From e54b82d739d4a2ef992976c8c0692cdf89286420 Mon Sep 17 00:00:00 2001 From: "Michael S. Tsirkin" Date: Thu, 10 Aug 2006 10:46:56 -0700 Subject: IB/mthca: Make fence flag work for send work requests The fence bit needs to be set in the doorbell too, not just the WQE. Signed-off-by: Michael S. Tsirkin Signed-off-by: Roland Dreier --- drivers/infiniband/hw/mthca/mthca_qp.c | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) (limited to 'drivers/infiniband/hw/mthca/mthca_qp.c') diff --git a/drivers/infiniband/hw/mthca/mthca_qp.c b/drivers/infiniband/hw/mthca/mthca_qp.c index cd8b6721ac9..157b4f8ac40 100644 --- a/drivers/infiniband/hw/mthca/mthca_qp.c +++ b/drivers/infiniband/hw/mthca/mthca_qp.c @@ -99,6 +99,10 @@ enum { MTHCA_QP_BIT_RSC = 1 << 3 }; +enum { + MTHCA_SEND_DOORBELL_FENCE = 1 << 5 +}; + struct mthca_qp_path { __be32 port_pkey; u8 rnr_retry; @@ -1502,7 +1506,7 @@ int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, int i; int size; int size0 = 0; - u32 f0 = 0; + u32 f0; int ind; u8 op0 = 0; @@ -1686,6 +1690,8 @@ int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, if (!size0) { size0 = size; op0 = mthca_opcode[wr->opcode]; + f0 = wr->send_flags & IB_SEND_FENCE ? + MTHCA_SEND_DOORBELL_FENCE : 0; } ++ind; @@ -1843,7 +1849,7 @@ int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, int i; int size; int size0 = 0; - u32 f0 = 0; + u32 f0; int ind; u8 op0 = 0; @@ -2051,6 +2057,8 @@ int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, if (!size0) { size0 = size; op0 = mthca_opcode[wr->opcode]; + f0 = wr->send_flags & IB_SEND_FENCE ? + MTHCA_SEND_DOORBELL_FENCE : 0; } ++ind; -- cgit v1.2.3 From a19aa5c5fdda8b556ab238177ee27c5ef7873c94 Mon Sep 17 00:00:00 2001 From: Roland Dreier Date: Fri, 11 Aug 2006 08:56:57 -0700 Subject: IB/mthca: Fix potential AB-BA deadlock with CQ locks When destroying a QP, mthca locks both the QP's send CQ and receive CQ. However, the following scenario is perfectly valid: QP_a: send_cq == CQ_x, recv_cq == CQ_y QP_b: send_cq == CQ_y, recv_cq == CQ_x The old mthca code simply locked send_cq and then recv_cq, which in this case could lead to an AB-BA deadlock if QP_a and QP_b were destroyed simultaneously. We can fix this by changing the locking code to lock the CQ with the lower CQ number first, which will create a consistent lock ordering. Also, the second CQ is locked with spin_lock_nested() to tell lockdep that we know what we're doing with the lock nesting. This bug was found by lockdep. Signed-off-by: Roland Dreier --- drivers/infiniband/hw/mthca/mthca_qp.c | 42 ++++++++++++++++++++++++---------- 1 file changed, 30 insertions(+), 12 deletions(-) (limited to 'drivers/infiniband/hw/mthca/mthca_qp.c') diff --git a/drivers/infiniband/hw/mthca/mthca_qp.c b/drivers/infiniband/hw/mthca/mthca_qp.c index 157b4f8ac40..2e8f6f36e0a 100644 --- a/drivers/infiniband/hw/mthca/mthca_qp.c +++ b/drivers/infiniband/hw/mthca/mthca_qp.c @@ -1263,6 +1263,32 @@ int mthca_alloc_qp(struct mthca_dev *dev, return 0; } +static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq) +{ + if (send_cq == recv_cq) + spin_lock_irq(&send_cq->lock); + else if (send_cq->cqn < recv_cq->cqn) { + spin_lock_irq(&send_cq->lock); + spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING); + } else { + spin_lock_irq(&recv_cq->lock); + spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING); + } +} + +static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq) +{ + if (send_cq == recv_cq) + spin_unlock_irq(&send_cq->lock); + else if (send_cq->cqn < recv_cq->cqn) { + spin_unlock(&recv_cq->lock); + spin_unlock_irq(&send_cq->lock); + } else { + spin_unlock(&send_cq->lock); + spin_unlock_irq(&recv_cq->lock); + } +} + int mthca_alloc_sqp(struct mthca_dev *dev, struct mthca_pd *pd, struct mthca_cq *send_cq, @@ -1315,17 +1341,13 @@ int mthca_alloc_sqp(struct mthca_dev *dev, * Lock CQs here, so that CQ polling code can do QP lookup * without taking a lock. */ - spin_lock_irq(&send_cq->lock); - if (send_cq != recv_cq) - spin_lock(&recv_cq->lock); + mthca_lock_cqs(send_cq, recv_cq); spin_lock(&dev->qp_table.lock); mthca_array_clear(&dev->qp_table.qp, mqpn); spin_unlock(&dev->qp_table.lock); - if (send_cq != recv_cq) - spin_unlock(&recv_cq->lock); - spin_unlock_irq(&send_cq->lock); + mthca_unlock_cqs(send_cq, recv_cq); err_out: dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size, @@ -1359,9 +1381,7 @@ void mthca_free_qp(struct mthca_dev *dev, * Lock CQs here, so that CQ polling code can do QP lookup * without taking a lock. */ - spin_lock_irq(&send_cq->lock); - if (send_cq != recv_cq) - spin_lock(&recv_cq->lock); + mthca_lock_cqs(send_cq, recv_cq); spin_lock(&dev->qp_table.lock); mthca_array_clear(&dev->qp_table.qp, @@ -1369,9 +1389,7 @@ void mthca_free_qp(struct mthca_dev *dev, --qp->refcount; spin_unlock(&dev->qp_table.lock); - if (send_cq != recv_cq) - spin_unlock(&recv_cq->lock); - spin_unlock_irq(&send_cq->lock); + mthca_unlock_cqs(send_cq, recv_cq); wait_event(qp->wait, !get_qp_refcount(dev, qp)); -- cgit v1.2.3