]> git.karo-electronics.de Git - linux-beck.git/commitdiff
mlx4_core: Add wrapper functions and comm channel and slave event support to EQs
authorJack Morgenstein <jackm@dev.mellanox.co.il>
Tue, 13 Dec 2011 04:13:58 +0000 (04:13 +0000)
committerDavid S. Miller <davem@davemloft.net>
Tue, 13 Dec 2011 18:56:06 +0000 (13:56 -0500)
Passing async events to slaves:
In SRIOV mode, each slave creates its own async EQ, but only the master can
register directly with the FW to receive async events.  Async events which
should be passed to slaves (such as a WQ_ACCESS_ERROR for a QP owned by a slave)
are generated at the slave by the master using the GEN_EQE FW command.

Wrapper functions: mlx4_MAP_EQ_wrapper
Only the master can map an EQ. The slave commands to map their EQs arrive
at the master via the comm channel.  The master then invokes the wrapper
function to do the work (and enter the resource in the tracking database).

New events: COMM_CHANNEL and FLR
The COMM_CHANNEL event arrives only at the master, and signals that
a slave has posted a command on the comm channel.
The FLR event is generated by the FW when a guest operating a VF
unexpectedly goes down.

Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/mellanox/mlx4/eq.c

index 9e5863dfa60aa6d80eac5f2cf524d3a895a18101..7416ef20c203686b16f6787da8ccd508fe58e1f4 100644 (file)
@@ -31,6 +31,7 @@
  * SOFTWARE.
  */
 
+#include <linux/init.h>
 #include <linux/interrupt.h>
 #include <linux/slab.h>
 #include <linux/export.h>
@@ -100,7 +101,9 @@ struct mlx4_eq_context {
                               (1ull << MLX4_EVENT_TYPE_SRQ_CATAS_ERROR)    | \
                               (1ull << MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE)    | \
                               (1ull << MLX4_EVENT_TYPE_SRQ_LIMIT)          | \
-                              (1ull << MLX4_EVENT_TYPE_CMD))
+                              (1ull << MLX4_EVENT_TYPE_CMD)                | \
+                              (1ull << MLX4_EVENT_TYPE_COMM_CHANNEL)       | \
+                              (1ull << MLX4_EVENT_TYPE_FLR_EVENT))
 
 static void eq_set_ci(struct mlx4_eq *eq, int req_not)
 {
@@ -123,13 +126,157 @@ static struct mlx4_eqe *next_eqe_sw(struct mlx4_eq *eq)
        return !!(eqe->owner & 0x80) ^ !!(eq->cons_index & eq->nent) ? NULL : eqe;
 }
 
+static struct mlx4_eqe *next_slave_event_eqe(struct mlx4_slave_event_eq *slave_eq)
+{
+       struct mlx4_eqe *eqe =
+               &slave_eq->event_eqe[slave_eq->cons & (SLAVE_EVENT_EQ_SIZE - 1)];
+       return (!!(eqe->owner & 0x80) ^
+               !!(slave_eq->cons & SLAVE_EVENT_EQ_SIZE)) ?
+               eqe : NULL;
+}
+
+/* dummies for now */
+void mlx4_delete_all_resources_for_slave(struct mlx4_dev *dev, int slave)
+{
+}
+
+int mlx4_get_slave_from_resource_id(struct mlx4_dev *dev,
+                                   enum mlx4_resource type,
+                                   int res_id, int *slave)
+{
+       return -ENOENT;
+}
+/*  end dummies */
+
+void mlx4_gen_slave_eqe(struct work_struct *work)
+{
+       struct mlx4_mfunc_master_ctx *master =
+               container_of(work, struct mlx4_mfunc_master_ctx,
+                            slave_event_work);
+       struct mlx4_mfunc *mfunc =
+               container_of(master, struct mlx4_mfunc, master);
+       struct mlx4_priv *priv = container_of(mfunc, struct mlx4_priv, mfunc);
+       struct mlx4_dev *dev = &priv->dev;
+       struct mlx4_slave_event_eq *slave_eq = &mfunc->master.slave_eq;
+       struct mlx4_eqe *eqe;
+       u8 slave;
+       int i;
+
+       for (eqe = next_slave_event_eqe(slave_eq); eqe;
+             eqe = next_slave_event_eqe(slave_eq)) {
+               slave = eqe->slave_id;
+
+               /* All active slaves need to receive the event */
+               if (slave == ALL_SLAVES) {
+                       for (i = 0; i < dev->num_slaves; i++) {
+                               if (i != dev->caps.function &&
+                                   master->slave_state[i].active)
+                                       if (mlx4_GEN_EQE(dev, i, eqe))
+                                               mlx4_warn(dev, "Failed to "
+                                                         " generate event "
+                                                         "for slave %d\n", i);
+                       }
+               } else {
+                       if (mlx4_GEN_EQE(dev, slave, eqe))
+                               mlx4_warn(dev, "Failed to generate event "
+                                              "for slave %d\n", slave);
+               }
+               ++slave_eq->cons;
+       }
+}
+
+
+static void slave_event(struct mlx4_dev *dev, u8 slave, struct mlx4_eqe *eqe)
+{
+       struct mlx4_priv *priv = mlx4_priv(dev);
+       struct mlx4_slave_event_eq *slave_eq = &priv->mfunc.master.slave_eq;
+       struct mlx4_eqe *s_eqe =
+               &slave_eq->event_eqe[slave_eq->prod & (SLAVE_EVENT_EQ_SIZE - 1)];
+
+       if ((!!(s_eqe->owner & 0x80)) ^
+           (!!(slave_eq->prod & SLAVE_EVENT_EQ_SIZE))) {
+               mlx4_warn(dev, "Master failed to generate an EQE for slave: %d. "
+                         "No free EQE on slave events queue\n", slave);
+               return;
+       }
+
+       memcpy(s_eqe, eqe, sizeof(struct mlx4_eqe) - 1);
+       s_eqe->slave_id = slave;
+       /* ensure all information is written before setting the ownersip bit */
+       wmb();
+       s_eqe->owner = !!(slave_eq->prod & SLAVE_EVENT_EQ_SIZE) ? 0x0 : 0x80;
+       ++slave_eq->prod;
+
+       queue_work(priv->mfunc.master.comm_wq,
+                  &priv->mfunc.master.slave_event_work);
+}
+
+static void mlx4_slave_event(struct mlx4_dev *dev, int slave,
+                            struct mlx4_eqe *eqe)
+{
+       struct mlx4_priv *priv = mlx4_priv(dev);
+       struct mlx4_slave_state *s_slave =
+               &priv->mfunc.master.slave_state[slave];
+
+       if (!s_slave->active) {
+               /*mlx4_warn(dev, "Trying to pass event to inactive slave\n");*/
+               return;
+       }
+
+       slave_event(dev, slave, eqe);
+}
+
+void mlx4_master_handle_slave_flr(struct work_struct *work)
+{
+       struct mlx4_mfunc_master_ctx *master =
+               container_of(work, struct mlx4_mfunc_master_ctx,
+                            slave_flr_event_work);
+       struct mlx4_mfunc *mfunc =
+               container_of(master, struct mlx4_mfunc, master);
+       struct mlx4_priv *priv =
+               container_of(mfunc, struct mlx4_priv, mfunc);
+       struct mlx4_dev *dev = &priv->dev;
+       struct mlx4_slave_state *slave_state = priv->mfunc.master.slave_state;
+       int i;
+       int err;
+
+       mlx4_dbg(dev, "mlx4_handle_slave_flr\n");
+
+       for (i = 0 ; i < dev->num_slaves; i++) {
+
+               if (MLX4_COMM_CMD_FLR == slave_state[i].last_cmd) {
+                       mlx4_dbg(dev, "mlx4_handle_slave_flr: "
+                                "clean slave: %d\n", i);
+
+                       mlx4_delete_all_resources_for_slave(dev, i);
+                       /*return the slave to running mode*/
+                       spin_lock(&priv->mfunc.master.slave_state_lock);
+                       slave_state[i].last_cmd = MLX4_COMM_CMD_RESET;
+                       slave_state[i].is_slave_going_down = 0;
+                       spin_unlock(&priv->mfunc.master.slave_state_lock);
+                       /*notify the FW:*/
+                       err = mlx4_cmd(dev, 0, i, 0, MLX4_CMD_INFORM_FLR_DONE,
+                                      MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
+                       if (err)
+                               mlx4_warn(dev, "Failed to notify FW on "
+                                         "FLR done (slave:%d)\n", i);
+               }
+       }
+}
+
 static int mlx4_eq_int(struct mlx4_dev *dev, struct mlx4_eq *eq)
 {
+       struct mlx4_priv *priv = mlx4_priv(dev);
        struct mlx4_eqe *eqe;
        int cqn;
        int eqes_found = 0;
        int set_ci = 0;
        int port;
+       int slave = 0;
+       int ret;
+       u32 flr_slave;
+       u8 update_slave_state;
+       int i;
 
        while ((eqe = next_eqe_sw(eq))) {
                /*
@@ -152,14 +299,68 @@ static int mlx4_eq_int(struct mlx4_dev *dev, struct mlx4_eq *eq)
                case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
                case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
                case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
-                       mlx4_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
-                                     eqe->type);
+                       mlx4_dbg(dev, "event %d arrived\n", eqe->type);
+                       if (mlx4_is_master(dev)) {
+                               /* forward only to slave owning the QP */
+                               ret = mlx4_get_slave_from_resource_id(dev,
+                                               RES_QP,
+                                               be32_to_cpu(eqe->event.qp.qpn)
+                                               & 0xffffff, &slave);
+                               if (ret && ret != -ENOENT) {
+                                       mlx4_dbg(dev, "QP event %02x(%02x) on "
+                                                "EQ %d at index %u: could "
+                                                "not get slave id (%d)\n",
+                                                eqe->type, eqe->subtype,
+                                                eq->eqn, eq->cons_index, ret);
+                                       break;
+                               }
+
+                               if (!ret && slave != dev->caps.function) {
+                                       mlx4_slave_event(dev, slave, eqe);
+                                       break;
+                               }
+
+                       }
+                       mlx4_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) &
+                                     0xffffff, eqe->type);
                        break;
 
                case MLX4_EVENT_TYPE_SRQ_LIMIT:
+                       mlx4_warn(dev, "%s: MLX4_EVENT_TYPE_SRQ_LIMIT\n",
+                                 __func__);
                case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR:
-                       mlx4_srq_event(dev, be32_to_cpu(eqe->event.srq.srqn) & 0xffffff,
-                                     eqe->type);
+                       if (mlx4_is_master(dev)) {
+                               /* forward only to slave owning the SRQ */
+                               ret = mlx4_get_slave_from_resource_id(dev,
+                                               RES_SRQ,
+                                               be32_to_cpu(eqe->event.srq.srqn)
+                                               & 0xffffff,
+                                               &slave);
+                               if (ret && ret != -ENOENT) {
+                                       mlx4_warn(dev, "SRQ event %02x(%02x) "
+                                                 "on EQ %d at index %u: could"
+                                                 " not get slave id (%d)\n",
+                                                 eqe->type, eqe->subtype,
+                                                 eq->eqn, eq->cons_index, ret);
+                                       break;
+                               }
+                               mlx4_warn(dev, "%s: slave:%d, srq_no:0x%x,"
+                                         " event: %02x(%02x)\n", __func__,
+                                         slave,
+                                         be32_to_cpu(eqe->event.srq.srqn),
+                                         eqe->type, eqe->subtype);
+
+                               if (!ret && slave != dev->caps.function) {
+                                       mlx4_warn(dev, "%s: sending event "
+                                                 "%02x(%02x) to slave:%d\n",
+                                                  __func__, eqe->type,
+                                                 eqe->subtype, slave);
+                                       mlx4_slave_event(dev, slave, eqe);
+                                       break;
+                               }
+                       }
+                       mlx4_srq_event(dev, be32_to_cpu(eqe->event.srq.srqn) &
+                                      0xffffff, eqe->type);
                        break;
 
                case MLX4_EVENT_TYPE_CMD:
@@ -172,13 +373,35 @@ static int mlx4_eq_int(struct mlx4_dev *dev, struct mlx4_eq *eq)
                case MLX4_EVENT_TYPE_PORT_CHANGE:
                        port = be32_to_cpu(eqe->event.port_change.port) >> 28;
                        if (eqe->subtype == MLX4_PORT_CHANGE_SUBTYPE_DOWN) {
-                               mlx4_dispatch_event(dev, MLX4_DEV_EVENT_PORT_DOWN,
+                               mlx4_dispatch_event(dev,
+                                                   MLX4_DEV_EVENT_PORT_DOWN,
                                                    port);
                                mlx4_priv(dev)->sense.do_sense_port[port] = 1;
+                               if (mlx4_is_master(dev))
+                                       /*change the state of all slave's port
+                                       * to down:*/
+                                       for (i = 0; i < dev->num_slaves; i++) {
+                                               mlx4_dbg(dev, "%s: Sending "
+                                                        "MLX4_PORT_CHANGE_SUBTYPE_DOWN"
+                                                        " to slave: %d, port:%d\n",
+                                                        __func__, i, port);
+                                               if (i == dev->caps.function)
+                                                       continue;
+                                               mlx4_slave_event(dev, i, eqe);
+                                       }
                        } else {
-                               mlx4_dispatch_event(dev, MLX4_DEV_EVENT_PORT_UP,
+                               mlx4_dispatch_event(dev,
+                                                   MLX4_DEV_EVENT_PORT_UP,
                                                    port);
                                mlx4_priv(dev)->sense.do_sense_port[port] = 0;
+
+                               if (mlx4_is_master(dev)) {
+                                       for (i = 0; i < dev->num_slaves; i++) {
+                                               if (i == dev->caps.function)
+                                                       continue;
+                                               mlx4_slave_event(dev, i, eqe);
+                                       }
+                               }
                        }
                        break;
 
@@ -187,7 +410,28 @@ static int mlx4_eq_int(struct mlx4_dev *dev, struct mlx4_eq *eq)
                                  eqe->event.cq_err.syndrome == 1 ?
                                  "overrun" : "access violation",
                                  be32_to_cpu(eqe->event.cq_err.cqn) & 0xffffff);
-                       mlx4_cq_event(dev, be32_to_cpu(eqe->event.cq_err.cqn),
+                       if (mlx4_is_master(dev)) {
+                               ret = mlx4_get_slave_from_resource_id(dev,
+                                       RES_CQ,
+                                       be32_to_cpu(eqe->event.cq_err.cqn)
+                                       & 0xffffff, &slave);
+                               if (ret && ret != -ENOENT) {
+                                       mlx4_dbg(dev, "CQ event %02x(%02x) on "
+                                                "EQ %d at index %u: could "
+                                                 "not get slave id (%d)\n",
+                                                 eqe->type, eqe->subtype,
+                                                 eq->eqn, eq->cons_index, ret);
+                                       break;
+                               }
+
+                               if (!ret && slave != dev->caps.function) {
+                                       mlx4_slave_event(dev, slave, eqe);
+                                       break;
+                               }
+                       }
+                       mlx4_cq_event(dev,
+                                     be32_to_cpu(eqe->event.cq_err.cqn)
+                                     & 0xffffff,
                                      eqe->type);
                        break;
 
@@ -195,13 +439,60 @@ static int mlx4_eq_int(struct mlx4_dev *dev, struct mlx4_eq *eq)
                        mlx4_warn(dev, "EQ overrun on EQN %d\n", eq->eqn);
                        break;
 
+               case MLX4_EVENT_TYPE_COMM_CHANNEL:
+                       if (!mlx4_is_master(dev)) {
+                               mlx4_warn(dev, "Received comm channel event "
+                                              "for non master device\n");
+                               break;
+                       }
+                       memcpy(&priv->mfunc.master.comm_arm_bit_vector,
+                              eqe->event.comm_channel_arm.bit_vec,
+                              sizeof eqe->event.comm_channel_arm.bit_vec);
+                       queue_work(priv->mfunc.master.comm_wq,
+                                  &priv->mfunc.master.comm_work);
+                       break;
+
+               case MLX4_EVENT_TYPE_FLR_EVENT:
+                       flr_slave = be32_to_cpu(eqe->event.flr_event.slave_id);
+                       if (!mlx4_is_master(dev)) {
+                               mlx4_warn(dev, "Non-master function received"
+                                              "FLR event\n");
+                               break;
+                       }
+
+                       mlx4_dbg(dev, "FLR event for slave: %d\n", flr_slave);
+
+                       if (flr_slave > dev->num_slaves) {
+                               mlx4_warn(dev,
+                                         "Got FLR for unknown function: %d\n",
+                                         flr_slave);
+                               update_slave_state = 0;
+                       } else
+                               update_slave_state = 1;
+
+                       spin_lock(&priv->mfunc.master.slave_state_lock);
+                       if (update_slave_state) {
+                               priv->mfunc.master.slave_state[flr_slave].active = false;
+                               priv->mfunc.master.slave_state[flr_slave].last_cmd = MLX4_COMM_CMD_FLR;
+                               priv->mfunc.master.slave_state[flr_slave].is_slave_going_down = 1;
+                       }
+                       spin_unlock(&priv->mfunc.master.slave_state_lock);
+                       queue_work(priv->mfunc.master.comm_wq,
+                                  &priv->mfunc.master.slave_flr_event_work);
+                       break;
                case MLX4_EVENT_TYPE_EEC_CATAS_ERROR:
                case MLX4_EVENT_TYPE_ECC_DETECT:
                default:
-                       mlx4_warn(dev, "Unhandled event %02x(%02x) on EQ %d at index %u\n",
-                                 eqe->type, eqe->subtype, eq->eqn, eq->cons_index);
+                       mlx4_warn(dev, "Unhandled event %02x(%02x) on EQ %d at "
+                                 "index %u. owner=%x, nent=0x%x, slave=%x, "
+                                 "ownership=%s\n",
+                                 eqe->type, eqe->subtype, eq->eqn,
+                                 eq->cons_index, eqe->owner, eq->nent,
+                                 eqe->slave_id,
+                                 !!(eqe->owner & 0x80) ^
+                                 !!(eq->cons_index & eq->nent) ? "HW" : "SW");
                        break;
-               }
+               };
 
                ++eq->cons_index;
                eqes_found = 1;
@@ -251,6 +542,36 @@ static irqreturn_t mlx4_msi_x_interrupt(int irq, void *eq_ptr)
        return IRQ_HANDLED;
 }
 
+int mlx4_MAP_EQ_wrapper(struct mlx4_dev *dev, int slave,
+                       struct mlx4_vhcr *vhcr,
+                       struct mlx4_cmd_mailbox *inbox,
+                       struct mlx4_cmd_mailbox *outbox,
+                       struct mlx4_cmd_info *cmd)
+{
+       struct mlx4_priv *priv = mlx4_priv(dev);
+       struct mlx4_slave_event_eq_info *event_eq =
+               &priv->mfunc.master.slave_state[slave].event_eq;
+       u32 in_modifier = vhcr->in_modifier;
+       u32 eqn = in_modifier & 0x1FF;
+       u64 in_param =  vhcr->in_param;
+       int err = 0;
+
+       if (slave == dev->caps.function)
+               err = mlx4_cmd(dev, in_param, (in_modifier & 0x80000000) | eqn,
+                              0, MLX4_CMD_MAP_EQ, MLX4_CMD_TIME_CLASS_B,
+                              MLX4_CMD_NATIVE);
+       if (!err) {
+               if (in_modifier >> 31) {
+                       /* unmap */
+                       event_eq->event_type &= ~in_param;
+               } else {
+                       event_eq->eqn = eqn;
+                       event_eq->event_type = in_param;
+               }
+       }
+       return err;
+}
+
 static int mlx4_MAP_EQ(struct mlx4_dev *dev, u64 event_mask, int unmap,
                        int eq_num)
 {
@@ -262,16 +583,16 @@ static int mlx4_MAP_EQ(struct mlx4_dev *dev, u64 event_mask, int unmap,
 static int mlx4_SW2HW_EQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
                         int eq_num)
 {
-       return mlx4_cmd(dev, mailbox->dma, eq_num, 0, MLX4_CMD_SW2HW_EQ,
-                       MLX4_CMD_TIME_CLASS_A,
+       return mlx4_cmd(dev, mailbox->dma | dev->caps.function, eq_num, 0,
+                       MLX4_CMD_SW2HW_EQ, MLX4_CMD_TIME_CLASS_A,
                        MLX4_CMD_WRAPPED);
 }
 
 static int mlx4_HW2SW_EQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
                         int eq_num)
 {
-       return mlx4_cmd_box(dev, 0, mailbox->dma, eq_num, 0, MLX4_CMD_HW2SW_EQ,
-                           MLX4_CMD_TIME_CLASS_A,
+       return mlx4_cmd_box(dev, dev->caps.function, mailbox->dma, eq_num,
+                           0, MLX4_CMD_HW2SW_EQ, MLX4_CMD_TIME_CLASS_A,
                            MLX4_CMD_WRAPPED);
 }
 
@@ -549,14 +870,16 @@ int mlx4_init_eq_table(struct mlx4_dev *dev)
        for (i = 0; i < mlx4_num_eq_uar(dev); ++i)
                priv->eq_table.uar_map[i] = NULL;
 
-       err = mlx4_map_clr_int(dev);
-       if (err)
-               goto err_out_bitmap;
+       if (!mlx4_is_slave(dev)) {
+               err = mlx4_map_clr_int(dev);
+               if (err)
+                       goto err_out_bitmap;
 
-       priv->eq_table.clr_mask =
-               swab32(1 << (priv->eq_table.inta_pin & 31));
-       priv->eq_table.clr_int  = priv->clr_base +
-               (priv->eq_table.inta_pin < 32 ? 4 : 0);
+               priv->eq_table.clr_mask =
+                       swab32(1 << (priv->eq_table.inta_pin & 31));
+               priv->eq_table.clr_int  = priv->clr_base +
+                       (priv->eq_table.inta_pin < 32 ? 4 : 0);
+       }
 
        priv->eq_table.irq_names =
                kmalloc(MLX4_IRQNAME_SIZE * (dev->caps.num_comp_vectors + 1 +
@@ -664,7 +987,8 @@ err_out_unmap:
                mlx4_free_eq(dev, &priv->eq_table.eq[i]);
                --i;
        }
-       mlx4_unmap_clr_int(dev);
+       if (!mlx4_is_slave(dev))
+               mlx4_unmap_clr_int(dev);
        mlx4_free_irqs(dev);
 
 err_out_bitmap:
@@ -689,7 +1013,8 @@ void mlx4_cleanup_eq_table(struct mlx4_dev *dev)
        for (i = 0; i < dev->caps.num_comp_vectors + dev->caps.comp_pool + 1; ++i)
                mlx4_free_eq(dev, &priv->eq_table.eq[i]);
 
-       mlx4_unmap_clr_int(dev);
+       if (!mlx4_is_slave(dev))
+               mlx4_unmap_clr_int(dev);
 
        for (i = 0; i < mlx4_num_eq_uar(dev); ++i)
                if (priv->eq_table.uar_map[i])
@@ -712,7 +1037,7 @@ int mlx4_test_interrupts(struct mlx4_dev *dev)
 
        err = mlx4_NOP(dev);
        /* When not in MSI_X, there is only one irq to check */
-       if (!(dev->flags & MLX4_FLAG_MSI_X))
+       if (!(dev->flags & MLX4_FLAG_MSI_X) || mlx4_is_slave(dev))
                return err;
 
        /* A loop over all completion vectors, for each vector we will check