IB/mlx4: Add support for WQ related verbs
authorGuy Levi <guyle@mellanox.com>
Tue, 4 Jul 2017 13:24:24 +0000 (16:24 +0300)
committerDoug Ledford <dledford@redhat.com>
Mon, 24 Jul 2017 14:45:01 +0000 (10:45 -0400)
Support create/modify/destroy WQ related verbs.

The base IB object to enable RSS functionality is a WQ (i.e. ib_wq).
This patch implements the related WQ verbs as of create, modify and
destroy.

In downstream patches the WQ will be used as part of an indirection
table (i.e. ib_rwq_ind_table) to enable RSS QP creation.

Notes:
ConnectX-3 hardware requires consecutive WQNs list as receive descriptor
queues for the RSS QP. Hence, the driver manages consecutive ranges lists
per context which the user must respect.
Destroying the WQ does not return its WQN back to its range for
reusing. However, destroying all WQs from the same range releases the
range and in turn releases its WQNs for reusing.

Since the WQ object is not a natural object in the hardware, the driver
implements the WQ by the hardware QP.

As such, the WQ inherits its port from its RSS QP parent upon its
RST->INIT transition and by that time its state is applied to the
hardware.

Signed-off-by: Guy Levi <guyle@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/hw/mlx4/main.c
drivers/infiniband/hw/mlx4/mlx4_ib.h
drivers/infiniband/hw/mlx4/qp.c
include/uapi/rdma/mlx4-abi.h

index 0944e224c0dfeb2e3d34dc1c34be751986205c01..7c6f929ebd3ebcfeff3196695188304eb7a993a1 100644 (file)
@@ -81,6 +81,8 @@ static const char mlx4_ib_version[] =
        DRV_VERSION "\n";
 
 static void do_slave_init(struct mlx4_ib_dev *ibdev, int slave, int do_init);
+static enum rdma_link_layer mlx4_ib_port_link_layer(struct ib_device *device,
+                                                   u8 port_num);
 
 static struct workqueue_struct *wq;
 
@@ -552,6 +554,11 @@ static int mlx4_ib_query_device(struct ib_device *ibdev,
        props->timestamp_mask = 0xFFFFFFFFFFFFULL;
        props->max_ah = INT_MAX;
 
+       if ((dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS) &&
+           (mlx4_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET ||
+            mlx4_ib_port_link_layer(ibdev, 2) == IB_LINK_LAYER_ETHERNET))
+               props->max_wq_type_rq = props->max_qp;
+
        if (!mlx4_is_slave(dev->dev))
                err = mlx4_get_internal_clock_params(dev->dev, &clock_params);
 
@@ -1076,6 +1083,9 @@ static struct ib_ucontext *mlx4_ib_alloc_ucontext(struct ib_device *ibdev,
        INIT_LIST_HEAD(&context->db_page_list);
        mutex_init(&context->db_page_mutex);
 
+       INIT_LIST_HEAD(&context->wqn_ranges_list);
+       mutex_init(&context->wqn_ranges_mutex);
+
        if (ibdev->uverbs_abi_ver == MLX4_IB_UVERBS_NO_DEV_CAPS_ABI_VERSION)
                err = ib_copy_to_udata(udata, &resp_v3, sizeof(resp_v3));
        else
@@ -2720,6 +2730,20 @@ static void *mlx4_ib_add(struct mlx4_dev *dev)
        ibdev->ib_dev.get_dev_fw_str    = get_fw_ver_str;
        ibdev->ib_dev.disassociate_ucontext = mlx4_ib_disassociate_ucontext;
 
+       if ((dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS) &&
+           ((mlx4_ib_port_link_layer(&ibdev->ib_dev, 1) ==
+           IB_LINK_LAYER_ETHERNET) ||
+           (mlx4_ib_port_link_layer(&ibdev->ib_dev, 2) ==
+           IB_LINK_LAYER_ETHERNET))) {
+               ibdev->ib_dev.create_wq         = mlx4_ib_create_wq;
+               ibdev->ib_dev.modify_wq         = mlx4_ib_modify_wq;
+               ibdev->ib_dev.destroy_wq        = mlx4_ib_destroy_wq;
+               ibdev->ib_dev.uverbs_ex_cmd_mask |=
+                       (1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
+                       (1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
+                       (1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ);
+       }
+
        if (!mlx4_is_slave(ibdev->dev)) {
                ibdev->ib_dev.alloc_fmr         = mlx4_ib_fmr_alloc;
                ibdev->ib_dev.map_phys_fmr      = mlx4_ib_map_phys_fmr;
index a6337f3161cf723f9a58fa1cc3f3ee279aa5636b..ba4e78c064d2e37253f5ccd3afb17c38f2238877 100644 (file)
@@ -88,6 +88,8 @@ struct mlx4_ib_ucontext {
        struct list_head        db_page_list;
        struct mutex            db_page_mutex;
        struct mlx4_ib_vma_private_data hw_bar_info[HW_BAR_COUNT];
+       struct list_head        wqn_ranges_list;
+       struct mutex            wqn_ranges_mutex; /* protect wqn_ranges_list */
 };
 
 struct mlx4_ib_pd {
@@ -289,8 +291,19 @@ struct mlx4_roce_smac_vlan_info {
        int update_vid;
 };
 
+struct mlx4_wqn_range {
+       int                     base_wqn;
+       int                     size;
+       int                     refcount;
+       bool                    dirty;
+       struct list_head        list;
+};
+
 struct mlx4_ib_qp {
-       struct ib_qp            ibqp;
+       union {
+               struct ib_qp    ibqp;
+               struct ib_wq    ibwq;
+       };
        struct mlx4_qp          mqp;
        struct mlx4_buf         buf;
 
@@ -329,6 +342,9 @@ struct mlx4_ib_qp {
        struct list_head        cq_recv_list;
        struct list_head        cq_send_list;
        struct counter_index    *counter_index;
+       struct mlx4_wqn_range   *wqn_range;
+       /* Number of RSS QP parents that uses this WQ */
+       u32                     rss_usecnt;
 };
 
 struct mlx4_ib_srq {
@@ -893,4 +909,11 @@ void mlx4_sched_ib_sl2vl_update_work(struct mlx4_ib_dev *ibdev,
 
 void mlx4_ib_sl2vl_update(struct mlx4_ib_dev *mdev, int port);
 
+struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
+                               struct ib_wq_init_attr *init_attr,
+                               struct ib_udata *udata);
+int mlx4_ib_destroy_wq(struct ib_wq *wq);
+int mlx4_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
+                     u32 wq_attr_mask, struct ib_udata *udata);
+
 #endif /* MLX4_IB_H */
index 247b9132e9de101e3bb08c3266bda1b2a6da1eb0..65e4ec5493685e5045cbbd92ec0892435fecd34c 100644 (file)
@@ -116,6 +116,11 @@ static const __be32 mlx4_ib_opcode[] = {
        [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]     = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
 };
 
+enum mlx4_ib_source_type {
+       MLX4_IB_QP_SRC  = 0,
+       MLX4_IB_RWQ_SRC = 1,
+};
+
 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
 {
        return container_of(mqp, struct mlx4_ib_sqp, qp);
@@ -330,6 +335,12 @@ static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
        }
 }
 
+static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type)
+{
+       pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n",
+                           type, qp->qpn);
+}
+
 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
 {
        /*
@@ -639,7 +650,91 @@ static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
        qp->counter_index = NULL;
 }
 
+/*
+ * This function allocates a WQN from a range which is consecutive and aligned
+ * to its size. In case the range is full, then it creates a new range and
+ * allocates WQN from it. The new range will be used for following allocations.
+ */
+static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
+                            struct mlx4_ib_qp *qp, int range_size, int *wqn)
+{
+       struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
+       struct mlx4_wqn_range *range;
+       int err = 0;
+
+       mutex_lock(&context->wqn_ranges_mutex);
+
+       range = list_first_entry_or_null(&context->wqn_ranges_list,
+                                        struct mlx4_wqn_range, list);
+
+       if (!range || (range->refcount == range->size) || range->dirty) {
+               range = kzalloc(sizeof(*range), GFP_KERNEL);
+               if (!range) {
+                       err = -ENOMEM;
+                       goto out;
+               }
+
+               err = mlx4_qp_reserve_range(dev->dev, range_size,
+                                           range_size, &range->base_wqn, 0,
+                                           qp->mqp.usage);
+               if (err) {
+                       kfree(range);
+                       goto out;
+               }
+
+               range->size = range_size;
+               list_add(&range->list, &context->wqn_ranges_list);
+       } else if (range_size != 1) {
+               /*
+                * Requesting a new range (>1) when last range is still open, is
+                * not valid.
+                */
+               err = -EINVAL;
+               goto out;
+       }
+
+       qp->wqn_range = range;
+
+       *wqn = range->base_wqn + range->refcount;
+
+       range->refcount++;
+
+out:
+       mutex_unlock(&context->wqn_ranges_mutex);
+
+       return err;
+}
+
+static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
+                               struct mlx4_ib_qp *qp, bool dirty_release)
+{
+       struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
+       struct mlx4_wqn_range *range;
+
+       mutex_lock(&context->wqn_ranges_mutex);
+
+       range = qp->wqn_range;
+
+       range->refcount--;
+       if (!range->refcount) {
+               mlx4_qp_release_range(dev->dev, range->base_wqn,
+                                     range->size);
+               list_del(&range->list);
+               kfree(range);
+       } else if (dirty_release) {
+       /*
+        * A range which one of its WQNs is destroyed, won't be able to be
+        * reused for further WQN allocations.
+        * The next created WQ will allocate a new range.
+        */
+               range->dirty = 1;
+       }
+
+       mutex_unlock(&context->wqn_ranges_mutex);
+}
+
 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
+                           enum mlx4_ib_source_type src,
                            struct ib_qp_init_attr *init_attr,
                            struct ib_udata *udata, int sqpn,
                            struct mlx4_ib_qp **caller_qp)
@@ -652,6 +747,7 @@ static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
        enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
        struct mlx4_ib_cq *mcq;
        unsigned long flags;
+       int range_size = 0;
 
        /* When tunneling special qps, we use a plain UD qp */
        if (sqpn) {
@@ -728,27 +824,69 @@ static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
 
 
        if (pd->uobject) {
-               struct mlx4_ib_create_qp ucmd;
+               union {
+                       struct mlx4_ib_create_qp qp;
+                       struct mlx4_ib_create_wq wq;
+               } ucmd;
+               size_t copy_len;
+
+               copy_len = (src == MLX4_IB_QP_SRC) ?
+                          sizeof(struct mlx4_ib_create_qp) :
+                          min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
 
-               if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
+               if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
                        err = -EFAULT;
                        goto err;
                }
 
-               err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
-                                 qp_has_rq(init_attr), qp, ucmd.inl_recv_sz);
-               if (err)
-                       goto err;
+               if (src == MLX4_IB_RWQ_SRC) {
+                       if (ucmd.wq.comp_mask || ucmd.wq.reserved1 ||
+                           ucmd.wq.reserved[0] || ucmd.wq.reserved[1] ||
+                           ucmd.wq.reserved[2]) {
+                               pr_debug("user command isn't supported\n");
+                               err = -EOPNOTSUPP;
+                               goto err;
+                       }
 
-               qp->inl_recv_sz = ucmd.inl_recv_sz;
-               qp->sq_no_prefetch = ucmd.sq_no_prefetch;
+                       if (ucmd.wq.log_range_size >
+                           ilog2(dev->dev->caps.max_rss_tbl_sz)) {
+                               pr_debug("WQN range size must be equal or smaller than %d\n",
+                                        dev->dev->caps.max_rss_tbl_sz);
+                               err = -EOPNOTSUPP;
+                               goto err;
+                       }
+                       range_size = 1 << ucmd.wq.log_range_size;
+               } else {
+                       qp->inl_recv_sz = ucmd.qp.inl_recv_sz;
+               }
 
-               err = set_user_sq_size(dev, qp, &ucmd);
+               err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
+                                 qp_has_rq(init_attr), qp, qp->inl_recv_sz);
                if (err)
                        goto err;
 
-               qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
-                                      qp->buf_size, 0, 0);
+               if (src == MLX4_IB_QP_SRC) {
+                       qp->sq_no_prefetch = ucmd.qp.sq_no_prefetch;
+
+                       err = set_user_sq_size(dev, qp,
+                                              (struct mlx4_ib_create_qp *)
+                                              &ucmd);
+                       if (err)
+                               goto err;
+               } else {
+                       qp->sq_no_prefetch = 1;
+                       qp->sq.wqe_cnt = 1;
+                       qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
+                       /* Allocated buffer expects to have at least that SQ
+                        * size.
+                        */
+                       qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
+                               (qp->sq.wqe_cnt << qp->sq.wqe_shift);
+               }
+
+               qp->umem = ib_umem_get(pd->uobject->context,
+                               (src == MLX4_IB_QP_SRC) ? ucmd.qp.buf_addr :
+                               ucmd.wq.buf_addr, qp->buf_size, 0, 0);
                if (IS_ERR(qp->umem)) {
                        err = PTR_ERR(qp->umem);
                        goto err;
@@ -765,7 +903,8 @@ static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
 
                if (qp_has_rq(init_attr)) {
                        err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
-                                                 ucmd.db_addr, &qp->db);
+                               (src == MLX4_IB_QP_SRC) ? ucmd.qp.db_addr :
+                               ucmd.wq.db_addr, &qp->db);
                        if (err)
                                goto err_mtt;
                }
@@ -853,6 +992,11 @@ static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
                                goto err_wrid;
                        }
                }
+       } else if (src == MLX4_IB_RWQ_SRC) {
+               err = mlx4_ib_alloc_wqn(to_mucontext(pd->uobject->context), qp,
+                                       range_size, &qpn);
+               if (err)
+                       goto err_wrid;
        } else {
                /* Raw packet QPNs may not have bits 6,7 set in their qp_num;
                 * otherwise, the WQE BlueFlame setup flow wrongly causes
@@ -891,7 +1035,9 @@ static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
         */
        qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
 
-       qp->mqp.event = mlx4_ib_qp_event;
+       qp->mqp.event = (src == MLX4_IB_QP_SRC) ? mlx4_ib_qp_event :
+                                                 mlx4_ib_wq_event;
+
        if (!*caller_qp)
                *caller_qp = qp;
 
@@ -918,6 +1064,9 @@ err_qpn:
        if (!sqpn) {
                if (qp->flags & MLX4_IB_QP_NETIF)
                        mlx4_ib_steer_qp_free(dev, qpn, 1);
+               else if (src == MLX4_IB_RWQ_SRC)
+                       mlx4_ib_release_wqn(to_mucontext(pd->uobject->context),
+                                           qp, 0);
                else
                        mlx4_qp_release_range(dev->dev, qpn, 1);
        }
@@ -1016,7 +1165,7 @@ static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
                return to_mpd(qp->ibqp.pd);
 }
 
-static void get_cqs(struct mlx4_ib_qp *qp,
+static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
                    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
 {
        switch (qp->ibqp.qp_type) {
@@ -1029,14 +1178,16 @@ static void get_cqs(struct mlx4_ib_qp *qp,
                *recv_cq = *send_cq;
                break;
        default:
-               *send_cq = to_mcq(qp->ibqp.send_cq);
-               *recv_cq = to_mcq(qp->ibqp.recv_cq);
+               *recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
+                                                    to_mcq(qp->ibwq.cq);
+               *send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
+                                                    *recv_cq;
                break;
        }
 }
 
 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
-                             int is_user)
+                             enum mlx4_ib_source_type src, int is_user)
 {
        struct mlx4_ib_cq *send_cq, *recv_cq;
        unsigned long flags;
@@ -1069,7 +1220,7 @@ static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
                }
        }
 
-       get_cqs(qp, &send_cq, &recv_cq);
+       get_cqs(qp, src, &send_cq, &recv_cq);
 
        spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
        mlx4_ib_lock_cqs(send_cq, recv_cq);
@@ -1095,6 +1246,9 @@ static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
        if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
                if (qp->flags & MLX4_IB_QP_NETIF)
                        mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
+               else if (src == MLX4_IB_RWQ_SRC)
+                       mlx4_ib_release_wqn(to_mucontext(
+                                           qp->ibwq.uobject->context), qp, 1);
                else
                        mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
        }
@@ -1102,9 +1256,12 @@ static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
        mlx4_mtt_cleanup(dev->dev, &qp->mtt);
 
        if (is_user) {
-               if (qp->rq.wqe_cnt)
-                       mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
-                                             &qp->db);
+               if (qp->rq.wqe_cnt) {
+                       struct mlx4_ib_ucontext *mcontext = !src ?
+                               to_mucontext(qp->ibqp.uobject->context) :
+                               to_mucontext(qp->ibwq.uobject->context);
+                       mlx4_ib_db_unmap_user(mcontext, &qp->db);
+               }
                ib_umem_release(qp->umem);
        } else {
                kvfree(qp->sq.wrid);
@@ -1200,8 +1357,8 @@ static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
                /* fall through */
        case IB_QPT_UD:
        {
-               err = create_qp_common(to_mdev(pd->device), pd, init_attr,
-                                      udata, 0, &qp);
+               err = create_qp_common(to_mdev(pd->device), pd, MLX4_IB_QP_SRC,
+                                      init_attr, udata, 0, &qp);
                if (err) {
                        kfree(qp);
                        return ERR_PTR(err);
@@ -1231,8 +1388,8 @@ static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
                        sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
                }
 
-               err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
-                                      sqpn, &qp);
+               err = create_qp_common(to_mdev(pd->device), pd, MLX4_IB_QP_SRC,
+                                      init_attr, udata, sqpn, &qp);
                if (err)
                        return ERR_PTR(err);
 
@@ -1303,7 +1460,7 @@ static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
                mlx4_ib_free_qp_counter(dev, mqp);
 
        pd = get_pd(mqp);
-       destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
+       destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, !!pd->ibpd.uobject);
 
        if (is_sqp(dev, mqp))
                kfree(to_msqp(mqp));
@@ -1626,12 +1783,15 @@ static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
        }
 }
 
-static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
+static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
                               const struct ib_qp_attr *attr, int attr_mask,
                               enum ib_qp_state cur_state, enum ib_qp_state new_state)
 {
-       struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
-       struct mlx4_ib_qp *qp = to_mqp(ibqp);
+       struct ib_uobject *ibuobject;
+       struct ib_srq  *ibsrq;
+       enum ib_qp_type qp_type;
+       struct mlx4_ib_dev *dev;
+       struct mlx4_ib_qp *qp;
        struct mlx4_ib_pd *pd;
        struct mlx4_ib_cq *send_cq, *recv_cq;
        struct mlx4_qp_context *context;
@@ -1641,6 +1801,28 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
        int err = -EINVAL;
        int counter_index;
 
+       if (src_type == MLX4_IB_RWQ_SRC) {
+               struct ib_wq *ibwq;
+
+               ibwq       = (struct ib_wq *)src;
+               ibuobject  = ibwq->uobject;
+               ibsrq      = NULL;
+               qp_type    = IB_QPT_RAW_PACKET;
+               qp         = to_mqp((struct ib_qp *)ibwq);
+               dev        = to_mdev(ibwq->device);
+               pd         = to_mpd(ibwq->pd);
+       } else {
+               struct ib_qp *ibqp;
+
+               ibqp       = (struct ib_qp *)src;
+               ibuobject  = ibqp->uobject;
+               ibsrq      = ibqp->srq;
+               qp_type    = ibqp->qp_type;
+               qp         = to_mqp(ibqp);
+               dev        = to_mdev(ibqp->device);
+               pd         = get_pd(qp);
+       }
+
        /* APM is not supported under RoCE */
        if (attr_mask & IB_QP_ALT_PATH &&
            rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
@@ -1674,11 +1856,11 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
        if (qp->inl_recv_sz)
                context->param3 |= cpu_to_be32(1 << 25);
 
-       if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
+       if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
                context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
-       else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
+       else if (qp_type == IB_QPT_RAW_PACKET)
                context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
-       else if (ibqp->qp_type == IB_QPT_UD) {
+       else if (qp_type == IB_QPT_UD) {
                if (qp->flags & MLX4_IB_QP_LSO)
                        context->mtu_msgmax = (IB_MTU_4096 << 5) |
                                              ilog2(dev->dev->caps.max_gso_sz);
@@ -1708,14 +1890,15 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
        if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
                context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
                context->xrcd = cpu_to_be32((u32) qp->xrcdn);
-               if (ibqp->qp_type == IB_QPT_RAW_PACKET)
+               if (qp_type == IB_QPT_RAW_PACKET)
                        context->param3 |= cpu_to_be32(1 << 30);
        }
 
-       if (qp->ibqp.uobject)
+       if (ibuobject)
                context->usr_page = cpu_to_be32(
                        mlx4_to_hw_uar_index(dev->dev,
-                                            to_mucontext(ibqp->uobject->context)->uar.index));
+                                            to_mucontext(ibuobject->context)
+                                            ->uar.index));
        else
                context->usr_page = cpu_to_be32(
                        mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
@@ -1759,7 +1942,7 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
                        steer_qp = 1;
                }
 
-               if (ibqp->qp_type == IB_QPT_GSI) {
+               if (qp_type == IB_QPT_GSI) {
                        enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
                                IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
                        u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
@@ -1776,7 +1959,7 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
        }
 
        if (attr_mask & IB_QP_AV) {
-               u8 port_num = mlx4_is_bonded(to_mdev(ibqp->device)->dev) ? 1 :
+               u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
                        attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
                union ib_gid gid;
                struct ib_gid_attr gid_attr = {.gid_type = IB_GID_TYPE_IB};
@@ -1791,7 +1974,7 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
                        int index =
                                rdma_ah_read_grh(&attr->ah_attr)->sgid_index;
 
-                       status = ib_get_cached_gid(ibqp->device, port_num,
+                       status = ib_get_cached_gid(&dev->ib_dev, port_num,
                                                   index, &gid, &gid_attr);
                        if (!status && !memcmp(&gid, &zgid, sizeof(gid)))
                                status = -ENOENT;
@@ -1848,15 +2031,14 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
                optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
        }
 
-       pd = get_pd(qp);
-       get_cqs(qp, &send_cq, &recv_cq);
+       get_cqs(qp, src_type, &send_cq, &recv_cq);
        context->pd       = cpu_to_be32(pd->pdn);
        context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
        context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
        context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
 
        /* Set "fast registration enabled" for all kernel QPs */
-       if (!qp->ibqp.uobject)
+       if (!ibuobject)
                context->params1 |= cpu_to_be32(1 << 11);
 
        if (attr_mask & IB_QP_RNR_RETRY) {
@@ -1891,7 +2073,7 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
                optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
        }
 
-       if (ibqp->srq)
+       if (ibsrq)
                context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
 
        if (attr_mask & IB_QP_MIN_RNR_TIMER) {
@@ -1922,17 +2104,19 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
                optpar |= MLX4_QP_OPTPAR_Q_KEY;
        }
 
-       if (ibqp->srq)
-               context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
+       if (ibsrq)
+               context->srqn = cpu_to_be32(1 << 24 |
+                                           to_msrq(ibsrq)->msrq.srqn);
 
-       if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
+       if (qp->rq.wqe_cnt &&
+           cur_state == IB_QPS_RESET &&
+           new_state == IB_QPS_INIT)
                context->db_rec_addr = cpu_to_be64(qp->db.dma);
 
        if (cur_state == IB_QPS_INIT &&
            new_state == IB_QPS_RTR  &&
-           (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
-            ibqp->qp_type == IB_QPT_UD ||
-            ibqp->qp_type == IB_QPT_RAW_PACKET)) {
+           (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
+            qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
                context->pri_path.sched_queue = (qp->port - 1) << 6;
                if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
                    qp->mlx4_ib_qp_type &
@@ -1965,7 +2149,7 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
                }
        }
 
-       if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
+       if (qp_type == IB_QPT_RAW_PACKET) {
                context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
                                        MLX4_IB_LINK_TYPE_ETH;
                if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
@@ -1975,7 +2159,7 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
                }
        }
 
-       if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
+       if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
                int is_eth = rdma_port_get_link_layer(
                                &dev->ib_dev, qp->port) ==
                                IB_LINK_LAYER_ETHERNET;
@@ -1985,14 +2169,15 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
                }
        }
 
-
        if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD  &&
            attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
                sqd_event = 1;
        else
                sqd_event = 0;
 
-       if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
+       if (!ibuobject &&
+           cur_state == IB_QPS_RESET &&
+           new_state == IB_QPS_INIT)
                context->rlkey_roce_mode |= (1 << 4);
 
        /*
@@ -2001,7 +2186,9 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
         * headroom is stamped so that the hardware doesn't start
         * processing stale work requests.
         */
-       if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
+       if (!ibuobject &&
+           cur_state == IB_QPS_RESET &&
+           new_state == IB_QPS_INIT) {
                struct mlx4_wqe_ctrl_seg *ctrl;
                int i;
 
@@ -2058,9 +2245,9 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
         * entries and reinitialize the QP.
         */
        if (new_state == IB_QPS_RESET) {
-               if (!ibqp->uobject) {
+               if (!ibuobject) {
                        mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
-                                        ibqp->srq ? to_msrq(ibqp->srq) : NULL);
+                                        ibsrq ? to_msrq(ibsrq) : NULL);
                        if (send_cq != recv_cq)
                                mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
 
@@ -2265,7 +2452,8 @@ static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
                goto out;
        }
 
-       err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
+       err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
+                                 cur_state, new_state);
 
        if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
                attr->port_num = 1;
@@ -3550,3 +3738,208 @@ out:
        return err;
 }
 
+struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
+                               struct ib_wq_init_attr *init_attr,
+                               struct ib_udata *udata)
+{
+       struct mlx4_ib_dev *dev;
+       struct ib_qp_init_attr ib_qp_init_attr;
+       struct mlx4_ib_qp *qp;
+       struct mlx4_ib_create_wq ucmd;
+       int err, required_cmd_sz;
+
+       if (!(udata && pd->uobject))
+               return ERR_PTR(-EINVAL);
+
+       required_cmd_sz = offsetof(typeof(ucmd), reserved) +
+                         sizeof(ucmd.reserved);
+       if (udata->inlen < required_cmd_sz) {
+               pr_debug("invalid inlen\n");
+               return ERR_PTR(-EINVAL);
+       }
+
+       if (udata->inlen > sizeof(ucmd) &&
+           !ib_is_udata_cleared(udata, sizeof(ucmd),
+                                udata->inlen - sizeof(ucmd))) {
+               pr_debug("inlen is not supported\n");
+               return ERR_PTR(-EOPNOTSUPP);
+       }
+
+       if (udata->outlen)
+               return ERR_PTR(-EOPNOTSUPP);
+
+       dev = to_mdev(pd->device);
+
+       if (init_attr->wq_type != IB_WQT_RQ) {
+               pr_debug("unsupported wq type %d\n", init_attr->wq_type);
+               return ERR_PTR(-EOPNOTSUPP);
+       }
+
+       if (init_attr->create_flags) {
+               pr_debug("unsupported create_flags %u\n",
+                        init_attr->create_flags);
+               return ERR_PTR(-EOPNOTSUPP);
+       }
+
+       qp = kzalloc(sizeof(*qp), GFP_KERNEL);
+       if (!qp)
+               return ERR_PTR(-ENOMEM);
+
+       qp->pri.vid = 0xFFFF;
+       qp->alt.vid = 0xFFFF;
+
+       memset(&ib_qp_init_attr, 0, sizeof(ib_qp_init_attr));
+       ib_qp_init_attr.qp_context = init_attr->wq_context;
+       ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
+       ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
+       ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
+       ib_qp_init_attr.recv_cq = init_attr->cq;
+       ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
+
+       err = create_qp_common(dev, pd, MLX4_IB_RWQ_SRC, &ib_qp_init_attr,
+                              udata, 0, &qp);
+       if (err) {
+               kfree(qp);
+               return ERR_PTR(err);
+       }
+
+       qp->ibwq.event_handler = init_attr->event_handler;
+       qp->ibwq.wq_num = qp->mqp.qpn;
+       qp->ibwq.state = IB_WQS_RESET;
+
+       return &qp->ibwq;
+}
+
+static int ib_wq2qp_state(enum ib_wq_state state)
+{
+       switch (state) {
+       case IB_WQS_RESET:
+               return IB_QPS_RESET;
+       case IB_WQS_RDY:
+               return IB_QPS_RTR;
+       default:
+               return IB_QPS_ERR;
+       }
+}
+
+static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state)
+{
+       struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
+       enum ib_qp_state qp_cur_state;
+       enum ib_qp_state qp_new_state;
+       int attr_mask;
+       int err;
+
+       /* ib_qp.state represents the WQ HW state while ib_wq.state represents
+        * the WQ logic state.
+        */
+       qp_cur_state = qp->state;
+       qp_new_state = ib_wq2qp_state(new_state);
+
+       if (ib_wq2qp_state(new_state) == qp_cur_state)
+               return 0;
+
+       if (new_state == IB_WQS_RDY) {
+               struct ib_qp_attr attr = {};
+
+               attr.port_num = qp->port;
+               attr_mask = IB_QP_PORT;
+
+               err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
+                                         attr_mask, IB_QPS_RESET, IB_QPS_INIT);
+               if (err) {
+                       pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
+                                ibwq->wq_num);
+                       return err;
+               }
+
+               qp_cur_state = IB_QPS_INIT;
+       }
+
+       attr_mask = 0;
+       err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
+                                 qp_cur_state,  qp_new_state);
+
+       if (err && (qp_cur_state == IB_QPS_INIT)) {
+               qp_new_state = IB_QPS_RESET;
+               if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
+                                       attr_mask, IB_QPS_INIT, IB_QPS_RESET)) {
+                       pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
+                               ibwq->wq_num);
+                       qp_new_state = IB_QPS_INIT;
+               }
+       }
+
+       qp->state = qp_new_state;
+
+       return err;
+}
+
+int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
+                     u32 wq_attr_mask, struct ib_udata *udata)
+{
+       struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
+       struct mlx4_ib_modify_wq ucmd = {};
+       size_t required_cmd_sz;
+       enum ib_wq_state cur_state, new_state;
+       int err = 0;
+
+       required_cmd_sz = offsetof(typeof(ucmd), reserved) +
+                                  sizeof(ucmd.reserved);
+       if (udata->inlen < required_cmd_sz)
+               return -EINVAL;
+
+       if (udata->inlen > sizeof(ucmd) &&
+           !ib_is_udata_cleared(udata, sizeof(ucmd),
+                                udata->inlen - sizeof(ucmd)))
+               return -EOPNOTSUPP;
+
+       if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
+               return -EFAULT;
+
+       if (ucmd.comp_mask || ucmd.reserved)
+               return -EOPNOTSUPP;
+
+       if (wq_attr_mask & IB_WQ_FLAGS)
+               return -EOPNOTSUPP;
+
+       cur_state = wq_attr_mask & IB_WQ_CUR_STATE ? wq_attr->curr_wq_state :
+                                                    ibwq->state;
+       new_state = wq_attr_mask & IB_WQ_STATE ? wq_attr->wq_state : cur_state;
+
+       if (cur_state  < IB_WQS_RESET || cur_state  > IB_WQS_ERR ||
+           new_state < IB_WQS_RESET || new_state > IB_WQS_ERR)
+               return -EINVAL;
+
+       if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
+               return -EINVAL;
+
+       if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
+               return -EINVAL;
+
+       /* Can update HW state only if a RSS QP has already associated to this
+        * WQ, so we can apply its port on the WQ.
+        */
+       if (qp->rss_usecnt)
+               err = _mlx4_ib_modify_wq(ibwq, new_state);
+
+       if (!err)
+               ibwq->state = new_state;
+
+       return err;
+}
+
+int mlx4_ib_destroy_wq(struct ib_wq *ibwq)
+{
+       struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
+       struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
+
+       if (qp->counter_index)
+               mlx4_ib_free_qp_counter(dev, qp);
+
+       destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, 1);
+
+       kfree(qp);
+
+       return 0;
+}
index bf3bdba2f3262e2444828a2a19496dfc4a6f5d5a..c9702a5f0bda5ce561291e23a6618f6cb1106cdd 100644 (file)
@@ -105,4 +105,18 @@ struct mlx4_ib_create_qp {
        __u8    reserved;
 };
 
+struct mlx4_ib_create_wq {
+       __u64   buf_addr;
+       __u64   db_addr;
+       __u8    log_range_size;
+       __u8    reserved[3];
+       __u32   comp_mask;
+       __u32   reserved1;
+};
+
+struct mlx4_ib_modify_wq {
+       __u32   comp_mask;
+       __u32   reserved;
+};
+
 #endif /* MLX4_ABI_USER_H */