From c4d49776e8f5bf2d900d2b6d4855c1670a535ac5 Mon Sep 17 00:00:00 2001 From: Steve Wise Date: Fri, 2 May 2008 10:57:09 -0700 Subject: RDMA/cxgb3: Silently ignore close reply after abort. Remove bad BUG_ON() that can trigger in correct operation from close_con_rpl(). It is possible to get a close_rpl message on a dead connection. The sequence is: - host refs ep for close exchange - host posts close_req - hw posts PEER_ABORT from incoming RST - host marks ep DEAD - host posts ABORT_RPL and releases ep resources - hw posts CLOSE_RPL - host derefs ep and ep freed. Signed-off-by: Steve Wise Signed-off-by: Roland Dreier --- drivers/infiniband/hw/cxgb3/iwch_cm.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/infiniband/hw/cxgb3') diff --git a/drivers/infiniband/hw/cxgb3/iwch_cm.c b/drivers/infiniband/hw/cxgb3/iwch_cm.c index d44a6df9ad8..32e290e5661 100644 --- a/drivers/infiniband/hw/cxgb3/iwch_cm.c +++ b/drivers/infiniband/hw/cxgb3/iwch_cm.c @@ -1650,8 +1650,8 @@ static int close_con_rpl(struct t3cdev *tdev, struct sk_buff *skb, void *ctx) release = 1; break; case ABORTING: - break; case DEAD: + break; default: BUG_ON(1); break; -- cgit v1.2.3 From 77a8d5741f3ee2c79554382179cca7b5893d6ae9 Mon Sep 17 00:00:00 2001 From: Steve Wise Date: Fri, 2 May 2008 10:57:09 -0700 Subject: RDMA/cxgb3: Bump up the MPA connection setup timeout. Testing on large clusters shows its way too short at 10 secs. Signed-off-by: Steve Wise Signed-off-by: Roland Dreier --- drivers/infiniband/hw/cxgb3/iwch_cm.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'drivers/infiniband/hw/cxgb3') diff --git a/drivers/infiniband/hw/cxgb3/iwch_cm.c b/drivers/infiniband/hw/cxgb3/iwch_cm.c index 32e290e5661..c325c44807e 100644 --- a/drivers/infiniband/hw/cxgb3/iwch_cm.c +++ b/drivers/infiniband/hw/cxgb3/iwch_cm.c @@ -67,10 +67,10 @@ int peer2peer = 0; module_param(peer2peer, int, 0644); MODULE_PARM_DESC(peer2peer, "Support peer2peer ULPs (default=0)"); -static int ep_timeout_secs = 10; +static int ep_timeout_secs = 60; module_param(ep_timeout_secs, int, 0644); MODULE_PARM_DESC(ep_timeout_secs, "CM Endpoint operation timeout " - "in seconds (default=10)"); + "in seconds (default=60)"); static int mpa_rev = 1; module_param(mpa_rev, int, 0644); -- cgit v1.2.3