]> git.karo-electronics.de Git - karo-tx-linux.git/blobdiff - drivers/infiniband/hw/bnxt_re/ib_verbs.c
Merge tag 'xfs-4.12-fixes-4' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux
[karo-tx-linux.git] / drivers / infiniband / hw / bnxt_re / ib_verbs.c
index 7ba9e699d7abc65cb82fbd588e9ea12e7f71433b..c7bd68311d0c5317973ecf9c3fb8fce8fe8ffad4 100644 (file)
 #include "ib_verbs.h"
 #include <rdma/bnxt_re-abi.h>
 
+static int __from_ib_access_flags(int iflags)
+{
+       int qflags = 0;
+
+       if (iflags & IB_ACCESS_LOCAL_WRITE)
+               qflags |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
+       if (iflags & IB_ACCESS_REMOTE_READ)
+               qflags |= BNXT_QPLIB_ACCESS_REMOTE_READ;
+       if (iflags & IB_ACCESS_REMOTE_WRITE)
+               qflags |= BNXT_QPLIB_ACCESS_REMOTE_WRITE;
+       if (iflags & IB_ACCESS_REMOTE_ATOMIC)
+               qflags |= BNXT_QPLIB_ACCESS_REMOTE_ATOMIC;
+       if (iflags & IB_ACCESS_MW_BIND)
+               qflags |= BNXT_QPLIB_ACCESS_MW_BIND;
+       if (iflags & IB_ZERO_BASED)
+               qflags |= BNXT_QPLIB_ACCESS_ZERO_BASED;
+       if (iflags & IB_ACCESS_ON_DEMAND)
+               qflags |= BNXT_QPLIB_ACCESS_ON_DEMAND;
+       return qflags;
+};
+
+static enum ib_access_flags __to_ib_access_flags(int qflags)
+{
+       enum ib_access_flags iflags = 0;
+
+       if (qflags & BNXT_QPLIB_ACCESS_LOCAL_WRITE)
+               iflags |= IB_ACCESS_LOCAL_WRITE;
+       if (qflags & BNXT_QPLIB_ACCESS_REMOTE_WRITE)
+               iflags |= IB_ACCESS_REMOTE_WRITE;
+       if (qflags & BNXT_QPLIB_ACCESS_REMOTE_READ)
+               iflags |= IB_ACCESS_REMOTE_READ;
+       if (qflags & BNXT_QPLIB_ACCESS_REMOTE_ATOMIC)
+               iflags |= IB_ACCESS_REMOTE_ATOMIC;
+       if (qflags & BNXT_QPLIB_ACCESS_MW_BIND)
+               iflags |= IB_ACCESS_MW_BIND;
+       if (qflags & BNXT_QPLIB_ACCESS_ZERO_BASED)
+               iflags |= IB_ZERO_BASED;
+       if (qflags & BNXT_QPLIB_ACCESS_ON_DEMAND)
+               iflags |= IB_ACCESS_ON_DEMAND;
+       return iflags;
+};
+
 static int bnxt_re_build_sgl(struct ib_sge *ib_sg_list,
                             struct bnxt_qplib_sge *sg_list, int num)
 {
@@ -149,8 +191,8 @@ int bnxt_re_query_device(struct ib_device *ibdev,
        ib_attr->max_total_mcast_qp_attach = 0;
        ib_attr->max_ah = dev_attr->max_ah;
 
-       ib_attr->max_fmr = dev_attr->max_fmr;
-       ib_attr->max_map_per_fmr = 1;   /* ? */
+       ib_attr->max_fmr = 0;
+       ib_attr->max_map_per_fmr = 0;
 
        ib_attr->max_srq = dev_attr->max_srq;
        ib_attr->max_srq_wr = dev_attr->max_srq_wqes;
@@ -410,6 +452,158 @@ enum rdma_link_layer bnxt_re_get_link_layer(struct ib_device *ibdev,
        return IB_LINK_LAYER_ETHERNET;
 }
 
+#define        BNXT_RE_FENCE_PBL_SIZE  DIV_ROUND_UP(BNXT_RE_FENCE_BYTES, PAGE_SIZE)
+
+static void bnxt_re_create_fence_wqe(struct bnxt_re_pd *pd)
+{
+       struct bnxt_re_fence_data *fence = &pd->fence;
+       struct ib_mr *ib_mr = &fence->mr->ib_mr;
+       struct bnxt_qplib_swqe *wqe = &fence->bind_wqe;
+
+       memset(wqe, 0, sizeof(*wqe));
+       wqe->type = BNXT_QPLIB_SWQE_TYPE_BIND_MW;
+       wqe->wr_id = BNXT_QPLIB_FENCE_WRID;
+       wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
+       wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
+       wqe->bind.zero_based = false;
+       wqe->bind.parent_l_key = ib_mr->lkey;
+       wqe->bind.va = (u64)(unsigned long)fence->va;
+       wqe->bind.length = fence->size;
+       wqe->bind.access_cntl = __from_ib_access_flags(IB_ACCESS_REMOTE_READ);
+       wqe->bind.mw_type = SQ_BIND_MW_TYPE_TYPE1;
+
+       /* Save the initial rkey in fence structure for now;
+        * wqe->bind.r_key will be set at (re)bind time.
+        */
+       fence->bind_rkey = ib_inc_rkey(fence->mw->rkey);
+}
+
+static int bnxt_re_bind_fence_mw(struct bnxt_qplib_qp *qplib_qp)
+{
+       struct bnxt_re_qp *qp = container_of(qplib_qp, struct bnxt_re_qp,
+                                            qplib_qp);
+       struct ib_pd *ib_pd = qp->ib_qp.pd;
+       struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
+       struct bnxt_re_fence_data *fence = &pd->fence;
+       struct bnxt_qplib_swqe *fence_wqe = &fence->bind_wqe;
+       struct bnxt_qplib_swqe wqe;
+       int rc;
+
+       memcpy(&wqe, fence_wqe, sizeof(wqe));
+       wqe.bind.r_key = fence->bind_rkey;
+       fence->bind_rkey = ib_inc_rkey(fence->bind_rkey);
+
+       dev_dbg(rdev_to_dev(qp->rdev),
+               "Posting bind fence-WQE: rkey: %#x QP: %d PD: %p\n",
+               wqe.bind.r_key, qp->qplib_qp.id, pd);
+       rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
+       if (rc) {
+               dev_err(rdev_to_dev(qp->rdev), "Failed to bind fence-WQE\n");
+               return rc;
+       }
+       bnxt_qplib_post_send_db(&qp->qplib_qp);
+
+       return rc;
+}
+
+static void bnxt_re_destroy_fence_mr(struct bnxt_re_pd *pd)
+{
+       struct bnxt_re_fence_data *fence = &pd->fence;
+       struct bnxt_re_dev *rdev = pd->rdev;
+       struct device *dev = &rdev->en_dev->pdev->dev;
+       struct bnxt_re_mr *mr = fence->mr;
+
+       if (fence->mw) {
+               bnxt_re_dealloc_mw(fence->mw);
+               fence->mw = NULL;
+       }
+       if (mr) {
+               if (mr->ib_mr.rkey)
+                       bnxt_qplib_dereg_mrw(&rdev->qplib_res, &mr->qplib_mr,
+                                            true);
+               if (mr->ib_mr.lkey)
+                       bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
+               kfree(mr);
+               fence->mr = NULL;
+       }
+       if (fence->dma_addr) {
+               dma_unmap_single(dev, fence->dma_addr, BNXT_RE_FENCE_BYTES,
+                                DMA_BIDIRECTIONAL);
+               fence->dma_addr = 0;
+       }
+}
+
+static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd)
+{
+       int mr_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_MW_BIND;
+       struct bnxt_re_fence_data *fence = &pd->fence;
+       struct bnxt_re_dev *rdev = pd->rdev;
+       struct device *dev = &rdev->en_dev->pdev->dev;
+       struct bnxt_re_mr *mr = NULL;
+       dma_addr_t dma_addr = 0;
+       struct ib_mw *mw;
+       u64 pbl_tbl;
+       int rc;
+
+       dma_addr = dma_map_single(dev, fence->va, BNXT_RE_FENCE_BYTES,
+                                 DMA_BIDIRECTIONAL);
+       rc = dma_mapping_error(dev, dma_addr);
+       if (rc) {
+               dev_err(rdev_to_dev(rdev), "Failed to dma-map fence-MR-mem\n");
+               rc = -EIO;
+               fence->dma_addr = 0;
+               goto fail;
+       }
+       fence->dma_addr = dma_addr;
+
+       /* Allocate a MR */
+       mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+       if (!mr) {
+               rc = -ENOMEM;
+               goto fail;
+       }
+       fence->mr = mr;
+       mr->rdev = rdev;
+       mr->qplib_mr.pd = &pd->qplib_pd;
+       mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
+       mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
+       rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
+       if (rc) {
+               dev_err(rdev_to_dev(rdev), "Failed to alloc fence-HW-MR\n");
+               goto fail;
+       }
+
+       /* Register MR */
+       mr->ib_mr.lkey = mr->qplib_mr.lkey;
+       mr->qplib_mr.va = (u64)(unsigned long)fence->va;
+       mr->qplib_mr.total_size = BNXT_RE_FENCE_BYTES;
+       pbl_tbl = dma_addr;
+       rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, &pbl_tbl,
+                              BNXT_RE_FENCE_PBL_SIZE, false);
+       if (rc) {
+               dev_err(rdev_to_dev(rdev), "Failed to register fence-MR\n");
+               goto fail;
+       }
+       mr->ib_mr.rkey = mr->qplib_mr.rkey;
+
+       /* Create a fence MW only for kernel consumers */
+       mw = bnxt_re_alloc_mw(&pd->ib_pd, IB_MW_TYPE_1, NULL);
+       if (!mw) {
+               dev_err(rdev_to_dev(rdev),
+                       "Failed to create fence-MW for PD: %p\n", pd);
+               rc = -EINVAL;
+               goto fail;
+       }
+       fence->mw = mw;
+
+       bnxt_re_create_fence_wqe(pd);
+       return 0;
+
+fail:
+       bnxt_re_destroy_fence_mr(pd);
+       return rc;
+}
+
 /* Protection Domains */
 int bnxt_re_dealloc_pd(struct ib_pd *ib_pd)
 {
@@ -417,6 +611,7 @@ int bnxt_re_dealloc_pd(struct ib_pd *ib_pd)
        struct bnxt_re_dev *rdev = pd->rdev;
        int rc;
 
+       bnxt_re_destroy_fence_mr(pd);
        if (ib_pd->uobject && pd->dpi.dbr) {
                struct ib_ucontext *ib_uctx = ib_pd->uobject->context;
                struct bnxt_re_ucontext *ucntx;
@@ -498,6 +693,10 @@ struct ib_pd *bnxt_re_alloc_pd(struct ib_device *ibdev,
                }
        }
 
+       if (!udata)
+               if (bnxt_re_create_fence_mr(pd))
+                       dev_warn(rdev_to_dev(rdev),
+                                "Failed to create Fence-MR\n");
        return &pd->ib_pd;
 dbfail:
        (void)bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl,
@@ -849,12 +1048,16 @@ static struct bnxt_re_qp *bnxt_re_create_shadow_qp
        /* Shadow QP SQ depth should be same as QP1 RQ depth */
        qp->qplib_qp.sq.max_wqe = qp1_qp->rq.max_wqe;
        qp->qplib_qp.sq.max_sge = 2;
+       /* Q full delta can be 1 since it is internal QP */
+       qp->qplib_qp.sq.q_full_delta = 1;
 
        qp->qplib_qp.scq = qp1_qp->scq;
        qp->qplib_qp.rcq = qp1_qp->rcq;
 
        qp->qplib_qp.rq.max_wqe = qp1_qp->rq.max_wqe;
        qp->qplib_qp.rq.max_sge = qp1_qp->rq.max_sge;
+       /* Q full delta can be 1 since it is internal QP */
+       qp->qplib_qp.rq.q_full_delta = 1;
 
        qp->qplib_qp.mtu = qp1_qp->mtu;
 
@@ -917,10 +1120,6 @@ struct ib_qp *bnxt_re_create_qp(struct ib_pd *ib_pd,
        qp->qplib_qp.sig_type = ((qp_init_attr->sq_sig_type ==
                                  IB_SIGNAL_ALL_WR) ? true : false);
 
-       entries = roundup_pow_of_two(qp_init_attr->cap.max_send_wr + 1);
-       qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
-                                       dev_attr->max_qp_wqes + 1);
-
        qp->qplib_qp.sq.max_sge = qp_init_attr->cap.max_send_sge;
        if (qp->qplib_qp.sq.max_sge > dev_attr->max_qp_sges)
                qp->qplib_qp.sq.max_sge = dev_attr->max_qp_sges;
@@ -959,6 +1158,9 @@ struct ib_qp *bnxt_re_create_qp(struct ib_pd *ib_pd,
                qp->qplib_qp.rq.max_wqe = min_t(u32, entries,
                                                dev_attr->max_qp_wqes + 1);
 
+               qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe -
+                                               qp_init_attr->cap.max_recv_wr;
+
                qp->qplib_qp.rq.max_sge = qp_init_attr->cap.max_recv_sge;
                if (qp->qplib_qp.rq.max_sge > dev_attr->max_qp_sges)
                        qp->qplib_qp.rq.max_sge = dev_attr->max_qp_sges;
@@ -967,6 +1169,12 @@ struct ib_qp *bnxt_re_create_qp(struct ib_pd *ib_pd,
        qp->qplib_qp.mtu = ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
 
        if (qp_init_attr->qp_type == IB_QPT_GSI) {
+               /* Allocate 1 more than what's provided */
+               entries = roundup_pow_of_two(qp_init_attr->cap.max_send_wr + 1);
+               qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
+                                               dev_attr->max_qp_wqes + 1);
+               qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe -
+                                               qp_init_attr->cap.max_send_wr;
                qp->qplib_qp.rq.max_sge = dev_attr->max_qp_sges;
                if (qp->qplib_qp.rq.max_sge > dev_attr->max_qp_sges)
                        qp->qplib_qp.rq.max_sge = dev_attr->max_qp_sges;
@@ -1006,6 +1214,22 @@ struct ib_qp *bnxt_re_create_qp(struct ib_pd *ib_pd,
                }
 
        } else {
+               /* Allocate 128 + 1 more than what's provided */
+               entries = roundup_pow_of_two(qp_init_attr->cap.max_send_wr +
+                                            BNXT_QPLIB_RESERVED_QP_WRS + 1);
+               qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
+                                               dev_attr->max_qp_wqes +
+                                               BNXT_QPLIB_RESERVED_QP_WRS + 1);
+               qp->qplib_qp.sq.q_full_delta = BNXT_QPLIB_RESERVED_QP_WRS + 1;
+
+               /*
+                * Reserving one slot for Phantom WQE. Application can
+                * post one extra entry in this case. But allowing this to avoid
+                * unexpected Queue full condition
+                */
+
+               qp->qplib_qp.sq.q_full_delta -= 1;
+
                qp->qplib_qp.max_rd_atomic = dev_attr->max_qp_rd_atom;
                qp->qplib_qp.max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom;
                if (udata) {
@@ -1025,6 +1249,7 @@ struct ib_qp *bnxt_re_create_qp(struct ib_pd *ib_pd,
 
        qp->ib_qp.qp_num = qp->qplib_qp.id;
        spin_lock_init(&qp->sq_lock);
+       spin_lock_init(&qp->rq_lock);
 
        if (udata) {
                struct bnxt_re_qp_resp resp;
@@ -1129,48 +1354,6 @@ static enum ib_mtu __to_ib_mtu(u32 mtu)
        }
 }
 
-static int __from_ib_access_flags(int iflags)
-{
-       int qflags = 0;
-
-       if (iflags & IB_ACCESS_LOCAL_WRITE)
-               qflags |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
-       if (iflags & IB_ACCESS_REMOTE_READ)
-               qflags |= BNXT_QPLIB_ACCESS_REMOTE_READ;
-       if (iflags & IB_ACCESS_REMOTE_WRITE)
-               qflags |= BNXT_QPLIB_ACCESS_REMOTE_WRITE;
-       if (iflags & IB_ACCESS_REMOTE_ATOMIC)
-               qflags |= BNXT_QPLIB_ACCESS_REMOTE_ATOMIC;
-       if (iflags & IB_ACCESS_MW_BIND)
-               qflags |= BNXT_QPLIB_ACCESS_MW_BIND;
-       if (iflags & IB_ZERO_BASED)
-               qflags |= BNXT_QPLIB_ACCESS_ZERO_BASED;
-       if (iflags & IB_ACCESS_ON_DEMAND)
-               qflags |= BNXT_QPLIB_ACCESS_ON_DEMAND;
-       return qflags;
-};
-
-static enum ib_access_flags __to_ib_access_flags(int qflags)
-{
-       enum ib_access_flags iflags = 0;
-
-       if (qflags & BNXT_QPLIB_ACCESS_LOCAL_WRITE)
-               iflags |= IB_ACCESS_LOCAL_WRITE;
-       if (qflags & BNXT_QPLIB_ACCESS_REMOTE_WRITE)
-               iflags |= IB_ACCESS_REMOTE_WRITE;
-       if (qflags & BNXT_QPLIB_ACCESS_REMOTE_READ)
-               iflags |= IB_ACCESS_REMOTE_READ;
-       if (qflags & BNXT_QPLIB_ACCESS_REMOTE_ATOMIC)
-               iflags |= IB_ACCESS_REMOTE_ATOMIC;
-       if (qflags & BNXT_QPLIB_ACCESS_MW_BIND)
-               iflags |= IB_ACCESS_MW_BIND;
-       if (qflags & BNXT_QPLIB_ACCESS_ZERO_BASED)
-               iflags |= IB_ZERO_BASED;
-       if (qflags & BNXT_QPLIB_ACCESS_ON_DEMAND)
-               iflags |= IB_ACCESS_ON_DEMAND;
-       return iflags;
-};
-
 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev *rdev,
                                    struct bnxt_re_qp *qp1_qp,
                                    int qp_attr_mask)
@@ -1378,11 +1561,21 @@ int bnxt_re_modify_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
                entries = roundup_pow_of_two(qp_attr->cap.max_send_wr);
                qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
                                                dev_attr->max_qp_wqes + 1);
+               qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe -
+                                               qp_attr->cap.max_send_wr;
+               /*
+                * Reserving one slot for Phantom WQE. Some application can
+                * post one extra entry in this case. Allowing this to avoid
+                * unexpected Queue full condition
+                */
+               qp->qplib_qp.sq.q_full_delta -= 1;
                qp->qplib_qp.sq.max_sge = qp_attr->cap.max_send_sge;
                if (qp->qplib_qp.rq.max_wqe) {
                        entries = roundup_pow_of_two(qp_attr->cap.max_recv_wr);
                        qp->qplib_qp.rq.max_wqe =
                                min_t(u32, entries, dev_attr->max_qp_wqes + 1);
+                       qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe -
+                                                      qp_attr->cap.max_recv_wr;
                        qp->qplib_qp.rq.max_sge = qp_attr->cap.max_recv_sge;
                } else {
                        /* SRQ was used prior, just ignore the RQ caps */
@@ -1883,6 +2076,22 @@ static int bnxt_re_copy_wr_payload(struct bnxt_re_dev *rdev,
        return payload_sz;
 }
 
+static void bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp *qp)
+{
+       if ((qp->ib_qp.qp_type == IB_QPT_UD ||
+            qp->ib_qp.qp_type == IB_QPT_GSI ||
+            qp->ib_qp.qp_type == IB_QPT_RAW_ETHERTYPE) &&
+            qp->qplib_qp.wqe_cnt == BNXT_RE_UD_QP_HW_STALL) {
+               int qp_attr_mask;
+               struct ib_qp_attr qp_attr;
+
+               qp_attr_mask = IB_QP_STATE;
+               qp_attr.qp_state = IB_QPS_RTS;
+               bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, qp_attr_mask, NULL);
+               qp->qplib_qp.wqe_cnt = 0;
+       }
+}
+
 static int bnxt_re_post_send_shadow_qp(struct bnxt_re_dev *rdev,
                                       struct bnxt_re_qp *qp,
                                struct ib_send_wr *wr)
@@ -1928,6 +2137,7 @@ bad:
                wr = wr->next;
        }
        bnxt_qplib_post_send_db(&qp->qplib_qp);
+       bnxt_ud_qp_hw_stall_workaround(qp);
        spin_unlock_irqrestore(&qp->sq_lock, flags);
        return rc;
 }
@@ -2024,6 +2234,7 @@ bad:
                wr = wr->next;
        }
        bnxt_qplib_post_send_db(&qp->qplib_qp);
+       bnxt_ud_qp_hw_stall_workaround(qp);
        spin_unlock_irqrestore(&qp->sq_lock, flags);
 
        return rc;
@@ -2071,7 +2282,10 @@ int bnxt_re_post_recv(struct ib_qp *ib_qp, struct ib_recv_wr *wr,
        struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
        struct bnxt_qplib_swqe wqe;
        int rc = 0, payload_sz = 0;
+       unsigned long flags;
+       u32 count = 0;
 
+       spin_lock_irqsave(&qp->rq_lock, flags);
        while (wr) {
                /* House keeping */
                memset(&wqe, 0, sizeof(wqe));
@@ -2100,9 +2314,21 @@ int bnxt_re_post_recv(struct ib_qp *ib_qp, struct ib_recv_wr *wr,
                        *bad_wr = wr;
                        break;
                }
+
+               /* Ring DB if the RQEs posted reaches a threshold value */
+               if (++count >= BNXT_RE_RQ_WQE_THRESHOLD) {
+                       bnxt_qplib_post_recv_db(&qp->qplib_qp);
+                       count = 0;
+               }
+
                wr = wr->next;
        }
-       bnxt_qplib_post_recv_db(&qp->qplib_qp);
+
+       if (count)
+               bnxt_qplib_post_recv_db(&qp->qplib_qp);
+
+       spin_unlock_irqrestore(&qp->rq_lock, flags);
+
        return rc;
 }
 
@@ -2643,12 +2869,36 @@ static void bnxt_re_process_res_ud_wc(struct ib_wc *wc,
                wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
 }
 
+static int send_phantom_wqe(struct bnxt_re_qp *qp)
+{
+       struct bnxt_qplib_qp *lib_qp = &qp->qplib_qp;
+       unsigned long flags;
+       int rc = 0;
+
+       spin_lock_irqsave(&qp->sq_lock, flags);
+
+       rc = bnxt_re_bind_fence_mw(lib_qp);
+       if (!rc) {
+               lib_qp->sq.phantom_wqe_cnt++;
+               dev_dbg(&lib_qp->sq.hwq.pdev->dev,
+                       "qp %#x sq->prod %#x sw_prod %#x phantom_wqe_cnt %d\n",
+                       lib_qp->id, lib_qp->sq.hwq.prod,
+                       HWQ_CMP(lib_qp->sq.hwq.prod, &lib_qp->sq.hwq),
+                       lib_qp->sq.phantom_wqe_cnt);
+       }
+
+       spin_unlock_irqrestore(&qp->sq_lock, flags);
+       return rc;
+}
+
 int bnxt_re_poll_cq(struct ib_cq *ib_cq, int num_entries, struct ib_wc *wc)
 {
        struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
        struct bnxt_re_qp *qp;
        struct bnxt_qplib_cqe *cqe;
        int i, ncqe, budget;
+       struct bnxt_qplib_q *sq;
+       struct bnxt_qplib_qp *lib_qp;
        u32 tbl_idx;
        struct bnxt_re_sqp_entries *sqp_entry = NULL;
        unsigned long flags;
@@ -2661,7 +2911,21 @@ int bnxt_re_poll_cq(struct ib_cq *ib_cq, int num_entries, struct ib_wc *wc)
        }
        cqe = &cq->cql[0];
        while (budget) {
-               ncqe = bnxt_qplib_poll_cq(&cq->qplib_cq, cqe, budget);
+               lib_qp = NULL;
+               ncqe = bnxt_qplib_poll_cq(&cq->qplib_cq, cqe, budget, &lib_qp);
+               if (lib_qp) {
+                       sq = &lib_qp->sq;
+                       if (sq->send_phantom) {
+                               qp = container_of(lib_qp,
+                                                 struct bnxt_re_qp, qplib_qp);
+                               if (send_phantom_wqe(qp) == -ENOMEM)
+                                       dev_err(rdev_to_dev(cq->rdev),
+                                               "Phantom failed! Scheduled to send again\n");
+                               else
+                                       sq->send_phantom = false;
+                       }
+               }
+
                if (!ncqe)
                        break;
 
@@ -2822,6 +3086,12 @@ int bnxt_re_dereg_mr(struct ib_mr *ib_mr)
        struct bnxt_re_dev *rdev = mr->rdev;
        int rc;
 
+       rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
+       if (rc) {
+               dev_err(rdev_to_dev(rdev), "Dereg MR failed: %#x\n", rc);
+               return rc;
+       }
+
        if (mr->npages && mr->pages) {
                rc = bnxt_qplib_free_fast_reg_page_list(&rdev->qplib_res,
                                                        &mr->qplib_frpl);
@@ -2829,8 +3099,6 @@ int bnxt_re_dereg_mr(struct ib_mr *ib_mr)
                mr->npages = 0;
                mr->pages = NULL;
        }
-       rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
-
        if (!IS_ERR_OR_NULL(mr->ib_umem))
                ib_umem_release(mr->ib_umem);
 
@@ -2914,97 +3182,52 @@ fail:
        return ERR_PTR(rc);
 }
 
-/* Fast Memory Regions */
-struct ib_fmr *bnxt_re_alloc_fmr(struct ib_pd *ib_pd, int mr_access_flags,
-                                struct ib_fmr_attr *fmr_attr)
+struct ib_mw *bnxt_re_alloc_mw(struct ib_pd *ib_pd, enum ib_mw_type type,
+                              struct ib_udata *udata)
 {
        struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
        struct bnxt_re_dev *rdev = pd->rdev;
-       struct bnxt_re_fmr *fmr;
+       struct bnxt_re_mw *mw;
        int rc;
 
-       if (fmr_attr->max_pages > MAX_PBL_LVL_2_PGS ||
-           fmr_attr->max_maps > rdev->dev_attr.max_map_per_fmr) {
-               dev_err(rdev_to_dev(rdev), "Allocate FMR exceeded Max limit");
+       mw = kzalloc(sizeof(*mw), GFP_KERNEL);
+       if (!mw)
                return ERR_PTR(-ENOMEM);
-       }
-       fmr = kzalloc(sizeof(*fmr), GFP_KERNEL);
-       if (!fmr)
-               return ERR_PTR(-ENOMEM);
-
-       fmr->rdev = rdev;
-       fmr->qplib_fmr.pd = &pd->qplib_pd;
-       fmr->qplib_fmr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
+       mw->rdev = rdev;
+       mw->qplib_mw.pd = &pd->qplib_pd;
 
-       rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &fmr->qplib_fmr);
-       if (rc)
+       mw->qplib_mw.type = (type == IB_MW_TYPE_1 ?
+                              CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE1 :
+                              CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE2B);
+       rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mw->qplib_mw);
+       if (rc) {
+               dev_err(rdev_to_dev(rdev), "Allocate MW failed!");
                goto fail;
+       }
+       mw->ib_mw.rkey = mw->qplib_mw.rkey;
 
-       fmr->qplib_fmr.flags = __from_ib_access_flags(mr_access_flags);
-       fmr->ib_fmr.lkey = fmr->qplib_fmr.lkey;
-       fmr->ib_fmr.rkey = fmr->ib_fmr.lkey;
+       atomic_inc(&rdev->mw_count);
+       return &mw->ib_mw;
 
-       atomic_inc(&rdev->mr_count);
-       return &fmr->ib_fmr;
 fail:
-       kfree(fmr);
+       kfree(mw);
        return ERR_PTR(rc);
 }
 
-int bnxt_re_map_phys_fmr(struct ib_fmr *ib_fmr, u64 *page_list, int list_len,
-                        u64 iova)
+int bnxt_re_dealloc_mw(struct ib_mw *ib_mw)
 {
-       struct bnxt_re_fmr *fmr = container_of(ib_fmr, struct bnxt_re_fmr,
-                                            ib_fmr);
-       struct bnxt_re_dev *rdev = fmr->rdev;
+       struct bnxt_re_mw *mw = container_of(ib_mw, struct bnxt_re_mw, ib_mw);
+       struct bnxt_re_dev *rdev = mw->rdev;
        int rc;
 
-       fmr->qplib_fmr.va = iova;
-       fmr->qplib_fmr.total_size = list_len * PAGE_SIZE;
-
-       rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &fmr->qplib_fmr, page_list,
-                              list_len, true);
-       if (rc)
-               dev_err(rdev_to_dev(rdev), "Failed to map FMR for lkey = 0x%x!",
-                       fmr->ib_fmr.lkey);
-       return rc;
-}
-
-int bnxt_re_unmap_fmr(struct list_head *fmr_list)
-{
-       struct bnxt_re_dev *rdev;
-       struct bnxt_re_fmr *fmr;
-       struct ib_fmr *ib_fmr;
-       int rc = 0;
-
-       /* Validate each FMRs inside the fmr_list */
-       list_for_each_entry(ib_fmr, fmr_list, list) {
-               fmr = container_of(ib_fmr, struct bnxt_re_fmr, ib_fmr);
-               rdev = fmr->rdev;
-
-               if (rdev) {
-                       rc = bnxt_qplib_dereg_mrw(&rdev->qplib_res,
-                                                 &fmr->qplib_fmr, true);
-                       if (rc)
-                               break;
-               }
+       rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mw->qplib_mw);
+       if (rc) {
+               dev_err(rdev_to_dev(rdev), "Free MW failed: %#x\n", rc);
+               return rc;
        }
-       return rc;
-}
-
-int bnxt_re_dealloc_fmr(struct ib_fmr *ib_fmr)
-{
-       struct bnxt_re_fmr *fmr = container_of(ib_fmr, struct bnxt_re_fmr,
-                                              ib_fmr);
-       struct bnxt_re_dev *rdev = fmr->rdev;
-       int rc;
 
-       rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &fmr->qplib_fmr);
-       if (rc)
-               dev_err(rdev_to_dev(rdev), "Failed to free FMR");
-
-       kfree(fmr);
-       atomic_dec(&rdev->mr_count);
+       kfree(mw);
+       atomic_dec(&rdev->mw_count);
        return rc;
 }